Single-Index Model

⃝Kilenthong
c 2017

Single-Index Model 1 / 25
Main Issues

Implementation of portfolio theory: this process requires correlation

structure of security returns.

Single-Index Model 2 / 25
Correlation Structure of Security Returns is a Key Input
Recall from Mean-Variance Portfolio:
( )
Z = Σ−1 R̄ − RF 1 (1)

This implies that correlation structure of security returns, Σ−1 , is a

key input to an optimal portfolio problem.
Consider a portfolio P. Its expected return and expected variance (or
risk) are

∑
N
R̄P = Xi R̄i (2)
i=1
[ ]1
∑
N ∑
N ∑
N 2

σp = Xi2 σi2 + Xi Xj σi σj ρij (3)

i=1 i=1 j=1
j̸=i

Single-Index Model 3 / 25
Correlation Structure of Security Returns is a Key Input
How many correlations do we need to estimate when there are N
assets? Answer is N(N−1)
2 .
In order to estimate correlation of each pair of assets, ρij , we need to
have data of historical returns of both assets.
A classical problem is that a finance firm traditionally organize their
analysts along industry line. That is, each analyst will know very well
about one asset only.
Therefore, calculating correlation requires coordination effort (which
is costly as well).
In my opinion, this may not be an issue with the database technology
now a day?
But still this model may bring more predictive power than simple
covariance calculation!
Anyway, suppose it is still a problem for now!

Single-Index Model 4 / 25
Market Return is the Single Index
To solve this coordination requirement problem, we need to find a
model that all correlations can be calculated or derived from simple
statistics that each analyst can provide without coordinating with
each other.
One solution is: to assume that return of each asset i, Ri depends on
the market return, Rm :

Ri = ai + βi Rm (4)

where
ai is the component of security i’s performance — a random variable
(with ai = αi + ϵi ).
ϵi is a random error— a random variable.
Rm is the rate of return on the market index — a random variable.
βi is a constant that measures the expected change in Ri given a change
in Rm (sensitivity).

Single-Index Model 5 / 25
Regression Equation
The equation for the return on a security can be written as

Ri = αi + βi Rm + ϵi (5)

This is a linear regression equation.

Key Assumptions:
1 Error term ϵi is unrelated/uncorrelated to the market return Rm :

E [ϵi (Rm − E (Rm )] = 0, for all stocks i, (6)

2 Securities are related through common response to market only:

E (ϵi ϵj ) = 0, for all pairs of stocks i ̸= j, (7)

By construction:
1 Mean of ϵi is zero:

E (ϵi ) = 0, for all stocks i. (8)

Single-Index Model 6 / 25
Regression Equation

By definition:
1 Variance of ϵi is

E (ϵi )2 = σ 2 = σϵ2i , for all stocks i, (9)

2 Variance of Rm is
2
E (Rm − E (Rm )) = σm
2
, for all stocks i, (10)

Single-Index Model 7 / 25
Covariance Structure from the Single Index

The mean of a security i return is

E (Ri ) = E (αi + βi Rm + ϵi ) = αi + βi E (Rm ) (11)

The Variance of a security i return is

σi2 = βi2 σm
2
+ σϵ2i (12)

The covariance of returns between securities i and j,

2
σij = βi βj σm · (13)

Single-Index Model 8 / 25
Example

Table: Decomposition of Returns for the Single-Index Model

Month Return on Stock Return on Market

1 10 4
2 3 2
3 15 8
4 9 6
5 3 0

Suppose βi = 1.5, what are αi and σϵi2 (the size of the error)?
αi = 2 and σϵi2 = 2.8.

Single-Index Model 9 / 25
Covariance Structure of a Portfolio
Let (Xi )N
i=1 be a portfolio P. Its expected return is
∑
N
E (Rp ) = Xi E (Ri ) (14)
i=1
Using the result above, we have
∑
N ∑
N
E (Rp ) = Xi αi + Xi βi E (Rm ) (15)
i=1 i=1
Its variance is
∑
N ∑
N ∑
N
σP2 = Xi2 σi2 + Xi Xj σij (16)
i=1 i=1 j=1
j̸=i

Using the result above, we have

∑
N ∑
N ∑
N ∑
N
σP2 = Xi2 βi2 σm
2
+ 2
Xi Xj βi βj σm + Xi2 σϵi2 (17)
i=1 i=1 j=1 i=1
j̸=i
Single-Index Model 10 / 25
Characteristics of the Single-Index Model

Using the result in the previous slide, we can show that

∑
N
βp = Xi βi (18)
i=1

and

∑
N
αp = X i αi (19)
i=1

Hence, we can write

E (Rp ) = αp + βp E (Rm ) (20)

If P is the market portfolio, then its βm = 1 and αm = 0.

Single-Index Model 11 / 25
Diversifiable Risk

The risk of an individual portfolio is

∑
N ∑
N ∑
N ∑
N
σp2 = Xi2 βi2 σm
2
+ Xi2 σϵi2 + 2
Xi Xj βi βj σm (21)
i=1 i=1 i=1 j=1
j̸=i

which can be further rearrange as

( )
1 ∑
N
1 2
σp2 = βp2 σm
2
+ σ (22)
N N ϵi
i=1

The average residual risk term is going to be very small when the
number of securities N is sufficiently large. Hence, it is common to
refer to σϵi2 as diversifiable or nonsystematic risk.

Single-Index Model 12 / 25
Systematic Risk

On the other hand, if we assume that the diversifiable term is zero,

then portfolio P’s risk is
[ ]2
∑
N
σp2 = σm
2
Xi βi (23)
i=1

That is, βi measures security i’s systematic or nondiversifiable risk.

Single-Index Model 13 / 25
Estimating Historical βs

Recall:

Ri = αi + βi Rm + ϵi (24)

This is a linear regression equation.

We can use historical or past time-series data on security Rit and
market return Rmt to estimate βi :
∑60 [ ]
Cov (Ri , Rm ) t=1 (Rit − R̄it )(Rmt − R̄mt )
βi = 2
= ∑60 ( )2 (25)
σm Rmt − R̄mt
t=1
αi = E (Ri ) − βi E (Rm ) = R̄i − βi R̄m (26)

Of course, in practice, this is a very easy task now a day. You can just
run it on a statistical program such as STATA.

Single-Index Model 14 / 25
Accuracy of Historical βs
If the βi is accurate, we should see a high correlation between βi from
different periods. See the Table below (from Blume (1970))

Table: Association of Betas Over Time

Number of Securities Correlation Coefficient of
in the Portfolio Coefficient Determination (R 2 )
1 0.60 0.36
2 0.73 0.53
20 0.97 0.95
35 0.97 0.95
50 0.98 0.96

β on individual securities are not so accurate (corr is about 0.60).

β of a large portfolio is very accurate.
What should we do if we really need a β of an individual security?
Single-Index Model 15 / 25
βs in Thailand

Single-Index Model 16 / 25
βs in Thailand

Single-Index Model 17 / 25
βs in Thailand

Single-Index Model 18 / 25
βs in Thailand

Single-Index Model 19 / 25
βs in Thailand

Single-Index Model 20 / 25
Adjusting Historical βs

We will need to account for the fact that β last period may have a
limit power to predict β this period.
Blume’s Technique: this approach is based on a simple regression,
e.g.,

βi2 = 0.343 + 0.677βi1 (27)

where βi1 stands for the β of stock/asset i for the earlier period, and
βi2 stands for the β of stock/asset i for the later period. This is a
result of a (cross-sectional) regression.

Single-Index Model 21 / 25
Adjusting Historical βs: Vasecek’s Technique

Vasecek’s Technique: this approach is based on a Bayesian estimation

technique (use information on variances), i.e.,

σβ̂2 2
σβ1
¯
i1
βi2 = β̄1 + β̂i1 (28)
σβ̄2 + σ 2 σβ̄2 + σ 2
1 β̂i1 1 β̂i1

where
β̄1 stands for a weighted average of β over all securities
using historical data,
2
σβ1 stands for a cross-sectional variance of β̂i1 across all securities
¯
using historical data, and
2
σβ̂ stands for the square of the standard error of the estimate of β
i1
for a security i (variance of an estimator).

Single-Index Model 22 / 25
Note: Standard Error of a Regression Estimator

Standard error of the estimator β̂i1 , for population is

∑n
2 ϵ2
σβ̂ = ∑N t=1 it (29)
i1
N t=1 (Rmt − R̄m )2

Standard error of the estimator β̂i1 , for finite sample N is

∑n 2
2 t=1 ϵit
σβ̂ = ∑ (30)
i1
(N − K ) N t=1 (Rmt − R̄m )
2

where K is the number of parameters of the model (α, β) or N − K is

the degree of freedom of the model.

Single-Index Model 23 / 25
Accuracy of Adjusted βs

One way to see which approach work better is to compare the

predicted correlation of each approach.
Recall:
σij 2
βi βj σm
ρij = = (31)
σi σj σi σj

Klemkosky and Martin (1975) shows that both adjustment techniques

work better than the basic non-adjusted approach.
In addition, Elton, Gruber, and Urich (1978) show that Bayesian
technique predicts the correlation structure the best.

Single-Index Model 24 / 25
Fundamental βs

Another way to predict β is to use fundamentals as independent

variables or predicting variables.

βi = a0 + a1 X1 + a2 X2 + . . . + aN XN + ϵi (32)

where each Xi is one of the N variables hypothesized as affecting

Beta.
Popular fundamentals are
1 dividend,
2 asset growth,
3 leverage (senior securities divided by total assets),
4 liquidity (current assets divided current liabilities),
5 asset size,
6 earning variability.

Single-Index Model 25 / 25