Expected Credit Loss

(ECL): Issues & Studies

Compiled by
Gaby Frangieh
Risk Management, Finance and Banking – Senior Advisor
April 2025

https://www.linkedin.com/in/gaby-frangieh-1873aa11/
Expected Credit Loss (ECL)
Issues and Studies
Expected Credit Loss (ECL)
Issues and Studies

Compilation Contents
1. The Procyclicality of Expected Credit Loss Provisions by Jorge Abad and
Javier Suarez, May 2018

2. Practical insights on implementing IFRS 9 and CECL by Deloitte

3. Current Expected Credit Loss (CECL) Model and Banks’ External

Information Environment by Samuel B. Bonsall IV, Brent A. Schmidt and Biqin
Xie, February 2025

4. When are expected credit losses decision-useful and new to investors?

Evidence from CECL adoption by Kurt H. Gee, Jed J. Neilson, Brent A. Schmidt,
and Biqin Xie, March 2025

5. Estimating Lifetime Expected Credit Losses Under IFRS 9 by Xin Xu, August
2016

6. Still “Too Much, Too Late”: Provisioning for Expected Loan Losses by
Roman Goncharenkoa and Asad Rauf, October 2024

7. The Decision Usefulness of Current Expected Credit Losses: Users’ Views

about the Current Expected Credit Losses Model by Jordan M. Bable,
Christopher I. L. Wong, and Michael J. Wynes, April 2025

8. Calculating Lifetime Expected Loss for IFRS 9: Which Formula is Correct?

by Bernd Engelmann, April 2018

9. Guideline on the management of expected credit losses by AMF France,

June 2024
 The Procyclicality of Expected Credit Loss
Provisions
Jorge Abad
CEMFI, Casado del Alisal 5 28014 Madrid, Spain. Email: jorge.abad@cem….edu.es

Javier Suarez
CEMFI, Casado del Alisal 5 28014 Madrid, Spain. Email: suarez@cem….es (contact author)

May 2018

Abstract
The Great Recession has pushed accounting standards for banks’loan loss provision-
ing to shift from an incurred loss approach to an expected credit loss approach. IFRS 9
and the incoming update of US GAAP imply a more timely recognition of credit losses
but also greater responsiveness to changes in aggregate conditions, which raises pro-
cyclicality concerns. This paper develops and calibrates a recursive ratings-migration
model to assess the impact of di¤erent provisioning approaches on the cyclicality of
banks’pro…ts and regulatory capital. The model is used to analyze the e¤ectiveness of
potential policy responses to the procyclicality problem.

Keywords: credit loss allowances, expected credit losses, incurred losses, rating migra-
tions, procyclicality.

JEL codes: G21, G28, M41

This paper has been prepared for the Sveriges Riksbank, De Nederlandsche Bank, and Deutsche Bun-
desbank Annual Macroprudential Conference in Stockholm, 15-16 June 2018. It is a signi…cantly revised
and extended version of “Assessing the Cyclical Implications of IFRS 9: A Recursive Model,” which bene-
…ted from comments received from Alejandra Bernad, Xavier Freixas, David Grünberger, Andreas P…ngsten,
Malcolm Kemp, Luc Laeven, Christian Laux, Dean Postans, Antonio Sánchez, Rafael Repullo, Josef Zech-
ner, and other participants at meetings of the ESRB Task Force on Financial Stability Implications of the
Introduction of IFRS 9 and the ESRB Advisory Scienti…c Committee, and presentations at the Banco de Es-
paña, Banco de Portugal, CEMFI, De Nederlandsche Bank, EBA Research Workshop, ESSFM 2017, Finance
Forum 2017, Magyar Nemzeti Bank, SAEe2017, and University of Zurich. The contents are the exclusive
responsibility of the authors.
1 Introduction
Banks’delayed recognition of credit losses under the incurred loss (IL) approach to loan loss
provisioning was argued to contribute to the severity of the Global Financial Crisis (Financial
Stability Forum, 2009). By provisioning “too little, too late,”it might have prevented banks
from being more cautious in good times and reduced the pressure for prompt corrective
action in bad times. Based on this diagnosis, the G-20 called for a more forward-looking
approach. As a result, the International Accounting Standards Board and the US Financial
Accounting Standards Board developed reforms, namely IFRS 9 (entered into force in 2018)
and an update of US GAAP (scheduled for 2021), which, with some di¤erences, coincide in
adopting an expected credit loss (ECL) approach to provisioning.1
Under the new approach, loan loss provisions are intended to represent best unbiased
estimates of the discounted credit losses expected to emerge over some speci…ed horizons.
In the case of IFRS 9, the horizon depends on the credit quality of the exposures, varying
from one year for those without deteriorated credit quality (stage 1) to the residual lifetime
of the credit instrument for exposures with deteriorated credit quality (stage 2) or already
impaired (stage 3). In contrast, the so-called current expected credit loss (CECL) model
envisaged by US GAAP opts for using the residual lifetime horizon for all exposures. The
general perception is that the ECL approach will increase the reliability of bank capital as
a measure of solvency and facilitate prompt corrective action (Cohen and Edwards, 2017).
There are concerns, however, that, absent the capacity of banks to anticipate adverse
shifts in aggregate conditions su¢ ciently in advance, the point-in-time (PIT) nature of the
estimates of ECLs might imply a more abrupt deterioration of pro…ts and capital when
the economy enters recession or a crisis starts, as it will be then and not before when the
bulk of the implied future credit losses will be recognized (European Systemic Risk Board,
2017). The fear is that banks’or markets’reaction to such an increase (or to its impact on
pro…ts and regulatory capital) cause or amplify asset sales or a credit crunch, and end up
producing negative feedback e¤ects on the evolution of the economy, making the contraction
more severe.
1
See International Accounting Standards Board (2014) and Financial Accounting Standards Board (2016)
for details.

1
 This paper develops a recursive model with which to compare the impact on pro…ts
and capital of IFRS 9 and CECL relative to their less forward-looking alternatives (namely,
IL and the one-year expected loss approach behind the internal-ratings based approach to
capital regulation). We assess the importance of the procyclical e¤ects and the e¤ectiveness
of several policy options for their containment, including the reinforcement or active use of
the loss absorbing capital bu¤ers introduced in Basel III.
We address the modeling of ECLs in the context of a ratings-migration model with a
compact description of credit risk categories, the economic cycle represented as a Markov
chain, and loan maturity modeled as random. Each of these modeling strategies has a well-
established tradition in economics and …nance and their combination prevents us from having
to keep track of loan vintages or a more complex state space, producing a model which is
overall highly tractable.2 The model is calibrated to capture the evolution of credit risk in
a typical portfolio of European corporate loans over the business cycle and to compare the
cyclical behavior of impairment allowances, pro…t or loss (P/L), and common equity tier 1
(CET1) across the various provisioning approaches.3 For the calibration, we use evidence
on the sensitivity of rating migration matrices and credit loss parameters to business cycles,
as in Bangia et al. (2002). We …nd that the more forward-looking methods of IFRS 9 and
CECL imply signi…cantly larger impairment allowances, sharper on-impact responses to the
arrival of an average recession than the old IL and one-year expected loss used by IRB banks,
and the quicker recovery of P/L and CET1 after the initial impact.
The arrival of a typical recession implies on-impact increases in IFRS 9 and CECL provi-
sions equivalent to about a third of a bank’s fully loaded capital conservation bu¤er (CCB)
or, equivalently, about twice as large as those implied by the IL approach.4 This suggests
2
Ratings-migration models are extensively used in credit risk modeling (see Trueck and Rachev, 2009,
for an overview) and Grünberger (2012) provides an early application of them to the analysis of IFRS 9.
Hamilton (1989) showed the possibility of representing the cyclical phases and turning points identi…ed in
business cycle dating (e.g. by the NBER) using a binary Markov chain, and Bangia et al. (2002) and Repullo
and Suarez (2013), among others, have used such representation in applications regarding ‡uctuations in
credit risk. The modeling of debt maturity as random started with Leland and Toft (1996) and has been
recently applied in a banking context by He and Xiong (2012), He and Milbradt (2014), and Segura and
Suarez (2017), among others.
3
The case of a bank fully specialized in European corporate loans must be interpreted as a “labora-
tory case” with which to run “controlled experiments” about the performance of the di¤erent provisioning
methods.
4
Under Basel III, banks’reporting earnings must retain them until reaching a CCB (or bu¤er of capital

2
that the di¤erential impact under the new approaches is sizeable but still suitably absorbable
if banks’ CCBs are su¢ ciently loaded when the shock hits. As we show in extensions in-
cluded in an appendix, the results depend on the degree to which turning points imply bigger
or smaller surprises relative to what banks have anticipated in advance. We show that the
sudden arrival of a contraction that is anticipated to be more severe or longer than average
will tend to produce sharper responses, while forecasting a recession one or several years in
advance would allow to signi…cantly smooth away its impact on P/L and capital.
According to our analysis, if banks fail to anticipate turning points well in advance or
to adopt additional precautions during good times, the more forward-looking provisioning
methods may paradoxically mean that banks experience more sudden falls in regulatory
capital right at the beginning of contractionary phases of the business cycle. Banks might
accommodate these e¤ects by consuming the capital bu¤ers accumulated during good times,
by cutting dividends or by issuing new equity. However, when confronted with these choices,
banks often undertake at least part of the adjustment by reducing their risk-weighted assets
(RWAs), for example by cutting the origination of new loans, selling some assets or rebal-
ancing towards safer ones.5 This gives rise to the procyclical concerns emphasized in the
title of the paper.
As discussed in the section devoted to the policy analysis, a full assessment of the pro-
cyclical impact of the new provisions would require taking care of multiple moving parts,
including endogenous decisions and general equilibrium e¤ects. In this paper we only look
at some …rst-round, partial equilibrium e¤ects that, we think, may help gauge the direction
and intensity of the procyclical e¤ects. We measure such intensity through the unconditional
annual probability with which the bank described in the model needs to raise new capital
to avoid violating its minimum regulatory capital requirements. We examine potentially
mitigating policies such as increasing the target size of the CCB, actively using the coun-
tercyclical capital bu¤er (CCyB) of Basel III, introducing prudential bu¤ers based on stress
on top of the regulatory minimum) equivalent to 2.5% of their risk-weighted assets.
5
See, for example, Mésonnier and Monks (2015), Aiyar, Calomiris, and Wieladek (2016), Behn, Haselmann
and Wachtel (2016), Gropp et al. (2016), and the references therein. The evidence in these papers is
consistent with average bank responses to the ESRB Questionnaire on Assessing Second Round E¤ects that
accompanied the EBA stress test in 2016. The questionnaire examined the way in which banks would expect
to reestablish their desired levels of capitalization after exiting the adverse scenario.

3
test results, and cyclically smoothing the inputs used in the estimation of ECLs. For reasons
explained in the discussion of the results, we …nd that introducing a CCB add-on (possibly
calibrated on the basis of stress test results) is not only the simplest but also the most ef-
fective of the compared policy options (that is, the one able to achieve a more signi…cant
reduction in the unconditional probability of having to recapitalize the bank).
We are aware of just a few papers trying to assess the cyclical behavior of impairment
allowances under the new provisioning approaches, none of which addresses the analysis of
potential mitigating policies. Cohen and Edwards (2017) develop such analysis from a top-
down perspective and relying on the historical evolution of aggregate bank credit losses in a
number of countries. Chae et al. (2018) use a more bottom-up approach based on credit loss
data for …rst-lien mortgages originated in California between years 2002 and 2015. Krüger,
Rösch, and Scheule (2018) use historical simulation methods on portfolios constructed using
bonds from Moody’s Default and Recovery Database.
The …rst two papers posit the conclusion that, if banks are capable to perfectly foresee
the incoming credit losses several years in advance, the new approaches will show smaller
spikes in impairment allowances than the incurred loss approach when the economy situation
deteriorates.6 This is consistent with what we obtain in the extension in which banks can
anticipate the turning points in the evolution of the economy several periods in advance,
albeit such capacity is not consistent with the di¢ culties faced by econometricians and
professional forecasters in predicting recessions.7 The results in Krüger, Rösch, and Scheule
(2018) are closer to ours and lead the authors to conclude that the new provisioning rules
“will further increase the procyclicality of bank capital requirements”(p. 114).
The paper is organized as follows. Section 2 describes the model. Section 3 develops
formulas for measuring impairment losses under the various provisioning approaches and
for assessing their e¤ects on P/L and CET1. Section 4 calibrates the model and uses it to
analyze the e¤ects of the arrival of a typical recession under the various measures. Section 5
6
Chae et al. (2017) also show that, if instead the inputs of the credit loss model are predicted using
a high-order autoregressive (AR) model based on information available at the time of producing the ECL
estimate, the implied provisions (ALLL) spike only once the housing crisis starts. In their words, the “AR
forecast is not able to forecast the in‡ection point of home prices which leads to large increases in ALLL in
early 2009.”
7
See Section B.4 of Appendix B for further discussion.

4
discusses the …ndings on procyclicality and analyzes the e¤ectiveness of various policies that
might be used to mitigate the problem. Section 6 concludes the paper. Appendices A, B,
and C contain further details about the model, the calibration, and several complementary
results.

2 The model
We consider a bank operating in an in…nite-horizon discrete-time economy in which dates
(regarded year-end dates for accounting reporting purposes) are denoted by t: The bank’s
assets consist solely of a portfolio of loans whose individual credit risk, in the absence of
aggregate risk, is fully described by a rating category j. The dynamics of loan origination,
rating migration, default, and maturity of the loans that make up the bank’s loan portfolio are
described in the …rst subsection below. The assumptions on the capital structure of the bank,
the regulatory environment, and the pricing of the loans appear in the second subsection.
For expositional clarity, we …rst present the main assumptions and formulas of the model
in a version without aggregate risk. Then in the third subsection below we comment on
the way in which the complete model incorporates aggregate risk. The notationally more
intensive equations for the complete model are presented in Appendix A.

2.1 The bank’s loan portfolio

Each of the loans held by the bank can belong to one of three credit rating categories:
standard (j=1), substandard (j=2) or non-performing (j=3). We denote the measure of
loans of each category in the portfolio of the bank at date t as xjt : In each date, the bank
originates a measure e1t > 0 of standard loans with a principal normalized to one, and an
endogenous contractual interest payment per period equal c: Each loan’s principal remains
constant and equal to one up to maturity, which for performing loans (j=1; 2) is assumed to
8
occur randomly and independently at the end of each period with a constant probability j.

This analytically convenient assumption implies that, conditional on remaining in rating j,

a loan’s expected life span is 1= j and that, by the law of large numbers, the stream of cash
8
Allowing for 1 6= 2 may help capture the possibility that longer maturity loans get early redeemed
with di¤erent probabilities depending on their credit quality.

5
‡ows due to maturing loans is very similar to the one that would emerge with a portfolio of
perfectly-staggered …xed-maturity loans.
The tree in Figure 1 summarizes the contingencies over the life of a loan.

δ3/2 resolution payoff 1–λ

PDj
1 − δ3/2 continuation with j’=3

δj
1–PDj
full repayment payoff c + 1

j=1,2
δ3/2 resolution payoff 1–λ

PDj
1 − δ3/2 continuation with j’=3
1 – δj

a1j c + continuation with j’=1

a2j
c + continuation with j’=2

δ3 resolution payoff 1–λ

j=3

1−δ3 continuation with j’=3

Figure 1. Possible transitions of a loan rated j: Possible contingencies

between two dates and their implications for payo¤s and continuation value.
Variables on each branch describe marginal conditional probabilities.

Non-performing loans (NPLs, j=3) are also assumed to be resolved randomly and inde-
pendently with probability j per period, producing a terminal payo¤ 1 ; so is a loan’s
loss rate at resolution. NPLs pay no interest and never return to the performing categories,
so they accumulate in category j=3 up to their resolution.9
Performing loans at t, irrespectively of their maturity happening at t + 1 or not, default
independently with probability P Dj at t + 1: Each loan that defaults between t and t + 1
enters the stock of NPLs with an independent probability 1 3 =2 and gets resolved within
9
For calibration purposes, it is possible to account for potential gains from the unmodeled interest accrued
while in default or from returning to performing categories by adjusting the loss rate :

6
 10
the same period with the complementary probability 3 =2: Maturing loans that do not
default pay back their principal of one plus interest c.
Performing loans at t that do not mature at t + 1 pay interest c, migrate to rating
i 6= j (i=1,2) independently with probability aij ; and stay in rating j independently with
probability ajj = 1 aij P Dj :

2.2 The bank’s capital structure and loan pricing

The bank originating and holding the loans is assumed to be perfectly competitive, fully
solvent, and with access to funding from risk-neutral investors who face an opportunity cost
of funds between any two periods constant and equal to r: To keep things simple, we assume
that the bank …nances its activity with one-period debt dt and equity (or more technically,
CET1) kt ; and is subject to capital regulation as per a bank operating under the internal
ratings-based (IRB) approach of Basel III.11
We model the evolution of the bank’s capital structure by assuming that the bank is a
minimizer of the equity funding kt required to support its loan portfolio under the prevailing
capital regulation and provisioning regime.12 The latter determines the loan loss allowances
LLt associated with the bank’s loan portfolio at each date t.
In Section 3 we provide expressions for LLt under di¤erent provisioning approaches as
well as further details on the law of motion of kt ; and its implications for the dividend
payments and the recapitalization needs of the bank under our assumptions.

2.3 Adding aggregate risk

We introduce aggregate risk in the model by considering an aggregate state variable st whose
evolution a¤ects the key parameters governing loan portfolio dynamics and credit losses in
the structure described above. For our baseline results, we assume that st follows a Markov
10
We divide 3 by two to re‡ect the fact that, if loans default uniformly during the period between t
and t+1, they will, on average, have just half a period to be resolved. Given that the calibration relies on
one-year periods and the resolution rate is large, this re…nement is important to realistically describe NPL
dynamics. The model could easily accommodate alternative assumptions on same-period resolutions.
11
The case in which the bank follows the standardized approach (SA) is analyzed in Appendix C.
12
We could rationalize banks’ minimization of its CET1 as the result of imperfections that make equity
…nancing more costly than debt …nancing. However, for simplicity, we formally consider the limit case where
the excess cost of equity …nancing goes to zero so that loan pricing (as described below) does not depend on
the bank’s capital structure.

7
chain with two states s=1,2 and time-invariant transition probabilities ps0 s =Prob(st+1 =
s0 jst = s): We interpret s=1 and s=2 as identifying expansion and contraction periods,
respectively.

3 Analysis
In this section we develop the formulas describing the dynamics of the bank’s loan portfolio,
the measurement of impairment losses under the various provisioning approaches, the pricing
of the loans, and the evolution of the bank’s pro…ts and regulatory capital, under the model
assumptions introduced in the previous section.

3.1 Portfolio dynamics

By the law of large numbers, the evolution of the loan portfolio can be represented by the
following di¤erence equation:
x t = M xt 1 + et (1)

where 0 1
x1t
xt = @ x2t A (2)
x3t
is the vector that describes the loans in each rating category j=1,2,3;
0 1 0 1
m11 m12 m13 (1 1 )a11 (1 2 )a12 0
M = @ m21 m22 m23 A = @ (1 )a
1 21 (1 2 )a22 0 A (3)
m31 m32 m33 (1 3 =2)P D1 (1 3 =2)P D2 (1 3)

is the matrix that accounts for the migrations across categories of the non-matured, non-
resolved loans, and 0 1
e1t
et = @ 0 A (4)
0
accounts for the new loans originated at each date, which we write re‡ecting the fact that,
as previously speci…ed, all loans have rating j=1 at origination.
When computing some of the moments relevant for the calibration of the model, we weight
each rating category by its share in the steady-state portfolio x that would asymptotically
be reached, in the absence of aggregate risk, if the amount of newly originated loans is equal

8
at all dates (et = e for all t). Such steady-state portfolio can be obtained as the vector that
solves:
x = M x + e , (I M )x = e; (5)

that is,
x = (I M ) 1 e: (6)

3.2 Measuring impairment losses

In the following subsections we provide formulas for impairment allowances, LLt ; under
six di¤erent approaches: incurred losses, ILt ; discounted one-year expected losses, EL1Y
t ;

prudential expected losses under the IRB approach, ELIRB

t ; discounted lifetime expected
losses, ELLT CECL
t ; current expected credit losses, ELt ; and IFRS 9 impairment allowances,
ELIF
t
RS9
: De…ning EL1Y
t and ELLT
t is useful to understand the mixed-horizon approach in
IFRS 9 and to compare the size of the various measures. Eventually our quantitative analysis
will focus on ILt ; ELIRB
t ; ELCECL
t ; and ELIF
t
RS9
:

3.2.1 Incurred losses

Under the narrowest interpretation, the incurred loss approach prescribes the provisioning,
on an expected loss basis, of the exposures for which there is clear evidence of impairment,
which in our model would be the NPLs in x3t : Thus, provisions at year t under the IL
approach would be
ILt = x3t ; (7)

since the loss rate is the expected loss given default (LGD) of the bank’s NPLs at date t:13

3.2.2 Discounted one-year expected losses

As a reference for the criterion used for the measurement of the impairment allowances
applied to stage 1 exposures under IFRS 9, we de…ne the one-year discounted expected
losses of the bank as
EL1Y
t = [ (P D1 x1t + P D2 x2t ) + x3t ] (8)
13
Under our assumptions, the losses associated with loans defaulted between dates t 1 and t which are
resolved within such period, ( 3 =2)(P D1 x1t 1 + P D2 x2t 1 ), are directly recorded in the P/L of year t and
do not enter ILt .

9
where = 1=(1 + c) is the discount factor based on the contractual interest rate of the loans,
c: Under this approach, impairment allowances for loans performing at t (rated j=1, 2) are
computed as the discounted expected losses due to default events expected to occur within
the immediately incoming year. This approach is forward-looking, but the forecasting horizon
is limited to one year. Instead, for NPLs (j=3), the default event has already happened and
the allowances equal the (non-discounted) expected LGD of the loans, exactly as in ILt .
In matrix notation, which will be useful when comparing the di¤erent impairment al-
lowance measures later on, EL1Y
t can also be expressed as

EL1Y
t = ( bxt + x3t ) ; (9)

where
b = (P D1 ; P D2 ; 0): (10)

3.2.3 Prudential expected losses under the IRB approach

When the Basel Committee on Banking Supervision (BCBS) introduced the IRB approach
to capital requirements, the idea was to set capital requirements so as absorb the bank’s
unexpected credit losses (over a one year horizon) while assuming that the corresponding
expected losses would have been already recognized (and subtracted from available capital)
via provisions. This led to the introduction of a prudential notion of expected losses which
can be formally expressed as

ELIRB
t = [P D1 x1t + P D2 x2t + x3t ] = (bxt + x3t ) ; (11)

where the only di¤erence with respect to (9) is the absence of discounting. As described in
detail in Appendix A, in the presence of aggregate risk further di¤erences appear due to the
fact that the BCBS speci…es that instead of PIT estimates of P Dj and ; (11) must be fed
with so-called through-the-cycle (TTC) PDs and downturn LGDs.

3.2.4 Discounted lifetime expected losses

To capture discounted lifetime expected losses in the way IFRS 9 stipulates for stage 2
exposures, we de…ne

2 3 4
ELLT
t = b xt + M xt + M 2 xt + M 3 xt + ::: + x3t ; (12)

10
which considers the discounted expected losses due to default events occurring over the whole
residual lifetime of the existing loans. The formula above re‡ects that the losses expected
from currently performing loans at any future year t + ; with = 1; 2; 3::: can be found as
1 1
bM xt ; where b contains the relevant one-year-ahead PDs (see (10)) and M xt gives
the projected composition of the portfolio at each future year t + 1. It also re‡ects that
the allowance for the NPLs simply equals the expected LGD of the a¤ected loans, as in the
other approaches.
Equation (12) can also be expressed as

2 3
ELLT
t = b(I + M + M2 + M 3 + :::)xt + x3t ; (13)

where the parenthesis is the in…nite sum of a geometric series of matrices. Thus, we can
write ELLT
t as
ELLT
t = ( bBxt + x3t ) ; (14)

where
B = (I M ) 1: (15)

3.2.5 Current expected credit losses (CECL)

Under the CECL approach all performing loans (j=1,2) are provisioned on a discounted
lifetime basis but, di¤erently from IFRS 9, the discount rate is not the contractual interest
rate of each loan but a reference risk free-rate. This gives:

ELCECL
t = ( r bBxt + x3t ) ; (16)

where the only di¤erence with respect to (14) is at the use of r = 1=(1 + r) rather than
= 1=(1 + c):

3.2.6 Impairment allowances under IFRS 9

IFRS 9 adopts, for performing loans, a mixed-horizon approach that combines the discounted
one-year and lifetime expected loss approaches described above. Speci…cally, the allowances
for loans that have not su¤ered a signi…cant increase in credit risk since origination (“stage
1”loans or, in our model, the loans in x1t ) are computed as in EL1Y
t while the allowances for

11
performing loans with deteriorated credit quality (“stage 2”loans or, in our model, the loans
in x2t ) are computed as in ELLT
t . Finally, for NPLs (“stage 3” loans or, in our model, the

loans in x3t ), the allowance equals the (non-discounted) expected LGD, as under all the other
approaches. Combining the formulas obtained in (9) and (14), the impairment allowances
under IFRS 9 can be described as
0 1 0 1
x1t 0
ELIF
t
RS9
= b@ 0 A + bB @ x2t A + x3t : (17)
0 0

3.2.7 Comparing the various impairment measures

The de…nitions provided above clearly imply that

ILt EL1Y
t ELIF
t
RS9
ELLT
t ELCECL
t ; (18)

where the last inequality follows from the fact that c r 0; which means r : Regarding
ELIRB
t ; it is obviously the case that ELIRB
t ILt and that the absence of discounting would,
without aggregate risk, also imply ELIRB
t EL1Y
t ; while the comparison with other measures

would be generally ambiguous. However, as further described in Appendix A, once aggregate

risk is introduced, the usage of TTC PDs and downturn LGDs also produces ambiguity with
respect to the comparison with EL1Y
t and the remaining expected-loss-based measures.

3.3 Implications for pro…ts and the dynamics of regulatory capital

Under our assumptions, the bank’s only assets are its loans and its only liabilities are one-
period debt, dt ; loan loss allowances, LLt ; and CET1, kt : So the bank’s balance sheet at the
end of any period t can be described as

x1t dt
x2t LLt (19)
x3t kt

with the law of motion of xt described by (1) and the law of motion of kt given by

kt = kt 1 + P Lt divt +recapt ; (20)

12
where P Lt is the result of the P/L account at the end of period t; divt 0 are cash dividends
paid at the end of period t; and recapt 0 are injections of CET1 at the end of period t:
Under these assumptions, the dynamics of dt can be recovered residually from the balance
sheet identity, dt = j=1,2,3 xjt LLt kt :
The result of the P/L account can in turn be written as
( ) !
X 3
X
P Lt = c(1–P Dj ) P Dj xjt 1 – 3 x3t 1 –r xjt 1 –LLt 1 –kt 1 – LLt ;
j=1;2
2 j=1;2;3
(21)
where the …rst term contains the income from performing loans net of realized losses on
defaulted loans resolved during period t; the second term is the interest paid on dt 1 ; and
the third term is the variation in credit loss allowances between periods t 1 and t: So, other
things equal, any increase in LLt has a negative contemporaneous impact on P Lt and,
through (20), on the capital available for the bank to operate in the subsequent period.
Under the assumption that the bank minimizes the use of CET1 subject to the prescrip-
tions of Basel III, its dividends and equity injections are determined as

divt = max[(kt 1 + P Lt ) 1:3125k t ; 0]; (22)

recapt = max[k t (kt 1 + P Lt ); 0]: (23)

To explain these expressions, notice that kt 1 + P Lt would be the inertial CET1 of the bank
at date t without discretionary adjustments (divt =recapt = 0). However, the bank need
to operate with CET1 of at least k t at all times, so it has to recapitalize whenever such
minimum is otherwise not reached (equation (23)). On the other hand, dividends cannot
be paid until the barrier k t = 1:3125k t given by the fully loaded CCB is not reached.14 So
e¤ectively the bank manages its CET1 within the bands k t and k t ; thus following a simple
sS-rule based entirely on existing capital regulations.15
14
The CCB requires the bank to retain pro…ts, whenever feasible, until reaching a fully loaded bu¤er equal
to 2.5% of its RWAs. Regulatory RWAs equal 12.5 (or 1/0.08) times the bank’s minimal required capital k t :
Thus a fully loaded CCB amounts to a multiple 0:025 12:5 = 0:3125 of k t :
15
The working of the sS rule proposed here implicitly assumes the absence of …xed costs associated with
the raising of new equity. If such costs were to be introduced, the optimal rule would imply, as in Fischer,
Heinkel, and Zechner (1989), discrete recapitalizations to an endogenous level within the bands if the lower
band were to be otherwise passed.

13
 Finally, for corporate loan portfolios operated under the IRB approach, BCBS (2017,
paragraph 53) establishes that the minimum capital requirement is
X
k IRB
t = j xjt ; (24)
j=1;2

with
" ! #
1
1 + [(1= j ) 2:5]mj (P Dj ) + corj0:5 1
(0:999)
j = P Dj ; (25)
1 1:5mj (1 corj )0:5

where mj = [0:11852 0:05478 ln(P Dj )]2 is a maturity adjustment coe¢ cient, ( ) denotes
the cumulative distribution function of a standard normal distribution, and corj is a corre-
lation coe¢ cient …xed as corj = 0:24 0:12(1 exp( 50P Dj ))=(1 exp( 50)):16 So the
baseline results below will be based on assuming k t = k IRB
t :17

3.4 Loan rates under competitive pricing

Taking advantage of the recursivity of the model, we can obtain the bank’s ex-coupon value
of loans rated j at any given date, vj ; by solving the following system of Bellman-type
equations:

vj = [(1 P Dj )c + (1 P Dj ) j + P Dj ( 3 =2)(1 ) + m1j v1 + m2j v2 + m3j v3 ] ; (26)

for j=1; 2, and

v3 = [ 3 (1 ) + (1 3 )v3 ] ; (27)

where = 1=(1 + r) is the discount factor of the risk neutral bank.18 Intuitively, the square
brackets in (26) and (27) contain the payo¤s and continuation value that a loan rated j=1; 2
16
In (25) we measure the maturity of performing loans as the expected residual maturity 1= j implied by
our formulation.
17
Our analysis abstracts from the existence of “regulatory …lters” addressing possible discrepancies be-
tween accounting provisions (the relevant LLt ) and prudential expected losses such as ELIRB t : These …lters
currently establish that if ELIRB
t exceeds LL t ; the di¤erence, ELIRB
t LL t , must be subtracted from CET1.
By contrast, if ELIRB
t LLt < 0; the di¤erence can be added back as tier 2 capital up to a maximum of
0.6% of the bank’s credit RWAs. Therefore, we assess the impact of alternative provisioning methods on
bank capital dynamics in the case prudential regulators accept the corresponding method as the one used
to de…ne prudential expected losses too.
18
For calibration purposes, the discount rate r does not need to equal the risk-free rate. One might adjust
the value of r to re‡ect the marginal weighted average costs of funds of the bank or even an extra element
capturing (in reduced form) a mark-up applied on that cost if the bank is not perfectly competitive.

14
or j=3, respectively, will produce in the contingencies that, in each case, can occur one
period ahead (weighted by the corresponding probabilities).
In (26), contingencies producing payo¤s are, in order of appearance, the payment of
interest on non-defaulted loans, the repayment of principal by the non-defaulted loans that
mature, and the recovery of terminal value on defaulted loans resolved within the period.
The last three terms contain continuation value under the three rating categories that can
be reached one period ahead. Similarly, (27) re‡ects the terminal value recovered if an NPL
is resolved within the period and the continuation value kept otherwise.
Under perfect competition, the value of extending a unit-size loan of standard quality
(j=1) must equal the value of its principal, v1 = 1; so that the bank obtains zero net present
value from its origination. Thus we obtain the endogenous contractual interest rate of the
loan, c; as the one that solves this equation.

4 Baseline quantitative results

4.1 Calibration

Table 1 describes the calibration of the baseline model under a parameterization intended
to represent a typical portfolio of corporate loans issued by EU banks. Given the absence
of detailed publicly available microeconomic information on such a portfolio, the calibration
relies on matching aggregate variables taken from recent European Banking Authority (EBA)
reports and European Central Bank (ECB) statistics using rating migration and default
probabilities consistent with the Global Corporate Default reports produced by Standard
& Poor’s (S&P) over the period 1981-2015.19 The probabilities of default (PDs) and yearly
probabilities of migration across our standard and substandard categories are extracted from
S&P rating migration data using the procedure described in Appendix B. These probabilities
are consistent with the alignment of our standard category (j=1) with ratings AAA to BB
in the S&P classi…cation and our substandard category (j=2) with ratings B to C.
In a nutshell, to reduce the 7 7 rating-migration probabilities and the seven PDs
extracted from S&P data to the 2 2 migration probabilities and two PDs that appear in
19
We use reports equivalent to S&P (2016) published in years 2003 and 2005-2016, which provide the
relevant information for each of the years between 1981 and 2015.

15
matrix M (equation (3)), we calculate weighted averages that take into account the steady-
state composition that the loan portfolio would have under its 7-ratings representation in
the absence of aggregate risk. To obtain this composition, we assume that loans have an
average duration of 5 years (or 1 = 2 =0.2) as re‡ected in Table 1, that they have a rating
BB at origination, and that they then evolve (through improvements or deteriorations in
their credit quality before defaulting or maturing) exactly as in our model, but with the
seven non-default rating categories in the original S&P data.

Table 1
Calibration of the baseline model
Parameters without variation with the aggregate state
Banks’discount rate r 1.8%
Persistence of the expansion state (s=1) p11 0.852
Persistence of the contraction state (s=2) p22 0.5
Expansion Contraction
Parameters that can vary with aggregate arrival state (s0 =1) (s0 =2)
Yearly probability of migration 1 ! 2 if not maturing a21 6.16% 11.44%
Yearly probability of migration 2 ! 1 if not maturing a12 6.82% 4.47%
Yearly probability of default if rated j=1 P D1 0.54% 1.91%
Yearly probability of default if rated j=2 P D2 6.05% 11.50%
Loss given default 30% 40%
Average time to maturity if rated j=1 1= 1 5 years 5 years
Average time to maturity if rated j=2 1= 2 5 years 5 years
Yearly probability of resolution of NPLs 3 44.6% 44.6%
Newly originated loans per period (all rated j=1) e1 1 1

Following the model formulation with aggregate risk described in Appendix A and the
interpretation of the aggregate state as describing expansion vs. contraction periods, we allow
for state-variation in the probabilities of loans migrating across rating categories and into
default in a way consistent with the historical correlation between those variables (as observed
in S&P rating-migration data) and the US business cycle as dated by the National Bureau
of Economic Research (NBER).20 The dynamics of the aggregate state as parameterized in
Table 1 imply that the average duration of expansion and contraction periods is 6.75 years
and 2 years, respectively, meaning that the system spends about 77% of the time in state
s=1. Expansions are characterized by signi…cantly smaller PDs among both standard and
substandard loans than contractions. During a contraction, the probability of standard loans
20
See http://www.nber.org/cycles.html.

16
being downgraded (or, under IFRS 9, moved into stage 2) is almost double than during an
expansion and the probability of substandard loans recovering standard quality (or returning
to stage 1) is reduced by about one-third. See Section B.2 of Appendix B for further details.
Under this calibration, the unconditional average yearly PDs for our standard and sub-
standard categories are 0.9% and 7.3%, respectively. As shown in Table 2, given the com-
position of the ergodic portfolio, the unconditional average annual loan default rate equals
1.9%, which is below the average 2.5% PD for non-defaulted corporate exposures that EBA
(2013, Figure 12) reports for the period from the …rst half of 2009 to the second half of 2012
for a sample of EU banks operating under the IRB approach. Conditional on being in an
expansion and in a contraction, our calibration implies average annual loan default rates for
performing loans of 1.36% and 3.43%, respectively.
We also allow for state variation in the loss rate experienced at resolution. We set
equal to 40% during contractions and 30% during expansions. This is consistent with
the cyclical evolution of average realized LGDs on European unsecured loans to large non-
…nancial corporations reported by Brumma and Winckle (2017, Exhibit 3). An LGD of
40% in contractions is also consistent with the (downturn) LGD prescribed by BSBC (2017,
paragraph 70) for unsecured corporate loans under the foundation IRB approach. The
cyclical variation is also consistent with that documented by Bruche and González-Aguado
(2010) for senior unsecured corporate bonds.
To keep the potential sources of cyclical variation under control, we maintain the pa-
rameters determining the e¤ective maturity of performing loans, the speed of resolution of
NPLs, and the ‡ow of entry of new loans as time invariant.
Banks’discount rate r is …xed at 1.8% so as to obtain an unconditional average of the
contractual loan rate c equal to 2.49%, which is very close to the 2.52% average interest
rate of new corporate loans made by Euro Area banks in the period from January 2010 to
September 2016.21
21
We use the Euro area (changing composition), annualised agreed rate/narrowly de…ned ef-
fective rate on euro-denominated loans other than revolving loans and overdrafts, and con-
venience and extended credit card debt, made by banks to non-…nancial corporations (see
http://sdw.ecb.europa.eu/quickview.do?SERIES_KEY=124.MIR.M.U2.B.A2A.A.R.A.2240.EUR.N).

17
 Table 2
Endogenous variables under the baseline calibration
(IRB bank, percentage of mean exposures unless indicated)
Conditional means
Mean St. Dev. Expansions Contractions
Yearly contractual loan rate c (%) 2.47 2.57
Share of standard loans (%) 81.35 3.48 82.68 76.85
Share of substandard loans (%) 15.46 1.90 14.59 18.42
Share of NPLs (%) 3.19 1.05 2.73 4.73
Realized default rate (% of performing loans) 1.89 0.90 1.36 3.43
Impairment allowances:
Incurred losses 1.04 0.37 0.87 1.60
Prudential expected losses under IRB 2.00 0.47 1.80 2.69
CECL expected losses 4.36 0.58 4.06 5.36
IFRS 9 allowances 2.43 0.61 2.14 3.42
Stage 1 allowances 0.22 0.05 0.20 0.32
Stage 2 allowances 1.17 0.20 1.07 1.51
Stage 3 allowances 1.04 0.37 0.87 1.60
IRB minimum capital requirement 9.05 0.08 9.04 9.10
IRB minimum capital requirement + CCB 11.88 0.10 11.86 11.94

Regarding the resolution of NPLs, we set 3 equal to 44.6% in order to produce an un-
conditional average fraction of NPLs consistent with the 5% average PD, including defaulted
exposures that the EBA (2013, Figure 10) reports for the earliest period in its study, namely
the …rst half of 2008.22 This value of 3 implies an average time to resolution for NPLs
of 2.24 years, which is very close to the 2.42-year average duration of corporate insolvency
proceedings across EU countries documented by the EBA (2016, Figure 13).
Finally, the assumed size of the ‡ow of newly originated loans, e1 =1, only provides a
normalization and solely a¤ects the average size of the bank’s total exposures. Further, we
report most variables as a percentage of the bank’s total mean exposures (assets) making
the absolute value of those exposures irrelevant in the analysis.
22
We take this observation, right before experiencing the full negative impact of the Global Financial
Crisis, as the best proxy in the data for the model’s steady state. As shown in Table 2, with this procedure,
we obtain an average 3.2% share of defaulted exposures in the ergodic portfolio, right inbetween the 2.5%
and 4.4% reported by the EBA (2013, Figure 8) for corporate loans in the second half of 2008 and the
…rst half of 2009, respectively. Conditional on being in an expansion and in a recession, the mean share of
defaulted exposures equals 2.73% and 4.73%, respectively.

18
4.2 Size, volatility and cyclicality of the impairment measures

Table 2 reports unconditional means, standard deviations, and means conditional on each
aggregate state for a number of endogenous variables. The variation in the aggregate state
causes a signi…cant variation in the composition of the bank’s loan portfolio. Not surprisingly,
the shares of substandard and non-performing loans increase during contractions, and the
overall realized default rate is more than double during contractions than during expansions.
The mean relative sizes of the various impairment allowances are ranked as predicted
above. While for the considered portfolio, impairment allowances under IFRS 9 (ELIF
t
RS9
)
essentially double those associated with the IL approach (ILt ), the incoming CECL approach
(ELCECL
t ) more than quadruples them. Note that higher level of allowances associated with
IFRS 9 comes mostly from stage 2 loans in spite of the fact that these loans only represent a
modest 15.5% in the loan portfolio. Interestingly, impairments measured under CECL and
IFRS 9 are the ones exhibiting greater volatility and mean variation across states (130 and
128 basis points of mean exposures, respectively), followed by the prudential expected losses
used by IRB banks and the IL impairments (91 and 73 basis points, respectively).
The decomposition by stage shown for IFRS 9 reveals that allowances associated with
NPLs, followed by those associated with substandard loans, are those that contribute most to
cross-state variation in loan loss provisions (73 and 44 basis points, respectively). However,
NPLs are treated in the same way by all measures, which implies that the di¤ering mean
cross-state variation of the alternative measures must stem from the treatment of standard
loans (which is similar in ELIRB and ELIF RS9 , but quite di¤erent in IL and ELCECL )
and stage 2 loans (which is similar in ELCECL and ELIF RS9 , but quite di¤erent in IL and
ELIRB ) or from the cyclical shift of loans across stages 1 and 2 (under ELIF RS9 ).
Finally, Table 2 also reports the descriptive statistics of the implied overall minimum
capital requirement (k) and the minimum requirement plus the CCB (k) that would apply
to the bank under our calibration.

4.3 Impact on the cyclicality of pro…ts and regulatory capital

Table 3 summarizes the impact of the various provisioning approaches on P/L and CET1.
The unconditional mean of P/L di¤ers across them, re‡ecting that di¤erent levels of pro-

19
visions imply de facto di¤erent levels of debt …nancing for the same portfolio and, hence,
di¤erent amounts of interest expense. Con…rming what one might expect after observing the
volatility ranking of the impairment measures in Table 2, P/L is signi…cantly more volatile
and variable across aggregate states under the more forward-looking ELCECL and ELIF RS9
than under ELIRB or IL:
The more forward-looking impairment measures are the ones that make the bank, on
average, more CET1-rich in expansion states and less CET1-rich in contraction states; that is,
those that render CET1 more procyclical in this sense. In any case, the reported quantitative
di¤erences for this variable are not huge, in part because under our assumptions on the
bank’s management of its CET1, the range of variation in CET1 under any of the impairment
measures is limited by the regulatory-determined bands of the sS-rule described in equations
(22) and (23). As explained above, the bank adjusts its CET1 to remain within those bands
by paying dividends or raising new equity.
Thus, a suitable way to assess the potential procyclicality associated with each impair-
ment measure is to look at the frequency and size (conditional on them being strictly positive)
of dividends and recapitalizations. Quite intuitively, under all measures we obtain that divi-
dend distributions only occur (if at all) during expansions, while recapitalizations only occur
(if at all) during contractions.
Dividends are most frequently paid under ELCECL , ELIF RS9 ; ELIRB ; and IL; in this
order, mostly re‡ecting the above mentioned di¤erences in leverage and interest expense
implied by the levels of provisioning. In terms of recapitalization needs, ELIF RS9 involves
a signi…cantly higher probability of having to recapitalize the bank in contractions (18.2%)
than any other the other measures (for which the probability ranges between 12.72% for
ELIRB and 13.42% for ELCECL ).23
23
However, these e¤ects become counterbalanced by the fact that, when strictly positive, the average size
of the recapitalizations needed under ELIF RS9 is slightly lower than that under EL1Y :

20
 Table 3
Endogenous variables under the baseline calibration
(IRB bank, percentage of mean exposures unless otherwise indicated)
IL ELIRB ELCECL ELIF RS9
P/L
Unconditional mean 0.18 0.20 0.25 0.21
Conditional mean, expansions 0.41 0.45 0.56 0.52
Conditional mean, contractions -0.59 -0.65 -0.81 -0.84
Standard deviation 0.42 0.47 0.60 0.59
CET1
Unconditional mean 11.33 11.33 11.37 11.31
Conditional mean, expansions 11.56 11.59 11.70 11.65
Conditional mean, contractions 10.52 10.43 10.21 10.14
Standard deviation 0.85 0.85 0.83 0.86
Probability of dividends being paid (%)
Unconditional 50.46 52.53 58.35 54.27
Conditional, expansions 65.40 68.07 75.62 70.33
Conditional, contractions 0 0 0 0
Dividends, if positive
Conditional mean, expansions 0.40 0.42 0.44 0.42
Conditional mean, contractions – – – –
Probability of having to recapitalize (%)
Unconditional 2.92 2.91 3.06 4.16
Conditional, expansions 0 0 0 0
Conditional, contractions 12.77 12.72 13.42 18.20
Recapitalization, if positive
Conditional mean, expansions – – – –
Conditional mean, contractions 0.53 0.56 0.46 0.48

4.4 E¤ects of the arrival of a contraction

Figure 2 shows the e¤ects of the arrival of a contraction at t=0 (that is, the realization
of s0 =2) after having spent a long enough period in the expansion state (that is, having
had st =1 for su¢ ciently many dates prior to t=0). The results shown in the …gure are
equivalent to those typical of impulse response functions in macroeconomic analysis. From
t=1 onwards the aggregate state follows the Markov chain calibrated in Table 1, thus making
the trajectories followed by the variables depicted in the …gure stochastic. The …gure depicts
the average trajectories resulting from simulating 10,000 paths.

21
 The higher amount of loans becoming substandard immediately after a recession arrives
makes the e¤ects of the arrival of a recession persistent over time, despite the relatively short
duration of the contraction state under our baseline calibration (2 periods on average). This
can be seen in Panel A of Figure 2, which depicts the evolution of NPLs.
The results regarding the evolution of the various impairment measures over the same
time span appear in Panel B of Figure 2. The average trajectories of the impairment al-
lowances ILt ; ELIRB
t ; ELCECL
t ; and ELIF
t
RS9
are reported as a percentage of the total initial
loans. The levels of the series at t=–1 re‡ect the di¤erent sizes of the impairment allowances
obtained after a long expansion phase under each of the compared measurement methods.
When the recession arrives at t=0, all the measures move upwards. The most forward look-
ing ones, ELCECL
t and ELIF
t
RS9
, peak on average after one period and then enter a pattern
of exponential decay, driven by maturity, defaults, migration of substandard loans back to
the standard category, and the continued origination of new standard-quality loans.24 The
measures based on incurred or one-year ahead expected losses, ILt and ELIRB
t react more
slowly, peaking on average after two periods, and then also fall gradually.
Therefore the on-impact responses of ELCECL
t and ELIF
t
RS9
to the arrival of a recession
are larger than those of ILt and ELIRB
t : In the case of ELIF
t
RS9
part of the reactivity to
the arrival of the recession is due to the so-called “cli¤ e¤ect” associated with the change
in the provisioning horizon when exposures shift from stage 1 to stage 2. In both cases
the additional reactivity is also related to the e¤ects of updating the PIT forecast of the
losses on currently performing loans which are expected beyond the one-year horizon (stage
2 exposures under IFRS 9 and all performing exposures under CECL).
The implications for P/L are described in Panel C of Figure 2. Each measure spreads over
time the (same …nal average) impact of the shock on P/L in a di¤erent manner. ELIF
t
RS9

and ELCECL
t front-load very similarly the impact of the shock: P/L becomes very negative
on impact, but then turns positive and returns to normal pretty quickly afterwards. With
ILt and, to a lesser extent, ELIRB
t , P/L is a¤ected much less on impact but remains negative
for several periods.
24
Variations of the experiment that simultaneously shut down or reduce origination of new loans for a
number periods can be easily performed without losing consistency. Experiencing lower loan origination
after t=0 delays the process of reversion to the steady state but does not qualitatively a¤ect the results.

22
 Panel A. Non-performing loans Panel C. P/L
4.5 0.5

0
3.5

3
-0.5

2.5

2 -1
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Panel B. Impairment allowances Panel D. CET1

6 12.5

12
5

11.5
4

11
3
10.5

2
10

1 9.5

0 9
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 2. E¤ects of the arrival of a contraction

Average responses to the arrival of s=2 after a long period in s=1
(IRB bank, as a percentage of average exposures).

Panel D of Figure 2 shows the implications for an IRB bank’s CET1. Before the shock
hits, at t=–1, the bank has a fully loaded CCB. The average positions of the bands k and
k re‡ected in the …gure exhibit some time variation as a result of the change in RWAs
that follows the temporary deterioration in the composition of the loan portfolio, implying
a bu¤er on top of the minimum required capital of more than 2.5% of total assets. The
di¤erent impacts of the alternative provisioning methods on CET1 essentially mirror the
previously commented impact on P/L.
The on-impact e¤ect of the arrival of contraction under CECL and IFRS 9 implies con-
suming about one percentage point of the roughly three percentage points of initial assets

23
represented by the CCB, but CET1 tends to recover on average afterwards. With ELIRB
and, especially, IL the on-impact e¤ect is much smaller (about half) than under the new
ECL provisions, but CET1 tends to deteriorate further in subsequent periods (though on
average not going below the trajectories obtained under ELCECL and ELIF RS9 ).
The fact that the trajectories depicted in Figure 2 are average trajectories is important
for interpreting these results. For example, in Panel D, the average trajectory of CET1 lies
within the bands determined by the average trajectories of k and k but this does not mean
that the bank does not need to recapitalize (or does not pay dividends) over all the possible
trajectories. On the contrary, many of the trajectories go upward and touch the upper band
for paying dividends (e.g. if the contraction ends and does not return), while a few other go
downward, touch the lower band, and force the bank to recapitalize (e.g. if the contraction
lasts long enough or another contraction follows soon after an initial recovery). This explains
the compatibility between the results in this …gure and the probabilities of positive dividend
payments and recapitalization needs reported in Table 2.

Panel A. CET1 under ELIRB Panel B. CET1 under ELIF RS9

12.5 12.5

12 12

11.5 11.5

11 11

10.5 10.5

10 10

9.5 9.5

9 9

8.5 8.5
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 3. CET1 after the arrival of a contraction (IRB bank)

500 simulated trajectories of CET1 under ELIRB and ELIF RS9
following the arrival of s=2 after a long period in s=1
(IRB bank, as a percentage of average exposures)

To further illustrate the di¤erence between average and realized trajectories, Figure 3
shows 500 simulated trajectories for CET1 under ELIRB and ELIF RS9 . Under IFRS 9, it
takes four consecutive years in the contraction state (s=2) for a bank to deplete its CCB

24
and require a recapitalization. By contrast, under the one-year expected loss approach, the
CCB would be used up only after …ve years in the contraction state.

5 Implications and policy analysis

The results in prior sections show that the loan loss provisions under CECL and IFRS 9
will imply a more up-front recognition of impairment losses upon the arrival of a recession.
This means that the on-impact declines in P/L and CET1 will be larger than under prior
provisioning approaches. In our model, a fall in CET1 …rst reduces a bank’s CCB (if positive),
forcing it to cancel its dividends. If the losses are large enough or persist for a number of
periods, the CCB may be fully depleted, pushing the bank to raise new equity in order to
continue complying with the minimum capital requirement without reducing the size of the
loan portfolio. In reality, if the bank owners dislike cancelling dividends, operating without
fully loaded bu¤ers or raising new equity capital, the above e¤ects might translate into loan
sales or a reduction in the origination of new loans. Speci…cally, imperfections pressing banks
to show up good capital ratios or a stable dividend ‡ow, or making equity issuance costly
(as in, e.g., Bolton and Freixas, 2006, Allen and Carletti, 2008, or Plantin, Sapra and Shin,
2008) might lead the bank to reduce its assets.25
Of course the potential disadvantages of a sharper contraction of credit or larger asset
sales at the beginning of a recession should be weighed against the advantages of having
…nancial statements that re‡ect the weakness or strength of the reporting institutions in a
more timely and reliable way, as there is evidence suggesting that this helps resolve bank
crises in a prompter, safer, and more e¤ective manner.26 In this sense, the policy analysis
contained below is made without prejudging or attempting to assess the net welfare e¤ects of
the new provisions. Instead, we explore the e¤ectiveness of several policy options in reducing
25
Concerns on this type of reaction are at the heart of the motivation for macroprudential policies. As put
by Hanson, Kashyap and Stein (2011, p. 5), “in the simplest terms, one can characterize the macroprudential
approach to …nancial regulation as an e¤ort to control the social costs associated with excessive balance sheet
shrinkage on the part of multiple …nancial institutions hit with a common shock.”
26
Laeven and Majnoni (2003) and Huizinga and Laeven (2012) document bank provisioning practices
during economic slowdowns and their implications for …nancial stability. Beatty and Liao (2011) document
that banks recognizing loan losses in a timelier manner experience lower reductions in lending during con-
tractionary periods. Bushman and Williams (2012, 2015) document the link between a timely and decisive
recognition of loan losses and banks’risk pro…les.

25
the impact of the new provisions on the bank’s need to be recapitalized. In other words,
we take the bank’s unconditional recapitalization probability as a proxy of the potential
procyclical e¤ects.27
As in discussions on the procyclicality of Basel capital requirements (Kashyap and Stein,
2004, and Repullo and Suarez, 2013), several factors can reduce the procyclical e¤ects asso-
ciated with the impact of loan loss provisions on recapitalization needs. First, banks may
react to the new provisioning methods by increasing their voluntary capital bu¤ers or by
undertaking less cyclical investments. Second, even if CECL and IFRS 9 provisions further
reduce banks’lending capacity during recessions, credit demand may also contract during re-
cessions, mitigating the implications. Yet recent evidence (including Mésonnier and Monks,
2015, Aiyar, Calomiris, and Wieladek, 2016, Behn, Haselmann, and Wachtel, 2016, Gropp et
al., 2016, Jiménez et al., 2017) suggests that banks tend to accommodate sudden increases in
required capital (or falls in available regulatory capital) by reducing bank lending, especially
during contractions, producing negative e¤ects on economic activity.
If this were the case, authorities could consider policies such as the ones explored in the
rest of this section, namely: (i) increasing the CCB, (ii) using the CCyB by activating it to
a level above zero during expansions and releasing it during contractions, (iii) introducing
prudential bu¤ers based on stress test results, and (iv) smoothing the credit risk parameters
used in the calculation of the new provisions so as to get closer to a TTC approach.

5.1 Increasing the capital conservation bu¤er

A …rst straightforward way to address the concern about the higher recapitalization prob-
ability implied by CECL and IFRS 9 is to increase the CCB. Our baseline results above
were obtained under the 2.5% of RWAs target speci…ed by Basel III. Panel A in Figure 4
compares the unconditional probability of the needing a recapitalization under each of the
new provisioning methods if such target is permanently increased with an add-on ranging
from 0% to 2.5% of the bank’s RWAs.
As shown in the …gure, such an add-on, by increasing the bank’s loss absorption capacity
27
In all the baseline results, recapitalization needs only emerge during contractions so using the conditional
probability of recapitalization during contractions would provide, up to a scale parameter, an equivalent
metric.

26
through earnings retention when pro…ts are positive, reduces the probability of having to
recapitalize the bank. For a total CCB target of 5%, the recapitalization probability would
be below 0.5% per year under both CECL and IFRS 9.

Panel A. CCB add-on Panel B. CCyB activation

4.5 4.5

4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Figure 4. E¤ects of increasing or activating the regulatory bu¤ers

Unconditional probabilities of recapitalization (%) as a function of the
add-on to or level of activation of the corresponding bu¤er,
measured as percentage points of RWAs

5.2 Activation of the countercyclical capital bu¤er

The CCyB of Basel III works similarly to a discretionary, time-varying add-on to the CCB.
Speci…cally, macroprudential authorities are intended to set such an add-on (called the CCyB
rate) in the range from 0% to 2.5% of RWAs on the basis of the evolution of the credit cycle.
Under Basel III, rises in the CCyB rate must be announced one year before their e¤ective
application, while CCyB reductions (releases) apply immediately.
To formally capture these features, we modify the upper band k t that drives the dynamics
of CET1 in the model by making it now equal to the sum of the time-invariant CCB target
of 2.5% of RWAs and, whenever activated, a CCyB rate of 2 [0; 0:025]:

CCyB
kt = k t [1:3125 + ( =0:08)(2 st )(2 st 1 ):::(2 st T )]: (28)

This formulation implies that the CCyB is activated whenever the economy is and has been
in an expansion for the last T periods (st = 1 for = 0; :::T ) , while the CCyB turns zero

27
otherwise.28
Panel B in Figure 4 shows the recapitalization probabilities obtained with a two year
implementation lag (T = 2) and CCyB rates ranging from 0% to 2.5% of RWAs. Interest-
ingly, the CCyB is e¤ective in reducing the need to recapitalize the bank but only down to
about a probability of 1.5% and 2% per year for CECL and IFRS 9, respectively. The lower
e¤ectiveness of the CCyB relative to a CCB add-on of the same maximum size is due to
the CCyB implementation lag, which delays its build-up, thus increasing the odds that the
arrival of a contraction catches the bank with an overall CCB+CCyB below its maximum
size.

Panel A. CCB add-on of 1% Panel B. CCyB activation at 1%

14 14

13 13

12 12

11 11

10 10

9 9

8 8
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 5. Evolution of CET1 under alternative bu¤er policies

500 simulated trajectories of CET1 under ELIF RS9 and alterative bu¤er
policies following the arrival of s=2 after a long period in s=1.
(IRB bank, as a percentage of average exposures)

Further intuition on the di¤erent behavior of a CCB add-on and an actively managed
CCyB of similar sizes can be obtained from Figure 5, which shows simulated paths for CET1
under IFRS 9 provisioning and each of the two countercyclical tools. Bu¤er releases upon
the arrival of contractions and implementation lags subsequent to the arrival of expansions
explain why the upper band and average trajectory of CET1 in Panel B ‡uctuate more than
their counterparts in Panel A. The following table re‡ects the quantitative di¤erence in the
28
To avoid a counterintuitive pay of dividends out of the release of the CCyB at the start of a recession,
we add in this part of the analysis the constraint that the bank is not allowed to pay dividends while in the
contraction state.

28
performance of the two tools for a common rate of 1%:

ELCECL ELIF RS9

Baseline model 3.06% 4.16%
CCB add-on of 1% 1.22% 1.59%
CCyB activation at 1% 1.83% 2.23%

5.3 Stress test based bu¤ers

Macroprudential stress testing has become part of supervisors’toolkit in the aftermath of the
Global Financial Crisis. What started in some jurisdictions as a one-o¤ exercise intended to
bring calm to investors concerned about the solvency of the banks, has consolidated as a sys-
tematic recursive way to assess and ensure banks’resiliency. At the risk of oversimplifying,
macroprudential stress testing consists on de…ning some adverse scenario going forward over
a number of years, estimating the capital that banks would need to resist such scenario, and
demanding banks (in a harder or softer manner) to have or otherwise raise or accumulate
such level of capital within a reasonable period of time. In the two parts of this subsection,
we consider two possible manners of formally de…ning the relevant adverse scenarios and
the implicit capital requirements (or capital bu¤ering requirements) associated with macro-
prudential stress testing. They di¤er in the frequency with which the size of the required
bu¤er gets recalibrated (recursively or just once at the peak of expansion periods) and in
the hardness of the attached capital constraint (which may work as a compulsory capital
requirement or, more softly, as a CCB add-on). We consider two variations without the aim
to be exhaustive but to illustrate the importance of the details even in a highly stylized setup
such as the one provided by our model.

5.3.1 Recursive stress test requirement

In this extension we assume that a macroprudential stress test is performed every year.
The stress test performed in year t consists in considering an adverse scenario in which the
economy is in recession in year t + 1 and up to, at least, year t + T; where T is the length
of the adverse scenario. The hypothetical dynamics of CET1 over the adverse scenario is
computed under the assumption that the bank starts at t with capital e
kt = kt 1 + P Lt and
does not make any discretionary capital adjustment up to year t + T so that its capital

29
follows the di¤erence equation e
kt+i = e
kt+i 1 + PfLt+i for i=1,...,T; where PfLt+i denotes the
P/L generated at year t+i under the projected trajectories of capital and the aggregate state
(which can be found by properly feeding the formula for P Lt+i used in the baseline model).
ST
In this context, the recursive stress test requirement is an add-on t to the minimum capital
requirement of the bank at t de…ned as

ST
t = max(e
kt + defST
t k t ; 0);

where defST
t = k t+T –e
kt+T is the capital de…cit detected in the stress test exercise. Intuitively,
ST
the add-on t is designed to guarantee that the bank can go through the adverse scenario
without violating the minimum capital requirement at the end of it. The e¤ect of the add-on
ST
is to e¤ectively rise the overall minimum capital requirement of the bank at t to k t + t ;

forcing it to adapt its dividend and equity raising policy at t accordingly.29

Under this formulation, we simulate the dynamics of the model and, in particular, P Lt
and kt assuming that the bank is subject to the corresponding stress test based add-on
ST
t to the minimum capital requirement in every period t. We consider adverse scenarios
of di¤erent severity as described by T: As a summary of the results, Figure 6 depicts the
ST
unconditional mean of the add-on t (Panel A) and the unconditional yearly probability
of having to recapitalize the bank (Panel B) for di¤erent values of T . Clearly, as di¤erent
ST
provisioning methods imply di¤erent trajectories for P Lt and kt ; the value of t can be
di¤erent for each provisioning method.
These results show that recursively requesting the bank to be able to cope with a su¢ -
ciently adverse scenario without needing a future recapitalization has the e¤ect of dramat-
ically increasing the probability of having to recapitalize the bank throughout its lifetime.
So, paradoxically, the attempt to avoid recapitalization needs along the hypothesized adverse
scenarios ends up producing much more frequent recapitalization needs along the bank’s life-
time than in the absence of the considered stress test requirement.30
29
Speci…cally, dividends and equity injections of the bank in year t would be driven by:
ST
divt = maxfmin[(kt 1 + P Lt ) 1:3125k t ; (kt 1 + P Lt ) (k t + t )]; 0g;

ST
recapt = max[(k t + t ) (kt 1 + P Lt ); 0];
which are immediate adaptations of (22) and (23) to the new requirement.
30
The ‡at sections in the curves represented in Panel B of Figure 6 evidence that the need to satisfy the

30
 Panel A. Add-on ST to required capital Panel B. Recapitalization probability
t

4.5 25

3.5 20

2.5 15

1.5 10

0.5 5
1 2 3 4 5 6 1 2 3 4 5 6

Figure 6. Recursive stress test requirement

Average add-on STt to required capital (in percent of RWAs) and unconditional
probability of recapitalization (%) for di¤erent lengths T of
a recursive adverse scenario (in years, on the x-axis)

5.3.2 CCB-like stress test requirement

We now consider a much softer stress test requirement: one that, instead of e¤ectively rising
the minimal required capital of the bank, takes the form of a CCB add-on such as the one
analyzed in subsection 5.1 but calibrated on the basis of the results of some one-o¤ stress
test performed at the peak of an expansion period. Speci…cally, we denote such CCB add-
ST
on as and …nd it by considering the situation reached by the bank at a reference year
t = 0 after spending a su¢ ciently large number of periods in the expansion state. In such
a situation, the bank would have both the reference CCB of Basel III and the new add-on
ST
fully loaded, so its regulatory capital would be e
k0 = k 0 1:3125 + ST
=0:08 . We then
suppose that at t=1, the economy shifts to the contraction state and remains in such state
up to, at least, year t = T; and the bank does not recapitalize or pay any dividend between
ST
t = 1 and t = T: In this setup, we numerically …nd the size of the add-on for which the
bank would complete the hypothesized trajectory at T with just enough capital to comply
with the minimum requirement, that is, with e
kT = k T .31 So, by construction, the CCB add-
large stress test based requirement during contractions, would essentially mean having to recapitalize the
bank in every contraction period.
31
We set ST = 0 whenever this procedure yields ST < 0; which explains the ‡at sections of the curves
depicted in Figure 7.

31
on guarantees that, by the end of this one-o¤ adverse scenario, the bank has just enough
CET1 to avoid being recapitalized.

Panel A. Required bu¤er size ST Panel B. Recapitalization probability

2.5 4.5

2 3.5

3
1.5
2.5

2
1
1.5

1
0.5
0.5

0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8

Figure 7. CCB-like stress test requirement

Required bu¤er size (in percent of RWAs) and unconditional probability
of recapitalization for di¤erent lengths T of a one-o¤ adverse
scenario (in years, on the x-axis)

ST
Figure 7 depicts the calibrated values for (Panel A) and the resulting unconditional
probabilities of recapitalization (Panel B) for di¤erent lengths T of the one-o¤ adverse sce-
ST
nario. The …gure shows that, for example, the CCB add-on that would leave the bank
with just enough capital not to need a recapitalization after a 5-year contraction scenario
(T =5) is 0.71% under IFRS 9 and 0.44% under CECL. As shown in Panel B of this …gure,
with this calibration of the CCB add-on, the bank would have a very similar unconditional
probability of being recapitalized under the two new provisioning methods.
Formally, a CCB-like stress test requirement as the one just described is entirely equiv-
alent to a CCB add-on of equal size, so they only di¤er in the narrative leading to their
calibration. However, in practice, a CCB add-on explicitly calibrated on the basis of a stress
test may have the advantage of adapting better to structural changes and bank heterogeneity
than a common CCB add-on …xed once and for all.

32
5.4 Smoothing the inputs

A key di¤erence between the provisions implied by CECL and IFRS 9 and the prudential
expected losses under the IRB approach is that the latter aim to avoid cyclicality by stipulat-
ing the use of TTC PDs and downturn LGDs, rather than PIT estimates of both of them. In
this subsection we consider how ELCECL and ELIF RS9 would perform if they were modi…ed
in that direction. Accordingly, both measures would still di¤er from ELIRB in the horizon
over which (some of) the expected losses are projected, but not in the state-variation of the
PDs and LGDs. These modi…cations can be interpreted either as a change in the accounting
standards or, less ambitiously, as a policy implemented in the form of prudential adjust-
ments (CET1 reductions or add-backs) based on the di¤erences between the PIT accounting
expected losses and the relevant TTC prudential expected losses.32
The following table shows the results associated with (i) smoothing only the PDs by
using TTC rather than PIT estimates of them, and (ii) smoothing also the LGDs by using
downturn LGDs rather than PIT estimates of them:
ELCECL ELIF RS9
Baseline model 3.06% 4.16%
Provisions based on TTC PDs 2.33% 3.17%
Provisions based on TTC PDs and downturn LGDs 2.31% 4.05%

Interestingly, smoothing the PDs is e¤ective in reducing the procyclicality of both CECL
and IFRS 9, while smoothing also the LGDs via the proposed procedure is, in incremental
terms, only very marginally countercyclical under CECL and rather procyclical under IFRS
9. To explain this last result, recall that part of the cyclical performance of IFRS 9 is due
to the cli¤ e¤ects associated with the reclassi…cation of exposures between stages 1 and
2. Relying on downturn LGDs rather than PIT LGDs appears to increase the di¤erential
provisioning needs behind the cli¤ e¤ects, possibly because under a PIT approach part of
the default losses associated with stage 2 exposures in the contraction state are estimated to
actually realize (and produce recoveries) when the economy is back in the expansion state,
thus implying LGDs lower than the downturn LGDs.
32
The second option would not require reforming the new accounting standards but would have the draw-
back of generating discrepancies between accounting and regulatory capital, which might erode investors’
con…dence in the reliability of both numbers. See BCBS (2016) for a description of alternatives for the
regulatory treatment of accounting provisions.

33
 As one can clearly see, smoothing the inputs relevant for the estimation of credit losses
helps reducing the procyclical e¤ects of CECL and IFRS 9 but not as much as the CCB
add-ons considered in Figure 4.

6 Conclusions
We have described a simple recursive model for the assessment of the level and cyclical
implications of the new ECL approaches to loan loss provisioning under IFRS 9 and the
incoming update of US GAAP. We have calibrated the model to represent a bank with a
portfolio of EU corporate loans, and compared its performance under alternative provisioning
methods: the old incurred loss approach, the prudential one-year expected losses associated
with IRB capital requirements, the lifetime CECL provisions of US GAAP, and the mixed-
horizon ECL provisions of IFRS 9.
Our results suggest that the loan loss provisions implied by IFRS 9 and the CECL ap-
proach will rise more suddenly than their predecessors when the cyclical position of the
economy switches from expansion to contraction. This implies that P/L and, without the
application of regulatory …lters, CET1 will decline more severely at the beginning of down-
turns. The baseline quantitative results of the paper suggest that the arrival of an average
recession might imply on-impact losses of CET1 twice as large as those under the incurred
loss approach and equivalent to about one third of the fully loaded CCB of the analyzed
bank.
As shown in Appendix C, the timing and the importance of the cyclical e¤ects depend
on the anticipated and unanticipated severity (duration) of recessions as well as the extent
to which cyclical turning points can be anticipated in advance. Greater severity exacerbates
the cyclical e¤ects, while greater capacity to anticipate the arrival of a contraction allows to
absorb part of the cyclical losses prior to the start of the contraction.
While the early and decisive recognition of forthcoming losses may have signi…cant ad-
vantages (e.g. in terms of transparency, market discipline, inducing prompt supervisory
intervention, etc.), it may also imply, via its e¤ects on regulatory capital, a loss of lending
capacity for banks at the very beginning of a contraction, potentially contributing, through
feedback e¤ects, to its severity. In this paper we have gauged the direction and intensity of

34
the procyclical e¤ects by looking at some …rst-round, partial equilibrium e¤ects through the
eyes of our simple model.
Speci…cally, in the discussion of potentially mitigating policies, we have focused the analy-
sis on the unconditional annual probability with which the bank described in the model needs
to raise new capital to avoid violating its minimum regulatory capital requirements. After
examining policies such as increasing the CCB, actively using the CCyB, introducing pruden-
tial bu¤ers based on stress test results or calculating the new provisions using a prudential
TTC approach rather than their current PIT approach, we conclude that introducing a
CCB add-on (possibly calibrated on the basis of stress test results) would be not only the
simplest but also the policy option with the highest e¤ectiveness in terms of reducing the
unconditional probability of having to recapitalize the bank.

35
 Appendices
A Model equations in the presence of aggregate risk
In this appendix we extend the equations presented in the main text to the case in which
aggregate risk a¤ects the key parameters governing credit risk and, potentially, loan origina-
tion. We capture aggregate risk by introducing an aggregate state variable that can take two
values st 2 f1; 2g at each date t and follows a Markov chain with time-invariant transition
probabilities ps0 s =Prob(st+1 = s0 jst = s): The approach can be trivially generalized to deal
with a larger number of aggregate states.
In order to measure expected losses corresponding to default events in any future date t,
we have to keep track of the aggregate state in which the loans existing at t were originated,
z=1,2; the aggregate state at time t; s=1; 2; and the credit quality or rating of the loan at
t; j=1; 2; 3. Thus, it is convenient to describe (stochastic) loan portfolios held at any date t
as vectors of the form 0 1
xt (1; 1; 1)
B xt (1; 1; 2) C
B C
B xt (1; 1; 3) C
B C
B xt (1; 2; 1) C
B C
B xt (1; 2; 2) C
B C
B xt (1; 2; 3) C
B
yt = B C; (A.1)
x (2; 1; 1) C
B t C
B xt (2; 1; 2) C
B C
B xt (2; 1; 3) C
B C
B xt (2; 2; 1) C
B C
@ xt (2; 2; 2) A
xt (2; 2; 3)
where component xt (z; s; j) denotes the measure of loans at t that were originated in aggre-
gate state z; are in aggregate state s and have rating j.33
Our assumptions regarding the evolution and payo¤s of the loans between any date t
and t + 1 are as follows. Loans rated j=1; 2 at t mature at t + 1 with probability j (s0 ),
where s0 denotes the aggregate state at t + 1 (unknown at date t). In the case of NPLs
(j=3), 3 (s0 ) represents the independent probability of a loan being resolved, in which case
it pays back a fraction 1 e(s0 ) of its unit principal and exits the portfolio. Conditional on
s0 ; each loan rated j=1; 2 at t which matures at t + 1 defaults independently with probability
P Dj (s0 ), being resolved within the period with probability 3 (s0 )=2 or entering the stock of
NPLs (j=3) with probability 1 3 (s0 )=2. Maturing loans that do not default pay back their
33
Along a speci…c history (or sequence of aggregate states), for any z and j, the value of xt (z; s; j) will
equal 0 whenever st 6= s.

39
principal of one plus the contractual interest cz , established at origination.
Conditional on s0 ; each loan rated j=1; 2 at t which does not mature at t+1 goes through
one of the following exhaustive possibilities. First, default, which occurs independently with
probability P Dj (s0 ); and in which case one of two things can happen: (i) it is resolved within
the period with probability 3 (s0 )=2; or (ii) it enters the stock of NPLs (j=3) with probability
0
1 3 (s )=2. Second, migration to rating i 6= j (i=1,2); in which case it pays interest cz
and continues for one more period; this occurs independently with probability aij (s0 ): Third,
continuation in rating j, in which case it pays interest cz and continues for one more period;
this occurs independently with probability

ajj (s0 ) = 1 aij (s0 ) P Dj (s0 ):

A.1 Portfolio dynamics under aggregate risk

Under aggregate risk, the dynamics of the loan portfolio between any dates t and t + 1 is no
longer deterministic, but driven by the realization of the aggregate state variable at t + 1;
st+1 : To describe the dynamics of the system compactly, let the binary variable t+1 = 1 if
st+1 = 1 and t+1 = 0 if st+1 = 2: The dynamics of the system can be described as

yt+1 = G( t+1 )yt + g( t+1 );

where
0 1
t+1 M (1) t+1 M (1)
B 06 6 C
1– t+1 M (2) 1– t+1 M (2)
G( t+1 )= B
@
C;
A
t+1 M (1) t+1 M (1)
06 6
1– t+1 M (2) 1– t+1 M (2)
T
g( t+1 ) = t+1 e1 (1); 0; 0; 0; 0; 0; 0; 0; 0; 1 t+1 e1 (2); 0; 0 ;
1 if ut+1 2 [0; p1st ];
t+1 =
0 otherwise,
st+1 = t+1 +2 1 t+1 ;
ut+1 is an independently and identically distributed uniform random variable with support
[0; 1]; e1 (s0 ) is the (potentially di¤erent across states s0 ) measure of new loans originated at
t + 1; and 06 6 denotes a 6 6 matrix full of zeros.

A.2 Incurred losses

Incurred losses measured at date t would be those associated with NPLs that are part of the
bank’s portfolio at date t: Thus, the incurred losses reported at t would be given by
X X
ILt = (s)xt (z; s; 3);
z=1;2 s=1;2

40
where (s) is the expected LGD on an NPL conditional on being at state s in date t: This
can be more compactly expressed as

ILt = bbyt ; (A.2)

where bb = (0; 0; (1); 0; 0; (2); 0; 0; (1); 0; 0; (2)):

The expected LGD conditional on each current state s can be found as functions of
the previously speci…ed primitives of the model (state-transition probabilities, probabilities
of resolution of the defaulted loans in subsequent periods, and loss rates e(s0 ) su¤ered if
resolution happens in each of the possible future states s0 ) by solving the following system
of recursive equations:
X h i
0 e 0 0 0
(s) = p s0 s 3 (s ) (s ) + (1 3 (s )) (s ) ; (A.3)
s0 =1;2

for s=1; 2:

A.3 Discounted one-year expected losses

Based on the loan portfolio held by the bank at t, provisions computed on the basis of
the discounted one-year expected losses add to the incurred losses written above the losses
stemming from default events expected to occur within the year immediately following. Since
a period in the model is one year, the corresponding allowances are given by

EL1Y b (A.4)
t = (b + b)yt ;

where b = ( 1 b; = 1=(1 + cz ); and b = (b11 ; b12 ; 0; b21 ; b22 ; 0), with

2 b); z
X n o
bsj = ps0 s P Dj (s0 ) [ 3 (s0 )=2] e(s0 ) + [1 3 (s 0
)=2] (s 0
) ; (A.5)
s0 =1;2

for j=1; 2: The coe¢ cients de…ned in (A.5) attribute one-year expected losses to loans rated
j=1; 2 in state s by taking into account their PD and LGD over each of the possible states s0
that can be reached at t + 1; where the corresponding s0 are weighted by their probability of
occurring given s: The losses associated these one-year ahead defaults are discounted using
the contractual interest rate of the loans, cz , as set at their origination. In Section A.10, we
derive an expression for the endogenous value of such rate under our assumptions on loan
pricing. As for the loans that are already non-performing (j=3) at date t; the term bbyt in
(A.4) implies attributing their conditional-on-s LGD to them, exactly as in (A.2).

41
A.4 Prudential expected losses under the IRB approach
Di¤erences between the BCBS prescriptions on expected losses for IRB loan portfolios and
the above de…nition of EL1Yt include the absence of discounting ( = 1), the preference
for using TTC (rather than PIT) PDs, and the usage of a downturn LGD that re‡ects the
depressed recoveries obtained under adverse circumstances. Thus, the prudential expected
losses de…ned by the BCBS for IRB portfolios can be found as

ELIRB
t = byt ; (A.6)

with b = (b1 ; b2 ; b3 ; b1 ; b2 ; b3 ; b1 ; b2 ; b3 ; b1 ; b2 ; b3 ) and bj = P Dj e(2); where

X
P Dj = i P Dj (si ) (A.7)
i=1;2

for j=1; 2; is the TTC PD for loans rated j (with i denoting the unconditional probability
of aggregate state i) and P D3 = 1.

A.5 Discounted lifetime expected losses

Impairment allowances computed on an lifetime-expected basis imply taking into account
not just the default events that may a¤ect the currently performing loans in the next year,
but also those occurring in any subsequent period. Building on prior notation and the same
approach explained for the model without aggregate risk, these provisions can be computed
as

ELLT
t = b yt + b M yt + b M 2 yt + b M 3 yt + ::: + bbyt
= b (I + M + M 2 + M 3 + :::)yt + bbyt
= b (I M ) 1 yt + bbyt = (b B + bb)yt ; (A.8)

with
1 Mp 06 6
M = ;
06 6 2 Mp

p11 M (1) p12 M (1)

Mp = ;
p21 M (2) p22 M (2)
0 1
m11 (s0 ) m12 (s0 ) 0
M (s0 ) = @ m21 (s0 ) m22 (s0 ) 0 A;
0 0 0 0 0
(1 3 (s )=2)P D1 (s ) (1 3 (s )=2)P D2 (s ) (1 3 (s ))
and mij (s0 ) = (1 j (s
0
))aij (s0 ).

42
A.6 Current expected credit losses (CECL)
CECL as stipulated by US GAAP relies on a discounted lifetime notion of expected losses
like ELLT
t above, but uses a discount factor r = 1=(1 + r) based on a reference risk free
rate rather than = 1=(1 + c); which is based on the contractual interest rate of the loans.
So, similarly to (A.8), we can write:

ELCECL
t = b r (I M r ) 1 yt + bbyt = (b r B r
+ bb)yt ; (A.9)

where b r
= ( r b; r b) and
r Mp 06 6
M = :
r
06 6 r Mp

A.7 Discounted expected losses under IFRS 9

IFRS 9 adopts a hybrid approach that combines the one-year-ahead and lifetime approaches
described in equations in (A.4) and (A.8). Speci…cally, it applies the one-year-ahead mea-
surement to loans whose credit quality has not increased signi…cantly since origination. For
us, these are the loans with j=1, namely those in the components xt (z; s; 1) of yt : By contrast,
it considers the lifetime expected losses for loans whose credit risk has signi…cantly increased
since origination. For us, these are the loans with j=2; namely those in the components
xt (z; s; 2) of yt :
As in the case without aggregate risk, it is convenient to split vector yt into a new auxiliary
vector
0 1
xt (1; 1; 1)
B 0 C
B C
B 0 C
B C
B xt (1; 2; 1) C
B C
B 0 C
B C
B 0 C
y^t = B
B
C;
C
B xt (2; 1; 1) C
B 0 C
B C
B 0 C
B C
B xt (2; 2; 1) C
B C
@ 0 A
0
which contains the loans with j=1; and the di¤erence

y~t = yt y^t ;

which contains the rest.

43
 This allows to express loan loss provisions under IFRS 9 as:34

ELtIF RS9 = b y^t + b B y~t + bbyt : (A.10)

A.8 Comparison between the various allowance measures

The above de…nitions clearly imply

ELIF
t
RS9
= ELLT
t b (B I)^
yt ELLT
t (A.11)

and
ELIF
t
RS9
= ELt1Y + b (B I)~
yt EL1Y
t ILt : (A.12)
Additionally, the de…nitions of and r; together with the fact that c r 0; imply

ELCECL
t ELLT
t : (A.13)

A.9 Implications for pro…ts and regulatory capital

By trivially extending the formula derived for the case without aggregate risk, the result of
the P/L account with aggregate risk can be written as
( )
X X (s )
P Dj (st )e(st ) xt-1 (z; st ; j)– 3 (st )e(st )xt-1 (z; st ; 3)
3 t
P Lt = cz (1–P Dj (st ))–
z=1,2 j=1,2
2
!
X X
r xt-1 (z; st ; j)–LLt-1 –kt-1 – LLt ; (A.14)
z=1,2 j=1,2,3

which di¤ers from (21) in the dependence of a number of the relevant parameters on the
aggregate state at the end of period t; st .
Dividends and equity injections are determined exactly as in () and ().
Finally, the minimum capital requirement under the IRB approach is now given by
X
k IRB
t = j (st )xjt ; (A.15)
j=1;2

where
h P i " ! #
1
1+ s0 p s0 st 0
j (s )
–2:5 mj 1
(P Dj ) + cor0:5 1
(0:999)
~ (s2 ) j
j (st )= –P Dj ;
1–1:5mj (1–corj )0:5
(A.16)
34
These de…nitions clearly imply ELtIF RS9 = ELLT
t b (B I)^
yt ELLT
t and ELIFt
RS9
= EL1Y
t +
1Y
b (B I)~
yt ELt : Additionally, the de…nitions of and r together with the fact that c r 0; imply
ELIRB
t EL1Y
t and ELCECL
t ELLT
t :

44
mj = [0:11852 0:05478 ln(P Dj )]2 ; and corj = 0:24 0:12(1 exp( 50P Dj ))=(1 exp( 50)):
Equation (A.7) implies assuming that the bank follows a strict TTC approach to the cal-
culation of capital requirements (which avoids adding cyclicality to the system through this
channel).35

A.10 Determining the contractual loan rate

Taking advantage of the recursivity of the model, for given values of the contractual interest
rates cz of the loans originated in each of the aggregate states z=1,2; one can obtain the
ex-coupon value of a loan originated in state z, when the current aggregate state is s and
their current rating is j, vj (z; s); by solving the system of Bellman-type equations given by:
X h
vj (z; s) = ps0 s (1 P Dj (s0 ))cz + (1 P Dj (s0 )) j (s0 ) + P Dj (s0 )( 3 (s0 )=2)(1 e(s0 ))
s0 =1;2

+m1j (s0 )v1 (z; s0 ) + m2j (s0 )v2 (z; s0 ) + m3j (s0 )v3 (z; s0 )] ; (A.17)

for j 2 f1; 2g and (z; s) 2 f1; 2g f1; 2g, and

X
vj (z; s) = ps0 s [ 3 (s0 )(1 e(s0 )) + (1 3 (s
0
))v3 (z; s0 )];
s0 =1;2

for j=3 and (z; s) 2 f1; 2g f1; 2g:

Under perfect competition and using the fact that all loans are assumed to be of credit
quality j=1 at origination, the interest rates cz can be found as those that make v1 (z; z) = 1
for z=1,2, respectively.

35
P
A PIT approach would imply setting P Dj (st ) = s0 ps0 st P Dj (s0 ) instead of P Dj and (st ) instead of
~ (s2 ) in (A.16).

45
B Calibration details
B.1 Migration and default rates for our two non-default states
We calibrate the migration and default probabilities of our two non-default loan categories
using S&P rating migration data based on a …ner rating partition. To map the S&P partition
into our partition, we start considering the 7 7 matrix A e obtained by averaging the yearly
matrices provided by S&P global corporate default studies covering the period from 1981
to 2015. This matrix describes the average yearly migrations across the seven non-default
ratings in the main S&P classi…cation, namely AAA, AA, A, BBB, BB, B and CCC/C.36
Under our convention, each element a ~ij of this matrix denotes a loan’s probability of migrat-
ing to S&P rating i from S&P rating j, and the yearly probability of default corresponding
g P7 e
to S&P rating j can be found as P Dj = 1 ~ij : With the referred data, we obtain A:
ai=1
0 1
0.8960 0.0054 0.0005 0.0002 0.0002 0.0000 0.0007
B 0.0967 0.9073 0.0209 0.0022 0.0008 0.0006 0.0000 C
B C
B 0.0048 0.0798 0.9161 0.0463 0.0034 0.0026 0.0022 C
B C
e
A=B 0.0010 0.0056 0.0557 0.8930 0.0626 0.0034 0.0039 C; (B.1)
B C
B 0.0005 0.0007 0.0044 0.0465 0.8343 0.0618 0.0112 C
B C
@ 0.0003 0.0009 0.0017 0.0082 0.0809 0.8392 0.1390 A
0.0006 0.0002 0.0002 0.0013 0.0079 0.0432 0.5752

which implies
T
g
P D = (0:0000; 0:0002; 0:0005; 0:0023; 0:0100; 0:0493; 0:2678):

To calibrate our model, we want to …rst collapse the above seven-state Markov process
into the two-state one speci…ed in our benchmark model without aggregate risk. We want to
obtain its 2 2 transition probability matrix, which we denote A, and the implied probabilities
P2
of default in each state, P Dj = 1 i=1 aij for j=1,2. To collapse the seven-state process
into the two-state process, we assume that the S&P states 1 to 5 (AAA, AA, A, BBB,
BB) correspond to our state 1 and S&P states 6 to 7 (B, CCC/C) to our state 2. We also
assume that all the loans originated by the bank belong to the BB category, so that the
vector representing the entry of new loans in steady state under the S&P classi…cation is
eeT = (0; 0; 0; 0; 1; 0; 0): Under these assumptions, we produce an average PD for the steady
state portfolio of the model without aggregate risk of 1.88%, slightly below the 2.5% average
PD on non-defaulted exposures of reported by the EBA (2013, Figure 12) for the period
from the …rst half of 2009 to the second half of 2012 for a sample of EU banks using the IRB
approach.
36
We have reweighted the original migration rates in S&P matrices to avoid having “non-rated” as an
eighth possible non-default category to which to migrate.

46
 The steady state portfolio under the S&P classi…cation can be found as z = [I7 7
f] 1 ee; where the matrix M
M f has elements m e ij = (1 aij and j is the independent
j )e
probability of a loan rated j maturing at the end of period t. For the calibration we set
j =0.20 across all categories, so that loans have an average maturity of …ve years. The
P P
“collapsed”steady state portfolio x associated with z has x1 = 5j=1 zj and x2 = 7j=6 zj .
For the collapsed portfolio, we construct the 2 2 transition matrix M (that accounts
for loan maturity) as
0 P5 P5 P7 P5 1
j=1 e ij zj
i=1 m j=6 e ij zj
i=1 m
0
B P5 Px71 P7 Px72 C
M =@ B j=1 i=6 e
m z
ij j j=6 i=6 e
m z
ij j C; (B.2)
x1 x2
0 A
(1 3 =2)P D1 (1 3 =2)P D2 (1 3)

where the probabilities of default for the collapsed categories are found as
P5 g
j=1 P Dj zj
P D1 = ; (B.3)
x1

and P7 g
P Dj zj
j=6
P D2 = : (B.4)
x2
Putting it in words, we …nd the moments describing the dynamics of the collapsed portfolio
as weighted averages of those of the original distribution, with the weights determined by
the steady state composition of the collapsed categories in terms of the initial categories.

B.2 State contingent migration matrices

Calibrating the full model with aggregate risk on which we base our quantitative analysis
requires calibrating state contingent versions of the matrix M found in (B.2), namely the
matrices M (s) for aggregate states s = 1; 2 that appear in the formulas derived in Appendix
A. We …nd M (1) and M (2) following a procedure analogous to that used to obtain M in (B.2)
e
but starting from state-contingent versions, A(1) e
and A(2), of the 7 7 migration matrix Ae
in (B.1). As described in B.1, we can go from each A(s)e to the maturity adjusted matrix
f
M (s) with elements m e ij (s) = (1 aij and then …nd the elements of M (s) as weighted
j )e
averages of the elements of M f(s): To keep things simple, we use the same unconditional
weights as in (B.2), implying
0 P5 P5 P7 P5 1
j=1 i=1 e ij (s)zj
m j=6 i=1 e ij (s)zj
m
0
B P5 P7x1 P7 P7x2 C
M (s) = B
@ j=1 i=6 e ij (s)zj
m j=6 i=6 e ij (s)zj
m
0
C
A
x1 x2
(1 3 (s)=2)P D1 (s) (1 3 (s)=2)P D2 (s) (1 3 (s))

47
where P5 g
P Dj (s)zj
j=1
P D1 (s) = ;
x1
P7 g
j=6 P D j (s)zj
P D2 (s) = ;
x2
g P7
with P Dj (s) = 1 i=1 a
~ij (s):
e
We calibrate A(1) e
and A(2) exploring the business cycle sensitivity of S&P yearly mi-
gration matrices previously averaged to …nd A: e We identify state s=1 with expansion years
and s=2 with contraction years. We use the years identi…ed by the NBER as the start of
the recession to identify the entry in state s=2 and assume that each of the contractions
observed in the period from 1981 to 2015 lasted exactly two years. This is consistent with the
NBER dating of US recessions except for the recession started in 2001, to which the NBER
attributes a duration of less than one year. However, the behavior of corporate ratings mi-
grations and defaults around such recession does not suggest it was shorter for our purposes
than the other three. To illustrate this, Figure B.1 depicts the time series of two of the
elements of the yearly default rates P g Dj and migration matrices A e whose cyclical behavior
is more evident: (i) the default rate among BB exposures (P gD5 ) and (ii) the migration rate
from a B rating to a CCC/C rating (~ a7;6 ). Year 2002 emerges clearly as a year of marked
deterioration in credit quality among exposures rated BB and B.

14

12

10

0
19 1
19 2
19 3
19 4
19 5
19 6
19 7
19 8
19 9
19 0
19 1
19 2
19 3
19 4
19 5
19 6
19 7
19 8
20 9
20 0
20 1
20 2
20 3
20 4
20 5
20 6
20 7
20 8
20 9
20 0
20 1
20 2
20 3
2014
15
8
8
8
8
8
8
8
8
8
9
9
9
9
9
9
9
9
9
9
0
0
0
0
0
0
0
0
0
0
1
1
1
1
19

Figure B.1. Sensitivity of default and migrations rates to aggregate states

Selected yearly S&P default and downgrading rates. Grey bars identify
2-year periods following the start of NBER recessions

e
In light of this, we estimate A(2) e extracted
by averaging the yearly counterparts of A
e
from S&P data for years 1981, 1982, 1990, 1991, 2001, 2002, 2008 and 2009, and A(1) by

48
averaging those corresponding to all the remaining years. This leads to
0 1
0.8923 0.0057 0.0005 0.0002 0.0002 0.0000 0.0000
B 0.1012 0.9203 0.0209 0.0023 0.0007 0.0003 0.0000 C
B C
B 0.0039 0.0668 0.9228 0.0500 0.0036 0.0025 0.0027 C
B C
e
A(1) = B C;
B 0.0010 0.0058 0.0495 0.8939 0.0668 0.0036 0.0043 C
B 0.0007 0.0002 0.0040 0.0429 0.8484 0.0679 0.0117 C
B C
@ 0.0000 0.0009 0.0020 0.0084 0.0680 0.8511 0.1548 A
0.0000 0.0002 0.0001 0.0009 0.0059 0.0360 0.5860

implying
g
P D(1)T = (0:0000; 0:0001; 0:0002; 0:0014; 0:0063; 0:0386; 0:2405);
and 0 1
0.9087 0.0044 0.0003 0.0005 0.0002 0.0000 0.0030
B 0.0786 0.8632 0.0209 0.0014 0.0013 0.0017 0.0000 C
B C
B 0.0077 0.1237 0.8936 0.0340 0.0026 0.0027 0.0009 C
B C
e =B
A(2) 0.0010 0.0050 0.0767 0.8899 0.0482 0.0028 0.0024 C;
B C
B 0.0000 0.0022 0.0057 0.0587 0.7865 0.0411 0.0095 C
B C
@ 0.0013 0.0007 0.0008 0.0076 0.1245 0.7988 0.0858 A
0.0027 0.0002 0.0006 0.0025 0.0143 0.0676 0.5389
implying
g
P D(2)T = (0:0000; 0:0005; 0:0014; 0:0054; 0:0224; 0:0853; 0:3596):
Finally, we set p12 =Prob(st+1 = 1jst = 2) equal to 0.5 so that contractions have an
expected duration of two years, and p21 =Prob(st+1 = 2jst = 1) equal to 0.148 so that
expansion periods have the same average duration as the ones observed in our sample period,
(35-8)/4=6.75 years.

B.3 Calibrating defaulted loans’resolution rate

The yearly probability of resolution of NPLs, 3 ; is calibrated so that the model without
aggregate risk fed with unconditional means of the credit risk parameters matches the 5%
average probability of default including defaulted exposures (P DID) that the EBA (2013,
Figure 10) reports for the second half of 2008, right before the stock of NPLs in Europe got
in‡ated by the impact of the Global Financial Crisis. The value of P DID for the steady
state portfolio obtained in the absence of aggregate risk can be computed as
P D1 x 1 + P D2 x 2 + x 3
P DID = P3 ; (B.5)
j=1 xj

where P D1 and P D2 are the unconditional mean probabilities of default for standard and
substandard loans obtainable from S&P data using (B.3) and (B.4), respectively (and the

49
procedure explained around those equations). Solving for x3 in (B.5) allows us to set a target
for x3 consistent with the target for P DID:
P D1 x1 + P D2 x2 (x1 + x2 )P DID
x3 = : (B.6)
P DID 1
The law of motion of NPLs evaluated at the steady state implies

x3 = (1 3 =2)P D1 x1 + (1 3 =2)P D2 x2 + (1 3 )x3 ; (B.7)

where it should be noted that the dynamic system in (1) allows us to compute x1 and x2
independently from the value of 3 : But, then, solving for 3 in (B.7) yields
2(P D1 x1 + P D2 x2 )
3 = ; (B.8)
P D1 x1 + P D2 x2 + 2x3
which allows us to calibrate 3 using x1 ; x2 ; and the target for x3 found in (B.6).

B.4 Can professional forecasters predict recessions?

The baseline quantitative results of the model are based on the assumption that changes
in the aggregate state st 2 f1; 2g cannot be predicted beyond what the knowledge of the
time-invariant state transition probabilities of the Markov chain followed by st allows (that
is, attributing some probability to the continuation in the prior state and a complementary
one to switching to the other state). At the other side of the spectrum, several papers assess
the cyclical properties of the new ECL approach to provisions using historical data and the
assumption of perfect foresight or that banks can perfectly foresee the losses coming in some
speci…ed horizon (e.g. two years in Cohen and Edwards, 2017, or two quarters in Chae et
al., 2017). In Section C.3 of the main text, we explore how our own results get modi…ed if
banks can foresee the arrival of a recession one year in advance. However, banks’capacity
to anticipate turning points in the business cycle, and especially switches from expansion
to recession, can be contended. There is a long research tradition in econometrics trying to
predict turning points but the state of the question can be summarized by saying that there
are a variety of indicators which allow to “nowcast” recessions (that is, to state that the
economy has just entered a recession) but have little or no capacity to “forecast”recessions
(see Harding and Pagan, 2010).
The same disappointing conclusion arises from the observation of the so-called Anxious
Index published by the Federal Reserve Bank of Philadelphia (at https://www.philadelphiafed
.org/research-and-data/real-time-center/survey-of-professional-forecasters/anxious-index).
Such index re‡ects professional forecasters’median estimate of the probability of experienc-
ing negative GDP growth in the quarter following the one in which the forecasters’views are
surveyed. The index, which can be traced back to mid 1968 thanks to the data maintained

50
by the Federal Reserve Bank of Philadelphia, is reproduced in Figure B.1. As one can see,
it does not systematically rise above good-times levels in the proximity of U.S. recessions
(marked as grey shaded areas), with the main exception of the second oil crisis in 1980.

Source: https://www.philadelphiafed.org/research-and-data/real-time-center/survey-of-professional-forecasters/anxious-index

Figure B.2. The Anxious Index

Professional forecasters’median probability of decline in Real GDP
in the following quarter. Grey bars identify NBER recessions.

51
C Complementary results
C.1 Quantitative results under SA capital requirements
For portfolios operated under the SA approach, the regulatory minimum capital requirement
applicable to loans to corporations without an external rating is just 8% of the exposure net
of its “speci…c provisions,” a regulatory concept related to impairment allowances (BCBS,
2017, paragraphs 1, 38 and 90). Assuming that all the loans in xt correspond to unrated
borrowers and that all the loan loss allowances LLt qualify as speci…c provisions, this implies
that !
X
k SA
t = 0:08 xjt LLt : (B.9)
j=1;2;3

Capital requirements for banks following the standardized approach (SA banks) apply
to exposures net of speci…c provisions and, hence, are sensitive to how those provisions
are computed. Thus, Table C.1 includes the same variables as Table 3 for IRB banks
together with details on the minimum capital requirement implied by each of the impairment
measurement methods. Except for the minimum capital requirement and the implied size
of a fully loaded CCB, all the other variables in Table 2 are equally valid for IRB and SA
banks.
The results in Table C.1 are qualitatively very similar to those described for an IRB bank
in Table 3, with some quantitative di¤erences that are worth commenting on. It turns out
that, in our calibration, an SA bank holding exactly the same loan portfolio as an IRB bank
would be able to support it with lower average levels of CET1 (between 130 basis points and
210 basis points lower, depending on the impairment measurement method). Therefore, in a
typical year, our SA bank features de facto higher leverage levels, and hence higher interest
expenses than its IRB counterpart. This explains why its P/L is slightly lower than that of
an IRB bank. This di¤erence explains most of the level di¤erences which can be seen in the
remaining variables in Table C.1.
When comparing impairment measurement methods in the case of an SA bank, the
di¤erences are very similar to those observed in Table 3 for IRB banks. The higher state-
dependence of the more forward-looking measures explains the higher cross-state di¤erences
in CET1, dividends and probabilities of needing capital injections under such measures. As
for IRB banks, the di¤erences associated with IFRS 9 relative to either the incurred loss
approach or the one-year expected loss approach are signi…cant, but not huge.

52
 Table C.1
Endogenous variables under SA capital requirements
(SA bank, as a percentage of mean exposures unless otherwise indicated)
IL ELIRB ELCECL ELIF RS9
P/L
Unconditional mean 0.15 0.17 0.20 0.17
Conditional mean, expansions 0.38 0.42 0.51 0.48
Conditional mean, contractions -0.62 -0.69 -0.85 -0.88
Standard deviation 0.43 0.47 0.60 0.59
Minimum capital requirement
Unconditional mean 7.75 7.52 6.95 7.42
Conditional mean, expansions 7.79 7.57 7.03 7.49
Conditional mean, contractions 7.62 7.35 6.71 7.18
Standard deviation 0.14 0.16 0.18 0.19
CET1
Unconditional mean 9.67 9.39 8.72 9.26
Conditional mean, expansions 9.90 9.65 9.06 9.61
Conditional mean, contractions 8.84 8.46 7.54 8.07
Standard deviation 0.88 0.88 0.84 0.90
Probability of dividends being paid (%)
Unconditional 50.92 50.47 57.35 52.32
Conditional, expansions 65.99 65.41 74.33 67.81
Conditional, contractions 0 0 0 0
Dividends, if positive
Conditional mean, expansions 0.34 0.37 0.38 0.38
Conditional mean, contractions – – – –
Probability of having to recapitalize (%)
Unconditional 3.68 3.92 4.45 4.66
Conditional, expansions 0 0 0 0
Conditional, contractions 16.13 17.18 19.50 20.40
Recapitalization, if positive
Conditional mean, expansions – – – –
Conditional mean, contractions 0.54 0.54 0.50 0.53

Comparing the results in Tables 3 and C.1 leads to the conclusion that the e¤ects on SA
banks of IFRS 9 and CECL are quantitatively very similar to those on IRB banks. As for
IFRS 9, this is further con…rmed by Figure C.1, which shows the counterpart of Figure 3
for a bank operating under the SA approach. It depicts 500 simulated trajectories for CET1
under IL and ELIF RS9 . As in Figure 3, it takes four consecutive years in the contraction
state (s=2) for an SA bank under IFRS 9 to use up its CCB and require a recapitalization,
while under the incurred loss method, the CCB would be fully depleted only after (roughly)

53
…ve years in the contraction state.37

Panel A. CET1 under IL Panel B. CET1 under ELIF RS9

10.5 10.5

10 10

9.5 9.5

9 9

8.5 8.5

8 8

7.5 7.5

7 7

6.5 6.5
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure C.1. CET1 after the arrival of a contraction (SA bank)

500 simulated trajectories of CET1 under IL and ELIF RS9
following the arrival of s=2 after a long period in s=1
(SA bank, as a percentage of average exposures)

C.2 Especially severe crises

In this section, we explore whether the severity of crises and the potential anticipation of a
particularly severe crisis make a di¤erence in terms of our assessment of the cyclicality of
the new more forward-looking provisioning methods (IFRS 9 and CECL) vis-a-vis the prior
less forward-looking measures (incurred losses and one-year expected losses). For brevity,
we again focus on IRB banks and on the comparison of one of the more forward looking
approaches. IFRS 9, with just one of the alternatives, namely the one-year expected loss
approach (the one so far prescribed by regulation for IRB banks).

C.2.1 Unanticipatedly long crises

We …rst explore what happens with the dynamic responses analyzed in the benchmark cal-
ibration with aggregate risk when we condition them on the realization of the contraction
state s=2 for four consecutive periods starting from t=0. So, as in the analysis shown in Fig-
ure 2, we assume that the bank starts at t=–1 with the portfolio and impairment allowances
resulting from having been in the expansion state (s=1) for a long enough period, and that
at t=0 the aggregate state switches to contraction (s=2).
37
In this case, the dashed lines that delimit the band within which CET1 evolves are averages across
simulated trajectories and across provisioning methods, since the sizes of the minimum capital requirement
and the minimum capital requirement plus the fully-loaded CCB depend on the size of the corresponding
provisions.

54
 Panel A. Non-performing loans Panel C. P/L under ELIRB and ELIF RS9
6 0.5

5.5

5
0
4.5

3.5
-0.5
3

2.5

2 -1
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Panel B. ELIRB and ELIF RS9 Panel D. CET1 under ELIRB and ELIF RS9
4 12.5

12
3.5
11.5

11
3

10.5

2.5
10

9.5
2
9

1.5 8.5
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure C.2. Unanticipatedly long crises

Average responses to the arrival of s=2 when the contraction is
unanticipatedly “long” (thick lines) rather than “average” (thin lines)
(IRB bank, as a percentage of average exposures)

In Figure C.2, we compare the average response trajectories already shown in Figure 2
(where, from t=1 onwards, the aggregate state evolves stochastically according to the Markov
chain calibrated in Table 1) with trajectories conditional on remaining in state s=2 for at
least up to date t=3 (four years).38
When a crisis is longer than expected, the largest di¤erential impact of ELIF RS9 relative
to ELIRB still happens in the …rst year of the crisis (t=0), since ELIF RS9 front-loads the
expected beyond-one-year losses of the stage 2 loans. In years two to four of the crisis
38
In the conditional trajectories, the aggregate state is again assumed to evolve according to the calibrated
Markov chain from t=4 onwards.

55
(t=1,2,3) the di¤erential impact of IFRS 9 (compared to one-year) expected losses on P/L
lessens before it switches sign (after t=4). In the …rst years of the crisis, ELIF RS9 leaves
CET1 closer to the recapitalization band and, in the fourth year (t=3), the duration of the
crisis forces the bank to recapitalize only under ELIF RS9 . However, ELIF RS9 supports a
quicker recovery of pro…tability and, hence, CET1 after t=4.

C.2.2 Anticipatedly long crises

We now turn to the case in which crises can be anticipated to be long from their outset.
To study this case, we extend the model to add a third aggregate state that describes “long
crises” (s=3) as opposed to “short crises” (s=2) or “expansions” (s=1). To streamline the
analysis, we make s=2 and s=3 have exactly the same impact on credit risk parameters as
prior s=2 in Table 2, and keep the impact of s=1 on credit risk parameters also exactly the
same as in Table 2. The only di¤erence between states s=2 and s=3 is their persistence,
which determines the average time it takes for a crisis period to come to an end. Speci…cally,
we consider the following transition probability matrix for the aggregate state:
0 1 0 1
p11 p12 p13 0:8520 0:6348 0:250
@ p21 p22 p23 A = @ 0:1221 0:3652 0 A; (B.10)
p31 p32 p33 0:0259 0 0:750

which implies an average duration of four years for long crises (s=3), 1.6 years for short
crises (s=2), and the same duration as in our benchmark calibration for periods of expansion
(s=1). The parameters in (B.10) are calibrated to make s=3 to occur with an unconditional
frequency of 8% (equivalent to su¤ering an average of two long crises per century) and to
keep the unconditional frequency of s=1 at the same 77% as in our benchmark calibration.
In Figure C.3 we compare the average response trajectories that follow the entry in state
s=2 (thin lines) or state s=3 (thick lines) after having spent a su¢ ciently long period in state
s=1. Therefore, the …gure illustrates the average di¤erences between entering a “normal”
short crisis or a “less frequent” long crisis at t=0. Opposite to ELIRB (whose credit risk
parameters are state-independent), ELIF RS9 behave di¤erently across short and long crises
from the very …rst period because it factors in the lower probability of a recovery at t=1
under s=3 than under s=2. ELIF RS9 is also more reactive in subsequent periods because,
for stage 2 loans, it takes into account the losses further into the future.
These di¤erence also explain the larger initial impact of a severe crisis on P/L and CET1
under IFRS 9 than under the prudential expected losses of the IRB approach. As a result, at
the onset of an anticipatedly long crisis, ELIF RS9 pushes CET1 closer to the recapitalization
band and the di¤erence with respect to ELIRB increases. Quantitatively, however, the e¤ect
on CET1 is still moderate, using up on impact less than half of the fully loaded CCB. Of

56
course, later on in the long crisis, ELIF RS9 results, on average, in a quicker recovery of
pro…tability and CET1 than ELIRB :

Panel A. Non-performing loans Panel C. P/L under ELIRB and ELIF RS9
5 0.5

4.5
0
4

3.5 -0.5

-1
2.5

2 -1.5
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Panel B. ELIRB and ELIF RS9 Panel D. CET1 under ELIRB and ELIF RS9
4 12.5

12
3.5
11.5

11
3

10.5

2.5
10

9.5
2
9

1.5 8.5
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure C.3. Anticipatedly long crises

Average responses to the arrival of a contraction at t=0 when it is anticipated
to be “long” (s0 =3, thick lines) rather than “normal” (s0 =2, thin lines)
(IRB bank, as a percentage of average exposures)

As a quantitative summary of the implications of an anticipatedly long crisis, the following

table reports the unconditional yearly probabilities of the bank needing equity injections,
under each of the impairment measures compared, in the baseline model and in the current
extension:

IL ELIRB ELCECL ELIF RS9

Baseline model 2.92% 2.91% 3.06% 4.16%
Model with anticipatedly long crises 3.83% 3.82% 5.18% 5.33%

57
 Note that the anticipatedly long crises signi…cantly increase the probability of having
to recapitalize banks under the CECL approach, making its procyclicality very similar (as
measured by this variable) to the one obtained under IFRS 9.

C.3 Better foreseeable crises

We now consider the case in which some crises can be foreseen one year in advance.39 Similar
to the treatment of long crises in the previous subsection, we formalize this by introducing
a third aggregate state, s=3, which describes normal or expansion states in which a crisis
(transition to state s0 =2) is expected in the next year with a larger than usual probability.
So we make s=3 identical to s=1 in all respects (that is, the way it a¤ects the PDs, rat-
ing migration probabilities, and LGDs of the loans, et cetera) except in the probability of
switching to aggregate state s0 = 2 in the next year.
To streamline the analysis, we look at the case in which s=3 is followed by s0 =1 with
probability one and assume that half of the crisis are preceded by s = 3 (while the other half
are preceded, as before, by s = 1; which means that they are not seen as coming). Adjusting
the transition probabilities to imply the same relative frequencies and expected durations
of non-crisis versus crisis periods as the baseline calibration in Table 1, the matrix of state
transition probabilities used for this exercise is
0 1 0 1
p11 p12 p13 0:8391 0:5 0
@ p21 p22 p23 A = @ 0:0740 0:5 1 A :
p31 p32 p33 0:0869 0 0

The thick lines in Figure C.4 show the average response trajectories to the arrival of
the pre-crisis state s0 =3 at t=–1 after having spent a long time in the normal state s=1.
We compare ELIRB and ELIF RS9 and include, using thin lines, the results of the baseline
model (regarding the arrival of s0 =2 at t=0 having been in s=1 for a long period). The
results con…rm the notion that being able to better anticipate the arrival of a crisis helps to
considerably soften its impact on IFRS 9 provisions and the subsequent e¤ects on P/L and
CET1.
39
See Section B.4 of Appendix B for a discussion of the degree to which professional forecasters are able
to predict recessions.

58
 Panel A. Non-performing loans Panel C. P/L under ELIRB and ELIF RS9
4.5 0.6

0.4
4
0.2

0
3.5
-0.2

-0.4
3
-0.6

2.5 -0.8

-1

2 -1.2
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Panel B. ELIRB and ELIF RS9 Panel D. CET1 under ELIRB and ELIF RS9
3 12.5

12

11.5
2.5

11

10.5

2
10

9.5

1.5 9
-1 0 1 2 3 4 5 6 7 8 9 10 11 12 -1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure C.4. Better foreseeable crises

Average responses to the arrival of pre-crisis state at t=–1 after long in s=1 (thick
lines). Thin lines describe the arrival of s=2 at t=0 in the baseline model
(IRB bank, as a percentage of average exposures)

Finally, as in the previous extension, the following table reports the unconditional yearly
probabilities of the bank needing equity injections under each of the compared provisioning
methods, in the baseline model and in the current extension. Indeed, crises that are better
anticipated imply a lower yearly probability that the bank needs an equity injection (and
the reduction in such probability relative to the baseline model is more sizeable under IFRS
9). Yet, the ranking of the various provisioning methods in terms of this variable remains
the same as in the baseline model, with IFRS 9 performing the worst:

IL ELIRB ELCECL ELIF RS9

Baseline model 2.92% 2.91% 3.06% 4.16%
Model with better foreseeable crises 2.43% 2.45% 2.45% 2.90%

59
C.4 Other possible extensions
In this section, we brie‡y describe additional extensions that the model could accommodate
at some cost in terms of notational, computational, and calibration complexity.

Multiple standard and substandard ratings Adding more rating categories within the
broader standard and substandard categories would essentially imply expanding the dimen-
sionality of the vectors and matrices described in the baseline model and in the aggregate-risk
extension. If loans were assumed to be originated in more than just one category, the need to
keep track of the (various) contractual interest rates for discounting purposes means we would
need to expand the dimensionality of the model further. Alternatively, an equivalent and
potentially less notationally cumbersome possibility would be to consider the same number
of portfolios as di¤erent-at-origination loans and to aggregate across them the impairment
allowances and the implications for P/L and CET1.

Relative criterion for credit quality deterioration This extension would be a natural
further development of the previous one and only relevant for the assessment of IFRS 9.
Under IFRS 9, the shift to the lifetime approach (“stage 2”) for a given loan is supposed to
be applied not when an absolute substandard rating is attained, but when the deterioration
in terms of the rating at origination is signi…cant in relative terms, for example because the
rating has fallen by more than two or three notches. This distinction is relevant if operating
under a ratings scale that is …ner than the one we have used in our analysis. As in the case
with the above-mentioned multiple standard and substandard ratings, keeping the analysis
recursive under the relative criterion for treating loans as “stage 1” or “stage 2” loans in
IFRS 9 would require considering as many portfolios as di¤erent-at-origination loan ratings
and to rewrite the expressions for provisions so as to impute lifetime expected losses to
the components of each portfolio whose current rating is signi…cantly lower than the initial
rating.

60
Practical insights on
implementing IFRS 9
and CECL
Practical insights on implementing IFRS 9 and CECL

Synopsis of the FASB’s and IASB’s to IAS 39 Objective Evidence of Impairment

We are pleased to present the fourth expected credit impairment standards [OEI]), stage 3 requires that, in addition to
publication in a series1 that highlights Both the impairment model in IFRS 9 and recognition of full lifetime expected credit
Deloitte Advisory’s point of view the FASB’s CECL 3 model are based on losses, the institution should measure
about the significance of the Financial expected credit losses. The IASB differs interest income by applying the EIR to the
Accounting Standards Board’s (FASB) from FASB in that IFRS 9 uses a three-stage net carrying amount of the debt instrument.
ASU 2016-13 – Measurement of Credit approach. Under IFRS 9, debt instruments,4 The indicators of observable evidence
Losses on Financial Instruments and excluding purchased or originated credit (e.g. default, significant financial difficulty,
related implementation considerations impaired financial instruments, move etc.) are consistent with how a financial
and IFRS 92 - Financial Instruments. through three stages as credit quality institution applies IAS 39 OEI today.
Thought pieces that provide perspective changes. Consequently, a financial
and discuss potential implications institution would measure expected credit For these purchased or originated
of FASB’s current expected credit losses and recognize interest income assets, a financial institution recognizes
loss (CECL) model will continue to be depending upon the following stages: only the cumulative change in lifetime
published at www.deloitte.com/us/cecl. expected losses since initial recognition
Stage 1: Assets that are performing. If credit as a loss allowance. Changes in lifetime
The standard setters—FASB and the risk is low as of the reporting date or the expected losses since initial recognition
International Accounting Standards Board credit risk has not increased significantly are recognized in profit or loss. Thus, any
(IASB)—have overhauled the accounting since initial recognition, an entity will favorable change in lifetime expected
models for credit impairment. Institutions recognize a loss allowance at an amount credit losses since initial recognition of a
required to issue financial statements under equal to 12-month expected credit losses. purchased or originated credit-impaired
either standard are likely to encounter This amount of credit losses is intended to financial asset is recognized as an
significant implementation and operational represent lifetime expected credit losses impairment recovery. Interest income
challenges. As institutions with dual filing that will result if a default occurs in the 12 recognition will be through a credit-adjusted
requirements under International Financial months after the reporting date, weighted EIR multiplied by the amortized cost.
Reporting Standards (IFRS) and US Generally by the probability of that default occurring.
Accepted Accounting Principles (GAAP) For these debt instruments, interest income In contrast, the FASB’s CECL model requires
begin their implementation efforts, they recognition will be the effective interest rate entities to recognize lifetime expected credit
should formulate a broad strategic plan (EIR) multiplied by the gross losses for all assets, not just those for which
that factors in both the similarities and carrying amount. there has been a significant increase in
differences between the FASB’s and IASB’s credit risk since initial recognition. Stated
approaches. A broad plan that effectively Stage 2: Assets that have significant differently, the FASB’s model follows a
leverages interdependencies in credit risk increases in default risk. In instances where single credit-loss measurement approach,
management practices, operations, and credit risk has increased significantly since whereas IFRS 9 follows a dual credit-loss
reporting can: initial recognition, an entity would measure measurement approach in which expected
a loss allowance at an amount equal to credit losses are measured in stages to
•• Lay the foundation for an integrated
full lifetime expected credit losses. That is, reflect deterioration over a period of time.5
framework
the expected credit losses that result from Discussed below are additional differences
•• Facilitate a lean rollout all possible default events over the life of and similarities in the FASB’s and IASB’s
the financial instrument. For these debt credit impairment models.
This point of view discusses the intersection instruments, interest income recognition will
of requirements between FASB’s and be based on the EIR multiplied by the gross Significant credit deterioration
IASB’s credit impairment models. It also carrying amount. A major point of divergence between the
identifies potential opportunities to gain FASB’s and IASB’s impairment models is
implementation efficiencies. Stage 3: Credit impaired. For debt the fact that credit deterioration affects the
instruments that have both a significant amount of loss allowance an entity would
increase in credit risk plus observable recognize under IFRS 9. Under IFRS 9, debt
evidence of impairment (e.g., similar concept instruments are transferred between stages

03
 Practical insights on implementing IFRS 9 and CECL

as credit quality changes. Consequently, a only be used in rare circumstances and Data requirements and credit modeling
critical decision point in implementing IFRS lifetime expected credit losses are generally For all financial institutions, IFRS 9 and
9 is determining whether there has been a anticipated to be recognized before a CECL will bring a fundamental change in
significant deterioration in credit risk since missed payment occurs. how impairment of debt instruments is
origination. Depending upon whether a measured. In addition to the requirement
financial asset is in stage 1 or stage 2/stage BCBS guidance provides that banks should to model lifetime expected losses, issues
3, expected credit losses will be measured “have processes in place that enable around data quality, availability, and
as 12 months or lifetime, respectively. them to determine [significant credit risk] collection will likely be at the forefront of
on a timely and holistic basis so that an implementation efforts. The following data
Under IFRS 9, the assessment of whether individual exposure, or a group of exposures will likely be necessary to measure expected
there has been a significant increase with similar credit risk characteristics, is credit losses under both IFRS 9 and CECL:
in credit risk is based on an increased transferred to [lifetime expected credit
•• Historical defaults, attrition, and
probability of default since initial recognition. losses] measurement as soon as credit risk
recovery data
While IFRS 9 permits the use of various has increased significantly, in accordance
approaches to assess whether credit risk with the IFRS 9 impairment accounting •• Risk grades
has increased significantly, it also includes requirements.”10 The BCBS guidance also
•• Delinquency data
a rebuttable presumption that credit risk recommends that banks establish policies
has increased significantly when contractual and specific criteria for what constitutes •• Internal indicators of the likelihood to pay
payments are more than 30 days past a “significant” increase in credit risk for
•• Prepayments
due (DPD). different types of lending exposures.
Regulators across multiple geographies •• Collateral information
will likely expect alignment of credit risk
•• Forward-looking economic scenarios
assessment across products, business units,
Participants of Deloitte UK’s Global
and jurisdictions. •• Macroeconomic variables
IFRS Banking Survey – Sixth Edition,6
were asked if they expect to rebut the •• Origination lifetime probability of default
As a practical expedient, IFRS 9 provides
presumption of significant increase in
an exception for low credit risk exposures, •• Loss given default estimates
credit risk if they are more than 30
where “entities have the option not to
DPD. Responses ranged from 30 •• Exposure at default estimates
assess whether credit risk has increased
percent “never” and 62 percent
significantly since initial recognition. [The •• EIR
“rarely” to 8 percent “often.”
low credit risk exemption] was included
•• Full repayment
to reduce operational costs.”11 However,
it is the BCBS’s expectation “that use of •• Data required for disclosures
The Basel Committee on Banking this exemption should be limited.”11 In
Supervision (BCBS) published guidance7 addition, the BCBS expects banks to assess Lifetime modeling of credit risk will be
in December 2015 on credit risk and significant increases in credit risk for all dependent upon historical risk grades and
accounting for expected credit losses.8 The lending exposures in a timely manner. expectations of performance across these
guidance sets out supervisory expectations risk grades. Data relating to historical loss
for banks relating to sound credit risk As noted in Deloitte UK’s Global IFRS given default rates and recovery curves will
practices associated with implementing Banking Survey – Sixth Edition, be critical to credit modeling. Key inputs
an expected credit loss framework. participants who expect to use the and assumptions (e.g., loss migration rates,
It also highlights three IFRS 9-specific “low credit risk” exemption plan to delinquencies, and defaults) are typically
requirements9 banks should consider apply it mostly for securities and used as inputs for various purposes in an
when designing and operationalizing corporate loans versus other type of organization, such as estimating expected
their implementation plan. With respect debt instruments. Respondents with losses under US GAAP and IFRS, capital
to defining and measuring significant over £100 billion of gross lending, adequacy tests, stress tests, and credit
deterioration in credit risk, the BCBS is however, are less likely to use any of risk management. As an implementation
of the view that delinquency data should the practical expedients under IFRS 9.

04
Practical insights on implementing IFRS 9 and CECL

leading practice, institutions should consider Groupings based on similar or shared credit Under the CECL model, estimates of
identifying any differences in inputs and risk characteristics is an area where banks expected credit losses must reflect the
tracking the underlying rationale for why can align their methodology for pooling debt time value of money explicitly only when
differences exist. This should lay the instruments. Consistent practices can be a discounted cash flow approach is used
foundation for an integrated approach used to group exposures to assess credit to estimate expected credit losses. Other
across the organization for data sourcing, risk (such as by product type, product terms methods (e.g., loss-rate methods, roll-rate
key assumptions, and drivers of credit risk and conditions, industry/market segment, methods, probability-of-default methods,
and modeling. geographical location, or vintages) for and an aging schedule using loss factors)
both US GAAP and IFRS. Other examples are acceptable, even though they do not
Unit of account of shared characteristics include type of explicitly incorporate time value of money.
Under the CECL model, entities are required customer—wholesale or retail industry, date Because IFRS 9 does not permit time
to evaluate debt instrument assets on a of initial recognition, term to maturity, the value of money to be reflected implicitly
collective (i.e., pool) basis when similar quality of collateral, and the loan to value while the CECL model does, differences in
risk characteristics are shared. If risk (LTV) ratio. The BCBS guidance provides measurement can arise.
characteristics of a given debt instrument that “[g]roups should be sufficiently
are not similar to the risk characteristics of granular to allow banks to group exposures In principle, measurement of expected
any of the entity’s other debt instruments, into portfolios with shared credit risk credit losses is conceptually the same
the entity would evaluate the financial characteristics so that banks can reasonably under the FASB’s CECL model and stage
asset individually. If the debt instrument is assess changes in credit risk.”12 Adopting 2/stage 3 debt instruments under IFRS 9.
individually evaluated for expected credit cohesive policies and practices for grouping However, IFRS 9 requires the measurement
losses, the entity would not be allowed to debt instruments for estimating expected of expected credit losses to reflect an
ignore available external information such as credit losses under US GAAP and IFRS unbiased and probability-weighted amount
credit ratings and other credit loss statistics. can eliminate redundancies and alleviate that is determined by evaluating the range of
Under IFRS 9, the expectation is the same— operational burden. possible outcomes, as well as incorporating
expected credit losses should be measured the time value of money. More specifically,
on a collective basis if the debt instruments Measuring expected credit losses IFRS 9 defines expected credit losses as
share similar credit risk characteristics. This Both IFRS 9 and the FASB’s CECL model the weighted average of credit losses, with
collective assessment is also applicable for provide latitude in how expected credit the weightings being respective risks of a
determining whether significant increase in losses are estimated—an entity can use a default occurring. Although consideration
credit risk has occurred. number of measurement approaches to of all possible scenarios is not required, at a
determine the impairment allowance. Under minimum, the risk or probability that a credit
Expected credit losses on individual IFRS 9 and the CECL model, information loss occurs must be considered, even if the
large exposures and credit-impaired about past events, current conditions, and probability of a credit loss occurring is low.
loans are more likely to be estimated reasonable and supportable forecasts In contrast, the CECL model allows for the
individually, as compared to retail of future economic conditions should be use of a single forecast and does not require
exposures and other similar considered when measuring expected a probability-weighted measurement of
exposures where there is a lack of credit losses. The models differ in terms expected credit losses. Entities, therefore,
borrower-specific information of how the time value of money should be will need to measure expected credit losses
(delinquency, collective historical reflected in the estimate of expected credit on assets that have a low risk of loss. Thus,
experience of losses, and losses. Under IFRS 9, for non-purchased or if an institution does not adopt a centralized
forward-looking macroeconomic). originated credit impaired debt instruments, approach, there may be instances where
This results in credit losses typically expected losses must be discounted to the the estimate under IFRS 9 differs from the
being estimated on a collective basis. reporting date using the effective interest CECL expected loss estimate. To optimize
rate of the asset (or an approximation implementation efforts for such debt
thereof) that was determined at initial instruments, banks should develop a unified
recognition (i.e., time value of money is methodology for estimating expected credit
required to be incorporated explicitly). losses.

05
 Practical insights on implementing IFRS 9 and CECL

•• Identifying synergies between modeling techniques

Differences between standards used for stress testing, capital, loan loss reserve, and
Contractual life for credit cards other regulatory requirements

An additional difference between the standards is in determining the •• Centralizing and aligning data sourcing requirements
contractual life for credit cards (or other cancellable corporate facilities). to simplify data architecture design, eliminate
redundancies, reduce cost, and enhance operational
Under CECL, if an entity has the unconditional ability to cancel the efficiencies
unfunded portion of a loan commitment, the entity would not be
required to estimate expected credit losses on that portion, even if the Efforts to comply with the new credit impairment models
entity has historically never exercised its cancellation right. will likely create downstream impact on an institution’s
current business processes, control environment, and
However, under IFRS 9, financial institutions will be required to measure operating model. Institutions should consider adopting
expected credit losses over the period for which they are exposed to a broad approach to end-to-end process redesign.
credit risk. For example, revolving credit facilities, such as credit cards The following can reduce implementation efforts and
and overdraft facilities, can be contractually withdrawn by the lender operational burden:
with as little as one day’s notice. In practice, however, lenders continue to •• Identifying multi-purpose processes and converging
extend credit for a longer period and may only withdraw the facility after control points to enhance use of existing credit risk
the credit risk of the borrower increases. Companies will have to models, governance framework, validation processes,
determine the behavioral life for these debt instruments under IFRS 9 and financial and regulatory reporting processes. For
and will not have the same determination under CECL. example, processes and controls related to Basel III,
Comprehensive Capital Analysis and Review (CCAR)/
Other minor differences Dodd-Frank Act Stress Testing (DFAST) processes, and
Modifications of financial assets is another known difference between eventually CECL have several points of convergence.
the standards. Under CECL a concession provided to a troubled borrower
is treated to be a continuation of the original lending agreement. •• Dual-purpose chart of account definitions and
However, the concept of a troubled debt restructuring does not exist classifications of both on- and off-balance-sheet
in IFRS 9. exposures can facilitate a holistic approach for balance
sheet mapping and ongoing exposure identification
In addition, IFRS 9 does not permit the application of nonaccrual activities.
practices while CECL does permit it.
Reap the benefits of an integrated approach
Adopting an integrated approach for end-to-end process
design and implementing standardized processes and
Building an integrated framework
controls can significantly alleviate implementation and
Identifying the intersection of requirements between the FASB’s and IASB’s
operational burden. Furthermore, a well-thought-out
credit impairment models and adopting a unified strategy for estimating
integrated approach should lend itself to a consistent
expected credit losses that effectively capitalizes on interdependencies can
framework and is more likely to be accepted by auditors
allow a financial institution to gain operational efficiencies and facilitate a lean
and regulators.
implementation program.

The following can facilitate a lean IFRS 9 implementation while laying the
foundation for CECL:

•• Identifying how existing capabilities (i.e., provision for loss framework and
parallel regulatory and reporting processes) can be leveraged in estimating
expected credit losses

06
 Current Expected Credit Loss (CECL) Model and Banks’ External Information Environment

Samuel B. Bonsall IV
[email protected]
The Pennsylvania State University

Brent A. Schmidt
[email protected]
The Pennsylvania State University

Biqin Xie
[email protected]
Office of Financial Research, U.S. Department of the Treasury

February 2025

This paper was previously circulated under the title “Current Expected Credit Loss (CECL) Model and Analyst
Forecasts.” We acknowledge helpful comments from Riddha Basu, Gauri Bhat, Jeffrey Burks, Kurt Gee, Apoorv
Gogar, Yadav Gopalan, Michael Iselin, Zachary Kaplan, Jed Neilson, Allison Nicoletti, Jacob Ott (discussant),
Stephen Penman, Joshua Ronen, Jalal Sani, Barbara Su, Ayung Tseng, Mihail Velikov, Barrett Wheeler, and
participants at the 2022 Accounting Design Project, 2023 FARS Midyear Meeting, 2023 Penn State Accounting
Research Conference, and 2024 St. Louis Fed Workshop on Financial Institutions Research. We acknowledge
financial support from the Smeal College of Business. All errors are our own. The views expressed in the paper do
not necessarily reflect those of the U.S. Treasury or the Office of Financial Research. The authors of this paper have
no conflicts of interest related to this research.
 Current Expected Credit Loss (CECL) Model and Banks’ External Information Environment

Abstract
We investigate how the adoption of the Current Expected Credit Loss (CECL) standard affects US
banks’ external information environment. To this end, we examine how the standard impacts the
information content of three important information events: earnings announcements, analysts’ loan
loss provision (LLP) forecast revisions, and 10-K/Q filings. Using return variability (U-statistic)
following Beaver (1968), we show the informativeness of earnings announcements and analyst
provision forecasts decreases, while the informativeness of 10-K/Q filings increases, after CECL
adoption. When we aggregate CECL’s impacts across the three events, we find an overall
deterioration in banks’ external information environment, which persists throughout our entire
post-period. Further analyses reveal that earnings announcements’ information content decreases
more when CECL is expected to increase LLP volatility the most and the bank is covered by lower-
quality analysts. We also find a deterioration in the properties of analyst forecasts (i.e., higher
forecast errors, higher dispersion, and lower analyst following) for both quarterly provisions and
one-year-ahead target prices. Finally, we provide evidence that the increase in 10-K/Q filings’
information content is stronger when banks’ vintage disclosures, an important element of the
CECL standard, contain more news. Collectively, our evidence broadens our understanding of the
consequences of expected credit loss standards.
1. Introduction

In June 2016, the Financial Accounting Standards Board (FASB) issued the current

expected credit loss (“CECL”) standard (ASU 2016-13), replacing the incurred loss (IL) model of

credit loss recognition. Under the IL standard, banks were required to delay credit loss recognition

until it was probable that a loss had been incurred. In contrast, CECL requires managers to

recognize, at loan origination, all credit losses expected to occur during the lifetime of a loan.

CECL has been called the “most sweeping change to bank accounting ever” 1 and has generated

significant controversy.2 According to the FASB, “The main objective of [CECL] is to provide

financial statement users with more decision-useful information about the expected credit losses

on financial instruments and other commitments to extend credit held by a reporting entity at each

reporting date” (FASB 2016). Motivated by this stated goal, we examine how CECL adoption

affects banks’ external information environment. Specifically, we investigate the impact of CECL

on the information content of three important information events: earnings announcements, analyst

loan loss provision (LLP) forecast revisions, and 10-K/Q filings.

It is unclear ex ante how CECL adoption affects banks’ external information environment.

On the one hand, we would expect an improvement if the FASB’s stated objective of CECL is

achieved. Prior literature suggests that researcher-constructed estimates of expected credit losses

(i.e., estimated before CECL adoption) are useful in predicting future credit losses and bank

failures (Harris, Khan, and Nissim 2018; Lu and Nikolaev 2022) and in pricing bank stocks

(Wheeler 2021). Furthermore, concurrent studies examining CECL adoption suggest that CECL

improves information production by banks (Kim, Kim, Kleymenova, and Li 2023) and that day-1

1
See “ABA Position” on CECL (https://www.aba.com/advocacy/our-issues/cecl-implementation-challenges).
2
For example, some members of Congress (e.g., Rep. Blaine Luetkemeyer), banking trade groups (e.g., American
Banker Association, Banking Policy Institute), and CEOs or CFOs of some large U.S. banks (e.g., Capital One, BB&T
Corp.) all called for either a complete removal or at least an implementation delay of CECL.

1
CECL allowance information is decision-useful for investors (Gee, Neilson, Schmidt, and Xie

2024).3 These findings would suggest an improvement in banks’ information environment

following CECL adoption.

On the other hand, critics raise several concerns that suggest CECL could lead to a

deterioration in banks’ external information environment. First, CECL requires banks to

incorporate forward-looking information when estimating expected credit losses. Given the

uncertainty surrounding such information, critics argue that CECL introduces additional

uncertainty, and thus volatility, into loan loss provisions (Pan, Wang, and Wu 2017; Maurer 2020;

Bable, Wong, and Wynes 2022). If investors are Bayesian, the usefulness of a signal will be based

on the inverse of its volatility (or, in Bayesian terminology, its precision), suggesting that the

introduction of additional uncertainty into the provision under CECL could diminish its usefulness

as a signal to investors (e.g., Hilary and Hsu 2013; Hilary, Hsu, and Wang 2014). This would be

the case even if forecasts of expected loan losses under CECL are more accurate. Further, CECL

allows an entity to use judgment in determining the estimation method it uses, rather than

prescribing specific methods for measuring expected credit losses (see ASC 326-20-30-3 and 326-

20-55-7).4,5 These aspects of CECL could significantly increase the complexity of interpreting

earnings for investors and forecasting LLPs for analysts. Consistent with this idea, Bable et al.

(2022) interview sell-side analysts and find that they believe CECL disclosures are confusing and

require more research than under the IL model, consistent with an increase in information

3
López-Espinosa, Ormazabal, and Sakasai (2021) draw similar conclusions using expected-credit-loss-based
provisions under IFRS 9, the international counterpart to CECL.
4
The different methods mentioned in ASU 2016-13 include vintage analysis, probability-of-default methods,
discounted cash flow methods, loss-rate methods, and roll-rate methods (see FASB 2016, p. 109, para. 326-20-30-3).
5
The potential subjectivity of CECL estimates has been one of the biggest concerns expressed about the standard.
Even equity investors, who generally support CECL, are concerned that CECL “opens considerable room for
manipulation” and that the estimation methodology chosen by an entity “may become too complex to understand and
to challenge” (https://www.cfainstitute.org/-/media/documents/comment-letter/2010-2014/20130910.ashx).

2
processing costs. Taken together, the concerns raised by critics would suggest a deterioration in

banks’ external information environment after CECL adoption.

Our definition of a bank’s external information environment is motivated by Beyer, Cohen,

Lys, and Walther (2010), who define three aspects of a firm’s information environment: mandated

disclosure, voluntary disclosure, and disclosure by financial analysts. Within these three

categories, we determine which information events are likely to be affected by CECL adoption.

For mandated disclosure, we examine the information content of earnings announcements and 10-

K/Q filings. For voluntary disclosure, we note that any bundled earnings guidance issued by

management would be captured by our proxy for the informativeness of earnings announcements

and that there are only 84 total instances of unbundled guidance in our sample. Thus, we do not

examine any additional information events within the voluntary disclosure category. For disclosure

by financial analysts, we examine the information content of analyst LLP forecast revisions. See

Section 3 for further discussion regarding the information events we examine.

Our measure of information content is the Beaver (1968) U-statistic, which captures stock

return variability during the information event relative to the variability of stock returns at other

times. The prior literature uses this measure to examine the change in information events across

time (e.g., Beaver McNichols, and Wang 2020). It also allows us to assess the overall change in

banks’ information environment by aggregating the changes in the information content across the

three events we investigate. This approach is consistent with Beyer et al. (2010), who note, “…it

is necessary to consider multiple aspects of the corporate information environment in order to

conclude whether it becomes more or less informative in response to an exogenous change.” (p.

298). We use a seemingly unrelated regression model (Zellner 1962) to aggregate the changes in

information content across our three events.

3
 Using a difference-in-differences design during 2017–2022, with a sample of 156 treatment

banks adopting CECL on January 1, 2020 and 136 control banks that are public smaller reporting

companies (SRCs) exempted from adopting CECL until 2023, we find that CECL adoption is

associated with a reduction (improvement) in the information content of earnings announcements

and analyst LLP forecast revisions (10-K/Q filings). When we aggregate CECL’s impacts across

the three events, we find an overall deterioration in banks’ external information environment. We

show these results are robust to alternative choices regarding the inclusion of control variables and

the fixed effects structure employed.

Next, we conduct several additional analyses aimed at better understanding the drivers

behind the baseline results. Consistent with critics’ concerns, we show the decrease in earnings

announcement informativeness is stronger in instances where CECL is expected to increase LLP

volatility the most. Our results also indicate that higher-quality analysts, as measured by pre-period

LLP forecast accuracy, mitigate the decrease in earnings announcement informativeness,

consistent with analysts’ information interpretation role (e.g., Livnat and Zhang 2012). We next

consider two alternative explanations for the decrease in earnings announcement informativeness:

a decrease in price discovery speed or an increase in preemption from other information sources.

Our tests provide no evidence that either of these alternative explanations account for our main

findings. Further analyses of analyst forecasts reveal a deterioration in the properties of both

quarterly LLP forecasts and longer-horizon target price projections, consistent with the decrease

in information content of analyst forecast revisions following adoption. 6 Finally, using cosine

similarity (e.g., Brown and Tucker 2011), we find the increase in the informativeness of 10-K/Q

filings is stronger when banks’ quarterly vintage disclosures contain more news (i.e., are more

6
Specifically, we find an increase in analyst forecast errors and dispersion, and a decrease in the number of analyst
forecasts, for both loan loss provisions and target prices.

4
dissimilar to the prior quarter’s disclosures).7 Taken together, our findings suggest a decline in

banks’ external information environment after CECL adoption, consistent with critics’ concerns.

It is worth discussing the differences between our findings and two related concurrent

papers. First, Kim et al. (2023) find that banks’ LLPs are timelier and better reflect changes in

local economic conditions following CECL implementation. While this may seem inconsistent

with our result that earnings announcements become less informative after adoption, there are two

important considerations we capture that are outside the scope of Kim et al. (2023). First, it is

unclear whether this type of information was already known to market participants (e.g., Wheeler

2021; Beatty and Liao 2021; Lu and Nikolaev 2022). To the extent this information was already

known to the market, we would not expect the information content of earnings announcements to

increase. Second, Kim et al. (2023) do not capture the extent to which CECL introduces additional

volatility into provisions, which our findings suggest reduces the informativeness of earnings.

To examine this issue more directly, we empirically test the trade-off between more timely

loan loss provisioning under CECL and provision volatility. Specifically, we first replicate the

findings of Kim et al. (2023) in our sample by demonstrating that provisions become timelier after

CECL adoption. Then, not only do we find that provision volatility increases for the average CECL

bank post-adoption, but also we show that this increase is stronger for banks with higher provision

timeliness. These results highlight the trade-off between provision timeliness and provision

volatility under CECL. In other words, although CECL appears to lead to timelier provisions (Kim

et al. 2023), this increase in timeliness also increases provision volatility, which our results suggest

leads to less informative earnings announcements. Our findings and those of Kim et al. (2023) are

consistent with the notion described in Hilary and Hsu (2013) and Hilary et al. (2014) that a less

7
CECL requires the disclosure of credit quality indicators (e.g., internal risk grades, ranges of FICO scores), in relation
to the amortized cost of financing receivables, disaggregated by year of origination, or “vintage” (FASB 2016).

5
biased and more accurate, but more volatile, forecast will be less informative to investors.

Second, Gee et al. (2024) find that expected credit loss information on day one of CECL

adoption in the U.S. is decision-useful for equity investors and that CECL credit loss allowances

provide new information about credit losses to investors. However, the authors focus on day-1

CECL implementation, and thus do not examine earnings announcements, analyst LLP forecasts,

nor 10-K/Q filings after CECL is adopted, as we do in this study.

One natural question that arises from our main findings is whether these results are

expected to persist into the future. For example, it is possible that the decline in banks’ information

environment will reverse as investors and analysts gain more experience under the CECL regime.

We examine the effect of CECL adoption by year and find that the overall decline in the

information environment persists throughout our entire post-period of 2020 through 2022 and that

there is no evidence the magnitude of the effect declines over time. 8 Thus, we find no evidence

that the overall deterioration in banks’ information environment is a temporary effect of CECL

adoption, at least during our sample period.

One concern is that our main results may be driven by differential effects of

macroeconomic uncertainty on treatment and control banks in the post-period, which includes the

onset of the COVID-19 pandemic. Specifically, the requirement of CECL adoption in 2020 is

based on a size threshold (i.e., SRC versus non-SRC status), leading treatment banks to be larger

than control banks. To the extent the larger treatment banks’ information environment is

differentially affected by changes in macroeconomic conditions relative to that of control banks,

our primary findings could be driven by changes in the macroeconomic environment in the post-

8
The point estimate of the decline in the information environment has the largest magnitude in 2022, although it is
not statistically different than any other year of the post-period. However, our results indicate that the information
content of earnings announcements and 10-K/Q filings in 2022 is not statistically different than the pre-period,
although we continue to document a decrease in the information content of analyst LLP forecasts (see Section 5.4).

6
period, rather than CECL. Our by-year analyses are inconsistent with this alternative explanation,

given the overall decline in the information environment is not exclusive to the year of the

pandemic’s onset (i.e., 2020) but rather persists throughout the entire post-period. Nonetheless, we

take two additional steps to further allay this concern. First, we provide evidence that our main

findings are robust to including additional controls for bank size. Second, we trim our sample by

removing the largest 50% of treatment banks and the smallest 50% of control banks. After applying

this trimming procedure, there is no statistical difference in total assets between treatment and

control banks, on average ($7.5 billion and $5.3 billion, respectively). We re-estimate our main

analyses using this subsample and find consistent results, mitigating concerns that our baseline

findings are driven by size differences between the treatment and control banks.

Finally, our difference-in-differences design relies on the assumption that differences in

the information environment between the treatment and control banks would have remained

unchanged had the treatment banks not adopted CECL (i.e., the parallel trends assumption). While

this assumption is untestable, we examine trends in the information content of our three

information events in the pre-period to gauge its plausibility. Overall, we find little evidence

inconsistent with the reasonableness of the parallel trends assumption.

Our study makes several contributions to the literature. First, we contribute to the literature

that examines the consequences of expected credit loss standards. Several recent studies

investigate the adoption of an expected credit loss standard (Gee et al. 2024; Kim et al. 2023;

López-Espinosa et al. 2021; Oberson 2021).9 While these studies examine specific components of

9
Gee et al. (2024) find that expected credit loss information on day one of CECL adoption is decision-useful for equity
investors and that CECL credit loss allowances provide new information about credit losses to investors; Kim et al.
(2023) document that CECL improves information production by banks; López-Espinosa et al. (2021) show expected-
credit-loss-based provisions under IFRS 9 are more predictive of future bank risk; and Oberson (2021) finds IFRS 9
adoption leads to timelier loan loss recognition and some increases in the informativeness of provisioning for CDS
pricing.

7
banks’ information environment, CECL’s overall implications for banks’ external information

environment are outside the scope of these studies. For example, none of these studies examine

analyst forecasts. To the best of our knowledge, we are the first to examine CECL’s impacts on

the external information environment of banks more broadly. Specifically, our results show a

deterioration in the external information environment of banks, driven by decreases in the

information content of earnings announcements and analyst forecasts. Our paper complements

concurrent studies examining CECL by showing that although specific aspects of the information

environment may improve, a broader examination of banks’ information environment is important

in understanding the overall impacts of CECL. Our findings should also be of interest to regulators

and standard setters. In particular, the FASB stated that CECL’s main objective was “…to provide

financial statement users with more decision-useful information about the expected credit losses

on financial instruments and other commitments to extend credit…” (FASB 2016). Our findings

of an overall decline in the external information environment of banks should be of use to the

FASB in assessing whether this primary goal of CECL was achieved.

Second, our study contributes to the literature on the impact of accounting standard changes

on financial analysts. Prior research examines how the properties of analyst forecasts are affected

by IFRS adoption (e.g., Byard, Li, and Yu 2011; Tan, Wang, and Welker 2011; Horton, Serafeim,

and Serafeim 2013) or by revenue disaggregation under ASC 606 (Hinson, Puendrich, and Zakota

2024; Ferreira, Jeong, and Landsman 2023). Since CECL represents the most important change to

bank accounting in decades, its impact on financial analysts merits examination. Our findings

suggest that CECL adoption is associated with a decrease in the information content of analyst

provision forecasts. Beatty and Liao (2021) provide evidence suggesting that, under the IL

standard, analysts sacrifice forecast accuracy to inform investors about expected credit losses. Our

8
study suggests that CECL adoption, which could reduce the need for such a tradeoff, does not

improve the informativeness of analyst LLP forecasts.

Finally, our study contributes to the literature examining the information content of

earnings announcements over time. Several prior studies examine whether the information content

of earnings announcements has increased over time and explore various explanations for this trend

(e.g., Landsman and Maydew 2002; Francis, Schipper, and Vincent 2002; Collins, Li, and Xie

2009; Landsman, Maydew, and Thornock 2012; Beaver et al. 2020). Consistent with critics’

concerns, we provide evidence of a decrease in earnings announcement informativeness following

CECL adoption, with this decrease being stronger in instances where CECL is expected to increase

LLP volatility the most.

2. Institutional Background, Prior Literature, and Empirical Predictions

2.1. Institutional Background

The IL model delayed loan loss recognition by banks until it was “probable” that the loss

had been “incurred.” In the aftermath of the 2007–2009 financial crisis, loan loss recognition under

the IL model was perceived to be “too little, too late,” a problem widely believed to have amplified

the depth and duration of the financial crisis (Basel Committee on Banking Supervision 2021). In

response to calls from regulators and investors, the FASB and the IASB explored forward-looking

alternatives for timelier recognition of credit losses. In June 2016, the FASB issued ASU 2016-13,

Financial Instruments—Credit Losses (Topic 326), which requires recognition of lifetime

expected credit losses on the day of loan origination or acquisition. 10 The new model is commonly

10
In June 2014, the IASB issued International Financial Reporting Standard 9 – Financial Instruments (IFRS 9), which
uses the expected credit loss framework and became effective for annual periods beginning on or after January 1,
2018. There are significant differences between CECL and IFRS 9. Most importantly, CECL requires banks to
recognize lifetime expected credit losses on all loans upon loan origination/acquisition; in contrast, for the vast
majority of loans (stage 1 loans), IFRS 9 requires banks to recognize credit losses expected within the next 12 months.

9
referred to as the current expected credit loss or “CECL” model. CECL brings about a sea change

to how banks should recognize credit losses from loans and other debt instruments and has been

called by the American Bankers Association the “most sweeping change to bank accounting ever.”

CECL was originally set to become effective for fiscal years beginning after December 15,

2022 for SRCs, and for fiscal years beginning after December 15, 2019 for non-SRCs. 11 However,

in response to the outbreak of the COVID-19 pandemic, the Coronavirus Aid, Relief, and

Economic Security (“CARES”) Act, signed into law on March 27, 2020, gave non-SRCs the option

to delay CECL adoption until 2021. For the non-SRCs that elected to delay adoption until 2021,

the Consolidated Appropriations Act (2021), signed into law on December 27, 2020, gave them

the further option to delay CECL adoption until 2022. 12,13

Since its proposal, CECL has generated significant debate among financial institutions,

regulators, academics, and members of Congress about its potential costs and benefits. Proponents

of CECL, including bank regulators, applaud its timelier recognition of credit losses and greater

transparency due to the incorporation of forward-looking information in managers’ credit loss

estimation. Opponents of CECL, particularly financial institutions and their trade groups, along

with some members of Congress, vehemently sought the total removal or at least implementation

delay of CECL. The main concerns expressed by the opponents of CECL include outsized

implementation costs outweighing any potential benefits, increased subjectivity and discretion in

credit loss estimation, increased earnings volatility, and increased procyclicality in loan loss

provisioning and, hence, in bank lending. Although hundreds of publicly listed banks have already

11
SRCs are companies that have (i) public float of less than $250 million or (ii) annual revenues less than $100 million
and either no public float or public float less than $700 million (https://www.sec.gov/smallbusiness/goingpublic/SRC).
12
See Sec. 540 a (2), Subtitle C of CAA at https://www.congress.gov/bill/116th-congress/house-bill/133/text.
13
As discussed in Section 3, our sample includes only CECL banks that adopted on 1/1/2020. Our main results are
unaffected if we include banks that elected to delay CECL adoption under the CARES Act (untabulated).

10
adopted CECL, the controversy about the potential costs and benefits of CECL remains.

2.2. Prior Literature

Several recent studies develop their own empirical models to estimate expected credit

losses using data under the IL model and conclude that loan loss allowances under expected credit

loss models are more informative than those under IL models. Covas and Nelson (2018) estimate

expected losses using a vector autoregression methodology to forecast lifetime charge-off rates.

Harris et al. (2018) develop a measure of one-year-ahead expected rate of credit losses that contains

incremental information about one-year-ahead realized credit losses relative to the IL model. Lu

and Nikolaev (2022) develop an empirical model that predicts long-term loan losses and

incorporates adjustments for macroeconomic forecasts. Wheeler (2021) develops a measure of

lifetime expected credit losses using vintage analysis and finds that stock prices partially reflect

his estimated unrecognized expected loss information. 14

While these previously discussed studies estimate expected credit losses under the IL

model (i.e., prior to the adoption of expected credit loss standards), a number of more recent papers

examine the impacts of the adoption of CECL and IFRS 9, with the latter being the standard issued

by the IASB that also uses the expected credit losses approach. Gee et al. (2024) find that expected

credit loss information on day one of CECL adoption in the U.S. is decision-useful for equity

investors and that CECL credit loss allowances provide new information about credit losses to

investors. Bonaldi, Liang, and Yang (2023) find that, during the first two quarters of 2020, analyst

earnings forecast dispersion and the level of discretionary earnings are higher for banks adopting

CECL. López-Espinosa et al. (2021) find that credit loss provisions under IFRS 9 are more

14
Relatedly, Beatty and Liao (2021) examine analyst forecasts of loan loss provisions as a measure of informed market
participants’ expected credit loss estimates and find that analyst provision forecasts are incrementally predictive of
expected losses (using measures from Harris et al. 2018 and Wheeler 2021) and actual future losses beyond IL
provisions, especially for banks facing greater IL model constraints.

11
predictive of future bank risk than provisions under the IL standard, especially in countries with

deteriorating credit conditions. Lejard, Paget-Blanc, and Casta (2021) find that, after IFRS 9

adoption, sovereign credit ratings (impaired loans) become a more (less) important determinant of

loan loss allowances, and that there is increased heterogeneity in the measurement of provisions.

Onali, Ginesti, Cardillo, and Torluccio (2021) document positive market reactions to IFRS 9

adoption events.

The primary way in which our paper differs from these studies is in its scope. Specifically,

while these papers investigate certain aspects of banks’ information environment, such as whether

provisions or allowances provide incremental information about future credit losses under CECL,

they do not assess CECL’s overall impacts on banks’ external information environment. For

example, none of these papers examine whether the information content of analyst LLP forecasts

changes after the adoption of CECL. The broader scope of our paper allows us to assess CECL’s

broader impacts on banks’ external information environment, building on these other studies that

examine components of it.

2.3. Empirical Predictions

It is unclear ex ante how CECL adoption affects banks’ external information environment.

It is possible that CECL leads to an improvement in the information environment. Under the IL

model, banks were required to delay credit loss recognition until it was probable that a loss had

been incurred. In a Bayesian sense, this led to a downward bias in the provision and allowance for

loan losses. For example, many observers cite the IL model’s “too little, too late” problem as

amplifying the depth and duration of the financial crisis (Basel Committee on Banking Supervision

2021). Consistent with this notion, concurrent studies examining CECL adoption suggest that

CECL improves banks’ information production, including the timeliness of the provision (Kim et

12
al. 2023) and that day-1 CECL allowance information is decision-useful for investors (Gee et al.

2024). The reduction of the downward bias in recognizing credit losses suggested by these studies

could lead to an improvement in banks’ external information environment following CECL

adoption.

However, it is possible that additional uncertainty introduced by CECL could lead to a

deterioration in banks’ external information environment. From a Bayesian standpoint, the

usefulness of a signal depends on the volatility of the signal’s error term (i.e., the inverse of the

signal’s precision). A signal with a predictable bias but high precision is more informative than an

unbiased signal with less precision (Hilary and Hsu 2013; Hilary et al. 2014). It is possible that the

former is representative of the IL model, whereas the latter is descriptive of CECL. Although the

IL model was biased downward, prior literature suggests that unrecognized expected credit losses

were predictable, at least to some extent, consistent with a predictable bias in the IL model signal

(Harris et al. 2018; Lu and Nikolaev 2022; Wheeler 2021; Beatty and Liao 2021). Critics argue

that CECL introduces additional volatility into provisions (Pan et al. 2017; Maurer 2020; Bable et

al. 2022). Thus, the more precise signal with predictable bias (IL model) may be more informative

than the less precise signal with less bias (CECL model). Taken together, it is possible that the

increased uncertainty in credit loss recognition under CECL outweighs the potential benefits of

decreased bias, leading to a deterioration in banks’ external information environment.

To summarize, there are competing effects regarding the impact of CECL adoption on

banks’ external information environment. While concurrent studies suggest CECL may decrease

the downward bias inherent in the IL model, which should positively impact banks’ information

environment, critics argue that CECL introduces additional volatility into credit loss recognition,

which should have a negative impact. Ultimately, how CECL affects banks’ external information

13
environment is an open empirical question.

3. Research Design

We employ a difference-in-differences research design to examine the impact of CECL

adoption on banks’ external information environment. The treatment banks are those that adopted

CECL on January 1, 2020 and the control banks are SRCs, which were not required to adopt CECL

until 2023.15 2017–2022 represents the sample period, with 2017–2019 serving as the pre-period

and 2020–2022 serving as the post-period.

Our primary tests examine the change in the information content of three events around

CECL adoption: (i) earnings announcements, (ii) analyst LLP forecast revisions, and (iii) filings

of 10-K/Qs with the SEC.16 We examine these events because their information content is likely

to be directly affected by the adoption of CECL, whereas other information events, such as media

articles, are unlikely to be directly affected by CECL. It is worth noting two information events

that we do not examine in our analyses. First, we do not examine management guidance in our

setting because it is not very prevalent in the banking sector. Specifically, only 6% of the bank-

quarter observations in our sample have management earnings guidance in IBES. Furthermore,

any guidance that is bundled with the earnings announcement would be captured by our earnings

announcement informativeness tests. Of the management guidance we observe for our sample

banks, 29% (i.e., 84 total observations) are unbundled (untabulated). Second, we do not examine

15
Companies classified as Emerging Growth Companies (EGC) by the SEC are not required to adopt CECL until
2023 unless they lose their EGC status before 2023 (https://www.sec.gov/smallbusiness/goingpublic/EGC). We
classify non-SRC EGCs that have not adopted CECL during our sample period as control banks.
16
While we refer to analyst LLP forecast revisions, we note that our measure of information content does not
distinguish between the particular financial statement line item(s) being forecasted, since a revision to a provision
forecast would be concurrent with a revision to an earnings forecast. However, in subsequent tests, we focus on the
properties of LLP forecasts, for example, by calculating forecast accuracy and dispersion using forecasts of LLP
specifically. Furthermore, in untabulated falsification tests, we do not observe the same changes in accuracy or
dispersion for analyst pre-provision earnings nor non-interest income forecasts, which helps tie the change in
information content of analyst forecasts to LLPs in particular (see Section 5.2.2 for further discussion).

14
the filing of regulatory reports (i.e., Y-9Cs/call reports). While the calculation of the provision and

allowance for loan losses in regulatory reports would be affected by CECL, this information would

be reported during the earnings announcement, which typically precedes the filing of the

regulatory report.17,18

Our measure of information content is the Beaver (1968) U-statistic (UStat). UStat captures

the ratio of the squared residual returns during the testing period to the variance of residual returns

during the estimation period. The testing period is the event window [0, +1] relative to the event

date (i.e., earnings announcement, analyst LLP forecast revision, or 10-K/Q filing). The estimation

period includes the windows [-130, -10] and [+10, +130] relative to the event date. To calculate

UStat, we first estimate the market model with daily stock returns in the estimation period, obtain

estimates of the intercept and slope coefficient (ai and bi), and calculate the residual returns and

variance (Vari). Using ai and bi to calculate the residual return ( i,t) during the testing period, we

construct UStat as follows:

i,t
UStati,t = (1).
Vari

To investigate the impact of CECL implementation on banks’ external information

environment, we estimate the following regression:

UStati,t = 0 + 1CECLi x Postt + Controls + Bank FE + Year FE + i,t (2),

where CECL is an indicator variable equal to one for our treatment bank observations and zero for

17
While the precise filing date of regulatory reports is not readily available, Badertscher, Burks, and Easton (2018)
find that most call reports (Y-9Cs) are filed close to their due dates, which is 30 (40) days following quarter-end. We
find that 96% of earnings announcements occur within 30 days of quarter-end, and we note that our results are
unaffected if we drop the 4% of earnings announcements that occurred more than 30 days after quarter-end
(untabulated).
18
We reviewed all 150 earnings announcements (i.e., 8-K filings) made by CECL banks during the first quarter of
2020, which is the first quarter after CECL adoption. At a minimum, all banks provided balance sheet and income
statement information. Furthermore, in no instances did the bank provide footnotes to the financial statements
(including vintage disclosures).

15
our control bank observations, and Post is an indicator variable equal to one for quarters after

January 1, 2020, and zero otherwise. We estimate equation (2) separately for earnings

announcements, analyst LLP forecast revisions, and filings of 10-K/Qs in the seemingly unrelated

regressions framework (Zellner 1962).19,20 A negative (positive) coefficient on 1 is consistent

with a decrease (increase) in information content. To examine CECL’s overall impact on banks’

information environment, we test the significance of the sum of 1 across the three samples.

In equation (2), we control for the natural logarithm of the number of analysts issuing

provision forecasts during the quarter (LogNumLLPEst). We also include the following control

variables, all measured one quarter before the measurement of the dependent variable: market-to-

book ratio (MTB); whether the bank’s financial statements are audited by a Big Four auditor

(BigN); whether the bank reported negative earnings during the quarter (Loss); the most recent

quarter-over-quarter change in the provision recognized by the bank (∆LLP); and bank size,

measured as the natural logarithm of total assets (LogAssets). As banks comprise the entirety of

our sample, we also control for the following bank-specific factors: capital adequacy (Tier1Ratio,

which is the Tier 1 regulatory capital ratio) and loan portfolio composition (ConsumerLoans,

which is the proportion of the bank’s total assets that is consumer loans; and RealEstateLoans,

which is the proportion of the bank’s total assets that is real estate loans). We include bank fixed

effects and year fixed effects in the model, which absorb the main effects of CECL and Post.

Finally, we cluster standard errors by bank.

19
We drop 10-K/Q filings (analyst forecast revisions) that occur during the event window of an earnings
announcement (an earnings announcement or a 10-K/Q filing).
20
For earnings announcements and 10-K/Q filings, there is one observation per bank-quarter. For analyst LLP forecast
revisions, we retain analysts’ most recent revisions for the quarter and we use the revision-bank-quarter unit of
observation.

16
4. Sample Selection, Data, and Summary Statistics

To construct our main sample, we start with all publicly traded U.S. banks (identified using

SIC codes 6000 – 6299) with calendar year-ends for which we can identify analyst loan loss

provision forecasts and other bank-level variables available from the “Companies” dataset in S&P

Capital IQ Pro (“SPCIQ,” formerly SNL Financial). We further require that the banks have CRSP

and Compustat coverage for calculating certain control variables (see Appendix A). We next

classify banks into SRC and non-SRC groups. To identify banks’ SRC status, we use the SEC’s

criteria for SRC status (see footnote 11 in Section 2.1) and approximate public float using market

value of equity from CRSP. To reduce the potential for misclassification, we hand collect SRC

status from the 10-K filing for 2019Q4 for the subset of banks for which our approximation has

the highest potential for misclassification.21 To identify CECL adoption dates, we first attempt to

extract the data from the “Regulated Depositories (U.S.)” dataset in SPCIQ. When the reported

CECL day-one impact in SPCIQ is either zero or missing for a non-SRC bank, we hand collect the

bank’s CECL day-one impact data and adoption date from its regulatory filings (i.e., FR Y-9C and

call reports) and SEC filings (e.g., 10-Q filings). 22

Our final samples consists of 4,764 earnings announcements, 16,157 analyst LLP forecast

revisions, and 4,520 10-K/Q filings from 156 treatment banks and 136 control banks. We winsorize

all continuous variables at the 1st and 99th percentiles to reduce the influence of outliers. Panel A

(B) of Table 1 presents summary descriptive statistics separately for the pre-period and the post-

21
We set our hand collection thresholds conservatively to minimize errors in approximating public float. For banks
with less than $100 million in annual revenue, we hand collect SRC status if the bank has less than $1 billion market
value of equity (from CRSP), relative to the SEC’s threshold of less than $700 million of public float. For banks with
at least $100 million in annual revenue, we hand collect SRC status if the bank has less than $400 million market
value of equity, relative to the SEC’s threshold of less than $250 million public float.
22
The CECL day-one impact captures the cumulative-effect adjustment for the changes in the allowance for credit
losses, net of any related deferred tax assets and excluding the initial allowance gross-up for any purchased credit-
deteriorated assets, recognized in retained earnings as of the adoption date (SPCIQ Keyfield 319094).

17
period for the CECL (control) banks. We separately report the descriptive statistics for UStat for

the three event samples; for the remaining variables, we report the descriptive statistics based on

the earnings announcement sample for brevity. For the average treatment bank in the pre-period,

UStat for earnings announcements is 9.0, which is comparable to that documented in Beaver et al.

(2020) of 7.6.23 For the same banks, the average UStat for analyst LLP forecast revisions (10-K/Q

filings) is 4.5 (1.5), consistent with earnings announcements being the most informative of the

three events, followed by analyst LLP forecast revisions and 10-K/Q filings, respectively. Our

univariate evidence in Panel A of Table 1 indicates a significant decline in informativeness of

earnings announcements and analyst LLP forecast revisions, and a significant increase in the

information content of 10-K/Q filings, for treatment banks following CECL adoption.

5. Empirical Results

5.1. Baseline Analyses

Table 2, Panel A presents the results from estimating equation (2). The coefficients on

CECL x Post are significantly negative in Columns 1 and 2 (earnings announcements and analyst

LLP forecast revisions, respectively) and significantly positive in Column 3 (10-K/Q filings),

consistent with the univariate evidence in Panel A of Table 1. These results suggest a decline

(increase) in the information content of earnings announcements and analyst LLP forecasts (10-

K/Q filings) following CECL adoption. In terms of economic magnitude, the coefficients in

Columns 1 and 2 imply a decrease in UStat of 36 and 59 percent, respectively, of the pre-period

means for CECL banks (see Table 1, Panel A). The coefficient in Column 3 indicates that the

information content of 10-K/Q filings nearly doubles relative to CECL banks’ pre-period mean.

These effect sizes represent 0.38, 0.43, and 0.31 standard deviations of the within-fixed-effect

23
Beaver et al. (2020) note that UStat is larger during the later years of their sample, reaching a peak of 11.7 in 2016.

18
variation of UStat in the three samples, respectively.24 Thus, these effects are both statistically and

economically significant. At the bottom of Panel A, we show that the sum of the coefficients on

CECL x Post across the three sample is negative and highly significant, consistent with an overall

deterioration in the information environment of banks following CECL adoption.

Panels B – D of Table 2 report the robustness of the results to changes regarding the

inclusion of control variables and fixed effects. Panel B (C) [D] reports the results including no

control variables and no fixed effects (including no control variables and only year fixed effects)

[including control variables and only year fixed effects]. We document similar results across these

three panels, suggesting our main results are robust to various design choices regarding control

variables and the fixed effects structure.

5.2. Additional Analyses

In the next three subsections, we discuss a series of additional analyses aimed at better

understanding the drivers behind the overall deterioration in banks’ information environment

following the implementation of CECL.

5.2.1. Earnings Announcements – Additional Analyses

We first conduct cross-sectional tests for the earnings announcement sample. One of the

main criticisms of CECL is that it could introduce additional uncertainty, and thus volatility, into

loan loss provisions, which could reduce earnings informativeness (Pan et al. 2017; Maurer 2020;

Bable et al. 2022). To test this idea, we estimate the following moderation model:

UStati,t = 0 + 1Moderatori x Postt + 2CECLi x Postt

+ 3Moderatori x CECLi x Postt+ Controls + Bank FE + Year FE + i,t (3),

24
Following Mummolo and Peterson (2018), we report economic significance in terms of the dependent variable’s
within-fixed-effect standard deviations. Specifically, we regress the dependent variable on the fixed effects and then
calculate the standard deviation of the residuals. The effect sizes are calculated by dividing the coefficients on CECL
x Post in Table 2 by the within-fixed-effect standard deviations of the dependent variable in the respective samples.

19
where Moderator is either PreLLPVol, the standard deviation of the bank’s quarterly loan loss

provisions during the pre-period, or PreRetVol, the average of the bank’s quarterly return volatility

during the pre-period. An important characteristic of a moderator variable is that it is unrelated to

the treatment variable (e.g., Baron and Kenny 1986). Since the implementation of CECL could

plausibly affect banks’ LLP and return volatility, we use pre-period variables, which cannot be

affected by adoption. This choice is important because if we were to use post-period volatilities,

our moderator variables could capture variation in the strength of treatment (i.e., how much the

bank is affected by CECL), rather than cross-sectional variation related to provision/return

volatility.25 Our expectation is that banks likely to experience greater increases in provision/return

volatility after CECL adoption (proxied by PreLLPVol and PreRetVol) will have a stronger

decrease in earnings announcement informativeness (i.e., 3 in equation (3) will be negative).

We also explore the role of analysts in this setting. Prior literature suggests that analysts

play an information interpretation role in the context of earnings announcements (e.g., Livnat and

Zhang 2012). Given our main results suggest that analyst forecasts become less informative

following CECL adoption, one contributing factor to the decline in the information content of

earnings announcements could be the decrease in forecast informativeness. To investigate this

idea, we estimate equation (3) with a third Moderator, PreLLPFcstAcc, the average analyst LLP

forecast accuracy during the pre-period. Our expectation is that banks followed by higher-quality

analysts (proxied by PreLLPFcstAcc) will have the decrease in earnings announcement

informativeness mitigated (i.e., 3 in equation (3) will be positive).

25
Our assumption with these moderator variables is that banks with higher provision/return volatilities in the pre-
period are likely to experience greater volatilities after adoption. We note the correlations between pre-period and
post-period provision and return volatility are 0.423 and 0.453 for CECL banks, respectively, providing support for
this assumption (untabulated). Furthermore, the average provision and return volatilities are significantly higher for
CECL banks compared to control banks in the post-period (untabulated).

20
 These results are presented in Table 3, Panel A. Consistent with our expectations, the

coefficient on Moderator x CECL x Post is significantly negative in Columns 1 and 2 (PreLLPVol

and PreRetVol), consistent with the decrease in earnings announcement informativeness being

stronger in instances where CECL is expected to increase LLP/return volatility the most. In

Column 3, where the moderator variable is PreLLPFcstAcc, we find the coefficient on the triple

interaction term is significantly positive, suggesting that the decrease in earnings announcement

informativeness is mitigated for banks followed by higher-quality analysts. Taken together, these

results are consistent with critics’ concerns that CECL’s introduction of additional

uncertainty/volatility into loan loss provisions leads to less informative earnings. Furthermore, our

results support the notion that higher-quality financial analysts can mitigate the deterioration in

information content of earnings announcements.

Next, we consider two alternative explanations for the change in earnings announcement

informativeness following CECL adoption. First, it is possible that our results reflect a decrease in

the speed of price discovery around earnings announcements, rather than a decrease in the

information content of earnings announcements. The event period for the calculation of UStat is a

two-day window [0, +1]. If the speed of price formation decreases for CECL banks after adoption,

it is possible that information from the earnings announcement may not be incorporated into price

until after day +1. We examine this possibility in three ways. For our first approach, we use

extended event periods to calculate UStat, [0, +5] and [0, +10]. An insignificant coefficient on

CECL x Post when we calculate UStat using these alternative event periods would be consistent

with the alternative explanation.

Next, we re-estimate equation (2) but replace the dependent variable with IPT, intraperiod

timeliness, calculated as follows:

21
 CARt
IPTt = ∑4t=0 ( ) (4),
CAR5

where CAR is the cumulative abnormal buy-and-hold return from day 0 through day t. If the main

results are driven by a decrease in price formation speed, we would expect a significantly negative

coefficient on CECL x Post when estimating the IPT regression.

Finally, we examine post-earnings announcement drift by estimating the following model:

DriftCARi,t = 0 + 1SUEi,t + 2CECLi x SUEi,t + 3CECLi x Postt

+ 4SUEi,t x Postt + 5CECLi x SUEi,t x Postt + Controls
+ Bank FE + Year FE + i,t (5),

where DriftCAR is the cumulative abnormal buy-and-hold return from the day after the current

earnings announcement date through the date of the subsequent earnings announcement date, and

SUE is standardized unexpected earnings, defined as actual earnings minus the mean consensus

forecast, scaled by the standard deviation of the consensus forecast; SUE is decile-ranked and

scaled to vary between -0.5 and 0.5. A significantly positive 5 would be consistent with more of

an underreaction to earnings announcements for CECL banks after adoption, which would be

consistent with the alternative explanation of slower price discovery.

The results of these tests are presented in Panel B of Table 3. The dependent variables in

Columns 1 and 2 are UStat calculated with the event period of [0, +5] and [0, +10], respectively.

Across both columns, we continue to document a significantly negative coefficient on CECL x

Post, with coefficient magnitudes that are larger than that shown in Column 1 of Table 2, Panel A.

In Column 3 (dependent variable IPT), the coefficient on CECL x Post is insignificant (t-stat 0.03),

indicating that we fail to find evidence of a change in intraperiod timeliness for CECL banks after

adoption relative to the control banks.26 Finally, in Column 4 (dependent variable DriftCAR), we

find the coefficient on CECL x SUE x Post is insignificant (t-stat -0.34); thus, we fail to find

26
We find similar results using adjusted intraperiod timeliness following Blankespoor, deHaan, and Zhu (2018).

22
evidence of a change in post-earnings announcement drift for CECL banks following adoption

relative to the control banks. Taken together, the results in Panel B of Table 3 are inconsistent with

the alternative explanation that a decrease in the speed of price discovery around earnings

announcements drives our main results.

The second alternative explanation we consider is that investors obtain information from

sources other than earnings announcements to a greater extent once CECL is adopted. Under this

alternative explanation, the decrease in UStat around earnings announcements is driven by

earnings announcement news being preempted by alternative information sources, rather than a

decrease in the informativeness of earnings. To examine this issue, we examine the extent to which

stock returns lead earnings (Beaver, Lambert, and Morse 1980; Kothari and Sloan 1992).

Specifically, we estimate the following annual regression:

CARi,[t-365,t-1] = 0 + 1Earningsi,t + 2CECLi x Earningsi,t + 3CECLi x Postt

+ 4Earningsi,t x Postt + 5CECLi x Earningsi,t x Postt + Controls
+ Bank FE + Year FE + i,t (6),

where CAR[t-365,t-1], is the cumulative abnormal buy-and-hold return from one year before the

earnings announcement date through the day before the earnings announcement date, and Earnings

is the bank's annual earnings scaled by lagged total assets. If earnings news is being preempted by

alternative information sources after adoption, we would expect 5 to be significantly positive.

The results of estimating equation (6) are presented in Panel C of Table 3. We present the

results using the full sample (Column 1) and with the sample excluding the year 2020 (Column 2).

The reason we include the Column 2 test is because for the year 2020, CAR[t-365,t-1] (Earnings) is

measured before (after) CECL is adopted. Thus, we present the results excluding 2020 in Column

2 to examine whether the difference in the accounting regimes for calculating stock returns versus

earnings for 2020 affects the results. Across both Columns 1 and 2, we find that the coefficient on

23
CECL x Earnings x Post is insignificant (t-stat -1.61 and -0.65, respectively), inconsistent with the

alternative explanation that earnings news is preempted by other information sources to a greater

extent for CECL banks under the CECL regime.

5.2.2. Analyst Forecasts – Additional Analyses

Next, we conduct additional analyses related to analyst LLP forecast revisions. First, we

explore whether the properties of analyst LLP forecasts deteriorate after the introduction of CECL,

consistent with the decrease in information content documented in Table 2. To explore this

possibility, we test whether CECL adoption affects the accuracy, dispersion, and coverage of

analyst provision forecasts. We re-estimate equation (2) but replace the dependent variable with

either forecast accuracy (|LLPForecastError|), dispersion (LLPForecastDispersion), or coverage

(LogNumLLPEst). Following prior literature (e.g., Dhaliwal, Radhakrishnan, Tsang, and Yang

2012), we use forecast error as an inverse measure of accuracy. We define |LLPForecastError| as

the absolute difference between the consensus (mean) analyst provision forecast and the actual

provision for the bank-quarter, multiplied by 100 and scaled by lagged total assets. 27 We define

LLPForecastDispersion as the standard deviation of the individual analyst provision forecasts

included in the consensus, scaled by lagged total assets. Finally, we define LogNumLLPEst as the

natural logarithm of the number of analyst provision forecasts included in the consensus.

We also consider whether CECL adoption affects analysts’ shorter horizon forecasts

differently than their longer horizon forecasts. For example, short-term forecasts of LLP may

become more difficult due to increased volatility in the provision following CECL adoption,

27
The prior literature (e.g., Dhaliwal et al. 2012) calculates EPS forecast errors as the difference between forecasted
and actual EPS, scaled by stock price: |ForecastEPS – ActualEPS|/Price. Since the two EPS numbers (stock price) are
calculated as earnings (market value of equity) divided by shares outstanding, this can be rewritten as follows:
|$ForecastEarnings – $ActualEarnings|/MVE. Thus, the forecast error is the difference between the forecasted and
actual earnings (in dollars) divided by firm size (i.e., the market value of equity). Our measure of provision forecast
error is analogous, except that we scale by lagged total assets (Beatty and Liao 2021).

24
whereas the forward-looking nature of CECL may enable analysts to provide better long-term

forecasts. To examine this possibility, we examine the effects of CECL on analyst target price

projections, which have a longer horizon than LLP forecasts (i.e., target prices have a one-year

horizon whereas LLP forecasts predict quarterly provisions). More specifically, we re-estimate

equation (2) but replace the dependent variable with either target price forecast accuracy

(|TPForecastError|), dispersion (TPForecastDispersion), or coverage (LogNumTPEst).

The results of these tests are presented in Table 4, with the results of the tests examining

the properties of LLP and target price forecasts presented in Panels A and B, respectively. Across

both panels, we find that the coefficients on CECL x Post are significantly positive for forecast

errors and dispersion, and significantly negative for coverage. Taken together, these results are

consistent with a deterioration in the properties of analyst forecasts, both for shorter-horizon LLP

forecasts and longer-horizon target price projections.

As discussed in Section 3, our main measure of the information content of analyst forecasts

does not distinguish between the particular financial statement line item(s) being forecasted. The

results in Table 4, Panel A support the notion that the decrease in informativeness of analyst

forecast revisions is driven by a deterioration in the properties of analysts’ LLP forecasts

specifically. To provide further support for this, in untabulated analyses, we examine the properties

of analyst forecasts of (i) earnings before the loan loss provision and (ii) non-interest income, both

of which should not be affected by CECL adoption. 28 Our results indicate an insignificant decrease

in forecast errors and dispersion following CECL adoption for forecasts of both earnings before

28
Non-interest income includes items such as fee and trading income. A benefit of examining these forecasts as a
falsification test is that, much like the provision, a bank’s non-interest income is typically highly sensitive to changes
in macroeconomic conditions. Thus, the falsification test can help rule out the alternative explanation that the changes
in analysts’ LLP forecast properties are driven by macroeconomic uncertainty caused by the COVID-19 pandemic,
rather than CECL.

25
the loan loss provision and non-interest income. In contrast, using the same samples, we continue

to document an increase in both errors and dispersion for loan loss provision forecasts.

5.2.3. 10-K/Q Filings – Additional Analyses

Finally, we consider the role of vintage disclosures, which are an important requirement of

CECL, in explaining the increase in 10-K/Q filings’ informativeness. CECL requires the disclosure

of credit quality indicators (e.g., internal risk grades, ranges of FICO scores), in relation to the

amortized cost of financing receivables, disaggregated by year of origination, or “vintage” (FASB

2016). The FASB believes that vintage disclosures “…will be useful for financial statement users

to better assess changes in underwriting standards and credit quality trends in asset portfolios over

time and the effect of those changes on credit losses” (FASB 2016).

To test this idea, we first gather data from CECL banks’ vintage disclosures using XBRL

data provided on the SEC’s website.29 To identify the tagged information related to vintage

disclosures, we first perform a keyword search for the standard tags used for the disclosures during

2020 – 2022.30 Using banks’ “XBRL instance document” from EDGAR, we also manually

collected each bank’s initial vintage disclosure from 2020Q1. We then used this hand-collected

data to ensure the validity of our broader search procedures. 31

Using the data from these search procedures, we construct a measure of vintage disclosure

information content based on cosine similarity (e.g., Brown and Tucker 2011). Specifically, for

29
https://www.sec.gov/about/dera_financialstatementandnotesdatasets.
30
Our Python regular expression for capturing standard XBRL tags from banks’ vintage disclosures was
“FinancingReceivable(?:.*Originated|Revolving|ConvertedToTermLoans” and resulted in coverage of more than 90
percent of the bank-quarter observations in our sample. Observations with no matches based on our regular expression
were overwhelmingly the result of banks using custom XBRL tags, which was revealed during our hand-collection
review process.
31
More specifically, we compared the two sets of data to find (i) instances of tagged items in the hand-collected data
that were missing from the broader dataset, which allowed us to modify our keyword search for instances when banks
used nonstandard tags, and (ii) instances of erroneous tags captured by our keyword search, which we removed from
the dataset.

26
each bank-quarter, we construct a list of each tagged data item (e.g., commercial and industrial

loans with an internal risk rating of “pass” originated in the current fiscal year) that appears in the

vintage disclosure for the current quarter and/or prior quarter. Then, we construct two vectors, one

for the current quarter and one for the prior quarter, whose elements represent the values of each

of the tagged data items.32 Finally, we calculate the cosine similarity between these two vectors.

To examine whether vintage disclosures explain the increase in informativeness of 10-K/Q

filings, we estimate the following regression:

UStati,t = 0 + 1HighNewsi,t x CECLi x Postt + 2LowNewsi,t x CECLi x Postt

+ Controls + Bank FE + Year FE + i,t (7),

where HighNews (LowNews) is an indicator variable equal to one if the cosine similarity between

the current and prior quarters’ vintage disclosures is below (above) the sample median for the

quarter, zero otherwise. Since HighNews and LowNews can only be calculated for CECL banks in

the post-period, lower-order interaction terms cannot be included in the model. Observations in

2020Q1 are dropped from this analysis because there are no prior quarters with which to construct

the cosine similarity measure. If vintage disclosures help explain the increase in information

content of 10-K/Q filings, we expect 1 to be significantly more positive than 2.

The results of this test are presented in Table 5. We find that both 1 and 2 are significantly

positive, consistent with Table 2. However, as noted in the F-test at the bottom of Table 5, we

show the coefficient on HighNews x CECL x Post is significantly more positive with a magnitude

that is nearly double that on LowNews x CECL x Post, consistent with the increase in 10-K/Q

32
If a data item is missing because it appears only in the prior or current quarter (but not both), we assign a value of
zero to that element of the vector. A data item would appear in only one quarter if the bank changed its vintage
disclosure breakdown (e.g., the bank changed the loan types disclosed by reporting a more or less granular breakdown
than before). However, this could also be caused by the bank changing its XBRL tagging for the same data item. We
hand-check each instance where we identify 50 or more new tags used in a quarter and adjust our matching procedure
to ensure the same data elements are matched across the two quarters. For these observations, if we are unable to
match the data items because of a change in the bank’s XBRL tagging, we drop them from the vintage disclosure tests.

27
filings’ informativeness being stronger when vintage disclosures convey more news.

5.3. Loan Loss Provision Timeliness and Volatility under CECL

Although concurrent studies suggest that the properties of loan loss provisions, such as

timeliness, have improved following CECL adoption (e.g., Kim et al. 2023), our predictions in

Section 2.3 suggest that CECL likely increases uncertainty in credit loss recognition, which our

results suggest leads to a deterioration in banks’ information environment.

To examine the potential trade-off between LLP timeliness and volatility, we conduct two

additional analyses. First, we examine how loan loss provision timeliness changes after CECL

adoption by estimating the following regression (Kim et al. 2023):

LLPi,t = 0 + 1NPLi,t- x CECLi x Postt + 2NPLi,t x CECLi x Postt

+ 3NPLi,t+ x CECLi x Postt + Additional Interaction Terms + Controls
+ Bank FE + Year FE + i,t (8),

where LLP is the bank's quarterly loan loss provision scaled by total assets, NPLt is the change

in nonperforming loans from the prior quarter to the current quarter, scaled by lagged total loans,

and NPLt- (NPLt+) is the change in nonperforming loans scaled by lagged total loans, averaged

over the prior (lead) two quarters. The lower-order interaction terms, as well as the same control

variables and fixed effects from equation (2), are included in the model.

These results are presented in Panel A of Table 6. Consistent with the Kim et al.’s (2023)

findings, we find that 2 and 3 from equation (8) are significantly positive, consistent with an

increase in provision timeliness for CECL banks post-adoption.

Second, we investigate how loan loss provision volatility changes after CECL is adopted,

and whether this change is related to provision timeliness. We first proxy for a bank’s provision

timeliness by estimating the following regression by bank, separately for the pre-period and post-

period:

28
 LLPi,t = 0 + 1NPLi,t- + 2NPLi,t + 3NPLi,t+ + i,t (9).

LLPTimeliness is the average of the estimated 2 and 3 from this regression.

Then, we estimate the following regression:

LLPVoli,t = 0 + 1LLPTimelinessi,t + 2LLPTimelinessi,t x CECLi

+ 3LLPTimelinessi,t x Postt + 4CECLi x Postt
+ 5LLPTimelinessi,t x CECLi x Postt + Controls
+ Bank FE + Time FE + i,t (10),

where LLPVol is the standard deviation of the bank's quarterly loan loss provisions, measured

separately over the pre-period and post-period. For this test, we include at most two observations

per bank (depending on data availability): one for the pre-period and one for the post-period. The

control variables are the same as those in equation (2) but are averaged over the pre- and post-

periods to match the level of observation for this test. The time fixed effect is an indicator for the

pre-period/post-period, rather than year fixed effects, due to the level of observation for the test.

The results are presented in Panel B of Table 6. We note two primary findings. First, we

find that 4 is significantly positive, suggesting that provision volatility increased for CECL banks

with an average level of LLPTimeliness post-adoption.33 Second, we show that 5 is significantly

positive, indicating that the increase in provision volatility is exacerbated for CECL banks with

timelier provisions. These findings highlight the trade-off between provision timeliness and

volatility under CECL. In other words, although CECL appears to lead to timelier provisions, this

increase in timeliness also increases provision volatility, which our main results suggest leads to

less informative earnings announcements.

5.4. Persistence of CECL’s Effects on Banks’ External Information Environment

One question that arises from our main findings is whether these results are expected to

33
LLPTimeliness is standardized to have a mean of zero and a standard deviation of one. Thus, 4 in equation (10)
captures the change in LLPVol for CECL banks with an average level of LLPTimeliness.

29
persist into the future. To examine this, in Table 7, we conduct by-year analyses of the post-period

to assess whether there is a reversal of our main results during our sample period. Such a reversal

would be expected if (i) our baseline results are only concentrated in 2020 because the onset of the

COVID-19 pandemic, which occurred in January/February 2020, drives our results or (ii) there is

a “learning effect” as investors and analysts gain more experience under the CECL regime.

Post2020, Post2021, and Post2022 are indicator variables set to one if the year is 2020, 2021, and

2022, respectively, and zero otherwise.

There are three main takeaways from Table 7. First, inconsistent with the notion that our

results are driven by the onset of the COVID-19 pandemic, we find that our results are not only

found in 2020 (i.e., the coefficients on CECL x Post2020 and on CECL x Post2021 are significant

and the same signs across all three samples). Second, there is evidence of a reversal of our results

for earnings announcements and 10-K/Q filings. Specifically, the coefficient on CECL x Post2022

is insignificant in Columns 1 and 3, indicating that we fail to find a difference in the information

content of earnings announcements and 10-K/Q filings in 2022 compared to the pre-period.

However, there is no evidence of a reversal for analyst LLP forecast revisions. Finally, and most

importantly, when we assess the overall change in banks’ information environment, we find no

evidence of a reversal, as the sums of the coefficients on CECL x Post2020, CECL x Post2021,

and CECL x Post2022 are -4.656, -4.062, and -5.392, respectively, with each of these sums being

statistically significant and none of them being statistically different from one another

(untabulated). Taken together, we fail to find evidence that the overall deterioration in banks’

external information environment is a temporary effect of CECL adoption.

5.5. Robustness Tests

One concern is that our main results may be driven by differential effects of

30
macroeconomic uncertainty on treatment and control banks in the post-period, which includes the

onset of the COVID-19 pandemic. Specifically, the requirement of CECL adoption in 2020 is

based on a size threshold (i.e., SRC versus non-SRC status), leading treatment banks to be larger

than control banks. To the extent the larger treatment banks’ information environment is

differentially affected by changes in macroeconomic conditions relative to that of control banks,

our primary findings could be driven by changes in the macroeconomic environment in the post-

period, rather than CECL.

While our by-year analyses discussed above, as well as the inclusion of the natural

logarithm of total assets (LogAssets) and bank fixed effects in equation (2), should help mitigate

this concern, we take two additional steps to address it. First, we include additional controls for

bank size in equation (2). Specifically, in addition to controlling for Assets, we control for the

square and cubic forms of Assets (AssetsSquared and AssetsCubed, respectively) and their

interaction terms with Post. We present the results in Panel A of Table 8. We find our main results

are robust to the inclusion of these additional size controls. Second, we examine whether our main

results in Table 2 are robust to trimming the sample based on total assets until there is no statistical

difference in the mean of total assets between treatment and control banks. Specifically, we first

calculate the bank’s largest value of total assets during the sample period. Then, using this value,

we drop the largest (smallest) 50% of treatment (control) banks. 34 Using this subsample of banks,

we re-estimate equation (2) and also conduct a test of the difference in the mean of total assets

across treatment and control banks. We report these results in Panel B of Table 8. Our results using

this trimmed sample are consistent with the results in Table 2. Furthermore, at the bottom of Panel

34
For this robustness test, we dropped the largest (smallest) treatment (control) banks in increments of 10% until the
difference in the mean of total assets between treatment and control banks was insignificant. Using this approach,
dropping the largest (smallest) 50% of treatment (control) banks was the fewest number of observations we could drop
while achieving no statistical difference in the mean of total assets between treatment and control banks.

31
B, we find that the difference in the mean of total assets between treatment and control banks in

the trimmed sample is statistically insignificant. In sum, the results in Table 8 mitigate concerns

that our baseline findings are driven by size differences between the treatment and control banks.

Finally, our difference-in-differences design relies on the assumption that differences in

the information environment between the treatment and control banks would have remained

unchanged had the treatment banks not adopted CECL (i.e., the parallel trends assumption). While

this assumption is untestable, we examine trends in the information content of our three

information events in the pre-period to gauge its plausibility. We split the pre-period into quarters

(e.g., the indicator variable 2017Q2 equals one for 2017Q2, and zero otherwise), with 2017Q1

serving as the benchmark.35 Our test variables are the interaction terms between the CECL

indicator and each of these quarter indicators (e.g., CECL x 2017Q2). We report the results from

this analysis in Table 9. Of the 30 coefficients of interest, only one is statistically significant and

the same sign as the main results: the coefficients on CECL x 2018Q1 in Column 2 (analyst LLP

forecasts sample). Given that this significantly negative coefficient occurs two years before CECL

implementation and none of the later interaction terms are significantly negative, we find little

evidence inconsistent with the reasonableness of the parallel trends assumption.

6. Conclusion

We examine how U.S. banks’ adoption of the CECL standard, which imposes the most

significant changes to bank accounting in decades, affects banks’ external information

environment. We find that CECL adoption is associated with a reduction in the information content

of earnings announcements and analyst provision forecasts, and an increase in the information

content of 10-K/Q filings. When we aggregate CECL’s impacts across the three events, we find

35
Given our focus on potential pre-period trends and to avoid including earnings announcements, LLP forecast
revisions, and 10-K/Q filings made after 1/1/2020, we include quarters through 2019Q3.

32
an overall deterioration in banks’ external information environment. We also show that this decline

in banks’ information environment persists throughout our post-period. Thus, we fail to find

evidence that the overall deterioration in banks’ information environment is a temporary effect of

CECL adoption.

Consistent with concerns of CECL’s critics, we show the decrease in earnings

announcement informativeness is stronger in instances where CECL is expected to increase LLP

volatility the most. Our results also indicate that higher-quality analysts, as measured by pre-period

LLP forecast accuracy, mitigate the decrease in earnings announcement informativeness. We fail

to find evidence that a decrease in price discovery speed or an increase in preemption from other

information sources explains our main findings regarding earnings announcements. Further

analyses of analyst forecasts reveal a deterioration in the properties of both quarterly LLP forecasts

and longer-horizon target price projections, consistent with the decrease in information content of

analyst forecasts following adoption. Finally, we find that the increase in information content of

10-K/Q filings is greater when vintage disclosures convey more news.

We also show the robustness of our results to various aspects of our research design. First,

we show the main results are robust to alternative choices regarding the inclusion of control

variables and the fixed effects structure employed. Second, we provide evidence that our main

findings are robust to including additional controls for bank size and using a trimmed sample with

no statistical difference in total assets between treatment and control banks. Finally, we find little

evidence inconsistent with the reasonableness of the parallel trends assumption.

Our findings contribute to the literature examining the consequences of expected credit

loss standards and should be of particular interest to regulators and standard setters. The FASB

stated that CECL’s main objective was “…to provide financial statement users with more decision-

33
useful information about the expected credit losses on financial instruments and other

commitments to extend credit…” (FASB 2016). Our findings of a decline in banks’ external

information environment should be of use to the FASB in assessing whether this primary goal of

CECL was achieved. Our study also contributes to the literature on the impact of accounting

standard changes on financial analysts by providing evidence that CECL adoption is associated

with a decrease in the information content of analyst provision forecasts. Finally, our study

contributes to the literature examining the information content of earnings announcements over

time by documenting a decrease in earnings announcement informativeness following CECL

adoption.

Despite our evidence that the deterioration in banks’ information environment after CECL

adoption occurs in 2020, 2021, and 2022, and our robustness tests using a trimmed sample where

treatment and control banks do not differ statistically in terms of size, we offer the caveat that we

cannot fully rule out the possibility that our main findings may be at least partly driven by the

pandemic and the resulting greater macroeconomic uncertainty that ensued. Thus, we caution

against the generalization of our findings to periods in a more “normal” economic environment.

34
References

Bable, J., C. Wong, and M. Wynes. 2022. The decision usefulness of CECL: Users’ views about the
current expected credit losses model. Working paper.
Badertscher, B. A., J. J. Burks, and P. D. Easton. 2018. The market reaction to bank regulatory reports.
Review of Accounting Studies 23: 686–731.
Baron, R. M., and D. A. Kenny. 1986. The moderator-mediator variable distinction in social
psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality
and Social Psychology 51 (6): 1173–1182.
Basel Committee on Banking Supervision. 2021. The procyclicality of loan loss provisions: A
literature review. Working paper.
Beatty, A., and S. Liao. 2021. What do analyst provision forecasts tell us about expected credit
loss recognition? The Accounting Review 96 (1): 1–21.
Beaver, W. H. 1968. The information content of annual earnings announcements. Journal of
Accounting Research 6: 67–92.
Beaver, W. H., R. Lambert, and D. Morse. 1980. The information content of security prices.
Journal of Accounting and Economics 2: 3–28.
Beaver, W. H., M. F. McNichols, and Z. Z. Wang. 2020. Increased market response to earnings
announcements in the 21st century: An empirical investigation. Journal of Accounting and
Economics 69 (1), 101244.
Beyer, A., D. A. Cohen, T. Z. Lys, and B. R. Walther. The financial reporting environment: Review
of the recent literature. Journal of Accounting and Economics 50, 296–343.
Blankespoor, E., E. deHaan, and C. Zhu. 2018. Capital market effects of media synthesis and
dissemination: Evidence from robo-journalism. Review of Accounting Studies 23: 1–36.
Bonaldi, P., P. Liang, and L. Yang. 2023. Did CECL improve banks’ loan loss provisions and
earnings quality during the COVID-19 pandemic? Working paper.
Brown, S., and J. Tucker. 2011. Large-sample evidence on firms’ year-over-year MD&A
modifications. Journal of Accounting Research 49 (2): 309–346.
Byard, D., Y. Li, and Y. Yu. 2011. The effect of mandatory IFRS adoption on financial analysts’
information environment. Journal of Accounting Research 49: 69–96.
Collins, D., O. Li, and H. Xie. 2009. What drives the increased informativeness of earnings
announcements over time? Review of Accounting Studies 14: 1–30.
Covas, F., and W. Nelson. 2018. Current expected credit loss: Lessons from 2007-2009. Working
paper.
Dhaliwal, D. S., S. Radhakrishnan, A. Tsang, and Y. G. Yang. 2012. Nonfinancial disclosure and
analyst forecast accuracy: International evidence on corporate social responsibility disclosure.
The Accounting Review 87 (3): 723–759.
Ferreira, P., J. Jeong, and W. R. Landsman. 2023. The effects of ASC 606 retrospective adoption
on a firm’s information environment. Working paper.
Financial Accounting Standards Board (FASB). 2016. FASB Accounting standards update No.
2016-13. June 2016. Financial Instruments—Credit Losses (Topic 326) Measurement of
Credit Losses on Financial Instruments.
Francis, J., K. Schipper, and L. Vincent. 2002. Earnings announcements and competing
information. Journal of Accounting and Economics 33: 313–342.
Gee, K. H., J. J. Neilson, B. Schmidt, and B. Xie. 2024. Are recognized expected credit losses
decision-useful and new to investors? Evidence from CECL adoption. Working paper.

35
Harris, T. S., U. Khan, and D. Nissim. 2018. The expected rate of credit losses on banks’ loan
portfolios. The Accounting Review 1 (5): 245–271.
Hilary, G. and C. Hsu. 2013. Analyst forecast consistency. The Journal of Finance 68 (1): 271–
297.
Hilary, G., C. Hsu, and R. Wang. 2014. Management forecast consistency. Journal of Accounting
Research 52 (1): 163–191.
Hinson, L. A., G. P. Puendrich, and M. Zakota. The decision-usefulness of ASC 606 revenue
disaggregation. The Accounting Review 99 (3): 225–258.
Horton, J., G. Serafeim, and I. Serafeim. 2013. Does mandatory IFRS adoption improve the
information environment? Contemporary Accounting Research 30: 388–423.
Kim, S., S. Kim, A. Kleymenova, and R. Li. 2023. Current expected credit losses (CECL) standard
and banks' information production. Working paper.
Kothari, S. P., and R. G. Sloan. 1992. Information in prices about future earnings: Implications for
earnings response coefficients. Journal of Accounting and Economics 15: 143–171.
Landsman, W. R., and E. L. Maydew. 2002. Has the information content of quarterly earnings
announcements declined in the past three decades? Journal of Accounting Research 40 (3):
797 – 808.
Landsman, W. R., E. L. Maydew, and J. R. Thornock. 2012. The information content of annual
earnings announcements and mandatory adoption of IFRS. Journal of Accounting and
Economics 53 (1): 34–54.
Lejard, C., E. Paget-Blanc, and J. Casta. 2021. The effects of the adoption of IFRS 9 on the
comparability and the predictive ability of banks' loan loss allowances. Working paper.
Livnat, J., and Y. Zhang. 2012. Information interpretation or information discovery: Which role
of analysts do investors value more? Review of Accounting Studies 17: 612–641.
López-Espinosa, G., G. Ormazabal, and Y. Sakasai. 2021. Switching from incurred to expected
loan loss provisioning: Early evidence. Journal of Accounting Research 59: 757–804.
Lu, Y., and V. Nikolaev. 2022. Expected loan loss provisioning: An empirical model. The
Accounting Review 97 (7): 319 – 346.
Maurer, M. 2020. Lawmakers press rule maker on economic impact of credit-loss standard. The
Wall Street Journal (January 15). https://www.wsj.com/articles/lawmakers-press-rule-maker-
on-economic-impact-of-credit-loss-standard-11579139500.
Mummolo, J., and E. Peterson. 2018. Improving the interpretation of fixed effects regression
results. Political Science Research and Methods 6 (4): 829–835.
Oberson, R. 2021. The credit-risk relevance of loan impairments under IFRS 9 for CDS pricing:
Early evidence. European Accounting Review 30 (5): 959-987.
Onali, E., G. Ginesti, G. Cardillo, and G. Torluccio. 2021. Market reaction to the expected loss
model in banks. Journal of Financial Stability, 100884.
Pan, Y., Y. Wang, and H. Wu. 2017. What do half a million loans say about the impact of CECL
on loan loss allowance? Moody’s Analytics Risk Perspectives 9 (July).
Tan, H., S. Wang, and M. Welker. 2011. Analyst following and forecast accuracy after mandated
IFRS adoptions. Journal of Accounting Research 49: 1307–1357.
Wheeler, B. 2021. Unrecognized expected credit losses and bank share prices. Journal of
Accounting Research 59: 805–866.
Zellner, A. 1962. An efficient method of estimating seemingly unrelated regressions and tests for
aggregation bias. Journal of the American Statistical Association 57 (298): 348-368.

36
Appendix A: Variable Definitions (in Alphabetical Order)

Variable Definition
2017Q2–2019Q3 2017Q2 (2017Q3, 2017Q4, 2018Q1, 2018Q2, 2018Q3, 2018Q4, 2019Q1, 2019Q2,
2019Q3) is an indicator variable equal to 1 if the quarter is 2017Q2 (2017Q3, 2017Q4,
2018Q1, 2018Q2, 2018Q3, 2018Q4, 2019Q1, 2019Q2, 2019Q3), 0 otherwise.
AssetsCubed The cubic form of total assets in millions (SPCIQ Keyfield 280297).
AssetsSquared The square form of total assets in millions (SPCIQ Keyfield 280297).
BigN An indicator variable equal to 1 if a bank is audited by a Big-N auditor as of the most
recent fiscal year, 0 otherwise (calculated using au from Compustat).
CAR[t-365,t-1] The cumulative abnormal buy-and-hold return (calculated using ret and vwretd in CRSP)
from one year before the earnings announcement date through the day before the
earnings announcement date.
CECL An indicator variable equal to 1 for banks adopting CECL on 1/1/2020, 0 otherwise.
ConsumerLoans Total consumer loans outstanding (SPCIQ Keyfield 290161), scaled by total assets
(SPCIQ Keyfield 280297).
DriftCAR The cumulative abnormal buy-and-hold return (calculated using ret and vwretd in CRSP)
from the day after the current earnings announcement date through the date of the
subsequent earnings announcement date.
Earnings Annual earnings (sum of Keyfield 329294, or 329255 if the former is missing, for the
four quarters during the year) scaled by lagged total assets (SPCIQ Keyfield 280297).
HighNews (LowNews) HighNews (LowNews) is an indicator variable equal to 1 if the cosine similarity score
between the current and prior quarters' vintage disclosures is below (above) the sample
median for the quarter, 0 otherwise. See Section 5.2.3 for further discussion.
IPT CARt
Intraperiod timeliness, calculated as IPTt = ∑4t=0 ( ), where CAR is the cumulative
CAR5
abnormal buy-and-hold return (calculated using ret and vwretd in CRSP) from day t0
through day t relative to the earnings announcement date.
LLP The loan loss provision in the current quarter scaled by total assets (SPCIQ Keyfield
280297). For the loan loss provision, we first use SPCIQ Keyfield 271740; if Keyfield
271740 is missing, we use Keyfield 272026 and then Keyfield 309178 as needed to
preserve sample size.
LLP The change in LLP from quarter t-1 to the current quarter.
LLPForecastDispersion Dispersion of analyst loan loss provision forecasts, defined as the standard deviation of
the individual analyst provision forecasts included in the consensus (SPCIQ Keyfield
332442), scaled by lagged total assets (SPCIQ Keyfield 280297).
|LLPForecastError| Forecast errors in analyst forecasts of loan loss provisions, defined as the absolute
difference between the consensus (mean) analyst provision forecast (SPCIQ Keyfield
332438) and the actual provision for the bank-quarter (defined the same as LLP), scaled
by lagged total assets (SPCIQ Keyfield 280297) following Beatty and Liao (2021).
LLPTimeliness LLPTimeliness is calculated in two steps. First, for each bank, we estimate the following
regression separately for the pre-period and post-period: LLPt = 0 + 1NPLt- +
2NPLt + 3NPLt+ + . Second, LLPTimeliness is the average of the estimated 2 and
3 from this regression.
LLPVol The standard deviation of LLP, measured separately over the pre-period and post-period.
LogAssets The natural logarithm of total assets in millions (SPCIQ Keyfield 280297). We report
the untransformed variable (i.e., Assets) in Table 1 and Table 8, Panel B.
LogNumLLPEst The natural logarithm of the number of analyst provision forecasts included in the
consensus (SPCIQ Keyfield 332443). We report the untransformed variable (i.e.,
NumLLPEst) in Table 1.
LogNumTPEst The natural logarithm of the number of analysts making a target price forecast with a
horizon equal to 12 months (horizon from IBES “ptgdet” table) during the quarter (based
on anndats from IBES “ptgdet” table).
Loss An indicator variable equal to 1 if net income (SPCIQ Keyfield 280344) is negative for
the quarter, 0 otherwise.

37
Appendix A (continued)
Variable Definition
Moderator In Table 3, Moderator is either PreLLPVol (Column 1), PreRetVol (Column 2), or
PreLLPFcstAcc (Column 3), which are defined below.
MTB The market value of equity divided by the book value of equity, obtained from
Compustat. The market value of equity is prccq x cshoq from the Fundamentals
Quarterly table; if those variables are missing, we use prccm x cshoq from the Security
Monthly table. The book value of equity is calculated as total shareholders’ equity (seqq
from Compustat; if missing, we use (i) ceqq + pstkq and then (ii) atq – ltq – mibq) minus
preferred shareholders equity (pstkq in Compustat). If the components of MTB are
missing from Compustat, we use the market value of equity (SPCIQ Keyfield 275838,
multiplied by 1,000) and book value of equity (SPCIQ Keyfield 280318) from SPCIQ.
NPLt NPLt is the quarterly change in nonperforming loans (sum of SPCIQ Keyfields 281530
(NPLt-, NPLt+) and 281489), scaled by lagged total loans (SPCIQ Keyfield 290178). NPLt- (NPLt+)
is the change in nonperforming loans scaled by lagged total loans, averaged over the
prior (lead) two quarters.
Post An indicator variable equal to 1 for quarters after 1/1/2020, 0 otherwise.
Post2020 Post2020 (Post2021, Post2022) is an indicator variable equal to 1 if the year is 2020
(Post2021, Post2022) (2021, 2022), 0 otherwise.
PreLLPFcstAcc The average analyst LLP forecast accuracy (calculated as -1 x |LLPForecastError|)
during the pre-period.
PreLLPVol The standard deviation of the bank's quarterly LLP during the pre-period.
PreRetVol The average of the bank's quarterly return volatility (calculated using daily ret from the
CRSP Daily Stock table) during the pre-period.
RealEstateLoans Total real estate loans outstanding (SPCIQ Keyfield 290155), scaled by total assets
(SPCIQ Keyfield 280297).
SUE Standardized unexpected earnings, defined as actual earnings (the first of the following
nonmissing SPCIQ Keyfields: 325470, 329294, or 329255) minus the mean consensus
forecast (SPCIQ Keyfield 317391), scaled by the standard deviation of the consensus
forecast (SPCIQ Keyfield 317395). SUE is decile-ranked and scaled to vary between -
0.5 and 0.5.
Tier1Ratio The tier 1 risk-based capital ratio (SPCIQ Keyfield 280216).
TPForecastDispersion Dispersion of analyst target price forecasts, defined as the standard deviation of analysts’
most recent target price forecast with a horizon equal to 12 months (horizon from IBES
“ptgdet” table) during the quarter (based on anndats from IBES “ptgdet” table).
|TPForecastError| Average forecast error in analyst target price forecasts made during the quarter (based
on anndats from IBES “ptgdet” table) with a horizon equal to 12 months (horizon from
IBES “ptgdet” table), calculated using analysts’ most recent target price forecast. The
forecast error for an individual analyst is defined as the absolute difference between the
target price forecast and the stock price 12 months later (absolute value of prc in CRSP),
scaled by the stock price 3 days before the forecast (absolute value of prc in CRSP).
UStat The ratio of the squared residual return during the testing period to the variance of
residual returns during the estimation period. The testing period is the event window [0,
+1] relative to the event date (i.e., earnings announcement, LLP forecast revision, or 10-
K/Q filing date); in Table 3, Panel B, Column 1 (Column 2), the testing period is [0, +5]
([0, +10]). The estimation period includes the windows [-130, -10] and [+10, +130]
relative to the event date. To calculate UStat, we first estimate the market model with
daily stock returns in the estimation period, obtain estimates of the intercept and slope
coefficient (i and i), and calculate the residual returns and variance (Vari). Using i
and i to calculate the residual return (i,t) during the testing period, we construct UStat
i,t
as: . Daily returns and market returns are obtained from CRSP (ret from the Daily
Vari
Stock table and vwretd from the Daily Stock Index table, respectively).

38
Table 1
Summary statistics

Panel A: CECL Banks

Pre-Period Post-Period

Variable N Mean Std. Dev. P25 P50 P75 N Mean Std. Dev. P25 P50 P75

Earn. Ann. UStatt 1,746 8.933 11.836 1.520 4.210 10.980 1,348 5.065*** 7.732 1.080 2.572 5.769
Forecast Rev. UStatt 7,714 4.520 7.140 0.679 1.874 4.953 5,289 3.167*** 5.337 0.515 1.454 3.431
10-K/10-Q UStatt 1,684 1.481 2.285 0.287 0.780 1.799 1,293 3.575*** 6.454 0.428 1.351 3.738
NumLLPEstt 1,746 6.792 3.219 5.000 6.000 8.000 1,348 6.237*** 3.186 4.000 6.000 8.000
MTBt-1 1,746 1.513 0.463 1.188 1.434 1.761 1,348 1.201*** 0.495 0.864 1.110 1.394
BigNt-1 1,746 0.643 0.479 0.000 1.000 1.000 1,348 0.612 0.487 0.000 1.000 1.000
Losst-1 1,746 0.005 0.068 0.000 0.000 0.000 1,348 0.042*** 0.200 0.000 0.000 0.000
DLLPt-1 1,746 0.019 0.489 -0.098 0.000 0.123 1,348 0.000 1.310 -0.494 -0.025 0.359
Tier1Ratiot-1 1,746 12.669 2.106 11.260 12.210 13.450 1,348 12.613 1.979 11.200 12.225 13.510
ConsumerLoanst-1 1,746 0.049 0.086 0.004 0.015 0.060 1,348 0.045** 0.083 0.004 0.013 0.051
RealEstateLoanst-1 1,746 0.456 0.180 0.360 0.487 0.580 1,348 0.415*** 0.169 0.328 0.430 0.538
Assetst-1 1,746 87,236.290 313,363.200 5,443.181 11,829.440 30,833.010 1,348 106,938.500*** 335,410.800 7,601.824 17,797.750 45,142.630

Panel B: Non-CECL Banks

Pre-Period Post-Period

Variable N Mean Std. Dev. P25 P50 P75 N Mean Std. Dev. P25 P50 P75

Earn. Ann. UStatt 927 4.533 7.155 0.532 1.856 5.372 743 3.330*** 5.706 0.434 1.481 3.683
Forecast Rev. UStatt 1,789 2.915 4.859 0.421 1.268 3.362 1,365 3.467** 6.644 0.435 1.213 3.401
10-K/10-Q UStatt 854 2.279 4.516 0.318 0.846 2.268 689 3.096*** 5.637 0.389 1.161 3.145
NumLLPEstt 927 3.188 2.133 2.000 3.000 4.000 743 2.685** 1.458 2.000 2.000 4.000
MTBt-1 927 1.392 0.326 1.169 1.336 1.559 743 1.032*** 0.317 0.808 0.996 1.161
BigNt-1 927 0.122 0.327 0.000 0.000 0.000 743 0.039*** 0.194 0.000 0.000 0.000
Losst-1 927 0.022 0.145 0.000 0.000 0.000 743 0.017 0.131 0.000 0.000 0.000
DLLPt-1 927 0.015 0.606 -0.113 0.000 0.137 743 0.008 0.948 -0.338 0.000 0.291
Tier1Ratiot-1 927 13.456 2.829 11.480 12.610 14.920 743 13.558 2.650 11.810 12.920 14.680
ConsumerLoanst-1 927 0.027 0.052 0.002 0.008 0.027 743 0.015** 0.023 0.002 0.007 0.016
RealEstateLoanst-1 927 0.602 0.128 0.523 0.597 0.701 743 0.565*** 0.124 0.483 0.568 0.643
Assetst-1 927 4,950.983 22,756.570 1,199.445 1,625.690 2,399.508 743 2,642.997 1,966.640 1,596.380 2,148.062 2,998.060

This table reports descriptive statistics and univariate tests of differences in the mean, with the sample of CECL banks reported in Panel A and non-CECL banks reported in Panel B. Tests of differences in the mean are based on OLS regressions
with the Post indicator included in the model to allow for clustering of standard errors by firm. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

39
Table 2
Information content of earnings announcements, LLP forecast revisions, and 10-K/10-Q filings

Panel A: Full Model

Dep. Var.: UStat t Col. (1): Earnings Col. (2): LLP Col. (3):
Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

(1) CECL x Postt -3.226*** -2.644*** 1.414***

(-7.21) (-6.09) (3.86)
(2) LogNumLLPEstt 0.773** 0.409** -1.724***
(2.26) (2.22) (-6.30)
(3) MTBt-1 0.326 0.720*** 0.157
(1.15) (4.24) (1.13)
(4) BigNt-1 1.551 -0.465 -0.086
(1.54) (-0.70) (-0.23)
(5) Losst-1 0.040 1.816*** -0.364
(0.05) (2.71) (-0.49)
(6) DLLPt-1 0.213* -0.094* 0.139
(1.80) (-1.72) (1.26)
(7) Tier1Ratiot-1 0.037 0.151 0.279*
(0.12) (0.85) (1.84)
(8) ConsumerLoanst-1 0.702 0.784 -0.898
(0.54) (0.94) (-1.18)
(9) RealEstateLoanst-1 1.849*** 1.461*** -1.543***
(3.06) (3.64) (-4.51)
(10) LogAssetst-1 3.119* 1.653 3.673***
(1.90) (1.57) (3.48)

c2 Test: Sum of Coefficients c2 Stat. (P-Val.)

S(1), Col. 1-3 = 0 -4.456*** 41.23 (0.000)

Bank FE Yes Yes Yes

Year FE Yes Yes Yes
N 4,764 16,157 4,520
Adjusted R2 11.51% 5.02% 7.70%

Panel B: No Fixed Effects and No Control Variables

Dep. Var.: UStat t Col. (1): Earnings Col. (2): LLP Col. (3):
Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

(1) CECL x Postt -2.665*** -1.906*** 1.277***

(-5.76) (-5.74) (4.49)

c2 Test: Sum of Coefficients c2 Stat. (P-Val.)

S(1), Col. 1-3 = 0 -3.294*** 26.39 (0.000)

Controls No No No
Bank FE No No No
Year FE No No No
N 4,764 16,157 4,520
Adjusted R2 5.52% 1.15% 3.28%

40
Table 2 (continued )

Panel C: Year Fixed Effects and No Control Variables

Dep. Var.: UStat t Col. (1): Earnings Col. (2): LLP Col. (3):
Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

(1) CECL x Postt -2.907*** -1.975*** 1.303***

(-6.21) (-5.97) (4.56)

c2 Test: Sum of Coefficients c2 Stat. (P-Val.)

S(1), Col. 1-3 = 0 -3.579*** 31.09 (0.000)

Controls No No No
Bank FE No No No
Year FE Yes Yes Yes
N 4,764 16,157 4,520
Adjusted R2 8.12% 1.87% 6.23%

Panel D: Year Fixed Effects and Control Variables

Dep. Var.: UStat t Col. (1): Earnings Col. (2): LLP Col. (3):
Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

(1) CECL x Postt -3.100*** -2.117*** 1.308***

(-6.84) (-6.07) (4.48)

c2 Test: Sum of Coefficients c2 Stat. (P-Val.)

S(1), Col. 1-3 = 0 -3.909*** 36.24 (0.000)

Controls Yes Yes Yes

Bank FE No No No
Year FE Yes Yes Yes
N 4,764 16,157 4,520
Adjusted R2 9.07% 2.21% 6.73%
This table reports the results of regressions examining the information content of earnings announcements, analysts' LLP
forecast revisions, and 10-K/10-Q filings following the adoption of CECL. Panel A presents the results using our full model
and Panels B - D present the results examining the robustness of the results to design choices regarding the inclusion of fixed
effects and control variables. CECL is an indicator variable equal to 1 for banks adopting CECL on 1/1/2020, 0 otherwise.
Post is an indicator variable equal to 1 for quarters after 1/1/2020, 0 otherwise. UStat is the squared residual return during the
testing period scaled by the variance of residual returns during the estimation period (Beaver et al. 2020). The testing period is
the event window [0, +1] relative to the earnings announcement, LLP forecast revision, or 10-K/10-Q filing date and the
estimation period is 130 to 10 days prior and 10 to 130 days after the event date. We estimate the market model during the
estimation period and use the estimated intercept and slope coefficients to calculate residual returns. The bottom of the table
presents the results of a chi-squared test examining whether the sum of the coefficients on CECL x Post across Columns 1 - 3
is significantly different from zero. Standard errors are clustered by bank. All continuous independent variables are
standardized to have a mean of zero and a standard deviation of one for ease of interpretation. ***, **, and * represent
statistical significance at the 1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

41
Table 3
Information content of earnings announcements: Cross-sectional tests and alternative explanations

Panel A: Cross-Sectional Tests

Dependent Variable: UStat t Col. (1): Col. (2): Col. (3):

Moderator Variable: PreLLPVol PreRetVol PreLLPFcstAcc

Moderator x Postt -0.124 -0.467* -0.250

(-0.43) (-1.76) (-0.65)
CECL x Postt -3.301*** -3.612*** -3.273***
(-7.44) (-8.16) (-7.37)
Moderator x CECL x Postt -0.582* -1.126*** 1.138**
(-1.68) (-2.60) (2.49)

Controls Yes Yes Yes

Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 4,764 4,764 4,764
2
Adjusted R 11.53% 11.79% 11.57%

Panel B: Speed of Price Discovery

Col. (1): UStat t Col. (2): UStat t Col. (3): Col (4):
Dependent Variable: 5-day window 10-day window IPT t DriftCAR t

SUEt 0.034**
(2.46)
CECL x SUEt -0.024
(-1.50)
CECL x Postt -4.663*** -4.269*** 0.010 0.016*
(-7.47) (-5.55) (0.03) (1.71)
SUE x Postt 0.090***
(3.92)
CECL x SUE x Postt -0.009
(-0.34)

Controls Yes Yes Yes Yes

Bank FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
N 4,764 4,764 4,295 4,167
2
Adjusted R 9.34% 5.32% 0.90% 20.87%

42
Table 3 (continued )

Panel C: Stock Returns Leading Earnings

Dependent Variable: CAR [t-365,t-1] Col. (1): Col. (2):

Sample: Full Sample Excluding 2020

Earningst 0.077*** 0.070***

(3.56) (2.69)
CECL x Earningst -0.043* -0.017
(-1.81) (-0.60)
CECL x Postt 0.046** -0.038*
(2.32) (-1.66)
Earnings x Postt 0.059** 0.045
(2.44) (1.51)
CECL x Earnings x Postt -0.045 -0.022
(-1.61) (-0.65)

Controls Yes Yes

Bank FE Yes Yes
Year FE Yes Yes
N 1,141 940
Adjusted R2 58.17% 54.51%
This table reports the results of additional analyses examining the impact of CECL adoption on the information content
of earnings announcements. CECL is an indicator variable equal to 1 for banks adopting CECL on 1/1/2020, 0
otherwise. Post is an indicator variable equal to 1 for quarters after 1/1/2020, 0 otherwise. Panel A presents the results
of cross-sectional tests. UStat is the squared residual return during the testing period scaled by the variance of residual
returns during the estimation period (Beaver et al. 2020). In Panel A, the testing period is the event window [0, +1]
relative to the earnings announcement and the estimation period is 130 to 10 days prior and 10 to 130 days after the
event date. We estimate the market model during the estimation period and use the estimated intercept and slope
coefficients to calculate residual returns. PreLLPVol is the standard deviation of the bank's quarterly loan loss
provisions during the pre-period. PreRetVol is the average of the bank's quarterly return volatility during the pre-period.
PreLLPFcstAcc is the average analyst LLP forecast accuracy during the pre-period. Panel B presents the results of tests
examining the impact of CECL adoption on the speed of price discovery around earnings announcements. In Columns 1
and 2, the dependent variable, UStat , is defined the same as Panel A except that in Column 1 (2), the testing period is
[0, +5] ([0, +10]). In Column 3, the dependent variable, IPT , is intraperiod timeliness, calculated as CAR t /CAR 5 , where
CAR is the cumulative abnormal buy-and-hold return from day 0 through day t, relative to the earnings announcement
date, and the measure is summed over days 0 to 4. In Column 4, the dependent variable, DriftCAR , is the cumulative
abnormal buy-and-hold return from the day after the current earnings announcement date through the date of the
subsequent earnings announcement date. SUE is standardized unexpected earnings, defined as actual earnings minus
the mean consensus forecast, scaled by the standard deviation of the consensus forecast. We decile-rank SUE and then
scale it to vary between -0.5 and 0.5. Panel C presents the results of tests examining the ability of stock returns to
predict future earnings. The dependent variable, CAR [t-365,t-1], is the cumulative abnormal buy-and-hold return from one
year before the earnings announcement date through the day before the earnings announcement date. Earnings is the
bank's annual earnings scaled by lagged total assets. Standard errors are clustered by bank. Except for SUE (as
discussed above), all continuous independent variables are standardized to have a mean of zero and a standard deviation
of one for ease of interpretation. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels,
respectively (two-tailed test). See Appendix A for variable definitions.

43
Table 4
Properties of analyst loan loss provision and target price forecasts

Panel A: Properties of Analyst Loan Loss Provision Forecasts

Column (1) Column (2) Column (3)

Dependent Variable: |LLPForecastError | LLPForecastDispersion LogNumLLPEst

CECL x Postt 0.023*** 0.013*** -0.173***

(5.29) (7.80) (-4.51)
LogNumLLPEstt -0.003 0.005***
(-1.15) (4.68)
MTBt-1 0.006*** 0.002* 0.032**
(2.83) (1.87) (2.07)
BigNt-1 0.007 0.003 0.023
(0.95) (0.93) (0.53)
Losst-1 0.116*** 0.027*** -0.001
(8.76) (3.99) (-0.04)
DLLPt-1 -0.002 0.002*** 0.013***
(-1.37) (5.04) (2.88)
Tier1Ratiot-1 0.005** 0.000 0.025
(2.19) (-0.05) (1.38)
ConsumerLoanst-1 0.007 0.006 0.007
(0.66) (1.58) (0.14)
RealEstateLoanst-1 0.018*** 0.007*** 0.089**
(3.45) (2.88) (2.45)
LogAssetst-1 -0.017 -0.014*** 0.348***
(-1.52) (-3.06) (3.75)
Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 4,777 4,339 4,777
2
Adjusted R 40.79% 40.18% 85.51%

44
Table 4 (continued )

Panel B: Properties of Target Price Forecasts

Column (1) Column (2) Column (3)

Dependent Variable: |TPForecastError | TPForecastDispersion LogNumTPEst

CECL x Postt 0.043* 1.186*** -0.144***

(1.91) (6.11) (-3.81)
LogNumTPEstt -0.023*** 0.583***
(-3.77) (6.77)
MTBt-1 0.004 1.027*** 0.023
(0.51) (6.18) (1.53)
BigNt-1 -0.024 -0.172 0.014
(-0.83) (-0.67) (0.22)
Losst-1 0.083*** 0.278 0.004
(2.84) (1.01) (0.10)
DLLPt-1 0.016*** -0.054* -0.015**
(5.22) (-1.74) (-2.52)
Tier1Ratiot-1 -0.024** -0.113 0.032
(-2.15) (-1.16) (1.56)
ConsumerLoanst-1 0.026 -0.043 0.124*
(0.76) (-0.09) (1.86)
RealEstateLoanst-1 0.009 0.021 0.061
(0.31) (0.05) (1.23)
LogAssetst-1 0.071 0.878 0.466***
(1.40) (1.25) (4.96)
Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 4,078 3,518 4,078
2
Adjusted R 45.12% 75.41% 79.50%
This table reports the results of regressions examining the impact of CECL adoption on the attributes of analysts' LLP
forecasts (Panel A) and the attributes of analysts' target price forecasts (Panel B). CECL is an indicator variable equal to 1
for banks adopting CECL on 1/1/2020, 0 otherwise. Post is an indicator variable equal to 1 for quarters after 1/1/2020, 0
otherwise. |LLPForecastError | is the absolute difference between the consensus analyst provision forecast (based on the
mean) and the actual provision for the quarter, multiplied by 100 and scaled by lagged total assets. LLPForecastDispersion
is the standard deviation of analyst provision forecasts included in the consensus, scaled by lagged total assets.
LogNumLLPEst is the natural log of the number of analyst provision forecasts included in the consensus.
|TPForecastError | is the absolute difference between the analyst's latest target price for the quarter and the one-year-ahead
stock price, scaled by the stock price 3 days prior to the target price forecast date; this measure is then averaged across all
analysts providing a target price forecast during the quarter. TPForecastDispersion is the standard deviation of analysts'
latest target price forecasts during the quarter. LogNumTPEst is the natural log of the number of analysts providing a target
price forecast during the quarter. Standard errors are clustered by bank. All continuous independent variables are
standardized to have a mean of zero and a standard deviation of one for ease of interpretation. ***, **, and * represent
statistical significance at the 1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

45
Table 5
Information content of 10-K/10-Q filings: Vintage disclosures

Dep. Var.: UStat t Coeff. Est. T-Stat.

(1) HighNews x CECL x Postt 1.631*** 3.49

(2) LowNews x CECL x Postt 0.840* 1.89

F-Test: Coeff. Diff. F-Stat. (P-Val.)

(1) - (2) = 0 0.791* 3.43 (0.065)

Controls Yes
Bank FE Yes
Year FE Yes
N 4,212
2
Adjusted R 8.51%
This table reports the results of analyses examining whether vintage disclosures affect the impact of CECL adoption
on the information content of 10-K/10-Q filings. CECL is an indicator variable equal to 1 for banks adopting
CECL on 1/1/2020, 0 otherwise. Post is an indicator variable equal to 1 for quarters after 1/1/2020, 0 otherwise.
UStat is the squared residual return during the testing period scaled by the variance of residual returns during the
estimation period (Beaver et al. 2020). The testing period is the event window [0, +1] relative to the 10-K/10-Q
filing date and the estimation period is 130 to 10 days prior and 10 to 130 days after the event date. We estimate the
market model during the estimation period and use the estimated intercept and slope coefficients to calculate
residual returns. HighNews (LowNews ) is an indicator variable equal to 1 if the cosine similarity score between the
current and prior quarters' vintage disclosures is below (above) the sample median for the quarter, zero otherwise.
Observations in 2020Q1 are dropped from this analysis because there are no prior quarters with which to construct
the cosine similarity score; additional CECL observations are dropped if we are unable to obtain vintage disclosure
data for the bank during the current or prior quarter. Since HighNews and LowNews can only be calculated for
CECL banks in the post-period, lower-order interaction terms cannot be included in the model. The bottom of the
table presents the results of an F-test examining whether the difference in the interaction terms is significantly
different from zero. Standard errors are clustered by bank. ***, **, and * represent statistical significance at the
1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

46
Table 6
Loan loss provision timeliness and volatility

Panel A: Loan Loss Provision Timeliness

Dependent Variable: LLP t Coeff. Est. T-Stat.

DNPLt- x CECL x Postt -0.004 -0.60

DNPLt x CECL x Postt 0.021*** 3.16
DNPLt+ x CECL x Postt 0.023*** 2.92

Additional Controls Yes

Bank FE Yes
Year FE Yes
N 3,513
2
Adjusted R 69.53%

Panel B: Loan Loss Provision Volatility

Dependent Variable: LLPVol t Coeff. Est. T-Stat.

LLPTimelinesst -0.018 -1.21

LLPTimelinesst x CECL -2.131*** -2.74
LLPTimelinesst x Postt -0.010 -0.23
CECL x Postt 0.644*** 4.82
LLPTimelinesst x CECL x Postt 1.941** 2.49

Additional Controls Yes

Bank FE Yes
Time FE Yes
N 382
2
Adjusted R 67.92%
This table reports the results of analyses examining the impact of CECL adoption on LLP timeliness (Panel A) and volatility
(Panel B). CECL is an indicator variable equal to 1 for banks adopting CECL on 1/1/2020, 0 otherwise. Post is an indicator
variable equal to 1 for quarters after 1/1/2020, 0 otherwise. LLP is the bank's quarterly loan loss provision scaled by total
assets. DNPL t is the change in nonperforming loans from the prior quarter to the current quarter, scaled by lagged total loans.
DNPL t- (DNPL t+) is the change in nonperforming loans scaled by lagged total loans, averaged over the prior (lead) two
quarters. LLPVol is the standard deviation of the bank's quarterly loan loss provisions, measured separately over the pre-period
and post-period. LLPTimeliness is calculated in two steps. First, for each bank, we estimate the following regression separately
for the pre-period and post-period: LLP t = a0 + a1DNPL t- + a2DNPL t + a3DNPL t+ + e. Second, LLPTimeliness is the average
of the estimated a2 and a3 from this regression. In Panel A, we start with the same sample used in the main earnings
announcement test (i.e., Column 1 of Table 2, Panel A) and then drop observations with missing nonperforming loans data.
Lower-order interaction terms and the same control variables and fixed effects in Panel A of Table 2 are included in the model
but are not reported for brevity. In Panel B, since LLPVol and LLPTimeliness are calculated once in both the pre- and post-
period for each bank, we use at most two observations per bank (depending on data availability). The control variables from
Panel A of Table 2 are included in the model but they are averaged across both the pre-period and post-period to match the
level of observation of this test; bank and time fixed effects are also included in the model, where the time effect is an indicator
for the pre-period/post-period (rather than year fixed effects) due to the level of observation for the test. ***, **, and * represent
statistical significance at the 1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

47
Table 7
Persistence of CECL's effects

Dep. Var.: UStat t Col. (1): Earnings Col. (2): LLP Col. (3):
Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

(1) CECL x Post2020t -4.148*** -2.413*** 1.905***

(-8.43) (-5.05) (3.51)
(2) CECL x Post2021t -3.228*** -2.527*** 1.693***
(-5.72) (-5.14) (5.10)
(3) CECL x Post2022t -0.053 -4.028*** -1.311
(-0.04) (-3.09) (-1.41)
2 2
c Tests: Sum of Coefficients c Stat. (P-Val.)

S(1), Col. 1-3 = 0 -4.656*** 26.38 (0.000)

S(2), Col. 1-3 = 0 -4.062*** 27.75 (0.000)
S(3), Col. 1-3 = 0 -5.392*** 7.61 (0.006)

Controls Yes Yes Yes

Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 4,764 16,157 4,520
2
Adjusted R 11.67% 5.05% 8.14%
This table reports the results of regressions examining the persistence of CECL's effects. CECL is an indicator
variable equal to 1 for banks adopting CECL on 1/1/2020, 0 otherwise. Post2020 (Post2021 , Post2022 ) is an
indicator variable equal to 1 for quarters during 2020 (2021, 2022), 0 otherwise. UStat is the squared residual
return during the testing period scaled by the variance of residual returns during the estimation period (Beaver et al.
2020). The testing period is the event window [0, +1] relative to the earnings announcement, LLP forecast revision,
or 10-K/10-Q filing date and the estimation period is 130 to 10 days prior and 10 to 130 days after the event date.
We estimate the market model during the estimation period and use the estimated intercept and slope coefficients to
calculate residual returns. The bottom of the table presents the results of chi-squared tests examining whether the
sum of the coefficients on CECL x Post2020 , CECL x Post2021 , and CECL x Post2022 across Columns 1 - 3 are
significantly different from zero. Standard errors are clustered by bank. ***, **, and * represent statistical
significance at the 1%, 5%, and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

48
Table 8
Robustness tests for bank size

Panel A: Additional Size Controls

Col. (1): Earnings Col. (2): LLP Col. (3):

Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

Dep. Var.: UStat t Coeff. T-Stat. Coeff. T-Stat. Coeff. T-Stat.

(1) CECL x Postt -3.175*** -6.54 -2.338*** -5.27 1.689*** 4.62

(2) Assetst-1 8.105 1.06 3.916 1.55 3.735 0.87
(3) AssetsSquaredt-1 0.571 0.06 -5.444** -2.29 -1.805 -0.63
(4) AssetsCubedt-1 -0.707 -0.71 2.607** 2.19 1.310*** 3.05
(5) Assetst-1 x Postt -3.588 -1.16 -2.462** -2.23 -2.887*** -3.07
(6) AssetsSquaredt-1 x Postt 4.063 1.13 3.997** 2.29 3.040** 2.53
(7) AssetsCubedt-1 x Postt 0.517 0.56 -2.648** -2.18 -0.990** -2.24
2 2
c Test: Sum of Coefficients c Stat. (P-Val.)

S(1), Col. 1-3 = 0 -3.824*** 28.01 (0.000)

Additional Controls Yes Yes Yes

Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 4,764 16,157 4,520
2
Adjusted R 11.45% 5.09% 7.53%

Panel B: Trimming Sample Based on Bank Size

Col. (1): Earnings Col. (2): LLP Col. (3):

Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

Dep. Var.: UStat t Coeff. T-Stat. Coeff. T-Stat. Coeff. T-Stat.

(1) CECL x Postt -2.114*** -3.30 -1.861*** -2.80 2.047*** 3.64

2 2
c Test: Sum of Coefficients c Stat. (P-Val.)

S(1), Col. 1-3 = 0 -1.928* 3.48 (0.062)

Controls Yes Yes Yes

Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 2,356 6,109 2,228
2
Adjusted R 10.20% 4.04% 9.21%

Mean Assets (Millions) Mean Assets (Millions) Difference:

Observation Level CECL = 0 CECL = 1 T-Stat (P-Val.)

Bank-Quarter 5,339.979 7,537.405 1.33 (0.184)

This table reports the results of size robustness tests for the main results. Panel A reports the results of the tests including additional size controls. Panel B
reports the results of tests using a trimmed sample based on bank size. CECL is an indicator variable equal to 1 for banks adopting CECL on 1/1/2020, 0
otherwise. Post is an indicator variable equal to 1 for quarters after 1/1/2020, 0 otherwise. UStat is the squared residual return during the testing period scaled
by the variance of residual returns during the estimation period (Beaver et al. 2020). The testing period is the event window [0, +1] relative to the earnings
announcement, LLP forecast revision, or 10-K/10-Q filing date and the estimation period is 130 to 10 days prior and 10 to 130 days after the event date. We
estimate the market model during the estimation period and use the estimated intercept and slope coefficients to calculate residual returns. The bottom of each
panel presents the results of a chi-squared test examining whether the sum of the coefficients on CECL x Post across Columns 1 - 3 is significantly different
from zero. In Panel B, the sample is trimmed based on bank size. Specifically, we drop the top 50% (bottom 50%) of treatment (control) banks based on asset
size and test the robustness of our main results to this trimmed sample. At the bottom of Panel B, we report the results of a test of differences in the mean of
bank size (i.e., total assets) across treatment and control banks; these results are based on an OLS regression with the CECL indicator included in the model
to allow for clustering of standard errors by bank. Standard errors are clustered by bank throughout the table. All continuous independent variables are
standardized to have a mean of zero and a standard deviation of one for ease of interpretation. ***, **, and * represent statistical significance at the 1%, 5%,
and 10% levels, respectively (two-tailed test). See Appendix A for variable definitions.

49
Table 9
Parallel trends test

Col. (1): Earnings Col. (2): LLP Col. (3):

Information Event: Announcements Forecast Revisions 10-K/10-Q Filings

Dep. Var.: UStat t Coeff. T-Stat. Coeff. T-Stat. Coeff. T-Stat.

2017Q2 -2.103** -2.13 0.664 0.68 0.198 0.56

2017Q3 -0.923 -0.77 0.244 0.57 0.981** 2.16
2017Q4 -1.140 -0.76 0.320 0.68 0.588* 1.74
2018Q2 0.808 0.60 -0.263 -0.35 -0.296 -1.07
2018Q3 1.346 1.27 -0.797 -1.11 0.334 1.01
2018Q4 0.222 0.18 0.808 1.07 0.215 0.68
2019Q2 0.186 0.15 0.143 0.37 0.745** 2.04
2019Q3 -1.732* -1.74 0.638 1.09 -0.892*** -3.52
CECL x 2017Q2 1.129 0.64 -0.237 -0.22 -0.798** -1.99
CECL x 2017Q3 -0.088 -0.05 0.559 0.95 -0.054 -0.10
CECL x 2017Q4 -2.657 -1.36 0.032 0.05 -0.200 -0.51
CECL x 2018Q1 5.852*** 2.90 -1.363** -2.11 -0.528 -1.28
CECL x 2018Q2 1.542 0.72 0.881 1.03 -0.132 -0.37
CECL x 2018Q3 4.422** 2.00 1.771** 2.01 -0.001 0.00
CECL x 2018Q4 0.314 0.15 0.290 0.36 -1.003*** -2.71
CECL x 2019Q1 -0.653 -0.34 2.816*** 3.02 -0.506 -1.30
CECL x 2019Q2 2.963 1.54 0.810 1.13 0.177 0.36
CECL x 2019Q3 2.215 1.15 0.879 0.91 0.323 0.99
Controls Yes Yes Yes
Bank FE Yes Yes Yes
Year FE Yes Yes Yes
N 2,450 8,875 2,319
2
Adjusted R 9.52% 6.65% 10.00%
This table reports the results of the parallel trends tests for the main results. 2017Q2 , 2017Q3 , 2017Q4 , 2018Q1 ,
2018Q2 , 2018Q3 , 2018Q4 , 2019Q1 , 2019Q2 , and 2019Q3 are indicator variables equal to 1 if the quarter is
2017Q2, 2017Q3, 2017Q4, 2018Q1, 2018Q2, 2018Q3, 2018Q4, 2019Q1, 2019Q2, and 2019Q3, respectively; the
indicators are equal to 0 for 2017Q1. CECL is an indicator variable equal to 1 for banks adopting CECL on
1/1/2020, 0 otherwise. UStat is the squared residual return during the testing period scaled by the variance of
residual returns during the estimation period (Beaver et al. 2020). The testing period is the event window [0, +1]
relative to the earnings announcement, LLP forecast revision, or 10-K/10-Q filing date and the estimation period is
130 to 10 days prior and 10 to 130 days after the event date. We estimate the market model during the estimation
period and use the estimated intercept and slope coefficients to calculate residual returns. Standard errors are
clustered by bank. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively (two-
tailed test). See Appendix A for variable definitions.

50
 When are expected credit losses decision-useful and new to investors?
Evidence from CECL adoption

Kurt H. Gee
[email protected]
The Ohio State University

Jed J. Neilson
[email protected]
Penn State University

Brent A. Schmidt
[email protected]
Penn State University

Biqin Xie
[email protected]
Office of Financial Research, U.S. Department of the Treasury

March 2025

We acknowledge helpful comments from Diana Choi (discussant), Seil Kim (discussant), Alina Lerman, Karl Muller,
Stephen Ryan, Hal Schroeder, Jake Thomas, Laura Wellman, Sunny Yang, Ying Zhou, Youli Zou, and workshop
participants at the AAA FARS Conference, George Mason University, Midwest Accounting Research Conference,
Office of Financial Research, Pennsylvania State University, Rutgers University, Santa Clara University, University
of Connecticut, University of Georgia, University of Missouri, Villanova University, and Yale Summer Accounting
Conference. The views expressed in the paper do not necessarily reflect those of the U.S. Treasury or Office of
Financial Research.
 When are expected credit losses decision-useful and new to investors?
Evidence from CECL adoption

Abstract
The Financial Accounting Standards Board replaced the “incurred loss” (IL) model with the
“current expected credit loss” (CECL) model, which requires entities to recognize lifetime
expected credit losses upon loan origination. We investigate four controversies surrounding CECL
to provide evidence on when expected credit losses are decision-useful and new for investors. We
examine whether (1) CECL allowances are only decision-useful for larger banks, (2) CECL
allowances provide new credit loss information, (3) credit losses expected to emerge beyond one
year are relevant for investors, and (4) separately reporting IL and CECL allowance amounts has
more information content than reporting a single CECL allowance amount. Contrary to the views
of community banks and evidence in prior literature, we find that CECL allowances are decision-
useful for both small and large banks. However, CECL allowances are new information only for
smaller banks. We also find that CECL allowances are decision-useful regardless of the loss
emergence period, despite later-emerging losses being unrecognized under IFRS 9, CECL’s
international counterpart. Finally, we find that reporting a single CECL allowance amount is
relatively less decision-useful than an alternative accounting regime that reports incurred and
expected future credit losses separately.
1. Introduction

The incurred loss (IL) standard that precluded banks from recognizing credit losses until

they were probable was roundly criticized as being “too little, too late” and contributing to the

depth and duration of the 2007-2009 financial crisis.1 In response, the FASB and IASB developed

separate and different standards of expected credit loss recognition, with the IASB standard (IFRS

9) implemented in 2018 and the FASB standard (ASU 2016-13 – the current expected credit loss

standard – CECL hereafter) implemented in 2020. Initial research indicated that future expected

credit losses, both under IFRS 9 and based on researcher-constructed estimates prior to CECL

adoption, are decision-useful (Wheeler 2021; Lu and Nikolaev 2022; López-Espinosa, Ormazabal,

and Sakasai 2021).

However, significant debates emerged about the informativeness and relevance of expected

credit loss measurement for different bank business models and loan types. These debates, which

have not been addressed by prior research, were reflected in contrasting commentary during the

rulemaking process and ultimately divergent IFRS and U.S. GAAP standards. We discuss these

debates and how they motivate four primary research questions that we address.

First, both bank trade groups and academics suggested that expected credit loss information

may be decision-useful only for larger banks. The Independent Community Bankers of America

(ICBA) questioned whether CECL’s complex modeling would produce meaningful information

for investors in community banks, as these banks’ “highly tailored, relationship-based lending

model does not easily link with identifiable environmental factors that can accurately depict future

economic risks” (ICBA 2013). Perhaps consistent with this concern, prior research found that

researcher-constructed estimates of expected losses are not relevant for the equity pricing of

1
Dugan (2009); Financial Stability Forum (2009); Bernanke (2009); G20 (2009); Basel Committee on Banking
Supervision [BCBS] (2011, 2021); FASB (2016); Bischof, Laux, and Leuz (2021).

1
smaller U.S. banks, i.e., those with less than $10 billion in assets (Wheeler 2021). With respect to

the IFRS 9 standard, López-Espinosa et al. (2021) note that they “expect that the effect of IFRS 9

is stronger among [systematically important banks], if not unique to these banks.” Our first

research question explores whether CECL allowances are decision-useful for both large and small

banks.

Second, prior research raises questions about whether CECL would produce new

information that investors could not already derive under the IL model. For example, prior research

finds that investors and analysts were informed about expected credit losses that were not reflected

by the accounting system under the IL regime (e.g., Wheeler 2021; Beatty and Liao 2021). This is

particularly true for larger banks, which are generally viewed to have a more robust information

environment. Banks with at least $10 billion in total assets had historically been subject to stress

testing by banking regulators (Schmidt 2024), which might reveal information about expected

credit losses similar to that provided by CECL (e.g., the disclosure of banks’ anticipated credit

losses under various stress scenarios). For such large banks, CECL allowances may be decision-

useful but not provide new information for investors, which would be consistent with the finding

of Wheeler (2021) that the pricing of researcher-constructed expected losses is more pronounced

among larger banks after the Dodd-Frank Act stress testing. This leads to our second research

question, which is whether CECL allowances provide investors with new information and if this

differs for smaller versus larger banks.

Third, the significant differences between IFRS 9 and CECL raise questions about whether

credit losses expected to emerge beyond one year are relevant to investors. While CECL requires

the recognition of lifetime expected credit losses for all loans, IFRS 9 requires only the recognition

of losses expected in the next 12 months for stage-1 assets (i.e., those with low credit risk or

2
without a significant increase in credit risk since loan origination). Since the vast majority of loans

are classified as stage 1 under IFRS 9, this represents a significant difference between CECL and

IFRS 9.2 Losses expected to emerge after one year (hereafter referred to as “later-emerging credit

losses”) may not be relevant for investors, given the difficulty of forecasting so far ahead (Bonsall,

Schmidt, and Xie 2025). Consistent with this concern, a group of U.S. banks submitted a comment

letter to the FASB proposing an alternative approach where losses expected within the next year

would be recognized in income and the remainder of lifetime losses would be recognized in other

comprehensive income.3 Despite recognizing this theoretical tradeoff, empirical studies on

researcher-constructed expected credit losses do not discriminate between losses with different

emergence periods. Thus, our third research question examines whether estimates of later-

emerging credit losses under CECL are relevant to investors.

Finally, we examine whether separately reporting incurred and expected future credit loss

amounts would have relatively more information content than reporting a single CECL allowance

amount. The prior literature on researcher-constructed estimates of expected credit losses

examines a time when incurred losses are observable. However, CECL replaces IL allowances

with CECL allowances, meaning incurred and future expected credit losses are not separately

observable. Our fourth research question investigates whether reporting CECL allowances as a

single amount is less useful than the separate reporting of IL allowances and incremental CECL

allowances.

2
This is consistent with the observations of the CFA Institute, who reiterated it was “disappointed by the FASB and
IASB’s decision to develop different accounting models” (CFA 2019) and López-Espinosa et al. (2021), who note
that CECL and IFRS 9 implementation likely could differ due to “the significant differences between the two
standards” (p. 761).
3
The comment letter can be found here: https://fasb.org/page/ShowPdf?path=AR-
2018.UNS.021.FINANCIAL%20INSTITUTIONS%20SEE%20LISTED,0.pdf

3
 Our research design leverages a powerful setting for examining our four research questions.

For a set of 200 publicly listed U.S. banks required to adopt CECL for fiscal years beginning after

December 15, 2019, we gather credit loss allowances under both the IL and CECL standards for

the same bank at the same time. Specifically, we identify the amount of expected credit losses

incremental to the IL allowance that banks recognized to comply with the new standard on day

one (hereafter, “the CECL day-1 impact”). Our design uniquely allows us to isolate allowance

differences between the IL and CECL regimes and hold constant all other differences across banks

(i.e., bank, time, and bank × time differences).

First, we examine the decision-usefulness of CECL allowances across bank size. We find

that, after controlling for IL allowances, CECL allowances incrementally predict future loan

losses, measured as future cumulative net charge-offs, and equity prices (Wheeler 2021) for both

community banks (i.e., less than $10 billion in total assets) and larger banks. These findings are

inconsistent with concerns raised by bank trade groups and contrast with existing evidence

documented in the academic literature, both of which suggest that CECL allowances may not be

decision-useful for investors of smaller banks. Instead, our evidence suggests that credit loss

information under CECL is relevant for both large banks and smaller community banks.

Second, we investigate (1) whether CECL allowances provide new information to investors

or merely confirm expectations investors could form using available information prior to CECL

(e.g., Wheeler 2021), and (2) if this differs for larger versus smaller banks. Specifically, we

examine how investors respond to changes in estimated CECL day-1 impacts disclosed by banks

leading up to CECL adoption, as required by SEC Staff Accounting Bulletin No. 74 (SAB 74). If

investors fully anticipate expected credit losses independent of CECL, we would not expect

investors to respond to changes in banks’ estimated CECL day-1 impacts. We find statistically

4
significant stock market reactions to CECL day-1 impact forecast revisions, on average, consistent

with investors perceiving increases (decreases) in banks’ estimated CECL day-1 impacts as bad

(good) news not previously impounded into stock prices. However, we find that CECL allowances

provide new information only for community banks. This is consistent with investors in larger

banks having sufficient information to estimate future credit losses in the IL regime, but CECL

providing new information for investors in smaller banks.

Third, we analyze the decision-usefulness of CECL allowances across the expected credit

loss emergence period. We exploit the fact that real estate loans have longer durations than other

loan types. Therefore, to identify expected credit losses that relate to a longer loss emergence

period, we use (1) U.S. banks with a relatively high proportion of real estate loans and (2) day-1

impacts specifically attributable to real estate loans. After controlling for IL allowances, we find

that CECL allowances relating to a longer loss emergence period are relevant for investors. A

potential implication of these findings is that IFRS 9 may omit relevant expected credit loss

information by not recognizing losses expected to emerge beyond one year.4 Furthermore,

researchers should be cautious in drawing inferences about CECL based on IFRS 9 evidence. Our

findings echo López-Espinosa et al. (2021), who state that “…difference[s] between IFRS 9 and

CECL could have nontrivial implications.”

Finally, we explore whether separately reporting incurred and expected future losses (i.e.,

presenting IL allowances and CECL day-1 impacts separately) has relatively more information

content than reporting a single CECL allowance amount. We find that separate reporting has more

explanatory power for future loan losses and equity prices than a single CECL allowance amount.

Subsequent analyses reveal that CECL day-1 impacts and IL allowances have statistically different

4
We note that it is not possible to examine this question with a sample of banks applying IFRS 9, since we are
interested in examining expected credit losses omitted under that standard.

5
relationships with future loan losses, with this difference generally being larger as we extend the

horizon over which loan losses are measured. These results suggest that decision-useful

information is lost when IL allowances are replaced by total CECL allowances, rather than

separately reporting the IL losses, which are incurred, and the expected future credit losses.

Our study makes three contributions. First, our examination of CECL contributes to the

academic literature on expected credit losses. Our results indicate that CECL allowances are

decision-useful for larger and smaller banks alike, but that they provide new information only for

smaller banks. This stands in contrast to the findings and assumptions in prior literature, which

suggest that researcher-constructed expected losses are not relevant for the pricing of smaller U.S.

banks (Wheeler 2021) and that credit loss information under IFRS 9 is expected to be decision-

useful only for larger banks (López-Espinosa et al. 2021). Additionally, we show that credit losses

expected to emerge beyond one year are decision-useful for investors, demonstrating that

differences between CECL and IFRS 9 are nontrivial. Finally, we extend prior research by

examining relative information content. We find that CECL allowances provide relatively more

information content than IL allowances, but relatively less information than a regime that reports

incurred losses and expected future credit losses separately. Such evidence could not be provided

in prior research due to the lack of CECL allowances.

Second, we contribute to the academic literature on accounting standard setting. A large

body of research has examined the implementation and consequences of specific, major accounting

standards.5 Our findings suggest that CECL allowances contain new, decision-useful information,

5
For example, researchers have specifically examined: SFAS 131 related to segment reporting requirements (e.g.,
Berger and Hann 2003; Botosan and Stanford 2005; Durney, Gee, and Wiebe 2024), SFAS 123 and 123R related to
stock compensation expense (e.g., Aboody, Barth, and Kasznik 2004, 2006; Choudhary, Rajgopal, and Venkatachalam
2009; Barth, Gow, and Taylor 2012), and SFAS 161 requiring new disclosures about derivatives and hedging (e.g.,
Chen, Dou, and Zou 2021). See Section 2 for a discussion of the literature on credit loss recognition.

6
especially for smaller banks, whose information environments are generally considered to be less

robust compared to larger banks. Additionally, we highlight the implications of differing decisions

made by different standard setters. By demonstrating that later-emerging credit losses are decision-

useful for investors, we provide evidence that IFRS 9 may omit relevant credit loss information

that is recognized under CECL.

Third, our study has policy implications for standard setters. Our findings are consistent

with CECL “provid[ing] financial statement users with more decision-useful information about

the expected credit losses” relative to IL allowances, which was the FASB’s stated goal (FASB

2016). We find, however, that the separate reporting of incurred losses and expected future losses

is more informative than the reporting of a single total CECL allowance, which is consistent with

incurred losses and expected future losses representing distinct components of credit loss

estimates. Our evidence also suggests that expected credit loss information is relevant regardless

of bank size. This offers support for the FASB’s decision to require community banks to adopt

CECL, despite requests to exempt them (ICBA 2013).

2. Background

2.1. Institutional Details about CECL and IL Models

Prior to CECL, accounting for credit losses followed ASC 450-20 (formerly SFAS 5) for

loans not individually identified as impaired and ASC 310-10-35 (formerly SFAS 114) for loans

individually identified as impaired. Most credit losses were covered by ASC 450-20 and treated

as loss contingencies, wherein credit losses were recognized only when “probable” and

“reasonably estimable.” These pre-CECL accounting treatments are commonly referred to as the

“incurred loss” (IL) model. Financial statement users criticized the IL model because it (1) delayed

recognition of credit losses that were expected but did not meet the threshold of being “probable,”

7
and (2) led to difficult comparisons of losses across entities due to diversity in entities’

determination of “probable” (FASB 2016, para. BC3–BC7).

ASU 2016-13 “Measurement of Credit Losses on Financial Instruments” (FASB 2016)

removes the “probable” threshold and requires banks to estimate and record a reserve for lifetime

expected credit losses on loans at the time of loan origination. CECL applies principally to

financial instruments measured at amortized cost, including loans held for investment, held-to-

maturity (HTM) debt securities, purchased credit-deteriorated (PCD) assets, and trade receivables,

as well as off-balance-sheet credit exposures (e.g., unfunded loan commitments). Because loans

are the most significant financial instrument class for most banks, the largest economic impact of

CECL is on banks’ loans held for investment. While the IL model generally only required loan

loss estimates to be based on past events and current conditions, CECL requires consideration of

reasonable and supportable forecasts of factors that could affect the collectability of the reported

amount (FASB 2016, p. 3). Banks are required to record a cumulative-effect “day-1” adjustment

to credit loss allowances and retained earnings as of the beginning of the first reporting period in

which CECL becomes effective.

2.2. Prior Literature on Expected Credit Losses

2.2.1 Researcher-Constructed Estimates of Expected Credit Losses

A stream of recent studies develops empirical models to estimate expected credit losses

using publicly available data under the IL regime and concludes that loan loss allowances under

these researcher-constructed expected loss models are more informative than those under IL

models. Covas and Nelson (2018) estimate expected losses using a vector autoregression

methodology to forecast lifetime charge-off rates. Harris, Khan, and Nissim (2018) develop a

measure of the one-year-ahead expected rate of credit losses that contains incremental information

about one-year-ahead realized credit losses relative to the IL model. Lu and Nikolaev (2022)

8
develop an empirical model that predicts long-term loan losses and incorporates adjustments for

macroeconomic forecasts. Wheeler (2021) develops a measure of lifetime expected credit losses

using vintage analysis and finds that stock prices partially reflect this estimated unrecognized

expected loss information.6 These studies offered ex ante perspectives on CECL’s potential

usefulness to investors, but they did not examine amounts recognized under CECL, which could

differ from estimates constructed by researchers in different time periods. Furthermore, these

studies do not examine whether CECL allowances provide new information, if the relevance of

CECL allowances varies based on bank size and the loss emergence period, or whether separately

reporting IL and CECL amounts provides more information than reporting a single CECL amount;

we build on these prior studies by investigating these unanswered questions.

2.2.2 Research Examining IFRS 9 Implementation

More recent studies examine the impact of IFRS 9, a standard issued by the IASB that also

uses an expected credit losses approach. Most relevant to our study is that of López-Espinosa et

al. (2021), who examine a sample of large, systemically important non-U.S. banks and find that

credit loss provision amounts under IFRS 9 are more predictive of future bank risk than IL

provisions, especially in countries experiencing deterioration in credit conditions, with the stock

market and the CDS market reacting to disclosures of the day-1 impact of IFRS 9.7 While both

CECL and IFRS 9 use expected credit losses, the standards differ significantly on when and how

expected credit losses are recognized. IFRS 9 categorizes financial assets into three “stages” based

6
Relatedly, Beatty and Liao (2021) examine analyst forecasts of loan loss provisions as a proxy for expected credit
loss estimates of informed market participants. They find that these forecasts are incrementally predictive of expected
losses (using measures from Harris et al. 2018 and Wheeler 2021) and actual future losses beyond IL provisions,
especially for banks facing greater IL model constraints.
7
The impact of IFRS 9 is also studied by Lejard, Paget-Blanc, and Casta (2021), who find IFRS 9 leads to less
comparable information regarding the allowance for loan losses across banks, and Onali, Ginesti, Cardillo, and
Torluccio (2024), who find positive market reactions around various IFRS 9 adoption events.

9
on their credit quality and applies different credit loss recognition criteria to each stage.8

Specifically, for loans that represent most of a typical bank’s assets (i.e., stage-1 assets), IFRS 9

requires recognition of only the portion of lifetime credit losses expected within the next 12 months

(i.e., 12-month expected credit losses). In contrast, CECL requires recognition of the entire amount

of lifetime expected credit losses for all loans upon origination.

The substantive differences between the two standards are highlighted in comment letters

submitted to the FASB by organizations that represent investors. For example, the CFA Institute

expressed concerns that “the difference in the approaches proposed by the FASB and the IASB

are likely to produce substantially different credit impairment and interest income recognition

patterns over the life of a financial instrument” (CFA 2013). After adoption, the CFA Institute

reiterated that it was “disappointed by the FASB and IASB’s decision to develop different

accounting models” (CFA 2019). Furthermore, López-Espinosa et al. (2021, p. 761) call for

research about the specific implementation effects of CECL, noting, “…it is also plausible that the

effect of implementing the ECL in the United States differs from that of implementing IFRS 9 in

other countries. The difference could be driven not only by the significant differences between the

two standards, but also by the unique institutional characteristics of the United States.” Our

examination of CECL’s decision usefulness across the loss emergence period is partly motivated

by the noted differences between IFRS 9 and CECL.

2.2.3 Research Examining CECL Implementation

Among the concurrent papers that examine the consequences of CECL implementation in

the U.S., Chen, Dou, Ryan, and Zou (2025) find that banks that adopted CECL prior to the COVID-

8
Stage-1 assets include those with a low credit risk at the reporting date or without a significant increase in credit risk
since initial recognition. Stage-2 assets include underperforming financial assets which exhibit a significant increase
in credit risk since initial recognition. Stage-3 assets are those in which credit risk increases to a point where it is
considered credit-impaired.

10
19 pandemic reduced loan growth during the accompanying recession more than other banks, and

the effect is stronger for CECL adopters with low regulatory capital and weaker for adopters with

low heterogeneous loans or large initial CECL adoption impacts. Kim, Kim, Kleymenova, and Li

(2023) find that CECL-adopting banks have loan loss provisions that are timelier and better reflect

future local economic conditions, have fewer loan defaults, disclose more forward-looking

information, and have lending that becomes less sensitive to economic uncertainty after adopting

CECL. They conclude that adopting a more forward-looking accounting standard improved the

quality of banks’ internal information production. Bable, Wong, and Wynes (2023) find that

analysts express generally negative perceptions about CECL. Bonsall et al. (2025) find that CECL

adoption is associated with a decrease in information content of earnings announcements and

analyst LLP forecasts, and an increase in the informativeness of 10-K/Q filings, after CECL is

adopted.

Our study differs from these papers in two main aspects. First, our specific research

questions are outside the scope of these papers. In particular, these studies do not examine whether

CECL allowances provide new information to investors, how the decision-usefulness of CECL

allowances differs based on bank size and the expected loss emergence period, nor whether

separately reporting IL and CECL amounts provides more information than reporting a single

CECL amount.

Second, we focus on day-1 CECL implementation, which allows us to compare

contemporaneous IL and CECL allowances for the same bank at the same time for our analyses

and thereby effectively eliminate bank differences or time differences as alternative explanations

for our results. Concurrent studies examining the impacts of CECL, along with López-Espinosa et

al. (2021), generally utilize a difference-in-differences design. Using such a design to study CECL

11
has important limitations because (1) adopting banks are larger than non-adopting banks, (2)

CECL adoption coincides with the onset of the COVID-19 pandemic, and (3) market effects or

banks’ responses to the COVID-19 pandemic could differ based on bank size, confounding

identification of the effects of CECL adoption. Our identification strategy uniquely avoids the

limitation of the difference-in-differences design, with a tradeoff that we are unable to examine

the decision-usefulness of CECL allowances in quarters after day-1 adoption, a period in which

banks do not simultaneously provide both IL and CECL allowances.

3. Research Design, Sample Selection, and Descriptive Statistics

3.1. Research Design

3.1.1. The Decision-Usefulness of CECL Information Across Bank Size

We first examine how the ability of CECL allowances to predict future credit losses differs

across bank size. Prior research examines how credit loss reserves in the IL regime predict future

credit losses by examining the behavior of banks’ quarterly or annual loan loss provision, i.e., the

changes in their allowances (Beatty and Liao 2011, 2014; Bushman and Williams 2012, 2015;

Nicoletti 2018). We deviate from this design because CECL and IL provisions are not observable

contemporaneously; after adopting CECL, banks do not disclose the IL allowances or provisions,

and prior to adopting CECL, banks do not disclose the CECL allowances or provisions.

Contemporaneous accounting numbers under both the CECL and IL models are available only on

day one of CECL adoption in the form of the IL allowance and the CECL day-1 impact on

allowances. Thus, we examine the extent to which CECL and IL allowances on the adoption date

predict future credit losses. Following Wheeler (2021), we use future net charge-offs to proxy for

future credit losses.

12
 We separately estimate the following equation for CECL adopters above versus below $10

billion in total assets using OLS:9

CECL_Impacti IL_Allowancei RELoansi ConsLoansi

∑ t+1:16 = 0 + 1 + 2 + 3 + 4
Assetsi Assetsi Assetsi Assetsi
RateSensitivei NIBPi
+ 5 + 6 + i (1)
Assetsi Assetsi

where subscript i indexes banks. The dependent variable is net charge-offs scaled by total assets
NCO
cumulated over the future 16 quarters (( )t+1:16). CECL_Impact is the CECL day-1 impact,
Assets

scaled by total assets as of the adoption date. IL_Allowance is the allowance for credit losses under

the IL model, scaled by total assets as of the adoption date. We include several variables

representing the bank’s asset composition and business model, including real estate loans scaled

by total assets ( ), consumer loans scaled by total assets( ), and rate-sensitive assets
Assets Assets

maturing within one year scaled by total assets ( ). We also control for net income
Assets

before taxes and credit loss provisions scaled by total assets ( ), which reflects overall financial
Assets

performance. We winsorize all variables at the 1st and 99th percentiles and cluster standard errors

by bank to address heteroskedasticity. A significantly positive coefficient on CECL_Impact

implies that CECL allowances have incremental information content beyond IL allowances in

predicting future loan losses.

We next examine whether CECL day-1 impacts are consistent with the information

investors use when valuing bank stocks, and whether this differs across bank size. Specifically, we

estimate the following equation for CECL adopters above versus below $10 billion in total assets

on their adoption date using OLS:

CECL_Impacti IL_Allowancei BVE_Adjustedi RELoansi

Pricei = 0 + 1 + 2 + 3 + 4
Sharesi Sharesi Sharesi Sharesi

9
$10 billion in total assets represents the Federal Reserve’s cutoff for defining community banks
(https://www.federalreserve.gov/supervisionreg/community-and-regional-financial-institutions.htm).

13
 ConsLoansi RateSensitivei NIBPi NPLi
+ 5 + 6 + 7 + 8 + i (2)
Sharesi Sharesi Sharesi Sharesi

where subscript i indexes banks. The dependent variable is stock price. We measure stock price

two quarters after the CECL adoption date (so the bank’s first Form 10-Q filing after CECL

adoption is available to investors).10 CECL_Impact and IL_Allowance are as defined previously.

BVE_Adjusted is the book value of equity before the IL allowance and is prior to any recognition

of the CECL day-1 impact. We include several variables representing the bank’s business model

and loan composition, including RELoans, ConsLoans, RateSensitive, as well as NIBP and

nonperforming loans (NPL), which reflect overall financial performance and information about

underlying loan quality that is available to investors both before and after CECL. Following Barth

and Clinch (2009), we deflate all variables by common shares outstanding (Shares).

Consistent with prior research (e.g., McInnis, Yu, and Yust 2018; Barth, Li, and McClure

2023), we measure value relevance based on the explanatory power of accounting information for

equity values. A significantly negative coefficient on CECL_Impact implies that CECL allowances

have incremental information content beyond IL allowances in explaining equity values.

3.1.2. Are CECL Allowances New Information for Investors?

Next, we address whether CECL allowance information is new to investors, or whether it

merely confirms their expectations (consistent with investors incorporating expected loan loss

information before CECL adoption, e.g., Wheeler 2021). We address this question by examining

investor responses to banks’ SAB 74 disclosures of their estimated CECL day-1 impacts leading

up to adoption. While this CECL day-1 impact is officially reported in the first Form 10-Q filed

after CECL adoption, entities provide information about the CECL day-1 impact earlier, as

required by SAB 74. SAB 74 requires entities to disclose a new standard’s anticipated impact on

10
We find CECL day-1 impacts are incrementally predictive of price when measured at fiscal year-end (untabulated).

14
the company’s financial statements.11 The SEC paid special attention to enforcing SAB 74

disclosures relating to CECL, urging banks to “not let their implementation planning or disclosure

of the anticipated effects of [CECL] lag during 2019.”12

A well-specified test of investor response to information requires a measure of investors’

expectations, as investors should only respond to unexpected CECL day-1 impacts. We hand

collect banks’ SAB-74 disclosures of their expected CECL day-1 impact estimates from their 10-

K/Q filings and focus on revisions to their estimated CECL day-1 impacts. If investors learn new

information from banks’ revisions of their estimated CECL day-1 impacts, then we should observe

significant investor responses to them.

We use an event study design around banks’ quarterly earnings announcements and

periodic filing cycles. Specifically, we examine how abnormal stock market returns are associated

with banks’ revisions to their expected CECL day-1 impact estimates using the following

regression:

Return_SAB74i,t = 0 + 1ΔSAB74_Estimate/MVE + kControlsi,t + tFEt + i,t (3)

where Return_SAB74 is market-adjusted abnormal return, calculated as the bank’s daily return

minus the daily value-weighted market return, cumulated using a buy-and-hold (B&H) or

cumulative (CAR) approach. We use two alternative event windows for this test. First, we use the

window from the earnings announcement day through one day after the bank’s 10-K/Q filing.13

We use this event window because although banks disclose their estimated CECL day-1 impacts

in their 10-K/Q filing, they could have disclosed the information earlier via other disclosure

11
https://www.sec.gov/interps/account/sabcodet11.htm#M
12
https://www.sec.gov/news/speech/speech-bricker-121018-1
13
We use different approaches to cumulating the return because the length of the event window varies for each
observation depending on the time lag between the earnings announcement and the 10-K/Q filing, and cumulative
(buy-and-hold) returns are customary for shorter (longer) windows.

15
mediums (e.g., earnings conference calls); while we do not know precisely when the information

was disclosed, our event window captures investors’ response to any disclosure made between the

earnings announcement day and the 10-K/Q filing date. Second, we manually read each bank’s

earnings announcement 8-K filing in EDGAR to determine whether the bank disclosed a SAB 74

estimate. For the subsample of banks where we identify a SAB 74 estimate during the earnings

announcement (66% of the sample), we use the two-day window [t0, t+1] relative to the earnings

announcement date. ΔSAB74_Estimate/MVE is the change in a bank’s estimated CECL day-1

impact, scaled by the market value of equity.

The vector Controls includes control variables intended to capture the banks’ general

liquidity and information environment, the impact of unexpected earnings for that quarterly

reporting cycle (UE), and properties of the estimated CECL day-1 impact. We include the natural

log of one plus the number of analysts following the firm (ln(1+Analysts)); the natural log of the

market value of equity (ln(MVE)); the market-to-book ratio (Market-to-Book); the cumulative bank

stock return starting 50 trading days before the earnings announcement until five days before the

announcement (Pre-EA Return); analyst earnings forecast error scaled by stock price (UE); and an

indicator for point forecasts (Point), which takes a value of 0 for a range forecast (in which case

we take the mean of the high and low ends of the range estimate for purposes of constructing

SAB74_Estimate). We also include each quarter’s loan loss provision (Provision/MVE) as a control

for other loan loss information released concurrently with the CECL day-1 impact information.

We estimate equation 3 for our full sample and separately for banks above versus below

$10 billion in total assets. We winsorize all continuous variables at the 1st and 99th percentiles, and

cluster standard errors by bank to account for banks with multiple estimated day-1 impact

revisions.

16
3.1.3. The Decision-Usefulness of CECL Information Across the Loss Emergence Period

Our third set of tests examines whether the decision-usefulness of CECL allowance

information differs across the expected loss emergence period. These tests are motivated by

differences between CECL and IFRS 9. For loans with a low credit risk at the reporting date or

without a significant increase in credit risk since initial recognition (i.e., stage-1 loans), IFRS 9

recognizes only credit losses expected to occur within the next 12 months and omits credit losses

expected beyond 12 months. In contrast, CECL recognizes lifetime expected credit losses for such

loans. Despite our earlier evidence on the relevance of CECL allowances, estimates of later-

emerging credit losses might be too inaccurate to be relevant, such that the relevance of CECL

could primarily come from losses expected only over the very near term.

To understand the potential relevance of credit losses expected beyond the next 12 months

for stage-1 loans, we cannot use banks reporting under IFRS 9 because the standard does not

require recognition of such losses; thus, our question requires the examination of banks reporting

under CECL. Among banks adopting CECL, we would ideally isolate the precise portion of banks’

CECL day-1 impacts representing losses expected to emerge beyond the next 12 months for loans

classified as stage-1 under IFRS 9. This approach is not possible, however, because banks do not

disaggregate CECL day-1 impacts based on the loss emergence period.

We, therefore, attempt two complementary approaches that provide the next best

alternatives to investigate our question. First, we identify a set of banks with a high proportion of

loans meeting two criteria: (1) the loans have a long duration and (2) the loans would likely be

classified as stage-1 under IFRS 9. We use real estate loans—comprised mainly of residential and

commercial real estate mortgage loans—to proxy for long-duration loans because mortgages

typically have longer durations than non-real-estate loans (e.g., auto loans and commercial and

17
industrial loans).14 We approximate stage-1 real estate loans as total real estate loans minus the

sum of real estate loans more than 30 days past due and nonaccrual real estate loans. For a bank

with a high proportion of such loans, we expect a substantial portion of its CECL day-1 impact to

reflect expected credit losses beyond the next 12 months, which would be omitted under IFRS 9

but are recognized under CECL. We partition our sample of banks based on the sample median;

81.0 (48.3) of the above-median (below-median) banks’ loans are identified as real estate loans

that would likely be classified as stage 1 under IFRS 9. For a bank with a longer estimated loss

emergence period, evidence that CECL day-1 impacts are decision-useful would be consistent with

IFRS 9 potentially omitting relevant expected credit loss information. We estimate equations 1

and 2 for each of the two subsamples to understand the extent to which CECL day-1 impacts

provide decision-useful information for each set of banks.

Second, most banks in our sample disclose the day-1 impact by loan type. We hand-collect

these disclosures from the banks’ initial 10-Q filings following CECL adoption. In particular,

similar to the test described above, we are interested in the day-1 impact for real estate loans, since

these loans have a longer duration than other loan types. We estimate equations 1 and 2 but replace

the day-1 impact variable with the day-1 impact from real estate loans (RE_CECL_Impact) and

from non-real estate loans (NonRE_CECL_Impact).

3.1.4. The Decision-Usefulness of Separate Reporting Versus Reporting Single CECL Amount

Our final set of analyses examines whether separately reporting IL and CECL allowance

amounts has relatively more information content than reporting a single CECL allowance amount.

It is important to distinguish between incremental and relative information content (Biddle, Seow,

14
For example, among our sample banks are TFS Financial Corporation (nearly 100% real estate loans) and American
Express (0% real estate loans); in their 2019 10-Ks, the former discloses that 82% of its loans mature beyond 10 years,
while the latter discloses that 0.2% of its loans mature beyond 5 years.

18
and Siegel 1995) in the context of CECL because standard setters replaced the IL model, rather

than supplementing it with incremental expected credit loss information. To examine this issue,

we estimate equations 1 and 2, which include IL_Allowance and CECL_Impact separately, and

compare the explanatory power of the model to one that replaces these two variables with

CECL_Allowance, defined as IL_Allowance plus CECL_Impact. To compare the explanatory

power of the two models, we use the Clarke (2003; 2007) test.15

3.2 Sample Selection

An entity’s required CECL adoption date depended on whether the entity was considered

a smaller reporting company (SRC) by the SEC.16 Non-SRCs were required to adopt CECL in

fiscal years beginning after December 15, 2019, while SRCs were required to adopt CECL for

fiscal years beginning after December 15, 2022. We focus our study on non-SRCs, the first

adopters of CECL, such that our sample comprises banks required to adopt CECL between January

1, 2020 and January 1, 2021.17

Table 1 presents the sample selection procedure. We begin with all publicly traded U.S.

banks (identified using SIC codes 6000–6299) available on S&P Capital IQ Pro (“SPCIQ,”

formerly SNL Financial) with a credit loss allowance balance and coverage in CRSP as of

December 31, 2019, yielding 402 banks. Removing banks not adopting CECL yields 212 banks.

For the analyses on the decision-usefulness of the CECL day-1 impact (as described in Sections

15
Given two non-nested models, the Clarke test determines whether a significantly greater number of observations
have lower unexplained variation based on one model or the other. Clarke (2007) finds that this nonparametric test is
more powerful in choosing the model with better explanatory power than the Vuong (1989) test, and the Clarke test
has been used in recent accounting research (e.g., Barth et al. 2012; Campbell, Gee, and Wiebe 2022).
16
SRCs are companies that have (i) public float of less than $250 million or (ii) annual revenues less than $100 million
and either no public float or public float less than $700 million (https://www.sec.gov/smallbusiness/goingpublic/SRC).
17
Non-SRCs were afforded an option to delay adoption under the Coronavirus Aid, Relief, and Economic Security
(“CARES”) Act, passed on March 25, 2020, and the Consolidated Appropriations Act (2021), signed into law on
December 27, 2020. We identify 31 banks in our sample who delayed their CECL adoption until on January 1, 2021.
Our main results are similar if we remove these later adopters from our sample.

19
3.1.1, 3.1.3, and 3.1.4), we classify banks with non-missing CECL day-1 impacts (CECL_Impact)

data in SPCIQ as CECL adopters.18 After dropping banks without the necessary bank-level

variables from the “Companies” dataset in SPCIQ, we arrive at our final samples for the value

relevance (future credit losses) analyses of 200 (174) publicly traded banks that adopted CECL

between January 1, 2020 and January 1, 2021.

For the SAB 74 investor response analyses, we begin by hand-collecting SAB 74 estimated

CECL day-1 impacts for the 212 CECL-adopting banks described above. We drop 37 banks that

did not have at least two consecutive quarters of interpretable estimates (two quarters are required

to first establish a baseline for a subsequent revision to investor expectations). We then drop 17

banks with missing data for the variables required in the regression, leaving 158 banks, from which

we have 244 quarterly observations.

3.3. Descriptive Statistics

The descriptive statistics for variables used in the decision-usefulness analyses are

presented in Panel A, Table 2. The average CECL day-1 impact on credit loss reserves is
CECL_Impact
approximately 0.2% of total assets ( ); for comparison, the average allowance for loan
Assets

IL_Allowance
and lease losses under the IL model is approximately 0.7% of total assets ( ). Thus, for
Assets

the average bank, day-1 adoption of CECL increased its credit loss allowance by close to 30%.

For nearly 90% of the 200 sample banks, the CECL day-1 impact on allowances was positive (i.e.,

CECL adoption increased credit loss reserves for all but a small minority of sample banks).

18
Day-1 impact variables are obtained from the “Regulated Depositories (U.S.)” dataset in SPCIQ. When CECL day-
1 impacts in SPCIQ are 0, we hand-collect day-1 impact data from regulatory filings (i.e., FR Y-9C reports and call
reports) and SEC filings (i.e., 10-Qs) to confirm banks’ adoption of CECL and day-1 impact amounts. We also hand-
collect day-1 impact data for non-SRCs (i.e., banks required to adopt CECL) with missing CECL day-1 impacts in
SPCIQ, because it is unclear whether these banks have adopted CECL or delayed adoption.

20
 The descriptive statistics for variables used in the SAB 74 investor response analysis are

presented in Panel B, Table 2. The average investor response is negative and the average SAB 74

disclosure revises the bank’s estimated CECL day-1 impact upward (i.e., larger expected credit

losses). The mean of Pre-EA Return is -0.221, which is largely attributable to significant declines

in bank valuations during the latter part of Q1 2020.

4. Results

4.1. The Decision-Usefulness of CECL Information Across Bank Size

Table 3 presents the results of estimating equation 1 separately for larger banks (column 1)
CECL_Impact
and smaller banks (column 2). In both columns, the coefficient on is significantly
Assets

positive. These results indicate that, for both sets of banks, CECL day-1 impacts incrementally

predict future loan losses beyond IL allowances. Table 4 presents the results of estimating equation

2 separately for larger banks (column 1) and smaller banks (column 2). The significant coefficient
CECL_Impact
on in both columns indicates that, for both sets of banks, CECL day-1 impacts provide
Shares

incremental information to investors beyond IL allowances. The evidence in Tables 3 and 4

indicates CECL allowances are decision-useful for both large and small banks, in contrast to

concerns expressed by community banks during the development of CECL (ICBA 2013).

4.2. Are CECL Allowances New Information for Investors?

Figure 1 illustrates the timing of revisions to banks’ SAB 74 CECL estimates in the quarters

leading up to adoption; revisions are more frequent as CECL adoption approaches. Table 5, Panel

A reports the OLS estimation of equation 3, which examines investor response to banks’ revisions

of estimated CECL day-1 impacts. Columns 1 and 2 report the results using the full sample of

banks using the longer-window return variables. Columns 3 and 4 report the results using the

subsample of banks disclosing a SAB 74 estimate during the earnings announcement, and thus use

21
the shorter-window return variables. Return_SAB74 is measured as buy-and-hold abnormal return

(cumulative abnormal return) in columns 1 and 3 (2 and 4). The coefficient on

ΔSAB74_Estimate/MVE is negative and significant in all four columns, which suggests that

investors perceive an increase (decrease) in banks’ estimated CECL day-1 impacts as bad (good)

news. This is consistent with CECL providing new information to investors; if investors already

knew expected credit loss information independently from CECL, we would not expect a market

reaction to revisions.

While we interpret our results as CECL providing new information about expected credit

losses to investors, it is possible that the market response is related to other (i.e., non-SAB 74)

information about expected credit losses released contemporaneously in earnings announcements

or periodic (i.e., 10-K/Q) filings. Any such information could confound our inferences only if it

exhibits a similar signed relation with returns as does our ΔSAB74_Estimate/MVE variable. In

untabulated analyses, we supplement equation 3 with a control related to analysts’ loan loss

provision forecasts, which Beatty and Liao (2021) find reflect information about expected credit

losses and which we use to represent potential non-SAB 74 information. Inferences under this

approach are identical to those in Panel A of Table 5, i.e., the coefficients on

ΔSAB74_Estimate/MVE remain significantly negative. This finding offers additional evidence

consistent with CECL providing new information about expected credit losses for investors.19

Figure 1 indicates that many day-1 impact revisions are disclosed in Forms 10-K for Q4

2019 and Forms 10-Q for Q1 2020. Both forms were filed in 2020, potentially after banks knew

of COVID-19’s onset. Thus, some of the market responses that we document in Table 5 could be

19
In subsequent tests, we find that CECL day-1 impacts are only new information for smaller banks. For other credit-
loss information to explain our results, smaller banks would have to be more transparent with such information than
larger banks.

22
attributed to banks “backdating” COVID-19 information into their estimated day-1 impacts. Such

a practice by banks would not be in accordance with CECL implementation guidance because

information after the CECL adoption date should be reflected in allowances in future periods

instead of in the day-1 impact. If banks improperly backdated COVID-19 information, we would

expect to see larger estimated day-1 impacts (i.e., larger increases in allowances) for revisions

made in Forms 10-Q for Q4 2019 or Q1 2020. In untabulated analyses, we find that the average

bank (1) reduced (rather than increased) its day-1 impact estimate in its 10-K filing for Q4 2019

and (2) maintained almost the exact same estimated magnitude in its 10-Q filing for Q1 2020.

These findings are inconsistent with banks’ day-1 impacts reflecting COVID-19 information.

Table 5, Panel B reports the OLS estimation of equation 3 separately for larger banks

(columns 1 and 2) and smaller banks (column 3 and 4). We find a significantly negative coefficient

on ΔSAB74_Estimate/MVE only for smaller banks, consistent with CECL day-1 impacts providing

new information only for community banks (differences in coefficients across the samples are

significant at the p < 0.05 level, untabulated).20 Our results in Panel B of Table 5 imply that CECL

allowances provide new information for investors when other relevant expected credit loss

information is less available.

4.3. The Decision-Usefulness of CECL Information Across the Loss Emergence Period

Table 6 presents the results of estimating equation 1 separately for banks with a low versus

high proportion of loans with a longer expected loss emergence period (columns 1 and 2,

respectively) and for banks’ day-1 impacts from real estate vs non-real estate loans (column 3). In
CECL_Impact
columns 1 and 2, the significantly positive coefficients on indicate that, regardless of
Assets

20
Our finding does not address whether other information under CECL available after adoption, such as vintage
disclosures, provide new information; our evidence is strictly related to credit loss allowances under CECL.

23
the expected loss emergence period, CECL day-1 impacts incrementally predict future loan losses
RE_CECL_Impact
beyond IL allowances. In column 3, the coefficient on is also significantly positive,
Assets

suggesting that the day-1 impacts from loans with longer loss emergence periods incrementally

predict future loan losses compared to IL allowances, consistent with columns 1 and 2.

Table 7 presents the results of estimating equation 2 separately for banks with a low versus

high proportion of loans with a longer expected loss emergence period (columns 1 and 2,

respectively) and for banks’ day-1 impacts from real estate vs non-real estate loans (column 3). In
CECL_Impact
columns 1 and 2, the significantly negative coefficients on indicate that, regardless of
Shares

the expected loss emergence period, CECL day-1 impacts provide incremental information to
RE_CECL_Impact
investors beyond IL allowances. In column 3, the coefficient on is also significantly
Shares

negative, suggesting that the day-1 impacts from loans with longer loss emergence periods are

incrementally value relevant compared to IL allowances, consistent with columns 1 and 2.

A potential implication of the findings in Tables 6 and 7 is that IFRS 9 may omit relevant

expected credit loss information by not recognizing losses expected to emerge beyond one year.

Furthermore, researchers should be cautious in drawing inferences about CECL based on IFRS 9

evidence.

4.4. The Decision-Usefulness of Separate Reporting Versus Reporting Single CECL Amount

Table 8, Panel A presents the results of estimating equation 1 using three different models,

where the inclusion of the IL and CECL allowance variables is varied. As a benchmark, in column

1, only IL_Allowance is included in the model. Column 2 includes IL_Allowance and

CECL_Impact as separate variables, whereas column 3 includes only CECL_Allowance.

Comparing columns 1 and 3, we find that CECL allowances have relatively more information

content than IL allowances. Specifically, CECL allowances better predict losses than IL

24
allowances for approximately 76 percent of banks (Clarke test p-value < 0.01). Comparing

columns 2 and 3, we find that CECL allowances have relatively less information content than a

reporting regime where incurred losses and expected future credit losses are presented separately.

Specifically, separate reporting of credit losses better predicts future net charge-offs for

approximately 60 percent of banks (Clarke test p-value < 0.01).21

One potential reason separate reporting, compared to reporting a single amount, better

predicts future loan losses is that IL allowances and incremental CECL allowances could have

different associations with future loan losses, particularly as they relate to the loss emergence

period. To examine this, we re-estimate equation 1 but consider four different outcome variables:

net charge-offs cumulated over the future (1) four, (2) eight, (3) twelve, and (4) sixteen quarters.

We estimate these models in the seemingly unrelated regressions framework (Zellner 1962) and

conduct two tests: whether the coefficients on IL_Allowance and CECL_Impact are statistically

different and whether the difference between the two coefficients increases as we extend the

horizon over which we measure loan losses.

These results are presented in Panel B of Table 8. In all four columns, we find that the

coefficients on IL_Allowance and CECL_Impact are significantly different. Furthermore, the

difference in the coefficients is generally larger when we measure loan losses over longer horizons

(i.e., between eight and sixteen quarters in columns 2–4) compared to when we measure loan losses

over a four-quarter horizon (column 1).22 Taken together, the results in Panel B of Table 8 are

21
In these and similar comparisons of relative explanatory power, we test the benefits of reporting separate credit loss
amounts relative to reporting a single CECL allowance. Because both were feasible choices standard setters could
have made, we do not penalize or adjust for differences in the number of variables between models.
22
We note that the comparison of the differences between columns 1 and 4 is not statistically significant (p-value
0.194). This is likely attributable to lower power, as the estimated magnitude of the difference is in between those
estimated in columns 2 and 3 (which are statistically different compared to column 1).

25
consistent with IL allowances and incremental CECL allowances having different associations

with future loan losses, particularly as the loss emergence period is extended.

Table 9 presents the results of estimating equation 2 using three different models,

analogous to those in Panel A of Table 8. Comparing columns 1 and 3, we find that CECL

allowances have relatively more information content than IL allowances. Specifically, CECL

allowances better predict price than IL allowances for approximately 77 percent of banks (Clarke

test p-value < 0.01). Comparing columns 2 and 3, we find that CECL allowances have relatively

less information content than a reporting regime where incurred losses and expected future credit

losses are presented separately. Specifically, separate reporting of credit losses better predicts price

for approximately 74 percent of banks (Clarke test p-value < 0.01).

Taken together, the results in Tables 8 and 9 suggest that decision-useful information is

lost when IL allowances are replaced with CECL allowances, rather than separately reporting the

two amounts. This is potentially due to the different information CECL day-1 impacts and IL

allowances provide regarding the timing of future expected credit loss emergence.

5. Conclusion

Motivated by the significant debates surrounding the terms and implementation of CECL

and IFRS 9, we examine four aspects of CECL that have not been addressed by the prior literature:

whether (1) credit loss information under CECL is only decision-useful for larger banks, (2) CECL

allowances provide new credit loss information, especially for larger banks, (3) credit losses

expected to emerge beyond one year are relevant for investors, and (4) separately reporting IL and

CECL allowance amounts would have relatively more information content than reporting a single

CECL allowance amount.

26
 We document four novel findings. First, contrary to criticisms raised by community banks

and the expectations of the prior literature, we document that CECL provides decision-useful

information regardless of bank size. Second, we show that CECL allowances provide new

information to investors, suggesting that CECL does not merely confirm investor expectations.

Furthermore, CECL only provides new information for smaller banks, suggesting that the market

reacts to CECL allowance information only when other information about expected credit losses

is less available. Third, we find that CECL allowances are decision-useful regardless of the

expected loss emergence period, which suggests that IFRS 9’s divergence from CECL may omit

relevant information from credit loss allowances, highlighting the important differences between

the two standards. Fourth, we find that allowances under CECL are relatively less decision-useful

than a regime that reports incurred losses and expected future credit losses separately, potentially

due to the different information CECL day-1 impacts and IL allowances provide regarding the

timing of future expected credit loss emergence.

While our focus on day-1 CECL implementation allows us to compare CECL and IL

allowances for the same bank at the same time, such that differences between these allowances

cannot be attributed to different banks or periods, we note that this strong identification precludes

us from examining implications beyond CECL’s adoption date, which future research can address.

27
 References

Aboody, D., M. E. Barth, and R. Kasznik. 2004. SFAS No. 123 stock‐based compensation expense and
equity market values. The Accounting Review 79 (2): 251–275.
Aboody, D., M. E. Barth, and R. Kasznik. 2006. Do firms understate stock option-based compensation
expense disclosed under SFAS 123? Review of Accounting Studies 11: 429–461.
Bable, J., Wong, C., Wynes, M., 2023. The decision usefulness of CECL: Users’ views about the current
expected credit losses model. Working paper. Available at
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4483117
Barth, M. E., and G. Clinch. 2009. Scale effects in capital markets-based accounting research. Journal of
Business Finance and Accounting 36: 253–288.
Barth, M. E., I. D. Gow, and D. J. Taylor. 2012. Why do pro forma and Street earnings not reflect changes
in GAAP? Evidence from SFAS 123R. Review of Accounting Studies 17: 526–562.
Barth, M. E., K. Li, and C. McClure. 2023. Evolution in value relevance of accounting information. The
Accounting Review 98 (1): 1–28.
Basel Committee on Banking Supervision (BCBS), 2011. Comment letter: financial instruments:
impairment. Available at https://www.bis.org/bcbs/commentletters/iasb32.pdf.
Basel Committee on Banking Supervision (BCBS). 2021. The procyclicality of loan loss provisions: a
literature review. Basel Committee on Banking Supervision working paper. Available at
https://www.bis.org/bcbs/publ/wp39.pdf.
Beatty, A., and S. Liao. 2011. Do delays in expected loss recognition affect banks’ willingness to lend?
Journal of Accounting and Economics 52 (1): 1–20.
Beatty, A., and S. Liao. 2021. What do analysts’ provision forecasts tell us about expected credit loss
recognition? The Accounting Review 96 (1): 1–21.
Bernanke, B., 2009. Financial reform to address systemic risk. Remarks at the Council on Foreign Relations,
March 10. Washington, DC. Available at
https://www.federalreserve.gov/newsevents/speech/bernanke20090310a.htm.
Berger, P.G., and R. Hann. 2003. The impact of SFAS No. 131 on information and monitoring. Journal of
Accounting Research 41: 163–223.
Biddle, G. C., G. S. Seow, and A. F. Siegel. 1995. Relative versus incremental information content.
Contemporary Accounting Research 12 (1): 1–23.
Bischof, J., C. Laux, C. Leuz. 2021. Accounting for financial stability: Bank disclosure and loss recognition
in the financial crisis. Journal of Financial Economics 141 (3): 1188–1217.
Bonsall, S., B. Schmidt, and B. Xie. 2025. Current Expected Credit Loss (CECL) Model and Banks’
Information Environment. Working paper. Available at
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4236547.
Botosan, C. A., and M. Stanford. 2005. Managers’ motives to withhold segment disclosures and the effect
of SFAS No. 131 on analysts’ information environment. The Accounting Review 80 (3): 751–772.

28
Bushman, R. M., C. D. Williams. 2012. Accounting discretion, loan loss provisioning, and discipline of
banks’ risk-taking. Journal of Accounting and Economics 54 (1):1–18.
Bushman, R. M., C. D. Williams. 2015. Delayed expected loss recognition and the risk profile of banks.
Journal of Accounting Research 53: 511–553.
Campbell, J. L., K. H. Gee, and Z. Wiebe. 2022. The determinants and informativeness of non-GAAP
revenue disclosures. The Accounting Review 97 (7): 23–48.
CFA Institute. 2013. Comment Letter on Financial Instruments: Expected Credit Losses (Impairments).
Available at
https://www.fasb.org/page/ShowPdf?path=AFIIMP.ED.0355.CFA%20INSTITUTE%20CDPC%20S
ANDRA%20J.%20PETERS%20AND%20ASHWINPAUL%20C.%20SONDHI.pdf.
CFA Institute. 2019. Cumulative Expected Credit Loss (“CECL”) Impairment Model. Available at
https://www.fasb.org/page/ShowPdf?path=CL.UNS.001.CFA%20INSTITUTE%20SANDRA%20J.
%20PETERS,0.pdf.
Chen, J., Y. Dou, Y. Zou. 2021. Information externalities of SFAS 161: Evidence from supply chains. The
Accounting Review 96 (4): 179–202.
Chen, J., Y. Dou, S.G. Ryan, and Y. Zou. 2025. The Effect of the current expected credit loss approach on
banks’ lending during stress periods: Evidence from the COVID-19 recession. The Accounting Review
100 (1): 113–138.
Choudhary, P., S. Rajgopal, and M. Venkatachalam. 2009. Accelerated vesting of employee stock options
in anticipation of FAS 123-R. Journal of Accounting Research 47 (1): 105–146.
Clarke, K. A. 2003. Nonparametric model discrimination in international relations. Journal of Conflict
Resolution 47 (1): 72–93.
Clarke, K. A. 2007. A simple distribution-free test for nonnested model selection. Political Analysis 15:
347–363.
Covas, F., and W. Nelson. 2018. Current expected credit loss: Lessons from 2007-2009. Working paper.
Available at https://ssrn.com/abstract=3215559.
Dugan, J.C. 2009. Loan loss provisioning and procyclicality. Remarks by J. C. Dugan, Comptroller of the
Currency, before the Institute of International Bankers. March 2, Washington, DC.
Durney, M., Gee, K., Wiebe, Z., 2024. Segment profit/loss and the limitations of a “management approach.”
Management Science, forthcoming.
Financial Accounting Standards Board (FASB). 2016. FASB Accounting standards updates No. 2016-13.
June 2016. Financial Instruments—Credit Losses (Topic 326) Measurement of Credit Losses on
Financial Instruments.
Financial Stability Forum. 2009. Report of the financial stability forum on addressing procyclicality in the
financial system. April 2. Basel, Switzerland. Available at https://www.fsb.org/wp-
content/uploads/r_0904a.pdf.
G20 (Group of Twenty). 2009. Declaration on Strengthening the Financial System – London Summit.
Declaration.

29
Harris, T.S., U. Khan, and D. Nissim. 2018. The expected rate of credit losses on banks’ loan portfolios.
The Accounting Review 93 (5): 245–271.
Independent Community Bankers of America (ICBA). 2013. Comment Letter on the CECL standard.
Available at
https://www.fasb.org/Page/ShowPdf?path=AFIIMP.ED.0025A.ICBA%20JAMES%20KENDRICK.
pdf&title=AFIIMP.ED.0025A.ICBA%20JAMES%20KENDRICK.
Kim, S., S. Kim, A.V. Kleymenova, and R. Li. 2023. Current Expected Credit Losses (CECL) Standard
and Banks’ Information Production. Working paper. Available at
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4117869.
Lejard, C., E. Paget-Blanc, and J.-F. Casta. 2021. The effects of the adoption of IFRS 9 on the comparability
and the predictive ability of banks’ loan loss allowances. Available at
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3782584.
López-Espinosa, G., G. Ormazabal, and Y. Sakasai. 2021. Switching from incurred to expected loan loss
provisioning: Early evidence. Journal of Accounting Research 59: 757–804.
Lu, Y., and V. Nikolaev. 2022. Expected loan loss provisioning: An empirical model. The Accounting
Review 97 (7): 319–346.
McInnis, J.M., Y. Yu, and C. G. Yust. 2018. Does fair value accounting provide more useful financial
statements than current GAAP for banks? The Accounting Review 93 (6): 257–279.
Nicoletti, A. 2018. The effects of bank regulators and external auditors on loan loss provisions. Journal of
Accounting and Economics 66 (1): 244–265.
Onali, E., G. Ginesti, G. Cardillo, and G. Torluccio. 2024. Market reaction to the expected loss model in
banks. Journal of Financial Stability 74: 100884.
Schmidt, B. 2024. Consequences of bank stress test disclosures. Working paper. Available at
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4649867.
Vuong, Q. H. 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica
57 (2): 307–333.
Wheeler, B. 2021. Unrecognized expected credit losses and bank share prices. Journal of Accounting
Research 59: 805–866.

Zellner, A. 1962. An efficient method of estimating seemingly unrelated regressions and tests for
aggregation bias. Journal of the American Statistical Association 57 (298): 348–368.

30
 Appendix: Variable Definitions (in Alphabetic Order)23

Variable Definition
The number of analysts contributing to the I/B/E/S consensus street
Analysts
forecast that is used in calculating unexpected earnings (UE).
Assets Total assets (SPCIQ Keyfield 280297).
Book value of equity excluding minority interest (SPCIQ Keyfield
BVE_Adjusted 280318) plus the allowance for loan and lease losses under the incurred
loss model (SPCIQ Keyfield 280287).
CECL_Allowance The total CECL allowance, calculated as IL_Allowance + CECL_Impact.
The effect of day-1 CECL adoption on allowances for credit losses on
loans and leases held for investment and held-to-maturity debt securities
CECL_Impact
(SPCIQ Keyfield 319096), including the initial allowance gross-up for
any purchased credit-deteriorated assets held as of the adoption date.
ConsLoans Total consumer loans outstanding (SPCIQ Keyfield 290161).
The allowance for loan and lease losses under the incurred loss model
IL_Allowance
(SPCIQ Keyfield 280287).
Market-to-Book Market value of equity scaled by book value of equity.
The market value of equity in $ millions. Market value of equity is
calculated as price per share from CRSP (abs(prc)/cfacpr) multiplied by
MVE
the number of shares outstanding from CRSP (shrout *cfacshr), divided
by 1,000.
Net charge-offs scaled by total assets cumulated over the next sixteen
quarters. Net charge-offs are equal to gross charge-offs (BHCK/RIAD
NCOt+1:16
4635) minus recoveries (BHCK/RIAD 4605). This variable is obtained
from regulatory reports (i.e., Y-9Cs and call reports).
Net income before taxes and loan loss provision, calculated as net income
NIBP before taxes (SPCIQ Keyfield 280344) plus loan loss provision (SPCIQ
Keyfield 280330).
The CECL day-1 impact for non-real estate loans, hand collected from
NonRE_CECL_Impact
banks’ first 10-Q filing after adoption.
Nonperforming loans, calculated as nonaccrual loans (SPCIQ Keyfield
NPL 281530) plus loans past due 90 days or more but still accruing (SPCIQ
Keyfield 281489).
Indicator equal to 1 if the bank’s CECL SAB 74 estimate is a point
Point Estimate
forecast, and zero otherwise.
The bank’s cumulative return in the [-50, -5] trading-day window prior
Pre-EA Return
to the earnings announcement.
Price Stock price per share from Compustat (prccq).
Provision Loan loss provision (SPCIQ Keyfield 280330).

23
With one exception, we extract bank financial data from the “Companies” dataset in S&P Capital IQ Pro, which
provides better coverage of the variables we need than the “Regulated Depositories (U.S.)” dataset. As a result, our
data item numbers (called “Keyfields”) for the bank financial variables often differ from those of other studies that
pull bank financial data from the “Regulated Depositories (U.S.)” dataset (e.g., Wheeler 2021). For example, the
Keyfield for total assets (the allowance for loan and lease losses under the incurred loss model) is Keyfield 280297
(280287) in the “Companies” dataset, and thus is what we use in our study, but is Keyfield 215382 (215372) in
Wheeler (2021). The one exception is that we extract the CECL day-1 impact variable (SPCIQ Keyfield 319096) from
the “Regulated Depositories (U.S.)” dataset because that variable is not available in the “Companies” dataset.

31
 Variable Definition
Rate-sensitive assets maturing within one year (SPCIQ Keyfield
RateSensitive
280515).
The CECL day-1 impact for real estate loans, hand collected from banks’
RE_CECL_Impact
first 10-Q filing after adoption.
RELoans Total real estate loans outstanding (SPCIQ Keyfield 290155).
Abnormal stock return around a bank’s disclosure of its estimates of
expected day-1 CECL impact as required by SAB 74, calculated over two
different windows: (1) beginning on the day of the earnings
announcement through one day after the bank’s 10-K/Q filing and (2)
Return
beginning on the day of the earnings announcement through one day after
the earnings announcement. Abnormal return is calculated as the bank’s
daily return minus the daily value-weighted market return, cumulated
using a buy-and-hold (B&H) or cumulative (CAR) approach.
Change in the bank’s estimate of the day-1 impact of CECL on its credit
loss allowances, calculated as the CECL SAB 74 estimate for quarter t
ΔSAB74_Estimate minus the most recently disclosed CECL SAB 74 estimate (with the final
quarter’s change calculated using the actual recognized CECL day-1
impact).
Number of shares outstanding from Compustat, in thousands
Shares
(cshoq*1,000).
Unexpected earnings, measured as I/B/E/S street earnings minus the
UE most recent mean consensus street EPS forecast from I/B/E/S, scaled by
stock price at fiscal quarter-end.

32
 Figure 1 – Timeline of CECL SAB 74 forecast revisions

Q2 2019 10-Q Q3 2019 10-Q Q4 2019 10-K Q1 2020 10-Q

10 forecast revisions 17 revisions 63 revisions 154* recognize CECL impacts
(All 10 are initial revisions) (7 initial revisions) (46 initial revisions) (95 initial revisions)

This figure shows the timeline and frequency of initial and subsequent CECL SAB 74 forecast revisions, which were collected from banks’ SEC filings. Forecast
revisions are the difference between the forecasted amount in quarter t and the forecasted amount in quarter t-1(e.g., Q2 2019 revisions are relative to the Q1 2019
forecast). *Although there are 158 unique banks in this sample, four banks did not have a usable final revision due to mergers, data availability, etc.

33
Table 1 – Sample Selection
Part I:
Sample Selection for the Analyses on Decision-Usefulness (Tables 3-4 and 6-9)
Number of unique publicly traded banks in SNL (SIC codes 6000-6299) with an incurred-
loss allowance balance and coverage in CRSP 402
Less: CECL non-adopters 190
Number of unique publicly traded banks that adopted CECL 212
Less: Banks with missing data for variables required in the regressions 12
Sample of banks used for Tables 4, 7, and 9 200
Less: Banks that were acquired or are otherwise missing 16-quarters-ahead NCO 26
Sample of banks used for Tables 3, 6, and 8 174

Part II:
Sample Selection for the Analyses on SAB 74 Investor Response (Table 5)
Number of unique publicly traded banks that adopted CECL 212
Less: Banks with less than two disclosed SAB 74 CECL estimates 37
Less: Banks missing data for variables required in the regressions 17
Sample of banks used for Table 5 158

Sample of quarterly observations for 158 banks represented in Table 5 244

Part I of this table presents the sample selection procedures for the sample used to explore whether the decision-
usefulness of the CECL day-1 impact varies across bank size, the expected loss emergence period, and combined
versus separate reporting of the CECL day-1 impact and the IL allowance. Part II presents the sample selection
procedures for the analyses that explore whether CECL allowances represent new information to investors (i.e., the
SAB 74 investor response analyses).

34
Table 2 – Descriptive Statistics
Panel A: Descriptive statistics for the sample used in decision-usefulness analyses (Part I tests)
Variable Name N Mean p25 p50 p75 SD
Dependent Variables
Σ(NCO/Assets)t+1:16 174 0.005 0.001 0.003 0.005 0.010
Pricet+2 200 34.667 16.795 25.420 39.780 31.456

Independent Variables
IL_Allowance/Assets 174 0.007 0.005 0.006 0.008 0.004
CECL_Impact/Assets 174 0.002 0.000 0.001 0.003 0.003
CECL_Allowance/Assets 174 0.009 0.006 0.008 0.010 0.006
RELoans/Assets 174 0.453 0.374 0.482 0.578 0.183
ConsLoans/Assets 174 0.054 0.005 0.015 0.061 0.110
RateSensitive/Assets 174 0.363 0.262 0.365 0.433 0.135
NIBP/Assets 174 0.004 0.004 0.004 0.005 0.002
CECL_Impact/Shares 200 0.507 0.036 0.244 0.600 0.870
IL_Allowance/Shares 200 1.861 0.794 1.298 2.077 1.926
CECL_Allowance/Shares 200 2.399 1.064 1.676 2.558 2.665
BVE_Adjusted/Shares 200 35.529 21.314 29.979 38.434 27.051
RELoans/Shares 200 113.255 66.977 100.180 143.794 76.475
ConsLoans/Shares 200 15.392 0.917 3.540 11.663 34.729
RateSensitive/Shares 200 115.256 49.354 74.847 129.358 130.312
NIBP/Shares 200 1.286 0.644 0.960 1.521 1.231
NPL/Shares 200 1.481 0.491 0.839 1.493 2.134

35
Panel B: Descriptive statistics for the SAB 74 investor response analyses (Part II tests)
Variable Name N Mean p25 p50 p75 SD
Dependent Variables
Return[EA day 0 to 10-K/Q+1]B&H 244 -0.043 -0.097 -0.043 0.008 0.085
Return[EA day 0 to 10-K/Q+1]CAR 244 -0.042 -0.097 -0.041 0.012 0.089
Return[EA day 0 to EA day+1]B&H 160 0.000 -0.038 -0.002 0.041 0.066
Return[EA day 0 to EA day+1]CAR 160 0.000 -0.038 -0.001 0.041 0.066

Independent Variables
ΔSAB74_Estimate(x100)/MVE 244 0.042 -0.030 0.000 0.175 0.480
ln(1+Analysts) 244 2.217 1.792 2.079 2.890 0.644
UE/MVE 244 -0.010 -0.013 -0.003 0.001 0.017
ln(MVE) 244 8.054 6.590 7.662 9.102 1.876
Market-to-Book 244 1.041 0.711 0.919 1.208 0.523
Pre-EA Return 244 -0.221 -0.388 -0.256 -0.005 0.218
Point Estimate 244 0.787 1.000 1.000 1.000 0.410
Provision/MVE 244 0.021 0.003 0.012 0.027 0.030
Panel A (B) of this table presents the summary statistics for the sample for the decision-usefulness (SAB 74 investor
response) analyses. In Panel A, the number of observations for each variable differs because we use the largest sample
possible for different analyses (see Table 1). In Panel A, except for variables with subscripts (e.g., Pricet+2), the
subscript for all other variables is quarter t (as of the end of the fiscal quarter immediately before CECL adoption) and
is omitted. In Panel B, except for the dependent variables (which are abnormal stock returns measured either (1) from
the earnings announcement day to one day after the 10-Q/K filing day or (2) from the earnings announcement day to
one day after the earnings announcement day), the subscript for all other variables is quarter t, as of the end of the
fiscal quarter. See the Appendix for variable definitions.

36
Table 3 – Predictive ability of CECL day-1 impact by bank size
Assets>$10 bln Assets<$10 bln
Σ(NCO/Assets)t+1:16
VARIABLES (1) (2)
CECL_Impact/Assets 0.685*** 0.962***
(4.00) (2.76)
IL_Allowance/Assets 0.670*** 0.378*
(4.15) (1.90)
RELoans/Assets -0.003 -0.007
(-1.41) (-0.87)
ConsLoans/Assets 0.054*** 0.002
(6.96) (0.14)
RateSensitive/Assets 0.004 0.005
(1.28) (1.47)
NIBP/Assets 0.035 -0.072
(0.13) (-0.17)
Constant -0.003* 0.002
(-1.89) (0.29)

Observations 92 82
Adjusted R-squared 0.945 0.208
This table presents analyses that examine the usefulness of the CECL day-1 impact in predicting future net charge-
offs, separately for banks with total assets above versus below $10 billion. The dependent variable is net charge-offs
scaled by total assets, cumulated over the next sixteen quarters. As described in Table 1, the sample consists of 174
publicly listed U.S. banks that adopted CECL from January 1, 2020–January 1, 2021. Except for CECL_Impact (which
is measured on the adoption date), all other independent variables are measured as of the fiscal quarter-end
immediately before CECL adoption (i.e., one day before the CECL adoption date). See the Appendix for variable
definitions. All continuous variables are winsorized at the 1st and 99th percentiles before estimating the regressions.
Standard errors are clustered by bank to address heteroskedasticity. Robust t-statistics are reported in parentheses
below coefficient estimates. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

37
Table 4 – Usefulness of CECL day-1 impact in valuing stocks by bank size
Assets>$10 bln Assets<$10 bln
Pricet+2
VARIABLES (1) (2)
CECL_Impact/Shares -8.031** -6.625**
(-2.40) (-2.49)
IL_Allowance/Shares -5.528*** -0.154
(-2.76) (-0.10)
BVE_Adjusted/Shares 0.204 0.912***
(1.07) (3.30)
RELoans/Shares 0.050 -0.027
(1.29) (-0.94)
ConsLoans/Shares -0.057 0.097
(-0.61) (1.08)
RateSensitive/Shares -0.034 -0.006
(-1.09) (-0.22)
NIBP/Shares 30.204*** 13.163**
(9.18) (2.00)
NPL/Shares -0.722 1.032
(-0.69) (0.63)
Constant 2.688 -5.230**
(0.87) (-2.11)

Observations 107 93
Adjusted R-squared 0.839 0.712
This table presents the analysis of the usefulness of the CECL day-1 impact in valuing stocks, separately for banks
with total assets above versus below $10 billion. The dependent variable is two-quarter-ahead stock price. As
described in Table 1, the sample consists of 200 publicly listed U.S. banks that adopted CECL from January 1, 2020–
January 1, 2021. Except for Pricet+2 (which is measured two quarters after the CECL adoption date—after the bank’s
first 10-Q filing would be available to investors following CECL adoption) and CECL_Impact (which is measured on
the adoption date), all other variables are measured as of the fiscal quarter-end immediately before CECL adoption
(i.e., one day before the CECL adoption date). See the Appendix for variable definitions. All continuous variables are
winsorized at the 1st and 99th percentiles before estimating the regressions. Standard errors are clustered by bank to
address heteroskedasticity. Robust t-statistics are reported in parentheses below coefficient estimates. ***, **, and *
indicate significance at the 1%, 5%, and 10% levels, respectively.

38
Table 5 – Investor response to revisions to CECL SAB 74 estimates
Panel A: Pooled sample
Full Sample EA Disclosure Sample
Long-Window Returns Short-Window Returns
Buy-and-Hold CAR Buy-and-Hold CAR
VARIABLES (1) (2) (3) (4)

ΔSAB74_Estimate/MVE -1.790* -1.930* -1.979* -1.980*

(-1.67) (-1.72) (-1.80) (-1.81)
ln(1+Analysts) 0.031* 0.032* 0.048*** 0.048***
(1.95) (1.95) (3.25) (3.29)
UE/MVE 0.433 0.593 0.316 0.341
(0.65) (0.86) (0.69) (0.75)
ln(MVE) 0.003 0.002 -0.013* -0.014*
(0.43) (0.35) (-1.86) (-1.94)
Market-to-Book 0.015* 0.016* 0.013 0.013
(1.78) (1.78) (0.93) (0.94)
Pre-EA Return -0.117** -0.131** -0.087* -0.083*
(-2.05) (-2.25) (-1.93) (-1.85)
Point Estimate 0.007 0.007 0.006 0.005
(0.60) (0.60) (0.49) (0.38)
Provision/MVE 0.275 0.285 -0.306 -0.292
(0.82) (0.81) (-1.25) (-1.20)

Observations 244 244 160 160

Adjusted R-squared 0.164 0.155 0.088 0.087

39
Panel B: Sample split based on size
Assets>$10 bln Assets<$10 bln
Long-Window Returns
Buy-and-hold CAR Buy-and-hold CAR
VARIABLES (1) (2) (3) (4)
ΔSAB74_Estimate/MVE 0.165 0.134 -4.468*** -4.784***
(0.13) (0.10) (-2.81) (-2.78)
ln(1+Analysts) 0.018 0.020 0.002 0.003
(0.77) (0.84) (0.10) (0.11)
UE/MVE -0.644 -0.529 2.307** 2.647***
(-0.93) (-0.74) (2.45) (2.65)
ln(MVE) 0.001 0.000 0.011 0.011
(0.07) (0.01) (0.62) (0.55)
Market-to-Book 0.016** 0.016* 0.047* 0.050
(1.96) (1.95) (1.67) (1.62)
Pre-EA Return -0.102 -0.115 -0.176** -0.193**
(-1.28) (-1.43) (-2.15) (-2.20)
Point Estimate 0.004 0.004 0.029* 0.030*
(0.26) (0.26) (1.79) (1.72)
Provision/MVE 0.206 0.203 0.208 0.308
(0.68) (0.64) (0.32) (0.45)

Observations 158 158 86 86

Adjusted R-squared 0.173 0.164 0.194 0.190
This table examines investor response to revisions to banks’ CECL estimates leading up to adoption, with Panel A
presenting the results using the pooled sample and Panel B presenting the results splitting the sample based on bank
size. The dependent variable is the bank’s abnormal stock return around its disclosure of its revision to its expected
CECL day-one impact estimate as required by SAB 74, calculated over two different windows: (1) beginning on the
day of the earnings announcement through one day after the bank’s 10-K/Q filing (columns 1 and 2 in Panel A and
all columns in Panel B) and (2) beginning on the day of the earnings announcement through one day after the earnings
announcement (columns 3 and 4 in Panel A). Abnormal return is calculated as the bank’s daily return minus the daily
value-weighted market return, cumulated alternatively using a buy-and-hold or cumulative (CAR) approach. The test
variable is ΔSAB74_Estimate/MVE, which is the revision to a bank’s CECL day-one impact estimate, scaled by market
value of equity. See the Appendix for variable definitions. All continuous variables are winsorized at the 1st and 99th
percentiles before estimating the regressions. Quarter fixed effects are included in all regressions. Standard errors are
clustered by bank. Robust t-statistics are reported in parentheses below coefficient estimates. ***, **, and * indicate
significance at the 1%, 5%, and 10% levels, respectively.

40
Table 6 – Predictive ability of CECL day-1 impact by loss emergence period
Shorter loss emergence Longer loss emergence
period period
Σ(NCO/Assets)t+1:16
VARIABLES (1) (2) (3)
CECL_Impact/Assets 1.229*** 0.488**
(4.89) (2.63)

RE_CECL_Impact/Assets 0.508***
(2.72)
NonRE_CECL_Impact/Assets 1.499***
(6.19)
IL_Allowance/Assets 0.505* 0.383* 0.489***
(1.89) (1.84) (3.31)
RELoans/Assets -0.009* 0.002 -0.004
(-1.91) (0.81) (-1.23)
ConsLoans/Assets 0.035*** 0.003 0.033***
(3.44) (0.38) (4.07)
RateSensitive/Assets 0.007* 0.005 0.004
(1.73) (1.51) (1.54)
NIBP/Assets 0.565* 0.140 0.360
(1.80) (0.60) (1.47)
Constant -0.004 -0.004 -0.004*
(-1.57) (-1.51) (-1.66)

Observations 87 87 166
Adjusted R-squared 0.888 0.131 0.869

This table presents analyses that examine the usefulness of the CECL day-1 impact in predicting future net charge-
offs, separately based on the expected loss emergence period. The dependent variable is net charge-offs scaled by total
assets, cumulated over the next sixteen quarters. In columns 1–2, we partition the sample into banks with below
(column 1) versus above (column 2) the median of real estate loans with no evidence of significant deterioration in
credit quality as a proportion of total loans. In column 3, we replace CECL_Impact with RE_CECL_Impact and
NonRE_CECL_Impact, which are defined as the day-1 impact specific to real estate and non-real estate loans,
respectively, obtained from banks’ first 10-Q filing after adoption. As described in Table 1, the sample consists of 174
publicly listed U.S. banks that adopted CECL from January 1, 2020–January 1, 2021, with 8 banks dropping from the
analyses due to a lack of data on the day-1 impact by loan type in column 3. Except for the CECL day-1 impact
variables (which are measured on the adoption date), all other independent variables are measured as of the quarter-
end immediately before CECL adoption (i.e., one day before the CECL adoption date). See the Appendix for variable
definitions. All continuous variables are winsorized at the 1st and 99th percentiles before estimating the regressions.
Standard errors are clustered by bank to address heteroskedasticity. Robust t-statistics are reported in parentheses
below coefficient estimates. ***, **, and * indicate significance at the 1%, 5%, and 10% levels, respectively.

41
Table 7 – Usefulness of CECL day-1 impact in valuing stocks by loss emergence period
Shorter loss emergence Longer loss emergence
period period
Pricet+2
VARIABLES (1) (2) (3)
CECL_Impact/Shares -9.970** -5.620*
(-2.39) (-1.75)

RE_CECL_Impact/Shares -7.029**
(-2.26)
NonRE_CECL_Impact/Shares -11.198***
(-4.14)
IL_Allowance/Shares -4.448*** -0.916 -2.433*
(-2.75) (-0.57) (-1.90)
BVE_Adjusted/Shares 0.240 0.433** 0.305**
(1.13) (2.31) (2.22)
RELoans/Shares -0.006 0.048 -0.001
(-0.17) (0.94) (-0.04)
ConsLoans/Shares -0.039 0.011 -0.073
(-0.37) (0.06) (-0.97)
RateSensitive/Shares -0.047* -0.009 -0.059**
(-1.68) (-0.26) (-2.42)
NIBP/Shares 27.111*** 17.897*** 28.355***
(6.66) (3.28) (8.42)
NPL/Shares 0.852 -1.516** 0.210
(0.64) (-2.29) (0.18)
Constant 11.028*** -2.416 4.118**
(2.78) (-1.18) (2.11)

Observations 100 100 189

Adjusted R-squared 0.749 0.907 0.832

This table presents the analysis of the usefulness of the CECL day-1 impact in valuing stocks, separately based on the
expected loss emergence period. The dependent variable is two-quarter-ahead stock price. In columns 1–2, we
partition the sample into banks with below (column 1) versus above (column 2) the median of real estate loans with
no evidence of significant deterioration in credit quality as a proportion of total loans. In column 3, we replace
CECL_Impact with RE_CECL_Impact and NonRE_CECL_Impact, which are defined as the day-1 impact specific to
real estate and non-real estate loans, respectively, obtained from banks’ first 10-Q filing after adoption. As described
in Table 1, the sample consists of 200 publicly listed U.S. banks that adopted CECL from January 1, 2020–January 1,
2021, with 11 banks dropping from the analyses due to a lack of data on the day-1 impact by loan type in column 3.
Except for Pricet+2 (which is measured two quarters after the CECL adoption date—after the bank’s first 10-Q filing
would be available to investors following CECL adoption) and the CECL day-1 impact variables (which are measured
on the adoption date), all variables are measured as of the quarter-end immediately before CECL adoption (i.e., one
day before the CECL adoption date). See the Appendix for variable definitions. All continuous variables are
winsorized at the 1st and 99th percentiles before estimating the regressions. Standard errors are clustered by bank to
address heteroskedasticity. Robust t-statistics are reported in parentheses below coefficient estimates. ***, **, and *
indicate significance at the 1%, 5%, and 10% levels, respectively.

42
Table 8 – Predictive ability of CECL day-1 impact by reporting regime
Panel A: Comparing separate reporting of IL allowance and CECL day-1 impact to single CECL allowance
Σ(NCO/Assets)t+1:16
VARIABLES (1) (2) (3)
CECL_Impact/Assets 1.058***
(5.91)
IL_Allowance/Assets 0.707*** 0.583***
(3.66) (3.54)
CECL_Allowance/Assets 0.758***
(5.62)
RELoans/Assets -0.002 -0.004 -0.004
(-0.63) (-1.36) (-1.33)
ConsLoans/Assets 0.055*** 0.037*** 0.040***
(6.54) (4.74) (4.63)
RateSensitive/Assets 0.005 0.004 0.004
(1.44) (1.51) (1.48)
NIBP/Assets 0.575** 0.531** 0.439**
(2.13) (2.45) (2.14)
Constant -0.006** -0.005** -0.005**
(-2.52) (-2.31) (-2.34)

Observations 174 174 174

Adjusted R-squared 0.809 0.846 0.843

43
Panel B: Difference between CECL day-1 impact and IL allowance by loss emergence period
Σ(NCO/Assets)
Future loan loss horizon: [t+1:t+4] [t+1:t+8] [t+1:t+12] [t+1:t+16]
VARIABLES (1) (2) (3) (4)
(1) CECL_Impact/Assets 0.427*** 0.704*** 0.942*** 1.058***
(4.10) (4.17) (4.07) (6.01)
(2) IL_Allowance/Assets 0.188*** 0.303*** 0.432*** 0.583***
(3.36) (3.33) (3.29) (3.60)
(3) RELoans/Assets -0.002 -0.002 -0.004 -0.004
(-1.48) (-1.42) (-1.59) (-1.39)
(4) ConsLoans/Assets 0.014*** 0.020*** 0.029*** 0.037***
(4.34) (3.84) (4.08) (4.82)
(5) RateSensitive/Assets 0.002 0.003 0.003 0.004
(1.57) (1.46) (1.00) (1.54)
(6) NIBP/Assets 0.138 0.294 0.471** 0.531**
(1.04) (1.60) (2.00) (2.49)
(7) Constant -0.002* -0.003** -0.004** -0.005**
(-1.84) (-2.04) (-1.98) (-2.35)

2 test: (1) – (2) = 0 0.239** 0.401** 0.510** 0.475**

(5.92) (6.25) (5.22) (4.11)

2 test: Col. (2) Diff. – 0.162**

Col. (1) Diff. = 0 (5.08)
 test: Col. (3) Diff. –
2 0.271**
Col. (1) Diff. = 0 (3.87)
 test: Col. (4) Diff. –
2 0.236
Col. (1) Diff. = 0 (1.68)

Observations 174 174 174 174

Adjusted R-squared 0.748 0.766 0.796 0.846

This table presents analyses that examine the usefulness of the CECL day-1 impact in predicting future net charge-
offs, separately based on the reporting regime. Panel A presents the results of tests comparing the separate reporting
of the IL allowance and CECL day-1 impact to the reporting of a single CECL allowance amount. Panel B compares
the difference between the IL allowance and the CECL day-1 impact’s relationship with future loan losses, where
future losses are calculated over different horizons. The dependent variable is net charge-offs scaled by total assets,
cumulated over the next sixteen quarters in Panel A and over the next four, eight, twelve, and sixteen quarters in
columns 1–4, respectively, in Panel B. As described in Table 1, the sample consists of 174 publicly listed U.S. banks
that adopted CECL from January 1, 2020–January 1, 2021. Except for the CECL day-1 impact/allowance variables
(which are measured on the adoption date), all other independent variables are measured as of the quarter-end
immediately before CECL adoption (i.e., one day before the CECL adoption date). See the Appendix for variable
definitions. All continuous variables are winsorized at the 1st and 99th percentiles before estimating the regressions.
Standard errors are clustered by bank to address heteroskedasticity. Except at the bottom of Panel B, robust t-statistics
are reported in parentheses below coefficient estimates (2 statistics are reported at the bottom of Panel B). ***, **,
and * indicate significance at the 1%, 5%, and 10% levels, respectively.

44
Table 9 – Usefulness of CECL day-1 impact in valuing stocks by reporting regime
Pricet+2
VARIABLES (1) (2) (3)
CECL_Impact/Shares -9.525***
(-3.54)
IL_Allowance/Shares -4.838*** -3.811***
(-2.63) (-2.63)
CECL_Allowance/Shares -5.015***
(-3.43)
BVE_Adjusted/Shares 0.256 0.328* 0.261
(1.34) (1.91) (1.46)
RELoans/Shares 0.055* 0.031 0.053*
(1.74) (1.15) (1.81)
ConsLoans/Shares -0.112 -0.020 -0.029
(-1.43) (-0.25) (-0.32)
RateSensitive/Shares 0.000 -0.037 -0.011
(0.01) (-1.35) (-0.34)
NIBP/Shares 24.166*** 25.491*** 25.308***
(5.43) (6.73) (6.31)
NPL/Shares -1.080 -0.162 -0.471
(-1.15) (-0.15) (-0.43)
Constant 0.555 3.357 1.277
(0.20) (1.50) (0.45)

Observations 200 200 200

Adjusted R-squared 0.772 0.804 0.787

This table presents the analysis of the usefulness of the CECL day-1 impact in valuing stocks, separately based on the
reporting regime. The dependent variable is two-quarter-ahead stock price. As described in Table 1, the sample
consists of 200 publicly listed U.S. banks that adopted CECL from January 1, 2020–January 1, 2021. Except for
Pricet+2 (which is measured two quarters after the CECL adoption date—after the bank’s first 10-Q filing would be
available to investors following CECL adoption) and the CECL day-1 impact/allowance variables (which are
measured on the adoption date), all variables are measured as of the quarter-end immediately before CECL adoption
(i.e., one day before the CECL adoption date). See the Appendix for variable definitions. All continuous variables are
winsorized at the 1st and 99th percentiles before estimating the regressions. Standard errors are clustered by bank to
address heteroskedasticity. Robust t-statistics are reported in parentheses below coefficient estimates. ***, **, and *
indicate significance at the 1%, 5%, and 10% levels, respectively.

45
 Estimating Lifetime Expected Credit Losses Under
IFRS 9
Xin Xu∗
This version: August 20, 2016

Abstract

We present an estimation framework of lifetime expected credit losses in accor-

dance with IFRS 9. Rooted in the literature of estimating multi-period default prob-
ability, the framework rests on a rigorous definition of “term structure of default
probability” and conditional expectation given forward-looking economic dynam-
ics. It is easy to implement and allows banks to adopt simplified and sophisticated
modeling strategies alike. We consider numerous modeling strategies within this
framework, and demonstrate examples of implementation.

JEL Codes: C41, G21, G32, M48.

Key Words: IFRS 9, Lifetime Expected Credit Losses, Term Structure of Probability of De-
fault, Multi-period Expected Credit Losses

∗
Email: [email protected].

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=2758513
https://ssrn.com/abstract=2758513
1 Introduction

In 2014, the International Accounting Standards Board (IASB) issued the completed ver-
sion of International Financial Reporting Standard 9 (IFRS 9). Under the impairment require-
ments of IFRS 9, banks shall measure the loss allowance equal to lifetime expected credit losses
(ECL) for all financial instruments for which there have been significant increases in credit risk
since initial recognition; see, IASB [2014, 5.5.3–5.5.4]. Unlike a related quantity, single-period
ECL whose measurement is well-accepted (see, for example, Chava, Stefanescu and Turnbull
[2011]),1 the concept and modeling practices of lifetime ECL is relatively new to the banking
industry, and may cause confusions. For example, it is generally unjustified to adopt naive
generalizations of the formula of single-period ECL to calculate lifetime ECL.
In the credit risk literature, the studies on the multi-period Probability of Default (PD) are
well-known and can be dated back from, at least, Duffie, Saita and Wang [2007] (henceforth,
DSW), Campbell, Hilscher and Szilagyi [2008]. Built on the foundation of this literature, we
propose a unified estimation framework of lifetime ECL, which is in accordance with FRS 9
and can also incorporate single-period ECL as a special case. In particular, this framework rests
on the rigorous definition of “term structure of PD” and conditional expectation given forward-
looking economic dynamics. It also allows banks to adopt, within this framework, simplified
and sophisticated modeling strategies alike. We consider a number of popular modeling strate-
gies within this context.
While this framework is theoretically justified, it is easy to implement. Based on specific
modeling strategies adopted, this framework can be implemented without additional complex-
ity. We demonstrate by illustrative examples how to use popular modeling techniques to mea-
sure lifetime ECL within the framework in straightforward manners.
We hope this paper, as one of the earliest attempts to properly measure lifetime ECL, con-
tributes to the banking industry which is currently implementing ECL to meet IFRS 9 impair-
ment requirements. To the best of our knowledge, prior publications about justified, IFRS9-
compliant methodology of measuring lifetime ECL are virtually non-existent. Hopefully this

1
Single-period ECL is measured as the present value of P D × LGD × EAD, where PD, LGD, EAD denote
Probability of Default, Loss Given Default and Exposure at Default respectively. If a period is defined as one year,
then the single-period ECL measures 12-month ECL as required by IFRS 9.

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=2758513
https://ssrn.com/abstract=2758513
paper sheds light on, and potentially initiates debates of, how to robustly estimate lifetime ECL
without undue cost and effort under IFRS 9.
Section 2 presents our proposed framework and shows how it satisfies IFRS 9 impairment
requirements. Section 3 elaborates key components of the framework and considers a num-
ber of popular modeling strategies for each component. Section 4 demonstrates hypothetical
examples to illustrate implementations of the framework in practice. Section 5 concludes.

2 A General Estimation Framework

To provide some perspectives of our estimation framework, we first cite the key require-
ments from IFRS 9 relating to lifetime ECL, with corresponding IFRS 9 paragraph numbers
included in brackets. Details of these requirements can be found in IASB [2014].

1. Credit losses is the present value of all cash shortfalls over the expected life of the fi-
nancial instrument. A cash shortfall is the difference between the cash flows that are due
to an entity in accordance with the contract and the cash flows that the entity expects to
receive. Both amount and timing of payments should be considered. [B5.5.28; Also see
B5.5.29–B5.5.35 for detailed examples]

2. Expected credit losses of a financial instrument shall reflect an unbiased and probability-
weighted amount that is determined by evaluating a range of possible outcomes. [5.5.17,
B5.5.41–B5.5.43]

3. When measuring expected credit losses, an entity need not necessarily identify every
possible scenario. However, it shall consider the risk or probability that a credit loss
occurs by reflecting the possibility that a credit loss occurs and the possibility that no
credit loss occurs, even if the possibility of a credit loss occurring is very low. [5.5.18]

4. Expected credit losses shall reflect the time value of money. In particular, they shall
be discounted to the reporting date using the effective interest rate (EIR), except for
purchased or originated credit-impaired financial assets, in which case the credit-adjusted
EIR is applied. [5.5.17, B5.5.44–B5.5.48]

Electronic copy available at: https://ssrn.com/abstract=2758513

 5. An entity shall measure expected credit losses using reasonable and supportable infor-
mation that is available without undue cost or effort at the reporting date about past
events, current conditions and forecasts of future economic conditions. [5.5.17, B5.5.49–
B5.5.54]

6. The maximum period to consider when measuring expected credit losses is the maximum
contractual period (including extension options) over which the entity is exposed to credit
risk and not a longer period, even if that longer period is consistent with business practice.
[5.5.19, B5.5.38 – B5.5.40]

We are now in the position to present and explain our estimation framework of lifetime
ECL. We denote time 0 as the reporting date, M as the time from reporting date to maturity
of a financial contract, and t ∈ [0, M ] as a time index in between.2 We introduce two random
variables, τ and η, as default time and non-default exit (henceforth, simply “exit”) time, re-
spectively. Note that the introduction of η is to facilitate embedded options of early exits (due
to pre-payments, refinancing, call options, etc.) in some financial contracts, and thus takes into
account of any uncertainty of actual maturity. Given η, M can always be treated as a fixed
number. We also denote r as the discount rate specified by IFRS 9, and Xt as a vector of state
variables at time t, t ∈ [0, M ]. Xt can be viewed as a full description of economic conditions
at time t.
In discrete-time, i.e., t = 0, 1, ..., M , lifetime ECL is

M
hX 1 i
Lifetime ECL = E LGDt × EADt × P r(t − 1 < τ ≤ t, η > t − 1)|X 0 , (1)
t=1
(1 + r)t

where P r(t − 1 < τ ≤ t, η > t − 1) denotes the probability of the outcome that there is no
default and exit prior to time t − 1 and a default during (t − 1, t], LGDt and EADt are Loss
Given Default and Exposure At Default, respectively, if such outcome is realized, and E[·|X0 ]
is the conditional expectation operator given X0 . Note that Equation (1) implicitly assumes
that defaults and exits occur at the end of each period, a typical discrete-time assumption made
in practice.
2
For revolving lending products, it is customary to set M = ∞ or to use the expected life of the contract.

Electronic copy available at: https://ssrn.com/abstract=2758513

 It may be informative to provide a continuous-time version of Equation (1). To proceed,
we assume there exists a joint probability distribution of τ and η, F (t, s) = P r(τ ≤ t, η ≤ s),
∂ 2 F (t,s)
and the corresponding probability density function (pdf), f (t, s) = ∂t∂s
. Consequently, for
continuous t and s, we have

hZ M Z M i
Lifetime ECL = E e−rt LGDt EADt f (t, s)dsdt|X0 , (2)
0 t

We note that the expression in Equation (1) (or (2)) is in accordance with IFRS 9 require-
ments summarized earlier, in point 1–6, explained as follows. To simplify notations in our
explanations, we define, using the quantity within Equation (1),

ELt = LGDt × EADt × P r(t − 1 < τ ≤ t, η > t − 1). (3)

First, ELt is the probability-weighted cash shortfalls discounted to the default time τ ∈
(t − 1, t], with probability reflecting both the possibilities that a shortfall occurs and that no
shortfall occurs. Particularly, (LGDt × EADt ) is the expected cash shortfalls given a default
occurring during (t − 1, t], as LGDt is commonly defined as the overall expected present value
(PV), at default time, of cash shortfalls occurring post-default divided by EADt . Because there
would be zero cash shortfalls for the period (t − 1, t] if no default occurs or there is an exit prior
to t − 1, ELt is the expected loss, which is the probability-weighted quantity considering both
nonzero and zero cash shortfalls, for the particular period. Therefore, ELt in Equation (3) is
consistent with the characterization of ECL by IFRS 9, in point 1–3 above.
Furthermore, because ELt represents the overall expected cash shortfalls discounted to the
ELt
default time, discounting ELt from the default time to the reporting time (time 0), (1+r)t
, meets
the requirements that the ECL shall be discounted to the reporting date; See point 4 above.
We also highlight two points relating to ELt in Equation (3). One, time t indicates when the
default occurs, not when the actual cash shortfalls occur. The cash shortfalls, whose expected
PV at default time is measured by ELt , could occur well after t. This point is consistent with
the implementation guidance emphasized by BCBS [2015]. Two, the introduction of exit time,
η, satisfies the IFRS 9 requirements summarized in point 6. Even though M is fixed, it is

Electronic copy available at: https://ssrn.com/abstract=2758513

evident that Equation (3) effectively considers expected losses only to the extent where there is
exposure to credit risk. Any exit in earlier periods contributes zero loss to ELt with a weight
equivalent to its associated probability.
PM ELt
In addition, summing up the PV of ELt for all t, i.e., t=1 (1+r)t gives an unconditional
lifetime ECL. The name comes from the fact that this expression has no reference to the current
economic conditions, X0 , and the conditional expectation. In light of requirements in point 5,
which explicitly mentions the preference to measure lifetime ECL using economic conditions,
we propose a conditional version of lifetime ECL given X0 , which is Equation (1) (or (2)).
It can also be shown, with some algebra and using LGD = 1−Recovery Rate , that Equation
(1) (or (2)) is equivalent to the difference between the EIR-discounted expected cash flows of
a default-free debt and those of a defaultable debt, which may default at any time during time
0 and M . This is indeed consistent with the IFRS-9 definition of “credit loss” as being the
difference between cash flows “that are due to an entity in accordance with the contract” and
those “that the entity expects to receive” discounted at original EIR.
Loosely speaking, there are two approaches of modeling ELt conditional on X0 in the
extant literature. One is similar to the approach adopted in Campbell, Hilscher and Szilagyi
[2008] who builds a model for each possible term from time 0 to a particular time t. Hence,
here we would need to build M models, each expressing ELt , t = 1, 2, ..., M , as a function of
X0 . Another approach is essentially the one adopted by Duffie, Saita and Wang [2007] (DSW),
which is to expand Equation (1) by the law of iterated expectations,

Lifetime ECL =
M
hX 1   i
E E LGDt EADt |F t P r(t − 1 < τ ≤ t, η > t − 1|F t )|X 0 , (4)
t=1
(1 + r)t

where Ft denotes the information set at time t, conceptually including all possible histories of
the state variables, i.e., Xt−1 , ..., X0 in discrete time and, denoting {Xt } as the process of Xt ,
the filtration of {Xt } in continuous time. The analogue to Equation (2) in continuous time is

hZ M Z M i
−rt
 
Lifetime ECL = E e E LGDt EADt |Ft f (t, s|Ft )dsdt|X0 . (5)
0 t

Electronic copy available at: https://ssrn.com/abstract=2758513

 There is further simplification of Equation (5) if we are willing to assume conditional in-
dependence of (non-default) exit time η and default time τ given Ft , the same assumption
adopted by DSW. Under this additional assumption, the joint pdf of (τ, η) can be written as
f (t, s) = g(s)h(t), where g(s) is the pdf for η with cumulative distribution function (cdf)
G(s), and h(t) is the pdf for τ with cdf H(t). Consequently, Equation (5) is simplified as

hZ M i
e−rt E LGDt EADt |Ft [1 − G(t|Ft )]h(t|Ft )dt|X0 .
 
Lifetime ECL = E (6)
0

RM
As will be shown in Section 3, P r(t−1 < τ ≤ t, η > t−1) in Equation (4) (or t
f (t, s)ds
in Equation (5), or [1 − G(t)]h(t) in Equation (6)) can be decomposed in such ways that only
a single model is required to express it as a function of Xt−1 , Xt−2 , ..., X0 .3
While both approaches have apparent advantages, they also impose challenges on the im-
plementation. In particular, the second approach requires modeling the dynamics of the entire
stochastic process {Xt }. While the first approach requires no knowledge of the dynamics of
{Xt }, it nevertheless imposes more difficulties in terms of satisfying the requirements of IFRS
9. One difficulty is that too many models may be required when M becomes large, particularly
when a bank has many different lending portfolios. This feature appears to be less desirable
based on the IFRS-9 principle of “without undue cost or effort” (see point 5 above). Another
consideration is that the first approach tends to be weak on the capacity to facilitate forecasts
of future economic conditions, a point that IFRS 9 alludes to (in point 5). It is for these rea-
sons that we choose the second approach to estimate the conditional lifetime ECL in this paper.
Hereafter, we adopt the expanded expression, Equation (4) (or Equation (6)), as our estimation
framework.
Finally, it is clear that Equation (4) becomes the well-accepted single-period ECL if M = 1
(i.e., t = 1, t − 1 = 0), noting that P r(0 < τ ≤ 1, η > 0|X0 ) = P r(τ ≤ 1|X0 ) is the
single-period PD, denoted as P D1 here. Hence, Equation (4) becomes

h 1 i
Single-period ECL = E LGD1 EAD1 P D1 |X0 . (7)
(1 + r)

3
Likewise, LGDt and EADt can also be modeled as functions of Xt−1 (or lagged values).

Electronic copy available at: https://ssrn.com/abstract=2758513

Therefore within our proposed framework, the single-period ECL in Equation (7) is a special
case, and a portion, of lifetime ECL in Equation (4).

3 Modeling Choices for Key Components of Lifetime ECL

We discuss a number of popular modeling strategies for the key components of our estima-
tion framework, Equation (4) (or Equation (6)), namely the term structure of PD, the models of
LGD and EAD, and the dynamics of the state variables Xt .

3.1 Modeling Term Structure of Default Probability

The most important component of lifetime ECL is perhaps P r(t − 1 < τ ≤ t, η > t − 1) or
[1 − G(t)]h(t). We need to specify them for all possible t. This multi-period default probability
or density is also known as the term structure of PD in the literature and industry.
We first discuss the cases in discrete time. Applying Bayes’ Rule, we can write P r(t − 1 <
τ ≤ t, η > t − 1) as

P r(t − 1 < τ ≤ t, η > t − 1) = P r(τ ≤ t|τ > t − 1, η > t − 1)P r(τ > t − 1, η > t − 1).
(8)

P r(τ > t − 1, η > t − 1) is known as the survival probability (including both non-default
and non-exit) until time t − 1. P r(τ ≤ t|τ > t − 1, η > t − 1) is the conditional probability
of default during time period (t − 1, t], and is the discrete-time version of the hazard rate
well-known to the literature of Survival Analysis. If we model this conditional probability
further conditional on the time-varying state variables (also known as “covariates” in Survival
Analysis), we have P r(τ ≤ t|τ > t − 1, η > t − 1, Xt−1 ). This expression is typically called
“point-in-time Probability of Default” (PIT PD) at time t, denoted as P Dt here. PIT PD is
a familiar concept to the banking industry, in the context of Basel internal rating based (IRB)
modeling; see BCBS [2006].4 The name comes from the fact that Xt−1 represents conditioning
information at a single point of time in the business cycle.

4
This definition of PIT PD is also applicable to P D1 , which becomes P r(τ ≤ 1|X0 ) as in Equation (7).

Electronic copy available at: https://ssrn.com/abstract=2758513

 There are numerous ways of modeling the (discrete-time) survival probability, P r(τ >
t − 1, η > t − 1). Two leading examples are Markov Chains methods with transition matrices
modeling the probability of being in different survival states prior to default, and estimation of
parametric or non-parametric survival probability as a functions of time.
We are particularly interested in expressing the survival probability using PIT PD, due to
the great popularity of PIT PD in the industry, and the fact that PIT PD models are required as
part of Equation (8) irrespective of whether they are used in the survival probability. Within
such a model design, we only require a single PIT PD model to estimate both components
of P r(t − 1 < τ ≤ t, η > t − 1) in Equation (8). Well-known methodologies to develop
PIT PD models include discrete-time Survival Analysis with time-varying covariates, credit
scoring, logit or probit regression, to name a few; See, for example, Lando [2004][Chapter 4,
5], Wooldridge [2010][Chapter 15, 22], Allison [2010].
To further proceed, we introduce some new notations. We let P Dk denote P r(τ ≤ k|τ >
k − 1, η > k − 1) and P Ek denote P r(η ≤ k|τ > k − 1, η > k − 1). It can be shown (in
Appendix A), by repeatedly applying Bayes’ rule, that survival probability is

P r(τ > t − 1, η > t − 1)

h i
= 1 − (P Dt−1 + P Et−1 ) P r(τ > t − 2, η > t − 2), (9)
t−1 h
Y i
By iteration, = 1 − (P Dk + P Ek ) . (10)
k=1

Substituting Equation (10) into Equation (8) and noting that P r(τ ≤ t|τ > t − 1, η >
t − 1) = P Dt , we obtain an expression of P r(t − 1 < τ ≤ t, η > t − 1) in terms of PIT PD,

t−1 h
Y i
P r(t − 1 < τ ≤ t, η > t) = P Dt 1 − (P Dk + P Ek ) . (11)
k=1

It is customary to parameterize P Dk as a time-homogeneous function of the state variables

Xk−1 through parameters β. Therefore, we only need to estimate a single set of parameters, β̂,
and develop a single PIT PD model, P Dk (Xk−1 ; β̂). Likewise, we may also estimate another
set of parameters γ̂ for P Ek , and develop a similar model P Ek (Xk−1 ; γ̂). If the (non-default)

Electronic copy available at: https://ssrn.com/abstract=2758513

exit rates exhibit little variation by empirical evidence, we may model P Ek as a constant (i.e.,
unconditional on Xk−1 ), P Ek = P E.5
Generally, plugging β̂, γ̂, and Xt−1 , Xt−2 , ..., X0 into Equation (11), we are able to obtain
P r(t − 1 < τ ≤ t, η > t) for arbitrary t, conditional on a realized path of {Xt }. As will be
shown in Section 3.3, we can obtain the paths of {Xt } by estimating its dynamics. Theoreti-
cally this approach is applicable whether Xt includes systematic factors like macro-economic
variables, or idiosyncratic risk factors like Loan-to-Value ratio, leverage, payment history.6
This is in fact the approach adopted by DSW who modeled the dynamics of both systematic
and idiosyncratic factors.
In practice, it may be deemed too onerous for an IFRS9-compliant bank to build dynamic
models of idiosyncratic factors. Hence, we suggest a number of simpler strategies which cir-
cumvents the requirement to model the dynamics of idiosyncratic factors. For notational conve-
nience in the following, we let the idiosyncratic risk factors be denoted as Zt , which is a subset
of Xt . The complement of Zt , i.e., the systematic risk factors or macro-economic factors, is
denoted as X−Z
t .

The key idea of our simplifications is to “integrates out” idiosyncratic factors, Zk−1 , of
P Dk , k > 1,7 resulting in the sole dependence of P Dt upon the macro-economic factors, X−Z
k−1 .

One way to implement this idea is to pool the data across different values of Zk−1 and estimate
the average partial effects (APE) of X−Z
k−1 , which are used to estimate P Dk (conditional on

survival). Another way is to explicitly take the weighted average of PD across all possible
outcomes of Zk−1 , weighted by the outcomes’ probability conditional on survival and X−Z
k−1 .

In other words, by iterated expectations, we may define P Dk , k = 2, ..., t, as

5
Equation (11) can also be used to find the expected length of lifetime for revolving products. Assuming
P Dt = p, P Et = q are constants and M = ∞, it can be shown that the expected length of survival for a
1
revolving loan is p+q .
6
The dynamics of idiosyncratic factors may also depend on (typically contemporaneous) macro-economic
variables.
7
Because Z0 are known at the reporting date, they can be used to model P D1 if available.

10

Electronic copy available at: https://ssrn.com/abstract=2758513

 h i
P Dk , EZ P r(τ ≤ k|τ > k − 1, η > k − 1, X−Z
k−1 , Zk−1 )|τ > k − 1, η > k − 1, X −Z
k−1
X h i
−Z −Z
= P r(τ ≤ k|τ > k − 1, η > k − 1, Xk−1 , Z)P r(Z|τ > k − 1, η > k − 1, Xk−1 ) ,
Z∈Zk−1

(12)

where EZ in the first line of Equation (12) denotes the conditional expectation operator with re-
spect to Zk−1 , and the second line elaborates how to compute this expectation. All the elements
in Equation (12) are easy to estimate, typically using panel data.
Further simplifications are possible if we also assume P Dk is independent of k given
macro-economic factors. Consequently, we can pool data across different lengths of survival
to estimate the elements in Equation (12), or to estimate P Dk directly, conditional on X−Z
k−1 .

Whether the assumptions reasonably approximate the reality is an empirical question, a hy-
pothesis testable using banks’ proprietary data.
The discussions thus far consider idiosyncratic factors within PD models alone. If a bank
models PD and LGD such that they depend on different idiosyncratic factors, Equation (4)
would remain unchanged. Otherwise, based on Equation (12), at time k = t > 1 we need to
include LGD within the conditional Zt−1 -expectation sign. Particularly, in context of Equation
(11), Equation (4) and various simplifications, our framework becomes

M
hX 1  −Z

Lifetime ECL = E EADt EZ LGDt (X t−1 )P Dt (X t−1 )|X t−1
t=1
(1 + r)t
t−1 t−1 i
Y Y
1 − P Dk (X−Z | X−Z
   
1 − P D1 (X0 ) k−1 ) 1 − P Ek 0 , (13)
k=2 k=1

where P Dk , k = 1, 2, ..., t, and EZ are defined in Equation (12).

The discrete-time method described above can be naturally extended to the cases in con-
tinuous time. Our focus is to obtain an expression for [1 − G(t)]h(t) in a way analogous
to Equation (11), which turns out to be the approach of DSW. With λt denoting the haz-
ard rate of default, αt denoting the hazard rate of exit, and using the well-known fact that

11

Electronic copy available at: https://ssrn.com/abstract=2758513

 Rt

h(t) = λt (1 − H(t)) = λt exp − 0
λu du , we find

h Z t
ih
Z t

i
[1 − G(t)]h(t) = exp −αu du λt exp − λu du (14)
0 0
h Z t
i
= λt exp −[λu + αu ]du . (15)
0

Rt
Rt

Note that exp − 0
λu du and exp − 0 αu du are the probability of non-default and non-exit
until time t, respectively. The standard practice in the literature is to model both λu and αu as
functions of time-varying covariates Xu , and the survival probability in Equation (15) is thus
Rt

exp 0 −[λu (Xu ) + αu (Xu )]du .
We note the similarity between Equations (11) and (15). In particular, λt ,
P r(t<τ ≤t+∆t|τ >t,η>t)
lim∆t→0 ∆t
by definition is the continuous-time limit of P Dt , and survival
Rt

probability, exp 0 −[λu (Xu ) + αu (Xu )]du , can be viewed as a continuous-time version
h i
of t−1
Q
k=1 1 − (P D k + P E k ) in Equation (11). The difference, however, is that Equation
(15) assumes conditional independence between τ and η, while Equation (11) makes no such
assumption, and thus allows P Dk and P Ek to correlate with unobservable factors.
Finally, while Equation (15) is a reduced-form model, our framework in this context, Equa-
tion (6), allows other choices of modeling strategies, for example, using structural models. In
Section 4.2, we will demonstrate one such example.

3.2 Modeling Exposure and Loss at Default Time

While theoretically both EADt and LGDt can be modeled as functions of the complete
history of {Xt }, it is a standard practice, in the literature and industry, to model both quantities
conditional on Xt−1 only. Given a particular realized path of {Xt }, we can then obtain EADt
and LGDt for all t ∈ [1, M ].
Typically, EADt models are required only for revolving lending products.8 In practice,
simple expedients exist, to model EADt as the credit limit or quantities independent of Xt−1 .
Once EADt is obtained, the measure of realized LGDt can be constructed using the historic
recovery (and recovery cost) data of defaulted exposures. Using the LGD datasets, we can
8
For non-revolving products, EADt can be estimated deterministically for all t based on the planned amorti-
zation/payment schedules.

12

Electronic copy available at: https://ssrn.com/abstract=2758513

develop models expressing expected LGDt as a function of Xt−1 . There is abundant literature
on identifying Xt−1 and studying its relation with LGD (or alternatively, recovery rate). See,
for example, Altman, Brady, Resti and Sironi [2005], Acharya, Bharath and Srinivasan [2007],
Bruche and González-Aguado [2010], Jankowitsch, Nagler and Subrahmanyam [2014], to cite
a few.

3.3 Modeling Dynamics of the State Variables

In order to compute the expectation at time 0 in Equation (4) or (6), we require the den-
sity or probability of all paths of the state variables. This amounts to modeling the complete
dynamics of {Xt }, t ∈ [0, M ). We stress that the knowledge of the density/probability of Xt
for a particular t is insufficient to calculate the lifetime ECL. This is because, as illustrated in
Equation (11), the computation of the survival probability typically requires the knowledge of
the full history of Xs , 0 ≤ s < t and its associated probability or density. Therefore, we require
path-by-path information of the process {Xt }, t ∈ [0, M ). There exist a number of modeling
options, with various levels of complexity and sophistication.
One of the most popular choices is perhaps a time-homogeneous Markov process. A leading
example is, as adopted by DSW, the first-order Gaussian vector auto-regression (VAR), which
models the discrete-time version of an Ornstein-Uhlenbeck process,

Xt = ΘXt−1 + t , (16)

where, assuming Xt is a L × 1-vector, Θ is a L × L parameter matrix to be estimated, and

t is a L × 1-vector Gaussian error term, t ∼ N (0, Σ), with Σ as the variance-covariance
matrix of t . Equation (16) can also be enhanced to take into account common stylized facts in
Econometrics, such as mean-reversion or regime-switching.
Once the parameters in Θ and Σ are estimated,9 Monte Carlo simulation is a convenient tool
to project simulated paths of {Xt } using Equation (16). Because the frequency of simulated
paths already reflects their density, a simple average of the lifetime ECL evaluated given each

9
As demonstrated by DSW in a Maximum Likelihood context, the estimation of parameters (Θ, Σ) here can
be separated from the estimation of parameters (β, γ) in Equation (11).

13

Electronic copy available at: https://ssrn.com/abstract=2758513

path, using the methods discussed in Section 3.1 and 3.2 above, provides an estimate of lifetime
ECL.
We may also construct Xt such that it has a finite number of values, with each value repre-
senting an economic state. In this design, simpler choices are available to model the dynamics
of {Xt }. For example, Markov Chains models, which may have time-invariant or time-varying
transition matrices, can generate the paths going through different economic states, and their
associated probability.
Forward-looking projections of Xt are not necessarily based on quantitative models that are
purely data-driven. A combination of economists’ judgmental forecasts, qualitative indicators
and econometric models may be used to obtain future scenarios describing paths of {Xt } and
their associated probability, possibly subjective. If these scenarios are deemed reasonable and
forward-looking, they can be used to calculate lifetime ECL in the same manner as simulated
paths are used. More precisely, the weighted average of scenario-dependent lifetime ECL,
weighted by the scenarios’ probability, gives a final estimate of lifetime ECL.

4 Illustrative Examples

In this section, we present two hypothetical examples to illustrate estimation of lifetime

ECL within our framework. The focus of these examples is to highlight the key elements
and estimation procedures of the framework, rather than specific component choices. To this
end, we may assume sometimes unrealistically simplified modeling strategies in the examples,
to underplay the less important features. Nonetheless, such simplifications are without loss
of generality, because practically, we can always change these modeling strategies to more
sophisticated ones, without impacting the overall structure of the estimation framework.

4.1 A Discrete-Time Lifetime ECL Model

In this example, we illustrate how to use a model of discrete-time term structure of PD,
namely Equation (11), among other elements, to measure lifetime ECL expressed in Equation
(4). We assume the estimation of lifetime ECL is deemed necessary, by bank B of country
A, for a group of debt exposures that is homogeneous in terms of risk characteristics, product

14

Electronic copy available at: https://ssrn.com/abstract=2758513

features and time to maturity. In particular, we assume that all the exposures within this group
have: i) time-to-maturity of around 10 years, ii) EIR close to 8% and, iii) pre-payment options,
i.e. possibility of early exit. The aggregate exposure of this group is $100 Million at the
reporting date. To provide more structured explanations, we present the estimation procedure
step by step.
1. State variables and their dynamics
We assume that bank B adopts a single variable, Xt , describing the states of economy. It
has three discrete values, defined as follows:





 D if the economy in a “Downturn” state,


Xt , N if the economy in a “Normal” state, (17)





B
 if the economy in a “Booming” state.

As at the reporting date, country A is in economic recession, i.e., X0 = D. After consulting

the economists of the bank, together with econometric forecasts using historic data, the follow-
ing scenarios regarding the paths of economic recovery in the next 10 years (i.e., X0 , X1 , ...,
X9 ) are deemed adequate and forward-looking,

Economic recovery scenarios, {Xt } =






 Slow recovery, with 40% probability: D, D, D, D, D, N, N, N, N, B;


Standard recovery, with 40% probability: D, D, D, N, N, N, B, B, B, N ; (18)





Fast recovery, with 20% probability:
 D, N, N, N, N, N, B, B, B, B.

The assumption on the scenarios in Equation (18) may be over-simplified, nonetheless it rep-
resents general cases of projecting future economic conditions. In reality, we can use scenarios
resulting from more sophisticated models, for example, Markov Chain models as discussed in
Section 3.3.
2. Term structure of PD
Suppose that bank B adopts Equation (11) to estimate P r(t − 1 < τ ≤ t, η > t), in

15

Electronic copy available at: https://ssrn.com/abstract=2758513

which PIT PD (P Dt ) and probability of exit (P Et ) for each period are required. We do not
prescribe the methodology of modeling PIT PD, P Dt , in this example – they are described in
Section 3.1 in details. We assume bank B collects historical debt and default datasets, from both
internal and external sources, to develop PIT PD models. For the particular group of exposures
of interest here, given their characteristics, their PIT PD is expressed as a function of Xt as
follows,





 2.25% if Xt−1 = D,


P Dt (Xt−1 ) = 1.50% if Xt−1 = N, (19)





0.80% if Xt−1 = B.


If bank B chooses to include idiosyncratic risk factors, Zt−1 , in the model of P Dt , Equation
(19) may be interpreted such that it already integrates out any Zt−1 , as discussed in Section
3.1. For simplicity, we additionally assume P D1 (Z0 , X0 = D) = 2.25% if there exists Z0 .
We further assume that bank B analyzes the historical exit rates of the debts similar to the
group of exposures studied here. The bank finds that the exit rates do not significantly correlate
with Xt−1 , nor with other time-varying characteristics like debt age or time to maturity. As
such, bank B decides to estimate P Et = P E, where P E is taken as a long-run average of
exit rate, 2.1% per annum in this case. In general, the methodologies of modeling P Et can be
similar to those of P Dt .
Given Equations (18), (19) and P E, it is straightforward to obtain the survival probability
until time t − 1 as

t−1 h
Y
i
SPt−1 = 1 − P Dk (Xk−1 ) + P E , (20)
k=1

and thus P r(t − 1 < τ ≤ t, η > t − 1) = P Dt × SPt−1 .

3. EAD and LGD
For simplicity, we assume that the total exposures will be reduced by $10 Million per year
according to the payment schedule, and the reductions only occur at each year-end. Therefore,
EADt is the outstanding balance at the beginning of, and throughout, year t. We may change

16

Electronic copy available at: https://ssrn.com/abstract=2758513

 Table 1: Lifetime ECL under the scenario of “slow economic recovery”

Year EADt P r(t − 1 < τ ≤ t, P V (ELt ) % in

Xt−1 P Dt SPt−1 LGDt
t ($ M) η > t − 1) ($ M) Lifetime
(1) (2) (3) (4) = (2)*(3) (5) (6) = (1)∗(4)∗(5)
(1+r)t ECL
1 100 D 2.25% 100.00% 2.25% 20% 0.417 28%
2 90 D 2.25% 95.65% 2.15% 22% 0.332 22%
3 80 D 2.25% 91.49% 2.06% 20% 0.261 17%
4 70 D 2.25% 87.51% 1.97% 20% 0.203 13%
5 60 D 2.25% 83.70% 1.88% 20% 0.154 10%
6 50 N 1.50% 80.06% 1.20% 15% 0.057 4%
7 40 N 1.50% 77.18% 1.16% 15% 0.041 3%
8 30 N 1.50% 74.40% 1.12% 15% 0.027 2%
9 20 N 1.50% 71.72% 1.08% 15% 0.016 1%
10 10 B 0.80% 69.14% 0.55% 10% 0.003 0%
Total 1.510 100%
This table demonstrates a discrete-time example of estimating lifetime ECL under one future scenario, “slow
economic recovery”. The notations are the same as in the prior steps (“M” in the labels of the second leftmost
and second rightmost columns standards for “Million”). In particular, SPt−1 denotes the survival probability until
time t − 1, and is computed according to Equation (20). The second rightmost column (the column labeled as
“P V (ELt )”) computes the present value of ELt for each year. The sum of this column (shown in the last row)
is the lifetime ECL estimated under this scenario. The rightmost column computes the proportion that each year’s
P V (ELt ) contributes to lifetime ECL.

the assumptions on EADt , but our framework remains virtually unchanged.

As discussed in Section 3.2, expected LGD for a time-t default can also be estimated as
a function of Xt−1 . Again, without prescribing a specific LGD modeling methodology, we
assume bank B develops the following LGD model, after conditioning on the characteristics of
the particular group,





 20% if Xt−1 = D,


E(LGDt |Xt−1 ) = 15% if Xt−1 = N, (21)





10% if Xt−1 = B.


4. Lifetime ECL
Putting together all the results from the prior steps, we are able to estimate lifetime ECL
under each economic scenario, and measure the final lifetime ECL as the scenario-probability-
weighted average, for this particular group of exposures. Table 1 presents the calculation of
lifetime ECL under one future scenario, “slow economic recovery”.
Most of the columns in Table 1 are self-explanatory. In particular, the column labeled as

17

Electronic copy available at: https://ssrn.com/abstract=2758513

“SPt−1 (3)” computes the survival probability until time t − 1, according to Equation (20).
The second rightmost column (the column labeled as “P V (ELt )”) computes the present value
of ELt for each year. The sum of this column (shown in the last row), $1.51 Million, is the
lifetime ECL estimated under the scenario of slow economic recovery. The rightmost column
tabulates the proportion that each year’s P V (ELt ) contributes to lifetime ECL. For example, it
shows that the single-period (i.e., 12-month) ECL is about 28% of lifetime ECL. This illustrates
that lifetime ECL could be in much greater magnitude than 12-month ECL, depending on
amortization schedule, time to maturity and economic outlook, among other factors.
We can also compute lifetime ECL under other scenarios using similar tables as Table 1,
which we skip for brevity. The resultant estimates are $1.285 and 0.988 Million under stan-
dard and fast recovery scenarios, respectively. Given the probability of scenarios in Equa-
tion (18), the final lifetime ECL is measured as $1.316 Million (=1.51×40% + 1.285×40% +
0.988×20%).

4.2 A Continuous-Time Lifetime ECL Model

This example considers the estimation of lifetime ECL using a continuous-time model,
Equation (6), which also assumes conditional independence between τ and η. While this frame-
work is most conveniently implemented using reduced-form models, like those of DSW, in this
section we consider an estimation approach using structural models. This treatment demon-
strates the generality of our framework.
In the following, we first consider the unconditional lifetime ECL, which is the part in
Equation (6) without the conditioning information, X0 ,

Z M  −rt 
Unconditional Lifetime ECL = e LGDt EADt (1 − G(t))h(t) dt. (22)
0

Then we discuss the issues and possibilities of extending Equation (22) to the conditional cases.
Within Equation (22), there are three components, h(t), (LGDt EADt ) and [1 − G(t)]. We
consider them within the structural-model framework of Black and Cox [1976], Leland [1994],
Leland and Toft [1996], who model the default mechanism as the first time that a debtor’s asset
value hits a default barrier. Compared to the Merton [1974] model, this framework is more

18

Electronic copy available at: https://ssrn.com/abstract=2758513

convenient to measure lifetime ECL, because it assumes that default may occur at any time
before maturity, and thus Equation (22) is applicable.
First, we find h(t), the pdf of the default time τ , within the first-hitting-time framework, in
which a debtor’s asset value, Vt , follows a geometric Brownian Motion with an initial value V0 ,

dVt
= µdt + σdWt , (23)
Vt

where W denotes a standard Brownian Motion, µ and σ are two constants representing the drift
and diffusion coefficient of the process. Denoting VB ∈ [0, V0 ) as the default barrier, and using
the well-known density of the first-passage time at which Vt hits VB , we obtain

n log(V /V ) − (µ − σ2
2
|log(VB /V0 )| B 0 2
)t o
h(t) = √ exp − . (24)
2πσt3/2 2σ 2 t

Apart from V0 which is typically taken as a normalizing factor,10 h(t) in Equation (24) has three
parameters, namely VB , µ and σ. VB can be further modeled as a function, parameterized by
other parameters, of the debtor’s liability, cash holdings, etc. at the reporting date (time 0). µ
and σ are typically implied, using an iterative process, from the debtor’s stock return and return
volatility (See, for example, DSW and references therein).
Second, to find LGDt EADt , note that the recovery within this theoretical model is simply
a haircut version of VB . Thus, LGDt EADt can be modeled as (EADt − VB ∗ haircut), where
VB is obtained during the estimation procedure of h(t) and “haircut” can be estimated as a
parameter based on past data.
Third, for [1 − G(t)], although the exit time η may be modeled in a similar structure as
the default mechanism, economically it is unclear why the decision to call the debt is directly
relevant to asset values. It thus makes more sense to adopt reduced-formed specifications for
h Rt
i
[1 − G(t)], for example, as in Equation (14), set [1 − G(t)] = exp 0 −αu du using popular
hazard functions such as proportional hazard (like Weibull hazard), log-logistic hazard, log-
normal hazard, and so forth.

10
Note that we can always take the initial value, without loss of generality, as 1 by normalizing all quantities
in the model by V0 . In practice, V0 is typically taken as the book value of collateral (for secured, senior debts) or
total assets (for unsecured, subordinated debts) as of the reporting date from the debtor’s financial statement.

19

Electronic copy available at: https://ssrn.com/abstract=2758513

 We denote all the parameters involved in the above three components as a vector ξ. Given a
specific value of ξ, all the components are determined, and thus it is straightforward to calculate
the unconditional lifetime ECL in Equation (22). Even if the integral in this expression may
not have a closed-form solution, it is easy to numerically evaluate the integral.
Things become tricky when we try to extend Equation (22) to the conditional cases in a
manner similar to the example in Section 4.1. Particularly, if either of the three parameters, VB ,
µ, σ, becomes time-varying, h(t) in Equation (24) no longer holds. In general, it is difficult,
if not impossible, to find h(t) for time-varying VB , µ, σ. Therefore, it may be infeasible to
specify h(t) as a function of the state variables, Xt . Admittedly, this challenge results from
the restricted assumptions (Equation (23)) of the structural model, which we need to face when
adopting such a model that also offers simplicity and convenience. One possible solution is to
relate the outlook of economic conditions to the values of ξ, instead of paths of Xt .
To be precise, denoting the entire parameter space as Ξ, each ξ ∈ Ξ may correspond to one
outcome of the future economic conditions. Therefore, forward-looking economic forecasts,
given X0 , amount to specifying the conditional probability P r(ξ|X0 ), possibly subjective.
Consequently, lifetime ECL in Equation (6) can be revised as a P r(ξ|X0 )-weighted average of
Equation (22),

hZ M i
e−rt LGDt EADt [1 − G(t)]h(t) dt|X0 .
 
Lifetime ECL = E (25)
ξ∈Ξ 0

Note that if all the parameters in the vector ξ are continuous-valued, we can specify the
conditional pdf of ξ, and the approach in Equation (25) becomes somewhat analogous to an
analysis using Bayesian Statistics. Also note that this structural-model example may better suit
corporate lending portfolios, especially public-listed debtors, because the parameters µ and σ
are typically implied from the stock market information.

5 Conclusions

We propose an estimation framework of lifetime ECL based on the foundation of the multi-
period PD literature. We show that this framework is in accordance with IFRS 9 and can also

20

Electronic copy available at: https://ssrn.com/abstract=2758513

incorporate single-period ECL as a special case. The framework rests on a rigorous definition of
“term structure of PD” and conditional expectation given forward-looking economic dynamics.
It also allows banks to easily implement, within this framework, simplified and sophisticated
modeling strategies alike. We consider a number of popular modeling strategies, and demon-
strate illustrative examples how to use popular modeling techniques to measure lifetime ECL
within the framework.
Lifetime ECL is one of the key elements, and perhaps the most difficult to measure, in IFRS
9 impairment requirements. As an early attempt to tackle this compliance and quantitative
problem, this paper hopefully contributes to the banking industry which intends to robustly
measure lifetime ECL without undue cost and effort under IFRS 9.

References

Acharya, V.V., Bharath, S.T. and Srinivasan, A., 2007, Does industry-wide distress affect de-
faulted firms? evidence from creditor recoveries, Journal of Financial Economics, 85, 787–
821.

Allison, P.D., 2010, Survival Analysis Using SAS R : A Practical Guide, Second Edition, SAS
Institute Inc.

Altman, E.I., Brady, B., Resti, A. and Sironi, A., 2005, The link between default and recovery
rates: Theory, empirical evidence, and implications, Journal of Business, 78, 2203–2227.

BCBS, 2006, International Convergence of Capital Measurement and Capital Standards, Basel
Committee on Banking Supervision, Bank for International Settlements.

BCBS, 2015, Guidance on credit risk and accounting for expected credit losses, Supervisory
guidance, Basel Committee on Banking Supervision, Bank for International Settlements.

Black, F. and Cox, J., 1976, Valuing corporate securities: Some effects of bond indenture
provisions, Journal of Finance, 31, 351–367.

Bruche, M. and González-Aguado, C., 2010, Recovery rates, default probabilities, and the
credit cycle, Journal of Banking and Finance, 34(4), 754–764.

Campbell, J.Y., Hilscher, J. and Szilagyi, J., 2008, In search of distress risk, Journal of Finance,
63, 2899–2939.

Chava, S., Stefanescu, C. and Turnbull, S.M., 2011, Modeling the loss distribution, Manage-
ment Science, 57, 1267–1287.

21

Electronic copy available at: https://ssrn.com/abstract=2758513

Duffie, D., Saita, L. and Wang, K., 2007, Multi-period corporate default prediction with
stochastic covariates, Journal of Financial Economics, 83, 635–665.

IASB, 2014, International Financial Reporting Standard 9 Financial instruments, International

Accounting Standards Board.

Jankowitsch, R., Nagler, F. and Subrahmanyam, M.G., 2014, The determinants of recovery
rates in the US corporate bond market, Journal of Financial Econometrics, 114, 155–177.

Lando, D., 2004, Credit Risk Modeling: Theory and Applications, Princeton University Press.

Leland, H.E., 1994, Corporate debt value, bond covenants, and optimal capital structure, Jour-
nal of Finance, 49, 1213–1252.

Leland, H.E. and Toft, K.B., 1996, Optimal capital structure, endogenous bankruptcy, and the
term structure of credit spreads, Journal of Finance, 51, 987–1019.

Merton, R.C., 1974, On the pricing of corporate debt: the risk structure of interest rates, Journal
of Finance, 29, 449–470.

Wooldridge, J.M., 2010, Econometric Analysis of Cross Section and Panel Data, Second Edi-
tion, MIT Press.

Appendices

A Derivation of Equation (9)

Applying Bayes’ rule repeatedly to the survival probability,

P r(τ > t − 1, η > t − 1) = P r(τ > t − 1, η > t − 1, τ > t − 2, η > t − 2)

=P r(τ > t − 1, η > t − 1|τ > t − 2, η > t − 2)P r(τ > t − 2, η > t − 2)
h i
= 1 − P r(η ≤ t − 1 or τ ≤ t − 1|τ > t − 2, η > t − 2) P r(τ > t − 2, η > t − 2)
h i
= 1 − (P Et−1 + P Dt−1 ) P r(τ > t − 2, η > t − 2), (26)

where the last line of Equation (26) is due to the fact that default and exit are mutually exclusive.
It gives Equation (9).

22

Electronic copy available at: https://ssrn.com/abstract=2758513

 Still “Too Much, Too Late”: Provisioning
for Expected Loan Losses∗

Roman Goncharenkoa and Asad Raufb

a
Central Bank of Ireland and KU Leuven
b
University of Groningen

The new accounting standards of IFRS 9 and U.S. GAAP

adopt the expected loss (EL) approach for loan loss recog-
nition. We investigate the effect of the EL approach on bank
loan supply and stability. When a bank is unable to anticipate a
downturn in the business cycle, it ends up recognizing the bulk
of expected losses after the arrival of a contraction. This aggra-
vates lending procyclicality and can potentially worsen bank
stability. We develop a dynamic model of a bank to quantita-
tively assess these effects and show that they are economically
significant.
JEL Codes: G21, G28, M41, M48.

1. Introduction

To ensure an accurate assessment of their overall financial positions,

banks periodically account for anticipated future loan losses through
loan loss provisions. In doing so, they must comply with accounting
standards for loan loss recognition. The recent financial crisis spurred

∗
Previously, this paper circulated under the title “Loan Loss Provision-
ing Requirements in a Dynamic Model of Banking.” For insightful comments,
we would like to thank Jorge Abad (discussant), Diana Bonfim, Mark Clat-
worthy, Hans Degryse, Theodoros Diasakos (discussant), Christian Laux, Mike
Mariathasan, David Martinez Miera, Robert Marquez, Alex Michaelides (discus-
sant), Bruno Parigi, Omar Rachedi (discussant), Rafael Repullo, Jonathan Smith
(discussant), Javier Suarez, Alonso Villacorta, and Toni Whited, as well as two
anonymous referees, the seminar participants at VGSF (WU Wien), KU Leuven,
CERGE-EI, University of Zurich, University of Bristol, Bank of Portugal, and the
conference participants of the 2nd Biennial Bank of Italy and Bocconi University
Financial Stability Conference, the Contemporary Issues in Banking Conference
in St. Andrews, the 27th Finance Forum in Madrid, and the 2nd Endless Sum-
mer Conference on Financial Intermediation and Corporate Finance in Athens.
Author e-mails: roman.goncharenko@kuleuven and [email protected].

415
416 International Journal of Central Banking October 2024

criticism of the then-existing standards, which were based on the

incurred loss (IL) approach. This approach limited loss recognition
only to those losses that were factually identified (i.e., incurred)
before the balance sheet date. As these standards led to delayed
provisioning and insufficient loan loss reserves, they were blamed for
contributing to the credit crunch (Financial Stability Forum 2009).
The policy response was to adopt a more “forward-looking” pro-
visioning approach based on expected rather than incurred credit
losses.1 Under the expected loss (EL) approach, banks’ provisions
constitute unbiased estimates of future losses over a specified hori-
zon. The new accounting standards of IFRS 9 and the new U.S.
GAAP replace the IL approach with the EL approach.
The objective of this paper is to quantify the long-term effect of
the EL approach on the cyclicality of bank lending and stability. As
a rationale for the adoption of the EL approach, the Financial Stabil-
ity Forum (2009) states that “earlier recognition of loan losses could
have dampened cyclical moves in the current crisis and is consistent
both with financial statement users’ needs for transparency regard-
ing changes in credit trends and with prudential objectives of safety
and soundness.” However, there is a shared concern among acade-
mics, policymakers, and market participants that the EL approach
may actually have a strong procyclical effect (Barclays 2017, Euro-
pean Systemic Risk Board 2017, Abad and Suarez 2018). If banks
fail to anticipate a downturn in the business cycle, they recognize
the bulk of expected losses after, and not before, the arrival of a
contraction. This leads to a spike in provisions right at the start of
a contraction, which erodes banks’ profit margins and, unless they
can swiftly raise fresh equity, reduces their lending capacity. Such a

1
The G-20 summit in London on April 2, 2009 resulted in signing the Declara-
tion on Strengthening the Financial System, which included the following reforms
among others: strengthen accounting recognition of loan loss provisions by incor-
porating a broader range of credit information and improve accounting stan-
dards for provisioning. The International Accounting Standards Board (IASB)
and the U.S. Financial Accounting Standards Board (FASB) set in motion a joint
project to improve accounting standards and, in particular, to develop methods
of accounting for credit losses that would give more timely recognition of those
losses, thereby helping to reduce lending procyclicality. This effort resulted in the
International Financial Reporting Standard (IFRS) 9 and the credit loss stan-
dard (ASC 326) under the U.S. GAAP (generally accepted accounting principles),
both of which adopt the expected credit loss approach.
Vol. 20 No. 4 Still “Too Much, Too Late” 417

sudden front-loading of losses at the dawn of a contraction could not

only force banks to cut new loans but also jeopardize their stability.
From a macroprudential point of view, bank procyclicality is
widely viewed as undesirable by both academics and policymak-
ers (Hanson, Kashyap, and Stein 2011).2 The effort to reduce the
procyclical effect of risk-based capital regulation led to the revision
of the Basel Accords in the form of the new Basel III regulation,
which includes policy instruments designed to reduce lending pro-
cyclicality.3 The EL approach can potentially undermine the post-
crisis regulatory effort to reduce bank procyclicality and is likely to
be inefficient from a macroprudential point of view.
To quantify the effect of the EL approach, we adopt a struc-
tural, rather than reduced-form, approach.4 We develop a dynamic
model of a bank. Our model features endogenous loan origination,
distribution, leverage, and default. The bank faces corporate taxes,
the cost of issuing external equity, and regulation. The regulatory
environment comprises a minimum capital requirement and provi-
sioning standards. The capital structure of the bank consists of fully
insured short-term deposits and equity. The asset side is composed
of risky long-term loans with stochastic and time-varying default
probabilities.
First, we calibrate our model under the benchmark provisioning
requirement, which is based on the IL approach of the International
Accounting Standards (IAS) 39. Next, we solve our model under
two variations of the EL approach, namely the expected credit loss
(ECL) of IFRS 9 and the current expected credit loss (CECL) of the

2
The literature on optimal time-varying capital requirements provides much
support in favor of a countercyclical capital regulation (i.e., procyclical capi-
tal requirements), which helps to smooth the cyclicality of credit supply (see
Kashyap and Stein 2004, Dewatripont and Tirole 2012, Repullo 2013, Gersbach
and Rochet 2017, and Malherbe 2020, among others). Empirical evidence further
suggests that a countercyclical capital regulation indeed helps to reduce credit
crunch (Jiménez et al. 2017).
3
Basel III instruments such as the countercyclical capital buffer, the con-
servation capital buffer, and contingent capital are all meant to reduce lending
procyclicality.
4
An empirical investigation using a reduced-form approach would require data
that include a full credit cycle under the EL approach. However, such data are
not available, since the accounting standards that adopt the EL approach either
have only recently been put in effect (i.e., IFRS 9) or are still planned to be
implemented (U.S. GAAP).
418 International Journal of Central Banking October 2024

new U.S. GAAP.5 We compare the solutions of the model under the
two versions of the EL approach to the benchmark case.
Our quantitative results indicate that the adoption of either ver-
sion of the EL approach results in a profound aggravation of lending
procyclicality in the long run. Our model predicts that, on average,
in a contraction, a bank originates about 6–7 percent fewer new loans
under the EL than the IL approach. At the same time, unconditional
on the aggregate state, the bank’s lending is only about 2–3 percent
lower under the EL approach. This highlights the strong procycli-
cality of the EL approach, as it disproportionately reduces lending
in a contraction.
We further examine the procyclicality of the EL approach when
a bank is subject to the countercyclical capital buffer (CCyB), which
is a new Basel III policy that explicitly aims at reducing procyclical-
ity. We find that the CCyB is unable to fully offset the procyclical
effect of the EL approach—that is, the simultaneous adoption of
the EL approach and CCyB also results in more procyclical lending
than under the benchmark. The magnitudes, however, are attenu-
ated. Furthermore, even when we allow the bank’s profits to respond
with a one-period delay to the arrival of a contraction, which effec-
tively allows the bank to anticipate the deterioration of its balance
sheet, the procyclical effect of the EL approach still persists.
Next, we show that when it comes to the effect of the EL
approach on banks’ stability there are two effects in play. On the one
hand, under the EL approach, a bank holds larger loan loss reserves
since on top of the incurred losses it must also recognize expected
losses. Larger reserves provide better loss-absorption capacity, thus
improving stability. On the other hand, the procyclicality of the
EL approach effectively increases the volatility of the bank’s profits.
This, in turn, increases the bank failure rate. The overall effect of the
EL approach on stability will depend on the relative strength of these
two effects. Our quantitative model suggests that IFRS 9 is more
likely to increase bank failure rate than U.S. GAAP. While their

5
The primary difference between these models is that they adopt different
horizons over which expected losses must be recognized. IFRS 9 is based on a
mixed-horizon approach such that, depending on the loan’s risk category, the
bank recognizes either one-year or lifetime discounted expected losses. The new
U.S. GAAP, on the other hand, requires banks to recognize lifetime discounted
expected losses on all loans.
Vol. 20 No. 4 Still “Too Much, Too Late” 419

procyclical effect on lending is similar, IFRS 9 produces smaller loan

loss reserves than U.S. GAAP, due to its mixed-horizon approach.
Our analysis indicates that earlier recognition of losses does not
per se help to smooth lending cyclicality. It matters how early in
advance future losses are recognized. If future losses were to be rec-
ognized before the arrival of a contraction, this would result in a pre-
cautionary capital buffer, which would then help to smooth lending
in a downturn. For example, the Spanish dynamic loan loss provi-
sioning approach, which prescribed higher provisions in expansions
relative to contractions, allowed the banks to effectively build up a
capital buffer during good times and consequently smooth lending
when the contraction arrived (Jiménez et al. 2017).6 In contrast,
under the EL approach, banks will recognize the bulk of expected
losses after the arrival of a contraction, provided they cannot antici-
pate the change in the aggregate state well in advance. Thus, forcing
banks to recognize their future losses based on the EL approach is
equivalent to imposing a more countercyclical capital requirement. It
is well understood that a countercyclical capital requirement results
in more procyclical lending (Kashyap and Stein 2004; Repullo, Sau-
rina, and Trucharte 2010).
Our paper contributes to several strands of the literature. First,
our paper relates to a large literature on the cyclical implications of
bank regulation. Kashyap and Stein (2004) provide a formal analy-
sis of the procyclical effect of capital regulation on bank lending
and advocate for more procyclical capital requirements than those
of Basel II. Similarly, in a dynamic model of banking, Repullo and
Suarez (2013) show that procyclical adjustment to the Basel II capi-
tal requirements are welfare-improving. In general, the theoretical
literature provides a vast support in favor of procyclical capital
requirements (Dewatripont and Tirole 2012; Repullo 2013; Gers-
bach and Rochet 2017; Malherbe 2020). Empirical evidence further
indicate that more procyclical capital requirements indeed help to
smooth bank lending better (Behn, Haselmann, and Wachtel 2016;
Jiménez et al. 2017). While this literature examines the cyclical effect
of capital regulation, we focus on the cyclical effect of loan loss pro-
visioning requirements. We show that adopting the EL approach for

6
This approach is very similar to a countercyclical capital buffer policies such
as CCyB.
420 International Journal of Central Banking October 2024

loan loss recognition is similar to imposing a more countercyclical

capital requirement.
Second, our paper contributes to the literature on bank loan loss
provisioning. While the empirical literature on loan loss provision-
ing is relatively large (Ahmed, Takeda, and Thomas 1999; Laeven
and Majnoni 2003; Beatty and Liao 2011; Bushman and Williams
2015; Huizinga and Laeven 2018), the theoretical literature is rather
scarce. To the best of our knowledge, our paper is the first one to for-
mally examine the effect of the provisioning requirement for future
losses on bank loan supply and stability.
Third, our paper contributes to the literature that studies the
cyclical impact of the expected credit loss approach of the new
accounting standards. Early concerns about the potential procycli-
cality of the expected loss approach can be found in Laux (2012),
Barclays (2017), and European Systemic Risk Board (2017). A few
papers provide a quantitative assessment of the cyclical implications
of the EL approach on capital and provisions. Krüger, Rösch, and
Scheule (2018) show that had the banks followed the EL approach,
they would have had lower levels of capital, especially during the
2007–08 crisis, and the bank capital would have been more procycli-
cal. Under the assumption that the bank can anticipate the turn in
the business cycle, Cohen and Edwards (2017) and Chae et al. (2019)
show that the EL approach achieves better smoothing of provisions
compared to the IL approach. Abad and Suarez (2018) quantify the
procyclical effect of the EL approach on bank capital in a dynamic
model of a bank with exogenous lending and heterogeneous loans.
In our paper, we evaluate the effect of the EL approach on loan
supply and bank stability under the bank’s optimal behavior, as our
settings allow for endogenous lending, financing, and default. Thus,
our analysis is less prone to the Lucas critique, since we allow the
bank to optimally respond to the adoption of the EL approach.
Finally, our paper broadly relates to the growing literature on the
interaction between accounting practice, on the one hand, and finan-
cial stability and prudential regulation, on the other (see Goldstein
and Sapra 2014 and Acharya and Ryan 2016 for a survey). Laux
and Leuz (2010) provide critical analysis of the role of fair-value
accounting in the recent financial crisis. Mahieux, Sapra, and Zhang
(2023) examine the effect of mandatory earlier loss recognition on
bank risk-taking. We contribute to this literature by pointing out
Vol. 20 No. 4 Still “Too Much, Too Late” 421

that provisioning rules and capital regulation are two sides of the
same coin and, thus, should not be designed independently.
The rest of the paper proceeds as follows. Section 2 provides insti-
tutional details on provisioning, the expected loss models of IFRS
9, and the new U.S. GAAP. Section 3 introduces a simple three-
date model of a bank to highlight the economic mechanism through
which the EL approach affects bank lending and stability. Section
4 presents a quantitative dynamic model of a bank, while the cal-
ibration of the model is found in Section 5. Section 6 contains the
quantitative analysis and results of the long-term effect of adopt-
ing provisioning requirements based on the EL approach. Finally,
Section 7 concludes.

2. Institutional Details

A loan loss provision is a non-cash expense set aside as an allowance

for impaired loans. It is an accounting entry that increases loan loss
reserves (a contra asset account on the balance sheet) and reduces
net income. Empirically, such provisions constitute a large fraction
of bank expenses (Huizinga and Laeven 2018). As a result, they
substantially reduce a bank’s profit in financial statements, thereby
affecting regulatory capital. In the future, when the losses realize,
they are charged off against the loss reserves. The rules for loan loss
provisioning for internationally active banks and U.S.-based banks
are formulated by the International Accounting Standards Board
(IASB) and the U.S. Financial Accounting Standards Board (FASB),
respectively.
The International Accounting Standards (IAS) 39, which was
effective until the end of 2017, employed the so-called incurred loss
model (ILM) for loan loss recognition. The ILM did not allow banks
to recognize credit losses based on the events expected to happen
in the future. Under the ILM, it was assumed that all loans would
be repaid until the evidence to the contrary is established. Only at
that point, the impaired loan (or portfolio of loans) could be writ-
ten down to a lower value. Therefore, only those losses that were
factually documented could be recognized.
In the aftermath of the financial crisis of 2007–08, IAS 39 was
criticized for potentially contributing to the credit crunch, as it
422 International Journal of Central Banking October 2024

did not allow for timely loss recognition (Financial Stability Forum
2009). The policy response was to adopt a more “forward-looking”
provisioning approach based on expected rather than incurred credit
losses. To reduce the procyclical effect of provisioning and improve
transparency, the IASB and the FASB created new accounting stan-
dards. Each of these standards introduces its own version of the
expected loss model (ELM). IFRS 9 replaced IAS 39 in January
2018, while the implementation of the new U.S. GAAP was planned
for the end of 2020 but has been delayed due to the COVID-19
pandemic.7
Under the ELM, banks’ provisions must constitute unbiased esti-
mates of future losses over a specified horizon. The defining feature
of the ELM is that it employs the so-called point-in-time (PiT) loan
default probabilities when estimating expected credit losses. That
is, expected losses are estimated not just based on historical data
but also with the incorporation of all presently available relevant
information.8 Thus, under the ELM, banks must employ statistical
inference to provide an unbiased estimate of expected loan losses
taking into account all currently available information.
The model of IFRS 9 adopts a mixed-horizon approach: either
one-year or lifetime discounted expected losses are recognized
depending on the risk category of loans. The bank must recog-
nize one-year discounted expected losses on stage 1 (good-quality)
loans and lifetime discounted expected losses on stage 2 (sub-quality)
loans.9 The current expected credit loss (CECL) model of the new
U.S. GAAP adopts a lifetime horizon for the entire portfolio of loans
irrespective of their credit risk. Moreover, whereas under IFRS 9
expected losses are discounted at the loan’s contractual interest

7
The adoption of IFRS 9 is still in the transitional period especially due to
the COVID-19 pandemic (see Borio and Restoy 2020 for details).
8
For example, according to the IFRS 9, “an entity shall adjust historical data,
such as credit loss experience, on the basis of current observable data to reflect
the effects of the current conditions and its forecasts of future conditions that
did not affect the period on which the historical data is based and to remove
the effects of the conditions in the historical period that are not relevant to the
future contractual cash flows” (paragraph B 5.5.52 in IASB 2014).
9
IFRS 9 also specifies stage 3 loans, which are non-performing loans (NPLs).
The accounting treatment of such loans under IFRS 9 is similar to that of incurred
losses under IAS 39.
Vol. 20 No. 4 Still “Too Much, Too Late” 423

rate, banks can use their own discount rates under the new U.S.
GAAP.

3. A Simple Model

3.1 Setup
There are three dates labeled as t = 0, 1, 2. At t = 0, the bank starts
with an initial equity endowment of E0 > 0. At t = {0, 1} the bank’s
risk-neutral manager maximizes the present value of expected future
dividends by investing in a risky portfolio of loans Lt , which matures
over one period at t + 1. The loan portfolio Lt has a stochastic net
repayment rate rt+1L
∼ N (μt , σt ). The loan portfolio is funded with
the bank’s equity capital Et and one-period deposits Bt . Deposits
are assumed to be fully insured, with the deposit insurance priced
at a flat rate normalized to zero.10 To keep the model simple, both
the deposit repayment rate and the discount rate are set to zero.
The bank operates in a regulatory environment characterized
by the capital regulation and provisioning requirement for expected
future loan losses. In the model, the minimum capital requirement
serves to minimize the probability of bank failure, thus minimizing
the implicit cost of the deposit insurance. When investing in loan
portfolio at t, at least a fraction of κt ∈ [0, 1] of the portfolio must
be financed with equity—that is, the minimum capital requirement
takes the standard form of

Et ≥ κt Lt . (1)

The provisioning requirement for future loan losses is specified

in terms of a requirement on loan loss reserves. Specifically, the
provisioning requirement stipulates that when the bank originates
new loans, Lt , the expected losses on the entire portfolio must be
recognized. Let the expected losses on portfolio Lt be θt Lt , where

10
When the bank fails, it defaults on deposits and their interest. In that case,
the deposit insurance agency fully repays the depositors the principal and the
interest. Therefore, although from the point of view of the insurance agency
deposits are risky, from the point of view of depositors these are risk-free claims.
424 International Journal of Central Banking October 2024

0 ≤ θt < 1 is the expected loss rate; then the bank’s loss reserves Rt
are given by11

Rt = θt Lt . (2)

Consequently, once the investment and provisioning have been

made, the bank’s balance sheet identity is given by

Lt − Rt = Bt + Et . (3)

The bank’s profits are subject to corporate taxes. To account

for loss limitations, the bank’s profits are taxed at the marginal tax
rate τ > 0 when the profits are positive, while the tax rate is zero
when the profits are negative.12 Therefore, the corporate tax rate is
a function of the bank’s profits πt is given by

τ (πt ) = τ Iπt >0 , (4)

where Iπt >0 denotes an indicator functions which is equal to one

when πt > 0 and zero otherwise.
Raising equity externally is assumed to be prohibitively expen-
sive. Thus, the bank can increase its equity capital only internally—
via profit retention. The bank is subject to limited liability
constraint—if the value of its equity drops below zero, the bank
optimally defaults generating a zero payoff to the shareholders.
Finally, we introduce the last two ingredients into our model,
which are important for the analysis of the expected provisioning
requirement. First, we assume that at t = 2 the economy is charac-
terized by the aggregate state s2 , which is either good, s2 = g, or
bad, s2 = b. The aggregate state affects the expected loan repay-
ment rates at t = 2: under the good aggregate state the expected

11
It is important to stress that in our model the loan loss reserves Rt are com-
posed of only provisions for expected credit losses and not of realized ones. In
practice, banks also keep loan loss reserves against realized losses. However, since
our analysis is on the requirement for recognition of expected loss, we assume
that the bank does not hold reserves against realized losses—that is, the realized
losses are written off immediately as they are realized without being accumulated
in the form of non-performing loans.
12
We follow Hennessy and Whited (2007), who adopt this parsimonious
approach to model a corporate tax schedule that accounts for loss limitations.
Loss limitations are introduced as a kink in the tax schedule producing convexity.
Vol. 20 No. 4 Still “Too Much, Too Late” 425

loan repayment is higher than under the bad one. Furthermore, we

assume that at t = 1 the bank receives a signal z1 over the aggre-
gate state at t = 2, which is either good, g, or bad, b. The signal is
informative in that P (s2 = a|z1 = a) > P (s2 = a) for a ∈ {g, b}. As
discussed in the previous section, under the expected provisioning
requirement the bank incorporates all presently available informa-
tion to estimate and recognize expected losses. Thus, the expected
provisioning rate at t = 1 depends on the signal and, thus, is denoted
as θ1|z . Since the expected provisioning rate effectively corresponds
to an expected loss rate, it follows that θ1|g < θ1|b .

3.2 Solution
The model is solved backwards starting at t = 2. The t = 2 profits
are given by
 
π2 = (1 − τ (π2 )) r2|s
L
L1|z + R1 , (5)

where the first term is the net repayment on loan portfolio, L1|z ,
and the second term captures the fraction of losses on the portfolio
that has already been recognized via provisioning at t = 1. Since the
world ends at t = 2, the manager uses all available funds to pay out
the dividend. Thus, the t = 2 dividend X2 = π2 + E1 . If X2 < 0,
L
which happens under a relatively low realization of r2|s , then the
bank fails at t = 2, in which case the value of the bank is zero due
to the limited liability constraint.
At t = 1, the bank’s manager maximizes the sum of the t = 1
dividend and the next period dividend, provided the bank does not
fail t = 1, which happens under a low realization of r1L .13 That is,
the problem is given by
 
V1 =max X1 + E [X2 |z1 , X2 > 0] ,
{L1 } , (6)
subject to X1 ≥ 0,

13
The bank fails at t = 1 if X1 < 0, even if the bank does not lend—that is,
L1|z = 0.
426 International Journal of Central Banking October 2024

where, from a basic accounting identity, the t = 1 dividend is given

by

X 1 = E 0 + π1 − E 1 , (7)

and the t = 1 after-tax profits are given by

 
π1 = (1 − τ (π1 )) r1L L0 + R0 − R1 , (8)

where the first term inside the square bracket is the net repayment
on the initial loan portfolio L0 , and the last two terms capture
provisioning for expected losses.
If the bank does not fail at t = 1—that is, if the realization of r1L
is sufficiently high—then conditional on the realization of the signal
z1 the bank chooses its optimal loan portfolio L1|z . Since the bank is
protected by the limited liability constraint, the expected t = 2 div-
idend is always positive, E [max{0, X2 }] > 0, and increasing in L1|z .
Therefore, it is straightforward to show that the manager invests
as much as possible in L1|z until X1 ≥ 0 is binding. For the same
reason, the minimum capital constraint is also binding at t = 0, 1.
The optimal L1|z is then derived by setting X1 = 0, in which case
one obtains

κ0 + (1 − τ (π1 )) θ0 + r1L
L1|z = L0 . (9)
κ1 + θ1|z (1 − τ (π1 ))
Similarly, one derives the optimal initial portfolio L0 . Since the min-
imum capital constraint is binding and since recognizing expected
losses at t = 0, θ0 L0 brings the available equity down to E0 − θ0 L0 ,
the optimal L0 is given by
E0
L0 = . (10)
θ 0 + κ0

3.3 Analysis and Discussion

3.3.1 Provisioning Requirement vs. Capital Requirement
Despite its simplicity, our stylized model can be used to understand
the implications of adopting expected provisioning requirement for
future losses on bank lending and stability. To provide better intu-
ition for the effect of the EL approach on bank lending and stability,
Vol. 20 No. 4 Still “Too Much, Too Late” 427

it is instructive to decompose the effect of the provisioning require-

ment for future losses into two channels: the capital requirement
channel and the tax channel, which is accomplished in a proposition
below.

Proposition 1. Let (Le0 , Le1|z ) and Pt denote the optimal lending

schedule and bank default probability at t = 1, 2, respectively, of a
bank that is subject to the minimum capital requirement Et ≥ κt Lt
and the provisioning requirement Rt = θt Lt . Then (Le0 , Le1|z ) and Pt
are also the optimal lending schedule and bank default probability at
t = 1, 2, respectively, of a bank that is subject to the minimum cap-
ital requirement Et ≥ (κt + θt )Lt and the provisioning requirement
Rt = 0, and receives a tax subsidy τ (πt ) (θt Lt − θt−1 Lt−1 ) at t.

Proof. The proof is straightforward. By solving for (Le0 , Le1|z ) when

the bank is subject to the minimum capital requirement Et ≥
(κt + θt )Lt and the provisioning requirement Rt = 0, and receives
a tax subsidy τ (πt ) (θt Lt − θt−1 Lt−1 ) at t, it is straightforward to
show that (Le0 , Le1|z ) are given by Equations (9) and (10). 
It follows from Proposition 1 that when the future loan losses
are not tax deductible, then the provisioning rate θt has the exact
same effect on bank lending as does the required minimum equity
rate κt . This is intuitive since accounting-wise the net income before
provisions is the source for both equity capital (through retention)
and loan loss reserves (LLRs) (through provisioning). Having to
recognize future losses limits the bank’s ability to increase equity
through retention and vice versa. Therefore, the cost of provisioning
is the same as those of increasing equity capital. At the same time,
both LLRs and equity capital serve as a buffer to absorb the loan
losses once they are realized. Our model, therefore, highlights that
the minimum capital requirement and the provisioning requirement
for future losses are in effect substitutes. This insight informs the
policy debate around macro- and micro-prudential regulation that
capital requirements (set by bank regulators) and accounting stan-
dards on provisioning (set by market regulators) cannot be isolated
from each other. Moreover, the tax treatment of provisions will also
influence the optimal lending policies. This has important policy
implications, especially given that both the accounting standards
428 International Journal of Central Banking October 2024

for loan loss recognition and the capital regulation under Basel III
are undergoing drastic modifications.
Therefore when future losses are not tax deductible, we can
understand the effect of θt on lending through κt . In particular,
increasing θt will lower lending, since it forces the bank to rely
more on equity financing, which raises the cost of capital. Moreover,
imposing a countercyclical (procyclical) provisioning requirement—
that is, θ1|g < θ1|b , (θ1|g > θ1|b )—will aggravate (mitigate) lend-
ing procyclicality. This follows immediately from a well-established
result in the literature: imposing a more countercyclical (procycli-
cal) capital requirement results in a more (less) procyclical lending
(Kashyap and Stein 2004; Repullo, Saurina, and Trucharte 2010).
The second channel of the effects of provisioning requirement
for future losses is due to the tax deductibility of these provisions.
Note that this channel is present even if the bank is not subject to
the minimum capital requirement and cannot be directly offset by
adjusting the minimum capital requirement.14 According to Propo-
sition 1, when provisions are tax deductible they generate the tax
subsidy. Due to the convexity of the tax schedule τ (πt ), the tax sub-
sidy has a higher value in good times, when the profits tend to be
higher. Therefore, this subsidy allows the bank to lend more aggres-
sively when times are good vis-à-vis when times are bad, thereby
amplifying lending procyclicality.15

3.3.2 Incurred Loss Approach vs. Expected Loss Approach

Recall that under the IL approach, banks are not allowed to recog-
nize losses that are based on events expected to happen in the future.
Therefore, we accommodate the IL approach in our model by set-
ting θt = 0.16 In contrast, under the EL approach the banks must

14
The capital regulation under Basel III is conditional on banks’ portfolio risk
and cannot be conditioned directly on banks’ profits.
15
Naturally, reducing the convexity of the tax schedule—that is, improving the
banks’ ability to shift losses intertemporally—could reduce the procyclical effect
of tax deductibility.
16
We assume incurred losses are provisioned for and charged off as soon as they
are incurred.
Vol. 20 No. 4 Still “Too Much, Too Late” 429

recognize expected losses on loans already at their inception. Impor-

tantly, expected losses must be estimated using all presently avail-
able information—that is, conditional on the current state. Without
going too much into details on how exactly θt is set under the EL
approach at this moment, we set θ1|a = θ0 > 0 and θ1|b = θ0 + ,
where  > 0, so that θ1|a < θ1|b , which reflects the fact that expected
losses are higher conditional on the bad signal about the aggregate
state.
Mathematically, replacing the IL with EL approach is equiva-
lent to simultaneously raising θ0 and . Proposition 1 then helps us
understand the implications of replacing the IL approach with the
EL approach on bank lending and stability. In line with Proposition
1, adopting the EL approach is equivalent to tightening the capi-
tal requirement (increasing θ0 ) and making it more countercyclical
(increasing ). Thus, even absent the tax deductibility of expected
provisions, adopting the EL approach will depress lending, worsen
lending procyclicality, and improve stability.
The following proposition formalizes our results with regard to
the implications of adopting the EL approach for loan loss recogni-
tion on bank lending and stability:

Proposition 2. Let (Li0 , Li1|z ) and (Le0 , Le1|z ) denote optimal lend-
ing under the IL and EL approaches, respectively. Let Pt denote the
probability of bank failure at t = 1, 2. Finally, define provisioning
rates for expected losses as θ1|g := θ0 > 0 and θ1|b := θ0 + , where
 > 0. Then replacing the IL approach with the EL approach

• lowers lending, dLe0 /dθ0 < 0 and dLe1|z /dθ0 < 0,

• amplifies lending procyclicality, d Le1|g − Le1|b /d > 0,
• improves stability, dP1 /dθ0 < 0 and dP2 /dθ1|z < 0.

Proof. See proof in Appendix A. 

That the model predicts a more procyclical lending under the

EL approach is problematic. One of the two objectives of adopt-
ing the more forward-looking EL approach is to reduce lending
430 International Journal of Central Banking October 2024

procyclicality.17 Intuitively, this cannot be achieved under the EL

approach for as long as the expected credit losses are computed
based on the point-in-time (PiT) default probabilities. By construc-
tion the PiT probabilities are countercyclical—defaults are relatively
more common in a recession—and thus, so are the expected credit
losses.18 While some of the countercyclicality of the EL approach
can be undone via adjustments to the capital requirements, it can-
not be fully eliminated due to the tax-deductibility channel dis-
cussed above. Moreover, such adjustments would make bank reg-
ulatory policy dependent on accounting rules, thus further raising
its complexity.
In the next section, we extend our simple model to quantita-
tively evaluate the effect of replacing the IL approach with EL one.
With a richer dynamic model, we can calibrate the model’s parame-
ters using their observed counterparts in real-world data and gen-
erate simulations to quantitatively compare the two provisioning
approaches.

4. Quantitative Dynamic Model

The is a partial equilibrium model and the bank takes all prices
as given. Time is discrete and the horizon is infinite. The timing
notation in the model is such that the predetermined (i.e., state)
variables at time t have subscript t − 1, while exogenous shocks real-
ized at t as well as the choice variables at time t are all indexed
by t.
The bank’s risk-neutral manager, acting on the behalf of share-
holders, invests in a risky and illiquid portfolio of long-term loans Lt
funding this investment with one-period deposits Bt and equity Et .

17
The second objective is to improve transparency via more timely loan loss
recognition. In our analysis, we do not analyze the effects of potential changes
in transparency. While this is an extremely interesting question, our dynamic
model cannot accommodate such complexity. Thus, our focus is solely on the
interaction of regulatory requirements—in the form of capital constraint—and
accounting standards—in the form of provisioning requirements.
18
Incidentally, the procyclical effect of PiT default probabilities was appre-
ciated when the internal ratings-based framework was introduced in Basel II,
which makes use of the so-called through-the-cycle (TTC) default probabilities
that reflect expected default rates under normal business conditions.
Vol. 20 No. 4 Still “Too Much, Too Late” 431

Deposits are assumed to be fully insured with the deposit insurance

priced at a flat rate normalized to zero.19 The bank provisions for
loan losses and, thus, holds LLRs Rt . The following balance sheet
identity holds:

Lt − Rt = Bt + Et . (11)

4.1 Aggregate State

The economic environment characterized by the aggregate state st .
The aggregates state follows a discrete-time Markov chain. The state
space of st consists of two values g and b corresponding to expansion-
ary and contractionary aggregate state, respectively. The transition
probability from state st to st+1 is denoted by qst ,st+1 .

4.2 Loan Portfolio

The bank’s loan portfolio consists of two types (categories) of loans:
stage 1 (good credit quality) and stage 2 (impaired credit quality)
loans.20 Let ξti ∼ F (ξti ; st ) denote a random fraction of stage i loans
that defaults at the beginning of period t. Conditional on aggregate
state, default rate ξti is iid. The cumulative distribution function
(CDF) is ranked in terms of first-order stochastic dominance with
respect to aggregate so that F (ξti ; st = g) ≤ F (ξti ; st = b) holds.21
All non-defaulted loans repay the same interest rsLt−1 at time t
and a fraction δ ∈ (0, 1) of them matures, repaying the principal.22
Defaulted loans are resolved and written off in the same period they

19
When the bank fails, it defaults on deposits and their interest. In that case,
the deposit insurance agency fully repays the depositors the principal and inter-
est. Therefore, although from the point of view of the insurance agency deposits
are risky, from the point of view of depositors these are risk-free claims.
20
Heterogeneity of loans based on quality is crucial for capturing regulatory
aspects of different versions of the EL approach.
21
In the calibration section, we show that under assumption that individual
loans are exposed to a single common factor, and thus have imperfectly correlated
defaults, F (ξti ; st ) takes the form of the Vasicek distribution (Vasicek 2002).
22
Every period, a loan matures with probability δ. Therefore, the average matu-
rity of the loan portfolio is then given by 1/δ > 1. Thus, the bank is engaged in
maturity transformation.
432 International Journal of Central Banking October 2024

default.23 Defaulted loans yield a recovery rate of 1 − λist ∈ [0, 1].

Thus, λist is a loss given default rate.
The fraction of type 1 loans is given exogenously and is denoted
by ωst . Therefore, the default rate for the portfolio of all loans is
given by ξt = ωst ξt1 + (1 − ωst )ξt2 .
Every period the bank originates new loans, Nt ≥ 0, thus, the
total portfolio of loans evolves according to the following law of
motion

Lt = (1 − ξt ) (1 − δ) Lt−1 + Nt . (12)

Loan origination is a costly process.24 Following De Nicolò,

Gamba, and Lucchetta (2014) and Mankart, Michaelides, and
Pagratis (2020), we assume a quadratic lending cost function

φ 2
C(Nt ) = N , (13)
2 t

where φ > 0. We restrict Nt to non-negative values—that is, the

bank is not allowed to sell its loans.

4.3 Provisioning Requirement for Future Losses

As in the simple stylized model, the provisioning requirement for
future loan losses is specified in terms of a requirement on LLRs.
The provisioning rate θst ∈ [0, 1] depends now on the aggregate
state. Under the assumption that the bank does not accumulate
defaulted loans in the form of NPLs and writes the losses off in the
same period they materialize, its loan loss reserves are given by

Rt = θst Lt . (14)

23
Assuming that a loan that defaulted during period t is resolved and is written
off during the same period t is a simplifying assumption, as it greatly reduces the
state space of the model. The consequence of this assumption is that the bank
does not accumulate NPLs—all defaulted loans are resolved and written down
immediately. For our purposes, this assumption is not restrictive because, as we
discuss it in the later section, the incurred and expected loss approaches treat
NPLs in the same way.
24
For example, the screening cost of processing new loan applications.
Vol. 20 No. 4 Still “Too Much, Too Late” 433

The law of motion of the bank’s loan loss reserves can then be
written as

Rt = Rt−1 − ξt λst Lt−1 + LLPt , (15)

where LLPt denotes the bank’s total loan loss provision: provisions
for incurred losses, ξt λst Lt−1 , and expected losses, Rt − Rt−1 .

4.4 Profits
The bank’s profits are given by loan repayments less interest
expense, operating expense, loan losses, and provisioning—that is,

πt := rsLt−1 (1 − ξt )Lt−1 − rt−1 Bt−1 − C(Nt ) − LLPt − ι. (16)

The first term above is the repayment on non-defaulted loans; the

second term is the interest expense on deposits, which repay risk-free
rate rt ; the third term is the loan adjustment costs associated with
new loans; the fourth term is the total loan loss provisions; the last
term, a constant ι, is the fixed cost of running the bank.
As in the simple stylized model, the bank’s profits are subject to
a convex corporate tax schedule

τ (πt ) = τ Iπt >0 , (17)

where Iπt >0 denotes an indicator function which is equal to one

when πt > 0 and zero otherwise. The bank’s net income—that is,
the after-tax profits—is given by (1 − τ (πt )) πt .25

4.5 Equity
The bank’s after-tax profits are either paid out as dividends or
retained to increase the stock of equity. Let Xt be a dividend pay-
out at time t; then the bank’s book equity evolves according to the
following accounting identity:

Et = Et−1 − Xt + (1 − τ (πt )) πt . (18)

25
If, however, provisions for future losses are not tax deductible, then the net
income is given by (1 − τ (πt + (Rt+1 − Rt ))) (πt +(Rt+1 −Rt ))−(Rt+1 −Rt ). We
proceed under the assumption that the provisions are tax deductible and state it
explicitly when it is not the case.
434 International Journal of Central Banking October 2024

Negative values of Xt mean that the bank is raising exter-

nal equity. We assume that raising external equity is costly. This
cost reflects the direct transactional costs (e.g., underwriter fees
(Altınkılıç and Hansen 2000)) and indirect costs of raising exter-
nal equity (i.e., debt overhang (Myers 1977 and Admati et al. 2018)
or signaling issues (Myers and Majluf 1984)). These costs do not
apply if banks retain earnings (in line with pecking-order theories).
Following Hennessy and Whited (2007), the cost of raising exter-
nal equity is modeled in a reduced form. In particular, for every
dollar raised in terms of equity, the bank will have to pay 1 + ηst ,
where ηst > 0 is a flotation cost for equity. Therefore, the cost of
external equity is given by
η(Xt ) := ηst Xt IXt <0 , (19)
where indicator function IXt <0 is equal to 1 when Xt < 0, and 0 oth-
erwise. Thus, η(Xt ) is strictly negative when the bank raises equity
and zero otherwise.

4.6 Capital Requirement

As in the simple stylized model, the bank is subject to the minimum
capital requirement. Every period t, the bank’s choice over the port-
folio of loans and equity must satisfy the following minimum capital
constraint
Et ≥ κst Lt , (20)
where κst ∈ [0, 1].26

4.7 Optimization Problem

The bank’s manager maximizes the present value of all future divi-
dends.27 The effective control variables are the next-period stock of

26
The current regulatory regime (i.e., Basel III) is the one with risk-based
capital requirements. Therefore, κst is an increasing function of loan default prob-
ability. We present the formula for κst in the calibration section. When constraint
(20) is not binding, we say that the bank holds a voluntary capital buffer.
27
It is straightforward to show that in our model maximizing the present value
of future dividends is equivalent to maximizing the present value of the future
free cash flows to equity.
Vol. 20 No. 4 Still “Too Much, Too Late” 435

equity, Et , and loans, Lt . The choice over these controls, in turn,

determines the bank’s dividend payout, Xt , and lending, Nt .
Formally, given the current state of the bank, Ξt = [Et−1 ,
Lt−1 , ξt , st , st−1 ], the bank’s manager maximizes the present value
of all future dividends net the cost of recapitalization subject to a
set of the constraints—that is, it solves
 
V (Ξt ) = max 0, Xt + η(Xt ) + βt E [V (Ξt+1 ) |st ] ,
{Et ,Lt }

subject to a) Et ≥ κst Lt ,
b) Lt − Rt = Bt + Et
c) Et = Et−1 − Xt + (1 − τ (πt )) πt ,
d) Lt = (1 − ξt ) (1 − δ) Lt−1 + Nt ,
e) πt = rsLt−1 (1 − ξt )Lt−1 − rt−1 Bt−1 − C(Nt )
− LLPt − ι,
e) Rt = (Rt−1 − ξt λst Lt−1 ) + LLPt ,
f ) Rt = θst Lt ,
g) Nt ≥ 0. (21)

The solution to the above problem is the policy functions

Et∗ : Ξt → R+ and L∗t : Ξt → R+ , which satisfy the above
system. Default takes place at time t when the bank finds itself insol-
vent. This happens when the sum of the bank’s current cash flows,
Xt + η(Xt ), and continuation value, βE [V (Ξt+1 ) |st ], is negative—
that is, when the limited liability constraint is binding.

5. Calibration

5.1 Loan Default Rate Distribution

Following Martinez-Miera and Repullo (2010), we assume that the
probability distribution of the aggregate default rate ξti is implied by
the single common risk factor model of Vasicek (2002). This spec-
ification allows for imperfectly correlated individual loan defaults:
the performance of an individual bank loan depends on the com-
mon and idiosyncratic factors, while the aggregate default rate ξti
436 International Journal of Central Banking October 2024

depends only on the common factor. Moreover, this specification is

adopted by the Basel Accords to provide a value-at-risk foundation
to the minimum capital requirements. Appendix B provides more
detailed information on this specification.
The CDF of ξti conditional on aggregate state is then given by
  
i 1 − ρist Φ−1 ξti − Φ−1 pist
F (ξt ; st ) = Φ , (22)
ρist

where Φ(.) is the standard normal CDF. We derive the distribu-

tion of ξti in Appendix B. Note, F (ξti ; st ) has two parameters: pist
and ρist ∈ (0, 1). The stage i loan default probability pist is identical
across all loans and is equal to the mean of ξti —that is, to E[ξt+1
i
|st ].
The loan default correlation ρst ∈ (0, 1) captures the dependence of
i

individual loan on the common risk factor and, thus, determines the
degree of correlation between individual loan defaults. To calibrate
the correlation coefficient ρist , we use the formula adopted by the
Basel framework (see Equation (B.5) in Appendix B).

5.2 Capital Requirement

The empirical counterpart of capital in our model is Tier 1 capital,
which primarily consists of common equity. Under the risk-based
approach of Basel capital regulation, κst is an increasing function of
loan default probability. We calibrate the capital requirement for a
bank that follows the internal ratings-based (IRB) approach. Most of
the largest banks adopt the IRB approach. Moreover, Basel Accords
specify an explicit formula for the capital requirement under the IRB
approach, which is a function of loan characteristics such as default
probability, maturity, and loan loss default rate. This allows us to
calibrate the minimum capital requirement so that it is consistent
with the characteristics of the bank’s loan portfolio. Under the IRB
approach, the capital requirement for corporate and bank exposures
is meant to ensure sufficient capital to cover loan losses with a con-
fidence level of 99.9 percent. The exact formula for κist is reported
in Equation (B.6) in Appendix B.
One of the defining elements of Basel III is the countercycli-
cal capital buffer (CCyB). The CCyB is a regulatory instrument
designed to smooth lending procyclicality, which requires banks to
Vol. 20 No. 4 Still “Too Much, Too Late” 437

build up an extra capital buffer during good times to increase their

loss-absorption capacity for bad times. Specifically, under the CCyB
a bank is required to hold the extra 2.5 percent of its risk-weighted
assets (RWA) in equity during an expansion. Practically, the release
and the accumulation of the CCyB should normally be implemented
stage-wise over some period of time. However, to keep our model
tractable, we assume that the CCyB is fully released once the aggre-
gate state deteriorates and it must be fully accumulated right upon
the improvement of the aggregate state. As such, for the calibration
purposes, we assume a smaller size of the CCyB, namely 1.5 per-
cent of its RWA. Thus, in our model, the CCyB is implemented by
raising the minimum capital requirement in expansion from κg to
1.188κg .28 We further provide robustness results when the required
CCyB is at 2.5 percent of RWA.

5.3 Provisioning Requirement

As discussed in Section 2, the IL approach does not allow recog-
nition of losses that are expected to happen in the future. How-
ever, aside from accounting provisions, banks that follow the IRB
approach must recognize one-year prudential expected losses on the
entire portfolio of loans. These prudential expected losses, however,
are computed in a different way than those under the EL approach.
In particular, the prudential expected losses are computed using the
so-called through-the-cycle (TTC) default probabilities and a con-
servative (downturn) estimate of loss given default. In our model,
the TTC default probabilities correspond to the unconditional on
the aggregate state default probabilities, p̄i , and the downturn esti-
mate of loss given default is given by λib . Thus, under the ILM the
bank recognizes a loss θIRB,i on a marginal loan i, where

θIRB,i = E[λib ξt+1 ] = λib p̄i . (23)

28
As per the Basel III’s formula for RWA, in our model these are given by
RW At =κst−1 12.5Lt at time t (see paragraph (53) of the section Internal Ratings-
Based Approach for Credit Risk in Basel Committee on Banking Supervision
2017). Therefore, the capital requirement conditional on an expansion increases
from κg to κg + 0.015 × 12.5κg = 1.188κg due to the CCyB.
438 International Journal of Central Banking October 2024

Thus, the provisioning rate for the entire portfolio of loans is given
by

θsIRB
t
= ωst θIRB,1 + (1 − ωst ) θIRB,2 . (24)

Note, however, that the prudential provisions are not accounting

losses and, thus, are not tax deductible.29
The discounted expected losses under the EL approach instead
employ the point-in-time (PiT) default probabilities—that is, the
expected losses are conditional on the current aggregate state. The
one-year discounted expected loss rate under the EL approach is
given by
1 1
θs1Y,i = E[λist+1 ξt+1 |st ] = ps E[λist+1 |st ], (25)
t
1 + d st 1 + d st t

while the lifetime discounted expected loss rate is given by

1
θsLT,i = E[λist+1 ξt+1 + (1 − ξt+1 )(1 − δ)θsLT,i |st ]. (26)
t
1 + d st t+1

The discount rate dst is equal to the contractual interest rate rsLt
under IFRS 9, while under the U.S. GAAP it is implied by the
bank’s discount factor βt . Note that in Equations (25) and (26), the
expectations are conditional on the aggregate state, which reflects
the PiT estimation of the EL approach. The closed-form solution to
Equation (26) is provided in Appendix B.
Since under IFRS 9 the bank recognizes one-year and lifetime
expected losses on stage 1 and stage 2 loans, respectively, the pro-
visioning rate for the entire portfolio of loans is given by

θsIF
t
RS9
= ωst θs1Y,1
t
+ (1 − ωst ) θsLT,2
t
. (27)

The provisioning rate for the entire portfolio of loans under the U.S.
GAAP is given by

θsGAAP
t
= ωst θsLT,1
t
+ (1 − ωst ) θsLT,2
t
, (28)

29
See Abad and Suarez (2018) for an in-depth discussion on prudential provi-
sions.
Vol. 20 No. 4 Still “Too Much, Too Late” 439

since in this case the lifetime expected losses are recognized on all
types of loans.
It is important to note that when we compare the IL and EL
approaches, we include the IRB provisions in both scenarios. That
is, regardless of the accounting approach the bank must still meet
the IRB requirement. We do, however, assume that if accounting
provisions for future loan losses are in excess of the prudential pro-
visioning (under IRB), then the regulator does not object to count
the accounting provisions for regulatory requirement of prudential
provisions.30

5.4 Parameter Values

The parameters of the model are listed in Table 1, along with their
values and sources of calibration. Below, we present a detailed sum-
mary of how we calibrate the model.
The model features aggregate and idiosyncratic uncertainty. We
set the transition probabilities for the aggregate state, qst ,st+1 , to
obtain contractions that last for 2 years, on average, and expan-
sions that last for 6.8 years, on average, which is consistent with the
National Bureau of Economic Research’s dating of business cycle.31
To that end, we set qg,g = 0.852 and qb,b = 0.5.32
Idiosyncratic uncertainty depends upon the aggregate state and
is captured by the loan default process. The absence of detailed
micro-level data on banks’ loan portfolios creates a problem for cal-
ibrating the bank loan default process. To circumvent this prob-
lem, we follow the approach in Abad and Suarez (2018) and use the
Global Corporate Default reports produced by Standard & Poor’s
(S&P) over the period 1981–2015 to calibrate the bank loan default
process. To that end, we set the default probability of stage 1
(2) loans to 0.54 percent (6.05 percent) and 1.9 percent (11.5 per-
cent) in expansion and contraction, respectively. These probabilities

30
This assumption ensures there is no double provisioning problem. For exam-
ple, the total provision rate under IFRS 9 is given by max{θsIF t
RS9
, θsIRB
t
}, which
may also require some adjustments to after tax profits since, unlike the EL
provisions, the IRB provisions are not tax deductible.
31
http://www.nber.org/cycles.html.
32
The unconditional probability of good and bad aggregate state are given by
qg = (1 − qb,b )/(2 − qg,g − qb,b ) = 0.77 and qb = 1 − qg = 0.23, respectively.
 Table 1. Model Parameters

Parameter Description Contraction Expansion Source

A. Parameters Set Outside the Model
440

Loan Default Process

λ1st Loss Given Default Rate Stage 1 Loans 40% 30%
λ2st Loss Given Default Rate Stage 2 Loans 40% 30% Standard & Poor’s
p1st Default Probability of Stage 1 Loans 1.90% 0.54% Abad and Suarez (2018)
p2st Default Probability of Stage 2 Loans 11.50% 6.05%
ρ1st Loan Default Correlation Stage 1 0.166 0.212
Loans
ρ2st Loan Default Correlation Stage 2 0.120 0.126 Equation (B.5)
Loans
Loan Portfolio and Repayment
1/δ Average Maturity of Loans (in Years) 5 5 Call Reports
ωst Fraction of Stage 1 Loans 0.81 0.85 Abad and Suarez (2018)
rsLt Loan Repayment Rate 5.00% 4.29% FRED: U.S. Net Interest Margins
Other Parameters
β Bank Discount Factor 0.95 0.95 De Nicolò, Gamba, and Lucchetta
(2014)
τ Corporate Tax Rate 0.20 0.20
ηs t Flotation Cost of Equity ∞ 0.06 De Nicolò, Gamba, and Lucchetta
International Journal of Central Banking

(2014), Dinger and Vallascas

(2016)
r Cost of Debt (Risk-Free Rate) 1.00% 1.00%
qst ,st+1 Transition Probability of the Aggregate 0.5 0.852 NBER Business Cycle Dating
State
B. Parameters Calibrated Inside the Model
ι Fixed Cost of Running the Bank 0.0045 0.0045 Calibrated to match the annual
bank failure rate
October 2024

φ Loan Adjustment Cost Parameter 0.60 0.60 Calibrated to match the loan
growth rate volatility

(continued )
 Table 1. (Continued)

Parameter Description Contraction Expansion Source

C. Residual Parameters
Capital Requirement
κ1st Minimum Capital Requirement for Stage 1 8.50% 8.50% IRB approach; see
Vol. 20 No. 4

Loans Equation (B.6)

κ2st Minimum Capital Requirement for Stage 2 14.4% 14.4%
Loans
κs t Total Minimum Capital Requirement 9.70% 9.40% Equation (B.7)
Provisioning Rates
θs1,ILM
t
Stage 1 Loan Provisioning Rate under ILM 0.34% 0.34% Equation (23)
(IRB)
θs2,ILM
t
Stage 2 Loan Provisioning Rate under ILM 2.92% 2.92%
(IRB)
θsILM
t
Average (Portfolio) Loan Provisioning Rate 0.84% 0.73% Equation (24)
under ILM (IRB)
RS9
θs1,IF
t
Stage 1 Loan Provisioning Rate under IFRS 9 0.44% 0.24% Equation (25)
2,IF RS9
θs t Stage 2 Loan Provisioning Rate under IFRS 9 8.84% 7.83% Equation (26)
RS9
θsIF
t
Average (Portfolio) Loan Provisioning Rate 2.06% 1.38% Equation (27)
under IFRS 9
θs1,GAAP
t
Stage 1 Loan Provisioning Rate under U.S. 1.35% 1.09% Equation (26)
GAAP
Still “Too Much, Too Late”

θs2,GAAP
t
Stage 2 Loan Provisioning Rate under U.S 8.68% 7.61%
GAAP
θsGAAP
t
Average (Portfolio) Loan Provisioning Rate 2.77% 2.07% Equation (28)
under IFRS 9
Note: This table summarizes the parameters of the model, their values, and the sources of their calibration. The values of some para-
meters vary with the aggregate state. Parameters listed in panel A have been calibrated outside the model due to the observability
of their data counterparts. Panel B lists the parameters which were calibrated inside the model to match the moments in the real
441

data; the data counterparts of these parameters are directly observable. Finally, panel C presents the residual parameters; the values
of these parameters are determined by the value of the parameters set outside the model.
442 International Journal of Central Banking October 2024

Figure 1. Loan Default Rate Density

Note: This figure plots the calibrated Vasicek density function for the aggregate
loan default rate ξti .

are consistent with the alignment of stage 1 loans with corporate

bonds with ratings AAA to BB in the S&P classification and stage
2 loans with ratings B to C. Furthermore, in line with Abad and
Suarez (2018), we set the fraction of stage 1 loans, ωst , to 0.85
and 0.81 in expansion and contraction, respectively. Finally, the loss
given default rates λist for both types of loans are set to 0.4 and
0.3 in contraction and expansion, respectively. Figure 1 depicts the
calibrated Vasicek density function of the aggregate default rate ξti .
Given the parameterization of the loan default process, we
can compute the implied values for the residual parameters:
(ρ1st , ρ2st , κ1st , κ2st , θs1t , θs2t ). Equation (B.5) from Appendix B implies
the value of the loan default correlation coefficient for stage 1
(2) loans, ρ1st (ρ2st ), is 0.166 (0.12) in contractions and 0.212 (0.126)
in expansions. Similarly, Equation (B.6) from Appendix B implies
the minimum capital requirement for stage 1 and 2 loans, κ1st and
κ2st , is 8.5 percent and 14.4 percent, respectively. The overall capital
requirement, κst , is then 9.4 percent in expansion and 9.7 percent
Vol. 20 No. 4 Still “Too Much, Too Late” 443

in contraction. Finally, the values of the provisioning rates, θst , are

assigned according to Equations (23)–(26).33
We set the interest rate on debt, r, to 1.0 percent. To calibrate
the interest on the loan portfolio, rsLt , we match the first and second
moments of the interest margins, rsLt − rt−1 , to their data counter-
parts. For the U.S. banks, the mean and the standard deviation
of the bank interest margins are 3.45 percent and 0.29 percent,
respectively.34 Imposing a restriction that rg < rb , which reflects
the pricing of risk, we thus set rsLt to 4.29 percent and 5.00 percent
in expansion and contraction, respectively. Finally, consistent with
the U.S. banks’ Reports of Condition and Income (Call Reports), we
set the average loan maturity to five years, which implies δ = 0.20.
The corporate tax rate, τ , is 0.20.35
Following De Nicolò, Gamba, and Lucchetta (2014), we set the
flotation cost of equity conditional on expansion, ηg , to 0.06 and the
banker’s discount factor, βt , to 0.95. Empirical evidence suggests
that banks seldom issue new equity during a downturn (Dinger and
Vallascas 2016). To accommodate this stylized fact, we make issuing
equity in a bad aggregate state prohibitively expensive setting ηb to
infinity—that is, we effectively impose a non-negativity constraint
on Xt when the aggregate state is bad.
The remaining two parameters, namely φ and ι, are calibrated
inside the model to match the relevant data moments. The parame-
ter φ, which is from the loan lending cost function, directly relates
to the volatility of bank loans. The higher φ is, the costlier it is
to increase the stock of loans through new lending. This in turn
lowers the volatility of loan growth. This parameter is thus cali-
brated to match the volatility of the annual loan growth rate, which

33
Note that since provisioning rates under the EL approaches exceed those
under the IRB approach, we assume that the regulator accepts accounting-
expected provisions under the EL approach as prudential provisions of the IRB
approach—that is, under the EL approach the bank no longer has to recognize
the prudential losses since those are already covered by the accounting provisions.
34
Data source: Federal Reserve Economic Database (FRED) time series on U.S.
banks’ net interest margins “USNIM.”
35
It is important to note that since this is a partial equilibrium model, where
the bank takes prices as given, we cannot comment on how provisioning require-
ments may affect loan demand, and therefore the indirect effect coming from loan
prices that are general equilibrium outcomes.
444 International Journal of Central Banking October 2024

Table 2. Data and Model Moments

under Benchmark Calibration

Moments Model Data Source

Matched Moments

Mean Interest Margins 3.45% 3.45% FRED (Time Series: “USNIM”)

St. D. Interest Margins 0.29% 0.29% FRED (Time Series: “USNIM”)
Bank Failure Rate 0.39% 0.37% Mankart, Michaelides, and
Pagratis (2020)
St. D. of Loan Growth 3.90% 3.70% FRED (Time Series: “TOTLL”)
Rate
Not Matched Moments

Mean Charge-Off Rate 0.66% 0.86% FDIC: Charge-Off and

Delinquency Rates
St. D. Charge-Off Rate 0.49% 0.60% On Loans and Leases at
Commercial Banks
Mean ROE 10.2% 11.2% FRED (Time Series: “USROE”)
Mean ROA 0.95% 0.99% FRED (Time Series: “USROA”)

Note: This table presents the matched and not matched moments implied by the
model and real data. The moments implied by the model are computed under the
benchmark—that is, the ILM case. The model moments are computed based on
the data from simulating the model for 80,000 periods and excluding the first 200
observations. The sources for the real data moments are representative of U.S. banks.

is about 3.8 percent in the data.36 To that end, we set φ = 0.6.

Finally, the fixed cost of running the bank, ι, needs to be set suffi-
ciently high to ensure that the bank’s profits are not too large and,
thus, the bank occasionally fails. We set ι to 0.0045 to match the
bank (annual) failure rate of 0.37 percent (Mankart, Michaelides,
and Pagratis 2020).
To assess the results of our calibration, we report some relevant
moments implied by our model under the benchmark case of the
ILM and the corresponding real data moments in Table 2. Despite
its parsimonious structure, the model matches the data moments
reasonably well.

36
Data source: FRED, Federal Reserve Bank of St. Louis, time series
“TOTLL”—Loans and Leases in Bank Credit, All Commercial 1975–2019.
Vol. 20 No. 4 Still “Too Much, Too Late” 445

Note that under the calibrated parameter values, our model pre-
dicts that the minimum capital requirement is binding all the time.37
We verify this numerically using the simulated data from the model
when the bank’s choice consists of both Et and Lt . Therefore, to
increase the precision of the numerical solution of our model, we
solve the model by imposing that the capital constraint is binding.
See Appendix C for more details on this matter.

6. Quantitative Results

6.1 Cyclical Effect of the EL Approach

First, we examine the effect of the EL approach on the cyclicality of
the key endogenous variables of the model, such as loan loss provi-
sions (LLPs), profits, lending, and bank failure rate. Table 3 reports
the moments of the endogenous variables, conditional on the aggre-
gate state. These are obtained by simulating the model for 80,000
periods under the three scenarios: the ILM (benchmark) and two
variations of the ELM, namely, IFRS 9 and U.S. GAAP.
The last two columns of Table 3, which provide a relative compar-
ison between ILM and ELM, suggest a profoundly large procyclical
effect of the EL approach on bank lending in our model. For exam-
ple, while, on average, the bank lending is about 3.6 percent (2.7
percent) lower under IFRS 9 (U.S. GAAP) than ILM, conditional
on a contraction the bank originates, on average, as much as 7.1 per-
cent (5.9 percent) fewer new loans under IFRS 9 (U.S. GAAP). The
procyclicality of the EL approach can also be assessed by examin-
ing the ratio of new loans to outstanding loans or loan growth rate.

37
The bank’s shareholders are relatively impatient (low β) and deposits are
cheap (due to mispriced deposit insurance and the tax deductibility of the inter-
est expense on deposits). Thus, holding an equity buffer is costly. That is why in
our calibration the bank does not optimally hold buffer. The benefit of holding
the buffer comes from the insurance it provides against having to increase equity
following a large loss. Thus, when equity is costly to issue, and when the proba-
bility of facing a high loss is sufficiently high, then the bank might find it optimal
to hold equity buffer. If we increase the likelihood of large losses and make the
cost of issuing equity prohibitively expensive even in good times (st = g), then
we could generate some capital buffers. However, the limited liability constraint
makes it much harder to obtain a voluntary capital buffer in the model—the abil-
ity to walk away from the bank with insufficient capital reduces the incentives of
the bankers to hold the buffer.
 Table 3. Expected Loss Approach: Cyclicality and Stability

Aggregate
RS9 SGAAP
446

Description Variable State ILM IFRS 9 U.S. GAAP ΔIF

ILM ΔU
ILM

Unconditional 0.69% 0.70% 0.69% 0.00 pp 0.00 pp

Rt −Rt−1 +λst ξt Lt−1
Total LLP Lt
Contraction 1.68% 1.94% 1.93% 0.26 pp 0.25 pp
Expansion 0.41% 0.33% 0.34% –0.07 pp –0.07 pp

Unconditional 1.21% 1.22% 1.22% 0.01 pp 0.02 pp

Profits πt /Lt−1 Contraction 0.48% 0.28% 0.29% –0.20 pp –0.18 pp
Expansion 1.42% 1.49% 1.50% 0.07 pp 0.08 pp

Unconditional 0.1288 0.1241 0.1253 –3.64% –2.74%

New Loans Nt Contraction 0.1219 0.1132 0.1148 –7.13% –5.85%
Expansion 0.1308 0.1273 0.1283 –2.69% –1.89%

Unconditional 0.5983 0.5766 0.5819 –3.63% –2.74%

Total Loans Lt Contraction 0.5764 0.5497 0.5555 –4.64% –3.63%
Expansion 0.6047 0.5844 0.5897 –3.35% –2.48%

New Loans/ Unconditional 21.60% 21.60% 21.60% 0.00 pp 0.00 pp

Outstanding Nt /Lt−1 Contraction 20.78% 20.14% 20.21% –0.64 pp –0.57 pp
Loans Expansion 21.87% 22.07% 22.05% 0.20 pp 0.18 pp

Unconditional 0.00% 0.00% 0.00% 0.00 pp 0.00 pp

International Journal of Central Banking

Loan Growth Rate ΔLt Contraction –2.08% –2.73% –2.71% –0.66 pp –0.63 pp
Expansion 0.77% 0.97% 0.95% 0.20 pp 0.18 pp

Unconditional 0.39% 0.49% 0.35% 0.10 pp –0.04 pp

Failure Rate P(Vt < 0) Contraction 1.73% 2.16% 1.54% 0.43 pp –0.18 pp
Expansion 0.00% 0.00% 0.00% 0.00 pp 0.00 pp

Note: This table presents unconditional and conditional on the aggregate state first moments of various endogenous variables implied
by the model under the incurred (ILM) and expected (IFRS 9 and U.S. GAAP) loss approach. The model is solved and simulated
October 2024

under the ILM (benchmark), IFRS 9, and U.S. GAAP provisioning approaches. The moments are constructed based on simulating
the model for 80,000 periods (with the first 200 observations excluded). Δij denotes the relative difference between approach i and j,
which is measured either in terms of percentage (%) or percentage points (pp).
Vol. 20 No. 4 Still “Too Much, Too Late” 447

Both of these measures also indicate increased procyclicality under

the EL approach.
Table 3 also helps to understand the mechanics behind the pro-
cyclicality of the EL approach. As we discussed earlier, the PiT
default probabilities of the EL approach amplify the cyclical move-
ment of the total LLPs. This, in turn, raises the volatility of the
bank’s profits over the cycle, which can be seen in Table 3. With the
combination of the minimum capital requirement and costly exter-
nal equity issuance, increased profit volatility translates into more
severe lending procyclicality.
Our model further predicts that IFRS 9 is slightly more procycli-
cal than U.S. GAAP. While the volatility of the provisioning rate,
θst , under both variants of the EL approach is roughly the same,
which is about 0.29 percent, because of its mixed-horizon approach,
IFRS 9 produces smaller loan loss reserves (LLRs), thus providing
a weaker loss-absorption capacity than U.S. GAAP. As a result, the
bank lends more procyclically under IFRS 9. Relatedly, our model
also predicts that U.S. GAAP does a better job in terms of improving
bank stability than IFRS 9. In fact, we find that with the adoption
of IFRS 9, the bank failure rate may even increase. Again, this has
to do with IFRS 9 being characterized by both larger procyclicality
and lower loss-absorption capacity relative to U.S. GAAP. Recall
that, on the one hand, since the bank holds larger LLRs under the
EL approach, it should lower the bank failure rate. On the other
hand, the procyclicality of the EL approach effectively increases the
volatility of banks’ profits, which increases bank failure rate.
Next, we examine the cyclical implication of the EL approach
when the bank is subject to the CCyB. The CCyB is one of the
most prominent features of Basel III that has been introduced to
combat lending procyclicality by requiring the banks to hold extra
capital during good times to support their lending activities upon
the arrival of a contraction. Thus, it is particularly important to
assess the joint cyclical effect of the EL approach and the CCyB.
To do so, we report in Table 4 the moments of various endogenous
variables under the ILM (benchmark), IFRS 9, and U.S. GAAP sce-
narios (conditional on the aggregate state), when the bank is subject
to the CCyB.
First, Table 4 predicts a sizable effect of the CCyB on bank lend-
ing. By comparing the first (“ILM”) and second (“ILM+CCyB”)
 448

Table 4. Expected Loss Approach and the CCyB: Cyclicality and Stability

Aggregate ILM+ IFRS 9 U.S. GAAP

ILM −CCyB IF RS9+CCyB U SGAAP +CCyB
Description Variable State ILM CCyB + CCyB + CCyB ΔILM ΔILM ΔILM

New Loans Nt Unconditional 0.1288 0.1260 0.1232 0.1225 –2.16% –4.34% –4.88%
Contraction 0.1219 0.1251 0.1176 0.1179 2.61% –3.49% –3.27%
Expansion 0.1308 0.1263 0.1248 0.1238 –3.45% –4.57% –5.31%

Total Loans Lt Unconditional 0.5983 0.5853 0.5723 0.5691 –2.16% –4.34% –4.88%
Contraction 0.5764 0.5719 0.5529 0.5510 –0.79% –4.09% –4.41%
Expansion 0.6047 0.5893 0.5780 0.5744 –2.55% –4.42% –5.01%

New Loans/ Nt /Lt−1 Unconditional 21.60% 21.60% 21.60% 21.60% 0.00 pp 0.00 pp 0.00 pp
Outstanding Contraction 20.78% 21.65% 20.93% 21.07% 0.87 pp 0.15 pp 0.30 pp
Loans Expansion 21.87% 21.58% 21.80% 21.75% –0.29 pp –0.07 pp –0.12 pp

Loan Growth ΔLt Unconditional 0.00% 0.00% 0.00% 0.00% 0.00 pp 0.00 pp 0.00 pp
Rate Contraction –2.08% –1.32% –2.06% –1.92% 0.76 pp 0.02 pp 0.15 pp
Expansion 0.77% 0.48% 0.70% 0.65% –0.29 pp –0.07 pp –0.12 pp

Failure Rate Default Unconditional 0.39% 0.17% 0.16% 0.12% –0.22 pp –0.23 pp –0.27 pp
International Journal of Central Banking

Contraction 1.73% 0.76% 0.69% 0.55% –0.97 pp –1.04 pp –1.18 pp

Expansion 0.00% 0.00% 0.00% 0.00% 0.00 pp 0.00 pp 0.00 pp

Note: This table presents unconditional and conditional on the aggregate state first moments of various endogenous variables implied by the
model under the incurred (ILM) and expected (IFRS 9 and U.S. GAAP) loss approach with and without the countercyclical capital buffer
(CCyB). The model is solved and simulated under the ILM (benchmark) with and without the CCyB, IFRS 9 with the CCyB, and U.S. GAAP
with the CCyB. The moments are constructed based on simulating the model for 80,000 periods (with the first 200 observations excluded).
Δij denotes the relative difference between approach i and j, which is measured either in terms of percentage (%) or percentage points (pp).
October 2024
Vol. 20 No. 4 Still “Too Much, Too Late” 449

columns in the table, we can see that while unconditional on the

aggregate state, the bank originates about 2 percent fewer loans, it
is able to increase its lending in a contraction by about 2.6 percent,
on average. Thus, the model suggests that the countercyclical cap-
ital buffer quantitatively smooths aggregate loan dynamics and it
also attenuates bank failures in the contractionary aggregate state.
Second, the last two columns in Table 4 suggest that the CCyB is
indeed able to considerably dampen the procyclical effect of the EL
approach on bank lending. In particular, when the bank is subject
to both IFRS 9 (U.S. GAAP) and the CCyB, it originates, on aver-
age, about 3.5 percent (3.3 percent) fewer new loans in a contraction
compared to that under the ILM without the CCyB. Nevertheless,
our model does suggest that the introduction of the EL approach
considerably reduces the efficacy of the CCyB to smooth lending
dynamics over the cycle.

6.2 Effect of the Arrival of a Contraction

Next, we examine a dynamic (i.e., multi-period) response of the key
endogenous variables of our model to the arrival of a contraction.
Since our model is solved fully non-linearly, we analyze the dynamic
response to the arrival of a contraction using the generalized impulse
response analysis (Koop, Pesaran, and Potter 1996).
A generalized impulse response of the variable Yt to a contrac-
tionary aggregate shock at time t = 0 constitutes a sequence of
conditional expectations of the form Yi = E[Yi |L−1 , s−1 = g, s0 = b]
for i = 0, 1, 2 . . . The bank’s endogenous initial state is set to the
average values in an expansion—that is, L−1 = E[Lt |st = g]. The
condition s−1 = g implies that prior to the arrival of a contraction
the bank is in expansionary aggregate state. Appendix C provides
more details on the numerical procedure to evaluate {Yi }T0 .
Figure 2 depicts the impulse response functions of the total LLPs,
profits, and new and outstanding loans to the arrival of a contrac-
tion at t = 0. The impulse response functions are evaluated under
the following three scenarios: ILM (benchmark), IFRS 9, and U.S.
GAAP. Panel A shows that the total LLPs react much stronger to
the arrival of a contraction under the EL approach. When learn-
ing about the deterioration of the aggregate state, the bank revises
its point-in-time estimates of default probabilities upward and has
450 International Journal of Central Banking October 2024

Figure 2. Effect of the Arrival of a Contraction

Note: This figure depicts the (generalized) impulse response functions of total
loan loss provisions, profits, and new and outstanding loans to the arrival of
a contraction at date t = 0. The impulse response functions are presented for
three scenarios: incurred loss model (ILM) and two variants of the EL approach,
namely, IFRS 9 and U.S. GAAP. The impulse response of outstanding loans
is in terms of relative changes to the (unconditional) mean E[L]. The impulse
responses are evaluated by taking the average across 30,000 simulated paths of
the variables of interest, with each path having the length of 11 periods and the
identical initial condition (L−1 , s−1 ). The initial condition is such that prior to
t = 0 the bank is in an expansionary aggregate state, s−1 = g, and its endoge-
nous state is given by the conditional on an expansionary aggregate state mean
value—that is, L−1 = E[Lt |st = g].

to abruptly front-load the increased expected losses. As a result,

the bank’s profits drop sharper under the EL approaches upon the
arrival of a contraction, as can be seen in panel B. Since the bank
is subject to the minimum capital requirement and issuing external
equity is too costly, we see in panel C that the bank cuts on new
loans more aggressively under the EL approaches. As a result, the
Vol. 20 No. 4 Still “Too Much, Too Late” 451

bank’s outstanding loans plunge deeper under the EL approaches

relative to their unconditional mean, as can be seen in panel D. We
show that these results are qualitatively robust to setting the cost
of external equity issuances to zero (Figure A.2), reducing an aver-
age duration of a contraction (Figure A.3), and imposing symmetric
transition probabilities of the aggregate state (Figures A.4).
To gain better understanding of the dynamics, Figure 3 depicts
the (generalized) impulse response functions (in red) and their confi-
dence bounds (in the shades of blue) of total LLPs, profits, and new
and outstanding loans to the arrival of a contraction at date t = 0.
The confidence bounds of the impulse responses are computed as
25–75, 10–90, and 5–95 percentiles of the response to the arrival of
a contraction. This figure allows us to see that not only do the EL
approaches worsen the procyclicality of bank lending, but they also
increase the volatility of the responses, which can be seen from the
widened bounds.
To follow up on our earlier discussion about the two channels of
the provisioning requirement for future losses, we try to disentangle
the procyclical effects of the EL approaches that are due to the tax
deductibility and the minimum capital regulation. Figure 4 plots
the impulse response functions of the new and outstanding loans
under the two EL approaches when the expected provisions are tax
deductible and when they are not. As seen from the figure, lending
procyclicality is only modestly increased when the tax deductibility
of the expected provisions is assumed. Thus, we conclude that the
procyclicality of an EL approach comes primarily from the capital
regulation channel rather than the tax channel.
Lastly, we examine the impulse responses to the arrival of a con-
traction when the bank is subject to the CCyB. Figure 5 depicts the
impulse response functions of profits, and new and outstanding loans
to the arrival of a contraction at t = 0 when the bank must hold the
CCyB of 1.5 percent.38 First, by comparing the impulse responses
of the new and outstanding loans under the ILM with and without
the CCyB, we note a quantitatively strong countercyclical effect of
the CCyB on bank lending. For example, the on-impact effect of
the contraction on the ratio of new to total loans improves from

38
Note, we do not plot the impulse response of the LLPs since their response
is not directly affected by the CCyB.
452 International Journal of Central Banking October 2024

Figure 3. Effect of the Arrival of

a Contraction: Confidence Bounds

Note: This figure depicts the (generalized) impulse response functions (in red)
and their confidence bounds (in shades of blue) of total loan loss provisions, prof-
its, and new and outstanding loans to the arrival of a contraction at date t = 0.
The impulse response functions are presented for three scenarios: incurred loss
model (ILM) and two variants of the EL approach, namely, IFRS 9 and U.S.
GAAP. The impulse response of outstanding loans is in terms of relative changes
to the (unconditional) mean E[L]. The impulse responses are evaluated by taking
the average across 30,000 simulated paths of the variables of interest, with each
path having the length of 11 periods and the identical initial condition (L−1 , s−1 ).
The initial condition is such that prior to t = 0 the bank is in an expansionary
aggregate state, s−1 = g, and its endogenous state is given by the conditional on
an expansionary aggregate state mean values—that is, L−1 = E[Lt |st = g].

20.25 percent to 21.2 percent. At the same time, panel A shows that
even when the introduction of an EL approach is accompanied by
the simultaneous adoption of the CCyB, new lending falls sharper
Vol. 20 No. 4 Still “Too Much, Too Late” 453

Figure 4. Effect of the Arrival of a Contraction:

Tax vs. Capital Regulation Channel

Note: This figure depicts the (generalized) impulse response functions of new
and outstanding loans to the arrival of a contraction at date t = 0 when the
expected provisions are tax deductible (blue and yellow plots) and when they
are not (red and purple plots). The impulse response functions are presented for
two variants of the EL approach: IFRS 9 and U.S. GAAP. The impulse response
of outstanding loans is in terms of relative changes to the (unconditional) mean
E[L]. The impulse responses are evaluated by taking the average across 30,000
simulated paths of the variables of interest, with each path having the length of
11 periods and the identical initial condition (L−1 , s−1 ). The initial condition is
such that prior to t = 0 the bank is in an expansionary aggregate state, s−1 = g,
and its endogenous state is given by the conditional on an expansionary aggregate
state mean values—that is, L−1 = E[Lt |st = g].

on impact. Figure 6 depicts the same impulse response when the

CCyB is set at a higher level of 2.5 percent of RWA.39 In this case,
while the procyclical effect of the EL approach is largely mitigated

39
Since in our model the CCyB is fully released once the aggregate state deterio-
rates and must be fully accumulated right upon the improvement of the aggregate
state, a higher level of the CCyB suppresses lending activity following the net
period after the shock.
454 International Journal of Central Banking October 2024

Figure 5. Effect of the Arrival of a Contraction

under the 1.5 Percent CCyB

Note: This figure depicts the (generalized) impulse response functions of new
and outstanding loans, and profits to the arrival of a contraction at date t = 0.
The impulse response functions are presented for four scenarios: incurred loss
model (ILM) with and without the 1.5 percent CCyB, and two variants of the
EL approach, namely, IFRS 9 and U.S. GAAP with the CCyB. The impulse
response of outstanding loans is in terms of a relative changes to the (uncondi-
tional) mean E[L]. The CCyB is characterized by an increase in the minimum
capital requirement conditional on the aggregate state being good. The impulse
responses are evaluated by taking the average across 30,000 simulated paths of
the variables of interest, with each path having the length of 11 periods and the
identical initial condition (L−1 , s−1 ). The initial condition is such that prior to
t = 0 the bank is in an expansionary aggregate state, s−1 = g, and its endoge-
nous state is given by the conditional on an expansionary aggregate state mean
values—that is, L−1 = E[Lt |st = g].
Vol. 20 No. 4 Still “Too Much, Too Late” 455

Figure 6. Effect of the Arrival of a Contraction

under the 2.5 Percent CCyB

Note: This figure depicts the (generalized) impulse response functions of new
and outstanding loans, and profits to the arrival of a contraction at date t = 0.
The impulse response functions are presented for four scenarios: incurred loss
model (ILM) with and without the 2.5 percent CCyB, and two variants of the
EL approach, namely, IFRS 9 and U.S. GAAP with the CCyB. The impulse
response of outstanding loans is in terms of a relative changes to the (uncondi-
tional) mean E[L]. The CCyB is characterized by an increase in the minimum
capital requirement conditional on the aggregate state being good. The impulse
responses are evaluated by taking the average across 30,000 simulated paths of
the variables of interest, with each path having the length of 11 periods and the
identical initial condition (L−1 , s−1 ). The initial condition is such that prior to
t = 0 the bank is in an expansionary aggregate state, s−1 = g, and its endoge-
nous state is given by the conditional on an expansionary aggregate state mean
values—that is, L−1 = E[Lt |st = g].
456 International Journal of Central Banking October 2024

(at least in the case of U.S. GAAP), this comes at a large contrac-
tion in the average level of outstanding loans of 3–4 percent, which
we do not report in the paper.

6.3 Effect of the Arrival of a Prolonged Contraction

We have so far shown that having to recognize the bulk of expected
losses at the start of a contraction can impede the bank’s lending.
It does, however, improve the bank’s loss-absorption capacity in the
following periods, thus allowing the bank to lend more later on. It
is natural then to examine the bank’s lending activities when a con-
traction persists for a longer period. Thus, we proceed to examine
the lending response to a prolonged contraction that lasts for at least
two periods—that is, a contraction that arrives at t = 0 and persists
at least until t = 1.
Figure 7 reports the impulse responses of the new and out-
standing loans to the arrival of the prolonged contraction at t = 0
under the same three provisioning approaches, with and without the
CCyB. The figure shows that the EL approach produces less pro-
cyclicality during the later stages of a contraction, which is consistent
with our intuition outlined above.

6.4 Effect of the Arrival of a Contraction under a Delayed

Response of Balance Sheet to Aggregate Shock
So far we have maintained an assumption that the arrival of a con-
traction has a contemporaneous effect on the distribution of the
bank’s losses. Under this assumption, the EL approach effectively
implies a “double blow” to the bank’s profitability, as both realized
and expected losses increase simultaneously upon the deterioration
of the aggregate state. This assumption can be questioned since
empirical evidence suggests that banks often report positive prof-
its at the start of a recession.40 Likewise, there might be some time
lag between the deterioration of the aggregate state and an increase
in consumer and corporate defaults. To account for this, we mod-
ify the timing of our model so that the distribution of the bank’s

40
For example, the return on average assets for U.S. banks was positive in 2007.
Even Lehman Brothers reported a net income of a record $4.2 billion in 2007.
Vol. 20 No. 4 Still “Too Much, Too Late” 457

Figure 7. Effect of the Arrival of a Prolonged

Contraction with and without the CCyB

Note: This figure depicts the (generalized) impulse response functions of new
and outstanding loans to the arrival of a contraction at date t = 0 that persists
for at least two periods—that is, s0 = s1 = b. The impulse response functions
are presented for six scenarios: incurred loss model (ILM) with and without the
CCyB, and two variants of the EL approach, namely, IFRS 9 and U.S. GAAP
with and without the CCyB. The impulse response of outstanding loans is in
terms of a relative changes to the (unconditional) mean E[L]. The CCyB is char-
acterized by an increase in the minimum capital requirement conditional on the
aggregate state being good. The impulse responses are evaluated by taking the
average across 30,000 simulated paths of the variables of interest, with each path
having the length of 11 periods and the identical initial condition (L−1 , s−1 ).
The initial condition is such that prior to t = 0 the bank is in an expansionary
aggregate state, s−1 = g, and its endogenous state is given by the conditional on
an expansionary aggregate state mean values—that is, L−1 = E[Lt |st = g].

losses responds with a one-period delay to the arrival of a contrac-

tion. While a one-year delay is admittedly an exaggeration, it should
really be viewed as an upper bound for the length of such a delay.
458 International Journal of Central Banking October 2024

Table 5. Provisioning Rates under

Delayed Response of Profits

ILM IFRS 9 U.S. GAAP

Stage 1 in Expansion θg1 0.34% 0.16% 1.19%

Stage 1 in Contraction θb1 0.34% 0.72% 1.99%
Stage 2 in Expansion θg2 2.92% 7.37% 8.38%
Stage 2 in Contraction θb2 2.92% 10.36% 11.54%
Average (Portfolio) in Expansion θg 0.73% 1.24% 2.27%
Average (Portfolio) in Contraction θb 0.84% 2.59% 3.83%
Difference across Aggregate State θb − θg 0.11% 1.35% 1.56%

Note: This table reports the calibrated values of the provisioning rates, θst (%),
under various provisioning approaches when the bank’s losses respond with a one-year
delay to the change in the aggregate state.

Formally, we denote the aggregate loan default rate at t by ξˆt

and assume that its distribution is given by

ξˆt ∼ F (ξˆt ; st−1 ) (29)

so that the distribution of the loss rate at time t now depends on the
previous-period aggregate state. This effectively implies that from
the point of view of the current period, the next-period losses are
determined up to the aggregate state.
Since the delay in the response of loan loss distribution affects
the conditional expected losses, we have to recompute provision-
ing rates under the EL approach. The one-year discounted expected
loss for a loan stage i is now given by θ̂s1Y,i t
= 1/(1 + dst )λist pist ,
whereas the lifetime discounted expected loss is given recursively by
i
θ̂sLT,i
t
= 1/(1 + dst )(λist pist + (1 − pist )(1 − δ)E[θ̂sLT,
t+1
|st ]).
Table 5 presents the recalibrated provisioning rates under IFRS 9
and U.S. GAAP, which reflects the modified version of the loan loss
distribution in Equation (29). As a result of this modification, the
provisioning rates under IFRS 9 and GAAP become even more coun-
tercyclical than before. Intuitively, the delay in the response of loan
losses implies that the bank anticipates the next-period losses bet-
ter; therefore, it provisions for expected losses more when times are
bad and less when they are good.
Figure 8 depicts the generalized impulse responses of the total
LLPs, and new and outstanding loans key to the arrival of a
Vol. 20 No. 4 Still “Too Much, Too Late” 459

Figure 8. Effect of the Arrival of a

Contraction under Delayed Losses

Note: This figure depicts the (generalized) impulse response functions of total
loan loss provisions and new and outstanding loans to the arrival of a contrac-
tion at date t = 0 when the bank’s losses respond with a one-year delay to the
change in the aggregate state. The impulse response functions are presented for
three scenarios: incurred loss model (ILM) and two variants of the EL approach,
namely, IFRS 9 and U.S. GAAP. The impulse response of outstanding loans is
in terms of a relative changes to the (unconditional) mean E[L]. The impulse
responses are evaluated by taking the average across 30,000 simulated paths of
the variables of interest, with each path having the length of 11 periods and the
identical initial condition (L−1 , s−1 ). The initial condition is such that prior to
t = 0 the bank is in an expansionary aggregate state, s−1 = g, and its endoge-
nous state is given by the conditional on an expansionary aggregate state mean
values—that is, L−1 = E[Lt |st = g].
460 International Journal of Central Banking October 2024

contraction at t = 0 under the ILM, IFRS 9, and U.S. GAAP when

the bank’s losses respond with a one-year delay to the change in
the aggregate state. Overall our results suggest that even when the
arrival of a contraction erodes the bank’s balance sheet with a one-
year delay, which allows the bank to anticipate the upcoming losses,
the ELM is still more procyclical than the ILM, at least, on impact.
Intuitively, as the bank learns about the increase in its expected
losses, it must recognize them. This erodes the bank’s profits and
forces it to cut new loans more than under the ILM. However, this
time the effect is not as pronounced, because the current losses are
smaller and, thus, they do not lower the current profits too much.
Thus, the procyclicality of the ELM is now partially mitigated, as
the bank recognizes the bulk of expected losses before these losses
actually realize. This allows the bank to originate more new loans
in the following periods after the arrival of a contraction.
Nevertheless, even with the delayed response of losses, the EL
approach produces a stronger procyclical effect on lending upon
the arrival of a contraction than the ILM. If we were to allow a
delayed response over a number of years, which would be equivalent
to assuming that the bank could anticipate or forecast the increase
in expected losses well in advance, then in this case the EL would
be able to smooth lending procyclicality. Intuitively, in this case, the
EL approach would be very similar to the CCyB, as it would allow
the bank to build up the reserves well in advance of the arrival of
a contraction. However, neither theoretically nor empirically would
it be possible to justify either such long delays in loss responses or
such great ability of banks to foresee the future losses.

7. Conclusion

Our quantitative dynamic model of a bank predicts that under the

expected provisioning approach of IFRS 9 and the new U.S. GAAP,
banks lend substantially more procyclically than under the incurred
loss approach. Moreover, the procyclicality of the EL approach
can worsen the bank’s stability despite providing an extra loss-
absorption capacity. Naturally, the ultimate question is whether the
ELM is better or worse than the ILM. Our partial equilibrium model
cannot answer this question, as it does not allow for welfare analysis.
Vol. 20 No. 4 Still “Too Much, Too Late” 461

However, the literature on optimal capital regulation helps to shed

some light on that issue.
It is rather well understood that the countercyclicality of capital
requirements is likely to amplify the business cycle (Kashyap and
Stein 2004; Repullo, Saurina, and Trucharte 2010). Moreover, most
scholars advocate in favor of procyclical capital requirements—that
is, to tighten the capital requirement during good times (Kashyap
and Stein 2004, Dewatripont and Tirole 2012, Repullo 2013,
Gersbach and Rochet 2017, and Malherbe 2020). In line with these
studies, the increased procyclicality of bank lending under the ELM
would lead to welfare losses.
The welfare analysis of expected provisioning is further compli-
cated by one important aspect of the ELM that we do not con-
sider in our analysis—that is, the information content of expected
provisions. One of the potential benefits of the ELM as argued in
Financial Stability Forum (2009) is that it “is consistent with finan-
cial statement users’ needs for transparency regarding changes in
credit trends.” When there is asymmetric information such that the
bank insiders know more about the state of the bank than other
market participants, expected provisions, provided that they are
properly estimated and truthfully disclosed, can be informative for
the outsiders. Therefore, any potential cost of expected provisioning
should be further compared to a potential benefit it may create by
increasing the transparency about the credit risk of the bank. How-
ever, the mere fact that there will be more information disclosed
under the ELM does not necessarily translate into welfare benefits
either. There is theoretical literature suggesting that more trans-
parency is not necessarily beneficial (Goldstein and Sapra 2014).
Mahieux, Sapra, and Zhang (2023) show that expected provisioning
may improve efficiency by allowing for timely regulatory intervention
to curb inefficient ex post asset substitution. They also argue that
banks, however, may respond to timely intervention by originating
riskier loans so that timely intervention induces timelier risk-taking.

Appendix A. Proof of Proposition 2

The first statement of the proposition is proved by taking the

derivatives of L0 and L1|z with respect to θ0 . Recall that we let
462 International Journal of Central Banking October 2024

θg := θ1|g = θ0 > 0, while θb := θ1|b = θ0 + , where  > 0. Using

Equations (9) and (10), it follows that
d e Le0
L0 = − < 0, (A.1)
dθ0 θ 0 + κ0
while
d e Le1|z dLe0 (1 − τ (π1 ))
L1|z = e + Le
dθ0 L0 dθ0 κ1 + θ1|z (1 − τ (π1 )) 0
(1 − τ (π1 ))
− Le . (A.2)
κ1 + θ1|z (1 − τ (π1 )) 1|z

The last term in the above equation is positive, and it is only present
when z = b. Thus, to prove that dθd0 Le1|z < 0, it is sufficient to show
that the sum of the first two terms in Equation (A.2) are negative,
which is the case since
Le1|z dLe0 (1 − τ (π1 ))
+ Le =
Le0 dθ0 κ1 + θ1|z (1 − τ (π1 )) 0
κ0 + (1 − τ (π1 ))(θ0 + r1L ) Le0 (1 − τ (π1 ))Le0
− + ∝
κ1 + θ1|z (1 − τ (π1 )) θ0 + κ0 κ1 + θ1|z (1 − τ (π1 ))
(1 − τ (π1 ))(θ0 + κ0 ) − κ0 − (1 − τ (π1 ))(θ0 + r1L ) =
−τ (π1 )κ0 − (1 − τ (π1 ))r1L < 0.

The second statement of the proposition is proved by taking the

derivative of L1|g − L1|b with respect to —that is,

d d (1 − τ (π1 ))L1|b e
Le1|g − Le1|g = − Le1|b = > 0. (A.3)
d d κ1 + θb (1 − τ (π1 ))
To prove the last statement of the proposition, first, write the prob-
ability of bank failure at t = 2 as

P2 := P (X2 < 0|z)

 
= P (1 − τ (π2 )) (r2L + θ1|z ) + κ1 ) < 0|z
 
κ1
= P r2 < −
L
− θ1|z |z
1 − τ (π2 )
Vol. 20 No. 4 Still “Too Much, Too Late” 463

Figure A.1. Distributions of the

Key Endogenous Variables

Note: This figure depicts distributions of the model’s key endogenous variables
obtained from simulating the model under the incurred loss approach for 80,000
periods (with the first 200 observations excluded).
 
r2L − μ1
=P < −T1 |z = Φ (−T1 ) = 1 − Φ (T1 ) ,
σ1
κ1
+θ1|z +μ1
where T1 = 1−τ (π2 )σ1 and Φ(.) denotes the standard normal
CDF. Then it follows that
d φ(T1 )
P2 = − < 0, (A.4)
dθ1|z σ1

where φ(.) denotes the standard normal density function.

464 International Journal of Central Banking October 2024

Figure A.2. Effect of the Arrival of a Contraction when

Cost of External Equity Issuance Is Zero

Note: This figure depicts the (generalized) impulse response functions of total
loan loss provisions, profits, and new and outstanding loans to the arrival of a
contraction at date t = 0 when the cost of external equity issuance are set to
zero—that is, ηst = 0. The impulse response functions are presented for three sce-
narios: incurred loss model (ILM) and two variants of the EL approach, namely,
IFRS 9 and U.S. GAAP. The impulse response of outstanding loans is in terms
of relative changes to the (unconditional) mean E[L]. The impulse responses are
evaluated by taking the average across 30,000 simulated paths of the variables
of interest, with each path having the length of 11 periods and the identical ini-
tial condition (L−1 , s−1 ). The initial condition is such that prior to t = 0 the
bank is in an expansionary aggregate state, s−1 = g, and its endogenous state is
given by the conditional on an expansionary aggregate state mean values—that
is, L−1 = E[Lt |st = g].

d
Similarly, one proves dθ0 P1 < 0. First, write

P1 := P (X1 < 0)
 
= P (1 − τ (π1 )) (r1L + θ0 ) + κ0 ) < 0
Vol. 20 No. 4 Still “Too Much, Too Late” 465

Figure A.3. Effect of the Arrival of a

Lower Persistency Contraction

Note: This figure depicts the (generalized) impulse response functions of total
loan loss provisions, profits, and new and outstanding loans to the arrival of
a contraction at date t = 0 when the probability of remaining in a contrac-
tion is 0.2, implying that an average duration of a contraction is 1.25 year. The
impulse response functions are presented for three scenarios: incurred loss model
(ILM) and two variants of the EL approach, namely, IFRS 9 and U.S. GAAP.
The impulse response of outstanding loans is in terms of relative changes to the
(unconditional) mean E[L]. The impulse responses are evaluated by taking the
average across 30,000 simulated paths of the variables of interest, with each path
having the length of 11 periods and the identical initial condition (L−1 , s−1 ).
The initial condition is such that prior to t = 0 the bank is in an expansionary
aggregate state, s−1 = g, and its endogenous state is given by the conditional on
an expansionary aggregate state mean values—that is, L−1 = E[Lt |st = g].
 
κ0
=P <− r1L − θ0 |z
1 − τ (π1 )
 L 
r1 − μ0
=P < −T0 = Φ (T0 ) = 1 − Φ (T0 ) ,
σ0
κ0
1−τ (π1 )
+θ0 +μ0
where T0 = σ0 .
466 International Journal of Central Banking October 2024

Figure A.4. Effect of the Arrival of a Contraction under

Symmetric Distribution of Aggregate States

Note: This figure depicts the (generalized) impulse response functions of total
loan loss provisions, profits, and new and outstanding loans to the arrival of a
contraction at date t = 0 when the probability of remaining in either a contrac-
tion or an expansion is 0.5. The impulse response functions are presented for
three scenarios: incurred loss model (ILM) and two variants of the EL approach,
namely, IFRS 9 and U.S. GAAP. The impulse response of outstanding loans
is in terms of relative changes to the (unconditional) mean E[L]. The impulse
responses are evaluated by taking the average across 30,000 simulated paths of
the variables of interest, with each path having the length of 11 periods and the
identical initial condition (L−1 , s−1 ). The initial condition is such that prior to
t = 0 the bank is in an expansionary aggregate state, s−1 = g, and its endoge-
nous state is given by the conditional on an expansionary aggregate state mean
values—that is, L−1 = E[Lt |st = g].

Then it follows immediately that

d φ(T0 )
P1 = − < 0. (A.5)
dθ0 σ0
Vol. 20 No. 4 Still “Too Much, Too Late” 467

Appendix B. Calibration Details

B.1 Vasicek Distribution

Following Vasicek (2002), we assume that the failure of an individ-
ual loan j at time t is driven by the realization of a latent random
variable:
√
yj = Φ−1 (pst ) + ρst zt + 1 − ρst ujt , (B.1)

where Φ(.) is the standard normal CDF, zt ∼ N (0, 1) is the common

risk, ujt ∼ N (0, 1) is idiosyncratic risk, and ρst is the correlation
coefficient that determines a correlation between the performance of
individual loans.
The loan defaults when yj < 0, which happens with probability
√
P (yj < 0) = P ρst zt + 1 − ρst ujt < −Φ−1 (pst )

= Φ Φ−1 (pst ) = pst . (B.2)

Since the probability of failure pst is identical for all loans, by

the law of large numbers, the failure rate ξt for a given realization
of the systematic risk factor zt equal to the probability of failure of
a (representative) project j conditional on zt . Thus,
√
ξt = ξt (zt , st ) = P ρst zt + 1 − ρst ujt < −Φ−1 (pst ) |zt
 −1 √ 
Φ (pst ) − ρst zt
=Φ √ . (B.3)
1 − ρst
We can now easily derive the distribution of ξt (zt , st ), which is given
by
  
1 − ρist Φ−1 ξti − Φ−1 pist
F (ξt |st ) = P (ξt (zt , st ) ≤ ξt ) = Φ ,
ρist
(B.4)

where the last equality comes from substituting Equation (B.3) for
ξt (zt , st ) above and rearranging terms.
Note that the distribution function in Equation (B.4) has two
parameters: pist and ρist ∈ (0, 1). pist is the stage i loan default
468 International Journal of Central Banking October 2024

probability. ρist ∈ (0, 1) is the correlation parameter, which cap-

tures the dependence of individual loan on the common risk factor
and, thus, determines the degree of correlation between individual
loan defaults. While we calibrate pist from the data, the correlation
coefficient is computed consistent with the Basel approach (para-
graph (53) of the section Internal Ratings-Based Approach for Credit
Risk in Basel Committee on Banking Supervision 2017) such that
ρist = ρ pist , where

1 − exp−50pst 1 − exp−50pst
i i

ρ pist = 0.12 + 0.24 1 − . (B.5)
1 − exp−pst 1 − exp−pst
i i

B.2 Capital Requirement

Under the internal ratings-based approach, the capital requirement
for corporate and bank exposures is meant to ensure sufficient cap-
ital to cover loan losses with a confidence level of 99.9 percent. The
formula for κist is taken from paragraph (53) of Internal Ratings-
Based Approach for Credit Risk in Basel Committee on Banking
Supervision (2017) and is given by
  −1   
Φ (p̄i ) ρ̄i −1
κist = λb Φ √ + Φ (0.999) − p̄λb
1 − ρ̄i 1 − ρ̄i
1 + (M − 2.5)bi
× , (B.6)
1 − 1.5bi

where p̄i := qg pig + qb pib is the through-the-cycle (i.e., uncondi-

tional on the aggregate state) default probability of stage i loans;
ρ̄i = ρ (p̄i ) is the through-the-cycle loan default correlation coef-
ficient of stage i loans; M is effective maturity in years, which in
our model is given by 1/δ; bi = [0.11852 − 0.05478 ln (p̄i )] is the
maturity adjustment coefficient. Under the IRB approach, Basel III
specifies the use of downturn loss given default in computing the cap-
ital requirement, which in our model corresponds to λb . The overall
capital requirement is then given by

κst = ωst κ1st + (1 − ωst )κ2st . (B.7)

Vol. 20 No. 4 Still “Too Much, Too Late” 469

B.3 Discounted Lifetime Losses

The discounted lifetime losses on a unit size stage i loan can be
written recursively as

1
θsLT = E[λst+1 ξt+1 + (1 − ξt+1 )(1 − δ)θsLT |st ]. (B.8)
t
1 + d st t+1

The above equation can be written in matrix form as

θ̄LT = A θ̄LT + μ, (B.9)

where
 
1−δ 1 − pg 1 − pb
A= Q◦ , (B.10)
1 + d st 1 − pg 1 − pb

and
 
1 λ g pg
μ= Q , (B.11)
1 + d st λ b pb

where Q is the 2 × 2 transition probability matrix and “◦” denotes

the Hadamard (element-wise) product. Thus,

−1
θ̄LT = (I2×2 − A) μ,

where I2×2 is a 2 × 2 identity matrix.

Appendix C. Numerical Solution Method

C.1 Model
We obtain a fully non-linear solution to the model numerically using
the value function iteration method. In general, the model has two
endogenous state variables, Et and Lt . However, given the calibrated
values of the parameters, we find that the minimum capital require-
ment Et ≥ κst Lt is binding on the simulation path. Therefore, we
solve our model under assumption that Et = κst Lt . Thus, effectively,
470 International Journal of Central Banking October 2024

the model has one endogenous state variable Lt . The grids for Lt−1
consist of 120 points and include 119 equispaced points between 0.17
and 0.65, and 0. Furthermore, we use a linear interpolation method
for the grid of choice variables Lt (implemented by applying the
interp1 MATLAB function to the original grids of Lt−1 with the
query point equal to 0.1). As a result of linear interpolation, the
grids for Lt consist of 1,201 points. The high density of the grid for
the choice variable is highly important to obtain a reliable approxi-
mation of the solution to our model. This is because the loss rate ξt
and, thus, the provisioning rate θst have relatively small magnitudes.
Therefore, to pick up any effect from a relatively small change in θst
(say the difference between IFRS 9 and U.S. GAAP), it is crucial
that the grid for Lt is sufficiently dense.
For the numerical representation of the exogenous state the ran-
dom variable ξt is discretized. The grid of the ξti ’s support consists of
41 points (in each aggregate state). As we show in Appendix B, the
default rate ξti can be written as a function of the standard normal
distribution:

Φ−1 (pist ) − ρist u
ξti = g(u; st ) = Φ ,
1 − ρist

where Φ(.) is the standard normal pist and ρist are defined in the
text, and u ∼ N (0, 1). Therefore the discrete approximation of ξti is
obtained by discretizing u which is performed using the approach of
Tauchen (1986) (we set the bounds of the support of u to [–3.5;3.5]
allowing for extreme realizations of u). Because we have two possi-
ble realizations of aggregate states and also record the one-period
history of the aggregate state (st−1 ), the space of exogenous state
consists of 2 × 2 × 41 = 164 points.
To compute the moments implied by the model, we simulate our
model for 80,000 periods. The first 200 observations are dropped
when computing the moments to avoid the initial value having any
effect. When the bank defaults on the simulation path, it is replaced
starting from the next period with a new bank with the average size
(i.e., with Lt given by unconditional means). Given that default is a
rare event, the replacement rule does not have any profound effect
on the moments.
Vol. 20 No. 4 Still “Too Much, Too Late” 471

C.2 Generalized Impulse Response Functions

The generalized impulse response is approximated using a simula-
tion method. That is, given the initial state variable, we perform
N = 30, 000 simulations of the model each with the length of T = 20
periods. Averaging the variable of interest across simulated paths for
each period t ∈ [0, T ] then produces its generalized impulse response
in that period. Taking p-th and 100-p-th percentile across the simu-
lated paths produces the p percent confidence bounds. In our figures,
we only plot the approximated responses for the first 11 periods.

References

Abad, J., and J. Suarez. 2018. “The Procyclicality of Expected

Credit Loss Provisions.” Working Paper.
Acharya, V. V., and S. G. Ryan. 2016. “Banks’ Financial Reporting
and Financial System Stability.” Journal of Accounting Research
54 (2): 277–340.
Admati, A. R., P. M. DeMarzo, M. F. Hellwig, and P. Peiderer.
2018. “The Leverage Ratchet Effect.” Journal of Finance 73 (1):
145–98.
Ahmed, A. S., C. Takeda, and S. Thomas. 1999. “Bank Loan Loss
Provisions: A Reexamination of Capital Management, Earnings
Management and Signaling Effects.” Journal of Accounting and
Economics 28 (1): 1–25.
Altınkılıç, O., and R. S. Hansen. 2000. “Are There Economies of
Scale in Underwriting Fees? Evidence of Rising External Financ-
ing Costs.” Review of Financial Studies 13 (1): 191–218.
Barclays. 2017. “European Banks: IFRS9 — Bigger than Basel IV.”
Industry Update.
Basel Committee on Banking Supervision. 2017. “Basel III: Finalis-
ing Post-Crisis Reforms.” Bank for International Settlements.
Beatty, A., and S. Liao. 2011. “Do Delays in Expected Loss Recogni-
tion Affect Banks’ Willingness to Lend?” Journal of Accounting
and Economics 52 (1): 1–20.
Behn, M., R. Haselmann, and P. Wachtel. 2016. “Procyclical Capital
Regulation and Lending.” Journal of Finance 71 (2): 919–56.
472 International Journal of Central Banking October 2024

Borio, C., and F. Restoy. 2020. “Reflections on Regulatory

Responses to the COVID-19 Pandemic.” FSI Briefs No. 1, Finan-
cial Stability Institute, Bank for International Settlements.
Bushman, R. M., and C. D. Williams. 2015. “Delayed Expected
Loss Recognition and the Risk Profile of Banks.” Journal of
Accounting Research 53 (3): 511–53.
Chae, S., R. F. Sarama, C. M. Vojtech, and J. Wang. 2019. “The
Impact of the Current Expected Credit Loss Standard (CECL)
on the Timing and Comparability of Reserves.” Finance and Eco-
nomics Discussion Series No. 2018-020, Board of Governors of the
Federal Reserve System.
Cohen, B. H., and G. A. Edwards, Jr. 2017. “The New Era of
Expected Credit Loss Provisioning.” BIS Quarterly Review 1
(March): 39–56.
De Nicolò, G., A. Gamba, and M. Lucchetta. 2014. “Microprudential
Regulation in a Dynamic Model of Banking.” Review of Financial
Studies 27 (7): 2097–2138.
Dewatripont, M., and J. Tirole. 2012. “Macroeconomic Shocks and
Banking Regulation.” Journal of Money, Credit and Banking 44
(s2): 237–54.
Dinger, V., and F. Vallascas. 2016. “Do Banks Issue Equity When
They Are Poorly Capitalized?” Journal of Financial and Quan-
titative Analysis 51 (5): 1575–1609.
European Systemic Risk Board (ESRB). 2017. “Financial Stability
Implications of IFRS 9.”
Financial Stability Forum. 2009. “Report of the Financial Stability
Forum on Addressing Procyclicality in the Financial System.”
Gersbach, H., and J.-C. Rochet. 2017. “Capital Regulation and
Credit Fluctuations.” Journal of Monetary Economics 90 (Octo-
ber): 113–24.
Goldstein, I., and H. Sapra. 2014. “Should Banks’ Stress Test Results
be Disclosed? An Analysis of the Costs and Benefits.” Founda-
tions and Trends in Finance 8 (1): 1–54.
Hanson, S. G., A. K. Kashyap, and J. C. Stein. 2011. “A Macropru-
dential Approach to Financial Regulation.” Journal of Economic
Perspectives 25 (1): 3–28.
Hennessy, C. A., and T. M. Whited. 2007. “How Costly Is External
Financing? Evidence from a Structural Estimation.” Journal of
Finance 62 (4): 1705–45.
Vol. 20 No. 4 Still “Too Much, Too Late” 473

Huizinga, H., and L. Laeven. 2018. “The Procyclicality of Banking:

Evidence from the Euro Area.” Working Paper.
International Accounting Standards Board (IASB). 2014. “Interna-
tional Financial Reporting Standard 9 Financial Instruments.”
Jiménez, G., S. Ongena, J.-L. Peydró, and J. Saurina. 2017. “Macro-
prudential Policy, Countercyclical Bank Capital Buffers, and
Credit Supply: Evidence from the Spanish Dynamic Provisioning
Experiments.” Journal of Political Economy 125 (6): 2126–77.
Kashyap, A., and J. Stein. 2004. “Cyclical Implications of the Basel
II Capital Standards.” Economic Perspectives (Federal Reserve
Bank of Chicago) 28 (1): 18–31.
Koop, G., M. H. Pesaran, and S. M. Potter. 1996. “Impulse Response
Analysis in Nonlinear Multivariate Models.” Journal of Econo-
metrics 74 (1): 119–47.
Krüger, S., D. Rösch, and H. Scheule. 2018. “The Impact of Loan
Loss Provisioning on Bank Capital Requirements.” Journal of
Financial Stability 36 (June): 114–29.
Laeven, L., and G. Majnoni. 2003. “Loan Loss Provisioning and Eco-
nomic Slowdowns: Too Much, Too Late?” Journal of Financial
Intermediation 12 (2): 178–97.
Laux, C. 2012. “Financial Instruments, Financial Reporting, and
Financial Stability.” Accounting and Business Research 42 (3):
239–60.
Laux, C., and C. Leuz. 2010. “Did Fair-Value Accounting Contribute
to the Financial Crisis?” Journal of Economic Perspectives 24
(1): 93–118.
Mahieux, L., H. Sapra, and G. Zhang. 2023. “CECL: Timely Loan
Loss Provisioning and Banking Regulation.” Journal of Account-
ing Research 61 (1): 3–46.
Malherbe, F. 2020. “Optimal Capital Requirements over the Busi-
ness and Financial Cycles.” American Economic Journal: Macro-
economics 12 (3): 139–74.
Mankart, J., A. Michaelides, and S. Pagratis. 2020. “Bank Capital
Buffers in a Dynamic Model.” Financial Management 49 (2):
473–502.
Martinez-Miera, D., and R. Repullo. 2010. “Does Competition
Reduce the Risk of Bank Failure?” Review of Financial Studies
23 (10): 3638–64.
474 International Journal of Central Banking October 2024

Myers, S. C. 1977. “Determinants of Corporate Borrowing.” Journal

of Financial Economics 5 (2): 147–75.
Myers, S. C., and N. S. Majluf. 1984. “Corporate Financing
and Investment Decisions when Firms Have Information that
Investors Do Not Have.” Journal of Financial Economics 13 (2):
187–221.
Repullo, R. 2013. “Cyclical Adjustment of Capital Requirements: A
Simple Framework.” Journal of Financial Intermediation 22 (4):
608–26.
Repullo, R., J. Saurina, and C. Trucharte. 2010. “Mitigating the
Pro-cyclicality of Basel II.” Economic Policy 25 (64): 659–702.
Repullo, R., and J. Suarez. 2013. “The Procyclical Effects of Bank
Capital Regulation.” Review of Financial Studies 26 (2): 452–90.
Tauchen, G. 1986. “Finite State Markov-Chain Approximations to
Univariate and Vector Autoregressions.” Economics Letters 20
(2): 177–81.
Vasicek, O. 2002. “The Distribution of Loan Portfolio Value.” Risk
15 (December): 160–62.
BEHAVIORAL RESEARCH IN ACCOUNTING American Accounting Association
Vol. XX, No. XX DOI: 10.2308/BRIA-2023-027
MONTH YEAR
pp. 1–17

The Decision Usefulness of Current Expected Credit

Losses: Users’ Views about the Current Expected
Credit Losses Model
Jordan M. Bable

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

Indiana University Bloomington

Christopher I. L. Wong
Wilfrid Laurier University

Michael J. Wynes
University of Saskatchewan

ABSTRACT: In 2016, the Financial Accounting Standards Board (FASB) issued ASU 2016-13, “Financial
Instruments—Credit Losses,” requiring firms to switch to a current expected credit losses (CECL) model. To assess
the impact of this new standard, we performed semistructured interviews with analysts, trade group members, and
financial journalists, all of whom have experience with CECL. Overall, interviewees shared the view that the CECL
standard-setting process was tumultuous and political. Interviewees also stated that CECL led to perceptions of
decreased decision usefulness of loan loss information and decreased comparability among reporting firms but had
little impact on firms’ lending operations. Our study answers the call from the FASB to perform research into the
impacts of CECL and also contributes to the literature on sell-side analyst decision making and the literature on the
determinants of decision usefulness for analysts.
Data Availability: Data are not available for confidentiality reasons.
JEL Classifications: G21; G28; M41; M48.
Keywords: current expected credit loss model (CECL); loan loss provisioning; FASB; standard setting; financial
analysts.

I. INTRODUCTION
So, is CECL something that has helped investors? No. It has harmed investors dramatically. Does CECL give
you any insight into what the bank operations are? No. It gives you an insight into what some crazy economist
thinks the future of the economy will be. Should CECL have ever been created? No, because it has only
harmed everyone that’s gotten involved with it over the past 24 months or so that it’s been put into effect.
—(SA 8)

We thank Nick Seybert (editor) and two anonymous reviewers for their detailed comments during the review process. We greatly appreciate our inter-
viewees for being generous with their time and thoughts. We are also appreciative of helpful comments from Rick Cazier, Mike Durney, Brian
Monsen, Jed Neilson, participants at the 2022 Brigham Young University Accounting Research Symposium, and round table participants at the 2023
Financial Accounting and Reporting Section Midyear Meeting. We also gratefully acknowledge the financial support of the University of Waterloo.
Jordan M. Bable, Indiana University Bloomington, Kelly School of Business, Accounting Department, Bloomington, IN, USA; Christopher I. L.
Wong, Wilfrid Laurier University, Lazaridis School of Business & Economics, Accounting Department, Waterloo, Ontario, Canada; Michael J.
Wynes, University of Saskatchewan, Edwards School of Business, Accounting Department, Saskatoon, Saskatchewan, Canada.
Editor’s note: Accepted by Nick Seybert, under the Senior Editorship of Victor S. Maas.

Submitted: July 2023

Accepted: March 2025
Early Access: April 2025

1
2 Bable, Wong, and Wynes

W
ith the issuance of Accounting Standards Update, ASU, 2016-13 the Financial Accounting Standards Board
(FASB) introduced a significant change in recognizing credit losses under U.S. generally accepted accounting
principles. Specifically, the FASB shifted from the previous incurred loss model (ILM) to a current expected
credit loss (CECL) model. Rather than waiting for losses to become probable as under the prior approach, typically as
the result of a trigger event, firms applying the new standard must now recognize the amount of all future losses they
expect on the day the loan is originated (FASB 2016). With most SEC filers beginning to implement CECL in fiscal year
2020, the first financial statements prepared under the new method are now available, and the FASB is undertaking a
post-implementation review. Accordingly, one of the primary questions being asked, and the one we seek to answer in
our study, is: Does CECL provide users with decision-useful information to assess expected credit losses for the underly-
ing portfolio?
Contemporaneous research is mixed regarding positive and negative effects of CECL on banks, investors, and ana-
lysts. Banks implementing CECL have timelier loan-loss provisions that are more predictive of future losses (S. Kim,

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

S. Kim, Kleymenova, and Li 2023) but also appear to engage in more procyclical lending (Chen, Dou, Ryan, and Zou
2025). CECL increases investors’ perceptions of loan-loss risks (Koonce, Mongold, Quaid, and Winchel 2024), but the
value relevance of allowances is only improved for small banks that did not previously perform stress tests (Gee,
Neilson, Schmidt, and Xie 2024). For analysts, CECL results in a deterioration of forecast properties (Bonsall, Schmidt,
and Xie 2023). Taken together, these results suggest that CECL results in superior information at the bank-level, but
that this information may not fully make its way to investors and analysts. Our paper seeks to address why.
We add to this discussion on the usefulness of CECL by conducting interviews with relevant professionals, including
sell-side analysts, trade group members, and financial journalists. Using a semistructured interview approach designed
based on the FASB Academic Research Webinar for Credit Losses, we asked our interviewees about their experiences
with the CECL standard, guided by interview prompts related to the overarching topic of the decision usefulness of
CECL. The interviews generated 490 minutes of source material transcribed into 162 pages of empirical data. All three
authors reviewed the data, coding them for various themes. In addition to the themes embedded in our interview
prompts, an additional theme emerged across our various interviews: the tension and politics in the CECL standard set-
ting process. Amongst these themes, our interviewees presented perceptions that also shine a light on some of the topics
addressed in the other papers discussed above: bank behavior, the usefulness of information for unsophisticated invest-
ors, and how analysts use this information.
Regarding tension and politics in implementing the CECL standard, our interviewees expressed feelings of being
heard but not listened to. This idea of “controversial linkages” between standard setters and users (Durocher and
Gendron 2011), or the notion that the standard-setting process was political, with different groups lobbying for their
views to be dominant in an arena of power, rationality, and legitimacy (Fogarty, Hussein, and Ketz 1994; Durocher,
Fortin, and C^ ote 2007; Pelger 2016) is not unique to CECL but highlights an ongoing issue where some believe that the
FASB does not adequately consider the needs of investors (Ho 2021). As succinctly expressed in a 2019 letter regarding
CECL to the FASB from the CFA Institute, “It is our view that the US GAAP model chosen was the preference of regu-
lators not investors” (Chartered Financial Analyst Institute 2019).
Regarding the decision usefulness of the standard, the analysts we interviewed conveyed that they do not believe
CECL provides decision-useful information to them. This belief was built on various concerns. For example, comparabil-
ity across issuers appears to have been eroded. Also, at many firms, macroeconomic forecasts used in CECL model esti-
mations for loan loss reserves rely on homogeneous inputs from a single information provider, Moody’s. Furthermore,
there is a perception that less sophisticated investors, such as retail investors, are harmed by the complexity of the new
standard. One interviewee even expressed concern that some sophisticated investors, like portfolio managers and non-
bank specialist analysts, struggle with understanding the CECL standard and its effects on the financial statements.
Finally, our interviewees talked about real operational impacts due to CECL adoption. Consistent with Kim et al.
(2023), none of them have seen nor expected to see any significant changes in banks’ lending behavior. However, there
was concern about the unintended consequences of CECL on lending behavior more generally. Specifically, there was a
concern about the movement of lending away from regulated banks to unregulated spaces, effectively avoiding CECL.
Our study makes several contributions. First, the results of our study answer the call from the FASB to explore the
consequences of CECL in this post-implementation review period to help inform standard setters, academics, and users
about the decision usefulness and real operational impacts of the standard. We find that in general users do not think
CECL has been helpful, and some believe it may in fact be harmful. These findings are in line with the negative conse-
quences documented by Bonsall et al. (2023) for degradation of analyst forecast usefulness and Chen et al. (2025) for the
procyclicality of banks’ lending. However, our results are contrary to the positive consequences documented by Gee
et al. (2024) and Koonce et al. (2024) for increased investor decision usefulness and Kim et al. (2023) for banks’ future
loan performance and information production quality. Thus, our findings contribute to the equivocal body of research

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 3

that shows the CECL standard in both positive and negative lights. Our study is unique among CECL studies in its use
of field-based research, specifically interviews with sophisticated users of financial information, rather than relying on
archival or lab-generated data. This method allows for a “greater understanding of the causal processes linking account-
ing practices and outcomes” (Ittner 2014, 545) and provides themes and quotes that “provide vivid images that promote
intuition” by asking “people directly about their objectives and constraints” (Helper 2000, 229) which can support exist-
ing research findings and guide future research in financial reporting generally, and CECL more specifically.1
Second, our paper contributes to our understanding of how accounting standards are interpreted and operationalized
by analysts. Prior studies have focused on the decision-making inputs of sell-side analysts (e.g., Soltes 2014; Brown, Call,
Clement, and Sharp 2015), the determinants of decision usefulness for sell-side analysts (e.g., Gassen and Schwedler 2010;
Campbell and Slack 2011; Georgiou, Mantzari, and Mundy 2021), and the effects of mandatory International Financial
Reporting Standards (IFRS) adoption on the properties of analysts’ earnings forecasts (e.g., Byard, Li, and Yu 2011;
Tan, Wang, and Welker 2011; Horton, G. Serafeim, and I. Serafeim 2013). We provide evidence that sell-side analysts

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

view CECL as irrelevant to their models because it does not provide decision-useful information, despite a robust prior
literature indicating that mandatory accounting standard adoption improves analyst forecasts.
The remainder of the paper proceeds as follows: Section II presents some background on the CECL standard and a
review of the relevant literature. In Section III, we explain our protocol for collecting evidence. In Section IV, we present
our findings through the views of our interviewees. Section V discusses and interprets what the interviewees said and
offers some concluding remarks.

II. BACKGROUND
Prior to the adoption of CECL, credit losses were recognized under the ILM which only recognized credit losses
when they were “probable” and “reasonably estimable” (ASC 450-20). In the aftermath of the financial crisis of the late
2000s, several groups including the G20 and Financial Stability Board posited that banks’ recognition of losses was “too
little, too late” and demanded reform to when credit losses would be recognized. ASU 2016-13, the standard that imple-
mented CECL, removed the “probable” threshold and required banks to look forward, estimate all expected future
credit losses, and establish a reserve on day-1 to cover those losses (FASB 2016). Although both the IL and CECL mod-
els have banks draw on historical information about loan loss rates to establish reserves, the CECL model also requires
banks to incorporate “reasonable and supportable forecasts” of future conditions (FASB 2016, 3).
The development and implementation of the CECL standard was a lengthy process involving both the IASB and
FASB (see Hashim, Li, and O’Hanlon 2019 for a summary). During the development of the standard, the FASB
received numerous comment letters from preparers (banks and credit unions), Congress, and members of the FASB
board opposing recognizing all losses on day-1. On the other hand, the FASB in their summary of the public comment
period (FASB 2013) noted that:
by a 3-1 margin, investors and other users prefer a model that recognizes all expected credit losses (as opposed
to maintaining a threshold that must be met before all expected credit losses are recognized or permitting rec-
ognition of only some expected credit losses). (emphasis in the original)
However, despite the majority of the 71 investors the FASB interviewed espousing positive feelings toward CECL, there
were still roughly 25 percent (18 of 71) who thought it was inappropriate to recognize all expected credit losses on day-1.
The FASB summary noted some of the reasons mentioned included:
a. It does not appropriately “match” credit losses against interest income.
b. It will result in too much reserves being recognized too soon.
c. It will require/allow forecasting beyond two years, which could result in unreliable (potentially cookie jar) esti-
mates and volatile revisions.
d. It will cause a major capital hit. (FASB 2013)
Thus, we see that experts’ views on CECL were not homogeneous and reflected a standard shrouded in controversy.2
Studies suggest that expected credit loss models can provide more timely information.3 For instance, Laeven and
Majnoni (2003) find that banks delay recognizing possible loan losses, whereas Vyas (2011) notes firms often delay

1
Kenno, McCracken, and Salterio (2017) notes that field-based research in financial reporting is relatively rare compared to other accounting
disciplines.
2
The fact that the standard setting process is difficult is not new. Prior research has documented similar controversy in the process related to standards
on stock compensation (Young 2014) and business combinations (Durocher and Georgiou 2022).
3
Addressing the issue of “too little, too late” in the old ILM models, CECL is directly related to the timeliness characteristic as defined in the
Conceptual Framework, “having information available to decision makers in time to be capable of influencing their decisions” (FASB 2018, QC29).

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
4 Bable, Wong, and Wynes

accounting write-downs when they have discretion. More timely credit loss recognition is shown to reduce bank risk-
taking (Bushman and Williams 2012, 2015). Additionally, recent research indicates that investors anticipate expected
credit losses before they are recognized, as evidenced by stock prices reflecting information not recognized in the finan-
cial statements (Wheeler 2021), analyst forecasts of loan loss provisions being predictive of future losses (Beatty and
Liao 2021), and credit loss provisions being predictive of future bank risk (L opez-Espinosa, Ormazabal, and Sakasai
2021). Looking specifically at CECL, Gee et al. (2024) find that banks’ CECL day-1 impacts (which reflect all expected
future losses) are associated with stock prices, future nonperforming loans, and future net write-offs. They also suggest
that CECL reveals new information to investors, reducing uncertainty.
Although CECL is designed to improve decision usefulness through timely loan loss provisioning, research also
warns that it may not be the credit crisis solution regulators hope for. Increased timeliness may reduce information accu-
racy, especially when recognizing lifetime losses at loan origination (Mahieux, Sapra, and Zhang 2023). Additionally,
the models used for estimating loan losses are prone to errors and manipulation. Covas and Nelson (2018) argue that

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

CECL could lead to procyclical behavior, exacerbating economic downturns by overestimating losses during weak peri-
ods and underestimating them during strong ones. This procyclicality is partly due to the inability of macroeconomic
forecasts to accurately capture business cycle turning points. Additionally, the choice of parameters, such as the discount
rate, can significantly impact loan loss provisioning, potentially forcing banks to recognize accounting losses that do not
reflect real economic losses (Ronen 2023).

III. COLLECTION OF EVIDENCE

Our analysis draws upon 11 semistructured interviews and two written correspondences with eight sell-side analysts
(hereafter simply analysts), three industry trade group members, and two financial journalists. The interviews were con-
ducted over video conferencing software (Zoom) between March and May 2022 and were recorded for later transcrip-
tion.4 Interviews lasted on average 49 minutes and generated 160 pages of transcripts (162 pages including the written
correspondences) in total (size 12 font, single spaced). Interviewees were asked about their experiences with the CECL
standard and their opinions on the standard across a variety of topics. See Appendix A for the semistructured script
used. To encourage a frank and candid discussion of the standard, interviewees were promised anonymity for the com-
ments and opinions shared, prior to providing consent to be recorded.
Interviewees were recruited through cold emailing individuals. In the case of our analyst interviewees, individuals
were identified by searching for analysts through a Bloomberg terminal. To be included in our study, analysts were
required to be actively covering one of the ten largest American banks.5 This approach ensured that the analysts we con-
tacted would have experience with CECL as the ten largest American banks were among the earliest adopters of the
standard. In total, contact information was obtained for 66 analysts.6 As noted in Georgiou (2018), these individuals
typically hold senior positions in their firms and their time is precious to them. The fact that we received responses from
eight of these individuals (representing a response rate of 12.1 percent) is a testament to the timeliness and importance of
the interview topic to these individuals. In the case of our financial journalist and trade group interviewees, contact
information was obtained through websites and reports.
Our analyst subsample had between nine and 60 years of experience, with an average of 30 years of sell-side experi-
ence. Everyone in our analyst subsample is cited regularly in the financial press, such as the Financial Times and the
Wall Street Journal, and/or appears on business news television outlets such as Bloomberg and CNBC. Our financial
journalist subsample has an average of 13 years of experience reporting on financial matters and has interviewed a con-
siderable number of analysts as part of their reporting on the passage and implementation of CECL. Our trade group
member subsample has an average of 32 years of experience representing their trade groups and hold senior positions
within their respective trade groups.7 Therefore, we believe that our interviewees possess the requisite knowledge and
experience to provide valuable insight with regard to our interview topic. Additional information about the interviewees
is provided in Table 1.

4
The authors received ethics approval at their respective universities to conduct interviews.
5
The list of the ten largest banks that we focused on, based on total assets: J.P. Morgan Chase, Bank of America, Citigroup, Wells Fargo, Goldman
Sachs, Morgan Stanley, U.S. Bancorp, PNC Financial Services, Truist Financial, and Capital One.
6
One of our analyst interviewees confirmed the small population of experts on CECL saying, “I’m probably one of 50 people who follow banks who
really understand how this works” (SA 6).
7
We define industry trade group members as individuals who have experience with and/or are associated with individuals who have experience with
preparing or using financial reports and disclosures relating to the CECL standard. This could include individual corporations, institutional investors,
investment banks, public accounting firms, and other representatives from the financial services industry.

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 5

TABLE 1
Interviews

Expressed Opinion to FASB Years of Length in # of

Interviewee Date during Pre-Implementation Experience Minutes Pages
Sell-Side Analysts
SA 1 March 7, 2022 No 35 NA (written) 1
SA 2 March 9, 2022 Yes 37 43 14
SA 3 March 23, 2022 No 30 NA (written) 1
SA 4 March 24, 2022 No 30 48 19
SA 5 March 30, 2022 No 9 26 10

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

SA 6 April 18, 2022 Yes 14 43 15
SA 7 May 24, 2022 ??? 24 50 14
SA 8 May 27, 2022 No 60 16 5
Financial Journalists
FJ 1 April 27, 2022 No 14 62 20
FJ 2 May 3, 2022 No 11 54 21
Trade Group Members
TG 1 May 3, 2022 Yes 20 107 31
TG 2 May 3, 2022 Yes 39
TG 3 May 25, 2022 Yes 36 41 11

Total 490 162

We developed a semistructured interview script based on concerns outlined in the FASB Academic Research Webinar for
Credit Losses. However, we approached our interviews with an open mind, and after each interview, analyzed the emerging
themes and iteratively updated our interview prompts and approach in each subsequent interview. We started the recruitment
process with analysts, recognizing their role as important financial intermediaries with specialized industry experience who are
accustomed to using bank financial reports and disclosures (Campbell and Slack 2011). After obtaining responses from our
first six analyst interviewees (SA 1–SA 6), we perceived that we had attained a degree of saturation as evidenced by a lack of
new emerging themes. Therefore, as suggested by Gurd (2008, 128), we decided to, “seek disconfirming cases which may con-
tradict parts of the present theory development,” by recruiting financial journalists and industry trade group members who we
expected to have different but relevant views relative to our analyst interviewees. We also decided to recruit and interview an
additional two analysts (SA 7 and SA 8) and did not discover any new emerging themes or insights. Therefore, we believe that
our interview protocol allowed us to attain a reasonable level of saturation and engagement with the field.8
Due to the cold calling method of gaining access to our interviewees, it is likely that the individuals responding to
our cold calls had strong opinions on the subject matter. However, in our examination of comment letters and the
FASB’s report summarizing the outcomes of their public consultations (FASB 2013), the opinions expressed by our
interviewees do not appear to be extreme or inconsistent. Another potential concern with our collection of evidence
could be that respondents, out of fear for reputation concerns or damaging their relationships with the FASB, would
exhibit a social desirability response bias (e.g. Zerbe and Paulhus 1987) preventing them from providing their honest
opinions if they were critical of the CECL standard. To safeguard against this, interviewees were promised anonymity
and encouraged to be truthful and objective. During the interview process, we did not detect a strong social desirability
response, instead, we felt that interviewees approached the interviews with candidness, openness, and honesty.

IV. FINDINGS
Prior literature investigates the decision usefulness and operational impacts of CECL on banks, analysts, and invest-
ors. We asked our interviewees for their views on these topics, generally finding unfavorable opinions concerning

8
According to Guest, Bunce, and Johnson (2006), saturation can be achieved with 12 interviews. Further, Francis et al. (2010) believe that saturation
can be attained with 13 interviews so long as there are no new insights in the last few interviews. Given the homogeneity within our analyst sample,
the small population from which to sample (66), our high response rate (12.1 percent), and the lack of new emergent themes within our analyst sub-
sample, we believe we have attained a reasonable degree of saturation.

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
6 Bable, Wong, and Wynes

CECL. In this section, we discuss results pertaining to decision usefulness and operational impacts. We also discuss a
key emerging theme from our interviews that there was a political nature to the CECL standard-setting process that
may have prevented a more useful standard from being developed. Consistent with Georgiou (2018, 1304), we make use
of illustrative quotes from the interview transcripts and present them verbatim to allow the reader to, “hear the inter-
viewees’ voices.”

Does CECL Provide Users with Decision-Useful Information?

General Thoughts on Decision Usefulness
One common response from our interviewees was that the complexities of CECL can be interpreted and disen-
tangled by industry expert analysts, but users with less expertise, such as the average retail investor, or even nonindustry
expert analysts, would struggle with using the information. Some interviewees went further, noting that CECL reporting

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

and disclosures were confusing, requiring investors to conduct more research than under the ILM to truly understand
effects of CECL on the financial statements:
It basically requires more research in order to interpret provisioning data than it did before. It isn’t necessarily
to say that before was very transparent to retail investors either; it just hasn’t really improved. If anything, it’s
gotten a little worse. (SA 5)
Some interviewees highlighted that only a small group of individuals outside of the FASB really understand the
impact of the CECL standard on bank reporting and disclosures, particularly in the context of an already complex regu-
latory environment:
I think there’s too many things in here that just make it too complicated for the average investor. I’ll say
something very purposely, I’m probably one of 50 people who follow banks who really understand how this
works. (SA 6)
FASB talked about how, “We talked to numerous investors about this.” As a reporter, you try and go out and
find investors to talk about it. That population is small. There are investors, and then there’s investors that pay
attention to accounting, and then there’s investors who not only pay attention to accounting but know what
CECL is. (FJ 1)
The above quote from the financial journalist speaks to the complexity of the CECL standard. However, the financial
journalist’s comment that the FASB “talked to numerous investors” also hearkens to our later discussion in the section
“Tension and Politics in the CECL Standard Setting Process” on the FASB’s pattern of soliciting feedback from constit-
uents but not providing them meaningful voice in the standard setting process.
Other interviewees noted that the danger with CECL is that it adds uninformative noise to reported earnings, which
distorts the true nature of the bank’s performance in any quarter, and that nonindustry experts, even generalist analysts
and portfolio managers, may be inappropriately influenced by these shocks to earnings, ultimately leading to losses in
wealth:
CECL, I think needlessly contorted the banks’ financials…It’s easy enough for me to back it all out, but there
were fairly sophisticated portfolio managers, generalist portfolio managers who just didn’t get it at all. Who,
when they saw the provisions in 2020, they just hit the panic button…“Oh my God. Bank X reported a loss.”
Well, it reported a loss, but that’s just because of accounting. People don’t want to hear it. (SA 2)
The retail investor is not going to get into the weeds. Just because you make an extra three pages of disclosure
doesn’t mean they’re going to read it. Sure as heck doesn’t mean they’re going to interpret it. (SA 4)
Some interviewees expressed concern that the way CECL is currently designed may actually reduce reliability either
through increased opportunity for earnings management or through a bias toward excessive conservatism:
It’s one more earnings management tool. It’s one more non-cash, discretionary estimation thing that’s in there.
To me it’s less linked to reality than the way we used to do it. (SA 2)
“Well, I know [under ILM] I used to lose 10 or 20 basis points on my real estate book. I better assume 35, because
you never know.” And you can talk yourself into much higher losses and that’s what CECL does. (SA 4)
Here, the analysts, respectively, describe how the uncertainty involved in forward-looking forecasting under CECL may
allow for more earnings management or induce excessive conservatism in accounting estimates. Analysts appear to be
expressing a sentiment that even if CECL could provide decision-useful information, its susceptibility to manipulation
renders it unreliable. Overall, our interviewees viewed the CECL standard as generally not decision useful, increasing

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 7

the complexity and noisiness of bank financial reports and disclosures, while reducing transparency and incorporating
excessive conservatism in banks’ loan loss provisions.

Decision Usefulness of Quantitative and Qualitative Disclosures

Analysts generally held the view that the quantitative information reported under the CECL standard was low in
information content. One analyst tersely responded: “No, the information is useless, and if used by investors will be a
lowest common denominator. Investors will react to the worst numbers presented” (SA 1). Another analyst shared a
similar sentiment:
“Oh, gee, Bank of America, you added $4.6 billion of reserves in the quarter. How did you come up with that
number?” Then immediately, of course, the CFO would put on his soft shoes and start dancing around and
say, “Well, we ran quite a few different scenarios. There was the central case, there was the aggressive case,

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

there was the conservative case, there was the extreme cases. We probability-weighted it and came to $4.6.” I
mean, it was completely opaque. The logic from one company to the other, they said similar things and came
out to completely different answers…Then when we had the reserve releases in 2021, the analyst questions
were the exact opposite…“Gee, how did you know that a $2.2 billion reserve release was just the right num-
ber?” It was obvious that there was no methodology…In the past, when people set the reserves, they would
look at things like, “Oh, our early delinquencies and credit cards are doing X, Y, and Z, therefore we need to
build some reserves,” or, “Oh, the energy sector is in trouble, therefore we need to set reserves for these kinds
of energy loans. Guess what? We should put aside a little bit of extra for the consumers in Houston because
they’re employed by energy companies.” Each company would have its methodologies, and they might be
flawed and not right, but the idea that you’re going to look through the five or 10 years of life or more in the
case of mortgages and figure it all out in a given quarter, it’s just asinine. (SA 2)
In the above quote, the analyst is clearly frustrated by the opacity surrounding the process through which banks model
loan loss reserves under CECL. Although CECL requires the incorporation of forward-looking information into the
modeling of loan loss reserves, the standard allows considerable variation in how the inputs to the modeling process are
disclosed in financial reports. For example, in examining J.P. Morgan’s 2021 CECL disclosures (J.P. Morgan Chase
2022), we note that point estimates for macroeconomic forecasts used in the bank’s models are provided and the infor-
mation is neatly displayed in a tabular format, increasing readability and usability of the disclosure. In contrast, consis-
tent with the sentiment described by the analyst above, Bank of America’s 2021 CECL disclosures (Bank of America
Corporation 2022) did not disclose point or even range estimates for macroeconomic forecasts used in the bank’s mod-
els, with information about the modeling process disclosed in paragraph format, decreasing its readability and usability.
Notably, neither of the banks identified are fully transparent about the modeling process, contributing to the opacity
described by the analyst above. However, some banks are opaque even about the inputs to their models, to the frustra-
tion of the users of the reports.
Commenting more specifically on the disclosure of macroeconomic forecasts related to modeling loan losses under
the CECL standard, the same analyst shares the following opinion:
I’m a bank analyst, not a macroeconomist. We’re going to go with something that is fairly close to consensus,
and most of those bank forecasts are also going to go with something that’s fairly consensus…They’re not
revealing anything you don’t know. (SA 2)
However, consistent with our discussion in the section “Capital and Other Regulatory Requirements” on the decision
usefulness of the CECL standard in general, another analyst opines that a sophisticated investor, but not a retail inves-
tor, may actually find the macroeconomic forecasts decision useful.
You know, for a sophisticated investor who does get those macroeconomic forecasts, I do think there is some
increase in transparency just because you can judge the inputs relative towards your own impression of what
those inputs should be. The actual way they get modeled is still a black box, but there is kind of a step up from
there. I think for a more unsophisticated investor, like a retail investor, it doesn’t really help in terms of
transparency. (SA 5)
Even for disclosures containing more detailed and complex information, such as the vintage disclosures, analysts felt
like they didn’t possess enough information to really make use of these disclosures:
They all give these vintage data, data now in the 10-Qs, but none of us know what to do with it because they’re
not telling us how they’re reserving against any of it. I’m on this committee with one of the industry trade
groups, and I made this point to them. I said, “You guys are giving us all this great data, but we don’t have

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
8 Bable, Wong, and Wynes

anything to do with it because all it’s just telling us is, ‘Hey, this is how much we originated and this is how
much delinquencies are.’” It’s not telling us how it’s reserving. (SA 6)
Therefore, despite the additional quantitative disclosures required under CECL, investor users do not perceive the dis-
closures under the CECL standard to be decision useful. This speaks to the CECL standard being less focused on deci-
sion usefulness in its design and more of an exercise in compliance for preparers.
However, specifically on the topic of vintage disclosures, one trade group member felt that the vintage disclosures
had the potential to provide useful information in coming years as more data becomes available.
I think a lot of investors want to say, “Okay, now that we’ll have this vintage information, once you start to
see it over five plus years. Then you can start to do some of that analysis, and then it becomes even more
useful going forward, because it gives that additional information about the underlying portfolio that then the
sophisticated analyst can then just apply their own judgment and perspectives to.” (TG 1)

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

Similar to their opinions about the decision usefulness of quantitative disclosures, analysts held the view that the
qualitative disclosures did not hold much informational value. One analyst puts this succinctly: “I don’t know if there’s
a lot of firm specific information that comes through on that truthfully. I would say it’s mostly boilerplate” (SA 5). One
analyst goes further to describe the mandatory qualitative disclosures as arcane and difficult to understand, requiring
them to obtain decision useful information from other channels, such as earnings calls and investor presentations.
What you read in the 10-K and 10-Q tends to be somewhat verbose…I call it gobbledygook. When you actu-
ally read the investor presentations, they’re like, “Hey, this is what we’ve got. This is our level of reserves.
Here’s the expected losses.” or “This is kind of where we’re tracking.”…It’s almost like you have to train your-
self to read the investor presentation and their disclosures. And then you’ve got the full picture…And I wish
you could just get the plain English explanation in the SEC filing. (SA 4)
Similar to our above discussion of the compliance driven nature of CECL’s quantitative disclosures, the above quotes
describing “boilerplate” disclosures are consistent with the idea of preparers engaging with the CECL standards on qualita-
tive disclosures with a focus on compliance, as opposed to being focused on providing decision useful information to users.

Ability of Banks to Forecast and External Influences

One factor contributing to the lack of decision usefulness of CECL disclosures is uncertainty over whether banks
are capable of, or even should be expected to be capable of, predicting economic turns. Consistent with Covas and
Nelson (2018), who find that macroeconomic forecasts are unable to reliably predict significant changes in the business
cycle, one of our trade group member interviewees discusses the difficulty faced by banks in determining loan loss provi-
sions in response to the COVID-19 pandemic in 2020:
So that overarching concern of forecasting ability to identify kind of turns is still an issue, and it’s not always
going to be as obvious as a global health pandemic…So it may be a more typical recession. We would
potentially still expect to see some of those significant pro-cyclicality concerns happen. And that being said,
look at what happened through 2020, right…banks built these huge reserves in Q1 and then particularly Q2,
and then you already started seeing release from some banks in Q3, and then throughout 2021…So point being
that they booked these huge provisions for losses that never happened. (TG 1)
This sentiment was shared by an analyst, who viewed the forecasting requirement of the CECL standard as unreason-
able and error prone.
I think the implementation of this kind of extreme crystal ball forward-looking kind of exercise when you’re
requiring it to be implemented by actual human beings is just going to exaggerate all the errors that you would
have in a normal process about where you would think losses are. (SA 2)
The same trade group member as above notes that the ability of banks to reliably provide reasonable and supportable
economic forecasts differed greatly based on their size and access to resources.
I mean, let’s talk about three different groups. You have the biggest banks that have their own economic
departments. And so those economists will be getting data and doing modeling and developing forecasts…
Then taking a step down from there…you have Moody’s data and that type of stuff…So within Moody’s
there’s…a range of different scenarios that can be pulled…the banks are looking at, “Okay, what are my key
variables? And how have losses interacted with those variables to the extent that I could then use those
interactions to help me predict future losses?”…and then this third group which largely isn’t on CECL yet…
Moody’s is…maybe out of their price range. (TG 1)

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 9

Substantially all our interviewees mentioned Moody’s as being instrumental in the preparation of reasonable and
supportable forecasts as required by the CECL standard. As one of our trade group member interviewees stated: “It’s
not like you have Moody’s; it’s like there is Moody’s” (TG 2). An analyst uses an analogy to share a similar sentiment:
Moody’s is the dominant player in this business. It’s almost like saying that we’re all going to drink soft
drinks, but you can only buy Coca-Cola, or it’s really preferred to buy Coca-Cola. You can’t do anything
else…But I know when we started this in spring of 2020…Moody’s was the only game in town. It was like the
good housekeeping seal of approval to have the Moody’s forecast in your CECL model. And I always had a
problem with that, because I’m like, “Who made Moody’s God? Why are they the best?” And again, I lived
through the financial crisis. Moody’s didn’t exactly get the ratings correct on the bonds…I just think they
stepped up, to their business credit. (SA 4)
In looking at Moody’s Analytics, we found that not only will the company provide forecasts of key macroeco-
nomic variables like gross domestic product (GDP) and unemployment, but it also offers services to estimate the

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

amount of loan losses that a bank should record (Moody’s Analytics 2022). This centralization of CECL estima-
tion with Moody’s exposes some banks, and thus by extension capital markets, to the whims of certain individuals.
As noted in a May 2021 Bloomberg article, “The guy at Moody’s has hemorrhoids, we all suffer” (White and
Hood 2021). One of our analyst interviewees noted that so many firms relying heavily on the same information
source can introduce information risk in the capital markets and ultimately make the estimates under CECL less
informative:
I think there’s a lot of concentration risk there where you have a homogeneity of…the assumptions that is
being taken from one source…So, obviously, the risk if they’re wrong can theoretically filter through pretty
materially. (SA 7)
Aside from the outsized influence that Moody’s has on CECL forecasts, some interviewees noted that many banks
wait to see what larger, earlier reporters like J.P. Morgan do and then mimic their disclosures:
When we saw those early examples of J.P. Morgan and Bank of America being first reporters, all those, it felt
like all the rest of those banks were like, “Oh, let’s include that type of information in our slides.” (FJ 2)
A lot of people when they were closing their books in the first and second quarter of 2020, I bet you they were
sitting there saying, “Let’s see what J.P. Morgan does.” (SA 2)
Potentially concerning, this herding behavior extended beyond mimicry in the types of disclosures, into actual parame-
ters and forecasts used, as one trade group member notes:
I think it might have been the…second quarter of 2020…credit cards had people at 300 basis points to 1,500
basis points. And there were two banks amongst the top 30 that had 10.93%…how did they come up with the
same exact amount? (TG 2)
Lastly, another trade group member noted that in their highly regulated space, there are many other external influ-
ences that impact CECL disclosure choices.
For regulated banks it’s not just the auditors. There are conversations that you’re having with your examiners
as well…CECL’s incredibly flexible. CECL says, “You can choose options A through Z. Whatever you think
as management is the best representation.” Then your auditor comes in and is like, “Eh, maybe A through F
are feasible.” And then your bank regulator comes in and says, “Okay, you’re doing C.” (TG 1)
As a result of constraints imposed by auditors and examiners, as well as the tendency toward mimicry and herding dis-
cussed above, one might expect CECL reporting and disclosures to exhibit a great deal of comparability between banks.
However, this is apparently not the case, as described in the next subsection.

Comparability of Information across Firms

One important characteristic of decision-useful financial information is comparability. The FASB classifies compa-
rability as an enhancing characteristic in the conceptual framework (FASB 2018) and is clear to define comparability as
something that allows “users to identify and understand similarities in, and differences among, items” (FASB 2018,
QC20). They also point out that comparability is not the same thing as consistency (“use of the same methods for the
same items” (FASB 2018, QC21), nor uniformity (“like things must look alike and different things must look different”
(FASB 2018, Q23). With these ideas in mind, we asked all our interviewees if they thought comparability had been
affected by the adoption of CECL.

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
10 Bable, Wong, and Wynes

When discussing comparability, analysts expressed that comparability of loan loss provisions between firms under
the CECL standard was low and may be worse than it was under the ILM. As one analyst shares brusquely: “No, the
quality of information has deteriorated, as comparability has been eviscerated. There is no longer even a semblance of
comparability” (SA 1). Another analyst was much softer in their tone but generally shared the sentiment that compara-
bility decreased as a result of the passage of CECL:
A lot of companies use three different scenarios: base, upside, downside. Some use five, some use one. So, the
number of scenarios that companies use is sometimes hard to compare. (SA 6)
Therefore, despite the mimicry in disclosure choices and herding in forecast sources discussed in the previous subsection,
there appears to be significant variation in the method employed in applying forecasts to the extent that it hinders com-
parability between firms. Another analyst points to even further between-firm differences that made it harder to compare
provisioning between firms:

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

I would say it’s harder [to compare]…on a quarter to quarter basis there’s not a lot to interpret how much of
that is just forecasting, how much of that is actual fundamental difference between the two portfolios. So I
don’t think it’s made it significantly easier for comparability and also…there’s this very significant gap
between firms with large credit card portfolios and ones without, just because the allowances on credit cards
are so high. (SA 5)
Clearly, analysts agree that the current CECL standard does not provide adequate guidance on what firms should
present related to their projections of credit losses nor how it should be presented. To be clear, this criticism of firms’ dis-
closure comparability is separate and distinct from analysts’ concerns above about the ubiquitous use of Moody’s fore-
cast data and the tendency of firms to mimic other firms’ forecasts, both of which do increase similarity across firms’
disclosures on that one aspect. However, as noted by TG1 in the last quote in the section “Ability of Banks to Forecast
and External Influences,” “CECL’s incredibly flexible,” allowing firms to create disparate disclosures that vary in their
levels of detail and overall structure.

Operational Impacts of CECL

Capital and Other Regulatory Requirements
Some interviewees talked about how CECL may interact negatively with existing regulatory requirements. One ana-
lyst discusses how CECL may cause banks to raise more capital as a result of increased loan-loss reserve levels:
This is more a Federal Reserve issue than it is a FASB issue, but…I’ll say it while we’re on the topic, CECL
was an accounting convention, it wasn’t created to trap capital in the banking…And effectively you told banks
you now have to hold reserves for the life of loans, but if you think about it, your allowance is supposed to
protect you for losses that you’re expected to incur over a certain loss emergence period. And then your capital
is supposed to protect you for anything unforeseen. And the way that CECL was written, you were now
needing to protect for the unforeseen and yet you weren’t being given any capital relief, which I think that’s
what created this notion of double accounting. (SA 6)
Therefore, regardless of whether CECL was intended to impact capital adequacy in banks, there is a perception that
it does.

Impact on Loan Origination

Another topic we discussed with our interviewees was the possibility of CECL affecting the quantity and quality of
loans being originated by banks. Our interviewees did not think that CECL had an impact on loan origination. The fol-
lowing quote from one of our analysts is representative of the views held by our interviewees:
Fortunately, I don’t think it really has changed meaningfully…Obviously, the economic effects are unchanged
by CECL. The revenue from the loan pool is going to be what it is and the losses are going to be what they
are. The accounting is just going to even out over time. However, if you think about it just conceptually, if it
were going to modify behavior, what would it do? It would make my 30-year mortgages less attractive to
make, it would make 10 or 12-year student loans a lot less attractive to make, it would make shorter term loans
more attractive. Nobody said, “We’re deemphasizing home mortgages now,” and you didn’t see a meaningful
change in trends. (SA 2)

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 11

Therefore, despite concerns from some individuals about loan origination being negatively impacted by the CECL stan-
dard (Schroeder 2023), overall, our interviewees did not seem to think that CECL has had any material impact on loan
origination practices or outcomes.

Proliferation of Nonregulated Lending

Another operational impact described by interviewees relates to a possible shift from traditional to nonregulated
lending. Specifically, CECL may make regulated banks less competitive and create opportunities for other market play-
ers not subject to the same CECL reporting requirements to benefit under this new accounting regime.
You’ve seen some companies go as far as say, “We’re going to sell loans, and we’re not going to balance sheet,
particularly in the student lending market.” And then you’ve seen others say, “We’re going to have to move to
an originate and sell model as opposed to a buy and hold”, because if you’re a FinTech where the market

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

doesn’t care about profitability, this is a great strategy for you. But if you’re a bank where the market upholds
you to a quarterly earnings number, it’s a lot more challenging. And I do think CECL…contributed to a lot of
the emergence of a lot of these FinTech oriented players trying to come into the market because in addition to
them not being subject to regulation, they’re also not subject to the same accounting standards, at least some
of them are not…it creates an arbitrage. (SA 6)
I think where it creates a lot of real-world impact is the move out of regulated banking space for certain things.
[There was a] private equity guy who was like, “Yeah, we love CECL. It’s going to move lending into private
equity and other spaces where you’re not going to have that [standard]”…When we think about consequences
that would be one thing I think should be identified…the relationship between regulated and unregulated
spaces, I don’t think CECL helps with that. (TG 1)
From our scan of the contemporary literature on the CECL standard, we did not observe the topic of a shift toward
unregulated spaces being mentioned. An Equifax report (deRitis 2018) documents the proliferation of nonregulated
lending even prior to the implementation of CECL and predicts that CECL will reduce the advantage of unregulated
lenders over regulated lenders in the credit market. This prediction is at odds with the opinions expressed above by our
interviewees. The emerging opportunity for lending in unregulated spaces described by our interviewees should be of
concern to regulators and standard setters as an unintended and potentially dangerous consequence of the CECL
standard.

Tension and Politics in the CECL Standard Setting Process

An overarching theme that emerged strongly from our interviews, that was not embedded in our original semistruc-
tured interview script, related to the political nature of the CECL standard setting process. Although the notion that the
FASB’s standard setting process is political is not new (e.g. Watts and Zimmerman 1978; Fogarty et al. 1994), interview-
ees noted that because the motivation for CECL was borne out of the 2007–2009 financial crisis, the political nature of
the CECL standard setting process was heightened. In fact, more than any other standard before, interference from
Congress spurred by special interests and lobbyists played a major role in the shaping and adoption of the CECL stan-
dard, potentially setting a precedent that accounting standards can be molded by power and money (Clor-Proell and
White 2020).
With public scrutiny at an all-time high, regulators and standard setters such as the FASB felt compelled to demon-
strate that they were willing to take decisive action to prevent such a crisis from occurring in the future.
Well, part of it when they had the financial crisis committees that were within the banking industry, or within
all governments or something back in the 2009 timeframe, accounting was made to be a scapegoat there, at
least in the US. (TG 2)
I feel like going into the design and the whole concept of CECL was kind of a bias that reserves were too low
for the biggest banks heading into the last financial crisis. (SA 4)
To address this issue of “too little, too late,” both the FASB and the IASB began drafting proposals for a new standard
to impose on the banking industry. However, after many years and attempts at producing converged standards, ulti-
mately, they could not agree on a converged standard on loan loss provisions, as recounted by one of our financial jour-
nalist interviewees:
By the time 2011 rolled around…they were still trying to write a converged accounting standard. But one year
later that effort fell apart…it was a big blow to this wider effort to create global accounting standards, but also

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
12 Bable, Wong, and Wynes

was a blow to well, “this is our one thing we needed to do after the financial crisis,” and they just couldn’t get
it done. (FJ 1)
As a result of the FASB/IASB’s inability to create a converged standard on loan loss reserves, the FASB decided to cre-
ate a standard on their own. However, their failure to collaborate with the IASB to create a converged standard despite
years of work may have undermined the FASB’s credibility as a standard setter in the eyes of their constituents (Fogarty
et al. 1994; Durocher, Fortin, Allini, and Zagaria 2019). This likely propagated the desire and motivation for the FASB
to push forward with a new standard for loan loss reserves.
A number of interviewees questioned whether a CECL-like standard was even necessary to prevent a future finan-
cial crisis given other regulatory changes in the banking industry:
The industry is a lot more capitalized than it used to be. Loans themselves tend to have less leverage than they
did in 2006 and 2007 heading into the great financial crisis. CECL never addressed that. CECL just created a

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

new way of reserving that didn’t look at the actual risk. (SA 4)
Clearly, the Federal Reserve made up their mind that they wanted more capital in the system, and the Dodd
Frank Act Stress Tests—DFAST, CCAR…was obviously one way to make that happen. But then CECL was
another way. (SA 6)
As these and other interviewees implied, in the post-financial crisis period, both legislation and regulatory controls were
enacted to address the perception of capital inadequacy in the U.S. banking system in the pre-financial crisis period.
Currently, large banks in the U.S. must submit annually to two different reviews by the Federal Reserve to assess their
capital reserves and their capital planning.9 Multiple interviewees suggested that CECL is yet another measure aimed at
increasing capital reserves in addition to these already adequate regulatory controls.
However, as one of our financial journalist interviewees notes, the inertia to pass a standard immediately may have
pushed the standard forward: “We have to get this done. This is our mandate post-financial crisis. And it’s 2016, eight
years past the crisis. Why haven’t we gotten this done?” (FJ 1). Consistent with this view, some other interviewees also
commented on the FASB’s seeming need to push something through to appease stakeholders that something had been
done to force banks to do better:
The past FASB leadership, which was around until spring of 2020, they were a little bit political or maybe a
lot political and a little self-serving and they just had an agenda that they were cramming through. And you
saw it in CECL. (SA 4)
This quote underscores that the politics surrounding the standard setting process is pervasive, affecting even the FASB
leadership. This analyst’s view is consistent with Allen and Ramanna (2013), who find associations between personal
characteristics (and their associated implicit biases and preferences) of FASB and SEC members and the standards that
they propose.
As part of the FASB’s standard setting process, they solicit feedback on exposure drafts from the public in the com-
ment letter period, in a sense inviting politicking and lobbying into the standard setting process (Fogarty et al. 1994).
Our financial journalist interviewees recount the chaos of the comment letter period:
Even though I think the standard is probably an improvement on the incurred loss model, no one in the
banking industry supported it, and…it made it very hard to implement this, to pass this with credibility, and
that’s probably something that FASB should be a little bit more mindful of. (FJ 2)
I’m pretty sure CECL, even though it was much watched…This is something that the board had to get done.
It didn’t pass through FASB unanimously. I think it was just two people who dissented. But that’s still
significant…that tells you something. (FJ 1)
Some interviewees noted that despite the FASB holding meetings and soliciting opinions from stakeholders, they did
not provide meaningful voice to these stakeholders. One analyst tersely stated, “The rules are what they are, and I can’t
do anything to change them” (SA 3). Other analysts shared similar views:
I think my personal view was that they wanted to get it in, and the FASB because they have their processes
made a decision they wanted. I found that they were a little standoffish and closed towards some of the

9
The first, the Comprehensive Capital Analysis and Review (CCAR) was mentioned by multiple interviewees. Instituted in 2010, CCAR tests banks’
capital adequacy and includes a review of banks’ internal processes regarding capital adequacy (Zhang 2013). It also allows the Federal Reserve to
limit banks’ dividends and share repurchases to maintain desired capital levels. The second capital adequacy review banks undergo is the Dodd
Frank Act Stress Tests (DFAST). Interviewees also mentioned DFAST, wherein banks are tested against three increasingly harsh scenarios to assess
their capital adequacy in the event of adverse economic conditions (Zhang 2013).

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 13

suggestions that the industry was making that would’ve made it more helpful for all of us…that would’ve
made it easier for readers of financial statements to be able to utilize. (SA 6)
This argument is not new, and in fact is reminiscent of a similar statement made in one of the CFA Society’s comment
letters presented to the FASB prior to CECL implementation: “It is our view that the US GAAP model chosen was the
preference of regulators not investors” (Chartered Financial Analyst Institute 2019). In a similar vein, one of our trade
group interviewees commented that with regards to the FASB soliciting feedback during the design and implementation
phase of CECL, “We feel heard, but not listened to” (TG 3).
The tension and politics in the design and implementation of CECL described above seemed to drive a generally
negative tone toward the standard that permeated analysts’ responses to our other interview questions. Although some
analysts initially presented diplomatic and/or more neutral tones toward CECL, “I think the spirit of the law has good
intentions” (SA 6), others were much clearer in their dislike of the standard, “I don’t like CECL. It’s dumb. It’s procycli-
cal” (SA 3). This generally negative sentiment reflected in the above quotes is consistent with Hashim, Li, and O’Hanlon

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

(2022) who document largely negative reactions to CECL in their analysis of comment letters. However, despite this
generally negative tone, many of our interviewees ultimately held very pragmatic views surrounding the impact of
CECL and had resigned to accept CECL. “We’re not going to fight city hall. We’re just going to pay the freight and go
on” (SA 4).

V. DISCUSSION
This study examines the decision-usefulness of CECL, an accounting standard once described as “the most sweeping
change to bank accounting ever.”10 Many of our interviewees expressed the belief that they were not listened to during
the FASB’s consultation period, resulting in a standard that does not suit their needs. Most importantly, the consensus
among our analyst interviewees was that CECL does not provide decision-useful information through either its qualita-
tive or quantitative disclosures. Our interviewees also voiced frustration that comparability across issuers has eroded
under CECL. This mirrors some research that suggests CECL made financial reports less informative and decreased
earnings quality (Bonaldi, Liang, and Yang 2023; Bonsall et al. 2023), but it is at odds with other research that shows
expected loss models generally (Wheeler 2021) and CECL specifically (Gee et al. 2024) provide decision-useful informa-
tion to capital markets.
This discrepancy may be resolved by the observation that even if information is intrinsically significant and relevant,
it is only useful to the extent users understand it. Our interviewees suggest that many users do not know how to interpret
the information being disclosed under CECL. Considering the mixed evidence, more research is needed, and our inter-
viewees provided insights into what could be examined. For example, many expressed that the standard was too com-
plex for most users to understand. However, now that CECL has been adopted across all public firms, has its ubiquity
forced users to increase their understanding of it? If so, has its decision-usefulness improved in recent periods?
Interviewees were also skeptical about CECL’s forward-looking nature and whether banks could accurately estimate
future loan losses. Some interviewees said that the use of forecasts in these estimates would spur procyclical behavior,
much like Covas and Nelson’s (2018) finding that retrofitting CECL onto banks in 2007–2009 would have exacerbated
problems during the housing crisis. Relatedly, it was noted that the sources of forecasts used by firms could influence
loan-loss provisions, and many expressed concern that Moody’s has a near-monopoly in the forecast market. All these
concerns seem apt given Vidinova’s (2023) discussion of macroeconomic forecasts used by banks being biased by fore-
casters’ overreaction to news. Does cross-sectional variability across forecast sources significantly affect banks’ finan-
cials? Is there a herding of forecasts across providers? Does Moody’s primacy in forecasting affect systemic risk in
banking? These are all questions future research could address.
We also asked our interviewees about the operational impacts of CECL. Topics from this discussion included the
effect on capital levels at banks, the negligible impact on loan origination, and the potential for a shift to nonregulated
entities in the lending space. Regarding the effect on capital levels, our interviewees perceived that, intended or not,
CECL would increase capital needs at banks. Future research could examine whether bank capital levels have risen
under CECL, as predicted here. On loan origination, our analysts had not seen and did not expect any impact from
CECL. However, some research suggests initial modest effects on lending with reductions in loan growth (Chen et al.
2025), fewer defaults (Kim et al. 2023), and minor increases in interest rates for some loans (Granja and Nagel 2024).
These effects could have been transitory phenomena for early adopters and should be reexamined as more data becomes
available. As to the potential for lending to shift to nonregulated entities, research has not yet explored this idea. There

10
See “ABA Position” on CECL: https://www.aba.com/advocacy/our-issues/cecl-implementation-challenges (last accessed May 31, 2023).

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
14 Bable, Wong, and Wynes

has been a proliferation of unregulated lending in recent years (deRitis 2018), but is there a loss in market share for lend-
ers that must comply with CECL? Will there be increased lending from firms that originate loans but do not hold them?
Consistent with prior research, our interviewees described the standard-setting process for CECL as political with dif-
ferent groups lobbying for their priorities (Fogarty et al. 1994; Durocher et al. 2007, 2019; Pelger 2016). Despite the repu-
tation of users of financial statements as docile participants in standard-setting (Young 2006; Durocher and Gendron
2011), the major reforms being proposed elicited numerous adversarial comment letters as the FASB developed proposal
drafts and even after the accounting standards update in 2016 (Hashim et al. 2022). To lend procedural legitimacy to the
process (Durocher et al. 2007; Jorissen, Lybaert, Orens, and Van der Tas 2013), the FASB delayed the passage of CECL
and solicited additional input from stakeholders, including users of financial statements. However, according to several
interviewees whose opinions were sought by the FASB, their input was “heard, but not listened to” (TG 3).
Finally, we note the limitations of our study. We recognize that the modest number of interviewees may not be rep-
resentative of the full population of sell-side analysts. Thus, although we provide evidence on sell-side analysts’ percep-

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

tions of CECL’s decision usefulness and make contributions to several different streams of accounting literature, future
field-based research could expand on our contributions by interviewing other financial statement users such as sell-side
analysts covering smaller banks, buy-side analysts, institutional investors, and retail investors. Another limitation of our
study is that to answer the FASB’s call to examine the decision usefulness and operational impacts of the CECL stan-
dard, we embedded specific themes in our interview prompts, which shaped and likely reduced the variety of themes dis-
cussed by interviewees. A final limitation is that the design of our questionnaire and the selection of participants was
focused on the banking industry as CECL was likely to have the most significant impact there. However, CECL applies
to a broad scope of financial instruments and is adopted by all publicly traded companies (i.e., banking and nonbanking
entities). Thus, although we believe the views of our interviewees apply generally toward the standard, it is possible that
they do not represent the views of all users, especially those outside the banking industry.

REFERENCES
Allen, A., and K. Ramanna. 2013. Towards an understanding of the role of standard setters in standard setting. Journal of
Accounting and Economics 55 (1): 66–90. https://www.sciencedirect.com/science/article/abs/pii/S0165410112000511
Bank of America Corporation. 2022. 2021 Annual Report. Charlotte, NC: Bank of America Corporation.
Beatty, A., and S. Liao. 2021. What do analysts’ provision forecasts tell us about expected credit loss recognition? The Accounting
Review 96 (1): 1–21. https://doi.org/10.2308/TAR-2018-0049
Bonaldi, P., P. Liang, and L. Yang. 2023. Did CECL improve banks’ loan loss provisions and earnings quality during the
COVID-19 pandemic? (Working paper). https://ssrn.com/abstract=4360356
Bonsall, S. B., B. Schmidt, and B. Xie. 2023. Current expected credit loss (CECL) model and analyst forecasts. (Working paper).
https://business.columbia.edu/sites/default/files-efs/imce-uploads/CEASA/Events%20Page/current_expected_credit_loss_cecl_
model_analyst_forecasts.pdf
Brown, L. D., A. C. Call, M. B. Clement, and N. Y. Sharp. 2015. Inside the ‘black box’ of sell-side financial analysts. Journal of
Accounting Research 53 (1): 1–47. https://doi.org/10.1111/1475-679X.12067
Bushman, R. M., and C. D. Williams. 2012. Accounting discretion, loan loss provisioning, and discipline of banks’ risk-taking.
Journal of Accounting and Economics 54 (1): 1–18. https://doi.org/10.1016/j.jacceco.2012.04.002
Bushman, R. M., and C. D. Williams. 2015. Delayed expected loss recognition and the risk profile of banks. Journal of
Accounting Research 53 (3): 511–553. https://doi.org/10.1111/1475-679X.12079
Byard, D., Y. Li, and Y. Yu. 2011. The effect of mandatory IFRS adoption on financial analysts’ information environment.
Journal of Accounting Research 49 (1): 69–96. https://doi.org/10.1111/j.1475-679X.2010.00390.x
Campbell, D., and R. Slack. 2011. Environmental disclosure and environmental risk: Sceptical attitudes of UK sell-side bank ana-
lysts. The British Accounting Review 43 (1): 54–64. https://doi.org/10.1016/j.bar.2010.11.002
Covas, F., and W. Nelson. 2018. Current expected credit loss: Lessons from 2007–2009. (Working paper). https://doi.org/10.2139/
ssrn.3215559
Chartered Financial Analyst Institute. 2019. CL-UNS Comment Letter No. 1. Cumulative Expected Credit Loss (“CECL”)
Impairment Model.
Chen, J., Y. Dou, S. Ryan, and Y. Zou. 2025. The effect of the current expected credit loss approach on banks’ lending during
stress periods: Evidence from the COVID-19 recession. The Accounting Review 100 (1): 113–138. https://doi.org/10.2308/
TAR-2022-0275
Clor-Proell, S., and B. White. 2020. Congress should leave accounting standards to FASB. Accounting Today (May 18). https://
www.accountingtoday.com/opinion/congress-should-leave-accounting-standards-to-fasb
deRitis, C. 2018. The potential impact of CECL. (October 2). Equifax Business Insights Blog. https://www.equifax.com/business/
blog/-/insight/article/potential-impact-of-cecl-consistency-in-loss-reserving-and-risk-management/

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 15

Durocher, S., and Y. Gendron. 2011. IFRS: On the docility of sophisticated users in preserving the ideal of comparability.
European Accounting Review 20 (2): 233–262. https://doi.org/10.1080/09638181003687869
Durocher, S., and O. Georgiou. 2022. Framing accounting for goodwill: Intractable controversies between users and standard set-
ters. Critical Perspectives on Accounting 89: 102357. https://doi.org/10.1016/j.cpa.2021.102357
Durocher, S., A. Fortin, and L. C^ ote. 2007. Users’ participation in the accounting standard-setting process: A theory-building
study. Accounting, Organizations and Society 32 (1–2): 29–59. https://doi.org/10.1016/j.aos.2006.03.004
Durocher, S., A. Fortin, A. Allini, and C. Zagaria. 2019. Users’ legitimacy perceptions about standard-setting processes.
Accounting and Business Research 49 (2): 206–243. https://doi.org/10.1080/00014788.2018.1526664
FASB. 2013. Proposed Accounting Standards Update, Financial Instruments-Credit Losses (Subtopic 825-15). Feedback
Summary. Norwalk, CT: FASB.
FASB. 2016. FASB Accounting standards updates No. 2016-13. June 2016. Financial Instruments—Credit Losses (Topic 326)
Measurement of Credit Losses on Financial Instruments. Norwalk, CT: FASB.
FASB. 2018. Statement of Financial Accounting Concepts No. 8 Conceptual Framework for Financial Reporting. August 2018.

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

Chapter 3, Qualitative Characteristics of Useful Financial Information. Norwalk, CT: FASB.
Fogarty, T. J., M. E. Hussein, and J. E. Ketz. 1994. Political aspects of financial accounting standard setting in the USA.
Accounting, Auditing & Accountability Journal 7 (4): 24–46. https://doi.org/10.1108/09513579410069830
Francis, J. J., M. Johnston, C. Robertson, L. Glidewell, V. Entwistle, M. P. Eccles, and J. M. Grimshaw. 2010. What is an adequate
sample size? Operationalising data saturation for theory-based interview studies. Psychology & Health 25 (10): 1229–1245. https://
doi.org/10.1080/08870440903194015
Gassen, J., and K. Schwedler. 2010. The decision usefulness of financial accounting measurement concepts: Evidence from an
online survey of professional investors and their advisors. European Accounting Review 19 (3): 495–509. https://doi.org/
10.1080/09638180.2010.496548
Gee, K., J. Neilson, B. Schmidt, and B. Xie. 2024. Are recognized expected credit losses decision-useful and new to investors?
Evidence from CECL adoption. (Working paper). https://ssrn.com/abstract=4038479
Georgiou, O. 2018. The worth of fair value accounting: Dissonance between users and standard setters. Contemporary Accounting
Research 35 (3): 1297–1331. https://doi.org/10.1111/1911-3846.12342
Georgiou, O., E. Mantzari, and J. Mundy. 2021. Problematising the decision usefulness of fair values: Empirical evidence from
UK financial analysts. Accounting and Business Research 51 (4): 307–346. https://doi.org/10.1080/00014788.2020.1814687
Granja, J., and F. Nagel. 2024. Current expected credit losses and consumer loans. (Working paper). https://ssrn.com/
abstract=4324118
Guest, G., A. Bunce, and L. Johnson. 2006. How many interviews are enough? An experiment with data saturation and variabil-
ity. Field Methods 18 (1): 59–82. https://doi.org/10.1177/1525822X05279903
Gurd, B. 2008. Remaining consistent with method? An analysis of grounded theory research in accounting. Qualitative Research
in Accounting & Management 5 (2): 122–138. https://doi.org/10.1108/11766090810888926
Hashim, N., W. Li, and J. O’Hanlon. 2019. Reflections on the development of the FASB’s and IASB’s expected-loss methods of
accounting for credit losses. Accounting and Business Research 49 (6): 682–725. https://doi.org/10.1080/00014788.2018. 1526665
Hashim, N., W. Li, and J. O’Hanlon. 2022. The development of expected-loss methods of accounting for credit losses: A review
with analysis of comment letters. Accounting Horizons 36 (3): 71–102. https://doi.org/10.2308/HORIZONS-19-117
Helper, S. 2000. Economists and field research: ‘You can observe a lot just by watching’. American Economic Review 90 (2): 228–232.
https://doi.org/10.1257/aer.90.2.228
Ho, S. 2021. Group tells SEC Chair Gensler that accounting standard-setter is ignoring investor views. Thomson Reuters Tax &
Accounting News (April 23). https://tax.thomsonreuters.com/news/group-tells-sec-chair-gensler-that-accounting-standard-set-
ter-is-ignoring-investor-views/
Horton, J., G. Serafeim, and I. Serafeim. 2013. Does mandatory IFRS adoption improve the information environment?
Contemporary Accounting Research 30 (1): 388–423. https://doi.org/10.1111/j.1911-3846.2012.01159.x
Ittner, C. 2014. Strengthening causal inferences in positivist field studies. Accounting, Organizations and Society 39 (7): 545–549.
https://doi.org/10.1016/j.aos.2013.10.003
J.P. Morgan Chase. 2022. 2021 Annual Report. New York, NY: J.P. Morgan Chase.
Jorissen, A., N. Lybaert, R. Orens, and L. Van der Tas. 2013. A geographic analysis of constituents’ formal participation in the
process of international accounting standard setting: Do we have a level playing field? Journal of Accounting and Public
Policy 32 (4): 237–270. https://doi.org/10.1016/j.jaccpubpol.2013.04.005
Kenno, S., S. McCracken, and S. Salterio. 2017. Financial reporting interview-based research: A field research primer with an
illustrative example. Behavioral Research in Accounting 29 (1): 77–102. https://doi.org/10.2308/bria-51648
Kim, S., S. Kim, A. Kleymenova, and R. Li. 2023. Current expected credit losses (CECL) standard and banks’ information pro-
duction. (Working paper). https://ssrn.com/abstract=4117869
Koonce, L., C. Mongold, L. Quaid, and J. Winchel. 2024. How do forward-looking estimates of credit losses affect the judgments
of financial-statement users? (Working paper). https://ssrn.com/abstract=4099729

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
16 Bable, Wong, and Wynes

Laeven, L., and G. Majnoni. 2003. Loan loss provisioning and economic slowdowns: Too much, too late? Journal of Financial
Intermediation 12 (2): 178–197. https://doi.org/10.1016/S1042-9573(03)00016-0
L
opez-Espinosa, G., G. Ormazabal, and Y. Sakasai. 2021. Switching from incurred to expected loan loss provisioning: Early evi-
dence. Journal of Accounting Research 59 (3): 757–804. https://doi.org/10.1111/1475-679X.12354
Mahieux, L., H. Sapra, and G. Zhang. 2023. CECL: Timely loan loss provisioning and bank regulation. Journal of Accounting
Research 61 (1): 3–46. https://doi.org/10.1111/1475-679X.12463
Moody’s Analytics. 2022. Current expected credit losses model (CECL). https://www.moodysanalytics.com/solutions-overview/
regulatory-accounting/current-expected-credit-loss-model-cecl
Pelger, C. 2016. Practices of standard-setting–an analysis of the IASB’s and FASB’s process of identifying the objective of finan-
cial reporting. Accounting, Organizations and Society 50: 51–73. https://doi.org/10.1016/j.aos.2015.10.001
Ronen, J. 2023. The discount rate used in the current expected credit loss standard creates accounting losses where there are no
economic losses. Journal of Accounting, Auditing and Finance 38 (2): 225–232. https://doi.org/10.1177/0148558X211034959
Schroeder, R. H. 2023. Research opportunities to confirm or dispel myths of CECL: Looking past the headlines to actual results.

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

Journal of Accounting, Auditing, & Finance 38 (2): 233–270. https://doi.org/10.1177/0148558X221148299
Soltes, E. 2014. Private interaction between firm management and sell-side analysts. Journal of Accounting Research 52 (1): 245–272.
https://doi.org/10.1111/1475-679X.12037
Tan, H., S. Wang, and M. Welker. 2011. Analyst following and forecast accuracy after mandated IFRS adoptions. Journal of
Accounting Research 49 (5): 1307–1357. https://doi.org/10.1111/j.1475-679X.2011.00422.x
Vidinova, H. 2023. Forward-looking loan loss provisioning under imperfect forecasts. (Working paper). https://ssrn.com/
abstract=4628656
Vyas, D. 2011. The timeliness of accounting write-downs by U.S. financial institutions during the financial crisis of 2007–2008.
Journal of Accounting Research 49 (3): 823–860. https://doi.org/10.1111/j.1475-679X.2011.00410.x
Watts, R. L., and J. L. Zimmerman. 1978. Towards a positive theory of the determination of accounting standards. The
Accounting Review 53 (1): 112–134.
Wheeler, B. 2021. Unrecognized expected credit losses and bank share prices. Journal of Accounting Research 59 (3): 805–866.
https://doi.org/10.1111/1475-679X.12353
White, N., and D. Hood. 2021. Loan-loss accounting rule worked in 2020, but banks hate it. Bloomberg Tax. https://news.bloom-
bergtax.com/financial-accounting/loan-loss-accounting-rule-worked-in-2020-but-banks-hate-it
Young, J. 2006. Making up users. Accounting, Organizations and Society 31 (6): 579–600. https://doi.org/10.1016/j.aos.2005.12.005
Young, J. 2014. Separating the political and the technical: Accounting standard-setting and purification. Contemporary
Accounting Research 31 (3): 713–747. https://doi.org/10.1111/1911-3846.12046
Zerbe, W. J., and D. L. Paulhus. 1987. Socially desirable responding in organizational behavior: A reconception. The Academy of
Management Review 12 (2): 250–264. https://doi.org/10.2307/258533
Zhang, J. 2013. CCAR and beyond: Capital assessment, stress testing and applications—An introduction. (White paper). https://books.
google.ca/books/about/CCAR_and_Beyond.html?id=XVKfngEACAAJ&redir_esc=y

APPENDIX A
Semistructured Interview Script

1. Can you give a description of what the CECL standard is and what it is trying to accomplish?
2. Is the information decision-useful?
a. Did CECL provide users like yourself with decision-useful information to assess expected credit losses?
i. Does having more forward-looking information (i.e., macroeconomic forecasts) give you a better
sense of what to expect with regards to loan collectability?
ii. Did CECL address the “too little, too late” concern? (i.e., Does recognizing the expected credit losses
on Day-1 provide users with better information regarding the measurement of the credit losses?)
b. Has the comparability of information across companies or loan types changed?
i. Did CECL increase comparability across all institutions previously subject to different regulatory
requirements?
c. Disclosures
i. Do the required disclosures achieve what you believe is the disclosure objective?
ii. Are entities’ qualitative disclosures informative or boilerplate?
iii. Do entities’ quantitative disclosures convey useful information?
iv. Are there disclosures that could be removed or should be added?

(continued on next page)

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
Decision Usefulness of CECL 17

APPENDIX A (continued)

v. Do the disclosures provide transparency on how an entity determines the allowance for credit losses?
vi. Are the disclosures used to improve the comparability of the allowance for credit loss information
across entities?
3. Have there been other changes as a result of implementing CECL?
a. Did CECL change the quantity/quality of loans?
b. Has CECL lead to operation improvements within banks?
c. Have you seen any other changes as a result of implementing CECL?
d. Did CECL increase volatility of loan loss provisions? If Yes, are banks able to successfully explain this
change in volatility?
e. Will CECL increase the pro-cyclicality in lending and risk-taking at banks?
f. Does CECL increase the transparency of bank operations?

Downloaded from http://publications.aaahq.org/bria/article-pdf/doi/10.2308/BRIA-2023-027/118958/bria-2023-027.pdf by guest on 17 April 2025

g. Does CECL serve an effective monitoring role for banks?
h. Have you observed any biased or opportunistic disclosures relating to CECL to manage: earnings, loan
loss provision, and/or investor expectations?
4. If you could make one change to improve the current CECL standard, what would it be?

Behavioral Research in Accounting

Volume XX, Number XX, 20XX
 Calculating Lifetime Expected Loss for IFRS 9:
Which Formula is Correct?

Bernd Engelmann1

This version: April 30, 2018

ABSTRACT

IFRS 9, the new accounting rules for financial instruments require banks to build pro-
visions for expected losses in their loan portfolios. A requirement which was not present
in previous regulation is the necessity to provision the expected loss over a loan’s lifetime
in case a loan shows a deterioration in credit quality. The IFRS 9 rules are formulated in
a qualitative way and no explicit formulas or precise parameter estimation methods are
prescribed. In this article, lifetime expected loss is computed as the difference in present
values of a loan’s cash flows. It is assumed that cash flows are risk-free in the first step
and expected present values including credit risk are subtracted in the second step to ar-
rive at lifetime expected loss. This is done under different modeling assumptions and the
outcome is compared to the weighted loss formula most commonly used in practice (e.g.
Deloitte (2017) or PricewaterhouseCoopers (2015)). It turns out that the formula used in
practice should at least be adjusted to be theoretically more sound or be replaced entirely
by present value formulas to assure accuracy.

JEL Classification: G12

Keywords: IFRS 9, Provisioning, Lifetime Expected Loss

1 Open University Ho Chi Minh City, e-mail: [email protected] and Open Source Investor
Services B.V., e-mail: [email protected]

Electronic copy available at: https://ssrn.com/abstract=3238632

 The accounting rules IFRS 9, published in 2014 (IASB 2014), require banks to improve
their quantitative modeling for calculating loan loss provisions. For loan exposures considered
as normally performing, expected loss is provisioned on a one-year basis using the well-known
formula ECL = PD · LGD · EAD where ECL is the expected credit loss, PD the probability
of a borrower default, LGD the loss given default and EAD the exposure at default. Note,
that under IFRS 9 forward looking risk measures are required, i.e. the estimates for PD,
LGD, and EAD have to be point-in-time estimates. This is in contrast to the Basel framework
where for minimum capital calculations under the IRB approach through-the-cycle default
probabilities are permitted and downturn LGD/EAD is required (Basel Committee on Banking
Supervision 2006). For estimating point-in-time versus through-the-cycle PD see Aguais,
Forest, Wong, and Diaz-Ledezma (2004) or Carlehed and Petrov (2012).
The more challenging part of IFRS 9 is the situation where the credit quality of a loan
deteriorates. In this case it is required to provision expected loss over a loan’s lifetime instead
of one-year expected loss. A common way do compute lifetime expected loss in practice is
using the formula for one-year ECL and computing lifetime ECL as the present value over all
one-year ECLs in future periods:
" #
n
1
ECL = E ∑ i
· Pr(τ = i) · LGDi · EADi , (1)
i=1 (1 + r)

where r is the discount rate, i = 1, . . . , n are a loan’s periods, i.e. the years or quarters until it
matures, and Pr(τ = i) is the probability that a borrower’s default time τ is in period i implying
that it survived the periods j = 1, . . . , i − 1, see e.g. Deloitte (2017), PricewaterhouseCoopers
(2015), Skoglund (2017) and Xu (2016).
In the sequel, ECL will be computed in terms of a term-structure pi which is the uncondi-
tional probability that a borrower defaults in one of the periods j = 1, . . . , i. To link Pr(τ = i)
with pi , note that

Pr(τ = i) = Pr(τ = i|τ > i − 1) · Pr(τ > i − 1) = (1 − Pr(τ > i|τ > i − 1)) · Pr(τ > i − 1)
= (1 − Pr(τ > i)/Pr(τ > i − 1)) · Pr(τ > i − 1) = Pr(τ > i − 1) − Pr(τ > i)
= Pr(τ ≤ i) − Pr(τ ≤ i − 1) = pi − pi−1 .

Furthermore, the term-structure of survival probabilities qi will play a major role which is
defined as qi = 1 − pi . Denote the default probability in period i conditional on survival until
period i − 1 with PDi := Pr(τ = i|τ > i − 1). The relation between qi and PDi is
i
qi = Pr(τ > i) = Pr(τ > i|τ > i − 1) · Pr(τ > i − 1) = ∏ Pr(τ > j|τ > j − 1) · Pr(τ > 0)
j=1
i i
= ∏ (1 − Pr(τ = j|τ > j − 1)) · 1 = ∏ (1 − PDi ).
j=1 j=1

Electronic copy available at: https://ssrn.com/abstract=3238632

 Unless no complex dependence structures are modeled between PD, LGD and EAD a
scenario for term-structures of default probabilities, LGD and EAD in each period are esti-
mated independently mostly using macroeconomic models and ECL is computed from these
quantities as
n
1
ECL = ∑ i
· (pi − pi−1 ) · LGDi · EADi , (2)
i=1 (1 + r)
One reason for the popularity of this approach is its transparency. Expected loss is computed
separately in each period and aggregated by a simple rule over lifetime. The estimation of risk
parameters for this purpose is discussed in the aforementioned articles of Skoglund (2017) and
Xu (2016).
The main purpose of this article is not to discuss the parameterization of (2) but to ask
whether this formula is appropriate for computing lifetime ECL and to shed light on its the-
oretical foundation. The starting point for answering this question is paragraph B5.5.29 in
IASB (2014). It says: ”For financial assets, a credit loss is the present value of the difference
between: (a) the contractual cash flows that are due to an entity under the contract; and (b) the
cash flows that the entity expects to receive.”. In the next section, two formulas will be derived
that compute lifetime ECL for a simple bullet loan applying B5.5.29 literally using present
values under different assumptions on the estimation of LGD. In the following section, a link
between these two formulas and (2) will be derived. It will be shown that (2) which is used
widely in practice has no theoretical justification but has to be adjusted for being consistent
with a present values approach. In Section 3 it will be outlined how the formulas will change
under the inclusion of prepayment probabilities. After that, the use of an effective instead of
the contractual interest rate for discounting will be discussed. This is a requirement accord-
ing to B5.5.44 in IASB (2014), where it says: ”Expected credit losses shall be discounted to
the reporting date, not to the expected default or some other date, using the effective interest
rate...”. It will be shown that the use of an interest rate different from the contractual rate for
discounting is inconsistent with a weighted loss formula like (2). A numerical example will
illustrate the impact of different assumptions and approaches.

1. Expected Lifetime Loss Based on Cash Flows

In this section a formula for ECL based on a loan’s future cash flows is derived. To keep the
exposition simple we use a bullet loan throughout this article. For the present value calculation
all future cash flows of the loan have to be considered. Future cash flows are the interest
rate payments, the payback of the loan’s balance at maturity, and the liquidation proceeds of
collateral in the case of a borrower default. We can split the present value into cash flows that

Electronic copy available at: https://ssrn.com/abstract=3238632

will be paid if a borrower survives and cash flows a bank receives in the case of a default. For
the survival part, the present value VS is given as
n
VS = ∑ N · z · δi · qi + N · δn · qn (3)
i=1

where N is the outstanding loan balance, z is the interest rate, and δi is the discount factor
corresponding to period i. The future interest periods until the end of a loan’s lifetime are
indexed i = 1, . . . , n as before. Since a borrower can pay interest and outstanding balance only
if he survives all discounted cash flows are weighted with survival probabilities.
For the default part, we distinguish two cases: In Case I a recovery rate was estimated
taking outstanding balance as the reference quantity, in the second case outstanding balance
plus the interest was used as exposure at default in recovery estimation. In Case I the resulting
recovery rate is denoted with R while in Case II we use R. An analogous notation will be
adopted for the corresponding LGD numbers. For Case I, we find
n
VD,I = ∑ N · Ri · δi · (qi−1 − qi ), (4)
i=1

while in Case II we obtain

n
VD,II = ∑ N · Ri · δi · (1 + z) · (qi−1 − qi ). (5)
i=1

In both cases it was assumed that a bank can claim the recovery payment at the end of the in-
terest period in case of a default. Recall that qi−1 −qi = pi − pi−1 = Pr(τ = i) is the probability
that a borrower defaults in period i.
The expected present value of all future cash flows is the sum of VS and VD . According to
B5.5.29 in IASB (2014) credit loss is defined as the difference of present values in contractual
cash flows and expected cash flows. The present value of contractual cash flows VC for a bullet
loan is simply
n
VC = ∑ N · z · δi + N · δn (6)
i=1
and ECL is defined as
ECL = VC −VS −VD . (7)

A crucial quantity in ECL calculation is the discount factor δ. For the moment, we assume
that discounting is done with the contractual interest rate z and discount factors are computed
as δi = (1 + z)−i . This is only approximately what IASB (2014) requires. We will comment

Electronic copy available at: https://ssrn.com/abstract=3238632

on this assumption later in Section 3. With this choice of the discount factor (6) simplifies
resulting in VC = N and
ECL = N −VS −VD . (8)

In the next section, we will use (8) as a starting point and derive an expression from it that
is using weighted losses in a similar form as the ECL formula used in practice (2).

2. Expected Lifetime Loss Based on Weighted Losses

We compute ECL for both versions of recovery rates (4) and (5) arriving at formulas similar
but not identical to (2). We start with Case I.

Proposition 1 When ECL is defined as ECL = N − VS − VD,I we can transform the present
values into a weighted sum of losses given as
n
1
ECL = ∑ i
· (pi − pi−1 ) · (LGDi + z) · N. (9)
i=1 (1 + z)

Proof:

ECL = N −VS −VD,I

n n
= N − ∑ N · z · δi · qi − N · δn · qn − ∑ N · Ri · δi · (qi−1 − qi )
i=1 i=1
n n n
= N − ∑ N · z · δi − N · δn + ∑ N · z · δi · pi + N · δn · pn − ∑ N · Ri · δi · (pi − pi−1 )
i=1 i=1 i=1

We have used qi = 1 − pi . The first part of the above expression N − ∑ni=1 N · z · δi − N · δn is

zero by construction because the discount rate is identical to the loan’s interest rate.
For the second term, we find
n n
∑ N · z · δi · pi + N · δn · pn = ∑ N · δi−1 · (pi − pi−1).
i=1 i=1

This expression can be proved by mathematical induction. For n = 1 we find

N · z · δ1 · p1 + N · δ1 · p1 = N · p1 · δ1 · (1 + z) = N · p1 = N · δ0 · (p1 − p0 ).

Electronic copy available at: https://ssrn.com/abstract=3238632

Note that δ0 = 1 and p0 = 0. Now assume, the relation is correct for n − 1 and show that under
this assumption it is also correct for n:
n n−1
∑ N · z · δi · p i + N · δn · p n = ∑ N · z · δi · pi + N · δn−1 · pn−1
i=1 i=1
− N · δn−1 · pn−1 + N · z · δn · pn + N · δn · pn
n−1
= ∑ N · δi−1 · (pi − pi−1) − N · δn−1 · pn−1 + N · δn−1 · pn
i=1
n
= ∑ N · δi−1 · (pi − pi−1 )
i=1

This allows us to finalize the calculation of the ECL formula using Ri = 1 − LGDi :
n n
ECL = ∑ N · δi−1 · (pi − pi−1 ) − ∑ N · (1 − LGDi ) · δi · (pi − pi−1 )
i=1 i=1
n

= ∑ N · (pi − pi−1 ) · δi−1 − δi + LGDi · δi
i=1
n
= ∑ N · (pi − pi−1 ) · δi · (1 + z − 1 + LGDi )
i=1
n
1
=∑ i
· N · (pi − pi−1 ) · (LGDi + z). 
i=1 (1 + z)

We find that the ECL formula is similar to (2) but not identical. Using identical risk
parameters leads to a higher ECL in (9) compared to (2) due to the correction for the loss in
interest.

Proposition 2 When ECL is defined as ECL = N − VS − VD,II we can transform the present
values into a weighted sum of losses given as
n
1
ECL = ∑ i−1
· (pi − pi−1 ) · LGDi · N. (10)
i=1 (1 + z)

ECL computed by (10) given identical risk parameters is smaller than ECL computed by (9).

Proof: Most steps for proving Proposition 1 can be reused for the proof of Proposition 2.
The only difference is in the term computing the present value of liquidation proceeds (5) in

Electronic copy available at: https://ssrn.com/abstract=3238632

case of a borrower default. Note, that δi · (1 + z) = δi−1 . Using this and results from the proof
of Proposition 1 we find

ECL = N −VS −VD,II

n n
= ∑ N · δi−1 · (pi − pi−1 ) − ∑ N · Ri · δi · (1 + z) · (qi−1 − qi )
i=1 i=1
n n
= ∑ N · δi−1 · (pi − pi−1 ) − ∑ N · (1 − LGDi ) · δi−1 · (pi − pi−1 )
i=1 i=1
n
= ∑ δi−1 · (pi − pi−1 ) · LGDi · N
i=1
n
1
=∑ i−1
· (pi − pi−1 ) · LGDi · N.
i=1 (1 + z)

If LGDi = LGDi the ECL in Proposition 2 is smaller. This can be easily seen by taking
differences of the terms containing LGDi :

LGDi + z LGDi LGDi + z − LGDi · (1 + z) z · (1 − LGDi )

− = = ≥ 0. 
(1 + z)i (1 + z)i−1 (1 + z)i (1 + z)i

In this section we have derived two versions of ECL formulas based on weighted loss
per period from present values of cash flows. Note, that both versions result in higher ECL
than the formula commonly used for lifetime ECL calculations under IFRS 9 in practice (2) if
identical risk parameters are used for the calculation. Whether the formula in Proposition 1 or
Proposition 2 is more appropriate depends on the way a bank is estimating LGD. If it is based
on outstanding balance only, then (9) should be used. If accrued interest is included, then (10)
is more accurate.

3. Generalizations

In this section two generalizations of the framework are discussed. The first is the inclusion of
prepayments. Prepayments are reducing the expected lifetime of a loan and therefore reduce
expected lifetime loss. Under IFRS 9 lifetime expected loss is only relevant for loans where
a deterioration of credit quality was observed. One would expect that voluntary prepayments
are not a big issue for these clients. However, in many loan markets one can empirically
verify that also loans that recently moved into an arrears status show material prepayments.
In these cases prepayments should be included in the framework to improve the accuracy of
ECL. In the second part of this section the discount rate used for calculating ECL is discussed
in more detail. In particular, the requirement in IFRS 9 of using an effective interest rate for

Electronic copy available at: https://ssrn.com/abstract=3238632

discounting which could be different from the loan’s contractual rate is reflected. A numerical
example for illustration concludes this section.

3.1. Including Prepayment Probabilities into ECL

The are two ways how prepayment can be modeled and prepayment probabilities can be de-
fined leading to different formulas and different ECL. Similar to the discussion on LGD mod-
eling on Section 1, one has to pick the framework that is consistent with the determination of
risk parameters to ensure consistent ECL calculation.
The first alternative is modeling prepayment conditional on survival, i.e. in each period i
the quantity PRi is the probability that a surviving borrower prepays. Denoting the random
prepayment time by η, we have PRi = Pr(η = i|τ ≥ i). The default probability of a borrower
conditional on neither default nor prepayment is PDi = Pr(τ = i|τ > i − 1, η > i − 1). In
this setup, a borrower prepays with probability (1 − PDi ) · PRi and continues the loan with
probability (1 − PDi ) · (1 − PRi ). The second alternative is modeling default and prepayment
as competing risks. Here, in each period a borrower can either default with probability PDi =
Pr(τ = i|τ > i − 1, η > i − 1), prepay with probability PRi = Pr(η = i|τ > i − 1, η > i − 1) or
continue servicing the loan with probability 1 − PDi − PRi . Both modeling alternatives lead
to different formulas for present values of future cash flows.
Similar to a term-structure of default probabilities pi = 1 − ∏ij=1 (1 − PDi ) a term-structure
of prepayment probabilities πi = 1 − ∏ij=1 (1 − PRi ) (or πi = 1 − ∏ij=1 (1 − PRi ), respectively)
can be estimated empirically using similar techniques as for PD models. Once this term-
structure is available, it can be included in the present value formulas. In this section only the
version using VD,I will be discussed. The result for VD,II can be derived easily from the result
for VD,I .
In the first step, (3) and (4) are generalized to include prepayment probabilities. Here, only
full prepayments are considered. The generalization of VS is given by
n
V̂S = ∑ δi · qi · (z · N · (1 − πi−1 ) + N · (πi − πi−1 )) . (11)
i=1

Here, 1 − πi−1 is the probability that there was no prepayment until period i − 1. Only if
the loan is not prepaid it can still pay interest and, therefore, interest rate payments have to
be weighted with this probability. The new term N · (πi − πi−1 ) is the expected prepayment
at time i, i.e. the outstanding balance weighted with the marginal prepayment probability in
period i. Since this term is multiplied with qi it represents the fraction of surviving borrowers
that prepay in period i. Note that πn = 1 since the outstanding balance has to be repaid at
maturity.

Electronic copy available at: https://ssrn.com/abstract=3238632

 For VD,I the extension is given as
n
V̂D,I = ∑ N · (1 − πi−1 ) · Ri · δi · (qi−1 − qi ) (12)
i=1

A loan can only default in period i if it has not yet been prepaid in a previous period. Therefore
the expected recovery has to be weighted with 1 − πi−1 .

Proposition 3 When including prepayment probabilities PR into ECL calculation using (11)
and (12) ECL can be computed as N − V̂S − V̂D,I and the present values can be transformed
into a sum of weighted losses given as
n
1
ECL = ∑ i
· (pi − pi−1 ) · (LGDi + z) · Ni , (13)
i=1 (1 + z)

where Ni is defined as Ni := N · (1 − πi−1 ).

Proof:
Recall that ECL was defined as the difference of the present value of contractual cash
flows and the expected present value of cash flows. First, it has to be verified that the present
value of contractual cash flows is still equal to N. Note, that prepayment risk is not a loss risk
and, therefore, prepayment probabilities have to be included in the present value calculation
of contractual cash flows:
n n
VC = ∑ δi · (z · N · (1 − πi−1 ) + N · (πi − πi−1 )) = ∑ δi · (z · Ni + Ni − Ni+1 )
i=1 i=1

Define M j = N j − N j+1 and note that Nn+1 = 0. We obtain

!
n n n j
n
VC = ∑ δi · z · ∑ M j + Mi = ∑∑ δi · z · M j + δ j · M j = ∑ M j = N1 − Nn+1 = N.
i=1 j=i j=1 i=1 j=1

At the second equality sign a change in summation order was performed. Similar steps are
applied to prove the main part of the proposition.
n
ECL = N − V̂S − V̂D,I = N − ∑ δi · qi (z · N · (1 − πi−1 ) + N · (πi − πi−1 ))
i=1
n
− ∑ N · (1 − πi−1 ) · Ri · δi · (qi−1 − qi )
i=1

Electronic copy available at: https://ssrn.com/abstract=3238632

 n n
= N − ∑ δi · qi (z · Ni + (Ni − Ni+1 )) − ∑ Ni · Ri · δi · (qi−1 − qi )
i=1
! i=1
n n n n n
= ∑ M j − ∑ δi · qi z · ∑ M j + Mi − ∑ ∑ M j · Ri · δi · (qi−1 − qi)
j=1 i=1 j=i i=1 j=i
" #
n j j
= ∑ M j − ∑ δi · qi · z · M j − δ j · q j · M j − ∑ M j · Ri · δi · (qi−1 − qi )
j=1 i=1 i=1
" #
n j
1
=∑ ∑ i
· (pi − pi−1 ) · (LGDi + z) · M j
j=1 i=1 (1 + z)
" #
n n
1
=∑ ∑ i
· (pi − pi−1 ) · (LGDi + z) · M j
i=1 j=i (1 + z)
n
1
=∑ i
· (pi − pi−1 ) · (LGDi + z) · Ni . 
i=1 (1 + z)

If an LGD using the convention in VD,II is applied, an analogous result could be obtained
carrying out exactly the same steps.
To derive the formulas for the second modeling alternative, define vi as the probability that
there is neither a default nor a prepayment in periods j = 1, . . . , i. As shown in Xu (2016),
it can be computed as vi = ∏ij=1 (1 − PD j − PR j ). Under these modeling assumptions, the
present value of future cash flows conditional on no default V̇S is given as
n

V̇S = ∑ δi · N z · vi + (1 + z) · vi−1 · PRi . (14)
i=1

The first term under the sum represent the interest rate payment in case the borrower continues
servicing the loan and the second term represents the payment of interest plus outstanding
balance in case of a prepayment. Note that in the final period n the loan must be repaid if the
borrower survives leading to PRn = 1 − PDn and vn = 0.
For the present values of cash flows in case of default V̇D,I , we find
n
V̇D,I = ∑ N · Ri · δi · vi−1 · PDi . (15)
i=1

A borrower can only default in period i if he has neither prepaid not defaulted earlier which is
captured in vi−1 .

Electronic copy available at: https://ssrn.com/abstract=3238632

Proposition 4 When including prepayment probabilities PR into ECL calculation using (14)
and (15) ECL can be computed as N − V̇S − V̇D,I and the present values can be transformed
into a sum of weighted losses given as
n
1
ECL = ∑ i
· (pi − pi−1 ) · (LGDi + z) · N i , (16)
i=1 (1 + z)

where N i is defined as N i := N · qvi−1

i−1
.

Proof:
The prepayment probabilities PRi and PRi are related by PRi = PRi · (1 − PDi ). Using this
equation allows a transformation of vi :
i i i
vi = ∏ (1 − PD j − PR j ) = ∏ (1 − PD j − PRi · (1 − PD j )) = ∏ (1 − PDi ) · (1 − PR j )
j=1 j=1 j=1
= qi · (1 − πi )

Using this to transform V̇S yields

n

V̇S = ∑ δi · N z · vi + (1 + z) · vi−1 · PRi
i=1
n

= ∑ δi · N z · vi−1 · (1 − PDi − PRi ) + (1 + z) · vi−1 · PRi
i=1
n

= ∑ δi · N z · vi−1 · (1 − PDi ) + vi−1 · PRi
i=1
n

= ∑ δi · N z · vi−1 · (1 − PDi ) − vi−1 · (1 − PDi − PRi − (1 − PDi ))
i=1
n
= ∑ δi · N (z · vi−1 · (1 − PDi ) − vi + vi−1 · (1 − PDi ))
i=1
n  
vi−1 vi vi−1
= ∑ δi · N · q i · z · − +
i=1 qi−1 qi qi−1
n
= ∑ δi · N · qi · (z · (1 − πi−1 ) + πi − πi−1 ) = V̂S
i=1

A similar consideration shows V̇D,I = V̂D,I . Since the present values of cash flows are identical
the same is true for ECL. Using the result of Proposition 3 leads us to the proof of this
proposition:
vi−1
Ni = N · (1 − πi−1 ) = N · = N i. 
qi−1

10

Electronic copy available at: https://ssrn.com/abstract=3238632

3.2. Modifying the Discount Rate

One requirement of the IFRS 9 rules is the use of an effective interest rate for discounting,
e.g. see B.5.5.44 in (IASB 2014). This means that the rate for discounting cash flows could
be different from the contractual interest rate of a loan. This requires some adjustments to the
ECL calculation formula of Section 1.
When the discount rate is different from the loan’s interest rate, it is no longer true that
the present value of contractual cash flows is equal to the loan’s outstanding balance. For this
present value, we find
n
ṼC = N · z · δ˜ i + N · δ˜ n
∑ (17)
i=1

where δ˜ is the discount factor using the effective instead of the contractual interest rate. For
VS and VD,I similar adjustments have to be made resulting in
n
ṼS = ∑ N · z · δ˜ i · qi + N · δ˜ n · qn (18)
i=1
n
ṼD,I = ∑ N · Ri · δ˜ i · (qi−1 − qi ) (19)
i=1

Lifetime expected loss is then computed as

ECL = ṼC − ṼS − ṼD,I . (20)

It is observed in practical applications that (2) is applied with an interest rate different
from the loan’s contractual interest rate. It was possible to find a link between present values
and weighted losses as in (2) where the derivation has shown that (2) needs modifications to
be consistent with present values when the discount rate and the contractual rate are equal.
When they are different it is no longer possible to come up with a weighted losses formula
consistent with present values. It was essential in deriving these formulas to have identical
contractual and discount rates. From this we can conclude that taking B5.5.29 and B5.5.44 in
IASB (2014) literally makes it impossible to use a weighted loss formula for computing ECL.
To be consistent with the text present value formulas have to be used instead.

3.3. Numerical Example

To illustrate the differences between the present values and the weighted loss formulas, a 20-
year fixed rate loan with a contractual rate of 5% is considered. The outstanding balance is
normalized to 100. For the risk parameters, assume for Year 1 a default probability PD = 1%
and compute pi iteratively as pi = (1 − pi−1 ) · PD + pi−1 . The term-structure of prepayment

11

Electronic copy available at: https://ssrn.com/abstract=3238632

rates is built in a similar way starting from a 1Y prepayment rate PR = 2%, and computing
π from πi = (1 − πi−1 ) · PR + πi−1 . For loss given default, a period-independent value of
40% is assumed. To illustrate the different ECL formulas we compute ECL with and without
prepayment, i.e. we set PR to 0% in the first test example and to 2% in the second. Besides
that, we use in one test set the contract rate for discounting while in the second test set we
discount with an assumed effective interest rate of 4.80% to outline the differences.
In total, seven different ECL numbers are computed. For completeness, ECL practice fol-
lowing (2) is reported. As the reference case, ECLwl,I which uses weighted losses under Case
I which was derived in (13) and ECL pv,I which is the corresponding ECL based on present val-
ues are used. Besides that, ECLwl,II computes the version of Proposition 3 with Case II for the
recovery payments which is presented together with its present values counterpart ECL pv,II .
Finally, ECLwl,I implements (16) from Proposition 4 together with its present values version
ECL pv,I . Note that conceptually it is not sensible to use all these formulas for the same set
of risk parameters since depending on the estimation of these parameters only one version is
consistent. The main purpose of this illustration is to measure the potential error of using the
wrong formula for ECL computation.
Scenario 1: PR = 0% and Discount Rate = Contract Rate
Scenario 2: PR = 2% and Discount Rate = Contract Rate
Scenario 3: PR = 0% and Discount Rate = Effective Rate
Scenario 4: PR = 2% and Discount Rate = Effective Rate
The results are displayed in Table 1 below.

ECL Scenario 1 Scenario 2 Scenario 3 Scenario 4

ECL practice 4.612 3.981 4.688 4.042
ECLwl,I 5.188 4.479 5.274 4.547
ECL pv,I 5.188 4.479 5.465 4.699
ECLwl,II 4.842 4.180 4.913 4.236
ECL pv,II 4.842 4.180 5.114 4.396
ECLwl,I 5.188 4.472 5.274 4.541
ECL pv,I 5.188 4.472 5.465 4.695
Table 1. Expected credit loss for a 20-year fixed rate loan with a contract rate of 5% under the
four risk parameter and discount rate scenarios.

From the first two scenarios it can be seen that there is no difference between the weighted
loss formula and the ECL based on present values since the weighted loss formulas are exact
when the discount rate and the contractual rate are identical. From ECLwl,II and ECL pv,II ,
respectively, we see that the LGD assumption has a non-negligible impact on ECL. If parts
of accrued interest are recovered ECL is lower when identical LGD numbers are used both in
ECLwl,I and ECLwl,II . Furthermore, prepayment effects if present in a credit portfolio have a

12

Electronic copy available at: https://ssrn.com/abstract=3238632

significant effect on ECL and cannot be ignored. However, if the inappropriate prepayment
formula is erroneously applied the effect seems to be immaterial as the values of ECL∗,I and
ECL∗,I are very similar. Formula (2) in all cases leads to the lowest numbers.
In Scenarios 3 and 4 the effect of changing the discount rate from the contractual rate to
an effective rate was illustrated. Since the derivation of the weighted loss formulas required in
many places that the discount rate equals the contractual rate differences should be expected.
We see in the examples that reducing the discount rate by 20 basis points results in ECL differ-
ences of about 5% between the weighted loss formulas and the present values framework for
the given set of risk parameters. This is considerable and one should think about abandoning
the weighted loss formulas when computing such a scenario.

4. Conclusion

In this article formulas for lifetime expected loss in the light of the IFRS 9 impairment rules
have been analyzed. Starting from a simple formula (2) that seems to be wide-spread in
practice, it was shown that this formula is inconsistent with differences in present values as
prescribed in IASB (2014), B5.5.29, but needs corrections depending on the way LGD is mod-
eled and estimated. An extension of the formula using weighted losses including prepayment
probabilities was derived. Finally, the case where discount rate and contractual rate are dif-
ferent was outlined and it was shown that only present values lead to a correct answer from a
conceptual viewpoint which was illustrated by a numerical example.
Overall, it was demonstrated that the application of weighted losses can be justified only
when discount and contract rates are equal. Although this approach is more intuitive since it
states the expected loss in every period and has a simple aggregation rule, it becomes problem-
atic when deviations from its core assumptions are introduced. In these cases it is unclear how
big the error compared to present values could become. A numerical example outlined that
the differences can be significant. In the view of the author, when developing IFRS 9 similar
steps should have been taken as in Basel Committee on Banking Supervision (2006) where for
the calculation of risk weights a simple formula based on a theoretically sound model was pre-
scribed and only the parameterization was left to the banks. This makes the implementation
easier and gives less room for discussions and interpretations.

References
Aguais, S. D., L. Forest, E. Wong, and D. Diaz-Ledezma, 2004, Point-in-Time Versus
Through-the-cycle ratings, in M. Ong, eds.: The Basel handbook: a guide for financial
practitioners (Risk Books, London ).

13

Electronic copy available at: https://ssrn.com/abstract=3238632

Basel Committee on Banking Supervision, 2006, International Convergence
of Capital Measurement and Capital Standards: A Revised Framework,
http://www.bis.org/publ/bcbsca.htm.
Carlehed, M., and A. Petrov, 2012, A methodology for point-in-time-through-the-cycle prob-
ability of default decomposition in risk classification systems, The Journal of Risk Model
Validation 6, 3–25.
Deloitte, 2017, IFRS 9: Probably Weighted and Biased?, https://www.business-
school.ed.ac.uk/crc-conference/wp-content/uploads/sites/46/2017/09/71-
Alexander Marianski.pdf.
IASB, 2014, IFRS Standard 9: Financial Instruments, .
PricewaterhouseCoopers, 2015, IFRS 9 - Credit Modelling and
Implementation, https://www.dico.com/design/Webinars En/DICO-
IFRS%209%20Modelling%20and%20Implementation.pdf.
Skoglund, J., 2017, Credit risk term-structures for lifetime impairment forecasting: A practical
guide, Journal of Risk Management in Financial Institutions 10, 177–195.
Xu, X., 2016, Estimating lifetime expected credit losses under IFRS 9, Working paper.

14

Electronic copy available at: https://ssrn.com/abstract=3238632

June 2024

Guideline on the management of

expected credit losses

1
Table of Contents

1. PRUDENTIAL EXPECTATIONS REGARDING CREDIT RISK AND ACCOUNTING FOR EXPECTED CREDIT
LOSSES UNDER THE INTERNAL RATINGS-BASED APPROACH .................................................................... 4
1.1 RESPONSIBILITIES OF THE BOARD OF DIRECTORS AND SENIOR MANAGEMENT ......................................................... 4
1.2 SOUND ECL METHODOLOGIES ..................................................................................................................... 5
1.2.1 Processes and systems ............................................................................................................. 5
1.2.2 Allowances................................................................................................................................ 5
1.2.3 ECL assessment and quantification methodology .................................................................... 5
1.2.4 Credit risk identification process .............................................................................................. 7
1.2.5 Allowance methodology ........................................................................................................... 9
1.3 CREDIT RISK RATING AND GROUPING ........................................................................................................... 10
1.3.1 Credit risk rating ..................................................................................................................... 10
1.3.2 Grouping based on shared credit risk characteristics............................................................. 10
1.4 ADEQUACY OF THE ALLOWANCE ................................................................................................................. 12
1.5 ECL MODEL VALIDATION........................................................................................................................... 13
1.6 EXPERIENCED CREDIT JUDGMENT ................................................................................................................ 14
1.7 COMMON DATA ...................................................................................................................................... 15
1.8 DISCLOSURE ........................................................................................................................................... 16
2. IMPAIRMENT OF LOAN EXPOSURES FOR STANDARDIZED FINANCIAL INSTITUTIONS ......................... 18
2.1 FORWARD-LOOKING INFORMATION............................................................................................................. 18
2.2 PAST-DUE INFORMATION .......................................................................................................................... 19

2
Introduction
This Guideline sets out the prudential expectations of the Autorité des marchés
financiers (the “AMF”) regarding sound and prudent management practices for credit
risk associated with expected credit losses (ECL).1
These best practices, presented in the form of expectations, are consistent with the
guidance on credit risk and accounting for ECL2 and core principles of credit risk
supervision 3 of the Basel Committee on Banking Supervision (the “BCBS”).
This guideline applies to financial services cooperatives, authorized trust companies,
savings companies and other authorized deposit institutions and to insurers that
engage in lending activities.
The first part of the guideline sets out the expectations for financial institutions that
use the Internal Ratings Based (IRB) approach (“Internal Ratings-Based Financial
Institutions”). The second part addresses the expectations applicable to insurers that
use the standardized approach (“Standardized Insurers”).
This guidance document does not relieve institutions from their obligations with
respect to current Canadian accounting standards.

1 The scope of this guideline is limited to those practices affecting the assessment and measurement of ECL
and allowances under current Canadian accounting standards. As used in this document, the term
“allowances” includes allowances on loans, and allowances or provisions on loan commitments and financial
guarantee contracts.
2 BANK FOR INTERNATIONAL SETTLEMENTS, Basel Committee on Banking Supervision, Guidance on

credit risk and accounting for expected credit losses, December 2015.
3 BANK FOR INTERNATIONAL SETTLEMENTS, Core Principles for Effective Banking Supervision,

September 2012. Core principles 17: Credit risk and 18: Problem assets, provisions and reserves.

3
1. Prudential expectations regarding credit risk and accounting for expected
credit losses under the internal ratings-based approach
1.1 Responsibilities of the board of directors and senior management4

The AMF expects senior management to implement sound and prudent credit risk
management practices and ensure that they are applied.

The AMF expects institutions to adopt and implement sound credit risk practices with
respect to identifying, assessing, quantifying, controlling, mitigating and adequately
monitoring credit risk consistent with its approved risk appetite and with sound
underwriting practices.
Senior management should develop and maintain appropriate processes, which
should be systematic and consistently applied. It should also establish and update a
strategy as well as policies and procedures to communicate the credit risk assessment
and quantification process to all relevant personnel.
The AMF believes that an effective internal control system for credit risk assessment
and quantification is essential to enable senior management to fulfill its
responsibilities. An effective internal control system should enable the institution to
consistently determine adequate allowances in accordance with its stated policies and
procedures and should include:
 measures to provide oversight of the integrity of information used and ensure
that the allowances reflected in the financial institution’s financial statements
and the reports provided to the AMF comply with current Canadian accounting
standards and the AMF’s expectations regarding the management of ECL;
 credit risk assessment and quantification processes that are independent from
the lending function. These processes should include, among other things:
o an effective credit risk rating system that is consistently applied, accurately
grades differing credit risk characteristics, identifies changes in credit risk
on a timely basis and prompts appropriate action;
o an effective process which ensures that all relevant and reasonable and
supportable information, including forward-looking information, is
appropriately considered in assessing and quantifying ECL;
o an assessment policy that ensures ECL quantification occurs not only at
the individual lending exposure level but also, when necessary, at the
collective portfolio level;5 6
o clear formal communication and coordination among credit risk staff,
financial reporting staff, the board, senior management and others who are
involved in the credit risk assessment and quantification process for an
ECL accounting framework.

4 The roles and responsibilities assigned to the board of directors and senior management are detailed in:
AUTORITÉ DES MARCHÉS FINANCIERS, Governance Guideline, April 2021.
5 By grouping exposures based on identified shared credit risk characteristics.

6 See sections 1.3 - Credit risk rating and grouping and 1.4 - Adequacy of the allowance.

4
1.2 Sound ECL methodologies

The AMF expects a financial institution to adopt, document and adhere to sound
methodologies that address policies, procedures and controls for assessing and
quantifying credit risk on all lending exposures. The measurement of allowances
should build upon those methodologies.

1.2.1 Processes and systems

The AMF expects a financial institution to have in place adequate processes and
systems to appropriately identify, assess, quantify, control, mitigate and monitor credit
risk. The financial institution should collect and analyze relevant information affecting
the assessment and quantification of ECL.
Credit risk assessment and quantification should provide relevant information for
senior management to make its experienced judgments about the credit risk of lending
exposures and the related estimation of ECL.
The AMF expects the financial institution to leverage processes that are used to
determine if, when and on what terms credit should be granted. The financial
institution is also expected to monitor credit risk at all stages of the loan’s life cycle
and quantify allowances for both accounting and capital adequacy purposes.
1.2.2 Allowances
A financial institution’s allowance methodologies should clearly document the
definitions of key terms related to the assessment and quantification of ECL.7 Where
different terms, information or assumptions are used across functional areas,8 the
underlying rationale for these differences should be documented. The rationale for
changes in assumptions that affect the quantification of ECL should also be
documented.
1.2.3 ECL assessment and quantification methodology
Robust and sound methodologies for assessing credit risk and quantifying levels of
allowances should:
 include a process enabling the institution to know the level, nature and drivers
of credit risk upon initial recognition of the lending exposure;
 include criteria to consider the impact of forward-looking information, including
macroeconomic factors. Whether the evaluation of credit risk is conducted on
a collective or individual basis, the methodology should demonstrate that this
consideration has occurred so that the recognition of ECL is not delayed. These
criteria should result in the identification of factors that affect repayment,
whether related to borrower incentives, willingness or ability to perform on the
contractual obligations, or to lending exposure terms and conditions;9
 include, for collectively evaluated exposures, a description of the basis for
creating portfolios of exposures with shared credit risk characteristics;

7 For example, loss and migration rates.

8 For example, accounting, capital adequacy and credit risk management.
9 Such as unemployment rates or occupancy rates. This may be at the international, national, regional or local

level.

5
  identify and document the ECL assessment and quantification methods10 to be
applied to each exposure or portfolio;
 document the reasons why the selected method is appropriate, especially if
different ECL quantification methods are applied to different portfolios and
types of individual exposures;11
 document the inputs, data and assumptions used in the allowance estimation
process,12 how the life of an exposure or portfolio is determined,13 the time
period over which historical loss experience is evaluated, and any adjustments
necessary for the estimation of ECL;
 document the methods used to validate models for impairment quantification;
 include a process for evaluating significant inputs and assumptions in the ECL
assessment and quantification method chosen. The AMF expects that the basis
for inputs and assumptions used in the estimation process will generally be
consistent from period to period. Where inputs and assumptions change, the
rationale should be documented;
 identify situations that would generally lead to appropriate changes in ECL
quantification methods, inputs or assumptions from period to period;14
 identify internal and external factors that may affect ECL estimates;15
 address how ECL estimates are determined.16 A financial institution should
have an unbiased view of the uncertainty and risks in its lending activities when
estimating ECL;
 identify what factors are considered when establishing appropriate historical
time periods over which to evaluate historical loss experience. A financial
institution should maintain sufficient historical loss data (ideally over at least
one full credit cycle) to provide a meaningful analysis of credit loss experience
for use as a starting point when estimating the level of allowances on a
collective or individual basis;
 determine the extent to which the value of collateral and other risk mitigants
affects ECL;
 outline the financial institution’s policies and procedures on write-offs and
recoveries;
 require that analyses, estimates or reviews that are inputs or outputs from the
credit risk assessment and quantification process are performed by competent
and well-trained personnel and validated by personnel who are independent of

10 Such as a loss rate method, probability of default method or another method.

11 An institution should be able to explain the rationale for any changes in measurement approach (e.g., a
move from a loss rate method to a PD/LGD method) and the quantitative impacts of such changes.
12 Such as historical loss rates, PD/LGD estimates and economic forecasts.

13 Including how expected prepayments have been considered.

14 For example, the institution may state that a loan that had been previously evaluated on a collective basis

using a PD/LGD method may be removed and evaluated individually using the discounted cash flow method
upon receipt of new, borrower-specific information such as the loss of employment.
15 Such as underwriting standards applied to a lending exposure at origination and changes in industry,

geographical, economic and political factors.

16 Such as historical loss rates or migration analysis as a starting point, adjusted for information on current

conditions, forward-looking information and macroeconomic factors.

6
 the institution’s lending activities. These inputs and outputs should be properly
recorded and well documented and the documentation should include clear
explanations supporting the analyses, estimates and reviews;
 ensure that ECL estimates appropriately incorporate forward-looking
information, including macroeconomic factors, that has not already been
factored into allowances measured on an individual exposure basis. A financial
institution should use its experienced credit judgment, in particular to consider
broad trends in the entire lending portfolio and changes in the financial
institution’s business model; and
 require a process to assess the overall adequacy of allowances.

1.2.4 Credit risk identification process

A financial institution’s credit risk identification process should ensure that factors that
impact changes in credit risk and estimates of ECL are properly identified on a regular
basis. Also, consideration of credit risk inherent in new products and activities should
be a key part of the risk identification process and the assessment and quantification
of ECL.
With respect to factors related to the type of borrower, the borrower’s capacity and
capital, the term of the loan and the value of assets pledged as collateral together
with other credit risk mitigants that may affect the full collectability of cash flows, the
financial institution could, depending on the type of exposure, consider:
 monitoring of its lending policies and procedures, including underwriting
standards and lending terms;
 a borrower’s sources of recurring income available to meet the scheduled
payments;
 a borrower’s ability to generate a sufficient cash flow stream over the term of
the commitment;
 the borrower’s overall leverage level and expectations of changes to leverage;
 unencumbered assets the borrower may pledge as collateral in the market or
bilaterally in order to raise funds and expectations of changes to the value of
those assets;
 potential one-off events and recurring behaviour that may affect the borrower’s
ability to meet contractual obligations; and
 timely evaluations of collateral value 17 and consideration of factors that may
impact the future value of collateral.
Where they have the potential to affect the financial institution’s ability to recover
amounts due, factors relating to the financial institution’s business model and
macroeconomic conditions should be considered. These factors include:
 competition and legal and regulatory requirements;
 trends in the financial institution’s overall volume of credit;

17 Bearing in mind that collateral values directly affect estimates of loss-given-default.

7
  the overall credit risk profile of the financial institution’s lending exposures and
expectations of changes thereto;
 credit concentrations to borrowers or by product type, segment or
geographical market;
 expectations on collection, charge-off and recovery practices;
 the quality of the financial institution’s credit risk review system; and
 other factors that may impact ECL such as expectations of changes in
unemployment rates, gross domestic product, benchmark interest rates,
inflation, liquidity conditions or technology.
The AMF expects that methodology will consider different potential scenarios and will
not rely purely on subjective, biased or overly optimistic considerations. A financial
institution should develop and document its process to generate relevant scenarios to
be used in the estimation of ECL. In particular:
 the financial institution should demonstrate and document how ECL estimates
would alter with changes in scenarios, including changes to relevant external
conditions that may impact ECL estimates or components of the ECL
calculation (such as probability of default and loss-given-default parameters);
 the financial institution should have a documented process for determining the
time horizon of the scenarios and, if relevant, how ECL is estimated for
exposures whose lives exceed the period covered by the economic forecast(s)
used;
 scenarios may be internally developed or vendor-defined:
o for internally developed scenarios, a financial institution should
have a variety of experts, such as risk management professionals,
economists, operational managers and senior management, assist
in the selection of scenarios that are relevant to the financial
institution’s credit risk exposure profile;
o for vendor-defined scenarios, a financial institution should ensure
that the vendor tailors the scenarios to reflect the financial
institution’s business and credit risk exposure profile;
 backtesting should be performed to verify that the most relevant economic
factors that affect collectability and credit risk are being considered and
incorporated into ECL estimates; and
 where market indicators of future performance (such as credit default swap
spreads) are available, senior management may consider them to be a valid
benchmark against which to check the consistency of its own judgments.
The AMF expects a financial institution to consider all information that is relevant to
the product, borrower, business model or economic and regulatory environment when
developing ECL estimates. It should consider the experience and information from
similar exercises.
Forward-looking information, including economic forecasts and related credit risk
factors used for ECL estimates, should be consistent with inputs to other relevant
estimates within the financial statements, budgets, strategic and capital plans, and
other information used in managing and reporting on the financial institution.

8
 1.2.5 Allowance methodology
A financial institution’s allowance methodology should build upon the accounting
framework for ECL assessment and quantification.
It should include the following criteria:
 Restructurings/modifications can take many forms, including a renewal or
extension of terms, other concessions to the borrower, or a modification of the
terms with or without concessions to the borrower;
 It should deliver a robust assessment and quantification of ECL such that the
allowance level continues to reflect the collectability of the substance of the
restructured exposure;
 It should also call upon lending staff to promptly notify the financial institution’s
accounting function when exposures are restructured or modified to ensure
appropriate accounting for the change. For more complex restructurings and
modifications, regular communication between this line of business and the
accounting function is warranted.
 The methodology should enable appropriate identification and accounting for
purchased or originated credit-impaired lending. The cash flow estimates for
these lending exposures should be reviewed each reporting period and
updated as necessary. Such updates should be properly supported and
documented.

9
1.3 Credit risk rating and grouping

The AMF expects a financial institution to have a process in place to appropriately

group lending exposures on the basis of shared credit risk characteristics.

1.3.1 Credit risk rating

The AMF believes that an effective credit risk rating process should capture the
varying level, nature and components of credit risk at every stage of the loan’s life
cycle. An effective rating system will ensure that all lending exposures are properly
monitored and that ECL allowances are accurately measured.
The credit risk rating process should include an independent review function. While
front-line lending staff may have initial responsibility for assigning credit risk grades
and ongoing responsibility for updating the credit grade to which an exposure is
assigned, the AMF expects this to be subject to the review of an independent review
function.
The credit risk grade a financial institution assigns upon initial recognition of a lending
exposure may be based on a number of criteria, including product type, terms and
conditions, collateral type and amount, borrower characteristics and geography or a
combination thereof. Existing credit risk grades assigned may subsequently change
on either a portfolio or an individual basis.18
The credit risk rating system should capture all lending exposures to allow for an
appropriate differentiation of credit risk and grouping of lending exposures within the
credit risk rating system, reflect the risk of individual exposures and, when aggregated
across all exposures, the level of credit risk in the portfolio as a whole. In this context,
an effective credit risk rating system will allow a financial institution to identify both
migration of credit risk and changes in credit risk grade.
In describing the elements of its credit risk rating system, a financial institution should
clearly define each credit risk grade and designate the staff responsible for the design,
implementation, operation and performance of the system as well as those
responsible for periodic testing and validation.
Credit grades should be reviewed whenever relevant new information is received or
a financial institution’s expectation of credit risk has changed. Credit risk grades
assigned should receive a periodic formal review (at least annually) to reasonably
ensure that those grades are appropriate and up to date. Credit risk grades for
individually assessed lending exposures that are higher-risk or credit-impaired should
be reviewed more frequently than annually. ECL estimates should be updated on a
timely basis to reflect changes in credit risk grades for either groups of exposures or
individual exposures.
1.3.2 Grouping based on shared credit risk characteristics
Groups should be sufficiently granular to allow banks to group exposures into
portfolios with shared credit risk characteristics so that the institution can reasonably

18 Such as changes in industry outlook, business growth rates, consumer sentiment and changes in economic
forecasts (interest rates, unemployment rates, commodity prices, etc.).

10
assess changes in credit risk and thus the impact on the estimate of ECL. A financial
institution’s method for grouping exposures to assess credit risk19 should be
documented and subject to review and appropriate approval.
Lending exposures should be grouped according to shared credit risk characteristics
so that changes in credit risk respond to the impact of changes in the current
environment, forward-looking information and macroeconomic factors. The basis of
grouping should be reviewed to ensure that exposures within the group remain
homogeneous in terms of their response to credit risk drivers.20
Exposures must not be grouped in such a way that an increase in the credit risk of
particular exposures is masked by the performance of the group as a whole.
A financial institution should have in place a robust process to ensure appropriate
initial grouping of their lending exposures. If relevant new information is received or a
financial institution’s changed expectations of credit risk suggest that a permanent
adjustment is warranted, the AMF expects the financial institution to re-evaluate and
re-segment the grouping of exposures.

19 Such as by product type, industry/market segment or geographical location.

20 Grouping implemented initially based on shared credit risk characteristics will not necessarily be appropriate
subsequently, given that the relevant characteristics and their impact on the level of credit risk for the group
may change over time.

11
1.4 Adequacy of the allowance

The AMF expects a financial institution’s aggregate amount of allowances,

regardless of whether allowance components are determined on a collective or an
individual basis, to be adequate and consistent with current Canadian accounting
standards.

A financial institution should implement sound and robust credit risk methodologies
with the objective that the overall balance of the allowance for ECL is developed in
accordance with current Canadian accounting standards and adequately reflects ECL
within that framework.
A robust assessment of allowances takes into account relevant factors and
expectations at the reporting date that may affect the collectability of remaining cash
flows over the life of a group of lending exposures or a single lending exposure.
Depending on its ability to incorporate forward-looking information into the ECL
estimate, a financial institution may use individual or collective assessment
approaches. The ECL estimation technique used should be the most appropriate in
the particular circumstances, and typically should be aligned with how the institution
manages its credit risk exposure.21
The use of individual versus collective estimation techniques should not result in
materially different allowance measurements. Regardless of the estimation technique
used, a financial institution should ensure this does not result in delayed recognition
of ECL.
When a financial institution does use the individual estimation technique, the ECL
estimate should always incorporate the expected impact of all reasonable and
supportable forward-looking information, including macroeconomic factors, that
affects collectability and credit risk. When an individual estimation technique is used,
the financial institution’s documentation should clearly demonstrate how forward-
looking information, including macroeconomic factors, has been reflected in the
individual ECL assessment.
In instances where a financial institution’s individual assessments of exposures do
not adequately consider forward-looking information, it would be appropriate to group
lending exposures with shared credit risk characteristics to estimate the impact of
forward-looking information, including macroeconomic factors.
This process allows identification of relationships between forward-looking
information and ECL estimates that may not be apparent at the individual exposure
level.

21 For example, collective assessment is often used for large groups of homogeneous lending exposures with
shared credit risk characteristics, such as retail portfolios. Individual ECL assessments are often conducted
for significant exposures, or where credit risks have been identified at the loan level, such as watch list and
past-due loans.

12
1.5 ECL model validation

The AMF expects a financial institution to have policies and procedures in place to
appropriately validate the models it uses to assess and quantify ECL.

ECL assessment and quantification may involve models and assumption-based

estimates for risk identification and quantification.
Models may be used in various aspects of the ECL assessment and quantification
process at both the individual transaction and overall portfolio levels. They may also
be used in credit grading, credit risk identification, quantification of ECL allowances
for accounting purposes, stress testing and capital allocation.
ECL assessment and quantification models should consider the impact of changes to
borrower and credit risk-related variables, such as changes in:
 PDs;
 LGDs;
 Exposure amounts;
 Collateral values;
 Migration of default probabilities;
 Internal borrower credit risk grades based on historical, current and forward-
looking information; and
 Macroeconomic factors.

13
1.6 Experienced credit judgment

The AMF expects a financial institution to use its experienced credit judgment,
especially in the consideration of reasonable forward-looking information, in
assessing and quantifying ECL.

The institution should have the necessary tools to ensure a robust estimate and timely
recognition of ECL.
Information on historical loss experience or the impact of current conditions may not
fully reflect the credit risk in lending exposures.
In that context, a financial institution should use its experienced credit judgment to
incorporate the expected impact of all reasonable and supportable forward-looking
information, including macroeconomic factors, on its estimate of ECL. A financial
institution’s use of its experienced credit judgment should be documented in its credit
risk policy and subject to appropriate oversight.
Estimates of the ECL amount should reflect the financial institution’s experienced
credit judgment and consider a wide range of possible outcomes.
To assess whether a loan should move to a lifetime expected credit loss (LEL), the
change in the risk of a default occurring over the expected life of the loan must be
considered.
Historical information provides a useful basis for the identification of trends and
correlations needed to identify the credit risk drivers for lending exposures. However,
ECL estimates should not ignore the impact of forward-looking events and conditions
on those drivers. The estimate should reflect the expected future cash shortfalls
resulting from such impact.
Macroeconomic forecasts and other relevant information should be applied
consistently across portfolios. When developing ECL estimates, a financial institution
should apply its experienced credit judgment to consider its point in the credit cycle.
This assessment may vary depending on the geographical data.
Additionally, the AMF believes that financial institutions are increasingly considering
a wide range of information, including that of a forward-looking nature, for risk
management and capital adequacy purposes. The AMF expects financial institutions
to consider information derived from the different stages in the credit risk management
process in developing their ECL estimates.

14
1.7 Common data

The AMF expects a financial institution to have a sound credit risk assessment
and quantification process that provides it with a common basis for systems, tools
and data.

There is commonality in the systems, tools and data used to assess credit risk,
quantify ECL for accounting purposes and determine expected losses for capital
adequacy purposes.
The use of common processes, systems, tools and data strengthens, to the maximum
extent possible, the consistency of the resulting estimates and minimizes
disincentives to following sound credit risk practices for all purposes.
A financial institution’s credit risk practices should meet fundamental requirements
and procedures, including having the appropriate tools to identify and assess credit
risk. These fundamental requirements are equally necessary for assessing credit risk
and fairly representing the financial institution’s financial position for both accounting
and capital adequacy purposes. These common processes are closely interrelated,
which strengthens the reliability and consistency of resulting ECL estimates. These
processes also increase transparency and provide incentives to follow sound credit
risk practices.
A financial institution’s credit risk monitoring system should be designed to include all
lending exposures when assessing the impact of changes in credit risk. The system
should include not only lending exposures that may have experienced significant
increases in credit risk, have incurred losses or are otherwise credit-impaired.
A financial institution should periodically review its credit risk practices to ensure that
relevant data available throughout a financial institution are captured and that systems
are updated as the institution’s underwriting or business practices change or evolve
over time.
The AMF expects feedback processes to be established to ensure that information on
estimates of ECL and changes in the credit risk is shared among credit risk experts,
accounting and regulatory reporting staff and, in particular, with loan underwriting
personnel.
Common processes, systems, tools and data include credit risk rating systems,
estimated PDs (subject to appropriate adjustments), past-due status, loan-to-value
ratios, historical loss rates, product type, amortization schedule, down payment
requirements, market segment, geographical location, vintage, and collateral type.

15
1.8 Disclosure

The AMF expects a financial institution’s public disclosures to promote

transparency and comparability by providing timely, relevant and decision-useful
information.

The objective of public disclosures is to provide decision-useful information, on a

financial institution’s financial position, performance and changes therein, to a wide
range of users in a clear and understandable manner. The AMF expects financial
institutions to continue to improve their disclosure with the aim of providing
information that is relevant and comparable information so that interested parties can
make timely, informed decisions.
Financial and credit risk management disclosures should be made in accordance with
current Canadian accounting standards. Accordingly, it is important that financial
institutions consider the disclosures needed to fairly depict an institution’s exposure
to credit risk, including its ECL estimates, and to provide relevant information on an
institution’s underwriting practices.
Senior management should apply judgment to determine the appropriate level of
aggregation and disaggregation of data disclosed, such that the financial institution’s
disclosures continue to meet current accounting and regulatory requirements. Senior
management should also provide insights into an entity’s exposure to credit risk for
interested parties to perform relevant peer group comparisons.
The quantitative and qualitative disclosures should clearly communicate to interested
parties the main assumptions/inputs used to develop ECL estimates.
Additionally, the AMF expects disclosures to highlight policies and definitions that are
integral to the estimation of ECL,22 factors that cause changes in ECL estimates, and
the manner in which management’s experienced credit judgment has been
incorporated. Disclosure of significant policies should be decision-useful and should
describe, in the specific context of the financial institution, how those policies have
been implemented.
The move to an ECL model requires that forward-looking information, including
macroeconomic factors, be incorporated into ECL estimates, in accordance with the
existing accounting framework. The AMF expects the financial institution to provide
qualitative disclosures on how this information has been incorporated into the
estimation process, particularly when the assessment is carried out on an individual
basis.
A financial institution’s decisions regarding the basis for grouping lending exposures
will normally reflect a combination of factors. The AMF expects disclosures in this
area to communicate how senior management satisfies itself that lending exposures
are appropriately grouped, such that these groups continue to share credit risk
characteristics.

22 Such as an institution’s basis for grouping lending exposures into portfolios with shared credit risk
characteristics and its definition of default, guided by the definition used for regulatory purposes - See
AUTORITÉ DES MARCHÉS FINANCIERS, Capital Adequacy Guideline, February 2024 (in French only).

16
To improve the quality and meaningfulness of information disclosed for ECL
estimates, the AMF expects financial institutions to provide an explanation of
significant changes to the estimation of ECL on a regular basis. This information
should include both relevant qualitative and quantitative disclosures. It should also
enhance the understanding of how ECL estimates have changed.
The AMF expects senior management to regularly review its disclosure policies to
ensure that the information disclosed continues to be relevant to the financial
institution’s risk profile, product concentrations, industry norms and current market
conditions. A financial institution should aim to provide disclosures that facilitate
comparisons with its peers. Such disclosures will enable interested parties to monitor
changes in the financial institution’s ECL estimates from period to period and will allow
users to perform meaningful analyses across peer groups.

17
2. Impairment of loan exposures for Standardized Financial Institutions23
2.1 Forward-looking information

The AMF expects a Standardized Financial Institution to incorporate forward-

looking information into its ECL assessment and quantification process.

A Standardized Financial Institution should use its experienced credit judgment to

incorporate the expected impact of reasonable and supportable forward-looking
information, including macroeconomic factors, on its estimate of ECL. A Standardized
Financial Institution’s use of its experienced credit judgment is integral to its credit
risk methodology and should be documented and subject to appropriate oversight.
A Standardized Financial Institution may incorporate forward-looking information in a
variety of ways, such as by using individual and/or collective estimates. This could
also be done through modelled approaches or the use of temporary adjustments.
Additionally, Standardized Financial Institutions are considering a wide range of
information, including forward-looking information, for risk management and stress
testing purposes. The AMF expects Standardized Financial Institutions to consider
reasonable and supportable information derived from the different stages in the credit
risk management process when developing their ECL estimates, such as information
and assumptions relevant to ECL used in stress testing, planning, etc.

23 Standardized Insurers should, in particular, apply this section in addition to the Guideline on Capital
Adequacy Requirements (Credit risk – balance sheet items) – AUTORITÉ DES MARCHÉS FINANCIERS,
Guideline on Capital Adequacy Requirements, January 2024.

18
2.2 Past-due information

Accordingly, the AMF expects a Standardized Financial Institution to limit its use of
the more-than-30-days-past-due rebuttable presumption as a primary indicator of
transfer to ECL quantification for the life of the loan.

The AMF expects that any assertion that the more-than-30-days-past-due

presumption is rebutted on the basis that there has not been a significant increase in
credit risk will be accompanied by a thorough analysis clearly evidencing that 30 days
past due is not correlated with a significant increase in credit risk.
Such analysis should consider both current and reasonable and supportable forward-
looking information that may cause future cash shortfalls to differ from historical
experience.
In the limited instances where past-due information is the best criterion available to a
Standardized Financial Institution to determine when exposures should move to the
lifetime ECL measurement category, Standardized Financial Institutions should pay
particular attention to their measurement of 12-month ECL to ensure that ECL are
appropriately captured in accordance with current Canadian accounting standards.
Moreover, Standardized Financial Institutions should recognize that significant
reliance on backward-looking information could introduce bias into the implementation
of an ECL framework, risking that the objectives of the impairment requirements under
the current Canadian accounting standards are not met.
Standardized Financial Institutions should notify the AMF of any material changes to
their ECL methodology and/or ECL level.
The AMF expects Standardized Financial Institutions to establish and maintain a
materiality definition with respect to modifications to its methodology for establishing
ECL allowances and the level of ECL. In arriving at a suitable assessment of
materiality, the Standardized Financial Institution should consider a combination of
factors, including impact to systems, data, and processes.

19