2024年5月5日发(作者:)
A Test of the Efficiency of a Given Portfolio
Author(s): Michael R. Gibbons, Stephen A. Ross, Jay Shanken
Source:
Econometrica,
Vol. 57, No. 5 (Sep., 1989), pp. 1121-1152
Published by: The Econometric Society
Stable URL: /stable/1913625
Accessed: 14/01/2010 08:12
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
/page/info/about/policies/. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
/action/showPublisher?publisherCode=econosoc.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@.
The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.
Econometrica, Vol. 57, No. 5 (September, 1989), 1121-1152
A TEST OF THE EFFICIENCY OF A GIVEN PORTFOLIO
BY
MICHAEL
R.
GIBBONS, STEPHEN
A. Ross,
AND JAY SHANKEN1
A test for the ex ante efficiency of a given portfolio of assets is analyzed. The relevant
statistic has a tractable small sample distribution. Its power function is derived and used to
study the sensitivity of the test to the portfolio choice and to the number of assets used to
determine the ex post mean-variance efficient frontier.
Several intuitive interpretations of the test are provided, including a simple mean-stan-
dard deviation geometric explanation. A univariate
test, equivalent to our multivariate-based
method, is derived, and it suggests some useful diagnostic tools which may explain why the
null hypothesis is rejected.
Empirical examples suggest that the multivariate approach can lead to more appropriate
conclusions than those based on traditional inference which relies on a set of dependent
univariate statistics.
KEYWORDS:
Asset pricing, CAPM, multivariate test, portfolio efficiency.
1. INTRODUCTION
The modern theory of finance has always been rooted in empirical analysis.
The mean-variance capital asset pricing model (CAPM) developed by Sharpe
(1964) and Lintner (1965) has been studied and tested in more papers than can
possibly be attributed here. This is only natural; the quality and quantity of
financial data, especially stock market price series, are the envy of other fields in
economics.
The theory is generally expressed in terms of its first-order conditions on the
risk premium. Expected returns on assets are linearly related to the regression
coefficients, or betas, of the asset returns on some index of market returns. In
other words, risk premiums in equilibrium
depend on betas. The standard tests of
the CAPM are based on regression
techniques with various adaptations. For
some notable examples, see Black,
Jensen, and Scholes (1972) and Fama and
MacBeth (1973). Usually, cross-sectional regressions are run of asset returns on
estimated beta
coefficients, and estimates of the slope are reported. Often the
data are grouped to reduce measurement
errors, and sometimes the estimation is
done at a
sequence
of time
points to create a time series of estimates from which
the precision of the overall
average can be determined.
Roll (1977, 1978), among others, has raised serious doubts whether these
procedures are, in fact, tests of the CAPM. Insofar as proxies are used for the
market
portfolio, the Sharpe-Lintner
theory is not being tested. Furthermore, as
Roll
emphasizes, the regression tests are probably of quite low power, and
We are
grateful
to Ted
Anderson,
Fischer
Black, Douglas Breeden,
Michael
Brennan, Gary
Chamberlain,
Dave
Jobson,
Allan
Kleidon,
Bruce
Lehmann,
Paul
Pfleiderer,
Richard
Roll,
and two
anonymous
referees as well as the seminar
participants
at Duke
University,
Harvard
University,
Indiana University, Stanford University, University of California at San Diego, University
of Illinois
at Urbana, and Yale
University
for
helpful
comments. We
appreciate
the research assistance of
Ajay
Dravid, Jung-Jin Lee,
and
Tong-sheng Sun.
Financial
support
was
provided
in
part by
the National
Science Foundation and the Stanford
Program
in Finance. This
paper supersedes
an earlier
paper
with the same title by Stephen Ross.
1121
1
1122
MICHAEL R. GIBBONS, STEPHEN A. ROSS,
AND JAY SHANKEN
These objections
leave the empirical
grouping may lower the power
further.
testing of the CAPM
in an odd state of limbo. If the proxy is not a
valid
are
investigations
then as tests of the CAPM the existing
empirical
surrogate,
beside the point.2
On the other hand, if the proxy
is valid, then
the
somewhat
of the tests are
unknown.
small sample
distribution and power
is one
opportunity. The CAPM
This is unfortunate and
indicative of a missed
couched
in
which suggest
quite specific
hypotheses
of many financial
theories
these hypotheses are an
terms of observables. The rich
data available for testing
we
at them.
In this paper
are
explicitly
directed
tests which
incentive
to develop
a test using multivariate statistical methods.
example of such
develop
a canonical
in tests of the CAPM.
one addressed
is the central
The problem
we consider
Since the theory is equivalent
to the assertion that the market portfolio
is
portfolio
is
mean-variance
efficient,
we wish to test whether any particular
efficient.
ex ante mean-variance
into seven sections, it also can be
viewed as
While the paper is organized
consisting of three parts.
The first part (Sections 2 through 4) considers
a
the prop-
efficiency and examines
for testing mean-variance
multivariate statistic
erties of such a test. The second part (Sections
5 and 6) studies the relation
based on a set of
approaches
between this multivariate test and alternative
the paper by
7 and 8) concludes
statistics. The third
part (Sections
univariate
for
and providing suggestions
to related
hypotheses
extending the framework
summary of each
section
follows.
detailed
future research. A more
portfolio.
for the efficiency of some
In Section 2 we recall a necessary condition
as a null hypothesis that can be tested using
a statistic
We use this implication
both the null and alternate
distribution under
sample
which
has a tractable finite
we relate
this statistic
to three alternative approaches
hypotheses.
In addition,
In the third section the multi-
which are based on asymptotic approximations.
deviation
interpretation in the mean-standard
variate test is given a geometric
to a data
are then applied
The method and geometry
space of portfolio theory.
we reaffirm and
in modern finance;
empirical papers
set from one of the classic
The fourth
section
and Scholes (1972).
Jensen,
the findings of Black,
complement
the sensitivity
of the test. Here
we consider
to the power
turns to issues relating
which is examined for efficiency and the
of the test to the choice
of the portfolio
frontier. A
of assets used to determine the ex post
efficient
effect of the number
that one's
new data base is analyzed
in this section, and we demonstrate
index can be altered by the type
of
the efficiency of a given
conclusions
regarding
assets used to construct the
ex
post
frontier.
when the multi-
results
actual
to contrast
empirical
The fifth section attempts
based on a set of dependent
inference
informal
variate method
is used versus
the multivariate test
rejects
where
we
provide
Here
examples
univariate statistics.
We also have
seem
to be
significant.
none of the univariate statistics
even though
tests of the
do consider
and Stambaugh (1987)
and Shanken (1987b)
2Recent
work by Kandel
and the true
market
the proxy
on an assumption about
the correlation between
CAPM
conditional
portfolio.
PORTFOLIO
EFFICIENCY
1123
the reverse situation where there are a seemingly large number of "significant"
univariate statistics; yet, the multivariate test fails to reject at the traditional
levels of significance. In this section we also introduce another data set which
allows us to re-examine the size-effect anomaly. Section 6 develops an alternative
interpretation of the multivariate test. The statistic is equivalent to the usual
calculation for a t statistic on an intercept term in a univariate
simple regression
model, with the ex post efficient portfolio used as the dependent variable and the
portfolio whose ex ante efficiency is under examination as the explanatory
variable. This section also develops some useful diagnostics for explaining why
the null hypothesis may not be consistent with the data. Most of the
empirical
work in this section focuses on the size effect only in the month of January.
Section 7 extends the analysis to a case where one wishes to
investigate the
potential efficiency of some linear combination of a set of portfolios, where the
weights in the combination are not specified. This turns out to be a minor
adaptation of the work
in Section 2.
2. TEST STATISTIC FOR JUDGING
THE EFFICIENCY
OF A GIVEN PORTFOLIO
We assume throughout that there is a given riskless rate of interest,
Rft,
for
each time period. Excess returns are computed
by subtracting
Rft
from the total
rates of return. Consider the following multivariate linear regression:
(1)
+
flipppt
+
iit
'it=aip
Vi
=
1,. .., N,
where ri
the excess return on asset i in period t;
Fpp-
the excess return on the
portfolio whose efficiency is being tested; and
9it-
the disturbance term for asset
i in period t. The disturbances are assumed to be jointly
normally distributed
each period with mean zero and
nonsingular covariance matrix 2, conditional on
the excess returns for portfolio
p. We also assume independence of the distur-
bances over time. In order that
2
be nonsingular, Fp and the N left-hand side
assets must be linearly
independent.
If a particular
portfolio is mean-variance efficient (i.e., it minimizes variance
for a given level of expected
return), then the following first-order
condition must
be satisfied for the given N
assets:
(2)
(Fit)
=
fiPeQPt).
Thus, combining the first-order
condition in (2) with the distributional assump-
tion
given by (1) yields
the
following parameter restriction,
which is stated in the
form of a null
hypothesis:
(3)
Ho:
aip
=
0
li
=19,...,9N.
Testing the above null hypothesis is essentially the same proposal as in the
work by Black, Jensen, and Scholes
(1972), except that they replace
ip,
by a
portfolio
which
they call the market portfolio and refer to their test as a test of
the CAPM. In
addition, they do not report the joint significance of the estimated
1124
MICHAEL R. GIBBONS, STEPHEN
A. ROSS, AND JAY SHANKEN
values for
a?1p
across all N equations;
instead, they report N univariate
t
based on each
equation.
statistics
Given the
normality assumption,
the null
hypothesis
in
(3)
can be tested
using
"Hotelling's
T2 test,"
a multivariate
generalization of the univariate t-test
(e.g.,
see Malinvaud
(1980, page 230)). A brief derivation of the
equivalent
F test is
included for completeness and as a
means
of
introducing some
notation
that will
be needed
later. If we estimate
the multivariate
system of (1) using
ordinary least
for each individual
squares
equation, the estimated intercepts
have
a
multivariate
normal
distribution, conditional on
rp,
(Vt= 1,..., T), with
(4)
/T(+k)ap
-
NtT(
a
sT;
p)Op;
where T
number of time series
observations on returns;
P);
mean of r
and
sp
sample
variance of
rp,
without
an
p
-plsp;
rp-
sample
adjustment for degrees of freedom.
Furthermore,
independent with
ap
and T are
(T
-
2)2 having a Wishart
distribution with parameters (T
-
2) and 2. These
facts imply (see Morrison
(1976, page 131)) that (T(T
-
N
-
1)/N(T
-
2))WJK has
a noncentral F
distribution with degrees
N
and (T
-
N
-
1), where
of freedom
(5)
-
apE lp/(l
WU
+
2
and 2-
unbiased
residual covariance matrix.3
(The
corresponding
statistic
based
on the maximum
likelihood estimate of
2
will be denoted
as W.)
The
noncentral-
ity parameter,
by
A,
is given
X
[T/(1+2
(6)
)]a
-
lap.
Under the null
hypothesis that
ap
equals
F
zero,
A=
0, and we have
a central
distribution. More
generally,
the distribution under the alternative
provides
a
way to study
the power
of the
test;
more will be said
about this
in a
later
section.
It is also interesting to note
that
under
the null hypothesis
the
Wu
statistic has a
central F distribution
for
the
parameters
unconditionally,
of this central F do not
depend
on rin
any
way.
However,
we do not know
the unconditional distribu-
tionofn
or
Wu
under the alternate, for the conditional
distribution depends on
the sample
values
of -p,
through 0.
has been viewed as
providing
Generally,
the
normality
assumption a
"good
working
approximation" to the distribution of
monthly
stock
returns
(see
Fama
(1976, Chapter
1)
for a
summary of the relevant
There
is some
empinrcal
work).
that the true distributions are
slightly
evidence,
however,
to
relative
leptokurtic
the normal
distribution. While from
normality
of the disturbances in
departures
(1) will affect the
small-sample
distribution of the test
statistic,
simulation
evidence
by MacKinlay
that the F test is
fairly
robust
to such
(1985) suggests
of standard
misspecifications.4 This is important, since the
application
asymp-
totic tests to the
efficiency problem
can result in
faulty inferences,
given
the
sample
sizes often used in financial
empirical work.
Tests for normality of the
residuals of the size and
industry portfolios, which
are used
below,
do
reveal
excess
kurtosis and some These
skewness as well.
results are available on
request to the
authors.
that N is less than
or equal to T
-
2 so that I is
nonsingular.
4We
assume
PORTFOLIO EFFICIENCY
TABLE 1
1125
A
COMPARISON OF FOUR ASYMPTOTICALLY EQUIVALENT TESTS OF EX ANTE EFFICIENCY OF A
GIVEN PORTFOLIO.
THE
W STATISTIC IS DISTRIBUTED AS A TRANSFORM OF A CENTRAL
F
DISTRIBUTION
IN FINITE SAMPLES. THE WALD
TEST, THE LIKELIHOOD RATIO TEST
(LRT),
AND
THE LAGRANGE MULTIPLIER TEST
(LMT)
ARE MONOTONE TRANSFORMS OF
W,
AND
EACH IS DISTRIBUTED
AS CHI-SQUARE WITH
N
DEGREES OF FREEDOM AS
T
APPROACHES
INFINITY.
N
T
P-Value Using
Exact Distribution
of W
P-Values Using
Asymptotic Approximations
Wald LRT
LMT
10
20
40
58
10
20
40
58
118
10
20
40
58
118
238
10
20
40
58
10
20
40
58
118
10
20
40
58
118
238
60
60
60
60
120
120
120
120
120
240
240
240
240
240
240
60
60
60
60
120
120
120
120
120
240
240
240
240
240
240
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.05
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.10
.008
.000
.000
.000
.023
.005
.000
.000
.000
.035
.109
.003
.000
.000
.000
.025
.000
.000
.000
.056
.017
.000
.000
.000
.076
.048
.009
.001
.000
.000
.027
.007
.000
.000
.038
.023
.003
.000
.000
.044
.035
.017
.006
.000
.000
.061
.019
.000
.000
.081
.053
.010
.000
.000
.090
.075
.041
.018
.000
.000
.071
.094
.173
.403
.060
.070
.094
.122
.431
.055
.059
.069
.079
.123
.451
.122
.146
.216
.404
.111
.122
.147
.175
.432
.106
.111
.122
.133
.178
.452
Note: N is the number of assets used together with portfolio p to construct the ex
post frontier, and T is the
number of time series observations.
Table I illustrates
this problem
for the Wald, likelihood
ratio, and Lagrange
is asymptotically distributed as
chi-square with N
multiplier
tests, each
of which
degrees of freedom as T
--
oo.S Since the small-sample distribution of W is
known (assuming
realization of W can be inferred from
normality), the implied
the information in the first three columns of Table I
(i.e., N, T, and the
hypothetical p-value).
The implied
asymptotic p-values given in the last three
the
5
Jobson
three tests
using
a simulation. They approximate
and Korkie also discuss these
(1982)
In -their
on Rao's
(1951) work.
distribution of the likelihood ratio
test with an F distribution based
the null hypothesis.
is available under
distribution
sample
they recognize that
a small
1985 paper
1126
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
columns are then obtained using the fact that each test statistic is a monotonic
function of W.6
Consistent with the results of Berndt and Savin (1977), the p-values are always
lowest for the Wald test and highest for the Lagrange multiplier test with the
likelihood ratio test in between. Clearly, the asymptotic approximation becomes
worse as the number of assets, N, approaches the number of time series
observations, T. Shanken (1985) reaches similar conclusions based on an approxi-
mation when the riskless asset is not observable.
3. A GEOMETRIC INTERPRETATION OF THE TEST STATISTIC, W
So far, the primary motivation for the W statistic has been its well-known
distributional properties. For rigorous statistical inference such results are an
absolute necessity. Just as important, though, is the development of a measure
which allows one to examine the economic significance of departures from the
null hypothesis. Fortunately, our test has a nice geometric interpretation.
It is shown in the Appendix that:
W=
?*2
(7)
-12
=2_ 1
where
&*
is the ex
post price
of risk
(i.e.,
the maximum excess
sample
mean
return per unit of sample standard deviation) and
Op
is the ratio of ex post
p~~~~~
average excess return on portfolio p to its standard deviation (i.e., p
-p/sp).
Note that
4
cannot be less than one since 0* is the slope of the ex post frontier
based on all assets used in the test (including portfolio p).
The curve in Figure la represents the (ex post) minimum-variance frontier of
the risky assets. When a riskless investment is available, the frontier is a straight
line emanating from the origin and tangent to the curve at m. 0* is the slope of
the tangent line whereas
0p
is the slope of the line through p.
An examination of (7) suggests that
42
should be close to one under the null
hypothesis. When 0* is sufficiently greater than 0
,
the return per unit of risk for
portfolio p is much
lower than the ex
post
frontier
tradeoff,
and we will
reject
the
hypothesis that portfolio p is ex ante
mean-variance efficient. In
Figure
la
4
is
just the distance along the ex post frontier up
to
any given
risk
level, a,
divided
by the
similar distance
along
the line from the
origin through p.
The reader may wonder why the test is based on the square
of the
slopes
as
opposed to
the actual
slopes.
The reason is
straightforward.
Our null
hypothesis
only represents
a
necessary
condition for ex ante
efficiency.
This condition is
satisfied even if portfolio p is on the negative sloping portion
of the minimum-
variance frontier for all assets (including the risk-free security). Thus, only the
The relations are LRT= T ln(1 + W) and LMT= TW/(1 + W). Shanken (1985) has
discussed
this result for the case where the riskless asset does not exist. A proof of
the result in the case where
the riskless asset does exist is available upon request to
the authors. Bemdt and Savin (1977) discuss
similar relationships among alternative asymptotic tests
in a more general setting.
6
PORTFOLIO
EFFICIENCY
1127
d
p~~~~~~~~~~~
0
Standard
Deviation
of Excess
Return
X
la.)
Geometric
intuition for W. Note the drstance
Oc
is
2.4
2.2
2/
I
+ 02,
and the distance Od is
V1
+*
--
-
1.2]
4)
0.8-
0.6-]
0
1
2-
81
0.4
0.2-
0I
0
2
4
6
8
10
Standard
Deviation
of Excess
Return
lb.)
Ex post efficient frontier based on 10
beta-sorted
portfolios
and tihe CRSP
Equal-Weighted
Index
using monthly
data,
1931-1965.
Point p repre-
sents the CRSP
Equal-Weighted
Index.
FIGURE
1.-Various plots of ex post mean variance efficient frontiers.
1128
MICHAEL R.
GIBBONS,
STEPHEN A.
ROSS,
AND JAY SHANKEN
1.7
1.6
1.5
1.4
1.3
1.2
0.9
*i
7
/
/
0.7-
XOB
*
0.8
N
0.5
0.4-
0.3-
0
0
2
4
6
8
Standard
Deviation of Excess
Return
ic.) Ex post efficient
frontier based on 12
industry
portfolios and the CRSP
Value-Weighted
1n(lex
using monthly
data,
1926-1982. Point
p repre-
sents the CRSP
Value-Weighted
Index.
1.2-
1.1
0.9
0.8
*
x
0.?
0.6
0.5
0.4-
0.3-
0.2-
0.
0 2
4
6
x
/
Standard
Deviation of Excess
Return
Id.) Ex post efficient frontier based on 10 size-sorted
portfolios and the CRSP
Value-Weighted
Index
using monthly
data,
1926-1982. Point
p repre-
sents the CRSP
Value-Weighted
Index.
FIGURE
1.-Continued.
PORTFOLIO EFFICIENCY
TABLE II
1129
SUMMARY
STATISTICS ON BETA-SORTED PORTFOLIOS BASED ON MONTHLY DATA,
1931-65
(T=420).
ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE FORM, AND THE
INDEX IS PORTFOLIO
CRSP
EQUAL-WEIGHTED
PARAMETER ESTIMATES
p.
THE FOLLOWING
10
AND
Vt =1.
ARE FOR THE REGRESSION MODEL:
Pit
=
alp
+
Pip Fpt
+
9,t
Vi
=
1.
420,
WHERE R2 IS THE COEFFICIENT OF DETERMINATION FOR EQUATION i.
Portfolio
Number
&,P
S(&,P)
/3P
s(Ap)
R2
0.94
0.97
0.98
0.98
0.98
0.97
0.97
0.96
0.92
0.85
1
2
3
4
5
6
7
8
9
10
-0.19
-0.19
-0.06
-0.09
-0.06
0.05
0.03
0.12
0.14
0.22
0.17
0.10
0.09
0.07
0.07
0.07
0.07
0.07
0.08
0.09
1.54
1.37
1.24
1.17
1.06
0.92
0.86
0.74
0.63
0.51
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
NOTE: For this sample period
Op
and
0*
are 0.166 and 0.227, respectively. These imply a value for Wu equal to
0.023, which has a
p-value
of 0.476. Under the
hypothesis that the CRSP Equal-Weighted Index is efficient,
'( WTu)
is 0.024 and SD( W,,) is 0 01 1.
is relevant for
our absolute value of the slope null
hypothesis, and our
test is then
based on the squared values.
Figure
lb is based
on a data set that is very similar to the one used by Black,
Jensen, and Scholes (1972) (hereafter,
BJS).7 Using monthly returns on 10
beta-sorted
portfolios
from
January, 1931 through 1965, 0*
=
0.227
December,
=
0.166. To
while the CRSP Equal-Weighted
which is portfolio
Index, p, has
0
p
judge whether these two slopes are statistically
different,
we can calculate
is 0.02333. Based on the results
in Section
(#2
_
1),
which
2,
we can use a central
F distribution
with degrees of freedom 10 and 409 to judge the statistical
significance of this
difference in slopes.
The
resulting F statistic is 0.96, which
has
a p-value of 0.48. Our
multivariate test confirms the conclusion
reached by BJS
in that the ex ante
efficiency of the CRSP Equal-
for their overall time
period
Index
cannot Weighted be rejected; equivalently, if this Index
is taken
as the true
market portfolio, then the Sharpe-Lintner version of the CAPM cannot be
rejected. Table
II
provides some summary statistics
on the beta-sorted portfolios
that were used for
Figure
lb. Table II, when compared
with Table
II
in BJS,
that our data
base is very
similar
verifies
to the one used by BJS.
BJS provide various scatter
plots of average returns versus
estimated betas to
data to the linear relation if the
CRSP
judge
the fit of the
expected
Equal-Weighted
of
Chicago,
it is not possible to replicate their The CRSP
the University data. tapes are continually
revised to reflect data errors, and one would need the same version of the CRSP file to perfectly
duplicate a data base. For
example,
we were able to find more firms per year than reported in Table 1
of BJS because of corrections to the data base. Also we relied on Ibbotson and Sinquefield (1979) for
the return of US Treasury Bills as the riskless rate. This latter data base was not used by BJS.
in forming the 10 portfolios that
were
However, we followed the grouping procedure outlined in BJS
1 and
Table
II.
used in constructing Figure
7
While
on the
data BJS relied from the Center for Research in Security Prices (hereafter, CRSP) at
1130
MICHAEL R. GIBBONS, STEPHEN
A. ROSS, AND JAY SHANKEN
Index is efficient. We view figures like our Figure lb as complementary to these
scatter plots, for they summarize the multivariate test in a manner familiar to
financial economists. The advantage of the scatter plots in BJS is that they may
provide some information as to which asset or which set of assets is least
consistent with the hypothesis that the index is efficient; figures like
Figure
lb
really do not provide such information. On the other hand, the scatter plots in
BJS can be difficult to interpret due to heteroscedasticity across the different
portfolios as well as contemporaneous cross-sectional dependence. Section 6 will
suggest some other types of diagnostic information based on the multivariate
framework.
To understand further the behavior of our measure of efficiency,
4?,
its small
sample distribution given in Section
2 is
helpful. Since a linear transform of
42
has a central F distribution with degrees of freedom N and (T
-
N
-
1), we can
use the first two moments of the central F to calculate:
(8)
and
-1) = [
T-N-3
(9)
SD(P2_1)
=[T-N-3
T-N-5
The first moment for
42
only exists if T
>
N + 3 while the second moment for
42
only exists
if T> N + 5. These last two
equations
for the moments can be
applied to the BJS data set for 1931-1965 where
N
=
10 and T=
420, so
1) and the standard deviation of
p2
are 0.024 and
0.011, respectively.
As
6,(+2-
the realized value of
42
_
1
is less than its
expectation, it is not surprising that
the ex ante efficiency of the
Equal-Weighted
Index cannot be
rejected
for this
time period.
This measure, 4, is a new variant of the geometry developed to examine
portfolio performance.
In
past procedures the efficient frontier has been taken as
given, and a distance such as mb
in
Figure la has been used as a measure of p's
performance. Note that mb is simply the return differential of the ex post
optimal portfolio over p, computed
at the
sample
standard deviation of the ex
post optimal portfolio.
Another
suggestion
has been to use the difference in their
slopes
*
-
p
as a measure of
p's
relative
performance.
How the true ex ante
frontier is to be known is unclear, and if the ex post frontier is
used,
then we face
the statistical
problem
of this
paper.
FOR EFFICIENCY
OF THE MULTIVARIATE TEST
4. THE POWER
The empirical illustration in the previous section fails to
reject
the ex ante
efficiency
of the
Equal-Weighted
Index when
using
10 beta-sorted
portfolios
as in
BJS.8 Such a result may occur because the null hypothesis is
in fact
true,
or it
the results of the multivariate test across these four
subperiods,
we can
reject
ex ante
efficiency
at
usual levels of
significance.
This confirms the conclusions reached
by
BJS.
8
We have also examined our data base
using
the same
subperiods
as in BJS. When we
aggregate
PORTFOLIO
EFFICIENCY
1131
may be due to the use
of a test which is not powerful enough to detect
of
of the Index. Questions
from efficiency
deviations
economically
important
power for various types of test statistics
have been a long standing
concern
will
others). This
section
among
financial economists (e.g.,
see Roll
(1977),
among
of the multivariate test.
focus on the power
a
both the null and alternate hypotheses
From Section
2 we know that under
N
of freedom
of
W,
or
42,
has an F distribution with degrees
simple transform
and T
-
N
-
1. The F distribution is noncentral with
the noncentrality parameter
the noncentrality parameter is
the null hypothesis
(6); under
given by equation
under the alternative is
zero. It deserves emphasis
that the F distribution
parameter
conditional
on the returns of portfolio p since the noncentrality
on a
.
Thus, we
will be
studying
the
power
function conditional
depends on
O2
function.
value for
0,
not the unconditional power
The probability of rejecting a false
null hypothesis increases as the noncentral-
holding constant the numerator and denominator de-
increases,
ity parameter
that
page 193)).
Studying the factors
(Johnson and Kotz
(1970,
grees
of freedom
will give
some
of
the power
the noncentrality parameter, X, guidance about
affect
sum of
the multivariate test. From equation (6) we can see that
A is a weighted
matrix is the inverse
deviations about the point
ap
=0. The weighting
squared
estimators for
ap.
Thus,
of the covariance matrix of the ordinary least squares
to the variability of
according
estimated
departures from the null are weighted
and the cross-sectional dependence among the estimators.
the estimator
economic interpre-
The noncentrality parameter can also be given
an intuitive
of equation
(23)
in the
Appendix would hold for the
tation. The derivation
of the
sample
that a'
-ap=
estimates, so it is also true
counterparts
population
9*2
-
a2
It follows
directly
that
X
=
[
T/(I
+
"p2
)] (
@*2
_
0p2)
.
of the test will increase as the ex ante inefficiency of
Not surprisingly, the power
of the
slope
of the relevant
as measured in terms
opportu-
portfolio p
increases
of the estimator for
ap
declines,
so the
the precision
nity sets. If
p2
increases,
power
of the test decreases.
When
of the test is affected by
0*
and
Op.
2 summarizes how the power
Figure
is
equal
to
one,
the null
the proportion of potential efficiency
(i.e.,
Op/O*)
approaches zero, the given portfolio
is
hypothesis is true.
As this proportion
2 is based
on values
for the
significance level,
N,
less efficient.
Figure
becoming
we
work on asset
pricing
for existing
models;
and T that are common
empirical
significance level.
have used N
=
10 or 20 and T= 60 or 120 and a five percent
estimates of
Op
between 0 and 0.4.
work on the CRSP Indexes
reports
Empirical
and 0*. In
We have used this range to guide our selection of a grid for
OP
=
Op
2 is based to eliminate one of the
on the assumption that
p
Figure
addition,
that
our
calculations of
power
that affect
A;
this
assumption suggests
parameters
of the
underlying population.
is
representative
the
sample
are for situations where
of
Even within the range of parameters that
we
consider,
the probability
ranges
from five
percent
to nearly 100 percent
rejecting the null hypothesis
measures of
slope.
For
depending
on the difference between the
two relevant
1132
MICHAEL R.
GIBBONS,
STEPHEN
A.
ROSS,
AND JAY
SHANKEN
,
01
C
4
~~~0
OTENTIAL
EFFICIENC'
2a.)
N
10
and rr
60.
A
ONI0
NT
1
0~
PRP
ENIA
7F
ION OF p0TrNTIAL
I
2b.) N
20 and
'r-
60.
FIGURE
2.-Sensitivity
of the
power
of the test to the choice of the index. Each
figure
is based on a
different combination of the number of assets
(N)
and the number of time series observations
(T).
In
all cases a critical level of five
percent
is used.
is
high
relative
example,
if
Op
to the
average
equals .2 (which
from
1926-1982)
and if N
=
20 and T
=
60,
then
the
probability
of
rejecting
a false null
hypothesis
from ten
percent
ranges
(for
9* =
.3)
to 98
percent
(for
9* =
1.0).
Given the data bases
that are
available,
an
empiricist
is
always
faced with the
question of the
appropriate
sizes for N and
T. For
example,
with the CRSP
monthly
file we have a data
base which
extends
back to 1926 for
every
firm
on
the New York Stock
Exchange.
This would
permit
the
empiricist
to use around
700 time series
observations
2000 firms
(i.e., T)
and
well over
(i.e., N). However,
the actual N used
may
be restricted
by T
to
keep
estimates of covariance
matrices
nonsingular,
and the actual T used is
constrained
by
concerns
over
PORTFOLIO
EFFICIENCY
1133
1
0
1
OF
PO
PROPORTION
4
ION
FOTENTIAL
EFFICIENC-y
2c.) N 1 arid
r
2d.)
N
-
20 and T
120.
FIGURE
2.-Continued.
stationarity. It is not uncommon
to see published work where
T is around 60
monthly observations and N is
between 10 and 20. While
these numbers for
N
and T are common, we
are not aware of any formal
attempts to study the
appropriate values to select. We
will now examine this issue in
the context of the
specific hypothesis of ex ante
efficiency. While the analysis is
focused on an
admittedly special case, our hope is
that it may shed some
light on other cases as
well.
To get more intuition
about the impact of N on equation
(6), consider a case
in
which
S
has a constant value down the
diagonal and a constant (but
different)
value for all off-diagonal
elements. Since
S
represents the
contemporaneous
covariances across assets after
the "market effect" has been
removed, sucha
1134
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
structure includes the Sharpe (1963) diagonal model as a special case when the
off-diagonal terms are zero. The more general case where the off-diagonal terms
are constant but are nonzero is motivated by the work of Elton and Gruber
(1973) and Elton, Gruber, and Urich (1978).9 Under this structure we can
parameterize T as:
(10)
S
= (1
-P)
2IN +
P2
'I,
where p
the correlation between
eit
and
ij,;
2
the variance of
eit; IA,
an
identity matrix of dimension N; and
l
I-
a 1 x N vector of l's. The inverse10 of
this patterned matrix is (Graybill (1983)):
(lP)2
[IN
1
+
(N-l)p
LNLNb
in equation (6) gives:
Substituting the above equation for .1
(11) =
T/(1+
(1p)W2
[N
)
[N2
(1-p)+Np
p
where y-
(l
ap)/N
and
2
(a'a
)/N. One could view
ju
as a measure of the
"average" misspecification across assets while
A2
indicates the noncentral disper-
sion of the departures from the null hypothesis across assets.
When N is relatively large and p is not equal to zero, equation (11) implies:
(12)
=
X/N_
1
_
A)y2
2)
=
(1(
)VAR(ap),
where VAR
(ap)
denotes the cross-sectional variance of the elements of
ap.
Thus,
X is approximately proportional to N and T."1 Alternatively, if either p = 0 or
Unfortunately,
this is still
Al
=
0,
then
X
is
exactly proportional
to N and T.12
not adequate to determine the impact of changing N and T, for these two
parameters affect not only the noncentrality parameter but also the degrees of
freedom.
We have evaluated the power of the multivariate test for various combinations
of X, N, and T.13
These numerical results provide some guidance on the proper
9Strictly speaking, the Sharpe diagonal model allows for heteroscedasticity in the disturbances of
the market model equations; our formulation assumes homoscedasticity. Also, the constant correla-
tion model of Elton and Gruber is usually applied to the correlation matrix for total returns; we are
assuming constant correlation after the market effect has been removed.
10
Necessary and sufficient conditions for this inverse to exist are that p
#
1 and p
#
(1
-
N)
';
see Graybill (1983, page 190-191). In addition, the matrix should be positive definite; this would
require that p>
-
l/(N
-
1).
"1 In general, since p < 1, X/N is less than or equal to the right side of (12) when p ? 0.
12
If the Equal-Weighted Index is portfolio p, then we would expect
Al
to approach zero as N
becomes large.
13
These numerical calculations require
evaluation of a noncentral F distribution and an inverse of
a central F distribution. The latter calculation relied
on the MDFI subroutine provided by IMSL. The
former calculations are based on a subroutine written by J. M. Bremner (1978), and a driver program
written by R. Bohrer and T. Yancey of the University of Illinois at Champaign-Urbana. Each
subroutine was checked by verifying its output with the published tables reported
in Tang (1938) and
Titu (1967).
PORTFOLIO
EFFICIENCY
10
1
1135
,
0 .8-
0.7
,-
0.6
SA=O.1T
-
0-4
;+
A
/
0.002NT
0.3
-
A
0.00002NT
,
02
0
0.2
0.4
0.6
0.8
1
3a.)
T
60.
Ratio of N Divided
by T
0.9
-
0.8
,
;:
0.7-
0.6-
0.5-
0.4-
s
W
0
.
0.3-
5
0.2
]
iA
0.002NT
0.1
-
0
]
0
A-
0.00002NT
A
.
I
,
I
0.2
0.4
.
0.6
I
O.B
I
Ratio
of N Divided
by T
3b.)
T
120.
of the number of assets
(N) gives
a
of the test to the choice
3.-Sensitivity of the power
FIGURE
fixed number of
time series observations (T).
that X is proportional to NT, and Figure
3
choice of
N and T. We assumed
We selected this
provides various
values for the
constant
of proportionality.14
based
on
equation
(11)
when
p
= 0. In this case, the
constant
of proportionality
the cross-sectional
+
)].
We then
is
2/[2(1
replaced
constant
I2
and
o2
with
data set. We also know
that
from an actual
of and
6i2,
respectively,
averages
Index (1931-1965) and 0.109 for
the
O
is 0.166 for the CRSP
Equal-Weighted
p
of the
(1987) studies the power of
the test using alternative parameterizations
MacKinlay
noncentrality parameter.
14
1136
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
z
0
0.9
0.7
A=O.1T
A
0.002NT
0
.
6
0.5
.
0.4
0.3
0.2
0
o.
-
0
A
0.2
0.002N
0.4
.
0.6 0.8
1
Ratio of N Divided by T
3c.) T
240.
FIGURE
3.-Continued.
guide to typical
This provides a rough
CRSP Index (1926-1982).
Value-Weighted
values for the constant of proportionality. The constant is
0.004 using the
beta-sorted
portfolios, and it is 0.002 using a set of industry
portfolios. For
size-sorted portfolios the constant is 0.004 using all months and 0.763 for
and
on how the
industry portfolios
January. (The
details
data only using
monthly
later
in this paper.)
were will be provided
created
size-sorted portfolios
is 0.00002
and 0.002, which
In Figure
3, we look at cases where
the constant
3
are small relative to the above calculations. For purposes of comparison, Figure
=.1T.
This
by N; instead
we set
X
also includes
a case where
X
is not affected
where an investigator has one asset that violates the null
a situation
represents
with the
hypothesis,
and all the remaining assets that are added are consistent
3 is based for
on specific values
Figure
portfolio p. While
efficiency of some
given
the general pattern that is observed is consistent
the constant of proportionality,
here.15
of choices but did not report
that we tried
with a wide range
of time series observations, Figure
3 demonstrates that
For a fixed number
Even
though
decision to be made
by
the
empiricist. there may be an important
it is not
necessarily appro-
increases as N
increases, the noncentrality parameter
N
possible.
Given our
particular parameterization
priate
to choose the maximum
a third to one half of
T,
or
be
roughly
of the problem, it appears that
N should
data are used, 20 to 30 assets may
be
appropriate.
when five years of monthly
of N
values
is so small
for all
possible
When the constant is very
low,
the
power
15
F for
very high
values
of
A,
so we have little
the noncentral
We were not able to evaluate
If the
is
high.
the constant of proportionality
of the
power
function when
knowledge about
the shape
= -
T 2 be
solution of setting N
it is conceivable that a corner
may
constant is large enough,
appropriate.
PORTFOLIO
EFFICIENCY
1137
that it is not an important decision.
Alternatively, if the noncentrality parameter
is proportional to T and not affected by N,
clearly
setting
N
=
1 is the preferred
In this case
adding
strategy. securities does not
provide
more information about
departures from the null hypothesis;
additional securities increase
however,
the
number
of unknown
parameters to be estimated. It deserves
emphasis that these
conclusions about
the
proper choice
of N
may
not be appropriate for all
possible
and models.
situations
The choices of N and T are not the
only decisions
facing the empiricist
in
the econometric
Since N must
always
designing
analysis. be less than T
(unless
highly structured covariance matrices
are entertained), the empiricist must also
decide
how to select the assets
to maximize the power
of the test.
Given N and T
we wish to maximize
the quadratic form
a-T'ap,
or equivalently 9*; however,
these parameters are unobservable. A common
approach
is to use
beta-sorted
in betas is useful in
decreasing
portfolios. While dispersion
the asymptotic
in estimates
standard error
of the expected return
on the zero-beta
asset (Gib-
bons (1980) and Shanken
(1982)), such sorting need not maximize
departures
from the null hypothesis as measured by X.16
Empirical examples presented below
illustrate the effect
that
different asset
sets
can have on the outcome of the test.
First,
we consider
a set of 12
industry
seems reasonable on economic
portfolios.17
An
industry
grouping
and
grounds
also captures some of the important correlations
To measure the
among
stocks.
return from a "buy-and-hold"
investment strategy, the relative market
values of
the securities are used
to weight the returns. Almost
return on the
every
monthly
CRSP
tape from
1926 is
included,
1982 which should
minimize
through
problems
with survivorship
bias.18
Table III provides some summary
statistics on the
industry
portfolios.
The multivariate F statistic rejects
the hypothesis of ex ante
efficiency at about
the one
percent
level.
The relevant
F statistic
is 2.13 with
degrees
significance
of
freedom 12 and 671; its p-value is 0.013.19
To complement
these numerical
results,
Figure Ic, which
is similar
to Figures
la and lb, provides
a geometrical
summary.
assets into N portfolios; we could
form N portfolios so that
they
have very little
dispersion in their
beta values with no impact
on the
power. This follows from
the
well-known result in the
multivariate
statistics
literature that our test is invariant to linear
transformations of the data (Anderson (1984,
pages 321-323)). Of course,
the selection of the original subset
of assets to be analyzed is important
even though the way they are into
portfolios is not (given
aggregated
that
the number of portfolios is
the same
as the number of original assets).
17
For the details of the data see Breeden, Gibbons, and
Litzenberger
base,
who
developed
(1987),
these data for tests of the consumption-based
asset pricing
model. The industry
grouping closely
follows a classification used
by Sharpe (1982).
18
with a SIC
number of 39 (i.e., miscellaneous manufacturing
However, all firms
industries) are
excluded to avoid any possible problems with a
singular
covariance matrix when the CRSP
Value-Weighted Index
is used as portfolio p.
19
While not
reported
we also
analyzed
this
data set across
various
Based on five
here,
subperiods.
for the F statistic is less than in 7 out
of 11
cases,
year five
subperiods, the p-value
is less than
percent
in 9 out of 11 cases,
10 percent
and rejects when across
aggregated the
subperiods.
the
rejection Thus,
of the overall
is confirmed by the subperiods as well.
period
16
In fact, for a given
set of N securities, the multivariate
test is invariant to how we group
these
1138
MICHAEL
R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
To understand this low p-value, consider
the fact that for this time period
=
0.109 while the slope of the opportunity
set using the ex post optimal
than double with a value
of 0.224. With these numbers we
portfolio, 9*, is more
For N= 12 and T=
684,
e(42)
iS
1.018 with
SD(
2)
can calculate
42
as 1.038.
three standard deviations from
of 0.007. Thus, the realized value
of
42
is
nearly
Index is truly ex ante efficient.
value
if the CRSP
Value-Weighted
its expected
of greater interest
the null
Perhaps is the fact that the multivariate test rejects
all 12 univariate t statistics fail to
hypothesis at the one percent level even though
builds on such contrasting
reject at even the five percent level.
The next section
to summarize across
results by analyzing why univariate test may be difficult
different assets.
0
5. THE PROBLEM WITH UNIVARIATE
TESTS
Table
II
suggests that high beta portfolios earn too little and low beta
Index is presumed
to be efficient;
portfolios too much if the Equal-Weighted
for the zero-beta version of
similar evidence was used by BJS to garner
support
is difficult to
interpret.
The
upper
the CAPM.
triangular portion
Yet, this pattern
of the market model
correlation matrix residuals
of Table
IV provides
the sample
in Table II. A
very distinctive
based on the regressions that are summarized
are
the residuals of portfolios with
similar betas
positively
pattern emerges in that
correlated
while those of portfolios with very different betas are negatively
for
ap
in equation
Based on the variance-covariance matrix
(4), it is
correlated.
the same pattern of correlation. Thus, it
clear that the estimators for
aip
will
have
in estimated
pattern values of a 's is
is difficult to infer whether the observed
TABLE III
ON INDUSTRY-SORTED PORTFOLIOS
BASED ON MONTHLY DATA,
SUMMARY STATISTICS
FORM,
1926-82 (T
=
684). ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE
AND THE CRSP
VALUE-WEIGHTED INDEX IS PORTFOLIO
PARAMETER
p.
THE FOLLOWING
ESTIMATES ARE
FOR THE REGRESSION MODEL:
,,=
a?ip
+/ApFpt
AND
+t
Vi
=1,.12
Vt=
1.684,
WHEREI
IS THE COEFFICIENT OF DETERMINATION FOR EQUATION i.
Industry Portfolio
P,
s(fi
a,
0.17
-0.05
0.03
0.00
0.12
-0.17
0.10
-0.17
0.05
0.00
0.43
-0.03
s(&,p)
R2
Petroleum
Financial
Consumer Durables
Basic Industries
Food and Tobacco
Construction
Capital Goods
Transportation
Utilities
Trade and Textiles
Services
Recreation
0.14
0.09
0.09
0.00
0.07
0.12
0.08
0.14
0.09
0.00
0.37
0.13
0.93
1.19
1.29
1.09
0.76
1.20
1.08
1.20
0.74
0.94
0.80
1.22
0.02
0.02
0.02
0.01
0.01
0.02
0.01
0.02
0.02
0.02
0.06
0.02
0.69
0.89
0.90
0.94
0.83
0.85
0.91
0.78
0.76
0.77
0.19
0.78
NOTE:
For this sample period
6,
and 6* are 0.109 and 0.224, respectively. These imply a value for Wu equal to
0.038, which has a p-value of 0.013. Under the hypothesis
that the CRSP
Value-Weighted Index is efficient, 9'( W")
is 0.018 and
SD(W")
is 0.007.
PORTFOLIO EFFICIENCY
TABLE IV
1139
SAMPLE
CORRELATION
MATRIX OF RESIDUALS FROM MARKET MODEL REGRESSIONS USING
EXCESS RETURNS.
THE UPPER TRIANGULAR
PORTION OF THE TABLE IS BASED ON 10 BETA-SORTED PORTFOLIOS
FOR THE DEPENDENT
VARIABLES AND THE
CRSP
EQUAL-WEIGHTED
INDEX FOR PORTFOLIO
p.
ALL MONTHLY
DATA FROM
1931-65
(T
=
420)
ARE USED. TABLE II
SUMMARIZES THE
OTHER
PARAMETER ESTIMATES FOR THIS REGRESSION
MODEL.
THE
LOWER
TRIANGULAR
PORTION OF THE TABLE IS BASED ON 10
SIZE-SORTED PORTFOLIOS FOR THE DEPENDENT
VARIABLES
AND THE
CRSP
VALUE-WEIGHTED
INDEX FOR PORTFOLIO
p.
ALL MONTHLY
DATA FROM
1926-82
(T
=
684)
ARE USED. TABLE V SUMMARIZES THE OTHER PARAMETER
ESTIMATES FOR THIS REGRESSION
MODEL.
1 2
3
4 5
Portfolio Number:
6 7
8
9 10
.52
.62
.72 .68
.66
.70
.66
.63
.61
.63
.41
.52
.39 .39
.28 .35
-.54 -.59
.39
.38
.75
.70
.68
.57
.51
.21
-.68
.03
.01
.08
.72
.65
.55
.46
.27
-.66
-.32
-.16
-.06
-.06
.72
.62
.50
.26
-.68
-.51
-.35
-.25
-.07
.21
.67
.59
.36
-.68
-.64
-.50
-.37
-.11
.25
.34
.52
.27
-.61
-.60
-.52
-.43
-.13
.12
.36
.43
.30
-.66
-.64
-.53
-.46
-.32
.03
.26
.46
.49
-.38
-.50
-.55
-.49
-.22
- .16
.06
.23
.27
.56
NOTE:
For the upper triangular portion of the table, portfolio 1 consists of firms with the highest values for
historical estimates of beta while portfolio 10 contains the firms with the lowest values. For the lower triangular
portion of the table, portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest
decile of the NYSE while portfolio 10 contains firms in the highest decile.
in the true
in the estimation or to the actual
pattern due to correlation error
parameters.
in financial economics could
also be cited
from
Other empirical work
examples
of Banz
difficult to interpret. Since the work and
(1981)
where univariate tests
are
a great
deal of attention. (For
has received
Reinganum (1981), the "size effect"
see Schwert
more information about this research
(1983),
who summarizes the
a useful bibliography.) While most of the
existing evidence and also provides
in this area
now focuses on returns in January, we begin
by looking at
research
which did
not distinguish between January and
non-January
evidence
the original
returns.
We have created
a data base of monthly
stock returns
using the CRSP file.
based on the relative
market value of their
Firms were sorted
into 10 portfolios
In other we ranked
firms
total equity
outstanding.
market values
by their
words,
in December, 1925 (say), and we then formed 10 portfolios where the first
in the lowest decile of firm
size and the tenth
portfolio contains all those
firms
in the
highest
of firm size on the New York
of
companies
decile
consists
portfolio
and the firms are
is
value-weighted,
Each of the ten
portfolios
Stock Exchange.
for five
years.
market values
not resorted Thus,
the returns on these 10
by their
1926
through December,
1930
represent
the returns
portfolios
from
January,
for five this
portfo- from a buy-and-hold strategy without years;
any rebalancing
to represent a low transaction cost investment strat-
lio formation was adopted
1140
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
TABLE V
SUMMARY STATISTICS ON
SIZE-SORTED PORTFOLIOS BASED ON MONTHLY
DATA,
1926-82
(T
=
684).
ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE FORM, AND THE
CRSP
VALUE-WEIGHTED INDEX IS PORTFOLIO
p.
THE FOLLOWING PARAMETER ESTIMATES
=
,P
+
,,fprpt
+
Eit
Vi =
1.
10 AND Vt =1.
MODEL:
ARE FOR THE REGRESSION
684,
Fit
WHERE R2
IS THE COEFFICIENT OF DETERMINATION FOR EQUATION i.
Portfolio
Number
/,p
s(6,p)
f3,lip
s(ip)
RI
1
2
3
4
5
6
7
8
9
10
0.28
0.34
0.25
0.18
0.19
0.18
0.08
0.08
0.00
-0.01
0.24
0.18
0.14
0.13
0.10
0.09
0.07
0.06
0.00
0.05
1.59
1.45
1.40
1.36
1.27
1.25
1.18
1.17
1.16
0.94
0.04
0.03
0.02
0.02
0.02
0.02
0.01
0.01
0.01
0.00
0.68
0.78
0.84
0.85
0.89
0.91
0.93
0.95
0.96
0.98
NOTE:
Portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest decile of the
NYSE while portfolio 10 contains firms in the highest decile. For this sample period
#
and #* are 0.109 and 0.172,
respectively. These imply a value for Wu equal to 0.017, which has a p-value of 0.301. Jnder the hypothesis that the
is 0.007.
CRSP Value-Weighted
Index is efficient, 6(W") is 0.015 and SD(PWV)
egy. The resorting and rebalancing
occurred in December of
1925, 1930,...,1980.
for the entire
Table
V summarizes the
behavior of the returns on these portfolios
time
period.
Given the existing evidence on the size effect, some readers
may find it
somewhat surprising
that,
in the overall
period
from 1926
through 1982, the
of the CRSP
Value-Weighted
Index at
multivariate test fails to reject
efficiency
VI
reports
The first row of Table the statistic and
the usual
levels of
significance.
a geometrical interpretation for
its corresponding p-value;
Figure
Id provides
this overall
period.20
The correlation
matrix of the market model residuals of the size portfolios
exhibits a distinctive of Table IV provides
pattern. The lower triangular portion
this information based on the overall
period. However, the pattern is identical
in Table
ten
year
across
every subperiod reported VI, and a similar pattern is also
described
by Brown, Kleidon,
and Marsh
(1983, page 47) and Huberman and
The correlation is positive
Kandel
(1985b).
and high
among
the low decile
firms.
The correlation declines as one compares portfolios from very different deciles.
Even more striking
is the fact that the highest decile portfolio has negative
with all other decile portfolios.
sample correlation (In some of the subperiods,
for the ninth decile as
well.) Thus,
if we
occurred
this negative correlation
observe that the lowest decile performs well (i.e., estimated alphas that are
that the highest decile would do poorly
we would then
expect positive), (and
vice
in the
period 1946-1955, where
This is the case, for example,
five out of
versa).
The subperiod results in Table VI are consistent with the conclusions of Brown, Kleidon, and
Marsh (1983) who find the size effect is not constant across all subperiods.
20
PORTFOLIO EFFICIENCY
TABLE VI
1141
TESTING THE EX ANTE
EFFICIENCY OF THE
CRSP VALUE-WEIGHTED INDEX (I.E., PORTFOLIO
PORTFOLIOS. ALL SIMPLE EXCESS
RETURNS ARE NOMINAL
p)
RELATIVE TO 10 SIZE-SORTED
AND IN PERCENTAGE FORM. OVERALL
PERIOD IS BASED ON ALL MONTHLY
DATA FROM
1926-82.
THE FOLLOWING MODEL IS ESTIMATED
AND TESTED:
r,
alp+
Aprpt
+
9,t
Vi
=
1.
10 AND Vt= 1,...,T.
Ho:
a,p
=0
Vi=
1,0.
Time
Period
(T)
p
W.
(P-Value)
of
Number
?t(ap)
I
>
1.96
1926-1982
(684)
1926-1935
(120)
1936-1945
(120)
1946-1955
(120)
1956-1965
(120)
1966-1975
(120)
1976-1982
(84)
0.109
0.065
0.146
0.308
0.216
-0.019
0.093
0.172
0.354
0.286
0.469
0.604
0.408
0.538
0.018
(0.301)
0.119
(0.227)
0.059
(0.765)
0.113
(0.264)
0.302
(0.001)
0.164
(0.065)
0.275
(0.039)
1
0
0
5
1
0
9
NOTE: on the CRSP
Value-Weighted Index divided by its
excess return
#P
is the ratio of the sample average
sample standard
deviation,
and #* is the maximum value
possible
of the ratio
of the
sample
average
excess return
divided
by the sample
standard deviation.
W_
(#*2
-
)/(1
+
2),
and it is distributed as a transform of a central
with
degrees
of freedom 10 and T
-
11 under the null
hypothesis.
W should
converge
F distribution
to zero as T
the
p-values
for the
W.
approaches
infinity
if the CRSP
Value-Weighted Index
is ex ante efficient.
By converting
for a standardized normal
statistics to an implied realization
random the results
across the 6
subperiods
variable,
and standardized normals
can be summarized by summing
and dividing root of
up
the 6
independent
by the square
in Shanken
6 as suggested This
quantity
is 2.87 which
implies
across
the
subperiods
a
rejection
at the usual
(1985).
levels of significance.
ten portfolios have
significant alphas
but the multivari-
(at
the five
percent level),
ate test cannot
reject
the
efficiency
of the
Value-Weighted
Index.
Even though summarizing the results of
univariate tests can be difficult,
work continues to
report
such statistics.
This is
only natural,
applied empirical
for univariate tests are more
intuitive
because
and
(perhaps
they
are used
more)
seem to give more
diagnostic information about
the nature of the
departure from
the null
hypothesis
when
it is
rejected.
Part of the
goal
of this
paper
is to
provide
some intuition
behind
multivariate tests. Section 3 has already done this to some
extent by
demonstrating that the multivariate test can be viewed
as a particular
in mean-standard deviation
space
of
portfolio theory.
The next
measurement
section
shows
that
the multivariate test is equivalent to a
"t
test" on the
intercept
in a particular regression which
should be intuitive. A way to generate
diagnostic
about the nature
information
of the departures from the null
hypothesis
is also
provided.
6. ANOTHER
INTERPRETATION OF THE TEST
STATISTIC,
W
=0
Vi is violated
if and
only
if some
linear The hypothesis combina-
that
aoip
of the N assets
has a
tion of the a's is
zero;
i.e.,
if and
only
if some
portfolio
1142
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
nonzero intercept when its excess
returns are regressed on those of portfolio
p.
the portfolio
which,
in a
given
With this in mind, it is interesting to consider
of the usual t statistic
the square for the intercept. It is well
sample,
maximizes
on multivariate statistics
that this maximum value is
known in the literature
T2
statistic,
our
TW,.
In this section
we focus on the
composition
of
Hotelling's
the maximizing portfolio,
a, and its economic
interpretation.
Thus, let
Frat
a'F2, where F2,
is an N x 1 vector with typical element
?,
Vi-1,..., N. Let a& be the N x
1
vector of regression intercept estimates. Then
aa
a
a'&t and
VAR(
a)
=
(1
+
0
)a'>a/T
by (4) above.21 Therefore,
[&A]2
(A'&)2AY
f
(13)
t2
=
[S(^)]=
AA
P
Since
we can multiply
a by any scalar
without of
t2,
we shall
changing the value
= c, where c is any constant
different
from
adopt the normalization that a'&t
zero. With this normalization,
T(a'&t)2
and (1 +
0p2)
in
(13) are fixed
given the
t2
is equivalent to the following minimization prob-
sample. Hence, maximizing
lem:
min:
a
a'Ta
=c.
a'ap
to:
subject
Since
the above
problem is similar to the standard portfolio
of
problem, the form
the solution
is:
a
A
a
p
a p
for a into equation
Substituting this solution
(13),
t2
becomes:
2
_
a
p
1
A
p
+
0
2
p
this equation with
(5)
establishes that
t2
=
TWa.
Not surprisingly,
Combining
the
of
ta
is not Student
t, for portfolio a was formed after
examining
distribution
the data.
The derivation of
t2
suggests some additional information to summarize
work on ex ante
efficiency. Given the actual
value of a based on the empirical
sample, one
will know the
particular
linear combination which led to the
of the null
hypothesis. If the
null
hypothesis is rejected, then a may
give
rejection
us some constructive information about a better
how to create
model.
Portfolio a has an economic
basis as well. When this portfolio is combined
with portfolio
properly p, the combination turns
out to be ex post efficient. In
Since we are
working
with returns in excess of the riskless
rate,
t'
a need not
equal
1,
for the
or
short)
asset
will be held
(long
so that all wealth is invested.
riskless
21
PORTFOLIO EFFICIENCY
1143
of k,
other words,
for some
value
(14)
= (1
-k)Fpt
+
krat,
Ft*
For convenience, we set
is the return on this ex post efficient portfolio.
where
Ft*
means of
rpt,
of
are all equal.
The
equivalence
c so that the sample
rat,
and
Ft*
means
requires that:
these three
aCf
a2rp
ap
-
Ip
r2
For the remainder of the paper, we will refer to portfolio a as the "active"
of our methodology, portfolio p will be a
portfolio. In many applications
our methods
"passive" portfolio, i.e., a buy-and-hold investment strategy. While
in its
where portfolio p is not passive, certainly
are applicable to situations
In such a
setting
to tests of the CAPM, portfolio p
will be
passive.
application
for it represents a way
portfolio,
portfolio a is naturally interpreted as an active
of "active" and
to improve the efficiency
of portfolio p. The terminology
has been used
by Treynor and Black
among others.
(1973),
"passive"
and portfolios
between the ex post efficient portfolio
To establish this relation
for the weights of an efficient portfolio,
a and p, we first recall the equation w*.
and setting
m in that
equation
to
rp:
Using equation (22) in the Appendix
(15)
w*
P
V can be parameterized as:
V- X
2tS
Using the formula for the
inverse of a partitioned matrix (see equation (24) in the
on the
last expression and substituting this
into equation (15) for
V-1,
Appendix)
but
straightforward,
after some
tedious,
algebra.
equation (14)
can be derived
The previous paragraphs have established
that the square of the usual t
statistic for the estimated intercept,
&a,
equals the
T2
test statistic,
TW,.
A
with the
can be established as well
for the ex
post efficient portfolio
similar result
a*,
from
same sample
mean as
portfolio
intercept, regress-
p, i.e.,
the estimated
-
is identical
t
statistic,
to
t
2.
Since
we will
t*2,
which
ing
*
on
rpt
has a squared
in what
follows,
we only note the fact here without
proof.22
not use this result
we return to
To illustrate the usefulness of the active
portfolio interpretation,
all
months)
is examined.
the size effect
(across
the example
of Section
5 where
V is roughly consistent with the findings of Brown,
The second column of Table
as well as the
disturbance in the regression of
F,*
on
Fp,;
these equalities hold for the estimates
k cancels. The two key facts can be
parameters. Since t* is essentially a ratio
of
&*
and SD(E^*),
established by working with
the moments of
Ft*
on equation (14).
based
22
The two key facts used in the proof are that a*
=
kaa
and
SD(9,*)
=
kSD(9a,),
where
9,*
is the
1144
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
TABLE VII
INFORMATION THE ACTIVE ARE BASED
DESCRIPTIVE ABOUT a. THESE STATISTICS
PORTFOLIO,
PORTFOLIOS USING MONTHLY 1926-82
(T
=
684).
ALL SIMPLE
ON SIZE-SORTED
RETURNS,
ARE NOMINAL AND IN PERCENTAGE
EXCESS
RETURNS
FORM,
AND THE CRSP
INDEX IS
PORTFOLIO
VALUE-WEIGHTED
p.
Monthly Returns
in all Months
(T
=
684)
Monthly Returns
in only January
(T-=
57)
Monthly Returns
Excluding January
(T
=
627)
&a
t2
ta
k
a,
a2
a3
a4
a.
a6
a7
a8
0.05
11.97
7.56
-0.02
1.36
71.57
0.87
0.09
0.05
11.09
7.95
-0.04
0.04
0.04
0.00
0.06
0.07
0.08
0.07
0.03
-0.01
0.03
-0.02
0.12
0.08
0.01
0.10
-0.04
0.10
0.04
0.06
0.03
0.09
a9
al0
aRF
0.04
0.60
0.12
1.00
-0.52
-0.16
1.25
1.00
0.01
0.63
0.09
1.00
Ea,
NOTE: Portfolio 1 is a value-weighted portfolio of firms whose market
capitalization is in the lowest decile of the
NYSE while portfolio 10 contains firms in the highest decile. The portion of wealth invested in the riskless asset is
denoted by
aRF.
Kleidon, and Marsh in that the estimated alphas are approximately monotonic in
the decile size rankings.
However,
such a result does not
imply that
an
optimal
portfolio should give large weight to small firms. As Dybvig and Ross (1985)
point out, alphas only indicate the direction of investment for marginal improve-
ments in a
portfolio.
The
portfolio
that is
globally optimal may
have a
very
different weighting scheme than is
suggested by the alphas. A comparison of
Tables V and VII verifies this.
For example, the portion of the active portfolio invested in the portfolio
of the
smallest firms
(i.e.,
a,)
has a sign which is opposite that of its estimated
alpha.
Furthermore, the active portfolio suggests spreading one's investment fairly
evenly across the portfolios in the bottom 9 deciles and then investing a rather
large proportion in the portfolio of large firms, not small firms. Table VII also
reports
a
'2,
and k for the overall period. Note that as k is much greater than
one (k
=
7.56), the ex post efficient portfolio has a huge short position in the
value-weighted index. Since this index is dominated by the largest firms, the net
large firm position in the efficient portfolio is therefore actually negative. It is
interesting that ex post efficiency is achieved by avoiding (i.e., shorting) large
firms rather than aggressively
investing
in small firms.
The reader should keep in mind that Tables IV through VI and the second
column of Table VII have examined the size effect across all months. Based on
just these results, the size effect seems to be less important than perhaps
originally thought. However,
if the data are sorted
by January returns versus
PORTFOLIO EFFICIENCY
TABLE VIII
1145
SUMMARY STATISTICS ON SIZE-SORTED PORTFOLIOS BASED ON JANUARY
RETURNS,
1926-82
(T=
57).
ALL SIMPLE EXCESS RETURNS ARE NOMINAL AND IN PERCENTAGE
FORM,
AND THE
CRSP
VALUE-WEIGHTED
INDEX IS PORTFOLIO
p.
THE FOLLOWING
PARAMETER ESTIMATES
=
aip +3,,iprpt
+
,it
Vi=.
MODEL:
ARE FOR THE REGRESSION
10,
WHERE
R2
IS THE
Fit
COEFFICIENT OF DETERMINATION FOR EQUATION
i.
Portfolio
Number
a,p
s(&,p) 3,,p s(I,P)
RI
1
2
3
4
5
6
7
8
9
10
6.12
4.60
3.43
2.88
1.79
1.79
0.85
0.80
0.31
-0.52
0.87
0.67
0.47
0.49
0.34
0.30
0.20
0.23
0.17
0.10
1.67
1.52
1.44
1.44
1.19
1.21
1.16
1.17
1.05
0.94
0.18
0.14
0.10
0.10
0.07
0.06
0.04
0.05
0.03
0.02
0.62
0.69
0.80
0.80
0.85
0.88
0.94
0.92
0.95
0.98
NOTE:
Portfolio 1 is a value-weighted portfolio of firms whose market capitalization is in the lowest decile of the
NYSE while portfolio 10 contains firms in the highest decile. For this sample period
#P
and
(*
are 0.259 and
1.197,
respectively. These imply a value for
WI,
equal to 1.256, which has a p-value of 0.000. Under the hypothesis that the
CRSP Value-Weighted Index is efficient, 4(W
,)
is 0.219 and
SL(W,,)
is 0.111.
non-January returns, the multivariate approach confirms
the importance of the
size effect-at least for the month VIII summarizes the
sample
of January. Table
characteristics of our 10 size-sorted portfolios when
using only returns in January
from
1926 through 1982. A comparison of Tables
V and
VIII reveals that the size
effect is much more pronounced in
January
than in other months; this is
consistent with the work
by Keim (1983).
This impression
from the univariate
is confirmed by the
multivariate
statistics
test of ex ante
efficiency, for the F test
is 5.99 with a p-value
of zero to three
decimal
places. In contrast, the F test
based on all months excluding
is 1.09 with a p-value of 0.36. The
January
of the active
weights
portfolio,
a, are presented in the last two columns
of Table
VII for January versus
non-January months. As in the first
column of Table
VII,
the active
portfolio
is not dominated by small
firms. For the month
of January,
one's investment
should be evenly spread
(roughly across the eight
speaking)
in the bottom
deciles
portfolios
in the top two
(or smaller firms); however, firms
deciles
(or larger
be shorted.23
firms) should
Results for non-January
months
are
similar to those based
on all monthly data.
The evidence in Table VII
suggests
that the optimal active
portfolio
is not
dominated firms even in the month
of
January-at
least based on the ex
by
small
in the
marketplace we see the development
post sample
moments.
Nevertheless,
of mutual funds which
specialize
in
holding the equities
of
just
small firms.24
The active portfolio for the month of
January
involves a rather
large position in the riskless asset
(CRF
equals 1.25). This investment in the riskless security is necessary to maintain a sample mean
return on the active portfolio equal to that of the CRSP
Value-Weighted Index.
24
Examples
of such funds include The Small
Company Portfolio of Dimensional Fund Advisors
and the Extended Market Fund of Wells Fargo Investment Advisors.
23
1146
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
Such funds suggest that efficient portfolios may be achieved by combining
indexes like the S&P 500 (or the CRSP Value-Weighted Index) with a portfolio
of small firms. We now turn to an examination of the ex ante efficiency of such a
linear combination in the next section. A multivariate statistical test of such an
investment strategy turns out to be a simple extension of the test developed in
Section 2.
7. TESTING THE EFFICIENCY OF A PORTFOLIO OF L ASSETS
If a portfolio of L other portfolios is efficient, then there exist parameter
restrictions on the joint distribution of excess returns similar to those considered
earlier. Specifically, if P
=
E
1xj/j,
(where
ELlxj
=
1) and if Fr
is efficient,
then
L
(16)
(r
it)=
E
Sii
j=1
(Fir),
where the 8 's are the coefficients in the following regression:
(17)
'Pit
=
io
+
E
8ij
ryt
+
r1it
j=1
L
vi
=
L
N.
(We will assume that the stochastic characteristics
of
nit
are the same as those of
git
in equation (1).) Conversely, (16) implies that some
portfolio of the given
L
portfolios is on the minimum variance frontier (Jobson and Korkie (1982)). Thus,
a necessary condition for the efficiency of a linear combination
(
r2t,...,
rLF)
with respect to the total set of N + L risky assets is:
(18)
Ho:
0
3io=
Vi
=
1,...,
N.
The above null hypothesis follows when the parameter restriction given by (16) is
imposed on (17).
In this
case, [T/N][(T-
N
-
L)/(T-
L
-
1)](1
+
Q2
-Irp)
l
2o'- 8
has a
noncentral F distribution with degrees of freedom N and (T
-
N
-
L), where
is a vector of sample means for
rpt
(i1t,
2t,...
Lt),
Q is the sample variance-
covariance matrix for
rpt,
So
has a typical element
Sio,
and
So
is the least squares
estimator for 83 based on the N regression equations in (17) above. Further, the
noncentrality parameter is given by [T/(1 + I-2
80)T-1
So.
Under the null
hypothesis (18), the noncentrality parameter is 0.
For an application of the methodology developed in this section, we return to
the results based on the size-sorted
portfolios using
returns
only during
the
month of January. In the previous section, we found that we could reject the ex
ante
efficiency
of the CRSP
Value-Weighted
Index. It could be that there exists a
linear combination of the lowest decile portfolio and the
Value-Weighted
Index
which is efficient. To consider such a case, we set L
=
2 and N
=
9.
(Since
portfolio
1
has become a
regressor
in a
system
like
(17),
we can no
longer
use it
as a
dependent variable.)
The F statistic to test
hypothesis (18)
is 1.09 with a
PORTFOLIO
EFFICIENCY
1147
at the usual
efficiency of this combination
p-value of 0.39, so we cannot reject
bias.
the obvious
pre-test
ignores
this inference
levels of significance. Of course,
riskless
that there
is an observable
this paper
we have assumed
Throughout
is simply
a line in mean-stan- case the efficient frontier
in which
rate of return,
a set of
now, that we wish to determine whether
space.
Suppose,
dard deviation
frontier determined
by
L + 1 portfolios (L
?
1) spans the
minimum-variance
The N + L + 1 asset
returns are assumed
and the N other assets.
these portfolios
If we observe the return on the "zero-beta" portfolio
independent.
to be linearly
(with
L = 1) naturally
we do not), this spanning hypothesis
(which in practice
of the CAPM due to Black
in the context of the zero-beta version (1972).25
arises
To formulate the test for spanning
for any L>
1,
consider the system of
regression equations,
L+1
(19)
(19
311,1
+
Rk=10
=
sio
+
E,
sji jt
Rit
j=1
it
=
1,...,5N,
Vi
i=,***,N
and Kandel
not excess returns.
Huberman
where
kit
denotes total returns,
(1985a) observe that the spanning
hypothesis
is equivalent to the following
restrictions:
(20)
3io=0
Vi=
... ,N
and
L+1
(21)
E
j=1
sij=l
Vi1
N.
in (19) and
letting
Fit
returns in excess
denote
(21) on the parameters
Imposing
of testing
of the returns on portfolio
L + 1, we derive
(17). Thus, the problem
case
(18) in the riskless
to that of testing
(20) in the context of (17) is identical
relevant
testing the riskless asset
case is equally
above.
All we have learned about
that "excess returns" are interpreted appro-
provided
to the spanning problem,
is
most importantly, the exact distribution of our test statistic
Perhaps
priately.
known under both the null and alternative hypotheses, permitting evaluation of
imposes (21) and then
the power of the test. Note that this test of spanning
in the
resulting
model are equal to
assesses whether the intercepts
regression
26
In contrast,
a joint F test of (20)
zero. Huberman and Kandel (1985a) propose
of the first
L
portfolios
with each
the
L
+ 1st portfolio is uncorrelated
More generally, suppose
A simple generalization of the
all such orthogonal portfolios.
variance among
and has minimum
for all
i.
It then follows
argument in Fama (1976,
page 373) establishes that
8,L+l=1-
:=
that the
L +
1
(1985a)
of Huberman and Kandel
on request) from
the results
(details are available
if and only if some combination of the first L
frontier
portfolios span the minimum-variance
hypothesis can be conducted as
in this
section
is on the frontier. Thus, a test of the latter
portfolios
orthogonal portfolio is observable.
that the minimum-variance
provided
26
An intermediate approach would
test (21) directly and then,
provided the null
is not
be to first
under
exact distribution
the
of (21) is an F test,
and
again, the
test
to test (20). Once
rejected, proceed
this test statistic
along
the lines
of our earlier analysis. Of course,
the alternative may
be determined
does require that
we observe the return on the L
+ 1
spanning portfolios.
25
1148
MICHAEL R. GIBBONS, STEPHEN
A. ROSS, AND JAY SHANKEN
and (21) against an unrestricted alternative; however, the distribution
of this
statistic has not been studied under the alternative.
AND FUTURE RESEARCH
8. SUMMARY
While this paper focuses on a particular hypothesis
from modern finance, this
apparently narrow view is adopted
to gain better insight about a broad class of
financial models which have a very similar structure to the
one that we examine.
The null hypothesis of this paper is a central hypothesis common
to all risk-based
The nature of financial data and theories suggests
the use
asset pricing theories.27
of multivariate statistical methods which are not necessarily
intuitive. We have
attempted to provide some insight into how
such tests function and to explain
why they may provide different answers relative to univariate
tests that are
applied in an informal manner.
In addition, we have studied the power of our
suggested statistic and have isolated factors
which will change the power of the
test. There are at least two natural extensions of this work,
and we now discuss
each in turn.
First, the multivariate test considered here requires that the number
of assets
under study always be less than the number of time series observations.
This
restriction is imposed so that the sample variance-covariance
matrix remains
nonsingular. A test statistic which could handle situations with a
large number of
assets would be interesting.28
Second, we have not been very careful to specify the information set on which
the various moments are conditioned. Gibbons and Ferson (1985), Grossman and
Shiller (1982), and Hansen and Singleton (1982, 1983) have emphasized the
importance of this issue for empirical work on positive models
of asset pricing.
Our methods provide a test of the ex ante unconditional efficiency of some
portfolio-that
is, when the opportunity set is constructed from
the uncondi-
tional moments, not the conditional moments.
When the riskless rate is changing
(as
it is in all of our data sets), then our methods provide a test of the conditional
efficiency of some portfolio given
the riskless rate. Of course, such an interpreta-
tion presumes that our implicit model for conditional moments given
the riskless
rate is correct. Ferson, Kandel, and Stambaugh (1987)
and Shanken (1987a)
provide more detailed analysis of testing
conditional mean-variance efficiency.29
If there is no riskless asset, then the null hypothesis becomes nonlinear
in the parameters, for the
intercept term is proportional
to (1
-
fl,p).
Gibbons (1982) has explored this hypothesis
using
statistics which only have asymptotic justification.
These statistics have been given an
elegant
geometric interpretation by
Kandel (1984). While we still do
not have a complete characterization of
the small sample theory, Shanken (1985,
1986) has provided some useful bounds for the
finite sample
behavior of these tests.
2X
See Affleck-Graves and McDonald (1988)
for some preliminary work on this problem.
29
As Hansen and Richard (1987)
emphasize, efficiency relative to
a given information set need not
imply efficiency relative to
a subset. This implication does
hold given some additional (and admittedly
restrictive) assumptions, however.
Let the information set, I, include the
riskless rate, and let p be
efficient, given I. Assume
betas conditional on I are constant and
9(Pi,
I
rp,
I) is linear in
rp,
It
=
lprpt,
where Rf, is
=
Rf,)
r,
follows that
(,t Ip,
I)
Irp,
iprpt,
and by iterated expectations
the riskless rate. Thus, p
is on the minimum-variance frontier, given Rf,, and the methods
of this
paper are applicable.
27
PORTFOLIO
EFFICIENCY
1149
Graduate School
of Business, Stanford
University,
Stanford, CA,
U.S.A.,
School of Organization and
Management,
Yale
University,
New
Haven, CT,
U. S.A.,
and
Simon School
of
Business
Administration, University of
Rochester, Rochester,
NY, U.S.A.
received
Manuscript
received
April, 1986;
final revision
1988.
November,
APPENDIX
DERIVATION OF
EQUATION (7)
To understand the derivation of
(7), first consider the basic portfolio
problem:
min: w'Vw
subject to w'r
=
m,
a mean
constraint,
where w
the vector of N
+ 1
portfolio weights;
V
the
variance-covariance matrix of N + 1
assets;
and r the vector of N + 1
sample
mean excess returns.
Without loss of generality, we assume that
p
itself is the first
component
of our excess return vector.
is a column vector of mean excess returns on the
original
N assets. The
Thus, r'
=
(r, i)
where
-5
first-order conditions for this
problem
are:
(22)
w
qV-
r
and
m
where
9p
is the
Lagrange multiplier. Hence,
mean
standard deviation
2
m2
w
w'V
m
2
Fr t
=
rV-lT
=
rV
0*2
Finally, to arrive at (7) we need to
establish that:
(23)
a
'
=*2
02
where in contrast to the rest of the
paper
S
is now the maximum likelihood
estimator. The last
and
S
and then
finding
V-
equality
follows from
rewriting
the elements of V in terms of
sp,
4p,
using the
formula for a
partitioned inverse. These
steps
lead to:
i[s
_
;1/.
4
I]
(24)
Then straightforward
algebra yields:
2-
l
r=
(-2/s2) +
[(
)-
-
)]
.
1150
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
Since
=2
-
fpp
and since the first term on the left-hand side of the above equation is
0*2
and the
we can rewrite the last equation as:
first term
on the right-hand side is
9p,
0*2
-
02 ? &p
p
p
- "&
p
or
2
2 Cp
ap
&=
*2
_
p
ap
Thus,
W=
0*2
_2
I
P
-2
1
and the equality
given in (7) has been justified.
REFERENCES
J.,
AND
B.
McDONALD
(1988): "Multivariate Tests of Asset Pricing: The Compar-
AFFLECK-GRAVES,
ative Power of Alternative Statistics," Working Paper, University
of Notre Dame.
to Multivariate Statistical Analysis, Second Edition. New
ANDERSON,
T. W. (1984): An Introduction
York: John Wiley.
BANZ,
R. W. (1981): "The Relationship between Return and
Market Value of Common Stock,"
Journal of Financial Economics, 9,
3-18.
BERNDT,
E. R.,
AND
N. E.
SAVIN
(1977): "Conflict among Criteria for Testing Hypotheses in
the
45, 1263-1277.
Multivariate Linear Regression Model," Econometrica,
Journal of Business,
BLACK,
F. (1972): "Capital Market Equilibrium
with Restricted Borrowing," The
45,
444-454.
BLACK,
F., M. C.
JENSEN, AND
M.
SCHOLES
(1972): "The Capital Asset Pricing Model: Some
of Capital Markets, ed. by M. C. Jensen. New York:
Empirical Findings," in Studies in the Theory
Praeger.
(1987): "Empirical Tests of the
BREEDEN,
D. T., M. R.
GIBBONS, AND
R. H.
LITZENBERGER
Consumption-Oriented CAPM," Research Paper #879, Graduate
School of Business, Stanford
University, Stanford, CA.
J. M. (1978): "Mixtures of Beta Distributions, Algorithm AS 123,"
Applied Statistics, 27,
BREMNER,
104-109.
BROWN,
P., A. W.
KLEIDON,
AND
T. A.
MARSH
(1983):
"New Evidence on the Nature of Size-Related
Anomalies in Stock Prices," Journal of Financial Economics, 12,
33-56.
P. H.,
AND
S. A. Ross (1985): "The Analytics of Performance Measurement Using
a
DYBVIG,
Security
Market Line," Journal of Finance, 40, 401-416.
ELTON,
E. J.,
AND
M. J.
GRUBER
(1973): "Estimating
the Dependence Structure of Share Prices-
Implications for Portfolio Selection," Journal of Finance, 28,
1203-1232.
ELTON,
E. J., M. J.
GRUBER, AND
T. J.
URICH
(1978):
"Are Betas Best?" Journal of Finance, 33,
1375-1384.
of Finance.
New York: Basic Books.
FAMA,
E. F. (1976): Foundations
FAMA,
E. F.,
AND
J. D.
MAcBETH
(1973): "Risk, Return,
and Equilibrium: Empirical Tests," Journal
of Political Economy, 81,
607-636.
(1987):
"Tests of Asset
Pricing with Time-Varying
FERSON, W.,
S.
KANDEL, AND
R.
STAMBAUGH
and Market Betas," Journal of Finance, 42, 201-220.
Expected Risk Premiums
GIBBONS,
M. R. (1980): "Estimating the Parameters of the Capital Asset Pricing
Model-
A Minimum Expected Loss Approach,"
Research
Paper #565,
Graduate School of Business,
Stanford University, Stanford,
CA.
(1982): "Multivariate Tests of Financial
Models: A New Approach," Journal of Financial
Economics, 10, 3-28.
(1985): "Testing Asset Pricing Models
with Changing Expectations
GIBBONS,
M. R.,
AND
W.
FERSON
and an Unobservable Market Portfolio," Journal of
Financial Economics, 14, 217-236.
GRAYBILL,
F. A. (1983): Matrices with Applications in Statistics, Second Edition. Belmont,
CA:
Wadsworth.
PORTFOLIO EFFICIENCY
1151
S.,
AND
R.
SHILLER
(1982): "Consumption Correlatedness and Risk Measurement in
GROSSMAN,
Economies with Non-Traded Assets and Heterogeneous
Information,"
Journal of Financial Eco-
nomics, 10,
195-210.
HANSEN,
L.,
AND
S.
RICHARD
(1987): "The Role of Conditioning Information in Deducing Testable
Restrictions Implied by Dynamic Asset Pricing
Models," Econometrica,
55, 587-614.
HANSEN,
L.,
AND
K.
SINGLETON
(1982): "Generalized Instrumental Variables Estimation of Nonlin-
ear Rational Expectations
Models," Econometrica, 50,
1269-1286.
(1983): "Stochastic
Consumption,
Risk Aversion, and the Temporal Behavior of Stock Market
Returns," Journal of Political
Economy,
91,
249-265.
HUBERMAN,
G.,
AND
S.
KANDEL
(1985A): "Likelihood Ratio Tests of Asset Pricing and Mutual Fund
Separation," Graduate School of Business, University
of
Chicago, Chicago,
IL.
(1985b): "A Size Based Stock Returns Model," Graduate School of Business, University of
Chicago, Chicago,
IL.
(1979): Stocks, Bonds, Bills and Inflation: Historical Returns
IBBOTSON,
R.,
AND
R.
SINQUEFIELD
(1926-1978).
Charlottesville,
VA: Financial Analysts Research Foundation.
JOBSON,
J.
D.,
AND
B.
KORKIE
(1982): "Potential Performance and Tests of Portfolio Efficiency,"
Journal of Financial
Economics, 10,
433-466.
(1985): "Some Tests of Linear Asset Pricing with Multivariate
Normality," Canadian
Journal
of Administrative Sciences, 2, 114-138.
JOHNSON,
N. L.,
AND
S. KOTZ (1970): Continuous Univariate Distributions, Volume 2. New York:
John Wiley.
KANDEL,
S. (1984): "The Likelihood Ratio Test Statistic of Mean-Variance Efficiency without a
Riskless
Asset," Journal of Financial Economics, 13, 575-592.
KANDEL, S., AND
R.
STAMBAUGH
(1987): "On Correlations and Inferences about Mean-Variance
Efficiency," Journal of Financial Economics, 18, 61-90.
KEIM,
D. B. (1983): "Size-Related Anomalies and Stock Return Seasonality: Further
Empirical
Evidence," Journal of Financial Economics, 12, 13-32.
LINTNER,
J. (1965): "The Valuation of Risk Assets and the Selection of Risky Investment in Stock
Portfolios and Capital
Budgets," Review of Economics
and
Statistics, 47,
13-37.
MAcBETH,
J. (1975): "Tests of Two Parameter Models of Capital Market
Equilibrium,"
Ph.D.
Dissertation, Graduate School of Business, University of Chicago, Chicago, IL.
MACKINLAY,
A. C. (1985): "An Analysis of Multivariate Financial
Tests," Ph.D. Dissertation,
Graduate School of Business, University of
Chicago, Chicago,
IL.
(1987): "On Multivariate Tests of the CAPM," Journal of Financial
Economics, 18, 341-372.
MALINVAUD,
E. (1980): Statistical Methods of Econometrics, Third Edition. Amsterdam: North
Holland Publishing.
MERTON,
R. (1973): "An Intertemporal Capital Asset Pricing
Model," Econometrica,
41, 867-887.
Statistical Methods, Second Edition. New York: McGraw-Hill.
MORRISON,
D. F.
(1976): Multivariate
RAo, C. R. (1951): "An Asymptotic Expansion of the Distribution of Wilks Criterion," Bulletin
of the
Institute of International
Statistics, 33, 33-38.
REINGANUM,
M. R. (1981): "Misspecification of Capital Asset Pricing: Empirical
Anomalies Based
on Earnings Yields and Market Values," Journal
of
Financial
Economics, 9,
19-46.
ROLL,
R.
(1977): "A Critique of the Asset Pricing Theory's Tests-Part
1: On Past and Potential
Testability of the Theory," Journal of Financial Economics, 4, 129-176.
(1978): "Ambiguity When Performance is Measured by the Securities Market
Line,
Journal
of Finance, 33, 1051-1069.
SCHWERT,
G. W. (1983): "Size and Stock
Returns,
and Other Empirical
Regularities," Journal of
Financial Economics, 12, 3-12.
SHANKEN,
J. (1982): "An Analysis of the Traditional Risk-Return Model," Ph.D.
Dissertation,
Graduate School of Industrial Administration,
Carnegie-Mellon University, Pittsburgh, PA.
(1985): "Multivariate Tests of the Zero-Beta CAPM," Journal of Financial Economics, 14,
327-348.
(1986): "Testing Portfolio Efficiency
when the Zero-Beta Rate Is Unknown: A
Note," Journal
of Finance, 41,
269-276.
(1987a): "The Intertemporal Capital
Asset
Pricing Model: An Empirical Investigation,"
Working Paper, University
of Rochester.
(1987b): "Multivariate Proxies and Asset Pricing Relations: Living with the Roll Critique,"
Journal of Financial Economics, 18,
91-110.
SHARPE,
W. F. (1963): "A Simplified Model of Portfolio Analysis," Management
Science, 9, 277-293.
1152
MICHAEL R. GIBBONS, STEPHEN A. ROSS, AND JAY SHANKEN
(1964): "Capital Asset Prices: A Theory of Market Equilibrium under
Conditions of Risk,"
Journal of
Finance, 19, 425-442.
(1982): "Factors in New York Stock Exchange Security
Returns, 1931-79," Journal of
Portfolio
Management, 8,
5-19.
SILVEY,
S. D. (1975): Statistical Inference. London: Chapman and Hall.
R. F. (1982): "On the Exclusion of Assets from Tests of the Two-Parameter
Model:
STAMBAUGH,
A Sensitivity
Analysis," Journal of
Financial
Economics, 10,
237-268.
TANG, P. C. (1938): "The Power Function of the Analysis of Variance
Tests with Tables and
Illustrations of Their
Use,"
Statistical Research
Memoirs, 2, 126-150.
TIKU, M. L. (1967): "Tables of the Power of the F-Test," Journal of the American Statistical
Association, 62, 525-539.
J.
L.,
AND F. BLACK (1973): "How to Use Security Analysis to
Improve Portfolio
TREYNOR,
Selection," Journal of Business, 46, 66-86.
发布评论