1

TrackingErrorandActivePortfolioManagement

NadimaEl-Hassan
SchoolofFinanceandEconomics
UniversityofTechnology,Sydney

and

PaulKofman
#

DepartmentofFinance
TheUniversityofMelbourne
Parkville,VIC3010
[email protected]

Persistentbear marketconditionshave ledto ashiftoffocusin thetrackingerrorliterature.

Until recently the portfolio allocation literature focused on tracking error minimization as a
consequence of passive benchmark management under portfolio weights, transaction costs
and short selling constraints. Abysmal benchmark performance shifted the literatures focus
towardsactiveportfoliostrategiesthataimatbeatingthebenchmarkwhilekeepingtracking
error within acceptable bounds. We investigate an active (dynamic) portfolio allocation
strategythatexploitsthepredictabilityintheconditionalvariance-covariancematrixofasset
returns. To illustrate our procedure we use Jorions (2002) tracking error frontier
methodology.WeapplyourmodeltoarepresentativeportfolioofAustralianstocksoverthe
periodJanuary1999throughNovember2002.

June25,2003

1.Introduction
Persistentbear marketconditionsonstock markets worldwidehavegeneratedabull market
in the tracking error literature. Practitioner-oriented journals (in particular the Journal of
Portfolio Management and the Journal of Asset Management, see our references) recently
devoted whole issues to implementation and performance measurement of tracking error
investment strategies. More technical mathematically-inclined journals (e.g., the
International Journal of Theoretical and Applied Finance) publish ever-faster optimization

Wearegratefulforhelpfulcommentsandsuggestionsfromtwoanonymousreferees,StanHurnand
participantsattheStockMarkets,Risk,ReturnandPricingsymposiumattheQueenslandUniversityof
Technology.
#
Correspondingauthor.
2
and risk measurement algorithms for increasingly realistic portfolio dimensions. Not
coincidentally, this surge in academic interest tracks the global stock market slump with
fundmanagerperformancecomingunderintensescrutinyfrominvestors.
With absolute return performance as a first-order condition of investors utility
preferences,trackingthebenchmarkascloselyaspossibleisnormallysufficientduringbull
marketconditions.Whenthebenchmarkfailstodeliver,however,fundmanagerswillhaveto
prove relative return performance against the benchmark. The aim is then to persistently
outperformthebenchmark.Ofcourse,suchperformancewillonlybefeasibleifthemanager
ispreparedtoacceptactiverisk,andhenceincurariskpenalty.Mostinvestmentfundsaccept
that investors want to cap this penalty and therefore set a maximum portfolio tracking error
accordingly. Thus, tracking error can either be the investment goal, or an investment
constraint.Thisleadstothefollowingtwointerpretationsofindextracking:

A passive strategy that seeks to reproduce as closely as possible an index or benchmark

portfoliobyminimizingthetrackingerrorofthereplicatingportfolio;
or,
an active strategy that seeks to outperform an index or benchmark portfolio while staying
withincertainriskboundariesdefinedbythebenchmark.

Whatdistinguishesthesestrategiesisthecompositionoftotalriskexposure.Bothactiveand
passive strategies will incur incidental risk, while the active strategy will also incur
intentionalrisk.Intentionalrisk mayconsistofstockspecificrisk(activestockselection)or
systematic risk (active benchmark timing). Interestingly, some index fund managers claim
that ex ante passive indexing generates persistent above average returns i.e., an active
portfolio outcome. Of course, what they really mean is that passive indexing often
outperformsthe averageactive strategy(whichis morelikelyareflectionofthepooractive
outcomes).Astandardmeasureusedtotradeoffactiveperformanceagainstintentionalriskis
the information ratio (a.k.a. appraisal ratio), defined as the portfolios active return the
alphadividedbytheportfoliosactiverisk.Thisstandardizedperformancemeasurecanbe
used to assess ex ante opportunity but it is more frequently used to assess ex post
achievement. Ex ante opportunity is defined by the maximum possible IR given a set of
3
forecast stock returns (and their forecast risk measures) and an inefficient benchmark
1
. A
passivemanager(minimizingtrackingerror)willhaveanexanteIRclosetozero.Anactive
manager (maximizing excess returns) will have a much larger IR. The ex post achieved IR
will depend on the realized excess return over the benchmark and as such depend on the
Information Coefficient (IC). The managers IC measures the correlation between forecast
excess returns and realized excess returns. Whereas the ex ante IR will (by necessity) be
strictly non-negative, the ex post IR can of course be negative. For a comprehensive
discussion of these performance measures, we refer to Grinold and Kahn (1999). Active
managers perceive that they need to frequently reallocate their portfolios either to capture
excess returns or to stay within a tracking error constraint. At times this may lead to a less
than perfectly diversified portfolio, and will incur substantial transaction costs and assume
hightotalrisk.Ofcourse,fewpassiveindexfundmanagersholdportfoliosthatexactlymatch
the index, e.g., due to liquidity constraints. Just as active management incurs transaction
costs,thepassivemanagerthenalsohastorebalancetheindexportfoliotomatchtheactual
indexreturnsascloselyaspossible.Thissuggeststhatthereisafairlyclosesymmetryinthe
treatment of active and passive tracking error strategies. The key difference in the
interpretation of passive and active ex post IR is that the best active manager will be
characterizedbyapersistentlylargepositiveIR,whereasthebestpassivemanagerwillhave
anexpostIRclosetozero.
2

The passive tracking practitioners have been well served by the academic literature.
Rudd(1980),ChanandLakonishok(1993)andChan,Karceski,andLakonishok(1999)are
butafewofthe manyexamplesofthisliterature.Theactivetrackingpractitionershavenot
yetattractedsimilarattention
3
.Tothebestofourknowledge,Roll(1992)andJorion(2002)
are the first papers that comprehensively derive and interpret active portfolio allocation

1
If the benchmark happens to be an efficient portfolio, the ex ante maximum possible IR will be zero! Of
course,thisishighlyunlikelyinpractice.TypicalbenchmarksliketheS&P500ortheASX200arecommonly
foundwellbelowtheefficientfrontier.
2
However,themeasurecanbepoorlydefined.Considerthepassivemanagerwhoactuallyholdsthebenchmark
portfolio.Forthismanager,theIRwillnotbedefined.
3
Moststandardinvestmenttextbooksstilldiscusstrackingerrorinapassiveportfoliomanagementcontext,see
e.g.,Eltonetal.(2003,p.677).Trackingerroristhentypicallydefinedasthestandarderrorofaregressionof
the passive portfolio returns on benchmark returns, see also Treynor and Black (1973). This regression
measure is appropriate if the beta in the regression equals 1 (as it would for passive portfolios), but it will
overstate tracking error when this is not the case (as it would for active portfolios). The same applies to the
correlationmeasuresuggestedinAmmannandZimmermann(2001).Wethereforedefinetrackingerrorasthe
square root of the second moment of the deviations between active portfolio returns and benchmark returns.
Alternatively, one can define tracking error as the mean absolute deviation between active portfolio and
benchmark returns, see e.g., Satchell and Hwang (2001). Both definitions can be used for ex ante tracking
error(usingforecastactiveandbenchmarkreturns)aswellasexposttrackingerror(usingrealizedactiveand
benchmarkreturns).
4
solutionswithinatrackingerrorcontext.Thereareafewpapers(e.g.,Clarkeetal.,2002)that
investigate different active strategies with or without constraints on weights and/or risk.
UnlikeRollandJoriontheydonotanalyticallytracethetrade-offbetweenactiverisk-taking
andexpectedexcessreturns.
In this paper we apply Jorions approach to active portfolio management within a
trackingerrorconstrainedenvironment.Weextendthemethodologybytakingacareful(and
practical) approach to compute the input list. We investigate the impact of seriously
inefficient benchmarks (which Jorion excludes) and the introduction of short selling
constraints.Weapplythemethodologytothetop-30stocksoftheAustralianStockExchange
duringathree yearsampleperiodcharacterizedbyastrongbullmarketfollowedbyasharp
and persistent bear market. We find, not surprisingly, that market conditions have a
substantial impact on the active portfolio allocation and its ex post performance. This
becomesevenmoreapparentwhenweallowforshortsellingconstraints.

The next section briefly describes our methodology, by reviewing the well-known
portfolio optimization algebra and the lesser-known tracking error analytical solutions. We
also describe how we operationalize the general portfolio allocation model. Section 3
summarizes the data from the Australian Stock Exchange and illustrates typical
implementation issues that confront portfolio managers. Section 4 discusses the empirical
resultsofourtrackingerroroptimization.Weconcludewithlessonslearntfromthisexercise
andpossiblevenuesforfurtherresearch.

2.Methodology
Our methodology is based on Jorion (2002) to derive a constrained tracking error frontier.
Defineanobservation,orestimation,period[t-j,t]fromwhichwederivetheinputlist.Based
onthisinputlistwefirstcomputetheglobalefficientfrontierwithoutrestrictionsonriskor
weights(exceptfortheusualfullinvestmentconstraint).Wetheninvestigatethereductionin
investment opportunities when we introduce a tracking error constraint, followed by a short
selling constraint. The introduction of a benchmark leads to a tracking error frontier, from
which we derive the (conditionally) optimal active portfolio allocation. We then track this
active portfolios performance over a subsequent tracking period [t+1,t+k] and compute the
realized tracking error. We dynamically update the active portfolio allocation at different
frequencies (daily, weekly, monthly). Each time we update the investment opportunity set,
5
wealsolocatethenewexantepositionofthepreviousperiodsactiveportfoliorelativeto
theupdatedtrackingerrorfrontier.
Computation of the input list (arguably the most important stage of portfolio
management, see Zenti and Pallotta, 2002) tends to be inconsistent in practice. Return
forecasting is a strictly separate exercise from risk forecasting. Stock analysts provide the
portfoliomanagerwithforecastreturns.Thesearetypicallypointestimateswithoutmatching
confidenceintervals,i.e.,stockanalystsdonotgeneratepredictionintervals,seeBlair(2002).
A stock analysts information set typically comprises accounting information, economic
information,managementinformation,etc.Itwouldbeextremelycomplicatedtocombinethe
uncertainty surrounding each information variable into a single confidence interval for the
stock return forecast. That is, the accuracy of the forecast is hard to define and measure.
Econometric forecast models are commonly based on a much smaller information set of
fairlyhomogeneousvariables(likehistoricalreturns,dividends,growthrates).Evenforthese
models it is still a challenge to find the joint confidence interval. In the absence of an
analytical solution, simulation techniques or scenario analysis may be used to generate the
uncertaintymeasure.ThehighestdensityforecastregionproposedinBlascoandSantamaria
(2001) or the bootstrapped prediction densities in Blair (2002) would be more promising
candidates to solve this problem. In the absence of analyst forecasts (as in our application),
the portfolio manager may apply some version of the beta pricing model along the lines of
RosenbergandGuy(1976),andRosenberg(1985).Wedonotusefundamentalinformation
as the portfolio manager would but instead opt for the following simplification of the
BARRAmodeltoforecastbetas:
( )
h t a t f
+ =
, ,
1 (1)
where
a
is the beta computed over the estimation period (effectively derived from the
forecast variance-covariance matrix as discussed below) and
h
is the long-run beta used to
smooth the forecast,
4
with a smoothing parameter . Of course, there are different
techniques to choose the smoothing parameter (e.g., Bayesian updating, Maximum
LikelihoodEstimation)andthehistoricallong-runbeta.Wedonotfocusontheselectionof
or
h
, but do investigate the sensitivity of our results to different values of . We then use
stocki'ssmoothedforecastbeta,
f
,togenerateitsforecastreturn

4
Inourapplication,
h
iscomputedasanexpandingaverage,i.e.,ateachoptimisationperiodwecomputethe
long-runbetaoverthefullsampleperioduptothatdate(alternativelyonecouldusetheaveragebeta
measuredovertheyearprecedingtheobservationperiod).
6
( ) ( ) [ ]
f
t t B t
i
t f
f
t t i t
R R E R R E + =
+ + 1 , , 1 ,
(2)
where the forecast return on the benchmark is its average return over the estimation period.
Ofcourse,ourchoiceofasinglefactor modelisafurthersimplification.GrinoldandKahn
(1999) discuss more appropriate multifactor generalizations (including the BARRA risk
factormodel).Chan,KarceskiandLakonishoks(1999)multifactormodelthoughrestricted
toforecast(co)variancesispotentiallysuitabletogeneratethecompleteinputlist,bothrisk
andreturnforecasts.GiventhatwedonotintendtomaximizetheexpostInformationRatio,
wedonotpursuethosemoreelaboratemodelsatthisstage.
As suggested in Blair (2002) and discussed above, portfolio managers often apply
forecast risk models in complete isolation from forecast return models. They might adopt a
number of modeling specifications. Most of these are based on the premise that a stocks
volatility changes over time (as does its beta). The ARCH model by Engle (1982) and its
many offspring have dominated the academic portfolio literature for the past two decades.
Much progress has been made in achieving ever better fitting specifications for univariate
time series. Multivariate extensions quite crucial for portfolio applications have not
witnessed similar advances, the main obstacle being the dimensionality problem. A
completely unrestricted time-varying variance-covariance matrix is almost impossible to
estimateforrealisticportfoliodimensions(letaloneachievingconvergenceinarealistictime
frame)
5
.Pragmaticsolutionsarethereforeneededandweoptfortheapproachadvocatedby
RiskMetrics (1995). Forecast volatility is an exponentially weighted moving average
(EWMA)ofpastsquaredreturnsforstocki:
( )

=
0
2
,
2
,
1
j
j t i
j
t i
R (3)
where , the decay factor, depends on how fast the mean level of squared returns changes
over time. The more persistent (autocorrelated) the squared returns will be, the closer
should be to 1. For highly persistent time series of financial returns, we find values of
between 0.9 and 1. Given the choice of decay factor, we then determine how many past
observations should be used to compute forecast volatility. For a tolerance level of 0.01
(when we consider the observations impact on forecast volatility to have sufficiently
decayed), and =0.9, we need about 40 historical observations. As in RiskMetrics , we
apply(3)toeachelementofthevariance-covariancematrix.Theoretically,withauniverseof
n stocks we should have as many as n+(n)x(n-1)/2 unique s and matching estimation
7
periods.Inourapplicationthisimplies465individualvarianceforecastprocesses.Whilestill
computationally feasible (and certainly faster than a multivariate GARCH estimation of the
samedimension),forourpurposesweimposeasingleonallvariance-covarianceelements.
Thepenaltyforthischoiceisreducedprecisionintheforecastvariance-covariancematrix.It
would be worthwhile to furtherinvestigate whether there is a diversification effect in this
additionalforecasterroracrossstocks,whichwouldeffectivelyreducethispenalty.
Which brings us to the next step, portfolio optimization. Rudd and Rosenberg (1980)
comprehensivelyderivetheinvestmentallocationproblemforarestrictedportfoliouniverse
(that is, for example, only the top 50 liquid stocks on a particular exchange) to match
empirical practice. Jorions (2002) derivation is similar in style and we follow his notation.
First,thestandardconstrainedportfoliovarianceminimizationproblem

=
=

E w
w t s
w w
P
p
P P
w
1 . .
min
(4)
where E is a vector of forecast returns and is the forecast variance-covariance matrix of
returnsandisatargetportfolioreturn.Optimizationoverportfolioweightsw
P
leadstothe
wellknownhyperbolainmeanstandarddeviationspace:
C
B
E E D
C
E B
C C
B
D
P P
2
1
1
1
2
2
1 1
=
=
=
+
(

(5)
The weights of the efficient portfolios that fall on the hyperbola are readily obtained for
different values of the target return constraint. Easy to apply, but of course, as soon as
additional constraints are added on (e.g., a short selling constraint), we lose this analytical
resultandhavetonumericallyoptimizetofindthefeasibleinvestmentopportunityset.Now,
ifweslightlyrefocustheoptimizationproblemtoreflecta search forportfoliovalueadded
(inexcessofabenchmarkportfoliovalue),weget

=

P P
P
P
a
a a
a t s
E a
0 . .
max
(6)

5
ArecentexceptionisTimmermannandBlake(2002)buttheirportfoliodimensionsaresmall.
8
where the active weights a
P
(in deviation from the benchmark weights, w
B
) have to add to
zero to satisfy our original full investment constraint. This guarantees that total active
portfolioweights,w
B
+a
P
,stilladduptoone.Activerisk(indeviationfrombenchmarkrisk)
maynotexceedtrackingerrortargetvariance.Jorionderivestheactiveportfoliosolutionif
wesetthetrackingerrorconstraintexactlyequaltothetargetexcessvarianceas
(

C
B
E
D
a
P
1
(7)
whereB,CandDareasdefinedin(4).AsJorionnotes,thesearesolutionsinexcessreturn
versusexcessriskspace.Foreaseofcomparisonwiththestandardportfoliosetup,wewould
rather have a representation in the mean (total return) versus standard deviation (total risk)
space:
( ) ( )
2
0 . .
max
P P B P B
P P
P
P
a
a w a w
a a
a t s
E a

= +

(6)
whichuponmaximizationresultsinthefollowingellipsoidalsolutions(
P
,
P
)
( ) ( ) ( )( )
0
1
4
4
1
4
2
2
2 2 2 2
2
2 2
=
(
(

|
.
|

\
|
|
.
|

\
|

|
.
|

\
|
|
.
|

\
|
+
C
B
C
D
C
B
C
D
B B
B P B P B B P B B P


(8)
as long as the benchmark (
B
,
B
) lies within the efficient set. The ellipse is vertically
centred around the benchmark expected return, but it is horizontally centred around the
benchmarkvarianceplustrackingerrorvariance.Theellipses principalaxis ishorizontalif
benchmark expected return coincides with the expected return of the global minimum
variance portfolio (
MVP
=B/C). If
B
>
MVP
typical in a bullish market it will have a
positive slope; if
B
<
MVP
in a bearish market it will have a negative slope. Jorion
providesanextensivediscussionofdisplacementsofthetrackingerrorellipseforchangesin
targettrackingerrorvariance.Forourpurposeswejustmentiontwomorerelevantmetrics:
|
.
|

\
|
|
.
|

\
|
+ + =
+ =
D C
B
D
B B MAX
B MAX


2
2 2
(9)
9
for the mean and variance of the maximum expected excess return portfolio. Note that the
portfoliomanagercanonlygenerateexcessreturnsbyassumingtrackingerror.Notealsothat
thepenaltyforthisactiveportfolioallocationnotonlyincreaseswithtrackingerrorvariance
butalsowitha moreefficientbenchmarkportfolio(anincreasinggapbetween
B
andB/C).
Themoreefficientthebenchmark,theharderitwillbetobeatitsperformance.
WetesttheJorionsmethodologyonrealdata.Twoempiricalfeaturesstandoutforour
application. The benchmark is frequently so inefficient, that its expected return falls below
theexpectedreturnoftheglobalminimumvarianceportfolio(B/C).Also,theunconstrained
weights take completely unrealistic values during bear market conditions with huge short
sellingimplications.Wethereforeaddashortsellingconstraint.Thiscomplicatesmatters,as
Jorionsuggests.Thenumericaloptimizationof(4)withsuchaconstraintposesnoparticular
problems. Unfortunately, we cannot simply distort the tracking error ellipse by taking the
intersectionofthetrackingerrorellipseandtheconstrainedmean-varianceefficientfrontier.
We have to perform an integrated numerical optimization of (6) including the short selling
constraint. The solution set then becomes considerably thinner. Ultimately we look for the
maximum excess return active portfolio (a single point) that satisfies all constraints
simultaneously. There is no analytic solution for this problem. Numerical optimization is,
however,quitefeasible.

3.DataIssues
Ourapplicationconsidersaportfolioof30AustralianstockswithareducedAustralianAll
Ordinaries Index (XAO) as our benchmark. The stock price data and risk-free interest rates
areobtainedfromDatastreamandIRESSMarketTechnology.Thetop-30stocksaccountfor
about 62% of the XAO index
6
. We standardize the top-30 weights to add up to 100% and
generate a new top-30 benchmark accordingly. We choose a sample that covers the rather
turbulent period from 4 January 1999 until 29 November 2002, a total of 991 trading days.
Whatmakesthissampleparticularlyappealingisthestrongbullmarketfromthestartofour
sample until April 2000, followed by a sequence of collapses and persistent bear market
conditionsuntiltheendofoursample.
Figure1indicatesthatourtop-30benchmarktrackstheXAOindexquitecloselyforthe
first year and a half of our sample. Mid 2000, however, the top-30 starts to diverge

6
Thetop-10stocksalreadyaccountforover40%oftheXAO.Forcomparison,thetop-10stocksintheS&P500
accountforabout25%.Thisdominanceofafewstocksshouldmakeittheoreticallyeasytokeepthetracking
10
substantially from the XAO index. The underperformance of the XAO index is easily
explainedbytheunderrepresentationoftechnologystocksinthetop-30,relativetotheXAO.
This is even more apparent when we include the S&P500 in the comparison. Whereas the
S&P500outperformsbothtop-30andXAOindicesuntilaboutJune2000,ithasbeenmuch
moreexposedtothetechstockscollapse.Acloserlookatthekey marketdropsinFigure1
highlightsthisphenomenon.

Figure1.AllOrdinariesversusTop-30BenchmarkIndexandS&P500
60
85
110
135
160
1
/
1
/
1
9
9
9
5
/
1
/
1
9
9
9
9
/
1
/
1
9
9
9
1
/
1
/
2
0
0
0
5
/
1
/
2
0
0
0
9
/
1
/
2
0
0
0
1
/
1
/
2
0
0
1
5
/
1
/
2
0
0
1
9
/
1
/
2
0
0
1
1
/
1
/
2
0
0
2
5
/
1
/
2
0
0
2
9
/
1
/
2
0
0
2
S&P500
Benchmark
XAO

Note:Indiceshavebeenstandardizedat100onthe1
st
ofJanuary1999.ThebenchmarkconsistsoftheTop-30
stocksintheAllOrdinaries.

RankedbythesizeoftheS&P500collapsetheyare:theNasdaqcollapseonthe14
th
of
April 2000, the 17
th
of September 2001 (following September 11), the 12
th
of March 2001,
the3
rd
ofSeptember2002,themillenniumbugonthe4
th
ofJanuary2000andthe19
th
ofJuly
2002.TheinitialNasdaqcollapseismirroredintheXAOcollapseonthefollowingtrading
day,buttoalesserextentforthetop-30.Subsequenttechstockspilloversarelessapparent
indicating the smaller presence of tech stocks in the Australian market. These US-led
collapseshaveamoderateimpactontheXAOindex,buthavevirtuallynoimpactonthetop-
30 index. Market wide collapses (like September 11, and the millennium bug), on the other
hand, are similarly reflected in the XAO and the top-30 index. The lesser exposure to
extremal price changes of the top-30 index does not imply less volatility. The standard
deviationofreturnsisalmostidenticalforbothAustralianindices(13%perannum),butboth
aresubstantiallylessvolatilethanS&P500returns(astandarddeviationof22%perannum).

error within acceptable bounds. Our ex post tracking error results show that this is true with a short-selling
11
Figure2stressestherelevanceoftime-variationintherisk-freerateofreturn(the90-day
bill rate). It suggests that the bull market returns were somewhat tempered by a rising risk-
freerateofreturn.Similarlybearmarketnegativereturnswereoffsetbyadeclineintherisk-
free rate of return (at least after October 2000). This offset lasts until April 2002, when the
risk-free rate again starts to increase and risk-taking performance was doubly penalized
(squeezingtheexcessreturns).

Figure2.Risk-FreeRateofReturn
4%
5%
6%
7%
1
/
4
/
1
9
9
9
5
/
4
/
1
9
9
9
9
/
4
/
1
9
9
9
1
/
4
/
2
0
0
0
5
/
4
/
2
0
0
0
9
/
4
/
2
0
0
0
1
/
4
/
2
0
0
1
5
/
4
/
2
0
0
1
9
/
4
/
2
0
0
1
1
/
4
/
2
0
0
2
5
/
4
/
2
0
0
2
9
/
4
/
2
0
0
2
90DayBill%

Table1summarizesthefullsampleinputlist.Thetop-30stocksarelistedbydescending
top-30 weight. Annualised mean returns and annualised standard deviations are given in
columns3and4,respectively.Volatilityvariesbetweenalowof15%(WFT)andahighof
53%(QBE),butisgenerallyaround30%perannum.Riskcompensationisrathermeagreata
low of 20% (LLC) mean return to a high of 19% (WES) mean return. Our benchmark
comparesreasonablywellwithameanreturnof5%pa.(medianreturnof11%pa.)againsta
standard deviation of 13% pa. Taking into account that the risk-free rate of return varied
between4%and7%pa.,thisdoesnotsuggestagenerousexcessreturn.
Table 1 also gives daily (not annualised) maximum return and minimum return in
columns 5 and 6, respectively. The maxima vary between 4% and 25%, while the minima
vary between 4% and 35%, truly extremal values. Not surprisingly, the empirical
distributions of daily returns are hugely kurtotic (fat-tailed). This causes problems when we
wanttooperationalisetheinputlistforportfoliooptimisationpurposes.Thenon-normalityof
stock returns suggests that mean-variance optimization might not reflect investors utility
tradeoffbetweenriskandreturn.

constraint,butisemphaticallynottruewhenshortsellingispossible.
12
Table1.DescriptiveStatisticsforTop30ASXStocks
Code Top-30
weight
Annualized
Average
return
Annualized
Standard
deviation
Daily
Maximum
return
Daily
Minimum
return
Beta

Ljung-Box
(p=12)
Test
NAB 12.60 7.31 24.10 5.00 -13.87 1.18 25.87
CBA 9.37 4.06 20.80 3.97 -7.13 0.95 12.27
BHP 8.94 16.43 29.93 6.87 -7.62 1.21 16.72
TLS 7.75 -13.09 25.67 8.34 -9.86 0.91 20.49
WBC 6.96 6.75 21.12 4.87 -5.21 0.92 11.00
ANZ 6.81 14.42 21.99 4.71 -5.60 0.99 22.07
NCP 4.85 4.18 46.49 24.57 -14.89 1.95 9.18
AMP 4.14 -11.55 26.61 7.65 -8.03 1.03 17.16
RIO 4.14 14.88 30.02 7.18 -5.78 1.13 16.19
WOW 3.11 18.17 23.41 5.16 -8.34 0.53 28.93
WSF 2.38 13.38 29.62 20.92 -6.55 0.86 16.35
WES 2.31 19.03 26.78 10.58 -6.42 0.71 17.49
FGL 2.29 1.35 21.95 4.83 -7.06 0.49 36.79
WMC 2.28 13.41 34.32 16.32 -10.40 0.97 46.10
SGB 2.18 14.57 18.93 7.17 -5.02 0.53 6.92
WPL 2.05 12.55 27.02 7.57 -11.13 0.62 13.06
BIL 1.92 -19.41 33.86 9.54 -35.25 0.83 25.12
CML 1.90 -6.72 27.77 12.83 -18.18 0.61 10.62
WFT 1.67 -1.01 14.98 3.64 -4.38 0.38 20.28
CSR 1.54 11.46 27.07 9.16 -6.37 0.74 9.91
PBL 1.47 3.51 29.22 10.41 -10.83 0.82 21.62
GPT 1.31 -1.55 16.60 4.55 -3.98 0.42 19.58
CCL 1.20 -1.16 35.65 12.42 -10.80 0.69 13.43
MBL 1.20 10.35 26.65 7.11 -11.46 0.91 37.72
TEL 1.13 -13.72 29.44 9.63 -9.12 0.77 13.88
CSL 1.13 10.84 39.41 26.76 -12.01 0.76 22.29
LLC 1.05 -20.15 29.68 5.32 -17.11 0.72 36.05
QBE 0.98 4.67 52.83 41.85 -52.62 1.25 131.26
AGL 0.92 -3.61 22.54 4.42 -8.48 0.38 33.24
AXA 0.79 -4.04 29.61 7.89 -6.12 0.78 17.97
Index 100.00 4.82 13.28 2.65 -5.14 1.00 11.83
Note: Columns 2-6 are in percentages. The top-30 weights are fixed as of November 2002 and are the
standardized(they sumto100%) XAO-weightsonthatdate. Weignore anyweight changesduringour
sample period. Betas are computed against the top-30 index. The Ljung-Box column gives the test
statisticforserialcorrelationinthereturnsupto12
th
orderlaglength,witha95%criticalvalueof21.03.

13
Campbelletal.(2001)illustrateanalternativeoptimisationprocedurethatbettercaptures
thekurtotic(andperhapsskewed)natureofstockreturns.Asamatterforfutureresearch,we
couldincorporateanequivalent trackingerrorconstraint intheiroptimisationprocedure.As
longastheempiricaldistributionsaresymmetric,weexpectverylittledifferenceintermsof
activeweightselection.
Unlike stock analysts, we do not have the fundamentals (growth forecasts, accounting
information,etc.)tovalueandthenrankstocksbyforecastreturntogeneratebuy/sellsignals.
Our information set is restricted to historical returns. For simplicity, consider the following
example whereourforecastreturnisasimpleaverageofthepast5tradingdaysreturns.If
wehappentoencounterasingleextremalreturninourinformationsample(say,10.58%)and
four zero returns, we generate an annualised forecast return of 535%. Clearly, the extremal
returnisnon-representativeforforecastpurposes.Evenforlongerinformationsamples(saya
month, or 20 trading days) this inflation of annualised forecast returns based on extremal
observationsremainsaproblem.Itreflectsthefactthatdailystockreturndistributionsarenot
normallydistributed,butarebettercharacterizedbysomefat-taileddistribution(like,e.g.,a
Student-t).Unlikethenormaldistribution,aStudent-tisnot closedunderaddition, i.e.,its
properties (e.g., the variance) cannot be simply scaled to derive equivalents at a different
samplingfrequency(sayfromdailytoannualised).
This problem highlights the difficulties that one encounters when optimising portfolios
of individual stocks based on simplistic input list rules. It might explain why the academic
literaturepreferstobuildefficientportfoliosfromportfoliosofindividualstocks
7
.Indexing
clearly normalizes the empirical distributions (just consider the descriptive statistics for the
top-30 index), which makes these portfolios much more suitable for portfolio optimisation
purposes. This may be feasible for academic purposes, but it will not typically be a
satisfactory solution for practical purposes. To somehow moderate the impact of extremal
returns on our input list, we therefore choose to forecast returns based on a smoothed beta
modelasexplainedintheprevioussection.
Ofcourse,thevalidityofusingpastaveragereturnstoforecastfuturereturns(eitherasa
simple average or as a market expectation) depends crucially on the stationarity of stock
returns.TheLjung-Boxportmanteaustatistic(autocorrelationuptolaglength12)incolumn
8, Table 1, indicates that quite a few series display significant autocorrelation (95% critical

7
Theauthorsarewellawareofmoreimportantreasons(liketheerrors-in-variablescorrection)forthischoice.
14
value
2
12
= 21.03). As suggested by Lawton-Browne (2001), this may lead to a downward
biasinexantetrackingerror.Weinvestigatethisbelow.

4.EmpiricalResults
Tostartouranalysis,wefirstcomputeourinputlist.Toupdatethevectorofforecastreturns,
E,weuseasimplified version oftheBARRAbetapricing modelencapsulatedin equations
(1) and (2). There is no real precedent for this procedure, and it obviously lends itself for
future improvement. We choose =0.34 for our application, but also investigate the
sensitivity of our results for different values of . This procedure generates fairly smoothly
evolvingforecastreturns.
To update the forecast variance-covariance matrix, we use the RiskMetrics
TM
EWMA
methodology. It is simple to understand, straightforward to implement, and generates
GARCH-likevarianceprocesses.Toillustratethispoint,considertheGARCH(1,1)outputin
Figure3forNAB.TheGARCH(1,1)parameterswereestimatedat
1
=0.17,
2
=0.74,while
the average decay factor in the EWMA was estimated to be =0.91 (with an effective
estimationsamplelengthof40days).TheEWMAprocessissomewhatsmoother,butthere
seemslittletoseparatethetwoprocessesasimilarpointismadeinRiskMetrics (1995).

Figure3.NABConditionalVolatilityGARCH(1,1)versusEWMA
0
0.25
0.5
0.75
1
/
4
/
1
9
9
9
5
/
4
/
1
9
9
9
9
/
4
/
1
9
9
9
1
/
4
/
2
0
0
0
5
/
4
/
2
0
0
0
9
/
4
/
2
0
0
0
1
/
4
/
2
0
0
1
5
/
4
/
2
0
0
1
9
/
4
/
2
0
0
1
1
/
4
/
2
0
0
2
5
/
4
/
2
0
0
2
9
/
4
/
2
0
0
2

Note: Annualized standard deviations for NAB are based on fitting a GARCH(1,1) model:
2
1 , 2
2
1 , 1 0
2

+ + =
t i t i it
R ,respectivelyanEWMAmodel:
( )

=
0
2
,
2
1
j
j t i
j
it
R
.

Whereas it is relatively straightforward to estimate a univariate GARCH(1,1) process

liketheoneabove,thecomputationalburdenbecomesexcessiveforamultivariateGARCH
process involving 30 stocks. Scowcroft and Sefton (2001) investigate the performance of a
EWMA
GARCH(1,1)
15
number of time-varying risk matrix specifications including the EWMA and GARCH
modelsandfindthatthetrackingerrorpredictionsagreedreasonablywell.
Figure 4 gives some insight in the dynamically updated input list for three stock
components in the top-30 benchmark: NAB (the largest weight, 12.6% and a full-sample
unconditionalbetaof1.18),NCP(thehighestfull-sampleunconditionalbeta,1.95)andAGL
(the lowest full-sample unconditional beta, 0.38). Figure 4 shows the time variation in their
betas,accordingtoequation(1).

Figure4.TimeVariationinBetas
0
0.5
1
1.5
2
2.5
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2

Note: Unconditional betas are respectively 1.18 (NAB), 1.95 (NCP), and 0.38 (AGL); Betas are measured
againstthetop-30benchmarkindex.

NABhasthemorestablebeta,whereasNCPandAGL(thestockswithmoreextremehigh
and low betas) have much more volatile intertemporal betas. This has obvious
repercussionsfortheforecastreturnsaccordingtoequation(2),wherethemorevolatilebeta
forecastswillgeneratemorevolatilestockreturnforecasts.

Figure5.TimeVariationinCorrelation

-1
-0.5
0
0.5
1
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2
-1
-0.5
0
0.5
1
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2

Note:NAB-WBCunconditionalcorrelation=0.46;NCP-AGLunconditionalcorrelation=0.06;

AGL
NCP
NAB
NAB-WBC
NCP-AGL
16
Figure 5 illustrates the time-varying intra-sector correlation between NAB and WBC,
two banking stocks in our universe; respectively between NCP and AGL for an inter-sector
illustration.Intra-andinter-sectortime-varyingcorrelationareequallyvolatile,butobviously
haveadifferentmean.Theintra-sectorcorrelationdisplaysswitchingbehaviourwithperiods
ofhighcorrelationalternatingwithperiodsoflowcorrelationandnotmuchinbetween.The
inter-sectorcorrelationdoesnotsharethisfeature.
Havingcompletedtheinputlist,wecanproceedtocomputetheefficientfrontiersolving
(4)with=0.91.Thenwesolve(6)toobtaintheactivetrackingerrorfrontierwithatracking
errortargetof5%.Ourtop-30benchmarkisfoundtobeseriouslyinefficient(throughoutour
sample period with a few exceptions when it is close to the global efficient frontier) which
suggestsactiveinvestmentopportunities.Or,inexanteInformationRatioterms,theyoffera
positive IR for active portfolio managers. To illustrate this, consider the following two
representative optimization periods. Figures 6a and 6b are representative for a bull market
episode, respectively a bear market episode. As expected in a bull (bear) market the active
investmentopportunitysettheellipseisupward(downward)sloping.Thebenchmark(the
squaresymbol)typicallyhaslowerriskthantheindividualstocks(thediamondsymbols)but
ofcourse,onlyaverageexpectedreturns.

Figure6.BullsandBears
-0.2
0
0.2
0.4
0.6
0.8
1
0 0.1 0.2 0.3 0.4 0.5
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0 0.1 0.2 0.3 0.4

EfficientFrontier ConstrainedActivePortfolio
ConstantTrackingErrorFrontier UnconstrainedActivePortfolio
NoShortSellingEfficientFrontier BenchmarkPortfolio

The location of the active portfolios chosen in the previous period optimisation is also
indicated in both graphs. Without rebalancing, the unconstrained active portfolio (circle
symbol outside ellipse) needs no longer be on the updated ellipse. As it turns out, the
constrainedactiveportfolio(circlesymbolinsideellipse)isalwaysinsidetheupdatedellipse,
17
but the unconstrained active portfolio is almost without exception outside the updated
tracking error frontier. That does not necessarily imply a violation of the tracking error
constraint over that period, but does imply an ex ante tracking error violation for the
subsequentperiod.
We summarize our dynamic optimisation exercise in Figure 7, which gives the active
portfolioweightsforthreestocks(NAB,NCP,AGL)inthe(un)constrainedactiveportfolios.
Perhaps surprisingly the unconstrained weights the black lines are as volatile (in both
positive and negative direction) during the bull market as they are during the bear market.
There is limited evidence of short-lived persistence in the active weights, suggesting that
portfolios need to be rebalanced fairly frequently (at least monthly). The short selling
constrained weights the grey lines are obviously much smoother which implies
substantiallylessrebalancingcosts.

Figure7.ActiveportfolioweightsforNAB,NCPandAGL
-1.0
-0.5
0.0
0.5
1.0
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2
-1.0
-0.5
0.0
0.5
1.0
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2

-1.0
-0.5
0.0
0.5
1.0
3
/
3
/
1
9
9
9
6
/
3
/
1
9
9
9
9
/
3
/
1
9
9
9
1
2
/
3
/
1
9
9
9
3
/
3
/
2
0
0
0
6
/
3
/
2
0
0
0
9
/
3
/
2
0
0
0
1
2
/
3
/
2
0
0
0
3
/
3
/
2
0
0
1
6
/
3
/
2
0
0
1
9
/
3
/
2
0
0
1
1
2
/
3
/
2
0
0
1
3
/
3
/
2
0
0
2
6
/
3
/
2
0
0
2
9
/
3
/
2
0
0
2

Note: Weights are given as fractions. Weights in the top-30 benchmark are respectively 0.126 (NAB), 0.049
(NCP),and0.09(AGL).Greylinesgivenoshortsellingweights;Blacklinesgiveunconstrainedweights.

A comparison across stocks is interesting. The stock with the highest top-30 benchmark
weightandamoderatebeta(NAB),switchesmostfrequentlyfrompositivetonegativeactive
positions. The stock with the lowest weight in the top-30 and the lowest beta (AGL), has
predominantly positive active weights (although with substantial volatility). The stock with
the highest beta (NCP) not surprisingly the least Australian of the 30 stocks has the
NAB NCP
AGL
18
most stable active weights (although they switch frequently from positive to negative and
viceversawithoutmuchpersistence).
Sincewerebalanceeverytimeperiod,ourexantetrackingerrorisalwaysexactlyequal
tothetarget.Thisisclearlynotthecaseexpost.Theproblemnowishowtomeasuretheex
post tracking error. We choose to measure ex post performance using a quadratic tracking
errormeasure,
8
thatisthesquareddeviationofportfolioreturnsfrombenchmarkreturns.We
first measure the ex post difference between portfolio return and benchmark return, i.e., the
activereturn.Thisactive/excessreturnperformanceisillustratedinFigure8.A comparison
suggests that the short selling constrained active portfolio only rarely wanders away from
the benchmark (Figures 6a and 6b are fairly typical of this phenomenon). In fact, the
constrainedportfolioseemsto hug thebenchmarkoverthetrackingperiod.

Figure8.ExPostActiveReturns
0
10
20
30
40
50
60
70
80
-
0
.
1
3
5
-
0
.
1
1
8
-
0
.
1
0
2
-
0
.
0
8
5
-
0
.
0
6
8
-
0
.
0
5
2
-
0
.
0
3
5
-
0
.
0
1
9
-
0
.
0
0
2
0
.
0
1
4
0
.
0
3
1
0
.
0
4
8
0
.
0
6
4
0
.
0
8
1
0
.
0
9
7
0
10
20
30
40
50
60
70
80
-
0
.
0
2
0
-
0
.
0
1
8
-
0
.
0
1
5
-
0
.
0
1
3
-
0
.
0
1
1
-
0
.
0
0
8
-
0
.
0
0
6
-
0
.
0
0
3
-
0
.
0
0
1
0
.
0
0
1
0
.
0
0
4
0
.
0
0
6
0
.
0
0
9
0
.
0
1
1
0
.
0
1
3

Note: the histograms have frequency count on the vertical axis and daily (fractional) excess returns on the
horizontalaxis.

The standard deviation of short selling constrained excess returns is almost exactly 5%
per annum. Not entirely surprising given the visual evidence in Figure 8, the unconstrained
standarddeviationis20%perannum.Aniceillustrationindeedofthetheoreticalresultofthe
downwardbiasinexantetrackingerrorasaforecastofexposttracking error,Satchelland
Hwang(2001).FurtherempiricalevidenceforthisbiasisgiveninZentiandPallotta(2002).
Among the possible reasons for this bias, Lawton-Browne (2001) and Scowcroft and
Sefton (2001) suggest autocorrelation in returns and volatility clustering. We therefore
investigatetheimpactofbothviolationsofthemean-varianceoptimizationassumptions.We
alsolookatthesensitivityofourresultstovariationsintrackingduration,andtrackingerror
target.TheresultsoftheseexercisesaregiveninTables2a,2b,and2c.

8
Rudolfetal.(1999)suggestthatameanabsolutedeviationmeasurewouldbepreferablesinceitmatchesfund
managerscompensationschedules.Weagreethatthisisworthinvestigating.
unconstrained noshortsales
19
Table2a.ExPostTrackingErrorVaryingtrackingduration
Tracking
Duration(days)
Constrained
TrackingError
Unconstrained
TrackingError
1 5.27% 19.67%
2 5.27% 19.93%
3 5.02% 19.68%
4 5.23% 19.77%
5 4.93% 20.54%
10 4.87% 19.76%
20 5.11% 19.85%
40 5.04% 19.79%
60 4.77% 18.20%
125 5.28% 22.25%
Note:trackingerrortarget,=0.05;volatilitypersistenceparameter=0.91.

Table2b.ExPostTrackingErrorVaryingtrackingerrortarget
TrackingError
Target(st.dev)
Constrained
TrackingError
Unconstrained
TrackingError
1% 1.08% 3.97%
2.5% 2.59% 9.93%
5% 5.11% 19.85%
10% 7.57% 39.70%
20% 12.52% 79.41%
Note:trackingduration=20days;volatilitypersistenceparameter=0.91.

Table2c.ExPostTrackingErrorVaryingvolatilitydecayfactor
EWMAdecay
Factor()
Constrained
TrackingError
Unconstrained
TrackingError
0.85 4.66% 50.84%
0.87 4.84% 36.40%
0.90 4.67% 15.25%
0.91 5.11% 19.85%
0.95 4.69% 12.62%
0.97 4.32% 11.66%
0.99 4.67% 15.25%
Note:trackingduration=20days;trackingerrortarget,=0.05.

Table2agivesexposttrackingerrors(asastandarddeviationofactivereturns)forarangeof
tracking durations (k=1, ,125 days). Apparently, the duration of tracking does not
materiallyaffecttherealizedtrackingerror.Perhapssurprisingly,thereisverylittlevariation
in ex post tracking error when rebalancing of the active portfolio occurs more or less
frequently. Short selling constrained active portfolio management generates tracking errors
20
that stay within the target tracking error. Unconstrained active portfolio management
generatestrackingerrorswellinexcessofthetarget.
Table 2b gives ex post tracking error for a range of tracking error targets (=1%,
,20%). We observe that the ex post short selling constrained tracking error marginally
exceeds(issubstantiallylessthan)theexantetrackingerrorfortargetsbelow(above)5%per
annum.Expostunconstrainedtrackingerror,however,generallyexceedstheexantetracking
erroratanincreasingratewithincreasingtargettrackingerror.
Table2csuggeststhesourceofthebiasinexantetrackingerrorexpectations.Itgivesex
post tracking errors for a range of EWMA decay factors (=0.85, ,0.99). For this
(admittedly limitedinscope)experiment, wefindthatat =0.97,thebiasis minimized. We
suspect that individual optimisation of the parameter for each and every element in the
variance-covariancematrixwouldallowanevenbetteroutcome.
We also experimented with the serial correlation in stock returns by varying the
smoothingparameterin(1).Thisdidnotaffectthebiasinexantetrackingerror.Itdoes,of
course,affectthelocationoftheexcessreturndistribution,butnotthescale!
Bringing the two results (time-variation in weights and ex post performance) together
suggeststhatthetypicalshortsellingconstraintactsasasafeguardonexposttrackingerror
while simultaneously reducing the cost of rebalancing. A shortselling constraint effectively
turns active portfolio management into something very close to passive portfolio
management. So why do we not observe the matching reduction in active returns? Simply
becausewe madenorealattempttoactivelyforecaststockreturns.Infact, wehavetakena
ratherpassiveapproachwhengeneratingtheinputlist
9
.Wecanimaginethatstockanalysts
(orsophisticatedeconometriciansmodelsforthatmatter)areabletomanipulatetheseactive
performance distributions to their benefit and hence change the location of the active return
distributionsinFigure8accordingly.

5.Conclusions
Althoughstraightforwardincontent,properimplementationofportfoliooptimizationtheory
can be notoriously complicated. Choices have to be made with regard to the input list,
constraints,estimationprocedure,andimpliedactions.Thispapergivesaflavourofsomeof
themanyissuesthathavetobedealtwithbyportfoliomanagers.Themainadvantageofour

9
We use a mechanistic rule as in equation (2) to generate expected returns. Our choice of the smoothing
parameterisnotbasedonpropermodelselectioncriteria,sincethiswasnotthemainfocusofourpaper.
21
procedureisitstransparentnaturewhichconsiderablyfacilitatescommunicationwithanever
moresophisticatedclientele.
Jorions (2002) main contribution is the visualization of the active investment
opportunityspace. Weillustratethattheintroductionofashortsellingconstrainteliminates
most of this opportunity set partly driven by persistent bear market conditions during our
sampleperiod.However,asaninvestmentadvicetool,successiveintroductionofinvestment
constraints clearly identifies the location of (and the reduction in) the relevant investment
opportunity set. The methodology also highlights the tradeoff between risk penalty and
excessreturngainwhenviolatingtheinvestmentconstraints.Fromthisperspective,Jorions
methodologyisaninvaluablecontributiontothepracticalinvestmentliterature.
Wefind(asdomanyothers,seee.g.,PlaxcoandArnott,2002)thatfrequentrebalancing
isanabsolutenecessitytokeepsomecontrolovertotalrisk(thoughnotnecessarilytracking
errorrisk)whenactivelymanagingportfolios.LarsenandResnick(2001)considerarangeof
optimization and holding periods, but do not consider transaction cost constraints. Clearly,
thecostsofrebalancinghavetobeoffsetagainstthegainsinriskcontrol,butitseemstous
thatcertain(threshold)levelsofriskwillsimplybeunacceptable.Theissue,ofcourse,ishow
tooptimallyrebalancesoastominimizethecontrolcosts.
Not surprisingly, we also find that ex ante tracking error expectations do not match ex
post realizations, see also Rohweder (1998). Satchell and Hwang (2001) show that we can
reasonably expect a worse ex post tracking error outcome due to the stochastic nature of
portfolioweights. Theyreportthatthisupwardbiasisnotrestrictedtoactiveportfolios,but
can also be found in passive portfolios where the weights are not stochastic (due to
rebalancing).Asimilarupwardbiasinexposttrackingerror(butduetoadifferentsource)is
causedbytheapparentserialcorrelation,notjustintheunderlyingstockreturns,butalsoin
theexcessreturnsandsquaredexcessreturns.FrinoandGallagher(2001),e.g.,findevidence
of seasonality in tracking error (partly driven by seasonality in dividend payments on the
benchmark).PopeandYadavs (1994)resultsindicatethat this will leadtoabiasedexante
estimateoftrackingerror.
Where to from here? There is plenty of scope to improve the selection of optimal
observation period and forecast period duration. Ultimately, this is an empirical matter. We
hinted at the possibility (and Table 2b underlines its importance) to individualize the
stochastic processes for every element of the variance-covariance matrix. A trade-off will
then have to be made between the improvement in goodness-of-fit and the increased
22
computationalburdenofsuchanexercise.Despitethis,wearguethatthereisnourgentneed
toresorttocomputationallyburdensomemultivariateGARCHspecifications.
Anotherconstraintworthinvestigatingisacaponthenumberofstocksinthe managed
portfolioortheminimumnumberofstocksinanactiveportfolio.JansenandvanDijk(2002)
illustrate the small portfolio constraint for a passive tracking portfolio. Ammann and
Zimmermann (2001) investigate admissible active weight ranges, which would guarantee a
limitonindividualstockweights.Alternatively,wecouldinvestigateacaponthenumberof
stocksinwhichtheportfoliomanagertakesactivepositions,whiletakingneutralpositionsin
the remaining benchmark component stocks. Yet another approach could be a factor-
neutrality constraint as in Clarke et al. (2002), which would fit typical style-type portfolio
constraints. It is quite possible that some of these constraints are internally conflicting. The
long-only constraint, e.g., tends to favour small capitalization active stock weights, which
wouldobviouslyclashwithalargecapitalizationstyleconstraint.
Another issue is the composition of the benchmark. Fund managers can frequently
choosetheirbenchmarks(withinreasonableboundaries,i.e.,amongapeergroup).Smallcap
fund managers would typically choose a representative small cap benchmark, like the ASX
SmallOrds.AsshowninLarsenandResnick(1998),themarketcapitalizationofcomponent
stocks in the benchmark index has a non-trivial impact on the tracking performance of
enhanced benchmark portfolios. Though their analysis quantifies the impact, they are not
explicit on the source. It could be a liquidity constraint, or perhaps the (related) excessive
non-normalityofthereturnsofthesestocks.
Afewpapershaverecentlyfocusedontrackingerrormeasurementthatbetterreflectsthe
incentive structure of the portfolio manager, see e.g., Kritzman (1987) and Rudolf et al.
(1999). Roll (1992) suggests that diversification of an investors portfolio across managers
reducestheimpactofexcessivelyriskyactiveportfolios.Jorion(2002)showsthatthisisnot
asatisfactorysolutionandinsteadfavoursadditionalconstraintsontotalrisktobettercontrol
forthefreeoptionprovidedtoportfoliomanagerswhoareonlyconstrainedbyactiverisk(or
tracking error). A closer look at investors utility functions and better integration of these
utility functions with constrained portfolio optimizations seems a worthwhile extension of
thispaper.
Perhapsmostimportantly,thereneedstobemoreresearchtowardsproperintegrationof
return and risk forecasts. Though it seems obvious that stock analysts do make risk
assessmentswhencomputingforecastreturns,itismuchlessobvioustoextractandcombine
theseriskassessmentsintoajustifiableriskmeasure.
23
References
Ammann, M. and H. Zimmermann, 2001. Tracking error and tactical asset allocation.
FinancialAnalystsJournal57,32-43.
Blair, B., 2002. Conditional asset allocation using prediction intervals to produce allocation
decisions.JournalofAssetManagement2,325-335.
Blasco, N., and R. Santamaria, 2001. Highest-density forecast regions: an essay in the
Spanishstockmarket.JournalofAssetManagement2,274-283.
Campbell,R.,Huisman,R.andK.Koedijk,2001.OptimalportfolioselectioninaValue-at-
Riskframework.JournalofBankingandFinance25,1789-1804.
Chan,L.K.C.,andJ.Lakonishok,1993.Arethereportsofbetasdeathpremature?Journalof
PortfolioManagement19,51-62.
Chan, L.K.C., J. Karceski, and J. Lakonishok, 1999. On portfolio optimization: Forecasting
covariancesandchoosingtheriskmodel.ReviewofFinancialStudies12,937-974.
Clarke,R.,H.deSilva,andS.Thorley,2002.Portfolioconstraintsandthefundamentallaw
ofactivemanagement.FinancialAnalystsJournal58,48-66.
Elton, E.J., M.J. Gruber, S.J. Brown, and W.N. Goetzmann, 2003. Modern portfolio theory
andinvestmentanalysis.JohnWiley&Sons,NewYork.
Engle, R.F., 1982. Autoregressive conditional heteroskedasticity with estimates of the
varianceofU.K.inflation.Econometrica50,987-1008.
Frino, A., and D.R. Gallagher, 2001. Tracking S&P 500 Index funds. Journal of Portfolio
Management28,44-55.
Grinold,R.C.,andR.N.Kahn,1999.Activeportfoliomanagement.2
nd
edition.McGraw-Hill:
NewYork,NY.
Jansen,R.,andR.vanDijk,2002.Optimalbenchmarktrackingwithsmallportfolios.Journal
ofPortfolioManagement29,33-39.
Jorion,P.,2002,Portfoliooptimizationwithconstraintsontrackingerror.FinancialAnalysts
Journal,forthcoming.
Kritzman, M., 1987. Incentive fees: Some problems and some solutions. Financial Analysts
Journal43,21-26.
Larsen, G.A., and B.G. Resnick, 1998. Empirical insights on indexing. Journal of Portfolio
Management25,27-34.
Larsen, G.A., and B.G. Resnick, 2001. Parameter estimation techniques, optimization
frequency,andportfolioreturnenhancement.JournalofPortfolioManagement28,27-34.
Lawton-Browne, C., 2001. An alternative calculation of tracking error. Journal of Asset
Management2,223-234.
Plaxco, L.M., and R.D. Arnott, 2002. Rebalancing a global policy benchmark. Journal of
PortfolioManagement29,9-22.
Pope,P.,andP.K.Yadav,1994.Discoveringtheerrorintrackingerror.JournalofPortfolio
Management21,27-32.
RiskMetrics
TM
,1995.TechnicalDocument.3
rd
Edition.
Rohweder,H.C.,1998.Implementingstockselectionideas:Doestrackingerroroptimization
doanygood?JournalofPortfolioManagement25,49-59.
Roll,R.,1992.Amean/varianceanalysisoftrackingerror.JournalofPortfolioManagement
18,13-22.
Rosenberg,B.,1985.Predictionofcommonstockbetas.JournalofPortfolioManagement11
(2),5-14.
Rosenberg, B., and J. Guy, 1976. Beta and investment fundamentals. Financial Analysts
Journal32,60-72.
Rudd,A.,1980.Optimalselectionofpassiveportfolios.FinancialManagement4,57-66.
24
Rudd,A.,andB.Rosenberg,1980.TheMarketModelininvestmentmanagement.Journal
ofFinance35,597-606.
Rudolf, M., H-J. Wolter, and H. Zimmermann, 1999. A linear model for tracking error
minimization.JournalofBankingandFinance23,85-103.
Satchell,S.E.,andS.Hwang,2001.Trackingerror:Ex-anteversusex-postmeasures.Journal
ofAssetManagement2,241-246.
Scowcroft, A., and J. Sefton, 2001. Do tracking errors reliably estimate portfolio risk?
JournalofAssetManagement2,205-222.
Timmermann, A., and D. Blake, 2002. International asset allocation with time-varying
investmentopportunities.ForthcominginJournalofBusiness.
Treynor,J.,andF.Black,1973.Howtousesecurityanalysistoimproveportfolioselection.
JournalofBusiness46,66-86.
Zenti,R.,andM.Pallotta,2002.Areassetmanagersproperlyusingtrackingerrorestimates?
JournalofAssetManagement3,279-289.