SOFR Derivative Pricing Using a Short Rate Model

Mingyang Xu

Abstract

In 2017 the Alternative Reference Rate Committee (ARRC) recommended the Secured Overnight
Financing Rate (SOFR) as the replacement for USD LIBOR as the reference rate for use in derivatives and
financial contracts. Since then SOFR-linked derivatives started to develop and their liquidities have also
improved gradually. Compared to LIBOR derivatives, pricing SOFR derivatives poses additional
challenges, for example, complexity due to rate averaging and convexity adjustment due to timing
mismatch. In this paper we propose Gaussian affine term structure models as a natural choice to
address these challenges. Because of the advantages of affine models and Gaussian models we derive
analytic pricing formula for forward SOFR term rates as well as future SOFR term rates, which can then
be used to price most linear SOFR products. Furthermore, since the short rate is normally distributed,
closed-form solutions based on Black-Scholes or Bachelier formula are also derived for European SOFR
options, including SOFR swaptions, SOFR cap/floor and options on SOFR futures. Although American
options cannot be priced analytically, efficient numerical methods are developed by combining lattice
and analytic formula.

Key words:

SOFR, Averaging, Affine, Hull-White, Swaption, Option, Futures, Callable Floater, Convexity Adjustment

1. Introduction

The Secured Overnight Financing Rate (SOFR) is the cost of borrowing overnight collateralized by
Treasury securities, i.e. overnight Treasury repo rate. The SOFR is published by the New York Federal
Reserve Bank publishes the SOFR on a daily basis. Unlike Libor, SOFR is an actual transaction rate. It is
calculated as the volume-weighted median rate of all eligible transactions. SOFR typically trades close to
overnight LIBOR, but with higher volatility due to its methodology and wide dispersion in bilateral repo
data.

As the SOFR is expected to replace Libor as the index of choice sometimes in 2021, the SOFR linked
derivative market started to develop in the past few years, and its liquidity greatly improved.

SOFR derivatives tend to use an averaged overnight rate over a given period, which is often either
simple interest (Arithmetic Average) or compound interest (Geometric Average) (ARCC, 2021)

𝑑𝑏 𝑟𝑖 × 𝑛𝑖 𝑁
𝑅𝐺 = [∏ (1 + ) − 1] ×
𝑖=1 𝑁 𝑑𝑐

1
 𝑑𝑏 𝑟𝑖 × 𝑛𝑖 𝑁
𝑅𝐴 = [∑ ]×
𝑖=1 𝑁 𝑑𝑐

where

db = the number of business days in the interest rate period

dc = the number of calendar days in the interest rate period

ri = the interest rate on the ith business day

ni = the number of calendar days for which ri is applied

N = the number of calendar days in the year

If the SOFR rate is a daily compounded interest rate, compounding is equivalent to rolling forward a
money market account, and thus

𝑑𝑏 𝑟𝑖 × 𝑛𝑖 𝑁 1 1
𝑅𝐺 = [∏ (1 + ) − 1] × =[ − 1] ×
𝑖=1 𝑁 𝑑𝑐 𝐷(𝑡, 𝑇) 𝑇−𝑡

with D(t,T) defined as

𝑇
1
𝐷(𝑡, 𝑇) = = exp⁡(− ∫ 𝑟(𝑢)𝑑𝑢 )
𝑀(𝑡, 𝑇) 𝑡

which is similar to the calculation of simple Libor rate.

Similarly, the simple average of SOFR rate is also approximately by its continuous version

1
𝑑𝑏 𝑟𝑖 × 𝑛𝑖 𝑁 𝑇
1 ln⁡
𝐷(𝑡, 𝑇)
𝑅𝐴 = [∑ ]× ≈ (∫ 𝑟𝑢 𝑑𝑢) =
𝑖=1 𝑁 𝑑𝑐 𝑡 𝑇−𝑡 𝑇−𝑡

which is similar to continuous compounded forward rate.

Although compound interest is the more economically correct convention, simple interest is easier to
calculate in practice and most existing system can accommodate it. On the other hand, simple interest
makes pricing much more complicated because of the extra convexity adjustment that requires a
stochastic interest rate model.

Moreover, the fixing of the floating rate of a SOFR derivative can be either in-advance or in-arrears. An
in-advance structure would reference an average of SOFR observed before the beginning of the current
interest period, while an in-arrears structure would reference an average of SOFR over the current
interest period. An average of SOFR in-arrears will reflect what actually happens over the period, but
loan borrowers usually prefer to know their payments ahead of time, i.e. in-advance fixing.

2
In the case of average SOFR fixing, extra time is also necessary in order to wait for the daily SOFR to be
published and the average SOFR rate to be calculated and posted (c.g. ARCC, 2021). This results in
additional convexity adjustment because of timing lags.

Thus, all of these, including averaging and timing lags, add to the complexity of pricing SOFR derivatives.
In order to deal with them a short rate model seems a natural choice. In the following we will discuss
pricing methods for various SOFR derivatives using a short rate model. The rest of the paper is organized
as follows: section 2 defines forward SOFR term rates and describes how to build a SOFR curve using
SOFR swaps; section 3 applies a Gaussian mean-reverting model to model the dynamics of continuous
SOFR rate and derives forward and futures SOFR term rates; section 4 derives analytic pricing formulas
for various SOFR derivatives based on the SOFR curve and SOFR short rate model; and section 5
concludes the paper.

2. Forward SOFR Term Rate and SOFR Curve

2.1 Forward SOFR Term Rate

While the overnight treasury repo market underlying SOFR is extraordinarily deep, term repo markets
are much thinner, and thus it would not be possible to build a robust forward-looking term rate off the
term Treasury repo market.

Suppose a future floating payment at Ti is linked to the compound SOFR rate, i.e.

𝑃𝑎𝑦𝑚𝑒𝑛𝑡 = 𝑁 ∙ 𝛿 ∙ 𝑅𝐺 (𝑇𝑖 )

With unit Its present value at t is given by

𝐷(𝑡, 𝑇𝑖−1 )
𝑉(𝑡) = 𝐸[𝛿 ∙ 𝑅𝐺 (𝑇𝑖 )𝐷(𝑡, 𝑇𝑖 )] = 𝐸 [(⁡ − 1) 𝐷(𝑡, 𝑇𝑖 )] = 𝑃(𝑡, 𝑇𝑖−1 ) − 𝑃(𝑡, 𝑇𝑖 )
𝐷(𝑡, 𝑇𝑖 )

where in term of money market account M(T) or discounting factor D(t, T) the SOFR term rate at Ti

1 𝑀(𝑇𝑖 ) 1 𝐷(𝑡, 𝑇𝑖−1 )

𝑅𝐺 (𝑇𝑖 ) = ( − 1) = ( − 1)
𝛿 𝑀(𝑇𝑖−1 ) 𝛿 𝐷(𝑡, 𝑇𝑖 )

Meanwhile, we have under the Ti-forward measure

𝑉(𝑡) = 𝑃(𝑡, 𝑇𝑖 )𝛿𝐸 𝑇𝑖 [𝑅𝐺 ]

and thus we have forward SOFR rate

1 𝑉(𝑡) 1 𝑃(𝑡, 𝑇𝑖−1 )

𝑓𝑅𝐺 (𝑡) = 𝐸 𝑇𝑖 [𝑅𝐺 ] = = ( − 1)
𝛿 𝑃(𝑡, 𝑇𝑖 ) 𝛿 𝑃(𝑡, 𝑇𝑖 )

which can be calculated from a discount curve {P(t, T)}.

3
Similarly, we can define the forward-looking SOFR term rate for the arithmetic average of SOFR rate as
the expected rate under the Ti-forward measure
𝑇𝑖
1
𝑓𝑅𝐴 (𝑡) = 𝐸 𝑇𝑖 [𝑅𝐴 ] = ∫ 𝐸 𝑇𝑖 [𝑟𝑢 ]𝑑𝑢
𝑇𝑖 − 𝑇𝑖−1 𝑇𝑖−1

which cannot be calculated directly from a discount curve without an interest rate model.

Thus, in order to calculate forward SOFR term rates at least a SOFR forward curve is required.

2.2 Build Forward SOFR Curve

Similar to any other yield curves, a SOFR curve can be built by matching the market quotes of SOFR-
linked instruments. However, during the LIBOR era and after the LIBOR era, the choice of instrument is
different.

Before the LIBOR completely retires, the SOFR market is not deep enough and a SOFR curve can be built
using SOFR futures and SOFR-LIBOR basis swaps.

After the LIBOR retires, a SOFR curve can be built using 3M SOFR futures (short-end) up to 20 months
and SOFR swaps (long-end) from 2 years on.

Since a SOFR forward rate is related to the corresponding SOFR futures rate

1 𝑃(𝑡, 𝑇0 )
𝑓𝑅 (𝑡) = 𝐸 𝑇 [𝑅] = ( − 1) = 𝐸[𝑅] + 𝐶𝑜𝑛𝑣𝑒𝑥𝑖𝑡𝑦⁡𝐴𝑑𝑗𝑢𝑠𝑡𝑚𝑒𝑛𝑡
𝛿 𝑃(𝑡, 𝑇)
= 𝐹𝑅 (𝑡) + 𝐶𝑜𝑛𝑣𝑒𝑥𝑖𝑡𝑦⁡𝐴𝑑𝑗𝑢𝑠𝑡𝑚𝑒𝑛𝑡

SOFR discount factors can be calibrated using SOFR futures rates together with convexity adjustment.

SOFR fixed vs floating swaps are valued in a similar way to LIBOR vanilla swaps and thus can be used to
calibrate long term rates. The valuation of SOFR swaps will be discussed in detail later.

3. SOFR Short Rate Model

Corresponding to SOFR rate, instantaneous short rate models, including both spot rate and forward rate,
seem to be a natural choice for modeling interest rate dynamics.

Assume the short rate follows a mean-reverting Gaussian process, i.e. Hull-White model, (c.g. Brigo and
Mercurio, 2001)

𝑟(𝑡) = 𝜑(𝑡) + 𝑥(𝑡)

𝑑𝑥𝑡 = −𝑘𝑥𝑡 𝑑𝑡 + 𝜎(𝑡)𝑑𝑊

4
whose model parameters can be first calibrated to swaption volatilities. Moreover, the drift parameter,
(t), is tuned to match the SOFR curve term structure before it is used to price SOFR instruments, i.e.
𝑇 𝑇
𝑒𝑥𝑝 (− ∫ 𝜑(𝑡)𝑑𝑡) 𝐸 [𝑒𝑥𝑝 (− ∫ 𝑥(𝑡)𝑑𝑡)] = 𝑃(𝑡, 𝑇)
0 0

One can show that under the Ti-forward measure the dynamics of the state variable x(t) follow as SDE

𝑟,𝑇𝑖
𝑑𝑥(𝑡) = (−𝐵(𝑡, 𝑇𝑖 )𝜎(𝑡)2 − 𝑘𝑥(𝑡))𝑑𝑡 + 𝜎(𝑡)𝑑𝑊𝑡

whose strong solution is given by

𝜏
𝑟,𝑇𝑖
𝑥(𝜏) = 𝐸 𝑇𝑖 [𝑥(𝜏)|ℱ𝑡 ] + exp⁡(−𝑘𝜏) ∫ exp⁡(𝑘𝑢)𝜎(𝑢)𝑑𝑊𝑢
𝑡

with
𝜏 𝜏 𝜏 𝑢
𝐸 𝑇𝑖 [𝑥(𝜏)|ℱ𝑡 ] = 𝑒𝑥𝑝 (− ∫ 𝜅(𝑢)𝑑𝑢) 𝑥(𝑡) − 𝑒𝑥𝑝 (− ∫ 𝜅(𝑢)𝑑𝑢) ∫ 𝑒𝑥𝑝 (∫ 𝜅(𝑣)𝑑𝑣 ) 𝐵(𝑢, 𝑇𝑖 )𝜎(𝑢)2 𝑑𝑢
𝑡 𝑡 𝑡 𝑡

where
𝑡 𝑇 𝑢 𝑇 𝑢
𝐵(𝑡, 𝑇) = exp (∫ 𝑘(𝑢)𝑑𝑢) ∫ 𝑒𝑥𝑝 (− ∫ 𝑘(𝑠)𝑑𝑠) 𝑑𝑢 = ∫ 𝑒𝑥𝑝 (− ∫ 𝑘(𝑠)𝑑𝑠) 𝑑𝑢
0 𝑡 0 𝑡 𝑡

Thus, the SOFR forward term rate for both types of averaging are given by

1 𝑃(𝑡, 𝑇𝑖−1 )
𝑓𝑅𝐺 (𝑡) = 𝐸 𝑇𝑖 [𝑅𝐺 ] = ( − 1)
𝛿 𝑃(𝑡, 𝑇𝑖 )
𝑇
𝑒𝑥𝑝 (∫𝑇 𝑖 𝜑(𝑢)𝑑𝑢 + 𝐴(𝑡, 𝑇𝑖−1 ) − 𝐴(𝑡, 𝑇𝑖 ) − (𝐵(𝑡, 𝑇𝑖−1 ) − 𝐵(𝑡, 𝑇𝑖 ))𝑥𝑡 ) − 1
𝑖−1
=
𝑇𝑖 − 𝑇𝑖−1

and
𝑇𝑖 𝑇𝑖 𝑇𝑖
1 1
𝑓𝑅𝐴 (𝑡) = 𝐸 𝑇𝑖 [𝑅𝐴 ] = 𝑇𝑖
∫ 𝐸 [𝑟𝑢 ]𝑑𝑢 = (∫ 𝜑(𝑢)𝑑𝑢 + ∫ 𝐸 𝑇𝑖 [𝑥(𝑢)|ℱ𝑡 ]𝑑𝑢)
𝑇𝑖 − 𝑇𝑖−1 𝑇𝑖−1 𝑇𝑖 − 𝑇𝑖−1 𝑇𝑖−1 𝑇𝑖−1

respectively.

In a special case where k is constant, we have

5
 𝑇𝑖
∫ 𝐸 𝑇𝑖 [𝑥(𝜏)|ℱ𝑡 ]𝑑𝜏
𝑇𝑖−1
𝑒𝑥𝑝(−𝑘(𝑇𝑖−1 − 𝑡)) − 𝑒𝑥𝑝(−𝑘(𝑇𝑖 − 𝑡))
= 𝑥(𝑡)
𝑘
𝑇𝑖 𝜏
exp⁡(𝑘𝑢) − 𝑒𝑥𝑝(−𝑘(𝑇𝑖 − 2𝑢))
− ∫ 𝑒𝑥𝑝(−𝑘𝜏)𝑑𝜏 ∫ 𝜎(𝑢)2 𝑑𝑢
𝑇𝑖−1 𝑡 𝑘

where

1 − exp⁡(−𝑘(𝑇 − 𝑡))
𝐵(𝑡, 𝑇) =
𝑘

and

𝜏
exp⁡(𝑘𝑢) − 𝑒𝑥𝑝(−𝑘(𝑇𝑖 − 2𝑢))
𝐸 𝑇𝑖 [𝑥(𝜏)|ℱ𝑡 ] = 𝑒𝑥𝑝(−𝑘(𝜏 − 𝑡))𝑥(𝑡) − 𝑒𝑥𝑝(−𝑘𝜏) ∫ 𝜎(𝑢)2 𝑑𝑢
𝑡 𝑘

In a more general case where neither k nor  is constant, we can calculate the 2-dimensional integration
by first change the order of integration and then numerically integrating by splitting into sub-intervals

4. SOFR Derivative Pricing

4.1 SOFR Futures

A single Eurodollar futures contract is similar to a forward rate agreement to borrow or lend
US$1,000,000 for three months starting on the contract settlement date. Buying the contract is
equivalent to lending money, and selling the contract short is equivalent to borrowing money.

SOFR futures are similar to Eurodollar futures, and their final settlement price are calculated as

𝑃𝑟𝑖𝑐𝑒 = 100 − 𝑅

where 1M SOFR futures use simple interest

∑𝑖 𝛿𝑖 𝑟𝑖 𝑑𝑏 𝑟 × 𝑛
𝑖 𝑖 360
𝑅= × 100 = [∑ ]× × 100
∑𝑖 𝛿𝑖 𝑖=1 360 𝑑𝑐

while 3M SOFR futures use compounded interest

𝑑𝑏 𝑟𝑖 × 𝑛𝑖 360
𝑅 = [∏ (1 + ) − 1] × × 100
𝑖=1 360 𝑑𝑐

where db denotes the number business days during the calculation period, either 1M or 3M, and ri is the
SOFR rate published by Fed NY next day.

6
The 3M futures are worth $2500 per index point and the 1M futures are worth $4167 per index point.
One can think of the 3M futures as having a $1 million face value and the 1M futures as having a $5
million face value, since

1 3
2500 = 1000000 × ×
100 12
1 1
4167 ≈ 1000000 × ×
100 12

Note that SOFR futures uses day counter Act/360 Fixed. Also note that for 3M SOFR futures the month
symbol refers to the beginning of the calculation period, not the month for final settlement, while for
1M SOFR futures the month symbol refers to the delivery month for that contract.

Define a daily rolling money market account

𝑑𝑏 𝑟𝑖 × 𝑛𝑖
𝑀(𝑇) = 𝑀(𝑇0 ) ∏ (1 + )
𝑖=1 360

where db is the number of business days between T0 and T. With rolling money market account the
compounded interest can be rewritten as

𝑀(𝑇) 1
𝑅𝐺 = [ − 1] × × 100
𝑀(𝑇0 ) 𝑇 − 𝑇0

where T0 denotes the beginning of the calculation period and T is the settlement date. For example, for
3M SOFR futures T-T0 = 3M.

Like any other futures contracts there is a convexity adjustment between futures and forward price or
rate due to daily settlement, and the convexity adjustment is computed by

𝐹𝑅𝐺 (𝑡) − 𝑓𝑅𝐺 (𝑡) = 𝐸[𝑅𝐺 ] − 𝐸 𝑇 [𝑅𝐺 ]

where by definition the forward rate is martingale under the T-forward measure

1 𝑃(𝑡, 𝑇0 )
𝑓𝑅𝐺 (𝑡) = 𝐸 𝑇 [𝑅𝐺 ] = ( − 1)
𝛿 𝑃(𝑡, 𝑇)

With daily settlement, the SOFR futures rate (without discounting) must be a martingale under the spot
risk-neutral measure, i.e.

1 𝑀(𝑇) 1 𝐸[𝑀(𝑇)|ℱ𝑡 ]
𝐹𝑅𝐺 (𝑡) = 𝐸[ − 1| ℱ𝑡 ] = ( − 1)
𝑇 − 𝑇0 𝑀(𝑇0 ) 𝑇 − 𝑇0 𝑀(𝑇0 )

which requires an interest rate model in order to compute the convexity adjustment.

Assuming the SOFR rate is governed by a 1-Factor Hull-White model, we can calculate the futures SOFR
rate for T0>t

7
 1 𝑃(𝑡, 𝑇0 )
𝐹𝑅𝐺 (𝑡) = ( − 1)
𝑇 − 𝑇0 𝑃(𝑡, 𝑇)
1 𝑃(𝑡, 𝑇0 ) 1
+ [exp (𝐴(𝑡, 𝑇) − 𝐴(𝑡, 𝑇0 ) + 𝐴(𝑇0 , 𝑇) + 𝐵(𝑇0 , 𝑇)2 𝑉𝑎𝑟(𝑦𝑇0 )) − 1]
𝑇 − 𝑇0 𝑃(𝑡, 𝑇) 2

where A and B are piecewise functions

1 1
𝐵(𝑡, 𝑇) = + (𝐵(𝑇𝑛 , 𝑇) − ) exp(−𝑘𝑛 (𝑇𝑛 − 𝑡)) > 0
𝑘𝑛 𝑘𝑛
1 2 1 − exp(−2𝑘𝑛 (𝑇𝑛 − 𝑡)) 𝜎𝑛2 (𝑇𝑛 − 𝑡)
𝐴(𝑡, 𝑇) = 𝐴(𝑇𝑛 , 𝑇) + 𝜎𝑛2 (𝐵(𝑇𝑛 , 𝑇) − ) +
𝑘𝑛 4𝑘𝑛 2𝑘𝑛2
𝜎𝑛2 1 1 − exp(−𝑘𝑛 (𝑇𝑛 − 𝑡))
+ (𝐵(𝑇𝑛 , 𝑇) − )
𝑘𝑛 𝑘𝑛 𝑘𝑛

and y(T0) is normally distributed

𝑇0 𝑇0 𝑇0 𝑢
𝑦𝑇0 ~𝑁 (𝑦𝑡 𝑒𝑥𝑝 (− ∫ 𝜅(𝑢)𝑑𝑢) , 𝑒𝑥𝑝 (− ∫ 𝜅(𝑢)𝑑𝑢) ∫ 𝑒𝑥𝑝 (∫ 𝜅(𝑣)𝑑𝑣 ) 𝜎(𝑢)2 𝑑𝑢)
𝑡 𝑡 𝑡 𝑡

If T0<t, we have
𝑡 𝑇
1
𝐹𝑅𝐺 = [𝑒𝑥𝑝 (∫ 𝑟𝑠 (𝑢)𝑑𝑢) 𝑒𝑥𝑝 (2 ∫ 𝜑(𝑢)𝑑𝑢) 𝑃(𝑡, 𝑇) − 1]
𝑇 − 𝑇0 𝑇0 𝑡

where rs is the realized SOFR from T0 through t and P(t,T) can be calculated from the SOFR curve.

For 1M SOFR futures simple interest formula is used instead, which can be approximated by

𝑑𝑏 𝑇
𝑟𝑖 × 𝑛𝑖 𝑁 1 1 𝑀(𝑇)
𝑅𝐴 (𝑇) = [∑ ] × × 100 ≈ ∫ 𝑟𝑠 (𝑢)𝑑𝑢 × 100 = 𝑙𝑛 × 100
𝑖=1 𝑁 𝑑𝑐 𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0 𝑀(𝑇0 )

Its corresponding forward SOFR term rate

𝑇 𝑇
1 1 1
𝑓𝑅𝐴 (𝑡) = 𝐸 𝑇 [𝑅𝐴 ] = 𝐸[ (∫ 𝑟𝑠 (𝑢)𝑑𝑢) 𝐷(𝑡, 𝑇)] = 𝐸 𝑇 [(∫ 𝑟𝑠 (𝑢)𝑑𝑢)]
𝑃(𝑡, 𝑇) 𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0 𝑇0

which can be calculated analytically.

The 1M SOFR futures rate is martingale under the spot risk-neutral measure, and thus we have
𝑇 𝑇
1 1 exp(−𝑘(𝑇0 − 𝑡)) − exp(−𝑘(𝑇 − 𝑡))
𝐹𝑅𝐴 = 𝐸 [∫ 𝑟𝑠 (𝑢)𝑑𝑢| ℱ𝑡 ] = (∫ 𝜑(𝑢)𝑑𝑢 + 𝑥(𝑡))
𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0 𝑇0 𝑘

where the expected short rate

𝑢
𝐸[𝑟𝑠 (𝑢)|ℱ𝑡 ] = 𝜑(𝑡) + 𝑥(𝑡)𝑒𝑥𝑝 (− ∫ 𝑘𝑑𝑡)
𝑡

8
4.2 SOFR Fixed vs Floating Swaps

Unlike an SOFR-linked OIS swap, the floating leg of a fixed vs SOFR swap makes multiple payment as the
floating leg of a LIBOR swap.

In a SOFR fixed-floating swaps, the floating leg pays at time Tj

𝑑𝑏,𝑗 𝑟𝑖 × 𝑛𝑖 𝑀(𝑇𝑗 )
𝛿𝑗 𝑅𝐺 (𝑇𝑗 ) = ∏ (1 + )−1= −1
𝑖=1 360 𝑀(𝑇𝑗−1 )

where db,j is the number of business days between [Tj-1, Tj]. It is essentially an SOFR linked OIS swap with
multiple floating payments with reset in-arrears.

Assuming compound average and no payment delay the PV of the floating leg at time t is then given by

𝑛 𝑀(𝑇𝑗 ) 𝑛
𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑡) = ∑ 𝐸 [( − 1) 𝐷(𝑡, 𝑇𝑗 )] = ∑ 𝐸[𝐷(𝑡, 𝑇𝑗−1 )] − 𝐸[𝐷(𝑡, 𝑇𝑗 )]
𝑗=1 𝑀(𝑇𝑗−1 ) 𝑗=1
= 𝑃(𝑡, 𝑇0 ) − 𝑃(𝑡, 𝑇𝑛 )

With payment delay, e.g. 2 business days, the convexity adjustment should be taken into account in
theory. However, in practice the convexity adjustment is quite small and thus neglectable
𝑛
𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑡) = ∑ 𝛿𝑗 𝑓𝑅𝐺 (𝑇𝑗 )𝑃(𝑡, 𝑇𝑗,𝑝 )
𝑗=1

where the forward SOFR term rate

1 𝑃(𝑡, 𝑇𝑗−1 )
𝑓𝑅𝐺 (𝑇𝑗 ) = ( − 1)
𝛿𝑗 𝑃(𝑡, 𝑇𝑗 )

The fixed leg pays at time Tj for unit notional

(𝑇𝑖 − 𝑇𝑖−1 )𝑆

and its PV of the ith fixed coupon

𝑚 𝑚
𝑃𝑉𝐹𝑖𝑥𝑒𝑑 (𝑡) = ∑ 𝐸[(𝑇𝑖 − 𝑇𝑖−1 )𝑆 ∙ 𝐷(𝑡, 𝑇𝑖 )] = ∑ (𝑇𝑖 − 𝑇𝑖−1 )𝑆 ∙ 𝑃(𝑡, 𝑇𝑖 )
𝑖=1 𝑖=1

Thus, the breakeven SOFR swap rate

∑𝑛𝑗=1 𝛿𝑗 𝑓𝑅𝐺 (𝑇𝑗 )𝑃(𝑡, 𝑇𝑗,𝑝 )

𝑆(𝑡) =
∑𝑚
𝑖=1 𝛿𝑖 𝑃(𝑡, 𝑇𝑗 )

In certain circumstances we ignore the payment delay and the faire swap rate becomes

9
 𝑃(𝑡, 𝑇0 ) − 𝑃(𝑡, 𝑇𝑛 )
𝑆(𝑡) =
∑𝑚𝑖=1 𝛿𝑖 𝑃(𝑡, 𝑇𝑗 )

Note that assuming the swap cash flows use compound interest and are discounted at SOFR itself, the
valuation of SOFR fixed-floating swaps with compounded interest are exactly the same as a vanilla IR
swap based on Libor.

For simple interest, we have to assume an interest rate model and use the forward simple SOFR term
rate 𝑓𝑅𝐴 (𝑇𝑗 ) in place of 𝑓𝑅𝐺 (𝑇𝑗 ).

4.3 SOFR Swaptions

The underlying of a SOFR swaption is an SOFR fixed vs floating swap, which is similar to a LIBOR vanilla
swap except for the index and resetting.

Similar to a LIBOR swap, the values of a floating leg and a fixed leg are

𝑛 𝑀(𝑇𝑗 )
𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑡) = ∑ 𝐸 [( − 1) 𝐷(𝑡, 𝑇𝑗 )] = 𝑃(𝑡, 𝑇0 ) − 𝑃(𝑡, 𝑇𝑛 )
𝑗=1 𝑀(𝑇𝑗−1 )

and
𝑚 𝑚
𝑃𝑉𝐹𝑖𝑥𝑒𝑑 (𝑡) = ∑ 𝐸[(𝑇𝑖 − 𝑇𝑖−1 )𝑆 ∙ 𝐷(𝑡, 𝑇𝑖 )] = 𝑆 ∑ (𝑇𝑖 − 𝑇𝑖−1 ) ∙ 𝑃(𝑡, 𝑇𝑖 )
𝑘=1 𝑖=1

Thus, ignoring the payment delay the SOFR swap rate is the same a LIBOR swap with a single curve

𝑃(𝑡, 𝑇0 ) − 𝑃(𝑡, 𝑇𝑛 )
𝑆(𝑡) =
∑𝑚𝑖=1 𝛿𝑗 𝑃(𝑡, 𝑇𝑖 )

which is martingale under the annuity measure, 𝐴(𝑡, 𝑇0 , 𝑇𝑚 ) = ∑𝑚

𝑖=1 𝛿𝑗 𝑃(𝑡, 𝑇𝑖 )

Thus, the PV of a SOFR swaption

𝑚
𝑉𝑠𝑤𝑎𝑝𝑡𝑖𝑜𝑛 (𝑡) = 𝐸 [(𝑆 − 𝐾)+ ∑ 𝛿𝑗 𝐷(𝑡, 𝑇𝑖 )] = 𝐴(𝑡, 𝑇0 , 𝑇𝑚 )𝐸 𝐴 [(𝑆(𝑇0 ) − 𝐾)+ ]
𝑖=1

with 𝐸 𝐴 [𝑆(𝑇0 )] = 𝑆(𝑡).

A Bermudan swaption has to be valued numerically on a lattice, e.g. trinomial tree. The fixed leg can be
calculated easily by backward deduction. The floating leg can be calculated similar to a callable SOFR
floater.

With geometric averaging, the PV of a floating payment at the beginning of a payment period

𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑇𝑖−1 ; 𝑥𝑗 ) = 𝑁 ∙ (𝑓𝑅 + 𝑠) ∙ 𝛿𝑖 ∙ 𝑃(𝑇𝑖−1 , 𝑇𝑖 ; 𝑥𝑗 )

10
with

1 1
𝑓𝑅𝐺 = 𝐸 𝑇 [𝑅𝐺 ] = ( − 1)
𝛿𝑖 𝑃(𝑇𝑖−1 , 𝑇𝑖 ; 𝑥𝑗 )

which is the same as a LIBOR vanilla swap.

With arithmetic averaging, the PV of a floating payment given a state variable x(Ti-1)

𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑇𝑖−1 ; 𝑥𝑗 ) = 𝑁 ∙ (𝑓𝑅 + 𝑠) ∙ 𝛿𝑖 ∙ 𝑃(𝑇𝑖−1 , 𝑇𝑖 ; 𝑥𝑗 )

with
𝑇𝑖
2𝑘𝛿𝑖 − 3 + 4 exp(−𝑘𝛿𝑖 ) − exp(−2𝑘𝛿𝑖 ) 1 − exp(−𝑘𝛿𝑖 )
𝑓𝑅𝐴 = 𝐸 𝑇 [𝑅𝐴 ] = ∫ 𝜑(𝑢)𝑑𝑢 − 𝜎 2 + 𝑥(𝑇𝑖−1 )
𝑇𝑖−1 2𝑘 3 𝑘

where x(Ti-1) is given at each node.

4.4 SOFR Caps/Floors

SOFR cap/floor is similar to LIBOR cap/floor except that index rates are reset in arrears using either
simple average rate or compounded average rate. Thus, different from LIBOR cap/floor the first floating
payment is unknown at inception.

The value of each caplet/floorlet is given by

𝑉𝐶𝑎𝑝𝑙𝑒𝑡 (𝑡) = 𝐸[(𝑅∙ (𝑇𝑖 ) − 𝐾)+ 𝛿𝑖 𝐷(𝑡, 𝑇𝑖 )]

where the average SOFR rate R at Ti is given by

𝑀(𝑇𝑖 ) 1
𝑅𝐺 (𝑇𝑖 ) = [ − 1] ×
𝑀(𝑇𝑖−1 ) 𝑇𝑖 − 𝑇𝑖−1

or

1 𝑀(𝑇𝑖 )
𝑅𝐴 (𝑇𝑖 ) ≈ 𝑙𝑛
𝑇𝑖 − 𝑇𝑖−1 𝑀(𝑇𝑖−1 )

with the relationship

𝛿𝑖 ∙ 𝑅𝐺 (𝑇𝑖 ) = [exp(𝛿𝑖 ∙ 𝑅𝐴 (𝑇𝑖 )) − 1]

Assuming the short rate follows a mean-reverting process, RA(Ti) as the integral of a normal process is
normally distributed, and thus RG(Ti) is lognormally distributed.

In order to value SOFR cap/floor, we compute the mean and variance of both RA(Ti) and RG(Ti)

11
 𝑃(𝑡, 𝑇𝑖−1 )
𝐸 𝑇𝑖 [exp⁡(𝛿𝑖 ∙ 𝑅𝐴 (𝑇𝑖 ))] =
𝑃(𝑡, 𝑇𝑖 )
𝑇𝑖
2𝑘𝛿𝑖 − 3 + 4 exp(−𝑘𝛿𝑖 ) − exp(−2𝑘𝛿𝑖 ) 1 − exp(−𝑘𝛿𝑖 ) 𝑇
𝐸 𝑇𝑖 [𝛿𝑖 ∙ 𝑅𝐴 (𝑇𝑖 )] = ∫ 𝜑(𝑢)𝑑𝑢 − 𝜎 2 + 𝐸 𝑖 [𝑥(𝑇𝑖−1 )]
𝑇𝑖−1 2𝑘 3 𝑘

𝑉𝑎𝑟 𝑇𝑖 [𝑅𝐴 (𝑇𝑖 ) ∙ 𝛿𝑖 ] = 2𝑙𝑛(𝐸 𝑇𝑖 [exp⁡(𝛿𝑖 ∙ 𝑅𝐴 (𝑇𝑖 ))]) − 2𝛿𝑖 ∙ 𝐸 𝑇𝑖 [𝑅𝐴 (𝑇𝑖 )]

𝑉𝑎𝑟 𝑇𝑖 [exp⁡(𝑅𝐴 (𝑇𝑖 ) ∙ 𝛿𝑖 )] = exp⁡(2𝛿𝑖 ∙ 𝐸 𝑇𝑖 [𝑅𝐴 (𝑇𝑖 )] + 𝛿𝑖 2 𝑉𝑎𝑟 𝑇𝑖 [𝑅𝐴 (𝑇𝑖 )])(exp(𝛿𝑖 2 𝑉𝑎𝑟 𝑇𝑖 [𝑅𝐴 (𝑇𝑖 )]) − 1)

Then, Black-Scholes (Blank and Scholes, 1973) or Bachelier formula can be applied to value the caplet
accordingly
+
𝑉𝐶𝑎𝑝𝑙𝑒𝑡 (𝑡) = 𝐸[(𝑅𝐺 (𝑇𝑖 ) − 𝐾)+ 𝛿𝑖 𝐷(𝑡, 𝑇𝑖 )] = 𝑃(𝑡, 𝑇𝑖 )𝐸 𝑇𝑖 [(exp(𝛿𝑖 ∙ 𝑅𝐴 (𝑇𝑖 )) − (1 + 𝛿𝑖 𝐾)) ]

or

𝑉𝐶𝑎𝑝𝑙𝑒𝑡 (𝑡) = 𝐸[(𝑅𝐴 (𝑇𝑖 ) − 𝐾)+ 𝛿𝑖 𝐷(𝑡, 𝑇𝑖 )] = 𝛿𝑖 𝑃(𝑡, 𝑇𝑖 )𝐸 𝑇𝑖 [(𝑅𝐴 (𝑇𝑖 ) − 𝐾)+ ]

4.5 Options on SOFR Futures

Options on SOFR futures are very similar to options on Eurodollar futures. There are American options
on both 3M SOFR futures and 1M SOFR futures. Each option is exercisable into one specified underlying
instrument, i.e. CME SOFR futures contract.

The SOFR futures price is a function of the corresponding SOFR futures rate

𝑃𝑟𝑖𝑐𝑒 = 100 − 𝐹𝑅

where at settlement the settlement SOFR rate

𝑑𝑏 𝑟𝑖 × 𝑛𝑖 360 𝑀(𝑇) 1
𝑅𝐺 (𝑇) = [∏ (1 + ) − 1] × × 100 = [ − 1] × × 100
𝑖=1 360 𝑑𝑐 𝑀(𝑇0 ) 𝑇 − 𝑇0

and

𝑑𝑏 𝑇
𝑟𝑖 × 𝑛𝑖 𝑁 1 1 𝑀(𝑇)
𝑅𝐴 (𝑇) = [∑ ] × × 100 ≈ ∫ 𝑟𝑠 (𝑢)𝑑𝑢 × 100 = 𝑙𝑛 × 100
𝑖=1 𝑁 𝑑𝑐 𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0 𝑀(𝑇0 )

for 3-month and 1-month, respectively.

As derived earlier, a forward SOFR term rate is a martingale under the T-forward measure, while a
futures SOFR term rate is martingale under the spot risk-neutral measure.

12
Suppose the expiration time of a European SOFR call option is T. Its value is

𝑉𝐶𝑎𝑙𝑙 (𝑡) = 𝐸[(𝐹𝑅 (𝑇) − 𝐾)+ 𝐷(𝑡, 𝑇)]

where the SOFR futures rate R is martingale under the spot risk-neutral measure
𝑇
1
𝐹𝑅 (𝜏) = 𝐸 [𝑒𝑥𝑝 (∫ 𝑟𝑠 (𝑢)𝑑𝑢) − 1| ℱ𝜏 ]
𝑇 − 𝑇0 𝑇0

for geometric averaging, or

𝑇
1
𝐹𝑅 (𝜏) = 𝐸 [∫ 𝑟𝑠 (𝑢)𝑑𝑢| ℱ𝜏 ]
𝑇 − 𝑇0 𝑇0

for arithmetic averaging.

Using a one-factor affine model, e.g. Hull-White model, we have for geometric averaging
𝑇
1 1
𝐹𝑅 (𝜏) = 𝐸 [𝑒𝑥𝑝 (∫ 𝑟𝑠 (𝑢)𝑑𝑢) − 1| ℱ𝜏 ] =
𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0

𝑇
1
= (𝑒𝑥𝑝 (∫ 𝜑(𝑢)𝑑𝑢 ) 𝐸[exp(𝐴(𝑇0 , 𝑇) − 𝐵(𝑇0 , 𝑇)𝑦𝑇0 )|ℱ𝑡 ] − 1)
𝑇 − 𝑇0 𝑇0
𝑃(𝑡, 𝑇0 ) 1
exp (𝐴(𝑡, 𝑇) − 𝐴(𝑡, 𝑇0 ) + 𝐴(𝑇0 , 𝑇) − 𝐵(𝑇0 , 𝑇)𝐸[𝑦𝑇0 ] + 2 𝐵(𝑇0 , 𝑇)2 𝑉𝑎𝑟(𝑦𝑇0 )) − 1
𝑃(𝑡, 𝑇)
=
𝑇 − 𝑇0

where the state variable y(t) = -x(t) that is governed by

𝑑𝑦𝑡 = −𝜅𝑦𝑡 𝑑𝑡 − 𝜎(𝑡)𝑑𝑊

and
𝑇0
𝐸[𝑦𝑇0 |ℱ𝜏 ] = 𝑦(𝜏)exp⁡ (− ∫ 𝜅𝑑𝑢)
𝜏

Because y() is normally distributed and Var(yT0) is deterministic, FR() will be a shifted lognormal
process, and eventually the Black-Scholes formula can be applied to value a European option.

In the case of arithmetic averaging, similarly we have

𝑇 𝑇
1 1
𝐹𝑅 (𝜏) = 𝐸 [∫ 𝑟𝑠 (𝑢)𝑑𝑢| ℱ𝜏 ] = ∫ 𝐸[𝑟𝑠 (𝑢)|ℱ𝜏 ]𝑑𝑢
𝑇 − 𝑇0 𝑇0 𝑇 − 𝑇0 𝑇0
𝑇
1 exp(−𝑘(𝑇0 − 𝜏)) − exp(−𝑘(𝑇 − 𝜏))
= (∫ 𝜑(𝑢)𝑑𝑢 + 𝑥(𝜏))
𝑇 − 𝑇0 𝑇0 𝑘

Because x() is normally distributed, FR() will be a normal process and thus the Bachelier formula can be
applied to value a European option on SOFR 1M futures

13
Usually, options on SOFR futures have American style, i.e. they can be exercised any time before
settlement date, and therefore a trinomial tree can be built for the state variable x(t) and then early
exercise can be modeled on the lattice.

4.6 SOFR Callable Floaters

A SOFR floating rate note (FRN) pays floating coupons linked to either geometric or arithmetic average
of SOFR rate. In general, the present value of a one floating coupon is given by

𝑉𝑖 (𝑡) = 𝑁 ∙ 𝛿𝑖 𝐸[(𝑅∙ (𝑇𝑖 ) + 𝑠)𝐷(𝑡, 𝑇𝑖 )] = 𝑁 ∙ 𝛿𝑖 ∙ (𝑓𝑅∙ (𝑇𝑖 ) + 𝑠) ∙ 𝑃(𝑡, 𝑇𝑖 )

where fR.(Ti) is the average SOFR rate for the ith period and the spread can be treated as a fixed rate
coupon.

If a SOFR-linked structured note has an embedded option, the lattice method can be applied to evaluate
the option similar to any other callable bonds. To that end, we build a trinomial tree according to the
short rate model for SOFR term structure.

On a lattice the discount factor and transition probability for the kth path at t can be written as

𝐷𝑖,𝑘 (𝑡, 𝑇) = ∏ 𝐷𝑖,𝑘 (𝑇𝑙−1 , 𝑇𝑙 )

𝑙

and

𝑝𝑖,𝑘 = ∏ 𝑝𝑖,𝑘,𝑙
𝑙

which provides us a way to calculate the Arrow-Debru price recursively.

For each node on the lattice, the present value of a fixed rate cash flow can be easily calculated as

𝑃𝑉𝐹𝑖𝑥 = 𝛿𝑖 𝑐 ∙ 𝐸 [𝐸[𝐷(𝑡, 𝑇)|𝑥(𝑇) = 𝑥𝑗 ]] = 𝛿𝑖 𝑐 ∙ 𝐸 [∑ 𝐷𝑖,𝑘 (𝑡, 𝑇)𝑝𝑖,𝑘 ]

𝑘

where Di,k(t,T) and pi,k represent the discount factor and transition probability, respectively, of the kth
path that ends at the node xj, and their product is actually the Arrow-Debru price.

The calculation of a floating rate coupon is more complicated. Since the floating cash flows are path-
dependent, and thus non-combining, its PV cannot be calculated by backward deduction. Instead,
floating cash flows’ PVs should be calculated at the beginning of each cash flow period either analytically
or numerically.

14
As mentioned earlier, there are two types of interest rate based on SOFR rate, namely compound
interest and simple interest. For compound interest, the coupon payment can actually be efficiently
computed at the beginning of the coupon period Ti-1

1
𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑇𝑖−1 ; 𝑥𝑗 ) = 𝐸 [(⁡ − 1) 𝐷(𝑇𝑖−1 , 𝑇𝑖 ) + 𝑠𝛿𝑖 𝐷(𝑇𝑖−1 , 𝑇𝑖 )| 𝑥(𝑇𝑖−1 ) = 𝑥𝑗 ]
𝐷(𝑇𝑖−1 , 𝑇𝑖 )
= 1 + (𝑠𝛿𝑖 − 1)𝐸[𝐷(𝑇𝑖−1 , 𝑇𝑖 )|𝑥(𝑇𝑖−1 ) = 𝑥𝑗 ]

where s denotes the spread over SOFR.

Similarly, the PV of the floating payment based on simple average of SOFR at the beginning of the period
is given by

𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑇𝑖−1 ; 𝑥𝑗 ) = 𝐸[−𝑙𝑛𝐷(𝑇𝑖−1 , 𝑇𝑖 ) ∙ 𝐷(𝑇𝑖−1 , 𝑇𝑖 )|𝑥(𝑇𝑖−1 ) = 𝑥𝑗 ]

If the Hull-White model is used to model the dynamics of the interest rate, i.e.

𝑑𝑥 = −𝜅𝑥𝑑𝑥 + 𝜎𝑑𝑊

and

𝑟 = 𝜑(𝑡) + 𝑥(𝑡)

we have the present value of the payment at the beginning of the accrual period calculated analytically
𝑇𝑖
𝐸[𝑅𝐴 𝛿 ∙ 𝐷(𝑇𝑖−1 , 𝑇𝑖 )] = 𝑃(𝑇𝑖−1 , 𝑇𝑖 )𝐸 𝑇 [−𝑙𝑛𝐷(𝑇𝑖−1 , 𝑇𝑖 )] = 𝑃(𝑇𝑖−1 , 𝑇𝑖 ) (∫ (𝜑(𝑢) + 𝐸 𝑇 [𝑥(𝑢)])𝑑𝑢)
𝑇𝑖−1

𝑇𝑖
2𝑘𝛿 − 3 + 4 exp(−𝑘𝛿) − exp(−2𝑘𝛿) 1 − exp(−𝑘𝛿)
= 𝑃(𝑇𝑖−1 , 𝑇𝑖 ) (∫ 𝜑(𝑢)𝑑𝑢 − 𝜎 2 + 𝑥(𝑇𝑖−1 ))
𝑇𝑖−1 2𝑘 3 𝑘

and thus the expected average rate

𝑇𝑖
2𝑘𝛿 − 3 + 4 exp(−𝑘𝛿) − exp(−2𝑘𝛿) 1 − exp(−𝑘𝛿)
𝐸 𝑇 [𝑅𝐴 ] ∙ 𝛿 = ∫ 𝜑(𝑢)𝑑𝑢 − 𝜎 2 + 𝑥(𝑇𝑖−1 )
𝑇𝑖−1 2𝑘 3 𝑘

where x(Ti-1) is given.

If there is no analytic solution for each floating payment, approximations can be applied to simplify the
𝑇
calculation. Let 𝐷(𝑇𝑖−1 , 𝑇𝑖 ) = 𝑒𝑥𝑝 (∫𝑇 𝑖 −𝑟(𝑢)𝑑𝑢) ≔ 𝑦. Then, we have
𝑖−1

𝑃𝑉𝐹𝑙𝑜𝑎𝑡 (𝑇𝑖−1 ; 𝑥𝑗 ) = 𝐸[−𝑙𝑛𝑦 ∙ 𝑦|𝑥(𝑇𝑖−1 ) = 𝑥𝑗 ]

which can be approximated by Taylor expansion around E[y]

𝐸[𝑅𝐴 (𝑇)𝛿𝑖 ∙ 𝐷(𝑇𝑖−1 , 𝑇𝑖 )] ≈ 𝐸[− ln(𝐸[𝑦]) ∙ 𝐸[𝑦]] + (1 − ln(𝐸[𝑦]))(𝐸[𝑦 2 ] − 𝐸[𝑦]2 )

15
where both expectation
𝑇𝑖
𝐸[𝑦] = 𝐸 [𝑒𝑥𝑝 (∫ −𝑟(𝑢)𝑑𝑢)] = 𝑃(𝑇𝑖−1 , 𝑇𝑖 )
𝑇𝑖−1

and
𝑇𝑖
𝐸[𝑦 2 ] = 𝐸 [𝑒𝑥𝑝 (− ∫ 2𝑟(𝑢)𝑑𝑢)]
𝑇𝑖−1

can also be computed on a recombining tree by backward deduction.

4.7 SOFR CMS with cap/floor

With a single curve the breakeven SOFR swap rate

∑𝑛𝑗=1 𝛿𝑗 𝑓𝑅𝐺 (𝑇𝑗 )𝑃(𝑡, 𝑇𝑗,𝑝 )

𝑆(𝑡) =
∑𝑚
𝑖=1 𝛿𝑖 𝑃(𝑡, 𝑇𝑖 )

where the average SOFR rate

1 𝑃(𝑡, 𝑇𝑗−1 )
𝑓𝑅𝐺 (𝑇𝑗 ) = ( − 1)
𝛿𝑗 𝑃(𝑡, 𝑇𝑖 )

Ignoring the payment delay the faire swap rate becomes

𝑃(𝑡, 𝑇0 ) − 𝑃(𝑡, 𝑇𝑛 )
𝑆(𝑡) =
∑𝑚𝑖=1 𝛿𝑖 𝑃(𝑡, 𝑇𝑖 )

which is the same as the LIBOR swap rate.

If the discount curve is different from the forecast curve, similarly we have the breakeven swap rate

𝑃𝑑 (𝑡, 𝑇𝑗,𝑝 )
∑𝑛𝑗=1 𝛿𝑗 (𝑃𝑓 (𝑡, 𝑇𝑗−1 ) − 𝑃𝑓 (𝑡, 𝑇𝑗 ))
∑𝑛𝑗=1 𝛿𝑗 𝑓𝑅𝐺 (𝑇𝑗 )𝑃𝑑 (𝑡, 𝑇𝑗 ) 𝑃𝑓 (𝑡, 𝑇𝑗 )
𝑆(𝑡) = =
∑𝑚𝑖=1 𝛿𝑖 𝑃𝑑 (𝑡, 𝑇𝑖 ) ∑𝑚
𝑖=1 𝛿𝑖 𝑃𝑑 (𝑡, 𝑇𝑖 )

where ZCB bond ratios are lognormally distributed under the Tpj-forward measure.

Both dual-curve and payment lag can be handled with the dual curve method, in which the SOFR swap
rate conditional on the state variable Z

16
 𝑃𝑓 (𝑡, 𝑇𝑘−1 ) −1𝛽𝑘−1 (𝑇𝑗)2 −𝛽𝑘−1 (𝑇𝑗)𝑍 𝑃𝑓 (𝑇𝑗 , 𝑇𝑘 ) −1𝛽𝑘(𝑇𝑗)2 −𝛽𝑘(𝑇𝑗)𝑍 𝑃𝑑 (𝑡, 𝑇𝑗,𝑝 )
∑𝑘 ( 𝑒 2 − 𝑒 2 )
𝑃𝑓 (𝑡, 𝑇𝑝𝑗 ) 𝑃𝑓 (𝑡, 𝑇𝑝𝑗 ) 𝑃𝑓 (𝑡, 𝑇𝑗 )
𝑆(𝑡; 𝑍) =
𝑃 (𝑇 , 𝑇 ) 1 2
𝑃 (𝑡, 𝑇𝑖 )
∑𝑖 𝛿𝑖 𝑓 𝑗 𝑖 𝑒 −2𝛽𝑖(𝑇𝑗) −𝛽𝑖(𝑇𝑗)𝑍 𝑑
𝑃𝑓 (𝑡, 𝑇𝑝𝑗 ) 𝑃𝑓 (𝑡, 𝑇𝑖 )

where the state variable Z is normally distributed under the Tpj-forward measure with volatility

𝑇 𝑇
𝛽(𝑇) = √∫ 𝜎𝐹𝐵 (𝑡)2 𝑑𝑡 = √∫ (𝐵(𝑡, 𝑇𝑗 ) − 𝐵(𝑡, 𝑇))2 𝜎(𝑡)2 𝑑𝑡
0 0

Given the distribution of the SOFR swap rate, the CMS rate and a CMS optionlet can be valued
numerically

𝐶𝑀𝑆(𝑡) = 𝐸 𝑇𝑝𝑗 [𝑆(𝑡; 𝑍)]

and

𝑃𝑉𝑜𝑝𝑡𝑖𝑜𝑛𝑙𝑒𝑡 (𝑡) = 𝑃𝑑 (𝑡, 𝑇𝑝 )𝐸 𝑇𝑝𝑗 [𝑚𝑎𝑥(𝜔(𝑆(𝑡; 𝑍) − 𝐾), 0)]

where the option type indicator 𝜔 = ±1 .

5. Summary

In 2017 the Secured Overnight Financing Rate (SOFR) was recommended as the replacement for USD
LIBOR, and the transition from USD LIBOR to SOFR is set to complete by the end of 2021, after which the
SOFR rate will become the primary benchmark floating rate used by most financial products. Since the
announcement SOFR-linked derivatives started to develop and their liquidities have also improved.
Because of daily SOFR rate averaging, arithmetic average in particular, and timing mismatch pricing
SOFR derivatives is more complicated than pricing LIBOR derivatives, even if a vanilla IR swap. In order
to address these challenges, we apply a mean-reverting Gaussian model, e.g. 1-factor Hull-White model,
to model the dynamics of the instantaneous SOFR rate and derive analytic formula for forward SOFR
term rates as well as future SOFR term rates, which can then be used to price most linear SOFR
products. Moreover, based on an affine interest rate model for the SOFR rate closed-form solutions are
also derived for European SOFR options, including SOFR swaptions, SOFR cap/floor and options on SOFR
futures. American options cannot be priced analytically, but efficient numerical methods are developed
by combining lattice and analytic formula. Although a 1-facotor Hull-White model is used in this paper, it
is almost straightforward to extend it to multi-factor affine models.

References

17
ARRC (2021) A User's Guide to SOFR.

Black, Fischer and Myron Scholes (1973). The Pricing of Options and Corporate Liabilities. Journal of
Political Economy. 81 (3): 637–654.

Damiano Brigo and Fabio Mercurio (2001) Interest Rate Models — Theory and Practice with Smile,
Inflation and Credit (2nd ed. 2006 ed.). Springer Verlag. ISBN 978-3-540-22149-4.

18