When Portfolios Well Diversified?: Will Mean-Variance Efficient

Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 25

THE JOURNAL O F FINANCE - VOL.XLVII, NO.

5 - DECEMBER 1992

When Will Mean-Variance Efficient


Portfolios Be Well Diversified?
RICHARD C. GREEN and BURTON HOLLIFIELD*

ABSTRACT
We characterize the conditions under which efficient portfolios put small weights on individual
assets. These conditions bound mean returns with measures of average absolute covariability
between assets. The bounds clarify the relationship between linear asset pricing models and well-
diversified efficient portfolios. We argue that the extreme weightings in sample efficient
portfolios are due to the dominance of a single factor in equity returns. This makes i t easy to
diversify on subsets to reduce residual risk, while weighting the subsets t o reduce factor risk
simultaneously. The latter involves taking extreme positions. This behavior seems unlikely t o be
at-tributable to sampling error.

THE MEAN-VARIANCE EFFICIENT frontier plays an important role in pedagogy and


applications in finance. The properties of efficient portfolios are also central in both static
and dynamic asset pricing.’ For all its simplicity and intuitive appeal, however, practical
implementation of mean-variance analy-sis has proved problematic. Portfolios
constructed using sample moments of returns often involve very extreme positions. As
the number of assets grow, the weights on individual assets do not approach zero as
quickly as suggested by naive notions of “diversification.”

Admittedly, the theoretical links between “diversification,” in the sense of putting


small weight on any asset, and variance minimization are tenuous. The Arbitrage
Pricing Theory (APT) implies that efficient portfolios are well diversified, but the
mean-variance Capital Asset Pricing Model (CAPM) requires this only insofar as the
value weighted market portfolio is diver-sified. In applications, however, the
connection between mean-variance efficiency and diversification is central. Black
and Eitterman (19901, for example, point to the inconsistency between naive
diversification and the weights generated by asset allocation models as a major
obstacle to their implementation. Practitioners are suspicious of portfolios that are not
naively

* Green is from Carnegie Mellon University and Hollifield is from the University of British Columbia.
We wish t o thank seminar participants at Carnegie Mellon University, Rutgers University, Queens
University, the University of Colorado at Boulder, the Stockholm School of Economics, and the 1990
Western Finance Association meetings for helpful comments on this work. The comments and guidance
of the editor, Re& Stulz, and of two anonymous referees significantly improved the paper. Hollifield
gratefully acknowledges the support of the Social Sciences and Humanities Research Council of Canada.

‘ S e e , for some recent examples, discussions in Chamberlain and Rothschild (19831,Hansen and
Richard (19871, and Huberman, Kandel, and Stambaugh (1987).

1785
1786 The Journal of Finance

diversified, so mean-variance methods are generally implemented with exten-sive


sets of constraints that enforce such diversification. Numerous academic researchers
have also documented problems in employing the weights that result from calculating
efficient frontiers using sample moments.‘
This paper studies the question of when mean-variance efficiency and “well
diversified” coincide, and when they do not. First, we provide a theoretical answer to
this question. Then, we use more specialized examples to advance possible
explanations for the “poorly diversified” weights found in efficient portfolios
calculated using sample return data.
Through duality theory, we provide simple and intuitive characterizations of
conditions under which minimum-variance portfolios will be well diversi-fied. These
characterizations involve bounds on the means of portfolio returns in terms of their
average absolute covariance with the individual assets. We use these results to study the
equivalence between diversified frontier portfo-lios and exact APT pricing, first
demonstrated in Chamberlain (1983). We show that an approximate factor structure is a
stronger condition than needed for diversification t o eliminate residual risk and ensure
APT pricing. Intuitively, the eigenvalues of the covariance matrix can grow, if they grow
more slowly than the rate at which the weights in the minimum-variance portfolios go to
zero.

We then develop some simple cases and examples to better understand why
efficient portfolios computed from sample moments tend to be so poorly diversified.
If the extreme behavior of the weights is due simply to meas-urement error in the
sample first and second moments, then one should employ Bayesian procedures, or
shrinkage estimators, or simply rely on non-statistical methods of estimating these
quantities. Examples of these ap-proaches include work by Frost and Savarino (1986,
1988) and Black and Litterman (1990). If, on the other hand, the behavior of the
weights is due to characteristics of the population moments, then perhaps
practitioners should question more critically the value of naive &versification.

Our analysis suggests the following explanation. The extreme weights in efficient
portfolios are due t o the dominance of a single factor in the covari-ance structure of
returns, and the consequent high correlation between naively diversified portfolios. With
small amounts of cross-sectional diversity in asset betas, well-&versified portfolios can
be constructed on subsets of the assets with very little residual risk and different betas. A
portfolio of these diversified portfolios can then be created that has a zero beta, thus
eliminat-ing the factor risk as well as the residual risk. Obviously, the greater the degree
of domination of the first factor, the easier it is to implement this strategy on a relatively
small asset universe.

As the number of assets grows to infinity, the weights that result from such a strategy
will eventually grow small. The rate at which they grow small,
’ See, for example, Kallberg and Ziemba (19831,Kroll, Levy, and Markowitz (19841, and Pulley
(1981).Similarly, Best and Grauer (1991) document the extreme sensitivity of portfolio weights to small
perturbations in means.
Mean-Variance Efficiency and Diversification 1787

however, will depend on how similar the betas of the underlying assets are. If they are
relatively constant in the cross section, portfolios diversified on subsets will have similar
betas, and extreme weighting of these will be required to reduce factor risk. Minimizing
variance may involve taking very large short positions in a diversified portfolio of assets
and using the proceeds to finance very large long positions in another diversified
portfolio.
These problems can become even more evident when the “assets” used in variance
minimization are themselves well-diversified portfolios of a large underlying set of assets,
as is often the case in applications. Strategic and tactical asset allocation procedures often
use variance minimization to &s-tribute funds across passively diversified portfolios such
as country, industry,
or size-based portfolios. Our results may shed some light on the difficulties encountered
in these application^.^
We illustrate this behavior through example, and argue that the resulting weights
correspond to those observed empirically. On the basis of this analy-sis, we argue that it is
difficult to dismiss the observed extreme weightings as simply a consequence of sampling
error. Given the dominance of a single factor, an extreme degree of similarity in the betas,
or loadings, on that factor would be required t o ensure that naive diversification
corresponds t o mean-variance efficiency for finite sets of assets. This is equally true of
the population moments as it is of the sample moments. Using standard multi-variate
methods we provide statistical evidence that sufficient diversity in the betas is discernible
even in the presence of sampling error.

The paper is organized as follows. Section I introduces notation, defines


“diversification,” and derives the dual to the system of inequalities that must have a
solution if efficient portfolios are well &versified. Section II studies the relationship
between diversification and APT pricing. Section 111 docu-ments the behavior of
efficient portfolios formed using sample moments and considers the extent to which the
first principal component of the sample covariance matrix is the source of the extreme
weights on sample mean-variance frontiers. Section IV briefly concludes.

I. Duality and Diversification


We consider the first two moments of the returns of n different securities. Let
R denote the random vector of returns, with typical element R;.Let V be the variance-
covariance matrix for the returns, assumed t o be strictly positive definite, with variances
a,‘ on the diagonal and covariances uiioff diagonal. We denote the n-vector of mean
returns as M , with typical element Mi.Let q denote an n-vector of ones. Then a
portfolio is a vector of weights for the securities, w,, satisfying wig = 1, so that the
weights sum to one. It has return R , = w i R , expected return E(R,) = wdM, variance o j
= wdVwp, and covariance with any other portfolio, s , ups= w ~ V W Alternatively,. we

For example Black and Litterman (1990), Frost and Savarino (1986, 1988),Kallberg and Ziemba
(19831,Kroll, Levy, and Markowitz (19841,and Pulley (1981).
1788 The Journal of Finance

can construct hedge portfolios that are zero-cost positions with "weights" that sum to
zero. Let z o be such a position. Then zAq = 0, and we denote the payoff on such a
position as IIo = zAR.
Let w, denote the portfolio that for a given mean p, has minimum variance. It
minimizes W'VW subject to w'q = 1 and w'M = p. The first
order condition for this problem, which is both necessary and sufficient, is the
existence of scalars A,, y, such that the following equation holds:
VW, = A, q + yPM.
Merton (1972) gives the following expressions for the two multipliers in this equation:
(M'V-lM)- ( q ' v - l M ) p
A, (2)
=
(M'V-'M)(q'v-'q)-(q'V-1M)2
(q'V-1q)p - (q'V-1M) (3)
Y, =

( M 'V-'M ) ( q'v-1q 1 ( q'v-'M l2


-

At the global minimum-variance portfolio, where the constraint on the means is not
binding, y, will be zero. Along the efficient part of the mini-mum-variance frontier l /
y , will be positive, and it will be negative along the lower boundary. For any mean
other than that of the global minimum-variance portfolio, let R, be the return on the
efficient portfolio with weights w, and let w, denote the weights on the minimum-
variance portfolio that is uncorrelated with R,. Premultiplying (1) by these weights
and setting the result equal t o zero gives as the "zero-beta" rate

Similarly, premultiplying (1)by wp gives

Equation (5) says that 1/ y, measures the excess return to variance ratio for efficient
portfolios.
We take the term "well diversified" to mean that weight in the portfolio is spread
reasonably evenly across many assets, and thus the weight on any one asset gets small
as the number of assets grow. This can be given formal content by imposing a bound
on the norm of the weights. The norm that proves tractable for our purposes here is
the sup-norm. We simply require lwil I K,. As an alternative, Chamberlain (1983)
identifies portfolio diversifi-cation with the sum of the squared weights, or the 1,-norm.
The differences between these approaches are discussed in Section 111. The bound we
impose on the absolute value of the weights depends on the size of the cross section. With
diversification, we typically intend that the weights on individual assets get small as the
universe of available assets expands.
Mean-Variance Eficiency and Diversification 1789

We can now pose formally the question raised in the Introduction. We say the
mean-variance frontier is well diversified a t mean p if there exists a portfolio that
satisfies:
P1
(6)
Vw,, = A p q + yFM
- K,l 5 wi I K , for i = l;.., n (7)

The next lemma provides the dual t o diversified frontier portfolios.


Lemma 1: There exists a solution to P1 if a n d only if there is no n-uector y
satisfying
D1
(A,(y'q) + y,(y'M)I > K,CYll lCov(y'R,Ri)l (8)

Proof: (10 Suppose the weights, wp, are poorly diversified. We will construct a
solution to the dual. Assume lwPjl> K , and let ej denote the j t h column vector of the
n-by-n identity matrix. It has zeros in all components but the j t h , where it has a one.
Note that
w W. = e!w.l@ = A,(eJV-'q) (9)
+
yp(e;V-'M)
Consider the portfolio yj = V-'ej. Since

IA,(yJ!q) + yp(yjlM)I = lAp(e;V-'q) + yp(eJV-'M)I = lwpjl (10)

the first two terms in (8), evaluated a t y J 7have absolute value greater than K,.
Since Vyj = ej, X i ICov(Rryj7Ri)l= 1, so the right side of (8) is simply

K,. Thus, (8) is dual to the system P1.


(Only ifJ Suppose that the efficient portfolio with mean p is well &versified, so Iwpil
_< K,. Let y be any n-vector. Premultiply (6) by y r to obtain
=
A,(y'q) + y,(y'M). (11)
y'Vwp = E l Cov(R'y, R,>wPi
Therefore,
1 A,+( y '4) + y,( Y ' M I I c , Cov( =

R ' Y , R wp 1 -< xiICodR'y, R,)llw,,I


-< K , C z ICov(R'Y,Ri)l (12)
Thus, if the frontier is well diversified at p, no n-vector that satisfies (8) exists.

Proceeding from the dual inequality (8) to the conclusion of the following theorem
is simply a matter of keeping track of the various ways in which the vector y can be
normalized. We use v to denote the mean return on the global minimum-variance
portfolio and :u its variance.
1790 The Journal of Finance

Theorem 1: The efficient portfolio with mean p # only u is well diversified i f and
if the return, R*, on every portfolio with satisfies weights that sum to one,

I E ( R * ) - E(R,)I Knn (13)


I - C,=lICov(R*,Ri)l
I YYI
and the p a y o f f , n*, on euery hedge position, with weights that sum to zero,
satisfies:

Proof: (10 Suppose the frontier is not well diversified at p # u . Then, by Lemma 1,
there exists y such that ( 8 ) is satisfied. Suppose y ' q # 0. Define R* = y ' R / y ' q . The
portfolio with this return has weights that sum to one. By
(8):
I[(A,/Y,) + ( Y ~ M ) / ( Y W ] Y , ( Y W ~ > K,C:=~IOV( Y ' R , ~ ~ 1 1(15),

divide both sides by 1 y,( y'q)l to get:


I -E(R,) + > I K , / ~ ,I
c In= 1 ICOV(R*,
R,)I, (16)

violating (13).
Suppose y ' q = 0. Define lI* = y ' R . By (8):
l(y'M)y,I > KnC:=lI C o v ( y ' R , R i ) l . (17)

Divide both sides by Iy,l to get:


IE(II*>l>IK,/yp/C:=l ICov(n*,R i ) l , (18)

which violates (14).

(Only i 0 Suppose either (13) or (14) are violated at p # u , where y, # 0. We will


construct a solution to the dual in these cases, implying the frontier is not well
diversified. First, suppose there is a portfolio, with return R* and weights that sum t o
one, violating (13). Let y be the weights on this portfolio. Thus, y ' q = 1, and using (4)
we will write the violation of (13) as:

Multiply the above by I ly, to get:

which shows y satisfies the dual, (8),at p.


Mean-Variance Efficiency and Diuersification 1791

Now suppose there exists a payoff PI*, with weights that sum to zero, that in this
violates (14). Let y be the weights y ' q position. Since (14) is violated and
= 0:

Multiply the above by Iy,l to get:


J(yf~)

+ h , ( y w ) > K , C ~ 1I C OV (Y' R , R ~ ) I ,
y,
(22)
1.=

and y satisfies the dual, (8) at p.


A corollary follows from consideration of the global minimum-variance portfolio.
At its mean, v , the constraint on the mean return is non-binding, s o y, = 0.
Premultiplying the first-order condition, (6), by the weights in the portfolio then gives the
variance of the minimum-variance portfolio as the multiplier for the constraint that the
weights sum to unity. That is, :a = A,,. Obviously, the conditions ensuring that the
weights in this portfolio are well diversified involve only the characteristics of the
covariance matrix, and not those of the means.

Corollary 1: The portfolio with the m i n i m u m feasible uariance is well diuersi-fied


if and only i f the return, R':, on euery portfolio with weights that s u m to one,
satisfies
q? I K , ICov(R*, Rill. (23)

Proof: Consider inequality (81, which characterizes solutions of the dual at any mean
return, with 7, = 0, and A, = uv2If. y solves the dual at this point,

Normalizing by I y '41 gives

where R* is the return on a portfolio with weights that sum to one. It thus violates
(23). Now suppose there is a portfolio violating (23). Such a portfolio obviously
satisfies (81, since A, = 4,' and y, = 0.
The results thus far characterize the first two moments of asset returns when the
efficient portfolio with a particular mean is well diversified. We now pose a broader
question: When will there be no portfolios anywhere on the mean-variance frontier
that are well diversified? We show that this will be the case if and only if there is a
single set of portfolio weights that solve the dual for all possible target mean returns.

Theorem 2: There are no well-diuersified portfolios anywhere o n the mean-uariance


frontier if and only if there is a portfolio y * with return R" =
1792 The Journal of Finance

R'y* that has the same mean as the global minimum-variance portfolio ( E ( R * )=
v), and that satisfies

Proof (ID Since q ' y * = 1, we can use the expressions for the multipliers,
(2) and (3) to write the left-hand side of (8) as:

1
- -D- ( E ( R * ) ( S ' V - ~ M )- ( M ' V - l M ) ) , (27)
where D = ( M ' V - l M ) ( q ' V - ' q ) - ( q ' V - ' M ) 2 . Since y, = 0, v = ( q ' V - ' M ) / (
q ' V - '4) and the coefficient multiplying /.L on the right-hand side is zero at E ( R * )= v.
The second term, at E ( R * )= v , is simply -A, = -9'. Thus, at y * , the left-hand side of
(8) is simply a;:, which clearly satisfies the dual if (26) is satisfied, and y * solves the
dual at all mean returns.
(Only if) Suppose the mean-variance frontier is nowhere well diversified. The set S =
{w,: - m < p < m} is, by the two-fund theorem, a convex set. By hypothesis, it is
disjoint from the set I = {w:lwil IK J , which is also convex. The two may therefore be
separated by a hyperplane, and there exists a vector b such that w; b > w ' b for all ,u
and all u' E I . In particular,
wLb = A,(q'V - lb) + y,(M'V - 'b) > supWEIw'b= K , C i lbL\, (28)
where the last step follows since w ' b is maximized by choosing w ,= K , or w . = -
K II depending on the sign of b,.4 Define j j = T ' b . Then, q ' j = q ' V - l b , M ' j = M ' V
- ' b , and ICov(Ri, R'jj)I = Ib,l. It follows then from (28) that j j satisfies the dual
inequality, (B), at every mean return, p. At p = v , y,, = 0, and by (28) A , q '9 > K , 1
bi]> 0. This implies q ' j # 0, so we can define y * = 4 / j ' q and R* = R'y''. Then for all p,

A, + y,E(R*) > K , C i ICov(Ri, R*)l, (29)

and in particular (26) is satisfied. The right-hand side of (29) does not depend on p.
Expressing the left-hand side as in (27), we find it is linear in p. The only way, then,
(29) can hold for all p is for the coefficient on p to be zero. As noted in connection
with (27), this implies E ( R * )= v.
Theorem 2 shows that when the frontier is nowhere well diversified, the portfolio
that solves the dual at every mean is one that solves it at the minimum-variance point.
In adhtion, it has the same mean as the global minimum-variance portfolio. Thus,
verifylng the existence of just one such
Note that it is in dealing with inequality (28) that the choice of the sup-norm provides trsctability. In
thec,definition of the set I, we do not impose w ' q = 1 so that we can set SUPW F I w ' b = K,, IbJ.
Mean-Variance Efficiency and Diversification 1793

portfolio obviates the need t o check for solutions to the dual problems at other means.
The intuition for this result is that the dual problem, like the primal, has a.“two-fund”
structure. Any hyperplane that separates the well-diversified portfolios from the efficient
frontier, which is a ray in Euclidean n-space, must meet only two criteria. It must separate
the well-diversified portfolio from the minimum-variance portfolio, and it must be
parallel t o the ray along whch all the other efficient portfolios lie.

II. Factors and Diversification


In this section we consider the duality results developed above when returns have a
factor structure. The issues we study here have been consid-ered within a Hilbert space
setting in Chamberlain and Rothschild (1983) and Chamberlain (1983). The finite
dimensional duality theory developed above provides a very straightforward intuition for
these relationships. It also provides a very simple means of proving Chamberlain’s result
that, given an approximate factor structure, well-diversified portfolios on the efficient
fron-tier imply that the mean of each asset converges to the APT expected return. This
result is stronger than the convergence in mean-square in Ross (1976). Finally, our
approach allows us to generalize the result to situations where the covariance matrix of
residuals has eigenvalues that are unbounded but grow at a relatively slow rate.

We begin by considering a “strict factor structure.” It is assumed that the n-vector of


returns, R,has the following structure:
(30)
R = E(R) t p6 + E

The K‘-vector of “factors,” 6, are taken to be mutually orthogonal with unit variance, and
the n-vector of “residuals,” E , has covariance matrix R, which is diagonal. We denote the
variance of the residual for asset i as w,“. The covariance matrix of returns can then be
decomposed as:
(31)
v = pp‘ + R
where the matrix /3 consists of n rows and K ” columns of “factor loadings.” The
variance, w’VW ,of a given portfolio, w ,consists of two terms. We will refer to the
term w’pp‘w as the “factor risk” of the portfolio, and the term W‘RWas its
“idiosyncratic risk.”
We say exact APT pricing holds if the mean returns are in the span of the betas and a
constant vector.

The portfolios R;,k = O; .. , K*,are “factor-mimicking” portfolios. We say


that exact APT pricing holds in the limit if, as the number of assets, n,
increases, there exist sequences of feasible factor-mimicking portfolios R:lz,
1794 The Journal of Finance

for any fixed j .


Consider, as an example, a single factor structure. In this case,

Cov( Ri,Rj) = P:+wJ~ i f i = j (34)


Pi Pj ifizj.

Using this we can bound the assets’ means. Assume the portfolio with mean
p is well diversified, and let y j be a portfolio that puts weight one on asset j. Equation
(13) from Theorem 1 implies:

where hj = w ~ / I pjl. It follows that if the efficient portfolio with mean p is well
diversified then

for every j = l;..,n.


Recall from (5) that 1/ y,, measures the tradeoff between risk and return along the
frontier. We assume that it is bounded away from zero and infin-ity =:as the cross
section grows. If the factor is “pervasive”, then we expect C I Pil to grow
without bound as n + M, while K , + 0. Otherwise there would be an infinitely large
subset of assets with infinitesimal factor risk. Factor risk, in such a case, could be
eliminated simply through diversifica-tion. The only term on the right-hand side of (36)
which is specific t o the asset is hj, which does not change with the cross section. If K n
goes to zero at the same rate as the sum of the absolute betas grows, then the right-hand
side of
(36) will go to a bound which is independent of the asset being considered.
Variation in the means outside of these bounds must be due to variation in the
betas. For a fixed cross section of assets, as asset j’s beta goes to zero, its
expected return must also be approaching the zero-beta rate for the frontier to
be well diversified. The larger the beta is in absolute value, the larger is the
allowable variation in expected returns. With multiple factors similar
Mean-Variance Efficiency and Diversification 1795

relations can be derived that bound mean returns in terms of weighted averages of the
betas across factors, as well as asset^.^

The bound in (36) above will be very tight for assets with very little factor risk. By
constructing such portfolios we can obtain much tighter bounds for all assets than
those above, and for each asset these converge to the expected return predicted by the
APT. Using a superscript n t o denote the efficient set parameters when there are n
assets, we make the following assumptions.
Assumption 1: There is a strict factor structure.
Assumption 2: For any # limn+ - un , y,” is uniformly bounded away from zero.

The first of these is self-explanatory. The second requires that the mean-variance
frontier not become vertical in the limit, and hence that there remains a meaningful
tradeoff between mean and variance. In Chamberlain and Rothschild (1983) this
condition is ensured by “absence of arbitrage” in their limiting, Hilbert space
economy. Thus, with the appropriate definitions, we could replace Assumption 2 with
“there is no asymptotic arbitrage.” The following theorem is analogous t o
Chamberlain’s (1983, Theorem 1 and Corollary) result for the Hilbert space setting.
The proof follows the standard strategy of constructing a return from the “factor-
mimicking” portfolios with the same betas as the return on some particular asset.

Theorem 3: Under Assumptions 1 and 2 a sufficient condition for exact APT pricing
to hold in the limit is that, when K , + 0 as n + m, every portfolio in the sequence of
efficient portfolios for some mean return p # limn a u n is well diuersified.
+

Proof: Let H, be the set of demeaned portfolios.


(37)
E ( R ) :R = En 1= 1 xiRi, cy=, xZ = I}.

We construct the factor-mimiclung portfolios by projecting factor k onto H,, , so that

where E( CnhRj)= 0 for j = I;.., n. Similarly, we construct R:o by projecting the zero
vector onto H , and let Cn0 be the projection error. (Note that R*,o will be the
minimum-variance portfolio which has zero betas.) For asset j ,
Consider the 1-norm of the betas for factor k , 11 Pklll = ci
1fi r k / .We can use these as
weights to compute a measure of asset j ’ s responsiveness to factor movements

The Cauchy-Schwartz inequality, and steps similar to those above now leads t o an inequality like ( 3 6 )
with b, substituting for pj and C, 11 P h ) ) l substituting for X i1 pi\on the right-hand side.
1796 The Journal of Finance

where

cj = E(Rj) - ((1- chPjh)E(R;o) + c,PjhE(Rx,)) . (41)

If the efficient frontier contains a well-diversified portfolio, the duality condi-tions


from Theorem 1 imply:

Consider the covariances in the bounds above:

Therefore

and since as n + to, K n + 0, y,” + y, > 0, the right-hand side goes to zero as
n-a. E

This theorem differs from Chamberlain’s in the way “&versification” is


formalized. Chamberlain requires that there be a portfolio on the efficient frontier
that is well diversified in the sense of having weights that converge to zero in the sum
of squares, or I,-norm. Our result requires that there be some point on the frontier
where portfolios are well diversified in the sense of the sup-norm. The sup-norm is a
weaker notion of diversification. We are dealing with sequences of portfolio weights
where the nth vector in the sequence is zero in all but the first n components. Thus,
for every finite n ,
Mean-Variance Efficiency and Diversification 1797

the norm-equivalence relation (see, e.g., Stewart (1973), p. 170) states

llwllm s IIwIIz 5 n1/211wIl=. (45)

The 1,-norm, which is what Chamberlain considers, will go to zero if K,, which
restricts the sup- or m-norm, goes to zero at rate n - l I 2 . Clearly, however, K, may go
to zero in situations where the I,-norm does not. Thus, portfolios that are well
diversified in Chamberlain’s (1983) sense will also be well diversified by our criteria,
but not the reverse. In this respect, the theorem is a generalization of Chamberlain’s
(1983) result. Using a weaker notion of diversification we still find the relationship
between efficient, diversified portfolios and APT pricing holds. On the other hand, as
stated above, our result relies on a strict factor structure whereas Chamberlain’s
simply requires that the eigenvalues of the residual covariance matrix have a uniform
bound, an “approximate factor structure.”

In fact, though, the duality conditions expressed in terms of the sup-norm reveal
that even an approximate factor structure is stronger than needed. Consideration of
equations (42)-(44) makes clear that t o ensure convergence to strict APT pricing
when there are well-diversified efficient portfolios, all that is really needed is that the
bound on the weights go to zero faster than the l-norm of the residual covariance
matrix. That is, we require

This can certainly happen in situations where the eigenvalues are not uniformly
bounded. Intuitively, diversification still “works,” in the sense that well-diversified
portfolios have smaller and smaller residual variances, even if the eigenvalues of
a,,the residual covariance matrix, are not bounded; they just cannot grow too fast
relative to K,. The empirical evidence in, for example, Trzcinka (1986) suggests such
situations may indeed be relevant. These studies document the existence of a
dominant eigenvalue which clearly grows linearly with the number of assets. Other
eigenvalues appear to be important, but they seem t o grow at a much slower rate. The
evidence on whether the additional factors are actually priced appears mixed (ShuMa
and Trzcinka (1990)). The following corollary to Theorem 3 gives an explicit
expression for the speed at which they can grow, although we show by example
below that for specific covariance matrices, faster rates of growth may be
permissible.
Corollary 3: Assume, in place of Assumption 1, that limn+ - Kni+hnn1I2= 0,
where i+h” denotes the largest eigenualue of the n-asset residual couariance matrix
a“.Th e n exact A P T p r i c i n g holds in the limit .
Proof: The arguments proceed as in the Theorem up to equation (38), from whch
we have,
COV(rI,”,Ri)= C0V(Ej, q ) . (47)
1798 The Journal of Finance

Therefore

The last inequality follows from the equivalence relations between matrix norms
(see, e.g., Stewart (1973),p. 183).
The following is an example of a residual covariance matrix that does not have any
bounded eigenvalues, and so does not have an approximate factor structure.
Nevertheless, diversification "works" in the sense that the equally weighted portfolio
has zero limiting residual variance. It also illustrates that the sufficient conditions in
Corollary 3, while weaker than bounded eigenval-ues, are still not necessary for well-
diversified efficient portfolios to imply APT pricing.

Example 1: Suppose, for 0 < a < 1,


(49)
wij = (1 + li - jl)" - (Ii - j l ) " .
The row o r column sums of this matrix "telescope," so that:
(50)

Note that both of the exponentiated terms on the right-hand side of (43) are
less than or equal t o n". Thus an upper bound on c .w i j is 2n". Further,
we have z 2 0 and n - z - 1 2 0, so i" + ( n - i - lj" 2 ( n - i - 1 + i)",
and ( n - 1)" - 1 gives a lower bound on cj
w L j .Thus we see that all of the
row sums grow at rate n", and since all terms in this covariance matrix
are positive, the row sums correspond t o the quantities in the bounds on
the means in, for example, equation (48) . Condition (46) will then hold if
Knn" + 0 or if K,, goes to zero faster than n-".For this covariance matrix a
well-diversified efficient frontier is equivalent to APT pricing as long as "well
diversified" implies bounds on portfolio weights that go to zero faster than
n-". For example, with a = 1 / 2 the row sums (and, as shown below, the
eigenvalues) of this matrix grow at rate n1I2 .To have APT pricing we must
have efficient portfolios with weights that go t o zero as the universe of assets
grows at least as fast as the square root of l / n . Thus, if a = 1/2 the
restriction that K , is o ( n - " ) strengthens the I , notion of &versification. For
(Y < 0.5, however, portfolios that are well diversified in the sense that n"K,,

goes to zero need not be well diversified in the I, sense. For such a case we
employ both a weaker definition of diversification and a weaker notion of
Mean-Variance Efficiency a n d Diversification 1799

“diversifiable risk” than Chamberlain employs, and still find the relationship between
efficient, &versified portfolios and APT pricing holds.
Now, consider the variance of the equally weighted portfolio:
-1(q’.n,q) = 21 CJ(El mi,)
n2 n

Thus, the equally weighted portfolio has zero residual variance in the limit. Nevertheless,
the maximal eigenvalue is unbounded. In fact, it grows faster than the sufficient
condition in the Corollary permits.
*, = sup,; 2 7- - C, ( C imi,) 2 C iw i j )
x ’l l , x ;min,,(

q‘R,q 1 n
x x 44 n
2 ( n - 1)“ - 1 (52)
so 9” grows at least at rate nm .The matrix in the example has the property that, as the
number of assets increases, the 1-norm grows at the same rate as the 2-norm, or the
maximal eigenvalue. In general, we can only be sure that the 1-norm grows at n 1 I 2 times
the rate at which the maximal eigenvalue grows.

111. Frontiers Calculated from Sample Moments


In this section we analyze the behavior of portfolio weights in minimum-variance
portfolios that are calculated using historical returns. We argue that the extreme
weightings seen in efficient portfolios are largely due to the cross-sectional dispersion of
the “loadings” associated with the first principal component, or “factor.”fiWe show that
there is reason to believe this is not simply a result of measurement error.7 We provide
examples that show behaviors similar to those encountered using sample data can be due
to the ability to diversify on subsets t o reduce idiosyncratic risks. These portfolios can
then be combined to simultaneously eliminate factor risk. This tends to result in extreme
portfolio weights.

In the first subsection we analyze conditions under which the efficient frontier fails to
appear well diversified. In the second subsection we will argue that these conditions are
an empirically plausible explanation for the poor diversification observed in sample
efficient portfolios.

We refer to an asset’s beta and its weight in the first principal component interchangeably. Strictly
speaking, they are only asymptotically equivalent as the number of assets grows. For the most part,
however, our arguments concern behaviors that arise as n grows large.
This is a different explanation than those made in Black and Litterman (199Oj,Best and Grauer
(19911, and Frost and Savarino (1986), for example.
1800 The Journal of Finance

A. The Role of a Dominant Factor and Diversification on Subsets


Suppose returns have a single-factor structure. If the betas of the assets are
constant across assets, ( /3 = cq, where c is a scalar) then any portfolio with w ' q = 1
must bear significant factor risk. Variance-minimizing portfo-lios will tend t o weight
assets equally. On the other hand, if there is sufficient diversity in the betas, portfolios
can be created with very little residual risk and different betas. These, in turn, can be
combined to eliminate factor risk, but the weightings involved in this second step
may be very extreme. We argue in this section that such behavior is the source of the
extreme weights observed in efficient portfolios calculated using historical estimates
of the covariance matrix. The effects of diversification on subsets can be illustrated
by example.

Example 2: Assume there is a strict factor structure governing returns with one factor
that has unit variance. All assets have constant idiosyncratic variance 0'. There are 2 n
assets, and we will be considering the conse-quences of letting n grow without bound. All
the assets have betas of unity, except for two finite subsets. These consist of otherwise
identical low-beta and high-beta assets.

p.
=
I,
1
1 - e
for i = l;.. ,N
+
for i = N + l;.., n (53) for i = n + l, ...n, + N
1 e

for i = n + N + l;.., 2n
Now consider two equally weighted portfolios, of the first n and second n assets,
respectively. They will each have idiosyncratic variance of w 2 / n . The first portfolio
will have PL = 1 - e ( N / n ) . The second, PH = 1 + e ( N / n ) . The minimum-
variance portfolio of these two equally weighted portfolios can be calculated directly. For
the first portfolio this weight is:

This weight does not become well diversified, that is close to one-half, as the number
of assets increases. Rather, it approaches (1/2) - ( N e / w 2 ) , which, for reasonably
large N , will be negative and large in absolute value. Note that this is true despite the fact
that for n large relative t o N , the two portfolios would appear t o be virtually identical.
Their betas both converge t o unity. They approach this limit, however, at the same rate as
the residual variance goes to zero. The role of the residual variance term is evident in the
denominator of the middle expression in (54). Were w 2 divided by a number that grew
faster than n , such as n2,the weights in the minimum-variance portfolio would explode
with n. If it were divided by n 1 I 2 the weights would approach one-half. As it is, the need
to take increasingly aggressive short
Mean- Variance Efficiency and Diversification 1801

positions in the low-beta portfolio, in order to eliminate factor risk as the betas
converge, just offsets the exposure to residual variance this entails. Therefore, the
weights approach a constant. This constant may be quite far from weights we would
think of as well diversified, however. Behavior of this sort provides a possible
explanation for the extreme weightings in efficient portfolios composed of other
portfolios we document below.
The weightings on individual assets in this example will not, of course, be variance
minimizing, but the intuition developed is still helpful in under-standing the behavior
of the individual weights on the frontier. For example, consider the simpler case,
where we omit the assets with unit betas. The weights in the global minimum-
variance portfolio can be computed directly €or this case. For the low-beta and high-
beta assets the weights are, respec-tively:

1 e 1 e (55)
+ W H = - -
W L =

2N
-
w2 + 2Ne2 2N w2 + 2Ne2
These weights do go to zero as N increases, but much more slowly if the difference
between the betas (or e ) is small. The second term, which summa-rizes the deviation
from equal weighting, has e in the numerator, while N in the denominator is
multiplied by e 2 . The tendency to offset the two subsets through aggressive shorting
of one will slow the speed with which the weights go to zero. For reasonable
parameter values the weights in (55) are similar in magnitude to the average absolute
values we report for sample frontiers below.'

To summarize, the analysis here provides conditions under which we might expect
to see extreme weightings in the global minimum-variance portfolio, and, by
continuity, in other portfolios on the efficient frontier. The source of the extreme
weightings would be the tendency to offset positions that are themselves well
diversified on subsets of the assets. These positions should be highly correlated with
each other. Furthermore, the extreme weightings should disappear if the heterogeneity
in betas is eliminated. The next section investigates these empirical predictions. Note
that, as the conditions we provide seem just as plausible for population as for sample
moments, our explanation contrasts with traditional ones that ascribe pathologies in
the mean-variance frontier completely to sampling ~ I T O ~ . ~

B. Characteristics of the Weights in Sample Minimum-Variance Portfolios


We first examine portfolios constructed using returns from the Center for Research
in Security Prices (CRSP) monthly returns file, from the period January 1971 to
December 1985. This fifteen-year time period was broken into three five-year
subperiods. Within each subperiod we randomly selected

For example, with n = 50, e = 0.1, and w = 0.1 the average absolute value of the weights is 9.5%. With
e reduced to 0.05,the average absolute value of the weights for 50 assets is 19%.
See, for example, Frost and Savarino (1986, 1988) and Black and Litterman (1990).
1802 The Journal of Finance

firms from the set of all CRSP firms, and then retained those with complete records
over that particular five-year subperiod. This procedure gave us just over 900 sixty-
month return series which were divided fairly evenly between the three subperiods.
Note, however, that a particular firm could easily appear in more than one subperiod.
Within each of the five-year subperiods, we constructed subsets of 10, 30, and 50
firms, and from these calculated sample means, variances, and covariances. The
larger subsets simply regroup the smaller subsets. For example, the first 30-stock
subset from 1/71 t o 12/75 is simply the firms from the first three 10-stock subsets.
For each subset we calculated efficient portfolio weights and other statistics. No
difference in the behaviors of interest was evident across the three subperi-ods, so the
descriptive results reported below are aggregated across subperi-ods as well as across
subsets of firms within subperiods. Over all three subperiods we had available 90
subsets of 10 stocks each, 30 subsets of 30 stocks each, and 18 subsets of 50 stocks
each.

The upper panel of Table I provides some summary measures of the degree of
diversification in portfolios constructed using sample moments, and how these measures
vary as the size of the asset universe grows. These are the weights in the global minimum-
variance portfolios. The first three columns concern the maximum absolute value of the
weights on individual assets within each of the portfolios. This is a measure of how big
the vectors of weights are in the sup-norm. In terms of the formulae in the paper, these
numbers tell us how big K, would have to be for the portfolios in question to be “well
diversified.” The second three columns report the same summary statistics for the average
of the absolute values of the weights in the portfo-lios. This measures how large the
portfolios are in the 1-norm.

By almost any criteria these portfolio weights seem large. On average across the
different subsets of assets, between 40% and 50% of the value of the portfolio is
invested in a single asset. The percentage of value invested in one asset does not seem
t o decrease as the number of assets in the portfolio grows.

Furthermore, this lack of diversification does not appear to be the result of simply
“loading up” on one or two assets. In the fifty-asset case, where equal weighting puts
2% in each asset, the average absolute value of the weights varies between 8.5% and
24%. Variance minimization appears to be achieved by selling a substantial number of
the stocks short and using the proceeds t o finance ever larger long positions in a
substantial number of other stocks.
Portfolios on the positively sloped portion of the frontier tend t o have weights that
are slightly more extreme, but of similar orders of magnitude and qualitatively
similar, t o those in the global minimum-variance portfolios. lo Since the global
minimum-variance portfolio is computed without reference t o the mean returns, this
suggests that estimation error in the

This is documented in the working paper versions of this paper, available from the authors on
request.
Table I
Characteristics of Weights in the Global Minimum-VariancePortfolio Using 60 m%
Months of Data for Three Random Samples of CRSP Firms from 1/71-12/75, R
1/76-12/80, and 1/81-12/85. P
The body of the table reports measures of t h e size of t h e weights, a n d t h e average correlation between components of t h e <
R2 .
portfolios for various asset universes of various sizes.
Correlation Between R
Maximum Absolute Average Absolute Long and Short m
Value of Weights Value of Weights Positions
Number of assets per subsets (n) 10 30 50 10 30 50 10 30 50

Total number of subsets 90 30 18 90 30 18 90 30 18


Using sample covariance matrix 0.113 0.896 0.974
Max. over all subsets 0.829 0.789 1.219 0.207 0.220 1.000 R
3
Min. over all subsets 0.164 0.201 0.295 0.100 0.062 0.085 NA" 0.855 0.992 R
Mean over all subsets 0.462 0.406 0.588 0.136 0.088 0.128 NA" 0.944 0.997
First component replaced with
constant vector
Max. over all subsets 0.493 0.118 0.073 0.102 0.033 0.020 0.560 0.634 0.654
Min. over all subsets 0.128 0.062 0.039 0.100 0.033 o.020 NA~ NA~ NA~
Mean over all subsets 0.229 0.080 0.050 0.100 0.033 0.020 NA~ NA~ NA'
~

'1 of 90 portfolios had no negative weights.


'Nl b u t one portfolio had no negative weights.
1804 The Journal of Finance

means is not the primary source of poor diversification in efficient portfolios. In


addition, the ability to diversify on subsets, and short one subset to finance large
positions in the other, boosts the mean at very little cost in variance. As a result, the
tradeoff between mean and variance along the frontier, as measured by 1/ yp, increases
dramatically as the number of assets increases.
The effects of diversification on subsets are clearly illustrated in the last three
columns of the upper panel in Table I. Here we take the short and the long positions in
the minimum-variance portfolio and renormalize each by the sum of the negative and
the sum of the positive weights, respectively. Thus, we end up treating the long and
short positions as separate portfolios. These portfolios appear to be very highly
correlated, and the average correlation grows to over 90% as the number of assets
increases to 50. This suggests that an important source of the extreme weights lies in
how the assets are related to some single source of risk, or factor. This factor can, in turn,
be associated with the “dominant” eigenvalue which various empirical studies have shown
tends to grow much faster than any of the others (see, e.g., Trzcinka (1986)).11

As a check on the importance of these phenomena in determining the weights, we


performed the experiment documented in the lower panel of Table I. Let +bk be the h th
largest eigenvalue of the covariance matrix and let
ak denote the associated eigenvector. We can decompose the covariance matrix as
follows (see, e.g., Morrison (1976)):

We constructed new matrices, denoted V*, by subtracting off the first princi-pal
component and replacing it with an equicorrelated component, a vector of constants
with the same norm as the original:”

Examination of the lower panel of Table I shows that with this one substitution the
frontier tends to look well diversified. The minimum-variance portfolio has all
positive weights in all but one case with 30 or 50

Brown (1989) argues that dominance of a single eigenvalue could arise even in a model with
multiple “equally important” factors. The “betas” in his model, however, are drawn from the same i.i.d.
cross-sectional distribution, so in large cross sections the factors are, in fact, close to linearly dependent.
I?.
This operation need not preserve the relative importance of the largest principal component. It may be that
the other eigenvectors will not be close t o orthogonal t o q , if the original eigenvector, al,is sufficiently
different from the constant vector. We calculated the percentage of trace due to the first eigenvalue both before
and after the substitution described above, to check whether the relative influence of this component was
affected appreciably. It does not appear to be.
Mean-Variance Efficiency and Diversification 1805

assets. Over the 18 cases with 50 assets, the maximum amount ever placed in a single
asset is 7.3%.13
We view these results as evidence that the “first factor,” or “dominant principal
component” is, t o a surprising degree, the source of the extreme weightings in
frontier portfolios. Substitution for the first principal compo-nent eliminated the high
correlation between long and short positions in the minimum-variance portfolio. We
emphasize that we are not “recommending” this substitution in applications. Given
the dominance of the first factor, as long as the cross-sectional heterogeneity in the
betas is real, and not simply due to sampl‘ng error, it can and should be exploited in
minimizing variance.
The use o i mean-variance analysis in “asset allocation” frequently involves
constructing portfolios of assets that are themselves reasonably well diver-sified. The
reduction in residual variance that is achieved by employing portfolios instead of
individual assets should reduce sampling error in the parameter estimates. Table I1
suggests that it does not reduce the extreme magnitudes of the weights; it aggravates
the phenomenon. Even if formation of portfolios reduces measurement errors in the
first principal component, there appears to be enough cross-sectional variation in the
loadings to construct portfolios that are uncorrelated with the first factor. There is very
little other variance that is not eliminated simply through diversification.

The weights reported in the upper panel of this table were calculated using
monthly data from twenty portfolios formed on the basis of size using all

Table I1
Characteristics of Weights in the Global Minimum-
Variance Portfolio

Maximum Absolute Value Average Absolute Value


20 Size-based portfolios
1/71-12/80 1.160 0.539
1/71-12/75 1.622 0.887
1/76-12/80 1.678 0.531
20 Beta-based portfolios
1/76-12/85 0.635 0.191
1/76-12/80 0.799 0.282
1/81-12/85 0.614 0.195

13The portfolios on the positively sloped portion of the frontier also have weights that are much less extreme with
this substitution, as is shown in working paper versions of this paper, available from the authors on request. The greater
degree of diversification in the weights with this substitution for the first principal component is similar t o a result
obtained by Frost and Savarino (1986). Their Bayesian procedure involves prior beliefs of a single factor model with
constant betas, and they find in simulations that the resulting efficient portfolios perform much better out of sample than
those which use the sample covariance matrix. In the simulation, however, the “true” moments are based on historical
estimates of the means and the covariance matrix. Thus the Bayesian pulls the covariance estimates away from the
“truth” covariance matrix and towards the covariances matrix employed in the lower panels of Table I. Whether this is
appropriate depends on the priors that the betas are, in fact, homogeneous.
1806 The Journal of Finance

firms on the CRSP daily returns file.14 The second set of portfolios was constructed
from a random sample of CRSP firms with continuous histories on the monthly
returns file from 1971-1985. For each of these firms the first sixty months of returns
were regressed on the equally weighted CRSP market index. The firms were then
ranked according to their “betas” and divided into twenty equally weighted portfolios
of eleven firms each. The sample means and sample covariances of these portfolios
over the subsequent ten years were used to compute the weights in the efficient
portfolios, which are reported in the lower panel. In each case, we find that portfolios
which are first diversified naively, and then combined to minimize variance, tend t o
be weighted in an extreme manner. The table also contains descriptive statistics for
the weights in a portfolio on the positively sloped portion of the frontier. These are, if
anything, slightly less extreme than those in the global minimum-variance portfolio,
again indicating that the basic problem is not with the estimated mean returns.

The analysis in Section 1II.A suggests that, at least in sample data, as the number
of assets grows it becomes possible to construct portfolios that are orthogonal to the
dominant principal component and yet diversify away most of the risk not associated
with this factor. The following question arises: Is this simply due to measurement error in
the “loadings” on the dominant factor, or is it a characteristic of the (unconditional)
population moments? In applications, if one believes this behavior is the result of purely
statistical error, one might try to “control” for it by constraining the weights. If it is
characteristic of the population moments, one might try to exploit the ability to diversify
on subsets and still reduce factor risk as a strategy to achieve low variance and high
return.

At a minimum we can ask whether there is so much noise in the data that we cannot
reliably attribute any of the heterogeneity we observe in the betas to the characteristics of
the population moments. Under the assumption of multivariate normality, we can test the
hypothesis that the first principal component is a constant vector, or an equicorrelated
component. Such tests account explicitly for sampling error in the components of the
eigenvector. We employ the test described in Morrison (1976, p. 295). Let a, denote the
ith eigenvector of the covariance matrix. Under the null hypothesis that a, = qn - l I 2 , the
following statistic has an asymptotic (as T + m) chi-squared distribution with n - 1 degrees
of freedom:

(59)
x 2 = T(+,n-lq‘S-’q + n-v,‘q‘Sq - 2)

Table I11 shows the results of a number of tests using this statistic. We first
computed this statistic for sets of 30 and 50 assets. The chi-squared statistics reported
in Table I11 are the minimum values attained over these subsets. The associated
probability values reject the hypothesis that the first principal component is the
equally weighted portfolio at very high levels of confidence. The marginal probability
that a given one of the chi-squares, across the
14 We are very grateful to Eric Sim for providing these data.
Mean-Variance Efficiency and Diversification 1807

Table 111
Tests for a Constant First Principal Component
The null hypothesis is that the first principal component is proportional to a constant
vector.

Number of d.f.
Subsets Minimum ,y2’
Subsets of 30 assets 30 160.8Zb 29
(time periods aggregated) 18 859.6Bb 49
Subsets of 50 assets
(time periods aggregated)
20 Beta-based portfolios 204.45 19
1/76-12/85 1
1/76-12/80 1 280.33 19
1/81-12/85 1 91.34 19
20 Size-based portfolios 317.59 19
1/71-12/80
1/71-12/75 231.48 19
1/76-12/80 191.81 19
2 Beta-based portfolios 1 112.98 1
1/76-12/85
1/76-12/80 1 101.61 1
1/81-12/85 1 26.04 1
2 Size-based portfolios 1 101.10 1
1/71-12/80
1/71-12/75 1 61.66 1
1/76-12/80 1 31.66 1

aAll the chi-squared statistics listed reject the null hypothesis of a constant first
eigenvector at extremely high levels of confidence.
bThe probability that all the x 2 ’ s across the subsets exceed this value is bounded
above by the marginal probability that any one exceeds this value. The marginal
probability under the null is very close t o zero.

subsets, would attain a value as large or greater than those in the table is virtually
zero; and this probability provides an upper bound on the probabil-ity that, under the
null hypothesis, all of the chi-squares would be above this critical value. These tests
are not independent across the 30-asset and 50-asset groupings, of course. The same
return data used to form the 30-asset subsets are simply regrouped to form the 50-
asset subsets.
Performing these tests with individual assets fails, t o some extent, to tell us what
we want to know: can diversity in the betas be maintained while diversifying on
subsets so as to eliminate residual variance? Accordingly, we also perform the tests
using large portfolios of stocks. These results are reported in the lower panels of
Table 111. For sets of size-based portfolios, the null hypothesis that the first principal
component is a constant vector is rejected at extremely h g h levels of confidence.
Beta-based portfolios were also used to perform tests such as those in Table 11. These
portfolios are constructed, using out-of-sample data, to have cross-sectional variation
in the betas on the equally weighted portfolio. The equally weighted portfolio is, in
1808 The Journal o f Finance

turn, highly correlated with the first principal component. In all these cases, the
hypothesis that the first principal component is constant is rejected. This evidence
suggests that the extreme weightings we see in variance-minimizing portfolios
constructed with sample data are not just a result of measurement error.

With the possible exception of the tests just reported, an important limita-tion of
our analysis is its descriptive nature, and the lack of formal tests of the hypothesis
advanced regarding the source of the extreme weightings observed in sample
efficient frontiers. Given the difficulties associated with posing this hypothesis in a
manner amenable to formal inference, we have attempted instead to take a range of
different approaches to documenting the phenomenon, and evaluating its robustness.

IV.Conclusions
We have shown that the existence of well-diversified portfolios on the frontier is
equivalent to conditions that bound excess returns, over the zero-beta rate, with
measures of the degree to which a portfolio covaries with the returns on all the
individual assets. For there t o be no well-diversified portfolios anywhere on the
frontier, there must exist portfolios for which these measures are less than the
variance of the minimum-variance portfolio. These bounds then provide simple means
of illustrating the relationship between mean-variance efficiency, APT pricing, and
diversification. We also investigate the sources of the empirical behaviors of efficient
portfolios’ weights. These appear to be attributable to variation across the elements of
the eigenvector associated with the first principal component, and the ability to form
well-&versified portfolios on subsets of stocks that have different loadings on this
factor. This allows for reduction of both factor risk and idiosyncratic risk, but also
tends to result in extreme portfolio weightings.

REFERENCES

Best, M. J. and R. R. Grauer, 1991, On the sensitivity of mean-variance efficient portfolios to changes in
asset means: Some analytical and computational results, Reoiew of Financial Studies 4, 315-342.

Black, F. and R. Litterman, 1990, Asset allocation: Combining investor views with market equilibrium,
Technical report, Goldman Sachs Fixed Income Research.
Brown, S. J., 1989, The number of factors in security returns, Journal of Finance 44,1247-1262.
Chamberlain, G.,1983, Funds, factors and diversification in arbitrage pricing models, Economet-
rica 51, 1305-1324.
_-__ and M. Rothachild, 1983, Arbitrage, factor structure, and mean-variance analysis on large asset
markets, Econometrica 51, 1281-1304.
Frost, P. A. and J. E. Savarino, 1986, An empirical Bayes approach to efficient portfolio selection, Journal
of Financial and Quantitatiue Analysis 21, 293-306.
, 1988, For better performance: Constrain portfolio weights, Journal of Portfolio Manage-ment
14, 29-34.
Hansen, L. P. and S. F. Richard, 1987, The role of conditioning information in deducing testable
restrictions implied by dynamic asset pricing models, Econometrica 55, 587-613.
Mean-Variance Efficiency and Diversification 1809

Huberman, G., S. Kandel, and R. F. Stambaugh, 1987, Mimicking portfolios and exact arbitrage pricing,
Journal of Finance 42, 1-10,
Kallberg, J. G. and W. T. Ziemba, 1983, Comparison of alternative utility functions in portfolio selection
problems, Management Science 29, 1257-1276.
Kroll, Y.,H . Levy, and H. M. Markowitz, 1984, Mean-variance versus direct utility maximiza-tion,
Journal of Finance 39, 47-61.
Merton, R. C., 1972, An analytic derivation of the efficient portfolio frontier, Journal of Financial a n d
Quantitatzue Analysis 7, 1851-1872.
Morrison, D. G., 1976, Multivariate Statistical Methods (McGraw-Hill, New York).
Pulley, L. B., 1981, General mean-variance approximation to expected utility for short holding periods,
Journal of Financial a n d Quantitatiue Analysis 16, 361-373.
Ross, S. A,, 1976, The arbitrage theory of capital asset pricing, Journal of Economic Theory 13, 341-360.

Shukla, R. and C. Trzcinka, 1990, Sequential tests of the Arbitrage Pricing Theory: A comparison of
principal components and maximum likelihood factors, Journal ofFinance 45, 1541-1564.
Stewart, G. W., 1973, Introduction to Matrix Computations (Academic Press, New York). Trzcinka, C . ,
1986, On the number of factors in the arbitrage pricing model, Journal of Finance
41, 347-368.

You might also like