by W. V.

Harlow

Asset Allocation in a Downside-Risk

Framework
A downside-risk approachto investmentdecisionsuses intuitivemeasuresof riskthatfocus
on returndispersionsbelowa specifiedtargetor benchmark return.Downside-risk measures
areattractivenot only becausetheyareconsistentwith investors'perceptionof risk,butalso
becausethe theoreticalassumptionsrequiredto justify theiruse are very simple. Equally
important,a numberof well known risk measures,including the traditionalvariance
(standarddeviation)measure,are special cases of the downside-riskapproach.Asset
allocationin a downside-risk
framework thereforedeterminesan investmentopportunityset
for downside-averse investorsthatis at leastas efficientas thatderivedusing conventional
techniques.
A set of internationalasset allocationexamplesdemonstratesthe benefits of the
downside-risk framework.Specifically,optimizationsbasedon downsidemeasuresproduce
portfoliostrategieswith realizedreturnsthat haveless downsideriskexposurethan those
determinedusing variance.Thusinvestorsaverseto below-target returndispersionsachieve
a more attractiverisk-returntradeoffwithin this framework.Moreover,in the asset
allocationexamplesconsidered,the downside-risk approachproducesa significantlyhigher
averagebondallocationrelativeto stocks. This differencein asset compositionincreases
downsideprotectionwhileofferingthe sameor a greaterlevelof expectedreturn.

C ENTRAL
TOMODERNportfoliotheory overall magnitude of a loss, should one occur.
is the premise that investment decisions
are made to achieve an optimal risk/
return tradeofffrom the availableopportunities.
In order to meet this objective, the portfolio
manager must first evaluate capital market in-
formationand quantify ex antemeasures of both
risk and expected return for the appropriateset
of assets. The next task is to isolate those
combinations of assets that are the most "effi-
cient," in the sense of providing the lowest level
of risk for a desired level of expected return, and
then to select one combinationthat is consistent
with the risk tolerance of the investor.
While the principle of identifying portfolios
with the requiredrisk and return characteristics
is certainly clear, the appropriate definition of
risk is more ambiguous. One manager might
view risk as the probability of shortfall below
some benchmark level of return, for example,
while another may be more sensitive to the
These seemingly disparate notions of risk, as
well as other possible definitions, serve as a
reminder that simple return variance (or stan-
dard deviation)-the traditional measure of
risk-is sometimes deficient for dealing with the
rich set of portfolio objectives and constraints
that investment managers often formulate.
This articlediscusses and demonstratesa gen-
eral approach to asset allocationbased on defi-
nitions of risk that are attractivealternatives to
variance.1These alternativesall capture the ap-
pealing notion of "downside risk" and provide
a more robust approach to portfolio optimiza-
tion. Using an asymmetric measure of risk that
focuses on the returns below a specified target
or benchmark return level, this framework in-
cludes as special cases such well known mea-
sures as the probability of loss, expected loss
1. Footnotes appear at end of article.

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 O 28

The CFA Institute

is collaborating with JSTOR to digitize, preserve, and extend access to
Financial Analysts Journal ®
www.jstor.org

Glossary
Asymmetric Measure of Risk: Any measure of
risk that focuses on a portion of the return
distributionratherthan the spread or dispersion
of the overall distribution. Downside-risk mea-
sures are asymmetric and isolate return devia-
tions in the left tail of the distributionthat fall
below a specified target rate.
Semivariance:An asymmetricmeasure of risk that
focuses on squared return deviations below the
mean of the distribution.Targetsemivarianceis
a similar but more general measure in which
return dispersions are considered below any
arbitrarytarget or benchmarklevel of return.
Target Shortfall: A concept of downside risk that
captures the severity of not achieving some
minimum target or benchmarkreturn. The tar-
get shortfall represents the expected deviation
of returns falling below the target rate.
Mean-VarianceFormulation:The traditionalfor-
mulation of the investment decision problem,
stated in terms of the expected mean returnand
variance of a portfolio of assets. In order to
achieve the most efficient portfolio, assets are
combined so as to minimize variancefor a given
level of return.
Downside-Risk Optimization:An alternativefor-
mulation of the investor's decision problem us-
ing a downside measure of risk as opposed to
the variance or standard deviation of asset re-
turns. Thus, instead of minimizing variancefor
a given level of return, a downside measure
such as target semivarianceis minimized.
and semivariance.2 It also includes the tradi-
tional variance measure as a special case, thus
ensuring that the downside-risk approach de-
termines an efficient frontier for a downside-
averse investor that is at least as optimal as that
derived using existing techniques.
Although the concept of downside risk is
older than modern portfolio theory itself, the
theoretical development and extensions of this
risk framework are relatively recent. Unfortu-
nately, empiricalinvestigations of this approach
within the context of asset allocation remain
somewhat limited. To obviate this deficiency,
the present study examines the downside-risk
frameworkby considering an internationalasset
allocation example across 11 countries, using
returns that span the last decade. In addition, it
compares the downside-risk results with those
derived using the mean-varianceapproach.
We find that empiricaloptimizationsbased on
downside measures are more efficient than
FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 I 29
mean-variancemeasures in the sense that the
selected portfolioshave less downside exposure
than those determined using variance. Thus
investors averse to below-target return disper-
sions achieve a more attractive risk/return
tradeoffwith a downside-riskframework.More-
over, in the asset allocation examples consid-
ered, the downside-risk approach produces a
significantly higher average bond allocation
than a traditional approach. This difference in
asset composition increases downside protec-
tion while offeringthe same or a greaterlevel of
expected return.
In view of its appealing theoreticaland intui-
tive features, the downside-risk framework
should provide a useful set of tools for portfolio
managers considering a broad set of problems.
Perhapsequally important,this approachcan be
implemented in a manner similar to the stan-
dard asset allocation procedures currently in
place, with little added complexity.
Investment Risk
In his seminal contributionto modern portfolio
theory, Harry M. Markowitz pointed out that
investors could reduce the overall risk of their
investments by forming well diversified portfo-
lios.3 As a part of his analysis, he considered
various alternativerisk measures in addition to
variance and concluded that the most theoreti-
cally robust measure was semivariance (i.e., the
expected value of the squared negative devia-
tions about a specified "target"rate of return).
Semivariance captures the notion of down-
side risk and is an appropriatecharacterization
of investment risk because investors are often
concerned about losses relative to a threshold
return level. It is important to note that, unlike
the variance measure, semivariance does not
increase with greater "upside potential." Up-
side potential is, rather, captured by the mean
of the return distribution.
Because of purported computational prob-
lems associated with calculating the semivari-
ance statistic, Markowitz adopted variance as
the risk measure in his analyses. As a result,
much of the initial research in finance focused
on the issues surrounding a simpler mean-
variance framework. Unfortunately, the theo-
retical assumptions necessary to support vari-
ance as a risk measure (discussed below) are
somewhat restrictive.Moreover, varianceis not
consistent with investors' actual perception of
risk.
Figure A Downside Riskfor a TargetRateof Returnof 6 Per Cent
0% 6%
Target
As researchersin finance, economics and psy-
chology have noted over the past three decades,
individuals view return dispersion in an asym-
metric manner;that is, losses weigh more heav-
ily than gains.4 The examples of risk mentioned
earlier-the probabilityof shortfall below some
benchmark level of return and the expected
magnitude of a loss-typify the concerns of
portfolio managers. While these measures differ
significantly,they both capturesome element of
downside risk. We focus our attentionon down-
side-risk measures in order to alleviate some of
the shortcomings of variance and establish a
more general approach to asset allocation.
Downside-Risk Measures
Several classes of downside-risk measures are
of particularinterest in finance. All involve the
tail of the relevant distributionof returns below
some specific threshold level or target rate.
These risk measures are referred to as "lower
partialmoments" (LPMs),because only the left-
hand tail of the returndistributionis used in the
calculation.5
Computationally, the LPM for an empirical
(discrete) distribution of portfolio returns, Rp,
with a target rate, i, is described by:
T computed around the mean of the distribution,
LPMn = > pp(t - Rp
Ro = - xc
FINANCIALANALYSTSJOURNAL/ SEPTEMBER-OCTOBER
15%
Mean
where pp is the probability that return, Rp,
occurs.6The type of "moment," n, specified in
Equation (1) captures an investor's preferences
by determining the manner in which the return
dispersion below the target is characterized.7
Figure A depicts a normal distributionwith a
mean of 15 per cent and a standarddeviation of
10 per cent. For a target rate of 6 per cent, the
applicabledownside portion of the distribution
is representedby the shaded area. Forn = 0, the
risk measure becomes a Oth-ordermoment (de-
noted LPMO),with the term in brackets being
raised to the 0th power (i.e., equal to 1). Hence
the measure is simply the probabilityof falling
below the target rate. For this particularcase,
LPMoequals 0.184. In other words, there is an
18.4 per cent chance that the return perfor-
mance will fall short of the desired minimum
level signified by the target rate.
Forhigher-ordermoments, the shaded areain
Figure A remains pertinent for the risk calcula-
tion. However, for n = 1, LPM1becomes the
expected deviation of returns below the target,
or the target shortfall.8 LPM1for this example
equals 1.01 per cent. For n = 2, LPM2is analo-
goits to variance, in that it is a probability
weighting of squared deviations. Rather than
(1) however, the deviations are determined with
respect to the target rate. LPM2 can thus be
1991El 30
referred to as a target semivariance. In our
example, this measure equals 9.38 (or a 3.06 per
cent target semideviation, using the square root
of the LPM2measure).
Many popular notions of risk are special cases
of the generalized LPMnmeasure. For example,
with n = 0 and a target rate equal to 0 per cent,
LPMois simply the probabilityof a loss. For n =
2 and a target rate equal to the mean of the
distribution, LPM2 becomes the traditional
semivariance measure. Furthermore, for nor-
mal, or symmetric, distributions, LPM2is ex-
actly proportional to variance [LPM2(r = R) is
equal to one-half of variance]. Using it as a risk
measure would be equivalent to using variance
and result in the same ordering of risky assets.
While the LPMnmeasure of risk has obvious
intuitive appeal, it is important to consider the
economic justificationfor its use and the general
conditions under which it is appropriate. The
distinction between this framework and the
traditional mean-variance approach lies in the
assumptions regarding the distributionalprop-
erties of returns and investor preferences. As
mentioned earlier, the use of variance as a
measure of risk requires a somewhat restrictive
set of assumptions. Specifically, either returns
must be normally distributed, or investors have
to exhibit behavior describable by a quadratic
utility function.9 Within the LPMnframework,
distributions can be any one of a class charac-
terized by a location and scale parameter(e.g.,
mean and standard deviation).10 In addition,
the downside-risk framework makes only gen-
eral assumptions regardinginvestor utility func-
tions (e.g., risk aversion and skewness prefer-
ence).
The power and flexibilityof the downside-risk
frameworkstem from the jointset of its assump-
tions regarding investor preferences and asset
return distributions. Unlike other frameworks,
which place restrictions on preferences or on
distributions, the LPMnapproach uses a com-
bined set of reasonable and less restrictive as- equal, two distributions with the same LPMn
sumptions. As a result, downside-risk analysis
is not only more attractivein terms of its con-
sistency with the way investors actually per-
ceive risk, but it is also valid under a broader
range of conditions.
Portfolio Optimization in a Downside-
Risk Framework
Having established a formaldefinition of down- cess. The probabilityof target shortfall, there-
side risk, we now consider its use in the context fore, cannot be used for the general purpose of
FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 O 31
of portfolio optimization. The objective within
this frameworkis essentially the same as in any
approach-i.e., select a portfolio of assets in
some combinationso as to minimize risk subject
to a specified level of expected return. In this
case, however, LPMnis used as the appropriate
characterizationof risk.
Figure B illustrates this decision problem
graphically.If Rp*is the investor's desired level
of return, then P* represents the portfolio that
provides the lowest risk. For alternativevalues
of Rp*,the resulting set of solutions trace out a
convex mean-LPMnefficient frontier (denoted
MLPMn), reflecting the optimal risk/return
tradeoffs in exactly the same manner as the
traditionalmean-varianceapproach.
Stated more formally, an investor who is
averse to downside risk and who has a target
rate of return, r, must determine the allocation
weight, Xj, for each relevant asset, j, to achieve
an efficient point within the investment oppor-
tunity set. The nonlinear MLPMn optimization
problem is represented by:
Select X to minimize: LPMn(T;X)
T
= E
Pp(T - Rp)n (2)
Rp< T
Subjectto: n = 1 or 2
{f;XPE(Rj)= Rp*}and
{EjXj= 1, Xj > 0} without short sales or
{NjXj= 1} with short sales,
where pp is the probabilityof portfolio return,
Rp*.
Once again, we note that the optimization
process uses the entire distribution of returns.
Information contained in the right tail of the
distribution (the returns above the target) does
not contribute to risk, but it is captured in the
mean of the distribution.Thus, all things being
but differentmeans are not the same. In such a
case, the distributionwith the higher mean has
a greater degree of positive skewness.
One necessary restrictionon the optimization
in Equation (2) is that LPMn must be of first
order or higher (i.e., n must be greater than
zero) if any form of risk aversion is to be
considered relevant in the decision-makingpro-
Figure B Downside-Risk(MLPMn)EfficientFrontier
E(R)
MLPMn Efficient Frontier
Rp* /
LPMn (Rp*;Target)
portfolio construction. The primary reason for
this is that the LPMomeasure does not differ-
entially weight the returns in the lower tail of
the distributionbased on their "distance"from
the target. In other words, the shortfall proba-
bility is an incomplete measure of risk, because
it fails to provide any indication of how severe
the shortfallwill be, should it occur. As such, it
is inappropriateto use this measure directly to
define an efficient frontier.11
Note that, while the probability of target
shortfallis not a complete measure of risk, it can
be incorporated into the asset allocation deci-
sion. Specifically, LPMo provides a means of
isolating the segment of the investment oppor-
tunity set most relevant to the investor. As
Leibowitz, Henriksson and Kogelman have
shown, the shortfallprobabilitycan be specified
as a constraint in a portfolio optimization prob-
lem that utilizes some other measure of risk-
for example, variance-to delineate the efficient
frontier.12This shortfall restriction essentially
provides information regarding the risk toler-
ance of the investor.
For the mean-variance formulation of Equa-
tion (2), the shortfall constraintis equivalent to
an upward-sloping line in Figure B, with an
intercept on the return axis representing the
investor's target rate and a slope tied to the
probability of falling below the target. The in-
vestor specifies both the probability and the
FINANCIALANALYSTSJOURNAL/ SEPTEMBER-OCTOBER
LPMn
target, thus establishing a level of aversion to
below-target exposure. This constraint inter-
sects or is tangent to the efficient frontier, indi-
cating which portfolios are attractive to the
investor. The probability of target shortfall,
therefore, can be thought of as a risk-tolerance
assessment tool when coupled with another
measure of risk. This is true whether the mea-
sure is variance or a downside-risk criterion
such as LPM1or LPM2.
The appeal of the downside-risk frameworkis
certainlynot diminished because target shortfall
probabilitiescannot be used directly as a risk
measure in a portfolio optimization problem.
Indeed, beyond LPMO,there exists a very large
class of theoreticallyattractiveand flexible risk
measures. LPM1 (target shortfall) and LPM2
(targetsemivariance)provide an intuitive set of
risk definitions that are more useful than tradi-
tional approachesand are valid under a broader
set of economic conditions.
Downside-Risk Optimization
To appreciate the downside-risk framework
more fully, it is useful to examine its implemen-
tation under alternative formulations. As with
any asset allocation decision, the analysis can
proceed in a variety of ways, based on the
manner in which informational inputs are de-
veloped. In particular,the optimizationproblem
stated in Equation (2) can be solved using ex-
1991 O 32
plicit return forecasts with probabilitiesof real-
ization derived through some independent
modeling process. Alternatively, the analysis
can be accomplished using historical data as
proxies for ex ante asset behavior.
Equation (2) can be modified to utilize histor-
ical returns, as follows:
Select X to minimize: LPMn(r;X)
T1
=E (r-R )n (3)
Rp T-
Subject to: n=1 or 2
{fjXjE(Rj)= Rp*}and
11AX= 1, Xj> 0},
where T is the number of return observations.
In other words, within the optimization pro-
cess, LPMncan be computed by constructingan
empirical portfolio distribution given a set of
asset weights, X, and T previously observed
returns. This value is then minimized subject to
an expected returnconstraint,Rp*.The expected
return for each asset, E(Rj),is determined using
its historicalmean risk premium in conjunction
with the current risk-free rate of interest, Rfo.
That is, the asset's excess return at time t, Rjt-
Rft, is first averaged over T observations. This
value is then added to Rfoto obtain the expected
return, so that E(Rj)= [NT (Rjt - Rft)/T] + Rfo.
Keep in mind that, for purposes of this dis-
cussion, we are interested primarilyin evaluat-
ing the usefulness of an alternative portfolio
constructiontechnique, as opposed to a method
of producing return forecasts. As most asset
allocationprocedures rely to some extent on the
information contained in historical returns, the
problem formulated in Equation (3) is repre-
sentative of typical implementationsand is thus
the focus of our attention.
We consider a global asset allocationproblem
relevant to many portfolio managers. We in-
clude fully currency-hedged equity and fixed
income markets in 11 countries-the United
States, the United Kingdom, Japan, West Ger-
many, Switzerland, France, the Netherlands,
Sweden, Australia, Canada and Hong Kong.
The returns used in this analysis span the 11-
year period from January 1980 to December
1990.
Figure C presents actual MLPMn efficient
frontiers for an investor who defines risk as any
FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991
loss of initialcapitaland thus has a target rate of
return, r, equal to 0 per cent. These frontiersare
computed using Equation (3) and 60 months of
recent historical return data (i.e., T = 60) to
derive estimates of the LPMnrisk measure and
expected asset returns. Panel A depicts the
MLPM1(target shortfall)frontier, while Panel B
illustrates the MLPM2frontier using the same
target rate of 0 per cent. In the latter case, the
square root of the LPM2risk value is plotted to
provide a measure comparable to percentage
return (i.e., target semideviation). Both panels
also indicate the risk and return of a global
capitalization-weighted benchmark composed
of 60 per cent equity and 40 per cent fixed
income securities, based on the Salomon-
Russell Global Equity Index and the Salomon
BrothersWorld Government Bond Index.13
In Panel A, the 60/40 benchmark has an
expected return of 13.5 per cent and a target
shortfall of 9.07 per cent. In other words, be-
cause the investor's target rate has been speci-
fied as 0 per cent, the magnitude of the expected
probability-weightedloss is 9.07 per cent. By
comparison, an MLPM1efficient portfolio deliv-
ering the same benchmark expected return of
13.5 per cent has only a 4.53 per cent expected
target shortfall. In Panel B, which uses target
semideviation to define the frontier, the risk of
the benchmark is estimated at 23.20 per cent,
while the risk of the comparable MLPM2effi-
cient portfolio is projected at 12.40 per cent.
Figure D indicates the effect that altering the
target rate of return, T, has on the magnitude of
the estimated risk. For ease of comparison, we
have included the previous frontiers, with tar-
get rates equal to 0 per cent as well as those
constructed using target rates of 8 and 16 per
cent. Forthe lattertwo cases, the specificationof
risk is based on measures of dispersion that
capture underperformance relative to higher
benchmark target levels of return. While the
investor with a target return equal to 0 per cent
is concerned with the loss of any principal,
these higher targets represent the risk of return-
ing less than the risk-freeasset (r = 8%)or less
than the market portfolio (r = 16%).
Inspection of the figures indicates that, as the
target rate increases, both the MLPM1 and
MLPM2frontiers shift to the right. This occurs
because more of the distributionof returns falls
below the specified target rate. The downside
component of returns thereforebecomes larger,
increasing the numerical value of the risk mea-
133
Figure C Downside - RiskEfficientFrontiers

Panel A: Target Shortfall (MLPM1) Efficient Frontier

20

MLPM,Frontier
15
8 - (Target=0%)

Benchmark(60/40)

x; 10

5
0 5 10 15

LPM1 Target Shortfall (%)

Panel B: Target Semideviation (MLPM2) Efficient Frontier

20 -

15 MLPM2Frontier
(Target=0%) /
Benchmark(60/40)

a)
a.)
X 10

5 11 11 1 I 1 11 1 11 11il11llll
0 5 10 15 20 25 30

LPM2 Target Semideviation (%)

sure. Accompanying this shift is an increase in
the allocation to stocks. This is certainly not an
intuitive result;rather,it represents the complex
interactionbetween the portfolio distributionof
returns and the target rate of return.
Another interesting result is that, for a given
target rate and expected return, the LPM1and
LPM2allocationsdiffersubstantially.For T-equal
to 0 per cent and a 15 per cent expected return,
for example, the LPM1stock allocation is ap-
proximately10 per cent higher than that for the
LPM2frontier (42.87 versus 32.83 per cent). In
fact, the relative stock allocations for the LPM1
optimization are higher for all values of ex-
pected return, with the differences becoming
smaller toward each end of the frontiers.
This disparity highlights the importance of
risk definition to the asset allocation decision.

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 O 34

Figure D Downside - RiskEfficientFrontiersfor AlternativeTargetRatesof Return

Panel A: Target Shortfall (MLPM1) Efficient Frontier

20
MLPM1EfficientFrontiers

Stocks:42.87%
Bonds:57.13%

154

av / / Stocks:43.72% S
x 10 Bonds:56.28% Bonds:53.59%

Target= 0% Target= 8% Target= 16%

5-
0 5 10 15
LPM1 Target Shortfall (%)

Panel B: Target Semideviation (MLPM2) Efficient Frontier

20 -
MLPM2EfficientFrontiers
Stocks:32.83%
Bonds:67.17%

Q) ~~~~~~~~~~~~~~~~~~~~~~~~Stocks:
36.73%
Q) 10 ; / / / \ |Bonds: 63.27%
X 10

Target= 0%/ Target= 8% Target= 16% Stocks:35.01%

Bonds:64.99%

5-

5 10 15 20 25 30
LPM2 Target Semideviation (%)

Simply changing the focus of attention from
target shortfall to target semideviation dramati-
cally alters the composition of the investor's
portfolio. Indeed, this choice has a much larger
effect on the allocation decision than the choice
of a target rate for a specific type of risk mea-
sure.
The actual benefits of the downside-risk
framework relative to the traditional mean-
variance setting must be determined empiri-
cally. If return distributions are approximately
normally distributed, then variance is a suffi-
cient measure of risk, and differences between
the two approaches will be small. But if returns
are not symmetrically distributed about the
mean, but instead possess some degree of skew-
ness, then the asset allocation decision reached
within the downside-risk framework can be

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 [ 35

Figure E Comparison of Downside-Risk and Mean-Variance Efficient Frontiers

PanelA: TargetShortfall(MLPM1)EfficientFrontier
20 -
- MLPM2Frontier(Target= 0%) .....*.

_ [ Stocks:32.83%
_ | Bonds:67.17%

Stocks:40.53%
Bonds:59.47%
;x J 10 Mean-Variance
Frontier

5-
5 10 15 20 25 30
LPM2TargetSemideviation(%)

significantly different from that obtained using normally distributed. The fact that they do not
mean-variancerelationships. It is to this poten- indicates the presence of skewness in the his-
tial distinction that we now turn our attention. torical return distribution. The variance risk
measure cannot capture the asymmetry in re-
Downside-Risk vs. Mean-Variance turns relative to the target rate.
Efficient Frontiers Figure E also indicates the stock/bond mix at
To make a fair comparison between the two the 15 per cent expected return level. For this
asset allocation frameworks, we use only the particularset of return data, the asset allocation
LPM2risk measure, because it, like variance, is under the MLPM2decision rule leads to a much
a second-order measure in which deviations higher bond allocation than that using the
from some return level are being squared. Fig- mean-varianceapproach(67.17 versus 59.47 per
ure E compares the MLPM2(r = 0%) efficient cent). The allocationdifferencesdiminish as the
frontier previously discussed with that com- frontiers converge toward each end. But at
puted using variance as the risk measure. (In intermediate return values, the two portfolio
order to display both frontiers on the same construction methodologies can lead to very
graph, we made LPM2 calculations for the differentasset allocation decisions.
mean-variance portfolios.) For both frontiers,
the square roots of the risk measures (i.e., target Downside-Risk vs. Mean-Variance
semideviation and standard deviation) are used Strategies
in order to reflect return-comparableunits. The differences between MLPM2and mean-
It is immediately apparent that the mean- variance portfolios, given a particular set of
variance frontier lies inside the MLPM2oppor- historicaldata, do not necessarily represent dif-
tunity set. Hence the MLPM2 portfolios are ferences in realized portfolio performance.It is
more efficient in that, for the same level of important to investigate the characteristicsof
expected return, they provide more downside these two alternative allocation strategies in
protectionthan those determined using a mean- terms of their ex post returns and risk.
variance decision rule. This holds over the en- To accomplish this task, we constructed port-
tire range of risk/return possibilities, with the folios using both optimization techniques. Be-
largest differences at intermediatevalues. ginning in January 1985, we used return data
Note that these frontierswould coincide if the over the preceding 60 months to develop the
historical returns used to construct them were required inputs. The return objective for both

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 O 36

strategies was the expected return on each of LO0 m w
O C m OC O O !
several global capitalization-weighted bench-
marks (e.g., a portfolio comprising 60 per cent
equity and 40 per cent fixed income securities).
In other words, for the MLPM2portfolio opti- nts O A S o o o t
mization in Equation (3), the expected return in \0 N N t

constraint, R *, was set equal to the global

benchmarkexpected return, and the downside-
risk measure was minimized. For the mean-
~
:3>~ io ~ ecnC,
tn t? cq t, c
variance strategy, the equivalent expected re-
turn constraintwas employed. Once again, only
c sM m m '.C
the LPM2target semideviation was considered s:ttt?N Ln

for purposes of direct comparison with the

mean-varianceapproach. tS-O N n mN p x r
XH
S Ce H~~~~H~H~~NNNen
To be consistent with the earlierexamples, we VCf

arbitrarilyset the target rate, r, equal to 0 per

cent. This represents an investor concerned
with the loss of any portion of initialinvestment 0, m mo nsoo t"Coo \D c

principal. This portfolio constructionprocedure

was applied each month, as the portfolios were
rebalanced over the next 72 months ending in a, nO t cn D ONa, " o
tt O O
t-,-
X
December 1990. In each case, the 60 months of N N s sN N, \0 \0 \?

return data immediately preceding the portfolio

formation month were used for the optimiza-
tions.
Table I summarizes the risk and return char-
acteristics of approximately1600 portfolio opti- 00 00 0 0 0 0 00000 N .C
mizations, using alternative benchmarks rang-
ing from 100 per cent fixed income to 100 per ON oo
tr>
qN
c>a
(3 a,\
en
cq
g

cent equities. It reports for each risk approach H ~ ~~~~~~

4Ua o,
6
t~ o,
a4 6 4 '6 0

the annualized geometric return, average

monthly return,standarddeviation, targetsemi- A t ~~~~U t- 00 0 C- , \c qo N t ?

deviation and minimum monthly return.

Columns (5) and (10) show that the MLPM2
approach outperforms the mean-variance ap- CZ 4o.1oq> Q)
proach in all cases by providing more downside
protection (as measured by smaller realized tar- :> 0! M
noNX~U ._rCN
get semideviation). Furthermore,in 10 of the 11 C E t b u: vo
un DNc t-s as Co<> o @,
cases, the minimum monthly return is smaller
for the downside-risk approach; this provides
an additional confirmationof risk reduction.
g E E AS g N eta t Na, t, 0 cv CDCD
Not only are the MLPM2portfolios less risky,
they also provide larger returns. A comparison
tVX m~~~~M
ooNN CI)9oo oo
NkOO o O

of columns (2) and (7) indicates that the annu-

alized MLPM2return is larger than the mean-
variancereturnfor all sets of portfolios, with the o~~~~~~~\
Q
St ,
t;-~~00
N cq cq O O00
00 a) r--
o" *m enC
c
N
N un
Cl
\C
N ,
Wv
difference ranging from 0.18 per cent (0/100
benchmark)to 1.58 per cent (100/0benchmark).
CZ
The average monthly returns, reported in col- tAn

umns (3) and (8), are also largerfor the MLPM2 S?SSSeE , =<
strategies. The higher ex post results for the \5 -
MLPM2approachappear to be an artifactof the n~~~~~~~~~c
U) m
t._ ,$CN
lower downside exposures of the MLPM2strat-
F
Q~~~4'
_ < s

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 0 37

Figure F Net PerformanceDifferenceBetween Downside-Riskand Mean-VarianceStrategies(1985-1990)

1.75 -
1.50 A
tJ~~~ A
> 1.25 -A
1.00-
A
0.75 -
0.50 - A
0.25 AA A
0.00
-0.25

1.4 -0.50
-0.75 -
-1.00 -
-1.25 -
-1.50
-1.75
-5.0 -4.0 -3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0

RiskDifference(%)MLPM2vs MV

egies when compounded over multiple holding
periods. 4
As a separate evaluation of the differences
between the MLPM2and mean-variance deci-
sion rules, the last two columns of Table I
summarize the below-targetreturns of both sets
of portfolios. Specifically, column (12) indicates
the percentage of times the MLPM2portfolios
outperformed the mean-variance portfolios
when the latterportfolios returnedless than the
target. For all benchmarks,the MLPM2portfolio
returns exceed the mean-variancereturns in a
majorityof the 72 months. A t-test of this ratio,
in column (13), shows that it is significantin six
of the 11 cases.
Figure F plots the annualized return differ-
ence for each set of strategiesnoted in TableI, as
well as the reduction in target semideviation
obtained by using the downside-risk optimiza-
tion. Note that all observations fall in the north-
west quadrantof the figure, indicating the ben-
efits of risk reduction and return enhancement
documented previously.
Table II compares details of the asset mixes
that generated the results reportedin TableI. In
particular,columns (2) through (7) indicate the
mean fixed income allocation over the 72
months, as well as the minimum and maximum
bond allocations for both strategies. Column (8)
shows the average monthly difference in the
fixed income allocation, while column (9) pres-
ents a t-test of this value's significance.
In all cases, the average allocationto bonds is
higher for the MLPM2approach than for the
mean-variancemethodology and is significantly
differentfrom zero. The differences range from
a low of 1.53 per cent (0/100 benchmark) to a
high of 2.68 per cent (90/10 benchmark). The
month-to-month differencesare as large as 9.49
per cent (90/10benchmarkin column (11)).
The MLPM2 approach's tendency to favor
bonds is also evident in the large percentage of
times in which the MLPM2fixed income alloca-
tion exceeds the mean-variance allocation (re-
ported in column (12) of the table). This per-
centage is in excess of 79 per cent for all
benchmarks, reaching a high of 90.3 per cent
(20/80 benchmark). A t-test of this ratio, re-
ported in column (13), shows that it is signifi-
cant for all sets of portfolios.
These results, and those in Table I, suggest
that asset allocation in a downside-risk frame-
work provides an attractive and powerful al-
ternative to the traditional mean-variance
approach. The differences between the down-
side-risk and mean-variance approaches arise
because returns are not strictlynormallydistrib-
uted. More important, the departuresfrom nor-
mality in past asset returns appear to contain
informationuseful in ex anteportfolio decisions.

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 0 38

 r- LO "o CA d H dS m O Ln The use of empirical asset distributions may
en Lf Lf .d
mt unL n Lf) o o) o
thus be beneficial in modeling risk in a down-
side-risk asset allocation framework. While all
~-
A) _q sO fin c O Nm n N C\ N x
these results do not necessarily generalize to
other asset classes or other periods of time, the
0 Aal x tn en czm N crN differences documented here indicate a poten-
tial benefit to be derived by closely examining
the methodology of portfolio construction and
the distributionsof returns used as inputs.
S: _ 00 00 enN00 O) ? CD t N
?,
ON
00 e
o N ao N r- ON LO) CP t Ln.

Conclusions
Downside risk is an intuitive measure of risk,
focusing on the asymmetry of returns about
< Q _
ocr
ec N rs N
oo a\cr o
et "D
\Cs N o oa, c\
ct CD \ N
L
some target level of return. The LPMn, as a
tn d I formal definition of downside risk, further en-
hances the appeal of this approach because of
u
the very simple theoretical assumptions re-
u u u Q
Q v
. U
l
U
q
U
f , U U
00 n 0
U
D
U U
:
.
quired to justify its use. Of equal import, a
broad set of well known risk measures, includ-
ing the traditionalvariance(standarddeviation)
> r ~~~~~~~~~~>
measure, are special cases of the LPMnframe-
work. Asset allocationin a downside-riskframe-
m x a,\ O H N 0 m 'IC N S
work, therefore, determines an investment op-
portunityset for downside-averse investors that
is at least as efficient as that derived using
conventional techniques.
o .E D ~~~~~N ri q \o \ q~o Ln o q More generally, the downside-risk approach
offers the potential for portfolios that are more
attractive than mean-variance portfolios. That
~~~~~~~aN ON ON aN ON (7 C?N x x x 00 o,
is, a downside-risk approach can lower risk
while maintaining or improving upon the level
of expected return offered by mean-variance
approaches.
u
4~~~~~~F t u o .
crtoN C: ooo o t, nH C
Beyond these appealing features, the down-
0 0 00~ ci t- lZ 1D 1D a? Cn tIn 0 b

>
side-risk approachallows the portfoliomanager
to define risk in a way appropriateto the objec-
W: S ~~~~
gC N 6 06 n ei 6 tf f- 0o UQ 14 -
tives and constraints of the portfolio. By being
able to specify a target rate of return and the
manner in which return deviations below this
target are considered, the manager can more
o~~~~~~~~~o
to - Ln 00 cn o, \0a fully appreciate the risk/returntradeoffs avail-
able from a given set of assets.'5 U
X~~~~~~~
C< C4 u6t C;L C;
Q e Wn ~~~N (OHNCN x x x x O O ;s
Footnotes
1. Extensionsof this frameworkto asset pricing are
discussed in W. V. Harlow and R. K. S. Rao,
"Asset Pricingin a GeneralizedMean-LowerPar-
X~~~~~~~~~~6L aL
, r
, 8L >t
tial Moment Framework,"Journalof Financialand
QuantitativeAnalysis, September 1989, and V.
co I t|t% Bawa and E. Lindenberg,"CapitalMarketEqui-
D
libriumin a Mean, Lower PartialMoment Frame-
U

work," Journalof FinancialEconomics,November

1977.
2. An alternativeapproachto returnshortfallrisk is
considered in M. L. Leibowitz and R. D. Hen-

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 [ 39

 riksson, "Portfolio Optimization with Shortfall 9. Quadraticutility functions are unappealing, be-
Constraints: A Confidence-Limit Approach to cause they imply increasing absolute risk aver-
Managing Downside Risk," FinancialAnalysts sion. That is, investors with this type of utility
Journal,March/April1989 and M. L. Leibowitz functionrequirehigher risk premiums for a given
and S. Kogelman, "Asset Allocation Under investment as their wealth increases. This is
ShortfallConstraints,"Journalof PortfolioManage- contraryto both intuition and observed investor
ment,Winter 1991. behavior. See G. J. Alexander and J. C. Francis,
3. See H. M. Markowitz, PortfolioSelection(New PortfolioAnalysis(Englewood Cliffs, NJ:Prentice-
York:John Wiley, 1959). Hall, 1986).
4. For a comprehensive survey of the early litera- 10. These distributionsbelong to the two-parameter
ture, see R. Libby and P. Fishburn, "Behavioral location-scalefamilyand include normaldistribu-
Models of Risk Taking in Business Decisions: A
tions, Student t-distributionswith the same de-
Survey and Evaluation," Journalof Accounting
gree of freedom, and stable distributionswith the
Research,Autumn 1977. See also D. Kahneman
and A. Tversky, "ProspectTheory:An Analysis same characteristicexponent and skewness pa-
of Decision under Risk," Econometrica,March rameter (not necessarily zero).
1979. 11. The LPMoefficient frontier is not convex as de-
5. For details on LPM analysis, see V. S. Bawa, picted in Figure B, which makes portfolio selec-
"Optimal Rules for Ordering Uncertain Pros- tion difficult to generalize across a set of risk-
pects," Journalof FinancialEconomics, March1975, averse investors. As noted in footnote 7, the
and V. S. Bawa, "Safety First, Stochastic Domi- LPMorisk measure is consistent with all utility
nance and Optimal Portfolio Choice," Journalof functions for which the first derivativeis positive
Financialand QuantitativeAnalysis,June 1978. (u' > 0), and it makes no assumptions regarding
6. For a continuous distribution with probability the risk aversion of the individual. Thus risk
density function dF(R), the analogous form of aversion can only be captured by using LPMn
Equation(1) is: measures of first order or higher.
12. See Leibowitz and Henriksson, "PortfolioOpti-
rt mization," op. cit. and Leibowitz and Kogelman,
LPMn I!c (T- R)ndF(R).
"Asset Allocation,"op. cit.
7. the
Theoretically, order (n) of the LPMn measure 13. See "Salomon-RussellGlobal Equity Indices: In-
determines the type of utility function, u, consis- dex Construction and Methodology," Salomon
tent with that risk measure. LPMois appropriate BrothersInc, October1990, and "Introducingthe
for investors who prefermore to less wealth (u' > World Government Bond Index," Salomon
0), while LPM1is valid with all risk-aversefunc- BrothersInc, November 1986.
tions (u' > 0 and u" < 0). The LPM2 measure 14. All results are qualitatively identical when the
allows for all risk-averse functions that display October1987stock marketcrash is excluded from
skewness preference(u' > 0, u"< 0 and u"'> 0). the analysis.
8. The target shortfallis not the mean of the returns 15. I thank Keith Brown, Gary Gastineau, John
below the target (the mean of the shaded area in Howe, Stan Kogelman, Marty Leibowitz, Eric
FigureA is 0.099 per cent). Rather,it is the mean Lindenberg, Ramesh Rao and Gary Shkedy for
of the deviations from the target-i.e., T - Rp. their helpful comments.

FINANCIAL ANALYSTS JOURNAL / SEPTEMBER-OCTOBER1991 0 40