THE JOURNAL OF FINANCE • VOL. LIV, NO.

2 • APRIL 1999

The Sampling Error in Estimates of

Mean-Variance Efficient Portfolio Weights

MARK BRITTEN-JONES*

ABSTRACT
This paper presents an exact finite-sample statistical procedure for testing hypoth-
eses about the weights of mean-variance efficient portfolios. The estimation and
inference procedures on efficient portfolio weights are performed in the same way
as for the coefficients in an OLS regression. OLS t- and F-statistics can be used for
tests on efficient weights, and when returns are multivariate normal, these sta-
tistics have exact t and F distributions in a finite sample. Using 20 years of data
on 11 country stock indexes, we find that the sampling error in estimates of the
weights of a global efficient portfolio is large.

MEAN-VARIANCE ANALYSIS IS IMPORTANT for both practitioners and researchers in

finance. For practitioners, theory suggests that mean-variance efficient1 port-
folios can play an important role in portfolio management applications. For re-
searchers in finance, mean-variance analysis is central to many asset pricing
theories as well as to empirical tests of those theories; however, practitioners
have reported difficulties in implementing mean-variance analysis.2 For ex-
ample, Black and Litterman ~1992! note that, “when investors have tried to
use quantitative models to help optimize the critical allocation decision, the
unreasonable nature of the results has often thwarted their efforts” ~p. 28!.
The main difficulty concerns the extreme weights that often arise when
sample efficient portfolios are constructed. This phenomenon is studied by
Best and Grauer ~1991!, who show that sample efficient portfolios are ex-
tremely sensitive to changes in asset means. The sensitive nature of port-
folio weights implies that sampling error in estimates of asset means and
covariances feeds through to the estimates of efficient portfolio weights. This
has prompted a variety of ‘fixes’ in order to reduce the inf luence of sampling
error, ranging from direct and extensive restrictions on portfolio weights
~see, e.g., Haugen ~1997!! to Bayesian shrinkage methods ~see Jorion ~1985!,
Frost and Savarino ~1986, 1988!, and Black and Litterman ~1992!!.

* London Business School. I thank Mark Grinblatt and Olivier Ledoit for many helpful dis-
cussions, and for pointing out an error in an earlier draft. I also thank the referee, Jay Shan-
ken, for helpful comments, and the editor, René Stulz, for suggesting an international focus for
the empirical section. Remaining errors are of course my responsibility.
1
Hereafter, the term efficient refers to mean-variance efficient.
2
Academics have also reported such difficulties. See Kroll, Levy, and Markowitz ~1984!,
Pulley ~1981!, and, in a slightly different context involving conditional moments, Gallant, Hansen,
and Tauchen ~1990!.

655
656 The Journal of Finance

Despite the important ~and possibly damaging! role of sampling error in

the construction of sample efficient portfolios, little formal analysis of sam-
pling error in the context of efficient portfolio weights has been published.
Jobson and Korkie ~1980! derive an asymptotic distribution for estimates of
efficient weights; however, Ledoit ~1995! and Gibbons, Ross, and Shanken
~1989! show that asymptotic results can be misleading when the number of
assets is not small. Ledoit ~1995! develops a general asymptotic theory that
bypasses some problems with existing asymptotic theory when the number
of assets is large, but his focus is the covariance matrix of returns, not port-
folio weights. A variety of Bayesian approaches have been developed for pur-
poses of testing the efficiency of a given portfolio, but these works have not
focused on the sampling error in estimates of efficient portfolio weights.3
For researchers in finance, mean-variance analysis can play an important
role in tests of asset pricing theories. For example, tests of the CAPM have
traditionally focused on the expected return-beta relation. But recent re-
search by Roll and Ross ~1994! and Kandel and Stambaugh ~1995! shows
that the linear relation between expected return and beta is not robust to
slight inefficiency in the market portfolio. These studies suggest that CAPM
tests should be based on the CAPM’s central prediction—the mean-variance
efficiency of the market portfolio.4 Fama ~1996! and Grinblatt and Titman
~1983! show that multifactor asset pricing models can also be tested by ex-
amining the mean-variance efficiency of linear combinations of certain key
portfolios.
Traditionally, researchers have thought about portfolio efficiency in terms
of returns. Such an approach is complicated by the fact that the location ~in
mean-variance space! of both the test portfolio and the efficient frontier are
subject to sampling error.5 For some purposes, a simpler way of thinking
about portfolio efficiency is in terms of portfolio weights. We present proce-
dures and test statistics enabling formal statistical inference on efficient
portfolio weights. We show that the Gibbons, Ross, and Shanken ~1989! ~GRS!
F-test of portfolio efficiency can be thought of ~and derived! as a test of the
restriction that the weights of the ‘tangency’ portfolio equal the weights of
the test portfolio. Rejection of this restriction implies rejection of the effi-
ciency of the test portfolio.
The tests we present can be implemented in a simple and intuitive man-
ner that is formally identical to the standard inference procedures used for
OLS regression coefficients. One result is a simpler implementation method
for the GRS F-test. For practitioners this procedure could be used to assess

3
See Kandel, McCulloch, and Stambaugh ~1995!, McCulloch and Rossi ~1990, 1991!, Harvey
and Zhou ~1990!, and Shanken ~1987!.
4
These studies also support the use of a generalized least squares ~GLS! regression to an-
alyze the linear relation between expected return and beta. However, the goodness-of-fit of this
regression is related in a simple manner to the degree of mean-variance efficiency of the mar-
ket portfolio. In other words, the GLS approach boils down to testing the efficiency of the
market portfolio.
5
See Roll ~1985! and Gibbons, Ross, and Shanken ~1989!.
 Sampling Error in Mean-Variance Efficient Portfolio Weights 657

the precision of the weights of a sample efficient portfolio. For researchers

in finance, this procedure provides a simple and f lexible way of testing a
given portfolio’s efficiency, and of testing hypotheses concerning the weights
of efficient portfolios.
In Section I we present the ‘regression’ procedure for inference on portfolio
weights. In Section II we demonstrate the procedure by using international
stock return data to analyze the weights of an international efficient port-
folio. Section III concludes. Technical proofs are contained in the Appendix.

I. The Regression Approach to Portfolio Analysis

In this paper a riskless asset is assumed available for both borrowing and
lending in each period. Excess returns are calculated by subtracting the re-
turn of this riskless asset from the total return.6 There are K risky assets
indexed by k 5 1, . . . , K. The excess returns on the K assets in some period t
in ~1, . . . ,T ! are denoted by the K elements of the vector x t :

x 't 5 @x 1t , . . . x kt , . . . , x Kt # . ~1!

The T observations of excess returns are contained in the T 3 K matrix X:

34
x '1
X5 I . ~2!
x 'T

Note that a portfolio of risky assets and a riskless asset has an excess return
that is determined solely by the weights and excess returns of the risky
assets. Thus, given a K-vector of risky asset weights b, the excess return of
this portfolio in period t is simply x 't b, where the weights in b need not sum
to one.
Let l represent a vector of ones with length conforming to the rules of
matrix algebra. Viewed as a portfolio excess return, the T-vector of ones l is
highly desirable as it has positive excess return with zero sample standard
deviation. The regression approach to portfolio selection7 is based on mini-
mizing the squared deviations between the excess returns on a constructed
portfolio and the excess returns in l. This minimization problem can be per-

6
Hereafter, the term return refers to excess return unless otherwise stated.
7
Cochrane ~1997!, building on the work of Hansen and Richards ~1987!, shows that the
identification of stochastic discount factors can be analyzed in terms of projections onto asset
payoffs without an intercept term. Cochrane’s results are in terms of population moments,
whereas the focus here is on statistical inference when sample moments are used. Future re-
search could attempt to analyze and identify stochastic discount factors using the statistical
inference procedures developed here.
658 The Journal of Finance

formed using an artificial ordinary least squares ~OLS! regression and the
following proposition states that such a regression recovers the weights of a
sample efficient portfolio.

THEOREM 1: OLS regression of a constant l onto a set of asset’s excess returns

X, without an intercept term,

l 5 Xb 1 u,
~3!
~T 3 1! ~T 3 k! ~k 3 1! ~T 3 1!

results in an estimated coefficient vector

bZ 5 ~X ' X!21 X ' l, ~4!

that is a set of risky-asset-only portfolio weights for a sample efficient

Z ' bZ
portfolio. The scaled (so that weights sum to one) coefficient vector b0l
is thus the familiar tangency portfolio

SO 21 xS
, ~5!
l ' SO 21 xS

derived from quadratic programming, where the sample mean xS 5 X ' l0T, and
the (maximum likelihood) sample covariance SO 5 ~X 2 lxS ' !' ~X 2 lxS ' !0T, are
used as parameters.
Proof: Using the updating formula for an inverse matrix,8 express the
coefficient vector bZ from the regression in equation ~3! in terms of the sam-
ple mean xS and sample covariance SO :

bZ 5 ~X ' X!21 X ' l

5 ~ SO 1 xS xS ' !21 xS

S
5 SO 21 2
SO 21 xS xS ' SO 21
1 1 xS ' SO 21 xS
D
xS

SO 21 xS
5 . ~6!
1 1 xS ' SO 21 xS

Scaling bZ so that the coefficients sum to one results in the tangency portfolio

bZ SO 21 xS
5 ~7!
l ' bZ l ' SO 21 xS

when sample means and covariances are used as parameters. Q.E.D.

8
See Greene ~1993!, p. 25.
 Sampling Error in Mean-Variance Efficient Portfolio Weights 659

Figure 1. Sample mean standard deviation diagram. The point b is the point on the line
0d that is closest to the point ~0,1! and the point a is the point on the line 0m that is closest to
~0,1!.

The regression in equation ~3! is unusual. There is no intercept, the de-

pendent variable is nonstochastic, and the residual vector u is correlated
with the regressors, which are stochastic. However, the regression has a
simple interpretation: The dependent variable l is a sample counterpart to
arbitrage profits—positive excess return with zero standard deviation; the
coefficients b represent the weights on risky assets in the portfolio; Xb rep-
resents excess returns on this portfolio; and the residual vector u shows
deviations in this portfolio’s return from 1.
The estimated portfolio weights bZ produce a portfolio return vector that is
closest in terms of least squares distance to the arbitrage return vector l.
This least squares distance can be illustrated using the familiar mean-
standard deviation diagram. The feasible set, constructed from the sample
mean and ~maximum-likelihood! sample covariance, has an efficient bound-
ary shown by the line 0d from the origin passing through the tangency port-
folio ~Figure 1!. The arbitrage return vector l is located at the point ~0,1!.
The residuals u can be split into a mean component9 ul S and a deviation
component e:

u 5 ul
S 1 e, ~8!

9
The lack of an intercept in the regression implies residuals need not sum to zero.
660 The Journal of Finance

where uS 5 ~10T !u ' l. Similarly, the mean squared residual ~MSR! can be split
into a squared mean and a squared standard deviation ~SD!:

MSR 5 Mean 2 1 SD 2

u ' u0T 5 uS 2 1 e ' e0T. ~9!

The residual’s mean uS equals the difference between the constructed port-
folio’s average return and 1; thus the constructed portfolio’s mean return
equals 1 2 u:S

xS ' bZ 5 1 2 u,
S ~10!

and this is measured on the vertical axis of Figure 1.

Since the residual and constructed portfolio excess returns sum to one, the
residuals’ standard deviation, SD, equals the constructed portfolio’s stan-
dard deviation, and this is measured on the horizontal axis. By Pythagoras’
theorem the root mean square residual equals the distance between the port-
folio’s location in mean–standard deviation space and the arbitrage portfo-
lio’s location ~0,1!. OLS thus finds a portfolio whose returns are located as
closely as possible in mean–standard deviation space to the point ~0,1!.
Such a portfolio is shown in Figure 1 by the point b. Note that this port-
folio is sample efficient.
The regression framework highlights the stochastic nature of optimal
weights calculated from sample data. Our next theorem states that the stan-
dard OLS F-statistic associated with a particular linear restriction applied
to the regression in ~3! implements the GRS F-test.
THEOREM 2: The GRS F-test for the efficiency of a given portfolio with weights
g can be implemented by the standard OLS F-statistic associated with the
linear restriction that portfolio weights in regression (3) are proportional to
the weights of the given portfolio. Under multivariate normality the F(OLS)-
statistic,

~SSR r 2 SSR u !0~K 2 1!

F~OLS! 5 , ~11!
SSR u 0~T 2 K !

is distributed in finite sample as central FK21,T2K where SSR u is the sum of

squared residuals from the unrestricted regression in (3), SSR r is the sum of
squared residuals from the regression in equation (3) estimated subject to the
restriction that b } g (g is a vector containing the weights of the given port-
folio), and T is the number of observations.
Proof: This theorem can be proved using Theorem 3 and its proof, but a
diagrammatic proof is simpler and provides greater economic intuition. In
the context of the regression in equation ~3!, consider the K 2 1 restrictions
 Sampling Error in Mean-Variance Efficient Portfolio Weights 661

that the weights b be proportional to the weights of the given portfolio g.

The proportionality restriction can be written as b 5 a 3 g, where a is a free
scalar. The restricted regression can thus be written as a regression with a
single independent variable, the given portfolio’s excess returns Xg, and a
single regression coefficient a:

1 5 ~Xg!a 1 u *. ~12!

This restriction is shown in Figure 1 by the line Om from the origin through
the given portfolio’s position shown as m.
The root mean squared residual associated with the restricted and un-
restricted regressions is given by the minimum distances between ~0,1! and
the lines Om and Od. The F~OLS!-statistic can thus be written

F~OLS! 5 SS D DS D
1a
R
1b
R 2
21
T2K
K 21
, ~13!

R and 1b
where 1a R are the distances shown in the diagram between 1 and a
and between 1 and b, respectively. Now GRS note that their test statistic,
which they denote W, is given by

W5 F GR
Od
Oc R
2
2 1, ~14!

where OdR and Oc

R are the distances shown in Figure 1. The point d shows the
average return attainable from taking a position on the efficient frontier
with a standard deviation of one and c shows the expected return attainable
from taking a position in the given portfolio with standard deviation of one.
Note that10

R
Od R
1a
5 . ~15!
Oc R 1b R

Thus the F~OLS!-statistic is proportional to GRS’ W-statistic:

F~OLS! 5 W S D
T2k
k 21
. ~16!

Under the assumption of multivariate normality, GRS show that W~T 2 K !0

~K 2 1! is distributed in a finite sample as a central F distribution with K 2 1

10
To see this write Od0R OcR as cos cO10cos
Z Z
dO1, where the angles are measured to the hori-
R 1bR can be expressed as cos O1a0cos
zontal axis. Similarly, 1a0 Z Z
O1b, where the angles are mea-
sured from the vertical axis. Since the angles cO1Z and O1a
Z are equal, and the angles dO1
Z and
Z are also equal, the equality holds.
O1b
662 The Journal of Finance

and T 2 K degrees of freedom. It therefore follows that, with multivariate nor-

mality, the F~OLS!-statistic in finite sample has an exact central F distribu-
tion with K 2 1 and T 2 K degrees of freedom:

F~OLS! ; Fk21,T2k . ~17!

Q.E.D.

It is worth pausing to examine the different interpretations of the GRS

F-test. The first interpretation ~which underlies the derivation and proof
contained in GRS! is as a joint test of zero intercepts in a multivariate re-
gression. The second interpretation is as a ratio formed from the squared
Sharpe ratios of the sample efficient portfolio and of the test portfolio. The
interpretation we offer is as a test of the restriction that the weights of the
tangency portfolio are proportional to the weights of the test portfolio. This
test can then be implemented using a standard OLS F-test for testing a set
of linear restrictions on regression coefficients. The next section shows that
we can extend this regression analogy further to examining other linear
restrictions.

A. Testing the Efficiency of a Linear Combination of Assets

Jobson and Korkie ~1989! and GRS show, in the context of testing the
efficiency of a subset of assets relative to a full set of assets, that an analo-
gous test statistic to the GRS F-test, based on the ratio of maximal Sharpe
ratios from the subset and the complete set of assets, has an exact central F
distribution under the null hypothesis of efficiency of the subset. This is an
extension of the GRS F-test, and the following proposition extends this re-
sult to any hypothesis that can be expressed as a linear restriction on effi-
cient portfolio weights.
THEOREM 3: Denoting a set of efficient (with respect to population moments)
portfolio weights by b, we can test the linear restriction

Rb 5 0, ~18!

where R is a ~q 3 K! matrix of restrictions, by the standard OLS F-statistic

associated with the restriction Rb 5 0 applied to the regression in equa-
tion (3). Thus,

~SSR r 2 SSR u !0q

, ~19!
SSR u 0~T 2 K !

has an exact central F distribution with q and T 2 K degrees of free-

dom. SSR u is the sum of squared residuals from the unrestricted regression
(3), SSR r is the sum of squared residuals from estimation of regression (3)
 Sampling Error in Mean-Variance Efficient Portfolio Weights 663

subject to the restriction Rb 5 0, q is the number of linear restrictions (the

number of rows of R ), K is the number of assets, and T is the number of
observations.
Proof: In the Appendix. Q.E.D.

B. Exact Statistical Inference on Efficient Portfolio Weights

Theorem 3 enables simple derivations of formal inference procedures for
efficient portfolio weights. The following corollary shows how a standard
t-test can be used to test whether an asset’s weight in an ex ante efficient
portfolio is different from zero.
COROLLARY 1: Denoting a set of ex ante efficient portfolio weights (for the case
of a riskless asset) by b, we can test whether the ex ante efficient weight bi on
a particular asset i is zero by testing the restriction

bi 5 0, ~20!

using the standard OLS t-statistic associated with this restriction in the re-
gression in equation (3). Thus,

bZ i
~21!
s!a ii

has an exact t distribution with T 2 K degrees of freedom, where s 2 is the

usual estimate of residual variance:

s 2 5 SSR u 0~T 2 K !, ~22!

and aii is the i th element on the principal diagonal of ~X ' X!21.

Proof: In the Appendix. Q.E.D.
Theorem 3 provides a very simple way of conducting formal exact statis-
tical inference on the weights of the efficient portfolio. The usual t- and
F-statistics associated with OLS regression can be used in the context of
regression equation ~3!, and with multivariate normality, these statistics
have exact t and F distributions in finite sample.
The only pitfall concerns the nature of testable hypotheses. The single-
regression procedure identifies sample efficient portfolio weights only up to
a scaling factor ~ref lecting the fact that the set of efficient portfolios is a
straight line in mean–standard deviation space!. So restrictions need to be
imposed in the form of proportionality constraints. For example, in order to
test the hypothesis that the tangency portfolio’s weight on the first asset is
equal to a number, say 0.1, one cannot simply test the restriction

b1 5 0.1. ~23!
664 The Journal of Finance

A meaningful restriction is that the scaled weight is equal to the number;

that is,

b1
5 0.1, ~24!
l' b

and this restriction is implemented by

b1 2 0.1 3 l ' b 5 0. ~25!

This restriction is still linear and can therefore be implemented using Theo-
rem 3. Consider the proportionality restriction in Theorem 2: b } g where
g ' 5 ~ g1 , . . . , gK ! is the vector of weights on the given portfolio. Since g ' 1 5 1,
this restriction can be expressed as

b1
5 g1 ~26!
l' b

b2
5 g2 ~27!
l' b

bk21
5 gK21 , ~28!
l' b

which is written in standard matrix notation as Rb 5 0, where the K 2 1 by

K restriction matrix R is

1 2 g1 2g1 2g1 J 2g1

R5
3 2g2
I
2gK21
1 2 g2
I
2gK21
2g2

J
L
J
L
1 2 gK21
2g2
I
2gK21
4 . ~29!

II. An International Mean-Variance Efficient Portfolio

The composition of an international efficient portfolio is of interest, not
only from the practical viewpoint of investors seeking to optimize their risk-
return tradeoff, but also from a theoretical viewpoint. International asset
pricing theories can have implications for the composition of a globally ef-
ficient portfolio,11 and the composition of such a portfolio can alert us to the
presence and importance of barriers to international investment.

11
A useful survey of the portfolio implications of various international asset pricing theories
is Stulz ~1995!.
 Sampling Error in Mean-Variance Efficient Portfolio Weights 665

A number of papers have constructed ex post global efficient portfolios,12

but these studies have been unable to formally assess the statistical relia-
bility of their results. The problem of statistical inference for this type of
analysis is well summarized by Adler and Dumas ~1983!, p. 945, who write
~concerning estimation of the weights of an international efficient portfolio
they call the log-portfolio!:

“. . . the estimates are plagued by major statistical problems which un-

dermine their significance. No statistical theory, to our knowledge, gives
the sample distribution of the estimated wlog @weights of international
efficient portfolio#. . . . We are, therefore, unable to build confidence in-
tervals for the optimal log-portfolio composition. . . .”
In this section, we use the inference procedures developed in the first
section to examine the magnitude and effect of sampling error in estimates
of the composition of an efficient international equity portfolio. Table I shows
the weights of an ex post international tangency portfolio for a U.S. investor.
We use monthly data from Morgan Stanley Capital International ~MSCI! for
the 20-year period from January 1977 to December 1996, for the equity
markets of 11 developed countries.13 Weights for the full 20-year period are
shown as well as weights for two 10-year subperiods. The estimates contain
several extreme positions such as a short position of 45 percent of portfolio
value in the Canadian market, and a short position of 18 percent in the
German market. Note that many weights change dramatically between the
two subperiods. For example, in the first subperiod the optimal weight in
Denmark is a short position of 29.6 percent, but in the second subperiod the
optimal weight in Denmark is a long position of 68.8 percent.
Using the statistical inference procedures developed in the first part of this
paper, we construct t-statistics for testing the hypothesis that a country’s weight
is zero. As pointed out earlier, hypotheses must be expressed in terms of a pro-
portional relation, and the zero restriction is very simply implemented. The
standard errors we show are derived imposing the null hypothesis. Thus they
are constructed by dividing the estimated weight ~scaled to sum to one! by the
associated t-statistic. This provides a useful indication of the sampling error,
but a formal confidence interval cannot be constructed for one weight alone
due to the required proportional nature of testable hypotheses.
Over the full 20-year period, the standard errors of the estimates are large.
The smallest standard error is for Japan at slightly more than 20 percent of
portfolio value. The largest standard error is for the United States, which at
47.0 percent is almost half of total portfolio value. Not surprisingly, the stan-
dard errors are generally larger in the two 10-year subperiods. The result of
such statistical imprecision is that none of the zero-weight restrictions can
be rejected at the standard significance level of 0.05.

12
See, for example, Adler and Dumas ~1983! and Solnik ~1982!.
13
We do not examine the issue of currency hedging, but assume all foreign positions are
unhedged with regard to currency exposure.
666 The Journal of Finance

Table I
Estimates of a Global Tangency Portfolio
This table contains estimates of the weights of a global tangency portfolio, from the viewpoint
of a U.S. investor. The monthly country returns are the Morgan Stanley Capital International
~MSCI! country stock index returns converted into U.S. dollar returns using the MSCI ex-
change rate series. Excess returns are calculated by subtracting the one-month T-bill return
obtained from Ibbotson Associates. Weights are in percentage form ~i.e., they sum to 100!. The
t-statistics test the hypothesis that the country weight is zero. Standard errors ~SE! are calcu-
lated under the null hypothesis of a zero weight, as the estimated weight divided by the t-statistic.

1977–1996 1977–1986 1987–1996

Weight Weight Weight

~t-statistic! SE ~t-statistic! SE ~t-statistic! SE

Australia 12.8 23.6 6.8 33.6 21.6 32.6

~0.54! ~0.20! ~0.66!
Austria 3.0 25.5 29.7 44.1 22.5 30.3
~0.12! ~20.22! ~0.74!
Belgium 29.0 35.1 7.1 46.8 66.0 54.5
~0.83! ~0.15! ~1.21!
Canada 245.2 38.9 232.7 51.0 268.9 62.7
~21.16! ~20.64! ~21.10!
Denmark 14.2 30.2 229.6 45.3 68.8 38.7
~0.47! ~20.65! ~1.78!
France 1.2 28.7 20.7 37.3 222.8 47.3
~0.04! ~20.02! ~20.48!
Germany 218.2 35.4 9.4 49.4 258.6 52.1
~20.51! ~0.19! ~21.13!
Italy 5.9 20.2 22.2 27.9 215.3 29.5
~0.29! ~0.79! ~20.52!
Japan 5.6 23.4 57.7 40.4 224.5 28.1
~0.24! ~1.43! ~20.87!
U.K. 32.5 32.1 42.5 42.9 3.5 49.8
~1.01! ~0.99! ~0.07!
U.S. 59.3 47.0 27.0 65.2 107.9 70.6
~1.26! ~0.41! ~1.53!

In order to examine the question of the benefits of global diversification

for a U.S. investor, we construct F-statistics for the restriction that weights
on all foreign ~i.e., non-U.S.! countries’ equity markets are zero. For the full
20-year period the F-statistic is 0.59 with an associated p-value of 0.82. For
the 1977 to 1986 subperiod the F-statistic is 0.81 with a p-value of 0.62, and
for the 1987 to 1996 subperiod the F-statistic is 0.69 with a p-value of 0.73.
Based on the data, we are unable to reject the hypothesis that the tangency
portfolio for a U.S. investor has no exposure to foreign equity markets.
These results provide no statistical support for the proposition that there
are benefits to global diversification for a U.S. investor. This is in line with
a recent study by Sinquefield ~1996!. However the magnitude of sampling
error tells us that the data ~by themselves! actually provide little informa-
 Sampling Error in Mean-Variance Efficient Portfolio Weights 667

tion for international portfolio construction. A possible remedy is to allow

prior information, perhaps derived from theory, to inf luence the estimates of
an efficient portfolio. Such an approach is pursued, within a Bayesian frame-
work, by Black and Litterman ~1992! who combine an international CAPM
with historical data to generate portfolio weights.

III. Conclusions and Further Research

In summary, the regression approach to portfolio analysis provides a new
and simple tool for the empirical analysis of mean-variance problems. Using
this approach, we derive exact formal inference procedures for hypotheses
about the weights of efficient portfolios. Furthermore, we show how to im-
plement the GRS F-test for portfolio efficiency using linear restrictions on a
single linear OLS regression. Finally, we use these inference procedures to
show the importance and magnitude of sampling error in estimates of the
weights of an international mean-variance efficient portfolio.
Further research could attempt to apply some of the regression estimation
techniques and inference procedures that are designed to deal with nonnor-
mal and nonindependent data. Another line of research could seek to extend
the approach developed here to the analysis of conditional mean-variance
efficiency14 by expanding the asset universe to include ‘managed’ portfolios
~excess returns scaled by instrumental variables!.

Appendix

Proof of Theorem 3: The sum of squared residuals ~SSR! from the un-
restricted regression

l 5 Xb 1 u ~A1!

is given by

SSR 5 l ' Ml, ~A2!

where

M 5 I 2 X~X ' X!21 X '. ~A3!

The restriction Rb 5 0 can be written as

R 1 b 1 1 R 2 b 2 5 0, ~A4!

14
See Cochrane ~1996! and Hansen and Richard ~1987!.
668 The Journal of Finance

where R 5 ~R 1 , R 2 !, b ' 5 ~b '1 , b '2 !, R 1 is a ~q 3 q! nonsingular matrix, and R 2

is a q 3 ~k 2 q! matrix. Solving for b 1 gives

b 1 5 2R21
1 R2 b2 . ~A5!

The restricted regression

l 5 Xb 1 u, ~A6!

Rb 5 0, ~A7!

can be rewritten with the restriction directly imposed as

l 5 X1 b1 1 X2 b2 1 u

5 2X 1 R21
1 R2 b2 1 X2 b2 1 u

5 ~X 2 2 X 1 R21
1 R 2 !b 2 1 u. ~A8!

If we define X * as X 2 2 X 1 R21
1 R 2 we can write the restricted regression
more compactly as

l 5 X * b 2 1 u. ~A9!

The sum of squared residuals from the restricted regression is given by

SSR r 5 l ' M * l, ~A10!

where

M * 5 I 2 X * ~X '* X * !21 X '* . ~A11!

Now X * is the set of excess returns on a set of hypothetical basis assets

which span the space of portfolio returns generated by the original assets
subject to the restrictions on portfolio weights. The restriction Rb 5 0 im-
plies that the maximal ex ante Sharpe ratio from the full set of assets is
equal to the maximal ex ante Sharpe ratio achievable from the restricted
set. Now the ex post Sharpe ratio for the complete set of assets is ~xS ' SO 21 x! S 102 .
Similarly, for the restricted case the maximal ex post Sharpe ratio is given
by ~xS '* SO 21
* xS * ! 102, where xS * 5 X '* l0T and SO * 5 ~X * 2 lxS '* !' ~X * 2 lxS '* !0T. Some
matrix algebra shows that the sum of squared residuals from the un-
restricted regression is

T
SSR 5 l ' Ml 5 , ~A12!
1 1 xS ' SO 21 xS
 Sampling Error in Mean-Variance Efficient Portfolio Weights 669

and the sum of squared residuals from the restricted regression is

T
SSR r 5 l ' M * l 5 . ~A13!
11 xS '* SO 21
* xS*

The standard OLS F-statistic can thus be written in terms of squared Sharpe
ratios:

~SSR r 2 SSR!0q
F~OLS! 5
SSR0~T 2 k!

~xS ' SO 21 xS 2 xS '* SO 21

* xS * !0q
5 . ~A14!
~1 1 xS '* SO 21
* x S * !0~T 2 k!

By replacing x 1 with xS * in Seber ~1982!, Theorem 2.11, p. 52, it can be seen

that this statistic is distributed as an exact central F distribution with q and
T 2 k degrees of freedom under the null hypothesis that the ex ante maxi-
mal Sharpe ratios are identical in the restricted and unrestricted cases. Mar-
dia, Kent, and Bibby ~1989!, p. 78, view this statistic as a ‘decomposition of
Mahalanobis distance’ and also provide a proof that the distribution is cen-
tral F under the null hypothesis.

Proof of Corollary 1: Define X 2i as the matrix X with the ith column

removed, and define its associated orthogonal projection matrix as

M 2i [ I 2 X 2i ~X '2i X 2i !21 X '2i . ~A15!

From standard least squares algebra15 we can express bZ i as

bZ i 5 ~x i' M 2i x i !21 x i' M 2i l, ~A16!

and we can express the estimate of bZ i ’s standard error as

s!a ii 5 s~x i' M 2i x i ! 102. ~A17!

Thus the t-statistic for bi 5 0 can be expressed as

x i' M 2i l
. ~A18!
s~x i' M 2i x i ! 102

15
The necessary algebra is contained in Davidson and MacKinnon ~1993!, pp. 19–24, 81–86.
670 The Journal of Finance

The square of the t-statistic is thus

~x i' M 2i l! 2
. ~A19!
s 2 x i' M 2i x i

From standard least squares algebra, the difference between the restricted
and unrestricted sum of squares can be expressed as

SSR r 2 SSR u 5 l ' Ml 2 l ' M 2i l

~x i' M 2i l! 2
5 ’ ~A20!
x i' M 2i x i

where M is the orthogonal matrix associated with the unrestricted regres-

sion:

M [ I 2 X~X ' X!21 X '. ~A21!

Thus the squared t-statistic can be expressed as

~x i' M 2i l! 2 SSR r 2 SSR u

5 . ~A22!
2 i'
s x M 2i x i
SSR u 0~T 2 K !

But this is the F-statistic from Theorem 3 for the special case of q 5 1. As the
square root of a random variable distributed as F1,T2k is distributed as t
with T 2 K degrees of freedom, we have proved the corollary.

REFERENCES
Adler, Michael, and Bernard Dumas, 1983, International portfolio choice and corporation fi-
nance: A synthesis, Journal of Finance 38, 925–984.
Best, Michael J., and Robert Grauer, 1991, On the sensitivity of mean-variance efficient port-
folios to changes in asset means: Some analytical and computational results, Review of
Financial Studies 4, 315–342.
Black, Fischer, and Robert Litterman, 1992, Global portfolio optimization, Financial Analysts
Journal 48, 28–43.
Cochrane, John H., 1996, A cross-sectional test of an investment-based asset pricing model,
Journal of Political Economy 104, 572–621.
Cochrane, John H., 1997, Asset pricing, Unpublished manuscript, Graduate School of Business,
University of Chicago.
Davidson, Russell, and James G. MacKinnon, 1993, Estimation and Inference in Econometrics
~Oxford University Press, New York, NY!.
Fama, Eugene F., 1996, Multifactor portfolio efficiency and multifactor asset pricing, Journal
of Financial and Quantitative Analysis 31, 441–465.
Frost, Peter A., and James E. Savarino, 1986, An empirical Bayes approach to efficient portfolio
selection, Journal of Financial and Quantitative Analysis 21, 293–306.
Frost, Peter A., and James E. Savarino, 1988, For better performance: Constrain portfolio weights,
Journal of Portfolio Management 14, 29–34.
 Sampling Error in Mean-Variance Efficient Portfolio Weights 671

Gallant, A. Ronald, Lars P. Hansen, and George Tauchen, 1990, Using conditional moments of
asset payoffs to infer the volatility of intertemporal marginal rates of substitutions, Jour-
nal of Econometrics 45, 141–179.
Gibbons, Michael R., Stephen A. Ross, and Jay Shanken, 1989, A test of the efficiency of a given
portfolio, Econometrica 57, 1121–1152.
Greene, William H., 1993, Econometric Analysis ~Macmillan Publishing Company, New York,
NY!.
Grinblatt, Mark, and Sheridan Titman, 1983, Factor pricing in a finite economy, Journal of
Financial Economics 12, 497–507.
Hansen, Lars P., and Scott F. Richards, 1987, The role of conditioning information in deducing
testable restrictions implied by dynamic asset pricing models, Econometrica 55, 587–613.
Harvey, Campbell R., and Guofu Zhou, 1990, Bayesian inference in asset pricing tests, Journal
of Financial Economics 26, 221–254.
Haugen, Robert A., 1997, Modern Investment Theory ~Prentice Hall, Upper Saddle River, N.J.!.
Jobson, John D., and Robert Korkie, 1989, A performance interpretation of multivariate tests of
asset set intersection, spanning, and mean-variance efficiency, Journal of Financial and
Quantitative Analysis 24, 185–204.
Jorion, Phillipe, 1985, International portfolio diversification with estimation risk, Journal of
Business 58, 259–278.
Kandel, Shmuel, Robert E. McCulloch, and Robert F. Stambaugh, 1995, Bayesian inference and
portfolio inefficiency, Review of Financial Studies 8, 1–53.
Kandel, Shmuel, and Robert F. Stambaugh, 1995, Portfolio inefficiency and the cross-section of
expected returns, Journal of Finance 50, 185–224.
Kroll, Yoram, Haim Levy, and Harry M. Markowitz, 1984, Mean-variance versus direct utility
maximization, Journal of Finance 39, 47–61.
Ledoit, Olivier, 1995, Linear-quadratic estimation of the covariance matrix under general as-
ymptotics, Working paper #25-95, The Anderson School, UCLA.
Mardia, Kanti V., Kent, John T., and John M. Bibby, 1989, Multivariate Analysis ~Academic
Press, Harcourt Brace Jovanovich, London, U.K.!.
McCulloch, Robert E., and Peter E. Rossi, 1990, Posterior, predictive, and utility-based ap-
proaches to testing the arbitrage pricing theory, Journal of Financial Economics 28, 7–38.
McCulloch, Robert E., and Peter E. Rossi, 1991, A Bayesian approach to testing the arbitrage
pricing theory, Journal of Econometrics 49, 141–168.
Pulley, Lawrence B., 1981, General mean-variance approximation to expected utility for short
holding periods, Journal of Financial and Quantitative Analysis 16, 361–373.
Roll, Richard, 1985, A note on Shanken’s CSR T 2 test for mean0variance efficiency, Journal of
Financial Economics 14, 349–357.
Roll, Richard, and Stephen A. Ross, 1994, On the cross-sectional relation between expected
returns and betas, Journal of Finance 49, 101–121.
Seber, G. A. F., 1982, Multivariate Observations ~John Wiley & Sons, New York, N.Y.!.
Shanken, Jay, 1987, A Bayesian approach to testing portfolio efficiency, Journal of Financial
Economics 19, 195–215.
Sinquefield, Rex A., 1996, Where are the gains from international diversification?, Financial
Analysts Journal 52, 8–14.
Solnik, Bruno, 1982, Optimal international asset allocation, Journal of Portfolio Management 9,
11–21.
Stulz, René M., 1995, International portfolio choice and asset pricing: An integrative survey; in
Robert A. Jarrow, V. Maksimovic, and William T. Ziemba, eds.: Handbooks in Operations
Research and Management Science, vol. 9 ~Elsevier Science B.V Amsterdam, Netherlands!.