4126989

Robust Portfolio Selection Problems Author(s): D. Goldfarb and G. Iyengar Source: Mathematics of Operations Research, Vol. 28, No.
1 (Feb., 2003), pp. 1-38 Published by: INFORMS Stable URL: http://www.jstor.org/stable/4126989 . Accessed: 28/07/2011 05:17
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=informs. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Mathematics of Operations Research.
http://www.jstor.org
OF MATHEMATICS OPERATIONSRESEARCH Vol. 28, No. 1, February2003, pp. 1-38 Printed in U.S.A.
ROBUST PORTFOLIO SELECTION PROBLEMS

D. GOLDFARB ANDG. IYENGAR
In this paper we show how to formulate and solve robust portfolio selection problems. The objective of these robust formulations is to systematically combat the sensitivity of the optimal portfolio to statistical and modeling errors in the estimates of the relevant market parameters.We introduce "uncertaintystructures"for the market parameters and show that the robust portfolio selection problems corresponding to these uncertainty structurescan be reformulated as secondorder cone programsand, therefore, the computationaleffort requiredto solve them is comparable to that requiredfor solving convex quadraticprograms. Moreover, we show that these uncertainty structurescorrespondto confidence regions associated with the statistical procedures employed to estimate the market parameters.Finally, we demonstratea simple recipe for efficiently computing robust portfolios given raw marketdata and a desired level of confidence.
1. Introduction. Portfolio selection is the problem of allocating capital over a number on of available assets in order to maximize the "return" the investment while minimizing the "risk."Although the benefits of diversification in reducing risk have been appreciated since the inception of financial markets,the first mathematicalmodel for portfolio selection was formulatedby Markowitz(1952, 1959). In the Markowitzportfolio selection model, the on "return" a portfolio is measured by the expected value of the random portfolio return, and the associated "risk"is quantified by the variance of the portfolio return. Markowitz showed that, given either an upper bound on the risk that the investor is willing to take or a lower bound on the returnthe investor is willing to accept, the optimal portfolio can be obtained by solving a convex quadraticprogrammingproblem. This mean-variancemodel has had a profound impact on the economic modeling of financial markets and the pricing of assets-the CapitalAsset Pricing Model (CAPM) developed primarilyby Sharpe (1964), Lintner(1965), and Mossin (1966) was an immediate logical consequence of the Markowitz theory. In 1990, Sharpe and Markowitz shared the Nobel Memorial Prize in Economic Sciences for their work on portfolio allocation and asset pricing. Despite the theoretical success of the mean-variance model, practitioners have shied away from this model. The following quote from Michaud (1998) summarizesthe problem: "AlthoughMarkowitzefficiency is a convenient and useful theoreticalframeworkfor portfolio optimality,in practice it is an error-prone procedurethat often results in error-maximized and investment-irrelevant portfolios."This behavior is a reflection of the fact that solutions in of optimization problems are often very sensitive to perturbations the parametersof the problem; since the estimates of the market parametersare subject to statistical errors, the results of the subsequent optimization are not very reliable. Various aspects of this phenomenon have been extensively studied in the literatureon portfolio selection. Chopra and Ziemba (1993) studies the cash-equivalentloss from the use of estimatedparametersinstead of the true parameters.Broadie (1993) investigates the influence of errors on the efficient frontier, and Chopra (1993) investigates the turnover in the composition of the optimal
Received January 1, 2002; revised May 25, 2002, and July 24, 2002. MSC 2000 subject classification. Primary:91B28; secondary:90C20, 90C22. OR/MSsubject classification. Primary:Finance/portfolio;secondary:programming/quadratic. Key words. Robust optimization, mean-varianceportfolio selection, value-at-riskportfolio selection, second-order cone programming,linear regression. 1 0364-765X/03/2801/0001/$05.00 1526-5471 electronic ISSN, ? 2003, INFORMS
D. GOLDFARB AND G. IYENGAR
portfolio as a function of the estimation error(see also Part II of Ziemba and Mulvey 1998 for a summaryof this research). Several techniqueshave been suggested to reduce the sensitivity of the Markowitz-optimal portfolios to input uncertainty:Chopra (1993) and Frost and Savarino (1988) propose constraining portfolio weights, Chopra et al. (1993) proposes using a James-Stein estimator (see Huber 1981 for details on Stein estimation) for the means, while Klein and Bawa (1976), Frost and Savarino(1986), and Black and Litterman(1990) suggest Bayesian estimation of means and covariances. Although these techniques reduce the sensitivity of the portfolio composition to the parameterestimates, they are not able to provide any guarantees on the risk-returnperformanceof the portfolio. Michaud (1998) suggests resampling the mean returns pi and the covariance matrix I of the assets from a confidence region around a nominal set of parameters,and then aggregating the portfolios obtained by solving a Markowitzproblem for each sample. Recently scenario-basedstochastic programming models have also been proposed for handling the uncertaintyin parameters(see Part V of Ziemba and Mulvey 1998 for a survey of this research). Both the sampling-based and the scenario-basedapproachesdo not provide any hard guaranteeson the portfolio performance and become very inefficient as the number of assets grows. In this paper we propose alternative deterministic models that are robust to parameter uncertainty and estimation errors. In this framework, the perturbationsin the market parametersare modeled as unknown, but bounded, and optimization problems are solved assuming worst case behavior of these perturbations.This robust optimization framework was introducedin Ben-Tal and Nemirovski (1999) for linear programmingand in Ben-Tal and Nemirovski (1998) for general convex programming(see also Ben-Tal and Nemirovski 2001). There is also a parallel literatureon robust formulations of optimization problems originatingfrom robust control (see El Ghaoui and Lebret 1997, El Ghaoui et al. 1998, and El Ghaoui and Niculescu 1999). Our contributionsin this paper are as follows: (a) We develop a robust factor model for the asset returns. In this model the vector of random asset returnsre Rn is given by r= +V + e,
where pL R is the vector of mean returns,f E R"' is the vector of random returns of the E is m(< n) factors that drive the market, V E RmXn the factor loading matrix and F is the vector of residual returns.The mean returnvector p, the factor loading matrix V, and the covariance matrices of the factor returnvector f and the residual error vector e are known to lie within suitably defined uncertaintysets. For this market model, we formulate robust analogs of classical mean-varianceand value-at-riskportfolio selection problems. (b) We show that the naturaluncertaintysets for the market parametersare defined by the statistical proceduresused to estimate these parametersfrom market return data. This class of uncertaintysets is completely parametrizedby the market data and a parameterw that controls the confidence level, thereby allowing one to provide probabilistic guarantees on the performance of the robust portfolios. In all previous work on robust optimization (Ben-Tal and Nemirovski 1998, 1999; El Ghaoui and Lebret 1997; El Ghaoui et al. 1998; Halld6rssonand Tiittincti2000), the uncertaintysets for the parametersare assumed to have a certain structurewithout any explicit justification. Also, there is no discussion of how these sets are parametrizedfrom raw data. (c) We show that the robustoptimization problems correspondingto the naturalclass of uncertaintysets (defined by the estimation procedures)can be reformulatedas second-order cone programs (SOCPs). SOCPs can be solved very efficiently using interior point algorithms (Nesterov and Nemirovski 1993, Lobo et al. 1998, Sttirm 1999) In fact, both the worst case and practical computationaleffort requiredto solve an SOCP is comparable to
ROBUST PORTFOLIOSELECTIONPROBLEMS
that for solving a convex quadraticprogram of similar size and structure;i.e., in practice, the computationaleffort requiredto solve these robustportfolio selection problems is comparable to that required to solve the classical Markowitz mean-varianceportfolio selection problems. In a recent related paper, Halld6rsson and Tiittincti (2000) show that if the uncertain mean returnvector tL and the uncertaincovariance matrix I of the asset returns r belong to the component-wise uncertaintysets Sm = {(J: IL <I I U} and S, = {I: I~ 0, L I <I lu}, respectively, the robust problem reduces to a nonlinear saddle-point problem that involves semidefinite constraints.Here A > 0 (resp. >-0) denotes that the matrix A is symmetricand positive semidefinite(resp. definite). This approachhas several shortcomings when applied to practicalproblems-the model is not a factor model (in applied work factor models are popular because of the econometric relevance of the factors), no procedure is provided for specifying the extreme values (LJ, Ixu) and (IL, ~U) defining the uncertainty structureand, moreover,the solution algorithm,although polynomial, is not practical when the number of assets is large. A multiperiod robust model, where the uncertaintysets are finite sets, was proposed in Ben-Tal et al. (2000). The organizationof the paper is as follows. In ?2 we introduce the robust factor model and uncertaintysets for the mean returnvector, the factor loading matrix,and the covariance matrix of the residual return.We also formulate robust counterpartsof the mean-variance optimal portfolio selection problem, the maximum Sharpe ratio portfolio selection problem, and the value-at-risk(VaR) portfolio selection problem. The uncertaintysets introduced in this section are ellipsoidal (intervals in the one-dimensional case) and may appear quite arbitrary.Before demonstratingthat these sets are, indeed, natural, we first establish in ?3 that the robust mean-varianceportfolio selection problem in markets where the factor loading and mean returns are uncertain,but the factor covariance is known and fixed, can be reformulatedas an SOCP. An SOCP formulation of the robust maximum Sharpe ratio problemin such marketsfollows as a corollary.In ?4 we develop an SOCP reformulationfor the robust VaR portfolio selection problem. In ?5 we justify the uncertaintysets introduced in ?2 by relating them to linear regression. Specifically, we show that the uncertaintysets correspondto confidence regions aroundthe least-squaresestimate of the marketparameters and can be constructed to reflect any desired confidence level. In this section, we collect the results from the preceding sections and present a recipe for robust portfolio allocation that closely parallels the classical one. In ?6 we improve the factor model by allowing uncertaintyin the factor covariance matrix and show that, for naturalclasses of uncertainty sets, all robustportfolio allocation problems continue to be SOCPs. We also show that these natural classes of uncertainty sets correspond to the confidence regions associated with maximum likelihood estimation of the covariance matrix. In ?7 we present results of some of preliminarycomputationalexperimentswith our robust portfolio allocation framework. 2. Market model and robust investment problems. We assume that the marketopens for trading at discrete instants in time and has n traded assets. The vector of asset returns over a single marketperiod is denoted by r E Rn, with the interpretation that asset i returns + ri) dollars for every dollar invested in it. The returnson the assets in different market (1 periods are assumed to be independent. The single period return r is assumed to be a random variable given by (1) r= + V'f + ,
where F.LeRn is the vector of mean returns,f -- \N(O,F) E Rm is the vector of returns of the factors that drive the market,V E Rmxnis the matrix of factor loadings of the n assets, and f - N(O, D) is the vector of residual returns. Here x - N(Fp, ) denotes that x is a multivariatenormal random variable with mean vector Fp and covariance matrix 1.
In addition, we assume that the vector of residual returnsF is independentof the vector of factor returnsf, the covariancematrix F >-0 and the covariance matrix D = diag(d) 0, _ i.e. di > 0, i = 1,.... , n. Thus, the vector of asset returnsr -- N(I, V'FV + D). Although not requiredfor the mathematicaldevelopment in this paper,the eigenvalues of the residual covariance matrix D are typically much smaller than those of the covariance matrix VTFV implied by the factors; i.e., the V'FV is a good low-rank approximationof the covariance to the asset returns. Except in ?6, the covariancematrix F of the factor returnsf is assumed to be stable and known exactly. The individualdiagonal elements di of the covariancematrix D are assumed to lie in an interval [_d,di], i.e., the uncertaintyset Sd for the matrix D is given by (2)
Sd = {D:D = diag(d), di E [di, di], i = 1,...,
n}.
The columns of the matrix V, i.e., the factor loadings of the individual assets, are also assumed to be known approximately.In particular,V belongs to the elliptical uncertainty set S, given by
(3)
S, = IV V = Vo+W, IWl
, i = ,...,n),
= where Wi is the ith column of W and Ilwllg N/WrGw denotes the elliptic norm of w with to a symmetric,positive definite matrix G. respect The mean returnsvector ti is assumed to lie in the uncertaintyset Sm given by = (F: = yi, i = 1, ..., n}; FI -Io + g, il j i.e., each component of 1L is assumed to lie within a certain interval. The choice of the uncertaintysets is motivated by the fact that the factor loadings and the mean returns of assets are estimated by linear regression. The justification of the uncertaintystructuresand suitable choices for the matrix G, and the bounds Pi, Yi, di, di, i = 1,... , n, are discussed in ?5. An investor's position in this market is described by a portfolio ( E R", where the ith component 4i representsthe fraction of total wealth invested in asset i. The return r, on the portfolio 4 is given by (4)
Sm
r, = r'
= T +f V + +V
-~ N(C ~
, T(VTFV + D)4).
The objective of the investor is to choose a portfolio that maximizes the "return"on the investment subject to some constraintson the "risk" of the investment. The mathematical model for portfolio selection proposedby Markowitz(1952, 1959) assumes that the expected value E[r] of the asset returns and the covariance Var[r] are known with certainty. In this model investment "return"is the expected value E[r,] of the portfolio return and the associated "risk"is the variance Var[r,]. The objective of the investor is to choose a portfolio 4* that has the minimum variance among those that have expected returnat least a i.e., 4* is the optimal solution of the convex quadraticoptimization problem minimize Var[r4,], subject to E[r,] > a, 1T = 1.
(5)
As was pointed out in the introduction, the primary criticism leveled against the Markowitz model is that the optimal portfolio 4* is extremely sensitive to the market parameters(E[r], Var[r])-since these parametersare estimated from noisy data, 4* often amplifies noise. By introducing "measures of uncertainty"in the market models, we are
attempting to correct this sensitivity to perturbations.The uncertainty sets Sm, S, and Sd representthe uncertaintyof our limited (inexact) informationof the marketparameters,and we wish to select portfolios that perform well for all parametervalues that are consistent with this limited information. Such portfolios are solutions of appropriatelydefined minmax optimization problems called robustportfolio selection problems. The robust analog of the Markowitzmean-varianceoptimizationproblem (5) is given by minimize (6) subject to max
{VESV, DESd)
Var[rr],
min E[r+] a, {tIESm1 > 1T+ = 1.
The objective of the robustminimumvarianceportfolio selection problem (6) is to minimize the worst case varianceof the portfolio subject to the constraintthat the worst case expected returnon the portfolio is at least a. We expect that the sensitivity of the optimal solution of this mathematicalprogramto parameterfluctuationswill be significantly smaller than it would be for its classical counterpart(5). A closely related problem, the robustmaximumreturnproblem, is the dual of (6). In this problem, the objective is to maximize the worst case expected returnsubject to a constraint on the worst case variance, i.e., to solve the mathematicalprogram maximize (7) min E[r], {JILESm max Var[rc] r subject to {VES,, DESdI 1T~ = 1. Another variant of the robust optimization problem (6) is the robust maximum Sharpe ratio problem. Here the objective is to choose a portfolio that maximizes the worst case ratio of the expected excess returnon the portfolio, i.e., the returnin excess of the risk-free rate rf, to the standarddeviation of the return.The correspondingmax-min problem is given by (8) maximize =
{W:1 lT=i
}
A,
VES,, {.LESm, DESd}
min
Var[ro] J
All these variants are studied in ?3. We show that for the uncertainty sets Sd, Sm, and S,, defined in (2)-(4) above, all of these problems reduce to SOCPs. Although the factor model (1) is crucial for the SOCP reduction,the assumptionthat factor and residual returns are normally distributedcan be relaxed. From the results in Bertsimas and Sethuraman (2000) it follows that all of our results continue to hold with slight modifications if one assumes that the distributionsare unknown but second moments of the factor and residual returnslie in the sets S, and Sd, respectively. We leave this extension to the reader. In ?4 we study robust portfolio selection with value-at-risk(VaR) constraintswhere the objective is to maximize the worst-case expected return of the portfolio subject to the constraint that the probability of the return r4 falling below a threshold a is less than a prescribedlimit; i.e., the objective is to solve the following mathematicalprogram. maximize subject to E[rm], min
,LESm, {VESv, P(r+ DESd}
max
< a) < /3.
VaR was introduced as a performance analysis tool in the context of risk management. Recently there has been a growing interest in imposing VaR-type constraints while optimizing credit risk (Kast et al. 1998, Mausser and Rosen 1999, Andersson et al. 2001). We
D. GOLDFARBAND G. IYENGAR
show that, for the uncertaintysets defined in (2)-(4), this problem can also be reduced to an SOCP and, hence, can be solved very efficiently. Our technique can be extended to another closely related performance measure called the conditional VaR by using the results in Mausser and Rosen (1999). El Ghaoui et al. (2002) presents an alternativerobust approach to VaR problems that results in semidefinite programs. The optimizationproblems of interesthere fall in the general class of robustconvex optimization problems. Ben-Tal and Nemirovski (1998, 2001) propose the following structure for generic robust optimizationproblem.
minimize
c'x,
subjectto
F(x, g) EX CR",
Vg E V ,
where g are the uncertainparametersin the problem, x e R" is the decision vector, X3is a convex cone and, for fixed ~ e U, the function F(x, g) is Y-concave. The essential ideas leading to this formalism were developed in robust control (see Zhou et al. 1996 and references therein). Robustness was introduced to mathematical by programming Ben-Tal and Nemirovski (1998, 1999, 2001). They establishedthat for suitably defined uncertaintysets 'U, the robust counterpartsof linear programs,quadraticprograms,and general convex programsare themselves tractableoptimizationproblems. Robust least-squaresproblems and robust semidefinite programswere independentlystudied by El Ghaoui and his collaborators(El Ghaoui and Lebret 1997, El Ghaoui et al. 1998). Ben-Tal et al. (2000) have studied robust modeling of multistage portfolio problems. Halld6rsson and Tiittincti(2000) study robustinvestmentproblems that reduce to saddle point problems. 3. Robust mean-variance portfolio selection. This section begins with a detailed analysis of the robust minimum variance problem (6). It is shown that for uncertainty sets defined in (2)-(4) this problem reduces to an efficiently solvable SOCP. This result is subsequently extended to the maximum return problem (7) and the robust maximum Sharpe ratio problem (8). 3.1. Robust minimum variance problem. r. -- N(IT4, Since the return
+D)+),
+(VrFV r
the robust minimum variance portfolio selection problem (6) is given by minimize (11) subject to
{VES,}+TVTFV+4
max {
+ max t(+D+4 , {DESdI
min IT'4) a, {I=ES 1. 1T,0= 1.
where D = diag(d). Also, since the The bounds di < di < di imply that +TD+ : :TD4, covariance matrix F of the factor f is assumed to be strictly positive definite, the function :x IlxlIf J/VFxdefines a norm on Rm. Thus, (11) is equivalentto the robust augmented least-squaresproblem minimize (12) max IIV|II ++ D4, } {VES subject to min I4T) > a, } J=ESm 1T4?= 1.
PORTFOLIO SELECTION PROBLEMS ROBUST
By introducingauxiliary variables v and 8, the robust problem (12) can be reformulatedas minimize subject to (13) v + 8, max IIV+ VES, } + 4D) < min IxrLT (FLESm 1T = v, 5, a,
1.
and S,m= {(i1, ,,r},
...
If the uncertainty sets S, and Sm are finite, i.e., S, = {V, ... V ..,
then (13) reduces to the convex quadraticallyconstrainedproblem minimize A+ 8,
to subject
(14)
iVkll2
T 'DT
4?Tpk
A, forall k = 1,..
8, > a, forallk=1 ..r.
s,
r,
14 = 1.
This problem can be easily converted to an SOCP (see Lobo et al. 1998 or Nesterov and Nemirovski 1993 for details). In El Ghaoui and Lebret (1997) it was shown that for
? = S, = V:V = V +W, IIWll Tr(WTW) p},

the problem can still be reformulatedas an SOCP. Methodologically speaking, our results can be viewed as an extension of the results in El Ghaoui and Lebret (1997) to other classes of uncertaintysets more suited to the application at hand. When the uncertaintyin V and p is specified by (3) and (4), respectively, the worst case mean returnof a fixed portfolio <4is given by,
(15)
{ILESm
m min
yT1;
and the worst case variance is given by,
maximize II(V0 +W)4II2

i= 1,..., n. < Since the constraints IIWll|g Pi, i = 1, ... n imply the bound, subject to
IjWill" Pi,
n1
(17)
i= IWIIg
i=1 g i=Y 1EEiWiV i=1 i
Wil=g i=l <
-i
(il1
the optimization problem,
least as large as that of (16).
maximize IIVo~+wI, subjectto I|wl -r where r = pT 141 Z =1Pi I0iI is a relaxation of (16), i.e., the optimal value of (18) is at =
AND G. IYENGAR D. GOLDFARB
The objective function in (18) is convex; therefore, the optimal solution w* lies on the n, define 1g boundaryof the feasible set, i.e., IIw* = r. For i = 1, .... A
i
*i
(19)
W =
I Ir PiW
r
Piw*,
otherwise. =
= and W* Then Pi, for all i= 1,... , n; i.e., W* is feasible for (16), n IIW*llg Therefore,the optimal value of (16) and (18) are, in fact, the same. = w*. E =1 iWf Thus, for a fixed portfolio 4, the worst-case variance is less than v if, and only if,
(20)
II {y: Ilyllg_<r IlYyo+
max
v,
where yo = Vo0 and r = pT1|. The following lemma reformulatesthis constraintas a collection of linearequalities, linear inequalities and restricted hyperbolic constraints (i.e. constraints of the form: zTz < xy, z E Rn, x, y ER and x, y 0). LEMMA Let r, v > 0, Y0, y E Rm and F, G E Rmxm be positive definite matrices. Then 1. the constraint
(21)
< ()y:llyllg<r}f max Ilyo+YII -V
is equivalent to either of the following: (i) there exist r, o > 0, and t E R, that satisfy
v >7 + 1't,
< or -
1
(H)' Amax Ai)ti,
(22)
r2
w?
<0o', (1 -
i= 1
<
...
m,
where QAQT is the spectral decomposition of H = G-1/2FG-1/2, A = diag(Ai), and w =

QTH1/2Gl/2yo;
(ii) there exist 7
>
0 and s E Rm that satisfy

r2
7(-
S)
i = 1 .. ,
m,
(23)
u < (1 - TOi)s, 1
Amax(K)'
where POPT is the spectral decomposition of K = F'/2G-'FI/2, pTFl/2yo.

PROOF. By setting y = ry, we have that (21) is equivalent to (24) (v - y Fyo) - 2ryTFy - r2 2TFy 0,
O = diag(Oi), and u =
for all y such that 1 - yTGy > 0. Before proceeding further,we need the following:
LEMMA 2 (SF-PROCEDURE). Let Fi(x) = XTAix+ 2bix + ci, i = 0, ... , p be quadratic functions of x E Rn. Then Fo(x) > 0 for all x such that Fi(x) > 0, i = 1, .... p, if there exist 7i > 0 such that T bT b [co o. [c:
bA Ao
Moreover,if p = 1 then the converse holds if there exists xo such that F, (xo) > 0. For a discussion of the SF-procedure and its applications, see Boyd et al. (1994). Since = 0 is strictly feasible for I - TGy > 0, the SF-procedure 1 implies that (24) holds for all 1 - yGy > 0 if and only if there exists a 7 > 0 such that (25) M=
_'i=l
A_
[v
7 yf Fyo
-rFyo
- ryoTF 2F
"G-- r2F-
0-.
define w = QTHI/2G/2y/2 = Al/2QTGl/2yo.
Let the spectral decomposition of H = G-'1/2FG-1/2be QAQ', where A = diag(k), and

Observing that yTFyo = WTW,we have that the
matrix M >- 0 if and only if
M-[
0
T
QTG-1/2
][I1
0
O][
G-1/2Q
v- r-w7Tw-rwTA1/2]
-rAl/2W
- r2A
The matrix M >-0 if and only if 7 > r2Ai, for all i = 1,..., m (i.e., 7 > r2,max(H)), wi = 0 for all i such that 7 = r2Ai, and the Schur complement of the nonzero rows and columns of 7I - r2A, i
\i:7#r2 i
Aiw
T i-r
iE
j)
i:oA
0, I
where a = r2/7. It follows that
max
{y:lyIIlg<r}
2 (Iyo0+ryI
v,
if and only if there exists 7, a > 0 and t E R' satisfying,

v r2 r= > 7"+lt, 7
(26)
w = (1 -
Ai)ti,
i=
..
m,
(H) Amax
It is easy to establish that there exist 7, oa> 0 and t E R' that satisfy (26) if and only if there exist 7, a-> 0, and t E that satisfy (26) with the equalities replaced by inequalities, R+ that satisfy (22). i.e., To establish the second representation,note that (21) holds if and only if (27) yTGy > r2,
v. Since Iyo +yll1 > v for all sufficiently large y, the for all y such that Il|y+yf -2 holds for all > v if and only if there exists a SF-procedureimplies that (27) 1yo0 +Yl2 7 > 0 such that
(28) M=
-r2- -
_"(y'Fyo -7"Fyo
- v)
-"yF F1 G- 7F -
0.
10
Let POPT be the spectral decomposition of K = FI/2G-'F'/2, where 0 = diag(O), and u = pTFl/2yo. Then M >-0 if and only if
S=PTF-1M
0 F-1/2p
--
7 0-1= 0 , m (i.e., 7 < 1/AmIax(K)), ui
However, M >- 0 if and only if 7 < 1/0i, for all i =
1,.... for all i such that 7rO= 1, and the Schur complement of the nonzero row and columns of 0-' - T1,
r2
V)
2 -
(i:,ruirl
7i:- 0i
l-
&
- )
T
2 2
r i:+i,
+7
- , _0
Toi
Hence,
<r) {y: IlYII,
max Ilyo +ryll
v,
if and only if there exists 7 > 0 and s e R' satisfying,

2 T u =??g u = (1 - 7)s,
r2 < 7(v-TS), 1 i= 1
.
(29)
Amax (K)
Completely analogous to the proof of part (i), we have that there exists 7 > 0 and s e R' satisfying (29) if and only if there exist 7 > 0 and s e R' satisfying (23). This proves the second result. O To illustrate the equivalence of parts (i) and (ii) of Lemma 1, we note that if F = KG, then K = H = KI and w = u. Hence, if (7, s) satisfies (23), then -, t) = (7, v - 1Ts, s) (r", satisfies (22). The restrictedhyperbolic constraints,zTz < xy, x, y > 0, can be reformulatedas secondorder cone constraintsas follows (see Nesterov and Nemirovski 1993, ?6.2.3):
(30)
< zTZ xy
4zz < (x + y)2 (xy)2 2
x-yz
< X y.
The above result and Lemma 1 motivate the following definition. 1. DEFINITION Given Vo E Rmxn, and F, GE Rmxm positive definite, define 'C(V0, G) F, to be the set of all vectors (r; v; 4) E R x R x Rn such that (r, v, yo = V04) satisfy (22); that satisfy i.e., there exist o, 7 > 0 and t e R+ 7 + rt v, < 1
[ 2r
1
-
(H) ' Amax
:!S--+
+
1
-
OtAi
)-<ti
i= OAi+? ti, 1,...9 m,
--
where QAQT is the spectral decomposition of H = G-1/2FG-1/2, A = diag(Ai) and w =

QTH1/2Gl/2Vo+.
11
From (15), (20), and Definition 1, it follows that (13) can be reformulatedas minimize subec. t subject to v +6,
2i1/20]11 -1-5 1)+8,

1
: rT,4)1
1?
(31)
py- I41 C,
1+ = 1,
(Vo, F, G).
(r; v; +) E
where the equality r = pT 14)1 been relaxed by recognizing that the relaxed constraint has will always be tight at an optimal solution. Although (31) is a convex optimizationproblem, it is not an SOCP. However, replacing 1I1 by a new variable *4e R" and adding the 2n linear constraints, I/ji > I I1,i = 1,... , n, leads to the following SOCP formulationfor the robust minimum variance portfolio selection problem (11):
minimize
subject to (32)
v + 8,
2D /2 IL
-yO>, i
-i
1+ ,
i= 1
- i i,
i
i = ..... ..., 1,
n,
n,
1T+ = 1,
(p'; v; cb) E i (Vo, F, G). Another possible transformationleading to an SOCP is to replace 4 by 4) = 40+- +_, e . An alternative SOCP formulation is obtained if one uses part (ii) of the ++, -_ R' lemma to characterizethe worst case variance. The derivationof these results is left to the reader. To keep the exposition simple, in the rest of this paper it will assumed that short sales are not allowed; i.e., 4) 0. In each case the result can be extended to general 4 by employing the above transformations. 3.2. Robust maximum return problem. The robustmaximum returnproblem is given by m maximize min ITb,
{ILESm)
subject to
max
{VESv,DESd}+T(VTFV
+D)+ < A,
1,T = 1,
4)>0,
or equivalently,from (15), by
maximize (UL--y)4),
subject to (33) max rTVTrFV+< A 8, 4TD4) 5 8, 1Tl = 1,
+>_o.
12
From Definition 1, it follows that the robust maximum returnproblem is equivalent to the
SOCP
maximize subject to (34)
(IL -
y) r, +-8, 1
1T+
=
<[2
1,
4 0>o, F, (p'4; A- ; +) E Y(Vo, G).

3.3. Robust maximum Sharpe ratio problem. problem is given by, The robust maximum Sharpe ratio
(35)
{:O,0
max minm IT4_rf ,DESdl (~:<b>O,1T~r)(VES,,tLES,, Var[r]

1T=I}
{VESeSmDESd
where rf is the risk-free rate of return.We will assume that the optimal value of this maxmin problem is strictly positive; i.e., there exists a portfolio with finite worst-case variance whose worst-case returnis strictly greater than the risk-free rate rf. In practice the worstcase variance of every asset is bounded; therefore, this constraint qualification reduces to the requirementthat there is at least one asset with worst-case returngreater than rf. Since the componentsof the portfolio vector + add up to 1, the objective of the max-min problem (35), _ (F rf1),4 pT4-r /Var[rj] /VI-ar[ro]' is a homogeneous function of the portfolio 0. This implies that the normalizationcondition = 1 can be dropped andthe constraint 1T4 (I - rfl)T+ = 1 addedto (35) without min(ism) loss of generality.With this transformation,(35) reduces to minimizing the worst case any variance;i.e., (35) is equivalentto minimize
(36) subject to
max ?TVT FV4 + +bTDf , {vES (O - y - rfl)T > 1,
+>0o,
= 1 has been relaxed by recognizing that the where the constraint mint,,s.)(xp rfl)T4 relaxed constraint will always be tight at an optimal solution. Consequently, the robust maximum Sharpe ratio problem is equivalent to a robust minimum variance problem with a = 1, Fp replacedby Fp rf1, and + no longer normalized.Exploiting this relationship,we have that (35) is equivalentto the SOCP minimize (37) subject to
v + 8,
[2D1/2]
(Fo - Y - 1'rfl)T
11
> 1,
(p'+; v; +) E W(Vo,F, G). The crucial step in the reductionof (35) to (37) is the realizationthat the objective function in (35) is homogeneous in 4 and, therefore, its numeratorcan be restricted to 1 without any loss of generality. This homogenization goes through even when there are additional inequality constraintson the portfolio choices.
13
Suppose the portfolio (4 is constrained to satisfy A4)? b (4) > 0 is assumed to be subsumed in A4 > b). Then, the robust maximum Sharpe ratio problem,
(38)
max
minm
1T4=1I} {VESv,,LESm,DESdd)
{t:A>b,
VVar[4]
is equivalent to the robust minimum variance problem

minimize max rTVTrFV4)+ I (VES, TFj),
subjectto (39)
(ixo
- rf1)' > 1 A4 > b,

1T4 =
>o,
where r is an auxiliary variable that has been introduced to homogenize the constraints A4) b and 1'4 = 1. The problem (39) can easily be converted into a second-ordercone problem by using the techniques developed above. 4. Robust Value-at-Risk (VaR) portfolio selection. problem is given by maximize min E[rm], ILESm The robustVaR portfolio selection
(40)
subject to t
VESv, {ILESm, DESd 1T+ = 1,
max
P(r+ < a) -
4)+.
This optimization problem maximizes the expected returnsubject to the constraintthat the shortfall probability is less than P. Note that for ease of exposition we again assume that no short sales are allowed; i.e., 4 > 0. The solution in the general case can be obtained by identical to the one detailed at the end of ?3.1. using a transformation For fixed (FL, D), the returnvector + = (T4T, T(VTFV+ D)4). Therefore, V, r, (41) P(r a)p
P (4+? (IP
P Y<
4v(VTFV+D)
< a) )
P,
-4V/T(VTrFV+ O)
<atT4)
ST(VTFV +D)4)
where 2 -- N(0, 1) is the standardnormal random variable and 3,(.) is its cumulative < density function. In typical VaR applications 3 << 1; therefore -l (13) 0. Thus, the probability constraintfor fixed (FI,V, D) is equivalent to the second-ordercone constraint
(42)
-y
I(p)
Fl/2V+12
+ IDI/2 112
a.
14
On incorporating(42), the problem (40) can be rewrittenas maximize minu IL

pIESm
subject (43)(43t{ESm, to
max VESV, DESd) SFl/2V4 -S(P) = 1, 1
+ IDl/2+2
<
4>o.
Assuming that the uncertaintysets Sd, S, and Sm are given by (2)-(4), (43) reduces to
maximize (FIo- y)T ,
(44)
to subject -9r1i 3) i (Ioa ,y - -8 I1/2 811 ,

max |1Fl/2Vll {VES}, v,
1T+ = 1,
>_o.
From part (ii) of Lemma 1, it follows that max IIFI/2Vll < V9
{VES, }
if and only if there exist I > 0 and s E R' such that,

r2 < 2
TS)
(45)
(1 - ?Oi)si, < ui < 1, Amax(K)F
i = 1 ...
m,
where K = POPT is the spectral decomposition of K = Fl/2G-'F/2, O = diag(0i), and On introducing the change of variables, 7 = vt and s = (1/v)s, (45) u = pTFl/2Vo0.
becomes
r2
(46)
u2 < (v Amax(K)7 < v.
r(v - TS),
7TOi)si,
i= 1
, m,
....
Therefore,the robust VaR portfolio selection problem (40) is equivalent to the SOCP (47) maximize subject to
(tio -
))T+, u = pTFl/2VT+,
_< (F(L) 8, ID1/'120
O_- )
15
E,<2pT s) (_ v+lr +
i
-(J _ ,lOi )i __Si)
...
(7
2u<(v
TS),
T
v-
(K) > 0, Amax 1T+ = 1,
+i=),1,..1m,
+>0o,
7 >0.
As in the case of the minimum variance problem, an alternativeSOCP formulationfor the robust VaR problem follows from part (i) of Lemma 1. 5. Multivariate regression and norm selection. In this section results from the statistical theory of multivariatelinear regression are used justify the uncertaintystructuresSd, S, and Sm proposed in ?2 and motivate naturalchoices for the matrix G defining the elliptic norm II-II, the bounds piyi, di, i = and n. 1,...., In ?2, the returnvector r is assumed to be given by the linear model, (48) r=
t
+ VTf + E,
where E is the residual return.In practice, the parameters(tp, V) of this linear model are estimated by linear regression. Given market data consisting of samples of asset return vectors and the correspondingfactor returns,the linear regression procedurecomputes the of (ix, V). In addition, the procedure also gives multileast squares estimates (ixo, Vo) dimensional confidence regions aroundthese estimates with the propertythat the true value of the parameterslie in these regions with a prescribed confidence level. The structureof the uncertaintysets introducedin ?2 is motivated by these confidence regions. The rest of this section describes the steps involved in parameterizingthe uncertainty structures,i.e., computing go, Vo, G, i, Ti, di, i = 1, ... , n. Suppose the market data consists of asset returns, {r': t = 1... , p}, for p periods and the correspondingfactor returns({f: t = 1 Then the linear model (48) implies that ...., p}. n = L+ i=1,...,n, t=1,...,p.
j=1
Vjifj+Et,
In linear regression analysis, typically, in addition to assuming that the vector of residual returnsE' in period t is composed of independent normals, it is assumed that the residual returnsof different marketperiods are independent.Thus, {ef:i = 1,..., n, t = 1..., p are all independent normal random variables and ef -i(0, of), for all t = 1, . . . , p; i.e., the variance of the residual returnof the ith asset is o-2. The independence assumption can be relaxed to ARMA models by replacing linear regression by Kalman filters (see Hansen and Sargent 2001). Let S = [r1, r2,... ,rP] E Rnxp be the matrix of asset returnsand B = [f, f2,... fP] E RmXPbe the matrix of factor returns.Collecting together terms correspondingto a particular asset i over all the periods t = 1,... , p, we get the following linear model for the returns {rit: t = 1, p}, .... Yi= Axi + Ei, where
.. Yi = [r ri2
rP]T ,
--" A = [1 B BT],
_)_
Xi
[Li
iVi
o V2i ...
,f Vmi]T ?
and Ei = [e, ... , eP']T the vector of residual returnscorrespondingto asset i. is
Vii
16
The least-squares estimate ii of the true parameterxi is given by the solution of the normal equations ATAii = ATy,; i.e., (49) ii = (ATA)-'ATyi,
if rank(A) = m + 1. Substitutingyi = Axi + Ei, we get xi - xi = (ATA)-'AT where Z = o.2(ATA)-'. Hence

(50)
--
i = N(O,q),
(i
- xi)T(ATA)(ii -
Xi)
2 " Xm+1 m l
is a X2 random variable with (m + 1) degrees of freedom. Since the true variance is o.2 unknown, (50) is not of much practical value. However, a standard result in regression theory states that if oa2in the quadraticform (50) is replaced by (m + 1)s?, where s2 is the unbiased estimate of of2given by (51) then the resulting randomvariable
(52) y =
2 Yi - Ai s2 = p-m-1 y-A
12
1
+ (m +1)s?
(j-
xi)T(ATA)(ii
xi)
is distributedaccordingto the F-distributionwith (m + 1) degrees of freedom in the numerator and (p - m - 1) degrees of freedom in the denominator (Anderson 1984, Greene 1990). Let 0 < o < 1, ~,r denote the cumulativedistributionfunction of the F-distributionwith J degrees of freedom in the numeratorand p - m - 1 degrees of freedom in the denominator and let cj,() be the w-critical value, i.e., the solution of the equation ?j (cj,(w)) = o.
Then the probability 9: _
Cm+,(w)
is o, or equivalently, (m + 1)cm+,l()s2)
_ = W.
(53) Define
P ((ji - xi)T(ATA)(ii - xi)
- xi) (m+ 1)cm+I(w)si1. (i, xi)(ATA)(ii < Then, (53) implies that Si(w) is a o-confidence set for the parametervector xi corresponding to asset i. Since the residual errors {ei: i = 1 ..., n} are assumed to be independent, it follows that
(54)
Si((o) = {xi:
(55)
S(W) = S (to) x S2((0) x
..
Sn(),
is a on-confidence set for (tp, V). Let Sm(o) denote the projectionof S(o) along the vector Ix; i.e.,
Sm(o)
(=l:
pR= I
< = 1,..., o + v, Ivil yi, i
n},
17
where
(56)
'0,=
fi,
V, =
(m+1)(ATA)-Im+1c(w)s,
i= 1...,n.
Then (55) implies that Sm(w) is an )on-confidence set for the mean vector FI. Note that the uncertaintystructure(4) for the mean Fpassumed in ?2 is identical to Sm(w). Let Q = [e2, e3 ,..., em+JlT E Rmx(m+') be projection matrix that projects xi along Vi. Define the projection S,(o) of S(w) along V as follows:
= S,(wo) {V:V= Vo+W, IIWl p ,i= 1,...,n g

where
Vo= V,], -[VI ."

(57)
1 G = (Q(ATA)-'QT)-1 = BBT- -(B1)(B1), P (m +

()CsI 1)cm+1) i = 1 ..., n.
Pi
Then S,(w) is an wo-confidence set for the factor loading matrix V. As in the case of Sm(to), the uncertaintystructure(3) for the factor covariance V assumed in ?2 has precisely the same structureas S,(w) defined above. The construction of the confidence regions can be done in the reverse direction as well; i.e., individual confidence regions can be suitably combined to yield joint confidence C Rmxmbe any J-confidence regions for p and V regions. Let Sm(c) C Rn and Sv,() respectively. Then
(58) P((fL, V) E Sm(ji) x S,(C)) = 1 -P((i,
V) V Sm(ji) x
Sm(Co))-
Sv(,()),
> 1 - P(I = 2 - 1;
P(V v S,(C)),
i.e., Cartesianproducts of individual confidence regions are joint confidence regions. This leads to an alternativemeans of constructingjoint confidence regions for marketparameters
L(, V).
Let Q E RJXm. Then (QXi- Qii) (Qxi - Qii)(Q(AA)-'QT)-I ~si is distributedaccording to the F-distribution with J degrees of freedom in the numerator and p - m - 1 degrees of freedom in the denominator,i.e., the probabilityY < cj,(O) is C5, or equivalently, (59) = (60) P ((Qxi - Qxi)'(Q(ATA)-'QT)-'(Qxi - Qii) < Jcj(C)s?) = 55.
1
the least squares estimate of the mean return of asset i and Set Q = eT. Then, Qxi = -'i, Qxi =/-ti, the true mean returnof asset i. Therefore, (60) implies that (61) P (i
-i
= (ATA) I'ci(6-)s2) ).
< Since the residual errors Ei are assumed to be independent,it follows that
Sm())
{"4: I
-, o v Ivil i<
, i = 1 ...,
n),
18 where
(62)
0,i =
y, = p, (ATA)-lc1
(1 )s2,
i= 1 ...
n,
is a ton-confidence set for the mean vector Lp. Next set Q = [e2, e3 ,... , em+,] E Rmx(m+l). Then, the projection Qxi = [Vii V2i ... Vim]T= Vi, the least squares estimate of the true factor loading Qxi = Vi, and (60) implies that S P ((Vi - Vi)T(Q(ATA)-'QT)-I(V -i Vi) mcm(C)s?) = < As in the case of the mean vector Ip, it follows that .
S,() = (V: V= Vo+ W, I ilg

Vo = V, (63) Pi is an = mcm()s, i= 1 ... n,
p i= 1, ... ,n},
(B1)(B1), - 1 P
G = (Q(ATA)-'QT)-1 = BBT
x S,( ) is a joint confidence region with (2Mn - 1) confidence. Sm(.) .n"-confidence
set for the factor loading matrix V. From (58), it follows that S(w) =
The two families of uncertaintysets defined by (56)-(57) and (62)-(63) are not entirely is an increasing function of J, it follows that, comparable.Since the critical value a fixed 0oe (0, 1), S(o) C S(w). cSjm(o) other hand, if (-0is chosen such that 2c"- 1 = On the for ton, the direction of inclusion depends on the dimension m. After determiningthe uncertaintysets for tp and V, all that remainsto be established is the upper bound di on the varianceof regression error oi2. From the analysis above, it follows that the naturalchoices for the bounds are given by the confidence interval aroundthe least squaresestimate s2 of the errorvariance oi2. Unfortunately,the regression procedureyields only a single unbiased sample s2 of the errorvariance. One possible solution is to construct bootstrapconfidence intervals(see Efron and Tibshirani 1993). Since the bootstrappingstep can often be computationallyexpensive and since the robust optimization problems only require an estimate of the worst case error variance, any reasonable estimate of the worstcase errorvariance could alternativelybe used as the bound di. Incorporatingthe results developed in this section, we have the following recipe for solving the robust portfolio selection problem: (1) Collect data on the returnsr of the assets and the returnsf of the factors. (2) Using (49) one asset at a time, evaluate the least-squaresestimates pIoand Vo of the mean Fp and the factor loading matrix V, respectively. (3) Choose a confidence thresholdw. (a) Constructa bootstrapo-confidence interval around -i2. Alternativelyuse any estimate of the worst case errorvariance. (b) Define Sm and S, using (56) and (57), respectively (or (62) and (63), respectively). (4) Solve the robust problem of interest. A word of caution about the choice of &j.If to is chosen very high, the uncertaintysets will be very large; i.e., implicitly one is demanding robustness with respect to a very large set of parametervalues. The resulting portfolio will be very conservative, and therefore its performancefor a particularset of parameters(I0o,Vo) will be significantly worse than the portfolio designed for that set of parameters.On the other hand, if o is chosen too low, the portfolio choice will not be robust enough. The typical choices of o lie in the range 0.95-0.99. See ?7 for more details on the implications of the choice of to and Appendix A
19
for a discussion of the probabilistic guarantees on the performanceof the optimal robust portfolio. Although we propose a strictly data-drivenapproachin this section, this analysis extends to Bayesian estimation methods such as those in Black and Litterman(1990), as well as empirical Bayes estimation. In the Bayesian setting, the confidence regions are given by the posterior distribution. Recall that the maximum likelihood estimate F,,mof covariance matrix F of the factors is given by,
1 [BBT (B1)(B1) p p-1 i.e., G = (p - 1)Fmi. Suppose the true covariancematrixF is approximated the maximum by likelihood estimate Fmi. Then F = KG, where K = 1/(p - 1). Recall that the worst case variance
(64)
Fm =-
(65) if and only if
max {
Ve S,
VrFV+
v2
<
{y:I1yI1g,<r)
max IIVo +yllIf
where r = p'T. Since F = KG, the above norm constraintis equivalent to
(66)
{u:Ilul_<_r-.,-}
max IF'/2V0++ul <v,
where I is now the usual Euclidean norm. It is easy to see that the maximum in (66) is 1-IJ attainedat U= constraint(65) is equivalent to the second-ordercone constraint
andthe corresponding maximum valueis I1F'/2Vo0ll variance + rJ-K. Thus,the worst-case
IIF'/2VoF1/2Vo0, ll
(67)
ll V FII'/2Vo - _ P 4
From the fact that v2 < 7 is equivalent to a second-order cone constraint, it follows that in the special case F = KG, the robust minimum variance reduces to the following simple SOCP minimize subject to
7 + 8,
(FLo y)T + > a,
1T+ = 1,
JjF'/2VO+
NKpT+,
(68)
2DI[/2+]
1 +,
1-5
11[- 12v, T
+7
,>__o. The robust maximum returnproblem and the robust maximum Sharpe ratio, as well as the robust VaR problem, can all be simplified in a similar manner. In practice, however, the covariance matrix F is assumed to be stable and is, typically, estimated from a much larger data set and by taking several extraneous macroeconomic indicators into account (see Ledoit 1996). If this is the case, then the above simplification cannot be used, and one would have to revert back to the formulationin (31).
20
6. Robust portfolio allocation with uncertain covariance matrices. The market model introducedin ?2 assumes that the covariance matrix F of the factors is completely known and stable. While this is a good first approximation,a more complete marketmodel is one that allows some uncertaintyin the covariance matrix and optimizes over it. For any uncertaintystructurefor covariance matrices to be useful, it must be flexible enough while at the same time restricted enough to admit fast to model a variety of perturbations and efficient optimization. Our goal in this section is to develop such an parameterization uncertainty structurefor covariance matrices. We assume the market structureis the one introducedin ?2, except that the covariance matrix is no longer fixed. 6.1. Uncertainty structure for covariance inverse. structurefor the factor covariance matrix F: Considerthe following uncertainty
(69)
Sf,
IF:F-' = Fo-'+
A_
0, A
AT,
Fo/2A1/2
= maxi lAi(A)I, where {Ai(A)} are in where F0 >- 0. The norm |IAlJ (69) is given by IAlthe eigenvalues of A. The uncertainty set Sf-, restricts the perturbationsA of the covariance matrix to be symmetric, bounded in norm relative to a nominal covariance matrix F0 and subject to F-' = Fo' + A - 0. Clearly, this uncertaintystructureis not the most general-one could, for example, allow for each element of the covariance matrix to lie in some uncertainty interval subject to the constraint that the matrix is positive semidefinite. However, the robustoptimizationproblem correspondingto this uncertaintystructureis not very tractable 2000). On the other hand, if the covariance uncertaintyis given (Halld6rssonand TiitOincti all of the robust portfolio allocation problems introduced in ?2 can be reformuby (69), lated as SOCPs. Moreover, as in the case of the uncertainty structuresfor Fp and V, the uncertaintystructureSf-, for F corresponds to the confidence region associated with the statistical procedure used to estimate F. In particular,we show that maximum likelihood estimation of the covariance matrix F provides a confidence region of the form Sf-, and yields a value of r that reflects any desired confidence level. All the robust portfolio selection problems introduced in ?2 can be formulated for this marketas well. For example, the robust Sharpe ratio problem in this marketmodel is given by
maximize
{#:1T4=1}
S,, {ILESmv DESd, FESf-I-
min
rf T 4/T(VTFV+D)4
Lemma 3, below, establishes that the worst-case variance for a fixed portfolio is given by a collection of linear and restrictedhyperbolic constraints-the critical step in reformulating robust portfolio selection problems as SOCPs.
LEMMA 3. Fix a portfolio 4 and let Sf- be given by (69). Then the following results hold. (i) If the bound q> 1, the worst-case variance is unbounded;i.e.,
{VES,,FESf-I }
maxs TVTFVsf
= oo.
(ii) If 71< 1, the worst-case variance max{vEs,,res_,)f TVT FV+ < v if and only if (pT+; (1 - i))v; +) EC(Vo, F0, G), where (, E (., .) is defined in Definition 1. PROOF. Define A = F/2AF1/2. Then
s,, = IF:F'
=F,1/2(I
)F
1/2
6 0, A =
T,
l|1 i
21
Fix V E S, and define x = V'4.
Then
FESf
{x (I sup Fx}= sup(Fo/2x)

I
1111 I+ (F/2x): 1= A , 14 )-'
I >_01.
First consider the case r>
I+
>= (1 - )I o. Thus,
1. Choose A = -l31 where 0 < </3 1. Then 1Ail= / : s
and
1 sup {xTFx} > sup (xTFox) = oo. {: 0:< l 1 - /3 FEsf-I Next, suppose 77< 1. Since q < 1 implies that I + Sf-, = Therefore sup {x'Fx} = sup
FESf-I
0 for all , =
IAl
F: F-1 =
Fo/2(I
)Fo'/2,
Ill
<
, 4ii
(Fo/2x)(I
+)-
(Fl/2X):
A =
i=1 1 + S(qrT
(F1/2x))2 (1i()
where {qi, i = 1,... , m} is the set of eigenvectors of A. Since IAi(A)I

FESfI
r , it follows that
sup {xTFx}1 -
il =1
q(F;
TXFoX.
Moreover, the supremumis achieved by A = - rl. Thus, - 1 max max T(VTFV)+ = 14TVTFoV4. {VESV,, 7 veS, FESf_ From part (i) of Lemma 1, we have that 1 -- max "TVTFoV4< v
if and only if (pT';
From Lemma 3 it follows that if r < 1, then all robust portfolio problems introduced in ?2 can be reformulatedas SOCPs. The next step is to justify the uncertaintystructureSf-, and show how it can be parameterized from marketdata. In ?5 we show that the naturaluncertaintystructuresfor (ix, V) and their parameterizations implied by the confidence regions associate with the statisare tical procedures used to compute the point estimates. Here, we show that the uncertainty structureSf-I and its parameterization,i.e., the choice of the nominal matrix Fo and the bound qr, arise naturally from maximum-likelihood estimation typically used to compute the point estimate of F. Suppose the covariance matrix of the factor returns is estimated from the data B = over p market periods. Then the maximum likelihood estimate Fml of the [f1, f2 ...fP] factor covariance matrix F is given by (64). Suppose one is in a Bayesian setup and assumes a noninformativeconjugate prior distribution for the true covariance F. Then the posterior distributionfor F conditioned on the data B is given by (70) FIB , W-11((p
-
(1 - r)v; 4)) e
(Vo, Fo, G).
1)Fml),
22
where denotes an inverse Wishart distribution with q degrees of freedom and Wq-'(A) parameterA (Anderson 1984, Schafer 1997); i.e., the density f(FIB) is given by
(71)
f(FIB)
J
-
0,
F-1I clFmll(p-m-2)/2 (p-1)/2e-[((P-1)/2)Tr(F-'Fmi)]
F >O,
otherwise,
where JAI= det(A) and c is a normalizationconstant. This density can be rewritten as follows:
(72) f(FIB)=
(m-1)/2 11/2F
l/2 (p-1)/2 -[((p-1)/2)Tr(F/2F-'1
clFmj S0,
12]
>
0,
otherwise.
From (72) it follows that the naturalchoice for the nominal covariance matrix FO in the definition of Sf-1 is FO= FmI. Let X E Rmbe the vector of eigenvalues of F/2 F- F/2. Then (72) implies that the density of X is given by
> 0
(73)
f(XB)=
p-1)/2e-((p-1)/2)?i (73) f(B) i=1otherwise,
0,
otherwise,
where cA is a normalization constant, i.e., the eigenvalues Ai, i = 1,... , m, are IID F((p + 1)/2, (p - 1)/2) distributedrandom variables. Let A = F-' - FO1 be the deviation of the F-' from the nominal inverse Fo' and let 1 < 1 + ; i.e., AF1/2 = F F-'F1/2 - I. Then IAII| < if and only if I - r1< Ai _
-
(74)
P(IA1 < ) = (P(1-
<
1+))m
(1+ ) - r(1-
where ?3~, F((p + 1)/2, (p - 1)/2) and is the correspondingCDF. rP If 7 < 1, F-1' = F' +A >- 0 for all AI|l j,, and therefore (74) implies that _ (75) P(F e Sf-,) = (rp (1 + q) r-,(1 q))m. Thus, in order to satisfy a desired confidence level wo, the parameterr7must be set equal to the unique solution of (76) = Yr, (1 + ) 5r, (1 q) 0.
Since the problem is unboundedfor states that the confidence level q'> 1, (76) implicitly Nw that can be supportedby p return samples is at most Womax(p)= (F-r(2))'. Figure 1 plots ofmax as a function of the data length p for m = 40 (the plot begins at p = m + 1 since at least m + 1 observations are needed to estimate the covariance matrix). From the plot, < one can observe that for p > 50, omax(p) > 0.995. Thus, the restriction wo"m max(p) is not likely to be restrictive. This methodology can be modified to accommodateprior informationabout the structure of F. Suppose the prior distributionon F is the informativeconjugate prior Wk1 - 1)F), ((k 1. Then, the posteriordistributionFIB -W-W1((p + (k - 1)F). In this case, the k> 1)Fmi nominal covariance matrix FO= (1/(k + p))((p - 1)Fmr+ (k - 1)F), and r is given by
(1 + rk+p
where
+r,
denotes the CDF of a F((k + p + 2)/2, (k + p)/2) random variable.
)-ruk+,
(1-7))=o,
ROBUST PORTFOLIO SELECTION PROBLEMS
23
0.995 .
0.99
0.985
....
40
60
80
100
120
140
160
180
200
p
FIGURE 1. Wmax VS p (m = 40).
From the results in this section, it follows that, as far as the portfolio selection problems are concerned, all that the uncertaintyin the covariance matrix does is "shrink"the MLE Fo' to (1 -, )Fo', where the shrinkagefactor (1 - rq)is a function of the desired confidence level w. Thus, this procedurehas the flavor of robust statistics (see Huber 1981). 6.2. Uncertainty structure for the covariance. for the factor covariance matrix is given by Another possible uncertaintystructure
(77)
Sf= F: F =
O,, + A >- A = Fo
N, IN-2N-1/2
where {i(A)} are the eigenvalues of A. Ai2(A), For (77) to be a viable uncertaintystructure,the correspondingrobustproblem should be efficiently solvable, and the structureought to be easily parameterizablefrom raw market data. The first requirementis established by the following lemma. Fix a portfolio E R" and let Sf be given by Equation (77). Then e v if and only if (pTr; v; +) E (Vo, (F0 + rN), G), where max{VEs,FESI} T(VTFV)4) < is defined in Definition 1. Ye(,., .)
LEMMA 4.
= = in or whereF0 >- 0. The normlIAII (77) is either IAllI maxi IAi(A)I, IIAil V/i
PROOF. Define A = N-1/2AN-1/2. Then Sf =
F: F=
F0 +N1/2N"/2
>
0,,
1=
T,
<
Fix V E S, and define x = VT . Then
max{x'Fx} = max x'Fox (78)

<
+ (Nl/2x)TA(N'/2x):
X A(N1/2X):
A=
AT,
I 11
Ail
,F0
+N+2AN1/21/2
x'F0ox + (N1/2x) XTFox
A =
~T,
(79)
<
+ ?(N1/2X)T(N1/2X),
where (79) follows from the propertiesof the matrix norm.
24 = or Since 1A11llmax{IAi(A)hI}
A2(A) i
and N >-0, the bound (79) is achieved by
(N/2X) (N1/2X)T
IIN1/2X112
unless x = 0. Thus, the right-handside of (78) is given by xr(Fo + rN)x and is achieved
by A* =
inequality (78) is, in fact, an equality; i.e., max {XTFx}= XT(Fo+ SN)x. FESf Thus, max 4T(VTFV)4 = max bTVT (Fo + SN)V4. VES, FESf} {vES,, From Lemma 1, we have that max4VT'(Fo + ?N)V
VESV
[ - (Nxx'N)/(x'Nx),
unless x = 0. Since Fo + N1/2I*N1/2
0, it follows that the
v,
_<
if and only if (pT'; v; 4) e (V0, (FO+ ?N), G) where (.-, , .) is defined in Definition 1. LO and do not require ( < 1. here we allow N F0o Notice that, unlike in the case Sf-1, The next step is to devise a method for parameterizingthis uncertainty structure. The will once again be implied by the statistical procedureused to estimate F. parameterization Let F be the true (unknown) covariance matrix of the factor returns and let Fm,,be the MLE of F computed from the returndata B. Then it is well known (Anderson 1984) that if p > m (i.e., the number of observationsis larger than the dimension of the matrix) then Fmi >-0 with probability 1 and
(80)
Fmi
((p Wp_1
1)F),
where Wq(A) denotes a Wishart distribution with q degrees of freedom centered at the matrixA; i.e., the density f(Fmi IF) is given by
(81)
f (FmF) = c(F)-'
I0,
Fml>- 0,
(P-')/2IFmlI(p-m-2)/2e-[((p-1)/2)Tr(F-'Fmi)' otherwise.
An argumentvery similar to the one in the previous section establishes that Fm,,belongs to the set
(82)
S= {F: F=F+A >-0,A= A, IIF-'/2AF-'/2II

- B) = o,
if with probabilitywom the bound 3 is set equal to the solution of (83) Sr, (1 +,B) r, (1
where I-, denotes the CDF of a F((p + 1)/2, (p - 1)/2) random variable. Note that this equation is identical to (76); i.e., the solution 3 of (83) is equal to the solution r of (76). Suppose the bound 3 < 1, or equivalently om < Wmax(P)= (2))m. Then we have that
(5rFp
(84)
IIF,"2(FFF/2
II
F/2/2F1/2(F
Fml)F-1/2F/22
Fml)F-1/2
II
II 1 F< '2F"2F F-1/2(F -1-/3'
I lF1/-/2' II,
25
where (84) follows from the fact that Fm,, (85)

5 =
Si
defined in (82). From (84) we have that

T' FJI/2 -AF I2 -11 I,
F:
>+ FmliA
9A
contains an to-confidence set for the (true) covariance matrix F; i.e., a natural parameterization of the uncertainty structurein (77) is Fo = N = Fm,, and r = /1(1- P). Thus, the uncertainty structure(77) is completely parameterizedby considering the confidence regions aroundthe MLE. Note that for N = FOand "= /81(1 -1), Lemma 4 implies that max 4T(VTFV)4 (VESv, FESf} if and only if
(p'4;
<v
v; +) e
(Vo, (1 + p/(1 - P))F0, G) = W(Vo, 1/(1 - P)Fo, G).
As noted above, the solution p of (83) is equal to the solution -a of (76), and from Definition 1 it follows that
(pT4);
v; 4) E
(V0, 1/(1 - P)F0, G) * (pT+4; (1 - 'q)v; 4) e Ye(Vo, F0, G).
Thus, max(VES,, FESf4)T(VTFV)+ < v if and only if (pT+; (1 - -)v; 4) E (V0, F0, G). Since this is precisely the condition in part (ii) of Lemma 3, it follows that, althoughthe two uncertainty sets (69) and (77) are not the same, they imply the same worst-case variance constraint.However, unlike the parameterization Sf- , the parameterization Sf cannot of of be biased to reflect prior knowledge about F. 7. Computational results. In this section we reportthe results of our preliminarycomputationalexperimentswith the robustportfolio selection frameworkproposed in this paper. The objective of these computationalexperiments was to contrast the performance of the classical portfolio selection strategies with that of the robust portfolio selection strategies. We conducted two types of computationaltests. The first set of tests comparedperformance on simulated data, and the second set compared sample path performance on real market data. In these experiments our intent was to focus on the benefit accrued from robustness; therefore, we wanted to avoid any user-defined variables, such as the minimum return a in the robust minimum variance problem or the probabilitythreshold 1 in the robust VaR problem. Thus, in our tests both the classical and robustportfolios were selected by solving the correspondingmaximum Sharperatio problem. We expect the qualitativeaspects of the computationalresults will carryover to the other portfolio selection problems. All the computations were performedusing SeDuMi V1.03 (Stiirm 1999) within Matlab6.1R12 on a Dell Precision 340 workstationrunning RedHat Linux 7.1. The details of the experimental procedureand the results are given below. 7.1. Computational results for simulated data. For our computationaltests on simulated data, we fixed the number of assets n = 500 and the number of factors m = 40. A symmetric positive definite factor covariance matrix F was randomlygenerated, except that we ensured that the condition numberof F, i.e., Amax(F)/Amin(F), was at most 20 by adding a suitable multiple of the identity. This factor covariance matrix was assumed to be known and fixed. The nominal factor loading matrix V was also randomly generated. The covariance matrix D of the residual returnsE was assumed to be certain (i.e., D = D = D) and set to D = 0.1 diag(VTFV); i.e., it was assumed that the linear model explains 90% of the asset variance.
26
The risk-free rate rf was set to 3, and the nominal asset returns were chosen indepenwe dently accordingto a uniformdistributionon [rf- 2, rf + 2]. Next, 1.,i generateda sequence of asset and factor returnvectors according to the marketmodel (1) in ?2 for an investment period of length p = 90 and used equations (62) and (63) in ?5 to set the parametersV0, G, FLo, p and Py.(Note that we did not estimate D from the data.) Next, the robust and the and +m were computedby solving the robust maximum Sharpe ratio classical portfolios +r problem (8) and its classical counterpart,respectively. Although the precise numbers used we in these simulationexperimentswere arbitrary, expect that the qualitativeaspects of the results are not dependent on the precise values. In the first set of simulationexperimentswe comparedthe performanceof the robust and classical portfolios as the confidence threshold w (see ?5 for details) was increased from 0.01 to 0.95. The experimentalresults for three independentruns are shown in Figures 2-4. In each of the three figures, the top plot is the ratio of the mean Sharpe ratio of the robust portfolio to that of the classical portfolio. The mean Sharpe ratio of any portfolio 4 is given by (FLo rfl),4
V;T
(VTFVo0)4
The classical portfolio 4m maximizes the mean Sharpe ratio. The bottom plot in Figures 2-4 is the ratio of the worst-case Sharperatio of the robust portfolio to that of the classical portfolio. The worst-case Sharpe ratio of a portfolio 4 is given by
minSDESd DESd)
pem', where Sd, S, and Sm are given by (2)-(4). The robust portfolio 4, maximizes the worstcase Sharpe ratio. The performanceon the three runs is almost identical--the ratio of the mean Sharpe ratios drops from approximately1 to approximately0.8 as o0 increases from 0.01 to 0.95 while the ratio of the worst-case Sharperatios increases from approximately 1 to approximately2. Thus, at a modest 20% reduction in the mean performance,the robust
Robust/Classical Mean Sharpe Ratios
0.95
{VES,,
/T(VFV + D)
(mI rf l)T4 + D)
0.85......
0.85
,.. . . . . . . . . ... . . . . . . . . .. . . . .
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Robust/Classical Worst Case Sharpe Ratios

1.8
1.2
0.1
0.2
0.3
0.4
0.5 w
0.6
0.7
0.8
0.9
of 2. as FIGURE Performance a function w (Run1).
27

0.98 . 0.96 0.94 0.92 0.9
0.88
0.86
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9

1.8
1.6
1.4
.
0
0.1 0.2 0.3 0.4 0.5
-....
1.2
0.6
0.7
0.8
0.9
w FIGURE Performanceas a function of w (Run 2). 3.
frameworkdelivers an impressive 200% increase in the worst-case performance.Notice that the drop in mean performancein run 3 (see Figure 4) is close to 25%, but the corresponding improvementin the worst-case performanceis close to 260%. The second set of simulation experiments compared the performanceof robust and classical portfolios as a function of the upper bound on D. For this set of experiments, we set = &j 0.95 and D = o02 diag(VJFVo), where 0-2 increased from 0.01 to 1. Since the performance of both the robust and classical portfolio was very sensitive to the sample pathparticularlyfor large values of 0-2-the results shown in Figures 5-7 were averaged over
0.95 . 0.9 . 0.85 0.8 0.75 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

2.6 2.4 . 2.2 2 1.8 1.6
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
w 4. FIGURE Performanceas a function of w (Run 3).
28

1.2 1.15 . 1.1 1.05
0.95
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4

7 6 5 4 3 2 0 0.05 0.1 0.15 0.2 a2 0.25 0.3 0.35 0.4
as of FIGURE Performance a function a (Run1). 5.
10 runs for every value of o2. As before, the top plot is the ratio of the mean performances of the robust and classical portfolios and the bottom plot is the ratio of the worst-case performances. (The three plots truncate at different values of o2 because we only compared performancesfor values of o2 for which the worst-case Sharpe ratio of the classical portfolio is nonnegative.)Again, the essential features of the performancewere independentof the particularrun-the mean performanceof the robust portfolio did not degrade signifi-

0.98 . 0.96 ". 0.942 0.92
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4

2.5
1.5
0.05
0.1
0.15
0.2 -2
0.25
0.3
0.35
0.4
FIGURE Performanceas a function of oa(Run 2). 6.
29
Robust/Classical
1.01
Mean Sharpe Ratios
0.99 0.98 0.97 0.96
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Robust/Classical
5 4 3 2
of Worst Case Sharpe Ratios
0.05
0.1
0.15
0.2 O-2
0.25
0.3
0.35
0.4
7. FIGURE Performanceas a function of tr (Run 3).
cantly with the increase in noise variance, if it did so at all, and the worst-case performance of the robust portfolio was significantly superiorto that of the classical portfolio as the data became noisy. This is not unexpected since the robust portfolios were designed to combat noisy data. Figure 8 compares the CPU time needed to solve the robust and classical maximum Sharperatio problem as a function of the numberof assets. For this comparisonthe number of factors m = [0. n], the risk-freerate rf was set to 3, 1pand F were generatedas described above, the noise covariance D was set to 0.1 diag(VTFV), the confidence level w was
600TT1
r Robust -A-Classical I
500
-.0o
300..................................................... S300
..
.200
..
100
200
400
600
800
1000
1200
1400
1600
1800
2000
Number of Assets n (m = [0.1n]) FIGURE Complexity of robust and classical strategies. 8.
30
set to 0.95, and the parameters V0, G, p, y) were computed from randomly generated (o0, factor and asset return vectors for an investment period p = 2m. These experiments were conducted using SeDuMi V1.03 (Stiirm 1999) within Matlab6.1R12 on a Dell Precision 340 machine. The runningtimes were averaged over 100 randomly generated instances for each problem size n. It is clear from Figure 8 that the computation time for the classical and robust strategiesis almost identical and grows quadraticallywith the problem size. We should note, however, that algorithmsbased on the active set method can be used to solve the classical problem, and these may be more efficient in some cases. 7.2. Computational results for real market data. In this section we compare the performanceof our robust approachwith the classical approach on real market data. The universe of assets that were chosen for investment were those currently ranked at the top of each of 10 industrycategories by Dow Jones in August 2000. In total there were n = 43 assets in this set (see Table 1). The base set of factors were five major marketindices (see Table 2), to which we added the eigenvectors corresponding to the 5 largest eigenvalues of the covariance matrix of the asset returns; i.e., the total number of factors used was m = 10. This choice of factors was made to ensure that the linear model (1) would have i.e., small values for the residualreturnvariances. (Selecting appropriate good predictability, factors to explain the covariancestructureof the returnsis a sophisticated industry,and our choice of factors is by no means claimed to be the most appropriate.)The data sequence
TABLE Assets 1. Aerospace Industry AIR BA LMT UTX AAR corporation Boeing Corp. Lockheed Martin United Technologies Semiconductor AMD INTC HIT TXN Applied Materials Intel Corp. Hitachi Texas Instruments ComputerHardware DELL PALM HWP IBM SUNW Dell ComputerCorp. Palm Inc. Hewlett Packard IBM Corp. Sun Microsystems Biotech and Pharmaceutical BMV CRA CHIR LLV MRK Bristol-Myers-Squibb Applera Corp.-Celera Chiron Corp. Eli Lilly and Co. Merck and Co. Chemicals AVY DD DOW EMN Avery Denison Corp. Du Pont Dow Chemical Eastman Chemical Co. FMC GE HON IR ENE DUK EXC PNW AKAM AOL CSCO NT PSIX ARBA CMRC MSFT ORCL T LU NOK MOT Telecommunication AT&T Lucent Technologies Nokia Motorola Computer Software Ariba Commerce One Inc. Microsoft Oracle Internetand Online Akamai AOL Corp. Cisco Systems NorthernTelecom PsiNet Inc. Utilities Enron Corporation Duke Energy Company Exelon Corp. Pinnacle West
IndustrialGoods FMC Corp. General Electric Honeywell Ingersoll Rand
31
2. TABLE DJA NDX SPC RUT TYX
Base factors
Dow Jones Composite Average Nasdaq 100 Standardand Poor's 500 Index (S&P500) Russell 2000 30-year bond
consisted of daily asset returns from January2, 1997 through December 29, 2000. Given that assets were selected in August 2000, the data sequence suffers from the survivorship bias; i.e., we knew a priori that the companies we were considering in our universe were the major stocks in their industrycategory in August 2000. It is expected that this bias would affect both strategies in a similar manner;therefore, relative results are still meaningful. A complete description of the experimental procedure is as follows. The entire data sequence was divided into investment periods of length p = 90 days. For each investment period t, we first estimated the covariance matrix XR of the asset returns based on the marketdata of the previous p tradingdays and extractedthe eigenvectors correspondingto the 5 largest eigenvalues (if an asset did not exist during the entire period it was removed from consideration). These eigenvectors together with the base market indices defined the factors for a particularperiod, and their returnswere used to estimate the factor covariance matrix F. Next, the equations (62) and (63) in ?5 were used to set Vo, FLO, p and y. G, The bound di on the variance of the residual return was set to di = syi, where s2 is given by (51), and the risk-free rate rf was set to zero. Once all the parameterswere set, the robustportfolio t' (resp. classical portfolio 4ot) was computed by solving the robust (resp. classical) maximum Sharpe ratio problem. The portfolio r' and +t were held constant for the period t and then rebalancedto the portfolio +(t+') and +(t+') in period t + 1. Let wt (resp. wt) denote the wealth at the end of period t of an investor who has an initial wealth w0 and employs the robust (resp. classical) strategy.Then
) r(t+ =
tp<k<_(t+1!)p
+t
(1+rk)T Wr,
(86)
tp<k<(t+l)p
Since both the robust and classical strategies require a block of data of length p = 90 to estimate the parameters,the first investmentperiod labeled t = 1 startsfrom the time instant p + 1. The time period January2, 1997-December 29, 2000 contains 11 periods of length p = 90; i.e., in all there are 10 investmentperiods. Figure 9 is a plot of the relative performanceof the robust strategy with respect to the classical strategy;i.e., it plots 100(wt/wt - 1) as a function of t, for a confidence threshold of to = 0.95. At the end of 11 periods the wealth generatedby the robustinvestmentstrategy is 40% greaterthan that generatedby the classical investment strategy.Notice that at t = 7 there is a precipitous drop in the relative performanceof the robust strategy.This drop is probably because the model estimate in period t = 6 is remarkablyclose to the realization in period t = 7; i.e., the robust strategy is being unnecessarily conservative. Therefore, one would expect that this problem would be somewhat mitigated if the threshold w were reduced. Figure 10 plots the relative performance of the robust and classical investment strategies for a confidence thresholdof to = 0.7. It is obvious that the drop in period 7 is reduced;however, the final wealth generatedby the robust strategy is, in fact, 9% less than that generated by the classical portfolio! Thus, it is not guaranteedthat the robust strategy will always outperformthe classical portfolio. Figure 11 plots the time evolution of the relative performancefor various values of to. As one would expect for small values of to,
32
80
Relative Performance of Robust vs Classical Portfolios
Ce
50 !
5e4 0 .......... .......... . ............ 30

11. 3 0 . . . . . . . . . .... .. .. .. . . . . . . . .
............
.................
"
.....
.......
. . . . . . . . . ... . . . . . . . . ... . . . . . . . ... . . . . . . . . . . . . . . . . . ... . . . . . . .
1
0 0
1.
.....
.................
...................
........................
10
Period = FIGURE Evolution of relative wealth for &w 0.95. 9.
the performanceof the robustand classical strategiesis quite close. For intermediatevalues, the dip at t = 7 is reduced, but unimpressive performance over the other periods drags the overall relative performanceof the robust strategy down. The performanceimproves as wo 1 (for w = 0.99 the robust strategygenerates a final wealth that is 50% larger than that generatedby the classical strategy).Figure 12 plots the final wealth ratio as a function of W (results were obtainedfor to = 0.1, 0.2, ... , 0.9, 0.95, 0.99). It is clear that the performance is not monotonic in &o. An importantaspect of any investment strategy is the cost of implementing it. Since we are interested in comparing the costs of implementing the robust strategy with that of 40
Relative Performance of Robust vs Classical Portfolios
. 0 . ........ ......... ................... .......................... . . . . .

.. -20
. 0
i....
CP
-20Pr
Period
10. FIGURE Evolution of relative wealth for w = 0.7.
33
of Relative Performance Robust vs ClassicalPortfolios

80
60 . . . . . . . . :.. . . . . . . . :.. . .. . . . . .. . . . .. . . . . . . . . . .
...
w=0.95. . . ... . . . . . . . . . . . . . . . . . . . . . .:.. . . . .
6= 40Ft
w = 0.8
4-D2
0. . .
S40
...... ..
20 . .
...
......
. .. . .. . . i. . .
a
c.
.w= 0.2 w =0.4
. . .. 0.6 . . -0SI. . . . . . . . . . . . . .. . . . . . . . . .. . . . ... .. .. .... . w. ........i........ -2

-40-
10
Period
FIGURE Evolution of relative wealth as a function of o. 11.
implementingthe classical strategywe will quantify the transactioncosts by I - t('-'1) In Figure 13 we plot the ratio of the costs, i.e. I1+r- rt-J/J for a confidence threshold of w = 0.95. The average cost is 0.9623; i.e., the transaction costs incurredby the robuststrategywere approximately4% less that that incurredby the classical strategy.Figure 14 plots the same quantity for w = 0.7 and now the average cost is 1.0057; i.e., as the robust strategy becomes less conservative it pays more in transaction costs. Figure 15 shows the average cost as a function of w. The average cost remains almost constant at a value slightly greaterthan 1 until w > 0.9, and then it decreases monotonically.
of Relative Performance Robust vs ClassicalPortfolios
60
50
. .........
.......... ..
2040
30 ......... . : o -. .......
20
S 20
ced
...4. .
.....
......:. 10 . .........
...... ....
......
..........
-20 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
w FIGURE Final relative wealth as a function of w. 12.
34
1.4
Relative Cost of Robust vs Classical Portfolios
1.2 ................
......
Average:Cost= 0.9623
-J-1-
0.6
0.4 2 3 4 5 6 7 8 9 10
Period 13. FIGURE Relative cost per period for w = 0.95.
7.3. Summary of computation results. The summary of our simulation experiments and the experimentswith real marketdata is as follows: (a) The mean performanceof the robust portfolios does not significantly degrade as the confidence level w is increased. Even at o = 0.95, the relative loss of the robust portfolios is only about 20%. On the other hand, the worst-case performanceof the robust portfolios is about 200% better. (b) Robust portfolios are able to withstand noisy data considerably better than classical portfolios.
Relative Cost of Robust vs Classical Portfolios
1.25
1.15
1.12
............
Average Cost = 1.0057 .. ............
S1.050
0.9 5
0 .........
0 .85 -...
.... ......... ......... .

............. ..............
..................
...........
0.8 0.75 I 5 . 6 7 8 9 10
Period 14. FIGURE Relative cost per period for w = 0.7.
35
Average Relative Cost of Robust vs Classical Portfolios

1.02
1.01 .
0.9. 0.99
m0.98. . ... . .... .. . ... ..... . . . .. . . ..~ .. .... ... .... ... .. ... ............. .... ... ... ... .. . .. . . .. .. ..
0.93
....
.....
0.92 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE Average cost vs w. 15.
(c) Summarizing the performanceof robust strategies on the real market data sequence is not easy. Based on the simulation data one would expect a monotonic improvement in performance as the threshold w is increased. However, the experimental results do not uphold this hypothesis. For the particulardata sequence used in our experiments,the robust strategywas clearly superior;i.e., it generateda largerwealth at a smaller cost when W was sufficiently large; whereas for small o there was no discernible trend. These computationalexperiments, particularlythose with the real data sequence, are by no means comprehensive.For one, the problem size-(m = 40, n = 500) for the simulation experiments and (m = 10, n = 43) for marketdata experiments-was small. For the robust strategy to be acceptable one would have to ensure that the complexity of the robust optimization problems is not significantly higher than that of the classical problems. A simple comparison of run times suggests that this is indeed the case. The experimentalresults on the real data suggests (see Figure 11) that if one wants the robust strategy to consistently outperformthe classical strategy one would want to choose wc- 1. However, such a choice of w makes the robust strategy extremely conservative, which would hamper its performanceif the noise in the model was low. Since the noise is not known a priori, the correct choice of w remains a vexing problem. Our preliminary experiments suggest that the factor model we used was probably noisy. More extensive experiments have to be conducted before one can assert that this is almost always the case (e.g., using "better"factors may significantly change the results), and therefore, setting w 1 is a good choice. We would expect that in practice one would have to adjust W dynamically by comparingthe robust strategy with the classical benchmark.Also, we have not tested robust portfolios based on our model in ?6 which incorporatesuncertaintyin the factor covariance matrix. These and other experimental studies are planned. Appendix: Probabilistic interpretation. In this section, we interpretthe choice of the uncertaintysets S,, S, and Sf in terms of the implied probabilistic guaranteeson the riskreturnperformance. First consider the case where the factor covariance F is fixed. Define Sm(co) and S(Wo) for a confidence level cv using (56) and (57). Let 4* be the optimal solution of the robust maximum Sharpe ratio problem (37) and let s*(o) be the correspondingSharpe ratio.
36
Since T'+* > 1 for all IF E Sm(W), it follows that 1TT4* 1 for (ti, V) in the joint > uncertaintyset S(o) defined in (55). Similarly, we have that
(4*)TVTFVI*
s*(WV) -:)E
VV e S,(o)
V(,V)ES(().
_r(5,)TVTFV*<
Therefore, we have that the Sharpe ratio of the portfolio 4*:
I 1
-(*)TVTFV+*
> *(),
( V)
);
with confidence &n. Similar probabilistic i.e., the Sharpe ratio is at least as large as s*(to) guaranteescan be providedfor the performanceof the optimal solutions of robust minimum variance, maximum return and VaR problems. Computing probabilistic guarantees on the performanceof the robust portfolio when the uncertaintysets are defined by (62)-(63) is left to the reader. Thus, in contrast to the classical Markowitz portfolio selection, the robust formulation allows one to impose confidencethresholdson the value of the optimizationproblem. Setting a low value for t results in a high value for the correspondingrobust Sharpe ratio s*(w), but the confidence that the correspondingoptimal portfolio +* would indeed achieve the Sharperatio is low. On the other hand, a high value for t would imply a lower Sharpe ratio s*(w), but it would ensure that the correspondingoptimal portfolio will achieve the Sharpe ratio with a higher confidence. Thus, the parameterw can be viewed as a surrogatefor risk aversion. Next, consider the case where the factor covariance matrix F is uncertain. Suppose one assumes the following Bayesian setup. The marketparameters(IL,V) and F are assumed to be a priori independentand distributedaccording to a noninformativeconjugate prior (see choices for the conjugate prior). Suppose also Greene 1990, Anderson 1984, for appropriate of Fp,V, and F. Then the marketmodel (1) implies that the residual return - is independent that the conditional distributionf(S, R I F, V, F) of the asset returns S = [r', r2...rP], can be factored as follows: and factor returnsB = [f', f2 ...fP] (87) V, f(S, R IA, V, F) = f(B I F)f(S I FL, B).
The a posteriori distributionf(ti, V, F IS, B) is, therefore, given by (88) f(, f(B F)f (F)f (S I i, V, B)f (p, V) , FIS, B)(S, SB)= I R) f (S,R) = (cf(B
IF)f(F)) - (c2f(S I p, V, B)f((x, V)),
where c1, c2 are suitable normalizingconstants. From (88) it follows that F and (Fi, V) are a posterioriindependent.Therefore,the uncertaintysets for (Fi, V, F) can be representedas a Cartesianproduct S x Sf-I. is the CDF of (2))m, where Suppose the confidence threshold is set to wo discussion of the upper bound on the a F((p + 1)/2, (p - 1)/2) random variable (for a(~<Fp p3achievable confidence, see ?6). Set FO= FmIand set q by solving (76). Then F e Sf - (q) with confidence wm.As before, let S(o) be the on confidence set for (Fp,V). Let f* be the optimal solution of the robust Sharperatio problem with uncertaincovariance and let s*(o) be the corresponding value. Then the confidence that (ix, V, F) e S(w) x Sf-i (q) is Cw(m+n), i.e., with a confidence level o(m+n), the realized Sharpe ratio of +* is at least as large as s*(o).
PROBLEMS SELECTION ROBUST PORTFOLIO
37
The latter developmentrelied on the fact that an independentBayesian prior ensures that the joint confidence set for (Fp,V, F) is the Cartesianproduct of the individual confidence sets for (iL, V) and F. Such a partition is not immediately obvious in the non-Bayesian setup and, consequently, it is not clear how the uncertaintyset Sf given by (77) could lead to probabilisticguarantees. Acknowledgments. The first author's research was partially supportedby DOE Grant GE-FG01-92ER-25126 and NSF Grants DMS-94-14438, CDA-97-26385, and DMS-0104282. The second author'sresearchwas partiallysupportedby NSF GrantsCCR-00-09972 and DMS-01-04282. References
Anderson,T. W. 1984. An Introductionto MultivariateStatistical Analysis, 2nd ed. John Wiley & Sons, New York. Andersson, F., H. Mausser, D. Rosen, S. Uryasev. 2001. Credit risk optimization with conditional value-at-risk criterion.Math. Programming,Ser B 89 273-291. Ben-Tal, A., T. Margalit, A. Nemirovski. 2000. Robust modeling of multi-stage portfolio problems. H. Frenk, K. Roos, T. Terlsky, S. Zhang, eds. High Performance Optimization. Kluwer Academic Publisher, Dordrecht,The Netherlands, 303-328. A. Nemirovski. 1998. Robust convex optimization. Math. Oper Res. 23(4) 769-805. , 1999. Robust solutions of uncertainlinear programs. Oper Res. Lett. 25(1) 1-13. 2001. Lectures on Modern Convex Optimization:Analysis, Algorithms, and Engineering Applications, MPS/SIAM Series on Optimization.SIAM, Philadelphia,PA. Bertsimas, D., J. Sethuraman.2000. Moment problems and semidefinite optimization. H. Wolkowicz, R. Saigal, L. Vandenberghe,eds. Handbook of SemidefiniteProgramming.Kluwer Academic Publishers, Dordrecht, The Netherlands. Black, F, R. Litterman. 1990. Asset allocation: Combining investor views with market equilibrium: Technical report. Fixed Income Research, Goldman, Sachs & Co., New York. Boyd, S., L. El Ghaoui, E. Feron, V. Balakrishnan.1994. Linear Matrix Inequalities in System and Control Theory. SIAM, Philadelphia,PA. Broadie, M. 1993. Computing efficient frontiersusing estimated parameters.Ann. Oper Res. 45 21-58. Chopra,V. K. 1993. Improving optimization.J. Investing 8(Fall) 51-59. 1, C. R. Hensel, A. L. Turner.1993. Massaging mean-varianceinputs: Returns from alternativeinvestment strategies in the 1980s. Management.Sci. 39(7) 845-855. on 1,W. T. Ziemba. 1993. The effect of errorsin means, variances and covariances optimal portfolio choice. J. Portfolio Management.(Winter)6-11. Efron, B., R. Tibshirani. 1993. An Introduction to Bootstrap, No. 57, Monographs on Statistics and Applied Probability.Chapman& Hall, New York. El Ghaoui, L., H. Lebret. 1997. Robust solutions to least-squares problems with uncertaindata. SIAMJ. Matrix Anal. Appl. 18(4) 1035-1064. , N. Niculescu eds. 1999. Recent Advances on LMI Methods in Control. SIAM, Philadelphia,PA. , M. Oks, F. Oustry.2002. Worst-casevalue-at-riskand robustportfolio optimization:A conic programming approach.Oper Res. Forthcoming. F, . Oustry, H. Lebret. 1998. Robust solutions to uncertain semidefinite programs. SIAM J. Optim. 9(1) 33-52. Frost, P. A., J. E. Savarino. 1986. An empirical Bayes approachto efficient portfolio selection. J. Fin. Quan. Anal. 21 293-305. . (1988). For betterperformance:Constrainportfolio weights. J. Portfolio Management.15 29-34. , W. H. Greene. 1990. EconometricAnalysis. Macmillan Pub. Co., New York. Halld6rsson,B. V., R. H. Tiittincti.2000. An interior-pointmethod for a class of saddle point problems: Technical report. Carnegie Mellon University, Pittsburgh,PA.
Hansen, L. P., T. J. Sargent. 2001. Robust control and model uncertainty. Amer Econom. Rev. 91 60-66. Huber, P J. 1981. Robust Statistics. John Wiley & Sons, New York. Kast, R., E. Luciano, L. Peccati. 1998. VaR and Optimization. Workshop on Preferences and Decisions. Klein, R. W., V. S. Bawa. 1976. The effect of estimation risk on optimal portfolio choice. J. Financial. Econom. 3 215-231. Ledoit, O. 1996. A well-conditioned estimator for large-dimensional covariance matrices: Technical report. Anderson School, UCLA, Los Angeles, CA. Lintner, J. 1965. Valuation of risky assets and the selection of risky investments in stock portfolios and capital budgets. Rev. Econom. Statist. 47 13-37. Lobo, M. S., L. Vandenberghe, S. Boyd, H. Lebret. 1998. Applications of second-order cone programming. Linear Algebra Appl. 284(1-3) 193-228.
38
Markowitz,H. M. 1952. Portfolio selection. J. Finance 7 77-91. . 1959. Portfolio Selection. Wiley, New York. Mausser,H., D. Rosen. 1999. Beyond VaR: From measuringrisk to managing risk. ALGO Res. Quart. 1(2) 5-20. Michaud, R. 0. 1998. EfficientAsset Management:A Practical Guide to Stock Portfolio Management and Asset Allocation, Financial ManagementAssociation, Survey and Synthesis Series. HBS Press, Boston, MA. Mossin, J. 1966. Equilibriumin capital asset markets.Econometrica 34(4) 768-783. Nesterov, Y., A. Nemirovski. 1993. Interior-Point Polynomial Algorithms in Convex Programming. SIAM, Philadelphia,PA. Schafer, J. L. 1997. Analysis of IncompleteMultivariateData. Chapmanand Hall, New York. Sharpe, W. (1964). Capital asset prices: A theory of marketequilibriumunder conditions of risk. J. Finance 19(3) 425-442. Sttirm,J. 1999. Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones. Optim. Methods and Software 11-12 625-653. Zhou, K., J. Doyle, K. Glover. 1996. Robust and Optimal Control, Prentice-Hall,Englewood Cliffs, NJ. Ziemba, W. T., J. M. Mulvey, eds. 1998. WorldwideAsset and Liability Modeling. Cambridge University Press, Cambridge,U.K. D. Goldfarb: IEOR Department, Columbia University, New York, New York 10027; e-mail: gold@ieor. columbia.edu G. Iyengar:IEOR Department,Columbia University, New York, New York; e-mail: garud@ieor.columbia.edu

4126989

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

4126989

Hochgeladen von

Copyright:

Verfügbare Formate

Robust Portfolio Selection Problems Author(s): D. Goldfarb and G. Iyengar Source: Mathematics of Operations Research, Vol. 28, No.

ROBUST PORTFOLIO SELECTION PROBLEMS

D. GOLDFARB AND G. IYENGAR

D. GOLDFARB AND G. IYENGAR

min E[r+] a, {tIESm1 > 1T+ = 1.

VES,, {.LESm, DESd}

< a) < /3.

+ max t(+D+4 , {DESdI

min IT'4) a, {I=ES 1. 1T,0= 1.

PORTFOLIO SELECTION PROBLEMS ROBUST

then (13) reduces to the convex quadraticallyconstrainedproblem minimize A+ 8,

? = S, = V:V = V +W, IIWll Tr(WTW) p},

and the worst case variance is given by,

maximize II(V0 +W)4II2

i=1 g i=Y 1EEiWiV i=1 i

Wil=g i=l <

the optimization problem,

least as large as that of (16).

AND G. IYENGAR D. GOLDFARB

II {y: Ilyllg_<r IlYyo+

< ()y:llyllg<r}f max Ilyo+YII -V

where QAQT is the spectral decomposition of H = G-1/2FG-1/2, A = diag(Ai), and w =

(ii) there exist 7

0 and s E Rm that satisfy

where POPT is the spectral decomposition of K = F'/2G-'FI/2, pTFl/2yo.

define w = QTHI/2G/2y/2 = Al/2QTGl/2yo.

Let the spectral decomposition of H = G-'1/2FG-1/2be QAQ', where A = diag(k), and

matrix M >- 0 if and only if

where a = r2/7. It follows that

if and only if there exists 7, a > 0 and t E R' satisfying,

7 0-1= 0 , m (i.e., 7 < 1/AmIax(K)), ui

However, M >- 0 if and only if 7 < 1/0i, for all i =

max Ilyo +ryll

if and only if there exists 7 > 0 and s e R' satisfying,

4zz < (x + y)2 (xy)2 2

(H) ' Amax

i= OAi+? ti, 1,...9 m,

where QAQT is the spectral decomposition of H = G-1/2FG-1/2, A = diag(Ai) and w =

2i1/20]11 -1-5 1)+8,

D. GOLDFARB AND G. IYENGAR

maximize subject to (34)

4 0>o, F, (p'4; A- ; +) E Y(Vo, G).

max minm IT4_rf ,DESdl (~:<b>O,1T~r)(VES,,tLES,, Var[r]

max ?TVT FV4 + +bTDf , {vES (O - y - rfl)T > 1,

is equivalent to the robust minimum variance problem

- rf1)' > 1 A4 > b,

VESv, {ILESm, DESd 1T+ = 1,

D. GOLDFARB AND G. IYENGAR

On incorporating(42), the problem (40) can be rewrittenas maximize minu IL

max VESV, DESd) SFl/2V4 -S(P) = 1, 1

maximize (FIo- y)T ,

to subject -9r1i 3) i (Ioa ,y - -8 I1/2 811 ,

if and only if there exist I > 0 and s E R' such that,

(1 - ?Oi)si, < ui < 1, Amax(K)F

u2 < (v Amax(K)7 < v.

_< (F(L) 8, ID1/'120

(K) > 0, Amax 1T+ = 1,

and Ei = [e, ... , eP']T the vector of residual returnscorrespondingto asset i. is

D. GOLDFARB AND G. IYENGAR

if rank(A) = m + 1. Substitutingyi = Axi + Ei, we get xi - xi = (ATA)-'AT where Z = o.2(ATA)-'. Hence

P ((ji - xi)T(ATA)(ii - xi)

S(W) = S (to) x S2((0) x

< = 1,..., o + v, Ivil yi, i