Barra Predicting Risk at Short Horizons

msci.
com
ModelInsight
PredictingRiskatShortHorizons
ACaseStudyfortheUSE4DModel
JoseMenchero,AndreiMorozovandAndreaPasqua
Jose.Menchero@msci.com
Andrei.Morozov@msci.com
Andrea.Pasqua@msci.com
January2013
MSCIAnalyticsResearch msci.com
2013MSCIInc.Allrightsreserved.
Pleaserefertothedisclaimerattheendofthisdocument.
ModelInsight
PredictingRiskatShortHorizons(USE4D)
January2013
2 of26
Contents
Introduction.................................................................................3
Evaluating the Accuracy of Risk Forecasts.......................3
Bias Statistics............................................................................................................4
Mean Rolling Absolute Deviation (MRAD)......................................................4
Adjusted MRAD.......................................................................................................7
Cross-sectional Bias Statistics.............................................................................7
Q-statistics.................................................................................................................7
The USE4 Model and Data Set..........................................10
Estimating Volatility................................................................11
Exponentially Weighted Moving Averages (EWMA).................................11
GARCH(1,1)..........................................................................................................13
Volatility Regime Adjustment (VRA)...............................................................13
Empirical Results....................................................................16
Setting the Correlation Half-Life..........................................22
Conclusion...............................................................................24
References...............................................................................25
Client Service Information is Available 24 Hours a Day............................26

Notice and Disclaimer.........................................................................................26
About MSCI............................................................................................................26
ModelInsight
January2013
3 of26
Introduction
Equityfactormodelsareafairlyrecentinvention.BarrRosenbergpioneeredtheuseofmultifactorrisk
modelsasarobustwaytoestimatetheassetcovariancematrix(1974).In1975hefoundedBarra,which
developedthefirstcommerciallyavailableriskmodelforUSequities,dubbedUSE1.
Initially,theUSE1Modelwasestimatedfromquarterlydata.Later,asdatabecomemorewidely
available,theobservationfrequencywasincreasedtomonthly.Formanyyears,usingmonthly
observationstoestimatethefactorcovariancematrixwasstandardpractice.Forinstance,theBarra
USE3Model,releasedin1998,usedmonthlyfactorreturnswithahalflifeof90monthsinthe
estimationprocess.
TheInternetBubblepresentedaseriouschallengeformodelsestimatedwithmonthlydata.Thecruxof
theproblemwasthatvolatilitychangesweretoorapidandextremetobereliablycapturedusinglow
frequencyobservations.Inresponsetothischallenge,BarraresearchersdevelopedtheUSE3SModel,
whichuseddailyfactorreturnsforestimatingthecovariancematrix.Thehigherfrequencyof
observationsallowedthemodeltoadaptmorerapidlytochanginglevelsofvolatility.
AlthoughtheUSE3SModelemployeddailyfactorreturns,itmaintainedapredictionhorizonofone
month.Thiswasaccomplishedbyexplicitlyaccountingfortheeffectsofserialcorrelationinfactor
returns,whichcancausesignificantdeviationsfromthefamiliarsquarerootoftimescaling.For
instance,thereturnstotheMomentumfactortypicallyexhibitpositiveserialcorrelation.Thismakesthe
factorsignificantlymorevolatileatthemonthlyhorizonthanwouldbesuggestedbyapplyingsquare
rootoftimescalingtodailyvolatility.
Formanyinstitutionalinvestors,however,therelevanthorizonmaybemuchshorterthanonemonth.
Inthispaper,wefocusononedaypredictionhorizons.Onepossibleapproachforpredictingriskatone
dayhorizonswouldbetotaketheforecastsfromamonthlymodelandsimplyapplysquarerootoftime
scalingtobringthepredictionhorizontoasingleday.However,therearetwoshortcomingswiththis
approach.First,themonthlymodeliscalibratedforalongerhorizonandwillnothavetheappropriate
responsivenessforaonedayforecast.Second,theserialcorrelationadjustmentsthatwereusedto
provideaccurateforecastsatamonthlyhorizonnowrepresentsourcesoferrorataonedayhorizon.In
otherwords,iftheobservationfrequencyissynchronizedwiththepredictionhorizon,thenserial
correlationadjustmentsshouldnotbeincorporated.
Inthispaper,wehighlightsomeofthemodelingissuesthatmustbeaddressedwhenconstructinga
modelwithaonedaypredictionhorizon.Centraltothischallengeistheidentificationofareliable
metrictoevaluatetheaccuracyofriskforecasts.
EvaluatingtheAccuracyofRiskForecasts
Buildingasoundriskmodelfirstrequiresareliablemeansofevaluatingtheaccuracyofriskforecasts.
Thisprovidesanessentialguideforsettingmodelparametersandevaluatingperformance.Ideally,our
measureofforecastingaccuracywillprovidebothtimeandportfolioresolution,meaningthatthe
measurecanbeusedtoevaluateriskforecastsforasingleportfolioacrosstime,orforacollectionof
portfolioswithinasingletimeperiod.

ModelInsight
January2013
4 of26
BiasStatistics
Onecommonlyusedmeasuretoevaluatetheaccuracyofriskforecastsisthebiasstatistic,which
conceptuallyrepresentstheratioofrealizedrisktoforecastrisk.Tocomputethebiasstatisticfora
portfolio n ,wefirstusetheriskmodeltopredicttheportfoliovolatility
nt
o atthestartofeveryperiod
t .Wethenobservetheoutofsamplereturnoftheportfolio
nt
R overthesubsequentperiod.The
standardizedreturnisdefinedby
nt
nt
nt
R
z
o
= , (1)
andexpressestheportfolioreturnasazscore.Thebiasstatisticisgivenbythestandarddeviationof
standardizedreturns,
( )
2
1
1
1
T
n nt n
t
B z z
T
=
=

, (2)
where T isthenumberofdaysinthetestingwindow.
Conceptually,thebiasstatisticmeasureswhethertheriskforecastswereaccurate,onaverage,fora
singleportfolioacrosstime.Foraccurateforecasts,weexpecttherealizedbiasstatistictobecloseto1.
However,duetosamplingerror,thebiasstatisticwillneverbeexactly1evenforperfectrisk
forecasts.Instead,itiscustomarytoidentifyaconfidenceinterval.Assumingnormallydistributed
returnsandperfectforecasts,the95percentconfidenceintervalisapproximately1 2/T .
Twoattractivefeaturesofthebiasstatisticarethatitissimpletointerpretandprovidesportfolio
resolution.Unfortunately,itdoesnotprovidetimeresolution,meaningthatitpossibletounderpredict
riskforsomesubperiodsandoverpredictitforotherswhilenonethelessobtainingabiasstatisticclose
to1.Inotherwords,thebiasstatisticmayallowcancellationoferrorsacrosstime.Thisbecomes
especiallyproblematicoverlongsampleperiodsencompassingmanyyearsandmultiplemarket
regimes.Ariskmodelusermustbeconfidentthatvolatilityforecastsarereliableforallmarketregimes
notjustonaverage.
MeanRollingAbsoluteDeviation(MRAD)
Onewaytopartiallymitigatethetimeresolutionproblemistodividethelongsampleperiodinto
smallersubperiods,andtocomputethebiasstatisticsoverthesesubperiods.Forinstance,ifweselect
12daysasthelengthofoursubperiod,weobtain,
( )
11
2 1
11
n nt n
t
B z z
t
t
t
+
=
=
, (3)
wheret denotesthestartofthe12daysubperiod.Thewindowisthenrolledforwardonedayata
timeuntilreachingtheendoftheentiresample.
TheconventionalMeanRollingAbsoluteDeviation,orMRAD,isdefinedasthemeanabsolutedeviation
ofthebiasstatisticsfromtheiridealvalueof1,

ModelInsight
January2013
5 of26
( )
1
1
11
n
n
MRAD B
N T
t
t
=

, (4)
where N isthenumberofportfoliosand ( 11) T isthetotalnumberof(overlapping)12daysub
periods.Byvirtueoftheabsolutevalue,MRADpenalizesbothunderpredictionandoverpredictionof
risk.Assumingperfectforecasts,normallydistributedreturns,and12daysubperiods,itmaybeshown
thattheexpectedvalueofMRADisapproximately0.17.
Realfinancialdata,ofcourse,tendtohavefattails.InFigure1,weplottheexpectedvalueofMRAD
versuskurtosisforperfectriskforecasts.Theseresultsweregeneratedviasimulationbydrawing12
randomlygeneratedreturnsfromafattaileddistributionwithstandarddeviationequalto1.Thefat
tailswereobtainedusingastudenttdistributionwiththeappropriatenumberofdegreesoffreedom.
Weseethatasthekurtosisrisestoevenmodestlevels,theexpectedMRADforperfectforecasts
increasessignificantly.
Figure 1: MRAD versus kurtosis for perfect risk forecasts. Kurtosis levels were varied using a student t-
distribution with the appropriate number of degrees of freedom.
Kurtosis
3 4 5 6 7 8 9 10
M
R
A
D
0.17
0.18
0.19
0.20
0.21
0.22
0.23
0.24
WhiletheMRADmeasurehelpsmitigatecancelationoferrorsoverthelongrun,itisnotsuitableforrisk
modelcalibration,asitmayspuriouslyfavoroverlyresponsiveforecasts.Thiscanbestbeseenbya
simulationexperiment.Wegeneratedonemillionsimulatedreturnsfromastandardnormal
distribution.Wethenestimatedthevolatilityusingexponentiallyweightedaverageswithagivenhalf
lifeparameter.Sincethesimulatedreturnsarestationary(i.e.,theirdistributiondoesnotchangeover
time),thetrueoptimalhalflifeisinfinite,asthiseliminatessamplingerrorandhencerecoversthe
perfectriskforecast.

ModelInsight
January2013
6 of26
InFigure2,wepresenttheresultsofthesimulationstudy.Whereasthetrueoptimalhalflifeisinfinite,
Figure2spuriouslypointstoanoptimalhalflifeofonlysixdays.Furthermore,theminimumMRADis
approximately0.16,wellbelowthetheoreticallowerboundof0.17forperfectforecasts.Howcanthese
puzzlingresultsbeexplained?
Figure 2: Conventional MRAD versus EWMA volatility half-life. Returns were generated using a standard
normal distribution. The conventional MRAD displays a spurious minimum at a six-day half-life, and exhibits
a value below the theoretical lower bound for perfect forecasts.
VolatilityHalfLife(Days)
0 5 10 15 20 25 30 35 40
C
o
n
v
e
n
t
i
o
n
a
l
M
R
A
D
0.15
0.16
0.17
0.18
0.19
0.20
0.21
0.15
0.16
0.17
0.18
0.19
0.20
0.21
0.22
SpuriousMinimum
ThisconundrumisresolvedbyrecognizingthatthebiasstatisticandbyextensiontheMRADisnot
atrueoutofsamplemeasure.Thismaybesurprisingatfirstglance,sincethebiasstatisticiscomputed
usingoutofsamplezscores.Nevertheless,becauseriskforecastsareupdateddaily,thebiasstatistic
usesinformationinthevolatilityforecaststhatwasnotknownatthestartofthe12dayperiod.
Thatthistendstofavoroverlyresponsiveriskforecastsmaybeseenbythefollowingthought
experiment.Assumeforsimplicitythatreturnsaredrawnfromastandardnormaldistribution.Suppose
that,bychance,theportfolioexperiencesseverallargestandardizedreturns.Inthiscase,the12day
biasstatisticislikelytoexceed1.However,byoverreactingtotheeventanddramaticallyincreasingrisk
forecasts,thesubsequentzscoreswillbeartificiallyreduced.Althoughtheriskforecastsfollowingthe
eventarebiasedupward,therealizedbiasstatisticiscloserto1.Conversely,astringofsmallreturnsis
likelytoleadtoabiasstatisticlessthan1.Byreducingriskforecasts,thesubsequentstandardized
returnswillbeartificiallyinflated,againleadingtoabiasstatisticcloserto1.Inotherwords,excessive
responsivenessproducesnoisyandinaccurateriskforecasts,althoughtherealizedbiasstatisticsmaybe
deceptivelycloseto1.

ModelInsight
January2013
7 of26
AdjustedMRAD
Astraightforwardwaytoremovetheinsampleeffectistosimplyuseriskforecaststakenfromthestart
ofthe12dayperiod,andtoholdtheseforecastsconstantoverthenext12days.Thisleadstothe
adjustedMRADmeasure.AnadvantageoftheadjustedMRADmeasureisthatitusesnoinformationin
thevolatilityforecaststhatwasnotknownatthestartofthe12daysubperiod,andtherefore
representsatrueoutofsamplemeasure.AdisadvantageoftheadjustedMRADisthatitrequiresusing
somewhatstaleforecasts.
CrossSectionalBiasStatistics
Anothermeasureofforecastingaccuracyisthecrosssectionalbiasstatistic,
( )
2
1
1
1
N
t nt t
n
B z z
N
=
=

. (5)
ThisissimilartothetimeseriesbiasstatisticofEquation2,exceptthatnowthestandarddeviationis
computedacrossportfoliosforasingleday t .Sincethecrosssectionalbiasstatisticusesnoinformation
forvolatilityforecaststhatwasunknownatthestartofthesubperiod,itrepresentsatrueoutof
samplemeasure.
Conceptually,thecrosssectionalbiasstatisticmeasureswhethertheriskmodelforecastswere
accurate,onaverage,foracollectionofportfoliosonasingleday.Itcannot,however,answerwhether
theriskmodelforecastswereaccurateforindividualportfolios.Inotherwords,thecrosssectionalbias
statisticprovidestimeresolution,butnotportfolioresolution.Consequently,itispossibletoobtaina
crosssectionalbiasstatisticcloseto1,evenifriskforecastsforindividualportfolioswerepoor.In
particular,thismayoccurifoverpredictionerrorsforsomeportfoliosarecanceledbyunderprediction
errorsforothers.
Qstatistics
Patton(2011)describesmeasuresofforecastaccuracyintermsoflossfunctions.Hedefinesaloss
functionasrobustiftherankingofanytwovolatilityforecastsbyexpectedlossisthesamewhether
therankingisdoneusingthetruevariance(unobservable)orsomeunbiasedvarianceproxy(e.g.,
squaredreturn).OneexampleofarobustlossfunctionistheQstatistic,definedforportfolio n and
time t as
( )
2 2
ln
nt nt nt
Q z z = . (6)
PattonfurthershowsthattheQstatisticistheuniquelossfunction(uptotrivialadditiveand
multiplicativeconstants)thatdependssolelyonstandardizedreturns(i.e.,zscores).ThismakestheQ
statisticidealforevaluatingriskmodelaccuracy,becauseitplaceseveryobservationonanequalfooting
(whetherthevolatilityishighorlow).
Anotherkeypropertyofrobustlossfunctionsisthattheyareminimizedinexpectationwhenthe
predictedvolatilityequalsthetruevolatility.Forotherlossfunctions,thisisnottrue.Thatis,abiased
volatilityforecastcanminimizethelossfunction.Thiswouldobviouslybeproblematicforriskmodel
calibrationpurposes.TheMRADisanexampleofalossfunctionthatisnotrobust.

ModelInsight
January2013
8 of26
AlsonotethattheQstatisticclearlysatisfiesbothtimeandportfolioresolution.Thatis,itisnotbased
onaveraging,soitisnotpossibletooffsetanoverforecastingerrorforoneobservationwithanunder
forecastingerrorforadifferentobservation.Ofcourse,theQstatisticcanbeaveragedeitheracross
timeoracrossportfoliostoobtainthedesiredresolution.
Intuitively,wecanthinkoftheQstatisticasbeingcomprisedoftwopenaltyfunctions.Thefirstterm,
2
nt
z ,becomeslargewhenriskforecastsaretoolowandthereforerepresentsanunderforecasting
penaltyfunction.Thesecondterm,
( )
2
ln
nt
z ,representstheoverforecastingpenaltyfunctionand
dominateswhenriskforecastsaretoohigh.
ItisimportanttounderstandtheexpectedvalueoftheQstatisticforperfectriskforecasts.Ifreturns
arenormallydistributed,theexpectedvalueoftheQstatisticisapproximately2.27.AswiththeMRAD,
however,itisarathersensitivefunctionofkurtosis.InFigure3weplottheexpectedvalueoftheQ
statisticversuskurtosis,assumingperfectriskforecasts.
Figure 3: Q-statistic versus kurtosis for perfect risk forecasts. Kurtosis levels were varied using a student t-
distribution with the appropriate number of degrees of freedom.
Kurtosis
3 4 5 6 7 8 9 10
Q
s
t
a
t
i
s
t
i
c
2.25
2.30
2.35
2.40
2.45
2.50
2.55
2.60
ItisalsointerestingtoconsiderhowbiasesinvolatilityforecastsaffectchangesintheQstatistic.
Assumethatreturnsaredrawnfromastandardnormaldistribution(i.e.,withtruevolatilityof1).Let
thevolatilityforecastsbedenotedby
P
o .Onemayshowthattheexpectedlossrelativetoperfect
forecastsisgivenby
| | ( )
2
1
2ln 1
P
P
E Q o
o
A = + . (7)
ThisfunctionisplottedinFigure4.Ifthepredictedvolatilityisequaltothetruevolatility ( ) 1
P
o = ,then
theexpectedlossiszero.Notealsothatthelossfunctionisasymmetric.Thatis,itpenalizesunder

ModelInsight
January2013
9 of26
predictionofriskmoreheavilythanoverpredictionofrisk.Forinstance,ifthepredictedvolatilityishalf
thetruevolatility,thentheincreaseinQstatisticisapproximately1.61.However,ifthepredicted
volatilityisdoublethetruevolatility,thentheincreaseinQstatisticisonlyabout0.64.Thisisanother
attractivefeatureoftheQstatistic,sincemostinvestmentmanagersregardunderpredictionofriskas
moreproblematicthanoverpredictionofrisk.
Figure 4: Increase in Q-statistic versus predicted volatility when the true volatility is 1. Results were obtained
assuming a normal distribution. The asymmetry indicates that the Q-statistic penalizes under-forecasting of
risk more heavily than over-forecasting of risk.
PredictedVolatility
0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
A
Q
0.0
0.5
1.0
1.5
2.0
ItisinstructivetorepeatthesimulationexerciseofFigure2,exceptnowusingtheadjustedMRADand
Qstatistics.TheresultsarepresentedinFigure5.Weseethatneitherofthesemeasuresexhibitsa
spuriousminimum,thuscorrectlypointingtoaninfiniteoptimalhalflife.
ModelInsight
January2013
10 of26
Figure 5: Q-statistic and adjusted MRAD versus EWMA volatility half-life. Returns were generated using a
standard normal distribution. Neither quantity exhibits a spurious minimum, thus correctly pointing to an
infinite half-life.
0 5 10 15 20 25 30 35 40
Q
2.2
2.3
2.4
2.5
2.6
M
R
A
D
0.15
0.16
0.17
0.18
0.19
0.20
0.21
0.22
0.23
MeanQ
MRAD(Adjusted)
TheUSE4ModelandDataSet
TheempiricaldatasetusedforthisstudyconsistsofthehistoryofdailyfactorreturnsfromtheBarraUS
EquityRiskModel(USE4).Thehistoryofdailyfactorreturnsstartsin1993,whilethevolatilityforecasts
inourempiricalstudybeginonJuly19,1995.Thedatasetcontains4,328tradingdays,endingon
September21,2012.
TheUSE4Modelisavailableinlonghorizon(USE4L)andshorthorizon(USE4S)versions.Bothmodels
sharethesamefactorstructureandfactorreturns,butdifferintheirresponsiveness.TheUSE4SModel
isdesignedtoprovidethemostaccurateforecastsataonemonthpredictionhorizon.TheUSE4LModel
istailoredforlongerterminvestorswhoarewillingtotradesomedegreeofforecastingaccuracyfor
greaterstabilityintheriskforecasts.WiththelaunchofUSE4D,theUSE4Modelisnowavailableina
thirdversionfordailyhorizonforecasts.Again,theUSE4DModelsharesthesamefactorstructureand
factorreturnsasthetwootherUSE4Modelvariants.
Here,webrieflyreviewsomehighlightsoftheUSE4Model.EmpiricalresultscanbefoundinYang,
Menchero,Orr,andWang(2011),whilemethodologydetailsaredescribedinMenchero,Orr,andWang
(2011).TheUSE4Modelcontains73factors,comprisedof:(a)theCountryfactor,whichisthe
regressionintercept,(b)60industryfactors,and(c)12stylefactors.Factorreturnsareestimatedby
performingdailycrosssectionalregressionsofstockreturnsagainstthestartofdayfactorexposures.
TheestimationuniverseistheMSCIUSAIMI,abroadindexrepresentingtheUSmarket.Themodel

ModelInsight
January2013
11 of26
employssquarerootofmarketcapitalizationastheregressionweights.Theexactcolinearitybetween
industryfactorexposuresandtheCountryfactorisremovedbyconstrainingtheindustryfactorreturns
tobecapweightedmeanzeroeveryperiod.
AsdescribedbyMenchero(2010),theestimatedfactorreturnsmaybeinterpretedasthereturnsof
factormimickingportfolios.TheCountryfactorportfolioessentiallyrepresentsthecapweighted
estimationuniverse(i.e.,MSCIUSAIMI).Theindustryfactorportfoliosaredollarneutralandcapture
theperformanceoftheindustrynetofthemarketandotherstyles.Thestylefactorportfoliosaredollar
neutralandhaveunittiltontheparticularstyle,withzeroexposuretoindustriesandotherstyles.
EstimatingVolatility
Thesimplestwaytoestimatevolatilityistocomputethesamplestandarddeviationoveratrailing
window.Ifstockreturndistributionswerestationary,thenusingthemaximumsamplesizeandequally
weightingeveryobservationwouldminimizesamplingerrorandhenceproducethemostaccurate
forecast.
Stockreturndistributions,however,arenotstationary.Eventsthatoccurredtenyearsagohavelittleto
dowithcurrentvolatilitylevels.Therefore,toreflectcurrentmarketconditions,wemustgivemore
weighttorecentobservations.Thecrucialquestionishowmuchmore?Ontheonehand,ifwegivetoo
muchweighttorecentobservations,thenwebaseourestimatesonveryfewdatapointsandweare
stronglypenalizedwithsamplingerror.Bycontrast,ifwegivetoomuchweighttodistantobservations,
thenweareharmedbyincorporatingstaledataintoourvolatilityestimates.Makinggoodriskforecasts
forrealfinancialdatarequiresanoptimaltradeoffbetweenthesetwoeffects.
ExponentiallyWeightedMovingAverages(EWMA)
Asimple,yeteffective,techniqueforattachingmoreweighttorecentobservationsisexponential
weightedmovingaverages(EWMA).Inthisapproach,thevarianceiswrittenasaweightedaverageof
laggedsquaredreturns.
2 2
t m t m
m
w r o

=
, (8)
wheretheweights
m
w sumto1,anddecreasebyfixedproportioneveryperiod.Morespecifically,
1 m m
w w
+
= ,where isthedecayfactor.TheEWMAestimatecanbeconvenientlyrewrittenin
recursiveformasaweightedaveragebetweenyesterdayssquaredreturnandyesterdaysvariance
forecast,
( )
2 2 2
1 1
1
t t t
r o o

= + . (9)
Therelativeweightweplaceonyesterdaysvarianceforecastdeterminestheresponsivenessofthe
modelandisrelatedtotheEWMAvolatilityhalflifeparameterasfollows,
( )
( )
ln 0.5
ln
HL
= . (10)

ModelInsight
January2013
12 of26
TheEWMAvolatilityhalflifeisacriticalmodelparameter.Toguideusinselectinganappropriatevalue,
wecomputetheaverageadjustedMRADandQstatisticforUSE4dailyfactorreturns,withtheaverages
takenoverallfactorsandtimeperiods.InFigure6,weplotthesequantitiesversustheEWMAvolatility
halflife.UnlikethesimulatedresultsinFigure5,weseethatrealfinancialdatadoexhibitaclear
minimum.Figure6showsthattheoptimalvolatilityhalflifebyboththeQstatisticandtheadjusted
MRADisabout21tradingdays,oronemonth.Reassuringly,bothoutofsamplestatisticspointtothe
sameoptimalhalflife.Notethata21dayvolatilityhalflifecorrespondstoadecayfactorof0.97
accordingtoEquation10.
Figure 6: Q-statistic, adjusted MRAD, and conventional MRAD versus volatility half-life for EWMA forecasts
using USE4 daily factor returns. The optimal half-life is about one month (21 trading days) by either the Q-
statistic or the adjusted MRAD. The conventional MRAD spuriously points to an optimal half-life of about five
trading days.
0 20 40 60 80 100
Q
2.35
2.40
2.45
2.50
2.55
2.60
M
R
A
D
0.2000
0.2200
0.2400
0.2600
0.2800
0.3000
MeanQ
MRAD(Conventional)
MRAD(Adjusted)
Forcomparisonpurposes,wealsopresenttheconventionalMRADinFigure6.Notethatthismeasure
suggestsanoptimalhalflifeofjustfivedays.AlsoobservethattheminimumMRADiswellbelowthe
minimumvalueoftheadjustedMRAD.Theseresultsillustratethistimeusingrealfinancialdata
thesameeffectseeninthesimulatedresultsofFigure2.Thatis,insampleeffectsspuriouslypointto
excessivelyshorthalflifeparameters,whichinturnproducedeceptivelyattractiveMRADvalues.

ModelInsight
January2013
13 of26
GARCH(1,1)
ThemethodofGeneralizedAutoRegressiveConditionalHeteroskedasticty,orGARCH,wasdevelopedby
Engle(1982),Bollerslev(1986)andothersinthe1980s.Bynow,therearemanyvariationsofGARCH.An
excellentsurveyofthesetechniquesisprovidedbyBauwens,Laurent,andRombouts(2006)
Inthispaper,wefocusonthemostwidespreadvariantofGARCH,namelyGARCH(1,1).Inthisapproach,
varianceestimatesaregivenby
2 2 2 2
1 1 t t t
r o o |o o

= + + . (11)
Thisexpressespredictedvarianceasaweightedaverageofthreeterms:(a)yesterdayssquaredreturn,
withweight o ,(b)yesterdaysvarianceforecast,withweight | ,and(c)alongrunvariance,with
weight .Thestabilityconditionrequiresthattheweightssumto1,
1 o | + + = . (12)
Wedeterminetheparametersusingmaximumlikelihoodestimationwithatwoyearlookbackwindow.
WeestimateGARCHparametersseparatelyforeveryfactorinthemodel.
GARCHhasseveralconceptuallyappealingattributes.First,themodelprovidesformeanreversionin
volatilityforecaststhroughthelongrunvarianceterminEquation11.Also,GARCHreplicatesthe
volatilityclusteringoftenfoundinfinancialtimeseries.Finally,bysettingthe parametertozero,
GARCH(1,1)encompassesEWMAasaspeciallimitingcase.
VolatilityRegimeAdjustment(VRA)
TheUSE4DModelutilizestheVolatilityRegimeAdjustment(VRA)techniquetoestimatefactor
volatilities.ThistechniquewasfirstintroducedwiththelaunchoftheBarraUSEquityModel(USE4),as
describedbyMenchero,Orr,andWang(2011).Thecentralconceptbehindthisapproachistousecross
sectionalobservationstocalibratethemodeltocurrentvolatilitylevels.
Riskmodelsbaseforecastsonhistoricalobservations.However,sincevolatilitylevelsvaryacrosstime,
thismayleadtocertainbiasesintheriskforecasts.Forinstance,enteringaperiodoffinancialcrisis,
volatilitylevelstendtorise.Inthiscase,riskmodelsusethelowervolatilitypasttopredictthehigher
volatilityfuture,thuscausingatendencytounderpredictriskduringtheseperiods.Conversely,
immediatelyfollowingaperiodoffinancialturmoil,therearemanyextremeeventsintheestimation
window,whichcreatesatendencytooverpredictriskatsuchtimes.Sinceriskmodelsarenotcrystal
balls,thereisnowaytocompletelyeliminatethesebiases.TheVRAtechnique,however,isdesignedto
mitigatethesebiasesbyusingcrosssectionalobservationstocalibratethemodeltocurrentvolatility
levels.
ThefirststepintheVRAtechniqueistoestimatefactorvolatilitiesusingEWMA,asinEquation8.This
estimateischaracterizedbythevolatilityhalflifeparameter,definedbyEquation10.Next,wecompute
thecrosssectionalbiasstatistic,asinEquation5.Thecomputationisperformedoverthezscoresof
dailyfactorreturns,usingtheEWMAvolatilityforecastsfromthefirststep.Wecanthinkofthecross
sectionalbiasstatisticasprovidinganinstantaneousmeasureofriskforecastingbias.Thefactor
volatilitymultiplier,attime t ,isdefinedintermsofthetrailingcrosssectionalbiasstatistics,
ModelInsight
January2013
14 of26
2 F
t m t m
m
u B

=

, (13)
where
m
u istheexponentialweightcharacterizedbytheVRAhalflifeparameter.
Ifthefactorvolatilitymultiplierisgreaterthan1,itsaystheoriginalEWMAestimatesweretoolow,and
riskforecastsshouldbeadjustedupward.Conversely,when
F
t
islessthan1,EWMAestimateswere
toohighandriskforecastsshouldbeadjusteddownward.Theadjustedvolatilityforecastforfactor k at
time t isgivenby
F
kt t kt
o o = . (14)
Thisadjustmentisdesignedtoremovetheaveragebias,withallfactorvolatilitiesscaledbythesame
proportion.
VolatilityestimationwiththeVRAtechniquethereforerequirestwohalflifeparameters:theEWMA
volatilityhalflifedefinedinEquation10,andtheVRAhalflifeusedinEquation13.Tofindtheoptimal
valuesforthesetwoparameters,weperformasearchfortheglobalminimumQstatisticoverthetwo
dimensionalspace.InFigure7,weplotcurvesofQstatistics(averagedacrossfactorsandtimeperiods)
versustheVRAhalflifeforseveralvaluesoftheEWMAvolatilityhalflife.Wefindthatthecombination
thatminimizestheQstatisticisa42dayEWMAvolatilityhalflifeinconjunctionwithatwodayVRA
halflife.
Figure 7: Q-statistic curves versus VRA half-life for three values of EWMA volatility half-life. The Q-statistic is
minimized at a VRA half-life of two days and an EWMA volatility half-life of 42 days.
VolatilityRegimeHalfLife(days)
0 5 10 15 20
Q
2.41
2.42
2.43
2.44
21dayVOLHL
84dayVOLHL
42dayVOLHL
ModelInsight
January2013
15 of26
Atfirstglance,atwodayVRAhalflifemayseemsurprisinglyshort.Recall,however,thatitisprimarily
thepenaltyofsamplingerrorthatimposeslimitsontheresponsivenessofthemodel.IntheEWMA
approach,wherevolatilitiesareestimatedonefactoratatime,samplingerrorisrelativelylarge.Inthe
VRAapproach,bycontrast,welargelymitigatesamplingerrorbyvirtueofusing73crosssectional
observationseveryday.
AlthoughatwodayVRAhalflifeisoptimalintermsofforecastingaccuracy,itmayleadtoratherlarge
dailyfluctuationsinpredictedvolatilities.Toreducesuchvariability,theUSE4DModelcombinesthe42
dayEWMAvolatilityhalflifewithafourdayVRAhalflife.Thisprovidesconsiderablymorestabilityin
volatilityforecasts,whileproducingonlyamodestincreaseintheQstatistic,asevidentfromFigure7.
Thefinancialcrisisperiodof2008and2009affordsaprimeexampletoillustratethebehaviorofthe
VRAtechnique.InFigure8,weplotthefactorvolatilitymultiplier
F
t
togetherwiththefactorcross
sectionalvolatility(CSV),definedas
2
1
t kt
k
CSV f
K
=

. (15)
Figure 8: Factor volatility multiplier and factor cross-sectional volatility (CSV) for the USE4D model. Results
were smoothed using 10-day rolling windows.
Year
2008 2009 2010
F
a
c
t
o
r
V
o
l
a
t
i
l
i
t
y
M
u
l
t
i
p
l
i
e
r
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
F
a
c
t
o
r
C
S
V
(
p
e
r
c
e
n
t
d
a
i
l
y
)
0
1
2
3
4
5
6
Factor
Volatility
Multiplier
Factor
CSV
ModelInsight
January2013
16 of26
InJanuary2008,weseethatfactorCSVspikedfrom60bpstoabout150bpsperday.Thefactor
volatilitymultiplierrespondedquicklytotheincreasedvolatilitylevels,reachingamaximumvalueof1.7
duringthemonth.ThenextbigspikeinfactorCSVcameatthestartofSeptember2008,whenlevels
stoodatabout100bpsperday.Overthenexttwomonths,theselevelshadmorethandoubledto250
bpsperday.Again,thefactorvolatilitymultiplierquicklydetectedtheincreasedvolatilitylevels,hitting
apeakof1.8inOctober2008.
BytheendofApril,2009,theworstofthefinancialstormwasover,andfactorvolatilitylevelsstoodat
about150bpsperday.Thenexteightmonthssawastrongseculardeclineinvolatilitylevels,withfactor
CSVhittingalowofabout50bpsperdaybytheendof2009.Overthistimeofrapidlydeclining
volatility,thefactorvolatilitymultiplierwaswellbelow1,hittinglowsofabout0.6duringthisperiod.
ThisexampledemonstratestheeffectivenessoftheVRAtechniqueforreducingvolatilityforecastsin
theaftermathofafinancialcrisis.
EmpiricalResults
Inthissection,wecompareforecastsforsixdifferentsetsofvolatilityestimates.Thedatasetusedfor
ourstudyisthefullhistoryofUSE4dailyfactorreturns,asdescribedpreviously.Thefirstsetofvolatility
estimatesistakendirectlyfromtheUSE4Dmodel,whichusesavolatilityhalflifeof42daysandaVRA
halflifeoffourdays.Next,weconsiderahighlyresponsiveEWMAforecast,withvolatilityhalflifeof
fourdays(i.e.,equaltotheVRAhalflifeofUSE4D).WerefertothisastheEWMA(4)forecast.Wealso
considertheEWMA(21)forecast,whichusestheoptimalvolatilityhalflifeof21days,asseeninFigure
6.ThelastEWMAforecastweconsiderisEWMA(42),whichusesthesamevolatilityhalflifeasUSE4D.
WealsoreportresultsfortheGARCH(1,1)model,describedabove.Finally,weexaminetheUSE4S
model,withvolatilityforecastsscaledtoaonedayhorizon.
Beforepresentingresultsonforecastingaccuracy,weconsiderthestabilityofthevolatilityforecasts.To
investigatethis,wefirstcomputeforfactor k andday t theabsolutechangeinvolatilityforecastover
thepreviousday,
, , 1
,
, 1
k t k t
k t
k t
v
o o
o

= . (16)
Wethendefinetheforecastvariabilityastheaverageoverallfactorsandalltimeperiods,
,
,
1
k t
k t
v v
KT
=

. (17)
ThevariabilityforeachvolatilityforecastisreportedinTable1.Thevariabilityrangedfromalowof63
bpsforUSE4S,toahighof7.96percentforEWMA(4).TheUSE4DModelhadanintermediatevariability
at2.82percent.
ModelInsight
January2013
17 of26
Table 1: Comparison for six different volatility estimates. Averages were computed across all 73 factors and
4328 trading days. Differences in Q-statistics are relative to the USE4D model.
Variability Conventional Adjusted QStatistic

Model (Percent) MRAD MRAD QStatistic (Difference)
USE4D 2.82 0.2124 0.2478 2.4182 0.0000
EWMA(4) 7.96 0.2059 0.3092 2.5231 0.1049
EWMA(21) 1.66 0.2239 0.2562 2.4455 0.0274
EWMA(42) 0.85 0.2403 0.2590 2.4564 0.0382
GARCH(1,1) 4.33 0.2147 0.2622 2.4486 0.0304
USE4S 0.63 0.2507 0.2647 2.4631 0.0450

InTable1,wealsoreportMRADandQstatisticsforthesixvolatilityforecasts.Bytheconventional
MRADmeasure,EWMA(4)hasthelowestscore.However,aswehaveseen,theconventionalMRAD
stronglyfavorsoverlyresponsiveforecastswhichcanmaketheMRADappearartificiallylow.Thetrue
outofsamplemeasures(adjustedMRADandQstatistic)paintadifferentpictureofforecasting
accuracy.
ByeithertheadjustedMRADortheQstatistic,theUSE4Dmodelproducedthemostaccuratevolatility
forecasts.WealsoreportinTable1thedifferenceinaverageQstatisticrelativetotheUSE4DModel.
Theaveragewascomputedoverall73factorsandtheentiresampleperiodof4,328tradingdays.The
secondmostaccurateforecastswereprovidedbytheEWMA(21)model,whichhadaQstatistic0.0274
aboveUSE4D.TheleastaccurateforecastswerefortheEWMA(4)model,whichscoredaQstatistic
0.1049abovetheUSE4Dmodel.
ItisausefulexercisetotranslatedifferencesinQstatisticsintoforecastingerrors.Thisisdifficulttodo
precisely,sinceitdependsontheforecasterrorsofeachestimateaswellasthereturndistributionsof
thetestportfolios.Nevertheless,aroughsensecanbegleanedfromEquation7andFigure4.Wesee
thatadifferenceinQstatisticof0.10translatesroughlyintoapredictedvolatilityof 0.81
P
o = (i.e.,19
percentunderforecastingbias)or 1.28
P
o = (28percentoverprediction).Similarly,aQstatistic
differenceof0.03correspondsroughlytoforecastsof 0.89
P
o = (11percentunderforecast)or
1.14
P
o = (14percentoverforecast).
Itisalsoimportanttoconsiderthestatisticalsignificanceofthesefindings.Onewaytogaugestatistical
significanceistocountthenumberoffactorsforwhichUSE4DproducedlowerQstatisticsthanthe
otherforecasts.ComparedwithGARCH(1,1),theUSE4DmodelproducedlowerQstatisticsfor71ofthe
73factors.RelativetoEWMA(42)orUSE4S,theUSE4Dmodeloutperformedfor72ofthe73factors.
Finally,versusEWMA(4)orEWMA(21),theUSE4Dmodeloutperformedforallfactors.Thelikelihoodof
theseresultsoccurringbymerechanceisexceedinglysmall.Thus,wecanconcludewithhighstatistical
confidencethattheUSE4Dmodelprovidedmoreaccurateriskforecaststhantheotherapproaches.
Although,onaverage,theUSE4DModelproducedmoreaccurateforecastsoverthefullsampleperiod,
itisimportanttoinvestigatethepersistenceoftheseresultsacrosstime.Forinstance,ifthe
outperformancewasdueprimarilytoaspecificsubperiod(suchasthe2008/2009financialcrisis),it
maycallintoquestiontherobustnessoftheVRAmethodology.Toexplorethispossibility,wedefinethe
cumulativeQincrement,

ModelInsight
January2013
18 of26
( ) ( )
1 1
1
K
A B
kt kt
k t
Q Q Q
KT
t
t
= =
A =
, (18)
where
A
kt
Q istheQstatisticforforecast A,factor k ,attime t ,and
B
kt
Q isthecorrespondingquantity
forforecast B .Wecanthinkof ( ) Q t A askeepingarunningtallyofthedifferenceinaccuracy
betweenforecasts Aand B .Attheendofthesampleperiod, T t = ,thecumulativeQincrement
becomessimplytheaveragedifferenceinQstatisticbetweenthetwoforecasts.Hereweletforecast B
representtheUSE4Dmodel,whereasforecast Aisusedtorepresentthealternative.
InFigure9weplotthecumulativeQincrementforallmodelsrelativetoUSE4D.Notethatallofthelines
haveapersistentupwardslope,indicatingthattheUSE4Dmodelconsistentlyprovidedmoreaccurate
forecastsacrossallmarketregimes.
Figure 9: Cumulative Q increment (relative to USE4D) versus time for five sets of volatility forecasts. The
persistent upward slope indicates that the USE4D consistently provided more accurate forecasts across
different market regimes.
Year
1995199719992001200320052007200920112013
C
u
m
u
l
a
t
i
v
e
A
Q
0.00
0.02
0.04
0.06
0.08
0.10
0.12
EWMA(4)
EWMA(21)
EWMA(42)
GARCH(1,1)
USE4S
Inordertodevelopgreaterintuitionforhowtheseforecastscompare,itisusefultocomparetime
seriesplotsofthevolatilityforecastsforspecificfactors.
InFigure10,weplotthepredictedvolatilityoftheUSE4CountryfactorcomputedwithEWMA(42)and
USE4D,overtheperiod20072010.Weseethatoverthelastfourmonthsof2008,theUSE4DModel
predictedsignificantlyhighervolatilityfortheCountryfactorthantheEWMA(42)forecast.
Conversely,forthelasteightmonthsof2009,theEWMA(42)forecastsweresignificantlyhigher.These
resultsareconsistentwiththegreaterresponsivenessofUSE4DrelativetoEWMA(42).

ModelInsight
January2013
19 of26
Figure 10: Predicted volatility of USE4 Country factor for USE4D and EWMA(42) forecasts.
Year
2007 2008 2009 2010
V
o
l
a
t
i
l
i
t
y
(
P
e
r
c
e
n
t
)
0
10
20
30
40
50
60
70
80
90
USE4D
EWMA(42)
InFigure11weplotthevolatilityoftheEarningsYieldfactorforGARCH(1,1)andUSE4Doverthetime
period20072010.Thiswasafactortowhichmanyquantitativeinvestorshadpositiveexposureover
thisperiod.
PriortotheQuantMeltdownofAugust2007,weseethattheGARCH(1,1)forecastswereverystable
andaboutatthesamelevelasUSE4Dforecasts.DuringtheheightoftheQuantMeltdown,the
GARCH(1,1)forecastswereabove12percent,versuseightpercentforUSE4D.
Forthenexttwoyears,whiletheQuantMeltdownremainedintheestimationwindow,theGARCH(1,1)
forecastswereveryjumpy.InAugust2009,astheeventexitstheestimationwindow,weseethatthe
GARCH(1,1)forecastssuddenlybecomemorestable.
ModelInsight
January2013
20 of26
Figure 11: Predicted volatility of USE4 Earnings Yield factor for USE4D and GARCH(1,1) forecasts.
Year
2007 2008 2009 2010
V
o
l
a
t
i
l
i
t
y
(
P
e
r
c
e
n
t
)
0
2
4
6
8
10
12
14
16
USE4D
GARCH(1,1)
Aswehaveseen,theQstatisticiscomposedofanoverforecastingandanunderforecastingpenalty
function.Forperfectriskforecasts,theexpectedvalueoftheunderforecastingpenaltyfunctionis
exactly1.Assumingnormaldistributionandperfectriskforecasts,theexpectedvalueoftheover
forecastingpenaltyfunctionis1.27.Plottingthesetwopenaltyfunctionsversustimeandcomparing
themtotheiridealvaluescanyieldimportantinsightintowhetherthevolatilityforecastsweretoohigh
ortoolow.
InFigure12,weplotthemeanQstatistic,togetherwiththeoverforecastingandunderforecasting
penaltyfunctions,averagedacrossall73factors,forEWMA(42)volatilityforecasts.Theplotcoversthe
financialcrisisperiodof2008and2009,andlinesweresmoothedusing10dayrollingaverages.TheQ
statisticexhibitsseveraldistinctpeaksoverthecourseof2008,withthelargestsuchpeakoccurringin
October2008.Weseethateachofthesepeaksin2008wascausedbyaspikeintheunderforecasting
penaltyfunction,implyingthattheEWMA(42)forecastwasnotresponsiveenoughduringthisperiod.
WecanevenusetheQstatistictoestimatethemagnitudeofthebias.InOctober2008,theQstatistic
wasatabout3.9,orabout1.6abovetheidealposition.FromFigure4,weseethistranslatesroughly
intoa50percentunderpredictionofrisk.Bycontrast,weseethatoverthelasteightmonthsof2009,
therelativelyhighvalueoftheQstatisticwasdominatedbytheoverforecastingpenaltyfunction.In
thiscase,theEWMA(42)modeldidnotadaptquicklyenoughtoreducedvolatilitylevelsinthewakeof
thefinancialcrisis.

ModelInsight
January2013
21 of26
Figure 12: Plot of Q-statistic together with over-forecasting and under-forecasting penalty functions for the
EWMA(42) forecast. The quantities represent averages across factors. Lines were smoothed using 10-day
rolling windows. The EWMA(42) forecast generally under-forecasts for much of 2008 and over-forecasts for
most of 2009.
Year
2008 2009 2010
Q
S
t
a
t
i
s
t
i
c
0
1
2
3
4
5
QStatistic
UnderforecastPenalty
OverforecastPenalty
InFigure13weplotfortheUSE4DforecaststhemeanQstatistic,averagedacrossallfactors,together
withoverforecastingandunderforecastingpenaltyfunctions.FromFigure13,itisclearlyevidentthat
theQstatisticforUSE4Dwasmuchclosertotheidealvalueof2.27thanthecorrespondingplotin
Figure12forEWMA(42).Furthermore,weseethattheoverforecastingandunderforecastingpenalty
functionsneverdeviatedtoofarfromtheiridealpositions.
ModelInsight
January2013
22 of26
Figure 13: Plot of Q-statistic together with over-forecasting and under-forecasting penalty functions for the
USE4D forecast. The quantities represent averages across factors. Lines were smoothed using 10-day rolling
windows. Compared to EWMA(42) forecasts, the USE4D Q-statistics are much closer to their ideal values.
Furthermore, the over-forecasting and under-forecasting penalties stay relatively close to their ideal values,
indicating that the model adapts well to changing volatility levels.
Year
2008 2009 2010
Q
S
t
a
t
i
s
t
i
c
0
1
2
3
4
5
QStatistic
UnderforecastPenalty
OverforecastPenalty
SettingtheCorrelationHalfLife
Thusfar,wehaveonlyconsideredvolatilityforecasts.Inbuildingafactorcovariancematrix,however,
wemustalsoestimatetheoffdiagonalcovariances.Thesearegivenbytheproductofthefactor
volatilitiesandthefactorcorrelations.Thecorrelations,inturn,areestimatedusingEWMAwitha
specifiedhalflifeparameter.
Thecorrelationhalflifeisacrucialmodelparameter.Ifitistoolong,theremaybeapenaltybyusing
staledatainourcorrelationestimates.Ifthecorrelationhalflifeistooshort,thenthecovariancematrix
maybecomenoisyandillconditioned,leadingtopoorperformanceandunderestimationofriskof
optimizedportfolios,asdescribedbyMenchero,Wang,andOrr(2011).
Onewaytodeterminetheoptimalcorrelationhalflifeistocomputethevaluethatproducesthebest
outofsampleperformanceofoptimizedportfolios.Tostudythis,weconstructtheminimumriskfully
investedfactorportfolioatthestartofeveryday.TheCountryfactorexposureandthesumofindustry
factorexposuresareconstrainedtoequal1.InFigure14weplottheoutofsamplerealizedvolatility
overthefullsampleperiodasafunctionofcorrelationhalflife.Weseethat200daysappearsoptimal,
asthisminimizestheoutofsamplerealizedvolatility.Ifthecorrelationhalflifeistoolong,wedoseea
penaltyinoutofsamplevolatilitybyusingstaledata.Note,however,thatthispenaltyisrelatively
modest:8.6percentat500daysversus8.3percentat200days.Bycontrast,whenthecorrelationhalf

ModelInsight
January2013
23 of26
lifebecomestooshort,thereisasteeppenaltyinoutofsamplevolatility.Inthiscase,modelestimates
areharmedbytoomuchnoiseinthecorrelationestimates.
Figure 14: Out-of-sample volatility for minimum-risk fully invested factor portfolio versus correlation half-life.
The out-of-sample period consisted of 4328 trading days. The volatility was minimized near a correlation half-
life of 200 days.
CorrelationHalfLife(Days)
0 100 200 300 400 500
V
o
l
a
t
i
l
i
t
y
(
A
n
n
u
a
l
i
z
e
d
)
8.0
8.5
9.0
9.5
10.0
10.5
MinRiskFullyInvested
Thereareseveralotherpossiblewaysofselectingthecorrelationhalflifeparameter.Forinstance,one
maystudythehalflifeparameterthatminimizestheoutofsamplevolatilitiesofstyleoptimized
portfolios.Anotherwaytosetthecorrelationhalflifeistoinvestigatethevaluethatproducesthebest
betaforecasts.Yetanotherapproachistodeterminewhichcorrelationhalflifeproducesthemost
accurateriskforecastsforfactorpairportfolios.
TheresultsoftheseinvestigationswerequalitativelysimilartotheresultsofFigure14.Namely,the
optimalcorrelationhalflifeiscloseto200days.Ifthecorrelationhalflifeistoolong,wefindaweak
penalty,butobserveamuchlargerpenaltyifthecorrelationhalflifeistooshort.Basedonthese
studies,weselecta200daycorrelationhalflifefortheUSE4Dmodel.

ModelInsight
January2013
24 of26
Conclusion
Wehaveinvestigatedtherelativeaccuracyofvariousvolatilityforecastsoveraonedayprediction
horizon.WeexaminedseveralEWMAforecasts,aGARCH(1,1)model,theUSE4SModel(scaledtoaone
dayhorizon),andfinallytheUSE4DModel.WefoundthattheUSE4DModelprovidedthemostaccurate
forecastsamongallmodelsconsidered.Furthermore,theoutperformancewasconsistentacrossfactors
aswellaspersistentacrosstime.

ModelInsight
January2013
25 of26
References
Bollerslev.1986.GeneralizedAutoregressiveConditionalHeteroskedasticity.JournalofEconometrics:
307327.
Bauwens,Laurent,andRombouts.2006.MultivariateGARCHModels:ASurvey.JournalofApplied
Econometrics,vol.21:79109.
Engle.1982.AutoregressiveConditionalHeteroskedasticitywithEstimatesofVarianceUKInflation.
Econometrica:9871008.
Menchero.2010.TheCharacteristicsofFactorPortfolios.JournalofPerformanceMeasurement,vol.
15,no.1(Fall):5262.
Menchero,OrrandWang.2011.TheBarraUSEquityModel(USE4).MethodologyNotes.
Menchero,Wang,andOrr.2012.ImprovingRiskForecastsforOptimizedPortfolios,FinancialAnalysts
Journal,May/June2012:4050.
Patton.2011.VolatilityForecastComparisonusingImperfectVolatilityProxies.Journalof
Econometrics,vol.160(1),246256.
Rosenberg,Barr.March1974.ExtraMarketComponentsofCovarianceinSecurityReturns,Journalof
FinancialandQuantitativeAnalysis,vol.9,263274.
Yang,Menchero,OrrandWang.2011.TheBarraUSEquityModel(USE4).EmpiricalNotes.
ModelInsight
January2013
26 of26
ClientServiceInformationisAvailable24HoursaDay
clientservice@msci.com
NoticeandDisclaimer
- This document and all of the information contained in it, including without limitation all text, data, graphs, charts (collectively, the Information) is the property of MSCI Inc. or its
subsidiaries (collectively, MSCI), or MSCIs licensors, direct or indirect suppliers or any third party involved in making or compiling any Information (collectively, with MSCI, the
InformationProviders)andisprovidedforinformationalpurposesonly.TheInformationmaynotbereproducedorredisseminatedinwholeorinpartwithoutpriorwrittenpermission
fromMSCI.
- TheInformationmaynotbeusedtocreatederivativeworksortoverifyorcorrectotherdataorinformation.Forexample(butwithoutlimitation),theInformationmaynotbeusedto
createindices,databases,riskmodels,analytics,software,orinconnectionwiththeissuing,offering,sponsoring,managingormarketingofanysecurities,portfolios,financialproductsor
otherinvestmentvehiclesutilizingorbasedon,linkedto,trackingorotherwisederivedfromtheInformationoranyotherMSCIdata,information,productsorservices.
- TheuseroftheInformationassumestheentireriskofanyuseitmaymakeorpermittobemadeoftheInformation.NONEOFTHEINFORMATIONPROVIDERSMAKESANYEXPRESSOR
IMPLIED WARRANTIES OR REPRESENTATIONS WITH RESPECT TO THE INFORMATION (OR THE RESULTS TO BE OBTAINED BY THE USE THEREOF), AND TO THE MAXIMUM EXTENT
PERMITTED BY APPLICABLE LAW, EACH INFORMATION PROVIDER EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES (INCLUDING, WITHOUT LIMITATION, ANY IMPLIED WARRANTIES OF
ORIGINALITY, ACCURACY, TIMELINESS, NONINFRINGEMENT, COMPLETENESS, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE) WITH RESPECT TO ANY OF THE
INFORMATION.
- Without limiting any of the foregoing and to the maximum extent permitted by applicable law, in no event shall any Information Provider have any liability regarding any of the
Informationforanydirect,indirect,special,punitive,consequential(includinglostprofits)oranyotherdamagesevenifnotifiedofthepossibilityofsuchdamages.Theforegoingshallnot
excludeorlimitanyliabilitythatmaynotbyapplicablelawbeexcludedorlimited,includingwithoutlimitation(asapplicable),anyliabilityfordeathorpersonalinjurytotheextentthat
suchinjuryresultsfromthenegligenceorwilfuldefaultofitself,itsservants,agentsorsubcontractors.
- Information containing any historical information, data or analysis should not be taken as an indication or guarantee of any future performance, analysis, forecast or prediction. Past
performancedoesnotguaranteefutureresults.
- None of the Information constitutes an offer tosell (or a solicitation of an offer to buy), any security, financial product or other investment vehicle or any trading strategy. You cannot
investinanindex.
- MSCIsindirectwhollyownedsubsidiaryInstitutionalShareholderServices,Inc.(ISS)isaRegisteredInvestmentAdviserundertheInvestmentAdvisersActof1940.Exceptwithrespect
to any applicable products or services from ISS (including applicable products or services from MSCI ESG Research Information, which are provided by ISS), neither MSCI nor any of its
products or services recommends, endorses, approves or otherwise expresses any opinion regarding any issuer, securities, financial products or instruments or trading strategies and
neitherMSCInoranyofitsproductsorservicesisintendedtoconstituteinvestmentadviceorarecommendationtomake(orrefrainfrommaking)anykindofinvestmentdecisionand
maynotbereliedonassuch.
- The MSCI ESG Indices use ratings and other data, analysis and information from MSCI ESG Research. MSCI ESG Research is produced by ISS or its subsidiaries. Issuers mentioned or
includedinanyMSCIESGResearchmaterialsmaybeaclientofMSCI,ISS,oranotherMSCIsubsidiary,ortheparentof,oraffiliatedwith,aclientofMSCI,ISS,oranotherMSCIsubsidiary,
includingISSCorporateServices,Inc.,whichprovidestoolsandservicestoissuers.MSCIESGResearchmaterials,includingmaterialsutilizedinanyMSCIESGIndicesorotherproducts,
havenotbeensubmittedto,norreceivedapprovalfrom,theUnitedStatesSecuritiesandExchangeCommissionoranyotherregulatorybody.
- Anyuseoforaccesstoproducts,servicesorinformationofMSCIrequiresalicensefromMSCI.MSCI,Barra,RiskMetrics,ISS,CFRA,FEA,andotherMSCIbrandsandproductnamesare
thetrademarks,servicemarks,orregisteredtrademarksorservicemarksofMSCIoritssubsidiariesintheUnitedStatesandotherjurisdictions.TheGlobalIndustryClassificationStandard
(GICS)wasdevelopedbyandistheexclusivepropertyofMSCIandStandard&Poors.GlobalIndustryClassificationStandard(GICS)isaservicemarkofMSCIandStandard&Poors.
AboutMSCI
MSCI Inc. is a leading provider of investment decision support tools to investors globally, including asset managers, banks, hedge funds and pension funds. MSCI
productsandservicesincludeindices,portfolioriskandperformanceanalytics,andgovernancetools.
Thecompanysflagshipproductofferingsare:theMSCIindiceswithclosetoUSD7trillionestimatedtobebenchmarkedtothemonaworldwidebasis
1
;Barramulti
assetclassfactormodels,portfolioriskandperformanceanalytics;RiskMetricsmultiassetclassmarketandcreditriskanalytics;IPDrealestateinformation,indices
andanalytics;MSCIESG(environmental,socialandgovernance)Researchscreening,analysisandratings;ISSgovernanceresearchandoutsourcedproxyvotingand
reporting services; FEA valuation models and risk management software for the energy and commodities markets; and CFRA forensic accounting risk research,
legal/regulatoryriskassessment,andduediligence.MSCIisheadquarteredinNewYork,withresearchandcommercialofficesaroundtheworld.
1
AsofMarch31,2012,aspublishedbyeVestment,LipperandBloomberginSeptember2012 Jan 2013
Americas
Europe,MiddleEast&Africa AsiaPacific
Americas
Atlanta
Boston
Chicago
Montreal
Monterrey
NewYork
SanFrancisco
SaoPaulo
Stamford
Toronto
1.888.588.4567(tollfree)
+1.404.551.3212
+1.617.532.0920
+1.312.675.0545
+1.514.847.7506
+52.81.1253.4020
+1.212.804.3901
+1.415.836.8800
+55.11.3706.1360
+1.203.325.5630
+1.416.628.1007
CapeTown
Frankfurt
Geneva
London
Milan
Paris
+27.21.673.0100
+49.69.133.859.00
+41.22.817.9777
+44.20.7618.2222
+39.02.5849.0415
0800.91.59.17(tollfree)
ChinaNorth
ChinaSouth
HongKong
Seoul
Singapore
Sydney
Tokyo
10800.852.1032(tollfree)
10800.152.1032(tollfree)
+852.2844.9333
798.8521.3392(tollfree)
800.852.3749(tollfree)
+61.2.9033.9333
+81.3.5226.8222

Barra Predicting Risk at Short Horizons

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Barra Predicting Risk at Short Horizons

Загружено:

Авторское право:

Доступные форматы

msci.

Client Service Information is Available 24 Hours a Day............................26

Variability Conventional Adjusted QStatistic

Вам также может понравиться