CH 6 Single Index
CH 6 Single Index
1
market risk, cannot be eliminated in a well diversi&ed portfolio. The random error
term εit has a similar interpretation as the error term in the CER model. In the single
index model, εit represents random news that arrives between time t − 1 and t that
captures micro or &rm-speci&c risk factors that affect an individual asset s return
that are not related to macro events. For example, εit may capture the news effects
of new product discoveries or the death of a CEO. This type of risk is often called
&rm speci&c risk, idiosyncratic risk, residual risk or non-market risk. This type of
risk can be eliminated in a well diversi&ed portfolio.
The single index model can be expanded to capture multiple factors. The single
index model then takes the form a k−variable linear regression model
where Fjt denotes the j th systematic factorm, β i,j denotes asset i0 s loading on the j th
factor and εit denotes the random component independent of all of the systematic
factors. The single index model results when F1t = RMt and β i,2 = · · · = β i,k = 0. In
the literature on multiple factor models the factors are usually variables that capture
speci&c characteristics of the economy that are thought to affect returns - e.g. the
market index, GDP growth, unexpected in! ation etc., and &rm speci&c or industry
speci&c characteristics - &rm size, liquidity, industry concentration etc. Multiple
factor models will be discussed in chapter xxx.
The single index model is heavily used in empirical &nance. It is used to estimate
expected returns, variances and covariances that are needed to implement portfolio
theory. It is used as a model to explain the normal or usual rate of return on an
asset for use in so-called event studies2 . Finally, the single index model is often used
the evaluate the performance of mutual fund and pension fund managers.
2
5. cov(εit , εjs ) = 0 for all t, s and i 6= j
6. εit is normally distributed
The normality assumption is justi&ed on the observation that returns are fairly
well characterized by the normal distribution. The error term having mean zero
implies that &rm speci&c news is, on average, neutral and the constant variance
assumptions implies that the magnitude of typical news events is constant over time.
Assumption 4 states that &rm speci&c news is independent (since the random variables
are normally distributed) of macro news and assumption 5 states that news affecting
asset i in time t is independent of news affecting asset j in time s.
That εit is unrelated to RMs and εjs implies that any correlation between asset i
and asset j is solely due to their common exposure to RMt throught the values of β i
and β j .
The proofs of these results are straightforward and utilize the properties of linear
combinations of random variables. Results 1 and 4 are trivial. For 2, note that
var(Rit ) = var(αi + β i,M RMt + εit )
= β 2i,M var(RMt ) + var(εit ) + 2cov(RMt , εit )
= β 2i,M σ 2M + σ 2ε,i
since, by assumption 4, cov(εit , RMt ) = 0. For 3, by the additivity property of
covariance and assumptions 4 and 5 we have
cov(Rit , Rjt ) = cov(αi + β i,M RMt + εit , αj + β j,M RMt + εjt )
= cov(β i,M RMt + εit , β j,M RMt + εjt )
= cov(β i,M RMt , β j,M RMt ) + cov(β i,M RMt , εjt ) + cov(εit , β j,M RMt ) + cov(εit , εjt )
= β i,M β j,M cov(RMt , RMt ) = β i,M β j,M σ 2M
3
Last, for 5 note that
Remarks:
• σ ij = 0 if β i = 0 or β j = 0 or both
• σ ij > 0 if β i and β j are of the same sign
• σ ij < 0 if β i and β j are of opposite signs.
4. The expression for the expected return can be used to provide an unconditional
interpretation of αi . Subtracting β i,M µM from both sides of the expression for
µi gives
αi = µi − β i,M µM .
4
1.1.2 Decomposing Total Risk
The independence assumption between RMt and εit allows the unconditional vari-
ability of Rit , var(Rit ) = σ 2i , to be decomposed into the variability due to the market
index, β 2i,M σ 2M , plus the variability of the &rm speci&c component, σ 2ε,i . This decom-
position is often called analysis of variance (ANOVA). Given the ANOVA, it is useful
to de&ne the proportion of the variability of asset i that is due to the market index
and the proportion that is unrelated to the index. To determine these proportions,
divide both sides of σ 2i = β 2i,M σ 2M + σ 2ε,i to give
5
Property 1 states that the expected return on asset i conditional on RMt = rMt
is allowed to vary with the level of the market index. Property 2 says conditional
on the value of the market index, the variance of the return on asset is equal to the
variance of the random news component. Property 3 shows that once movements in
the market are controlled for, assets are uncorrelated.
Σ = σ 2M ββ 0 + ∆
Consider forming a portfolio of these two assets. Let x1 denote the share of wealth
in asset 1, x2 the share of wealth in asset 2 and suppose that x1 + x2 = 1. The return
6
on this portfolio using (2) and (3) is then
Example 2 To be completed
The additivity result of the single index model above holds for portfolios of any
size. To illustrate, suppose the single index model holds for a collection of N assets:
Consider forming a portfolio of these N assets. Let xi denote the share of wealth
P
invested in asset i and assume that Ni=1 = 1. Then the return on the portfolio is
N
X
Rpt = xi (αi + β i,M RMt + εit )
i=1
N
ÃN ! N
X X X
= xi αi + xi β i,M RMt + xi εit
i=1 i=1 i=1
= αp + β p RMt + εpt
PN ³P ´ PN
N
where αp = i=1 xi α i , β p = i=1 xi β i,M and εpt = i=1 xi εit .
7
the portfolio cannot be completely diversi&ed away. The so-called beta of an asset
captures this covariance contribution and so is a measure of the contribution of the
asset to overall portfolio variability.
To illustrate, consider an equally weighted portfolio of 99 stocks and let R99 denote
the return on this portfolio and σ 299 denote the variance. Now consider adding one
stock, say IBM, to the portfolio. Let RIBM and σ 2IBM denote the return and variance
of IBM and let σ 99,IBM = cov(R99 , RIBM ). What is the contribution of IBM to the
risk, as measured by portfolio variance, of the portfolio? Will the addition of IBM
make the portfolio riskier (increase portfolio variance)? Less risky (decrease portfolio
variance)? Or have no effect (not change portfolio variance)? To answer this question,
consider a new equally weighted portfolio of 100 stocks constructed as
R100 = (0.99) · R99 + (0.01) · RIBM .
The variance of this portfolio is
σ 2100 = var(R100 ) = (0.99)2 σ 299 + (0.01)2 σ 2IBM + 2(0.99)(0.01)σ 99,IBM
= (0.98)σ 299 + (0.0001)σ 2IBM + (0.02)σ 99,IBM
≈ (0.98)σ 299 + (0.02)σ 99,IBM .
Now if
• σ 2100 = σ 299 then adding IBM does not change the variability of the portfolio;
• σ 2100 > σ 299 then adding IBM increases the variability of the portfolio;
• σ 2100 < σ 299 then adding IBM decreases the variability of the portfolio.
Consider the &rst case where σ 2100 = σ 299 . This implies (approximately) that
(0.98)σ 299 + (0.02)σ 99,IBM = σ 299
which upon rearranging gives the condition
σ 99,IBM cov(R99 , RIBM )
= =1
σ 299 var(R99 )
De&ning
cov(R99 , RIBM )
β 99,IBM =
var(R99 )
then adding IBM does not change the variability of the portfolio as long as β 99,IBM =
1. Similarly, it is easy to see that σ 2100 > σ 299 implies that β 99,IBM > 1 and σ 2100 < σ 299
implies that β 99,IBM < 1.
In general, let Rp denote the return on a large diversi&ed portfolio and let Ri
denote the return on some asset i. Then
cov(Rp , Ri )
β p,i =
var(Rp )
measures the contribution of asset i to the overall risk of the portfolio.
8
2.1 The single index model and Portfolio Theory
To be completed
rbit = α b
bi + β i,M rMt , t = 1, . . . , T
b
εit = rit − rbit = rit − α b
bi − β i,M rMt , t = 1, . . . , T
Now some lines will &t better for some observations and some lines will &t better for
others. The least squares regression line is the one that minimizes the error sum of
squares (ESS)
T
X T
X
b
b i, β
SSR(α εb2it = b
bi − β
(rit − α 2
i,M ) = i,M rMt )
t=1 t=1
9
These are two linear equations in two unknowns and by straightforward substitution
the solution is
b
b i = r̄i − β
α i,M r̄M
PT
t=1 (rit − r̄i )(rMt − r̄M )
βb i,M = PT 2
t=1 (rMt − r̄M )
where
T T
1X 1X
r̄i = rit , r̄M = rMt .
T t=1 T t=1
The equation for βb i,M can be rewritten slightly to show that βb i,M is a simple
function of variances and covariances. Divide the numerator and denominator of the
expression for βb i,M by T −1
1
to give
1 PT
t=1 (rit − r̄i )(rMt − r̄M ) d it , RMt )
cov(R
βb i,M = T −1
1 PT =
t=1 (rMt − r̄M )
2 d Mt )
var(R
T −1
which shows that βb i,M is the ratio of the estimated covariance between Rit and RMt
to the estimated variance of RMt .
The least squares estimate of σ 2ε,i = var(εit ) is given by
T T
1 X 1 X b
σb 2ε,i = eb2it = bi − β
(rt − α i,M rMt )
2
T − 2 t=1 T − 2 t=1
The divisor T − 2 is used to make σb 2ε,i an unbiased estimator of σ 2ε,ι .
The least squares estimate of R2 is given by
2
b 2 βb i,M σb 2M σb 2ε,i
Ri = = 1− ,
d it )
var(R d it )
var(R
where
T
1 X
d it ) =
var(R (rit − r̄i )2 ,
T − 1 t=1
and gives a measure of the goodness of &t of the regression equation. Notice that
b 2 = 1 whenever σ
R b 2ε,i = 0 which occurs when bεit = 0 for all values of t. In other
i
b 2
words, Ri = 1 whenever the regression line has a perfect &t. Conversely, R b2 = 0
i
2
when σb ε,i = var(R
d it ); that is, when the market does not explain any of the variability
of Rit . In this case, the regression has the worst possible &t.
10
3.2 Testing the Restriction α = 0.
Using the single index model regression,
Rt = α + βRMt + εt , t = 1, ..., T
εt ∼ iid N(0, σ 2ε ), εt is independent of RMt (4)
consider testing the null or maintained hypothesis α = 0 against the alternative that
α 6= 0
H0 : α = 0 vs. H1 : α 6= 0.
If H0 is true then the single index model regression becomes
Rt = βRMt + εt
and E[Rt |RMt = rMt ] = βrMt . We will reject the null hypothesis, H0 : α = 0, if
the estimated value of α is either much larger than zero or much smaller than zero.
Assuming H0 : α = 0 is true, α̂ ∼ N (0, SE(α̂)2 ) and so is fairly unlikely that α̂ will
be more than 2 values of SE(α̂) from zero. To determine how big the estimated value
of α needs to be in order to reject the null hypothesis we use the t-statistic
αb −0
tα=0 = d ,
SE(αb)
where α d α)
b is the least squares estimate of α and SE( b is its estimated standard error.
The value of the t-statistic, tα=0 , gives the number of estimated standard errors that
b is from zero. If the absolute value of tα=0 is much larger than 2 then the data cast
α
considerable doubt on the null hypothesis α = 0 whereas if it is less than 2 the data
are in support of the null hypothesis3 . To determine how big | tα=0 | needs to be to
reject the null, we use the fact that under the statistical assumptions of the single
index model and assuming the null hypothesis is true
If we set the signi&cance level (the probability that we reject the null given that the
null is true) of our test at, say, 5% then our decision rule is
11
Consider the estimated MM regression equation for IBM using monthly data from
January 1978 through December 1982:
b 2
b ε = 0.0524
R IBM,t =−0.0002 + 0.3390 ·RMt , R = 0.20, σ
(0.0068) (0.0888)
where the estimated standard errors are in parentheses. Here α b = −0.0002, which is
d
very close to zero, and the estimated standard error, SE(α̂) = 0.0068, is much larger
than αb . The t-statistic for testing H0 : α = 0 vs. H1 : α 6= 0 is
−0.0002 − 0
tα=0 = = −0.0363
0.0068
so that α b is only 0.0363 estimated standard errors from zero. Using a 5% signi&cance
level, |t58 (0.025)| ≈ 2 and
|tα=0 | = 0.0363 < 2
so we do not reject H0 : α = 0 at the 5% level.
H0 : β = 1 vs. H1 : β 6= 1.
The data cast doubt on this hypothesis if the estimated value of β is much different
from one. This hypothesis can be tested using the t-statistic
βb − 1
tβ=1 = d b
SE(β)
which measures how many estimated standard errors the least squares estimate of β
is from one. The null hypothesis is reject at the 5% level, say, if |tβ=1 | > |tT −2 (0.025)|.
Notice that this is a two-sided test.
Alternatively, one might want to test the hypothesis that the risk of an asset is
strictly less than the risk of the market index against the alternative that the risk is
greater than or equal to that of the market:
H0 : β = 1 vs. H1 : β ≥ 1.
Notice that this is a one-sided test. We will reject the null hypothesis only if the
estimated value of β much greater than one. The t-statistic for testing this null
12
hypothesis is the same as before but the decision rule is different. Now we reject the
null at the 5% level if
tβ=1 < −tT −2 (0.05)
where tT −2 (0.05) is the one-sided 5% critical value of the Student-t distribution with
T − 2 degrees of freedom.
13
Monthly Returns on IBM Monthly Returns on Market Index
0.2 0.2
0.1 0.1
0.0 0.0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
78 79 80 81 82 83 84 85 86 87 78 79 80 81 82 83 84 85 86 87
IBM MARKET
Figure 1
Notice that the IBM and the market index have similar behavior over the sample
with the market index looking a little more volatile than IBM. Both returns dropped
sharply during the October 1987 crash but there were a few times that the market
dropped sharply whereas IBM did not. Sample descriptive statistics for the returns
are displayed in &gure 2.
The mean monthly returns on IBM and the market index are 0.9617% and 1.3992%
per month and the sample standard deviations are 5.9024% and 6.8353% per month,
respectively.. Hence the market index on average had a higher monthly return and
more volatility than IBM.
14
Monthly Returns on IBM Monthly Returns on Market Index
12 30
10 25
8 20
6 15
4 10
2 5
0 0
-0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 -0.2 -0.1 0.0 0.1
Figure 2
Notice that the histogram of returns on the market are heavily skewed left whereas
the histogram for IBM is much more sysingle index modeletric about the mean. Also,
the kurtosis for the market is much larger than 3 (the value for normally distributed
returns) and the kurtosis for IBM is just slightly larger than 3. The negative skewness
and large kurtosis of the market returns is caused by several large negative returns.
The Jarque-Bera statistic for the market returns is 67.97, with a p-value 0.0000, and
so we can easily reject the hypothesis that the market data are normally distributed.
However, the Jarque-Bera statistic for IBM is only 0.1068, with a p-value of 0.9479,
and we therefore cannot reject the hypothesis of normality.
The single index model regression is
Rt = α + βRMt + εt , t = 1, . . . , T
15
Figure 3
where the estimated standard errors are reported underneath the estimated coeffi-
cients. The estimated intercept is close to zero at 0.0053, with a standard error of
d α)),
0.0069 (= SE( b and the estimated value of β is 0.3278, with an standard error of
d b
0.0890 (= SE(β)). Notice that the estimated standard error of βb is much smaller
than the estimated coefficient and indicates that β is estimated reasonably precisely.
The estimated regression equation is displayed graphically in &gure 4 below.
16
Market Model Regression
0.2
0.1
IBM
0.0
-0.1
-0.2
-0.3 -0.2 -0.1 0.0 0.1 0.2
MARKET
Figure 4
To evaluate the overall &t of the single index model regression we look at the R2 of
the regression, which measures the percentage of variability of Rt that is attributable
to the variability in RMt , and the estimated standard deviation of the residuals, σb ε .
From the table, R2 = 0.190 so the market index explains only 19% of the variability
of IBM and 81% of the variability is not explained by the market. In the single index
model regression, we can also interpret R2 as the proportion of market risk in IBM
and 1 − R2 as the proportion of &rm speci&c risk. The standard error (S.E.) of the
regression is the square root of the least squares estimate of σ 2ε = var(εt ). From the
above table, σb ε = 0.052. Recall, εt captures the &rm speci&c risk of IBM and so σb ε is
an estimate of the typical magnitude of the &rm speci&c risk. In order to interpret the
magnitude of σb ε it is useful to compare it to the estimate of the standard deviation
of Rt , which measures the total risk of IBM. This is reported in the table by the
standard deviation (S.D.) of the dependent variable which equals 0.057. Notice that
σb ε = 0.052 is only slightly smaller than 0.057 so that the &rm speci&c risk is a large
proportion of total risk (which is also reported by 1 − R2 ).
Con&dence intervals for the regression parameters are easily computed using the
reported regression output. Since εt is assumed to be normally distributed 95%
con&dence intervals for α and β take the form
α d α
b ± 2 · SE( b)
b d b
β ± 2 · SE(β)
17
The 95% con&dence intervals are then
Our best guess of α is 0.0053 but we wouldn t be too surprised if it was as low as
-0.0085 or as high as 0.0191. Notice that there are both positive and negative values
in the con&dence interval. Similarly, our best guess of β is 0.3278 but it could be as
low as 0.1498 or as high as 0.5058. This is a fairly wide range given the interpretation
of β as a risk measure. The interpretation of these intervals are as follows. In
repeated samples, 95% of the time the estimated con&dence intervals will cover the
true parameter values.
The t-statistic given in the computer output is calculated as
estimated coefficient − 0
t-statistic =
std. error
and it measures how many estimated standard errors the estimated coefficient is away
from zero. This t-statistic is often referred to as a basic signi&cance test because it
tests the null hypothesis that the value of the true coefficient is zero. If an estimate is
several standard errors from zero, so that it s t-statistic is greater than 2, then it is a
good bet that the true coefficient is not equal to zero. From the data in the table, the
b = 0.0053 is 0.767 standard errors from zero. Hence
t-statistic for α is 0.767 so that α
it is quite likely that the true value of α equals zero. The t-statistic for β is 3.684,
βb is more than 3 standard errors from zero, and so it is very unlikely that β = 0.
The Prob Value (p-value of the t-statistic) in the table gives the likelihood (computed
from the Student-t curve) that, given the true value of the coefficient is zero, the data
would generate the observed value of the t-statistic. The p-value for the t-statistic
testing α = 0 is 0.4465 so that it is quite likely that α = 0. Alternatively, the p-value
for the t-statistic testing β = 0 is 0.001 so it is very unlikely that β = 0.
18
Market Model Regression for IBM
0.2
0.1
0.0
0.15
0.10 -0.1
0.05
-0.2
0.00
-0.05
-0.10
-0.15
1978 1979 1980 1981 1982
Figure 5
Notice that the &tted values do not track the actual values very closely and that
the residuals are fairly large. This is due to low R2 of the regression. The residuals
appear to be fairly random by sight. We will develop explicit tests for randomness
later on. The histogram of the residuals, displayed below, can be used to investigate
the normality assumption. As a result of the least squares algorithm the residuals
have mean zero as long as a constant is included in the regression. The standard
deviation of the residuals is essentially equal to the standard error of the regression
- the difference being due to the fact that the formula for the standard error of the
regression uses T − 2 as a divisor for the error sum of squares and the standard
deviation of the residuals uses the divisor T − 1.
19
Residuals from Market Model Regression for IBM
8
Series: Residuals
Sample 1978:01 1982:12
Observations 60
6
Mean -2.31E-19
Median -0.000553
Maximum 0.139584
4
Minimum -0.104026
Std. Dev. 0.051567
Skewness 0.493494
2 Kurtosis 2.821260
Jarque-Bera 2.515234
Probability 0.284331
0
-0.10 -0.05 0.00 0.05 0.10
Figure 6
The skewness of the residuals is slightly positive and the kurtosis is a little less
than 3. The hypothesis that the residuals are normally distributed can be tested
using the Jarque-Bera statistic. This statistic is a function of the estimated skewness
and kurtosis and is give by
à !
T c − 3)2
(K
JB = Sb2 +
6 4
where Sb denotes the estimated skewness and K c denotes the estimated kurtosis. If
the residuals are normally distribued then Sb ≈ 0 and Kc ≈ 3 and JB ≈ 0. Therefore,
b c
if S is moderately different from zero or K is much different from 3 then JB will get
large and suggest that the data are not normally distributed. To determine how large
JB needs to be to be able to reject the normality assumption we use the result that
under the maintained hypothesis that the residuals are normally distributed JB has
a chi-square distribution with 2 degrees of freedom:
JB ∼ χ22 .
For a test with signi&cance level 5%, the 5% right tail critical value of the chi-square
distribution with 2 degrees of freedom, χ22 (0.05), is 5.99 so we would reject the null
that the residuals are normally distributed if JB > 5.99. The Probability (p-value)
reported by Eviews is the probability that a chi-square random variable with 2 degrees
of freedom is greater than the observed value of JB :
P (χ22 ≥ JB) = 0.2843.
For the IBM residuals this p-value is reasonably large and so there is not much data
evidence against the normality assumption. If the p-value was very small, e.g., 0.05 or
smaller, then the data would suggest that the residuals are not normally distributed.
20