GRS Review
Karl Diether
University of Chicago
Graduate School of Business
Nov 14, 2001
Empirical Implications of the ICAPM
Consider the regressions,
L
Rit R f t = i + i j (Fjt ) + it , i = 1. . .N,
j=1
where,
L 1 is number of priced state variables.
Fj is the excess return (or return from a zero-cost portfolio) on the
jth factor portfolio.
Ri R f is the excess return on a security or portfolio.
2
Empirical Implications of the ICAPM
If each factor portfolio is multifactor minimum variance in a S state
variable world then,
1. i = 0 i
2. Some linear combination of the factor portfolios is on the minimum
variance boundary.
There is actually only one implication because i = 0 i if and only if
some portfolio of the right-hand side portfolios is on the minimum
variance boundary.
The GRS Test
1. the GRS test is a statistical test of the hypothesis that i = 0 i.
2. Equivalently, it is a test that some linear combination of the factor
portfolios is on the minimum variance boundary.
3. Equivalently, it is also a test that each factor portfolio is multifactor
minimum variance in a S state variable world.
4
GRS Test, Sharpe Ratio Representation
q 2
T T N L 1 + 2
N+L
q 1 F(N, T N L)
N T L1 1 + L2
where,
1. N+L is the ex post maximum Sharpe ratio of the N test assets and the
L factor portfolios.
2. L is the ex post maximum Sharpe ratio of the L factor portfolios.
Thus the GRS statistic determines whether |N+L | is statistically greater
than |L |.
Graphical Representation
Mean Variance Frontier
L+N
L
E(R)
RF
6
GRS Test, = 0 Representation
0 1
T T N L
F(N, T N L)
N T L1 1 + 0 1
where,
1. is a N 1 vector of estimated intercepts.
2. is an unbiased estimate of the residual covariance matrix.
3. is a L 1 vector of the factor portfolios sample means.
4. is an unbiased estimate of the factor portfolios covariance matrix.
if i = 0 i, then the GRS statistic equals zero; the larger the s are in
absolute value the greater the GRS statistic will be.
Computing The GRS Statistic
The problem set will require you to compute a few GRS statistics. You
can compute it using any software package you wish. The following
software packages are commonly used:
Excel
A Matrix Language (i.e. MATLAB, Gauss, or Ox)
It is easier to compute a GRS statistic using a matrix language than it is
using Excel. However, it is probably not worth your time to learn a matrix
language just for this problem set.
8
GRS Recipe
Step 1: Time series regressions
Run the following regression for all N left hand side portfolios:
L
Rit R f t = i + i j (Fjt ) + it
j=1
If you are testing the Fama French three factor model then you would run,
Rit R f t = i + bi (RMt R f t ) + si SMBt + hi HMLt + it
GRS Recipe
Step 2: The intercept vector
Form the estimated intercepts into a N 1 vector ().
1
2
=
..
.
N
10
GRS Recipe
Step 3: The residual matrix
Calculate the residual for each regression,
L
it = (Rit R f t ) i i j (Fjt ) .
j=1
For example, if you are testing the Fama French three factor model then
you would compute,
it = (Rit R f t ) i bi (RMt R f t ) si SMBt hi HMLt .
11
GRS Recipe
Step 3: The residual matrix continued
Form the residuals into a T N matrix (note, that now the first subscript
refers to time period and the second refers to the test portfolio),
12 1N
11
21 22 2N
=
.. .. .. ..
. . . .
T 1 T 2 T N
12
GRS Recipe
Step 4: = cov()
Compute an unbiased estimate of the covariance matrix of residuals,
0
=
T L1
is a N N matrix. For example, if there are ten test portfolios, then
is 10 10 matrix.
In Excel you do not have to explicitly compute using the above
formula; you can just use the COVAR command. However, Excels
estimate is not unbiased. You need to multiply Excels estimate by
T /(T L 1).
13
GRS Recipe
Step 5: Factor mean vector
Calculate the sample means of the factor portfolios and form a L 1
vector () of sample means.
F1
F2
=
..
.
FL
14
GRS Recipe
Step 6: The factor matrix
Form the factor portfolio (excess) returns into a T L matrix (note, that
now the 1st subscript refers to time period and the 2nd refers to the factor
portfolio),
F F12 F1L
11
F21 F22 F2L
F = .. .. .. ..
. . . .
FT 1 FT 2 FT L
15
GRS Recipe
Step 7: Compute an unbiased estimate of the covariance matrix of the
factors (the dimension of the covariance matrix is L L),
(F F)0 (F F)
=
T 1
where,
F1 F2 FL
F1 F2 FL
F =
.. .. .. ..
. . . .
F1 F2 FL
Note: In Excel you do not have to explicitly compute using the above
formula; you can just use the COVAR command. However, Excels
estimate is not unbiased. You need to multiply Excels estimate by
T /(T 1).
16
GRS Recipe
Step 8: Compute the GRS statistic,
0 1
T T N L
F(N, T N L)
N T L1 1 + 0 1
Note: Both 0 1 and 0 1 are scalars. If you do not get scalars, then
you have done something wrong.
Step 9: Find the p-value of the GRS statistic. You can find the p-value in
Excel using the FDIST command.
17
