Index models

Kristien Smedts

KU Leuven
Readings

BKM - Chapter 8: Index models

Can mean-variance optimization be used in practice?

Kristien Smedts (KU Leuven) Index models 2 / 40

Can MV optimization be used in practice?

Even though many MV statistical software exists, serious questions can be

raised as to whether they work well:
1 Obtaining reliable and accurate estimates of returns is hard (cf GIGO)
Inconsistent/noisy inputs lead to unreasonable solutions or no solution

2 Estimating the covariance matrix requires many estimates (for a

N (N +1)
portfolio of N assets, the covariance matrix has 2 inputs)
Example: a portfolio of 200 assets has a covariance matrix with 20,010 inputs
3 When the number of assets to include is large, while the historical
data availability is limited (N > T ), the covariance matrix is singular
the MV optimization then detects arbitrage opportunities, with
unreasonably concentrated portfolios as a consequence
even when the number of assets to include is large relative to the
historical sample length (still N < T ), the MV optimization will
perceive near-arbitrage opportunities

Kristien Smedts (KU Leuven) Index models 3 / 40

Can MV optimization be used in practice?

Such difficulties and limitations have initiated a search for more practical
solutions to find optimal portfolios

The baseline of these solutions is to introduce more structure on

the return dynamics in order to shorten the input list
Different approaches can be distinguished in the literature
statistical-based index models (single-index models and multi-index
models)
equilibrium pricing models (CAPM and APT)
Both approaches share the idea that risk can be decomposed in
systematic risk and idiosyncratic risk

In the current course, we focus on the statistical models

Kristien Smedts (KU Leuven) Index models 4 / 40

Roadmap

1 Single-index model

2 Portfolio construction in a single-index world

3 Portfolio construction in a CAPM world

4 Portfolio construction in a pragmatic way

5 The mean-variance model: a critical note

Kristien Smedts (KU Leuven) Index models 5 / 40

Single-index model
A statistical model

Assume we have an asset i with returns Ri

Ri = β i m + ε i + µ

with
Ri are joint normally distributed
m is a common risk factor −→ systematic risk
β i is the response of asset i to the common factor/index
ε i is a random variable with mean 0 and variance σ2 and uncorrelated
with the common factor −→ idiosyncratic risk
µ is the baseline return

To make this model operational, we impose intuitive economic

structure: the common risk factor is proxied by a broad market index
(e.g. S&P500 for US assets; or Euronext100 for Belgian assets)
Kristien Smedts (KU Leuven) Index models 6 / 40
Single-index model
A statistical model

The single-index model can then be rewritten as a regression equation:

e e
Ri,t = β i Rm,t + ε i,t + αi

with
e excess returns for asset i over time t
Ri,t
e
Rm,t excess returns on a broad market index over time t (generating
systematic risk)
β i the response of asset i to the systematic risk factor
ε i,t a residual with mean 0 and variance σε2i and cov (ε i , Rm
e = 0, and
)
cov (ε i , ε j ) = 0 (capturing idiosyncratic risk)

e  R e = 0 the expected excess return when the market


αi = E Ri,t m,t
excess return is zero (= constant)

Kristien Smedts (KU Leuven) Index models 7 / 40

Single-index model
A statistical model

The security’s risk premium then equals:

e e



E Ri,t = β i E Rm,t + αi

This risk premium can be decomposed into two parts

1 The systematic risk premium β E R e


i m,t

2 A non-market premium αi or alpha: it cannot persist in equilibrium

as it is arbitraged away (superior security analysis lies in the
identification of αi investments)

Kristien Smedts (KU Leuven) Index models 8 / 40

Single-index model
A statistical model

The security’s variance then equals:

σ2 Ri,t
e
= σi2 = β2i σm
2
+ σε2i

An asset’s risk can be decomposed into two parts

1 Systematic variance β2i σm
2

2 Idiosyncratic variance σε2i

Kristien Smedts (KU Leuven) Index models 9 / 40

Single-index stock market
A statistical model

Characteristics of the index portfolio

What is the β, firm-specific risk and α of the index portfolio m?

Kristien Smedts (KU Leuven) Index models 10 / 40

Single-index stock market
Estimation

Estimation of a single index model

Assume a sample of monthly returns (Feb 2011 - Jan 2017) for an
individual stock IBM and the common risk factor S&P500 index, and 3
month T-bill as risk-free rate.

Kristien Smedts (KU Leuven) Index models 11 / 40

Single-index stock market
Estimation

Estimation of a single index model

IBM excess returns follow S&P500 excess returns, but there is
important firm-specific movement in returns → the correlation is 51%
implying that IBM does not track changes in the returns of the
S&P500 very closely
Volatility of IBM excess returns seems slightly larger → see the
sample volatility and min/max observations

ReIBM ReS&P500
Mean 0.0041 0.0102
St.Dev 0.0480 0.0329
Min -0.1340 -0.0703
Max 0.1557 0.1093
Correl 51.24%

Kristien Smedts (KU Leuven) Index models 12 / 40

Single-index stock market
Estimation

Estimation of a single index model

Estimate T-stat
α -0.0035 -0.6864
β 0.7482∗∗∗ 4.9917
2
R 25.20%
Observations 72

β is smaller than 1 indicating a relatively low market sensitivity

α is negative, though not statistically significant
2
R is only 25% which implies that the variation in the index returns
explains about 25% of the variation in the IBM returns; the remainder
is idiosyncratic risk

Kristien Smedts (KU Leuven) Index models 13 / 40

Single-index model
Advantages

This single-index model greatly shortens the input list in a MV

optimization: for a portfolio of N assets, only (3N + 2) estimates are
needed
N non-market premia
N betas
N firm-specific variances σε2i
e


1 market risk premium E Rm,t
2
1 systematic risk variance σm

This is clearly a great reduction in the number of parameters to estimate

What about the covariances of returns?

Kristien Smedts (KU Leuven) Index models 14 / 40

Single-index model
Advantages

No additional parameters are needed to obtain estimates of the

covariances. We can easily prove this, by deriving expressions for the
different dimensions of risk
Variance of asset i:
σ2 Ri,t
e
= σi2 = β2i σm
2
+ σε2i

Covariance between asset i and j:

e e 2


cov Ri,t , Rj,t = covij = β i β j σm
Correlation between asset i and j:
2
β i β j σm 2 σ2
β i β j σm
e e m


ρ Ri,t , Rj,t = ρij = = 2
= ρim ρjm
σi σj σi σj σm

Note that these risk expressions clearly indicate that specialization of

security analysis within e.g. industries in such index-model is possible
Kristien Smedts (KU Leuven) Index models 15 / 40
Single-index model
Advantages

The single-index model also offers important insights into the benefits of
diversification

Assume the single-index model for a portfolio of N assets:

Rie = αi + β i Rm
e
+ εi

For an equally weighted portfolio of N assets, it then holds that:

Rpe = αp + β p Rm
e
+ εp

with
1 N 1 N 1 N
N i∑ N i∑ N i∑
αp = αi βp = βi εp = εi
=1 =1 =1

Kristien Smedts (KU Leuven) Index models 16 / 40

Single-index model
Advantages

To understand the benefits of diversification, we need to analyze the

portfolio variance:
σp2 = β2p σm2
+ σε2p

1 The systematic risk component (β2p σm 2 ) depends on the individual

sensitivities and the variance of the common factor; this part persists
no matter how many or which assets are added
2 The idiosyncratic component (σε2p ) depends on the firm-specific risks;
as they are independent (all with zero mean), the law of averages
implies that for increasing N these components cancel out:
!
1 1 N 2 1 N −→∞
2
σε p = ∑
N N i =1
σε i = σ2 = 0
N

Kristien Smedts (KU Leuven) Index models 17 / 40

Single-index model
Advantages

Can we construct a portfolio with zero systematic risks?

Kristien Smedts (KU Leuven) Index models 18 / 40

Roadmap

1 Single-index model

2 Portfolio construction in a single-index world

3 Portfolio construction in a CAPM world

4 Portfolio construction in a pragmatic way

5 The mean-variance model: a critical note

Kristien Smedts (KU Leuven) Index models 19 / 40

Portfolio construction in a single-index world

The main advantage of the single-index model is that less parameters need
to be estimated:

1 It allows for a specialization and organizational decentralisation,

without loss of consistency
2 The estimation of the market risk premium is dis-entangled from the
estimation of the non-market risk premia
3 Common risk is dis-entangled from idiosyncratic risk

Kristien Smedts (KU Leuven) Index models 20 / 40

Portfolio construction in a single-index world

The input list of a single-index model in a portfolio allocation model

consists of
1 Macro-economic analysis to estimate the risk and risk premium of the
market index
2 Statistical analysis to estimate expected returns and risk of individual
securities, absent any security analysis
estimates of the β sensitivities of the individual portfolio constituents
estimates of the residual variances σε2i

3 Security analysis to generate incremental (alpha) risk premia or

non-market returns

Kristien Smedts (KU Leuven) Index models 21 / 40

Portfolio construction in a single-index world

Assume an investment set consisting of N individual risky securities

The optimal risky portfolio is found as:

E Rpe


N
maxSp = s.t. ∑ wi = 1
wi σp i =1

with:
N N
E Rpe = αp + E (Rm
e
∑ wi αi + E (Rme ) ∑ wi βi


) βp =
i =1 i =1
 !2  21
  21 N N
σp = β2p σm
2
+ σε2p = ∑ wi β 2
σm + ∑ wi2 σe2i 
i =1 i =1

Kristien Smedts (KU Leuven) Index models 22 / 40

Roadmap

1 Single-index model

2 Portfolio construction in a single-index world

3 Portfolio construction in a CAPM world

4 Portfolio construction in a pragmatic way

5 The mean-variance model: a critical note

Kristien Smedts (KU Leuven) Index models 23 / 40

Portfolio construction in a CAPM world

Mean-variance optimization, CAPM and passive investment

Assume a world in which:

1 Investors are alike: they are mean-variance optimizers with common
time-horizon and common set of information as reflected in input
estimates
2 Markets are well-functioning: there are no frictions and impediments
to trading

Kristien Smedts (KU Leuven) Index models 24 / 40

Portfolio construction in a CAPM world

It then follows that:

All investors calculate an identical efficient frontier

All investors identify an identical CAL
All investors arrive at an identical tangency portfolio

The tangency portfolio is the (market-weighted) market portfolio and

CAPM holds such that there is no α in equilibrium → investors can skip
the trouble of MV and security analysis and obtain an efficient portfolio by
just holding the market portfolio.
The passive strategy is efficient!

The CAL of the market portfolio is labeled CML = Capital Market Line

Kristien Smedts (KU Leuven) Index models 25 / 40

Roadmap

1 Single-index model

2 Portfolio construction in a single-index world

3 Portfolio construction in a CAPM world

4 Portfolio construction in a pragmatic way

5 The mean-variance model: a critical note

Kristien Smedts (KU Leuven) Index models 26 / 40

Portfolio construction in a pragmatic way

Does the CAPM hold, at all times? NO!

Is the CAPM a reasonable benchmark model? YES!

A pragmatic approach therefore consists of combining:

the diversification benefits and cost-efficiency of a passive portfolio

security analysis targeted at outperformance of an active portfolio

Kristien Smedts (KU Leuven) Index models 27 / 40

Portfolio construction in a pragmatic way

Assume an investment set consisting of N+1 securities:

1 market portfolio → the market portfolio can be thought of as the
passive portfolio
N individual risky assets → these individual securities can be
combined into an active portfolio

The portfolio problem then reduces to solving for:

1 the optimal weights of this passive and active portfolio
2 the optimal weights within the active portfolio

Kristien Smedts (KU Leuven) Index models 28 / 40

Portfolio construction in a pragmatic way

As before...
The optimal risky portfolio is found as:

E Rpe


N+1
maxSp = s.t. ∑ wi = 1
wi σp i =1

N+1 N+1
E Rpe = αp + E (Rm
e
) β p = ∑ wi αi + E (Rm
e
) ∑ wi β i


i =1 i =1
 !2  21
  21 N+1 N+1
σp = β2p σm
2
+ σε2p = ∑ wi β 2
σm + ∑ wi2 σe2 
i
i =1 i =1

Kristien Smedts (KU Leuven) Index models 29 / 40

Portfolio construction in a pragmatic way

Unlike before...
analytical expressions can be derived for the optimal risky portfolio,
making the traditional MV optimization superfluous.

It can be shown that the optimal risky portfolio in the single-index model
is a combination of

The active portfolio (denoted a)

The passive market portfolio (denoted m)

See BKM for a step-by-step procedure of the analytical optimization

procedure

Kristien Smedts (KU Leuven) Index models 30 / 40

Portfolio construction in a pragmatic way

The Sharpe ratio of a risky portfolio with weights wa and wm = 1 − wa

exceeds the Sharpe ratio of the passive-only portfolio as follows:
 2
αa
Sp2 = 2
Sm +
σε a

The active portfolio contributes to the Sharpe ratio by the ratio of its
alpha to its residual standard deviation, known as the information
ratio: active return over tracking error
It measures the extra return from security analysis, compared to its
idiosyncratic risk when we over-or underweight securities relative to
the passive portfolio → alpha comes at a cost of increased variance!

To achieve the highest Sharpe ratio possible, this information ratio

needs to be maximized

Kristien Smedts (KU Leuven) Index models 31 / 40

Portfolio construction in a pragmatic way

Characteristics of the optimal risky portfolio

What are the allocation implications for a security with negative α?

When is the market portfolio m efficient?

Kristien Smedts (KU Leuven) Index models 32 / 40

Roadmap

1 Single-index model

2 Portfolio construction in a single-index world

3 Portfolio construction in a CAPM world

4 Portfolio construction in a pragmatic way

5 The mean-variance model: a critical note

Kristien Smedts (KU Leuven) Index models 33 / 40

The mean-variance model: a critical note
µ σ S excess K Min Max
Panel A: Daily returns
CRSP index (value weighted) 0.044 0.82 -1.33 34.92 -18.80 8.87
CRSP index (equal weighted) 0.073 0.76 -0.93 26.03 -14.19 9.83
IBM 0.039 1.42 -0.18 12.48 -22.96 11.72
General Signal corp. 0.054 1.66 0.01 3.35 -13.46 9.43
Wrigley Co. 0.072 1.45 -0.00 11.03 -18.67 11.89
Interlake Corp. 0.043 2.16 0.72 12.35 -17.24 23.08
Raytech Corp. 0.050 3.39 2.25 59.40 -57.90 75.00
Ampco-Pittsburgh Corp. 0.053 2.41 0.66 5.02 -19.05 19.18
Energen Corp. 0.054 1.41 0.27 5.91 -12.82 11.11
General Host Corp. 0.070 2.79 0.74 6.18 -23.53 22.92
Garan Inc. 0.079 2.35 0.72 7.13 -16.67 19.07
Continental Materials Corp. 0.143 5.24 0.93 6.49 -26.92 50.00

Source: Campbell, Lo and MacKinlay [1998] p.21.

Variance as a proxy of risk is reasonable for a normal distributed

variable, but in reality, returns show evidence of skewness and
kurtosis.
A portfolio model that neglects these dimensions is then likely not
optimal
Kristien Smedts (KU Leuven) Index models 34 / 40
The mean-variance model: a critical note

"In particular, skewness and kurtosis are generally ignored as criteria for
evaluating portfolio decisions, even though studies generally admit that
asset returns are positively skewed and leptokurtic."
Source: You and Daigler (2010) p. 164

"In measuring these gains from diversifcation, previous studies assumed that
returns are normally distributed. However, this assumption is contradicted
by numerous empirical findings (e.g., Fama, 1965; SImkowitz and Beedles,
1978; Singleton and Wingender, 1986). These studies imply that the
two-parameter CAPM, which assumes either quadratic utilty functions for
investors or normally distributed stock returns, is insufficient in making
optimal investment decisions."
Tang and Choi (1998) p. 119

Kristien Smedts (KU Leuven) Index models 35 / 40

The mean-variance model: a critical note

How is the investor’s appetite for skewness and kurtosis?

An investor will be averse to left-skewness, but will be happy with right-
skewness. Also, an investor will be averse towards kurtosis

What are the benefits of diversification for skewness and kurtosis?

A number of studies find that the idea that a portfolio diversifies volatility
can not just be extended to the risk as captured in skewness and kurtosis.
1 Skewness: adding more assets seems to introduce undesired negative
skewness as compared to desired positive skewness
2 Kurtosis: adding more assets increases the kurtosis

Kristien Smedts (KU Leuven) Index models 36 / 40

The mean-variance model: a critical note

"Diversification then is a two-edged sword: it eliminates undesired variance

in return distributions, but also eliminates desired skewness."
Source: Mitton and Vorking (2007) p. 1256

"It is found that diversification does not reduce either skewness or kurtosis.
As the portfolio size increases, portfolio returns become more negatively
skewed and more leptokurtic."
Source: Tang and Choi (1998) p. 119

"We find that skewness risk and kurtosis may even increase (i.e.
skewness decrease) by building diversified portfolios. Hence, a general
assumption of risk mitigation by diversification does not hold if higher
moments are taken into account."
Source: Walther (2014) p. 255

Kristien Smedts (KU Leuven) Index models 37 / 40

The mean-variance model: a critical note
Also, diversification benefits also seem to vary over time due to
time-varying correlations

In particular, correlations seem to increase in a more volatile market,

implying that benefits of diversification disappear exactly when
needed most

"Correlations are higher during bear markets and tend to fall during
periods of recovery."
Source: Antoniou et al. (2007) p. 173

"In our study, most of the correlations are larger during the bear market
relative to the bull market, confirming previous findings that markets possess
higher correlation during more volatile times." and "Thus, one needs to
take the variability of correlations into account in order to evaluate the
effectiveness and stability of diversification."
Source: You and Daigler (2010), p.166 and 169
Kristien Smedts (KU Leuven) Index models 38 / 40
The mean-variance model: a critical note

MV optimization is mechanical: small differences in estimates can

have large effects on estimated allocations

This results in concentrated portfolios, that require lots of rebalancing

To overcome this a number of solutions exist:

Opt for the pragmatic approach in which passive and active

management are combined

Apply a robust resampling approach to estimate the efficient frontier

Kristien Smedts (KU Leuven) Index models 39 / 40

Next

How likely is to detect and exploit alpha returns?

Kristien Smedts (KU Leuven) Index models 40 / 40