2 Bank of Canada WorkingPaperSeptemberWhich Parametric Model for ConditionalSkewnessBruno FeunouMohammad R JahanParvarand Rom ID: 430216
Download Pdf The PPT/PDF document "Working Paper/Document de travail2013Whi..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Working Paper/Document de travail2013Which Parametric Model for ConditionalSkewness?by Bruno Feunou, Mohammad R. JahanParvar and Roméo Tédongap 2 Bank of Canada WorkingPaperSeptemberWhich Parametric Model for ConditionalSkewnessBruno Feunou,Mohammad R. JahanParvarand Roméo TédongapFinancial Markets DepartmentBank of CanadaOttawa, Ontario, Canada K1A 0G9feun@bankofcanada.caOffice of Financial Stability Policy and ResearchFederal Reserve Board of Governorsshington, DC 20551Corresponding author: Mohammad.JahanParvar@frb.govDepartment of FinanceStockholm School of EconomicsStockholm, SwedenRomeo.Tedongap@hhs.se Bank of Canada wor king papers are theoretical or empirical works - in - progress on subjects in economics and finance. The views expressed in this paper are those of the authors.No responsibility for them should be attributed to the Federal Reserve Board of Governors or the ank of Canada. ISSN © 2013Bank of Canada ii AcknowledgementsWe thank two anonymous referees, Chris Adcock (the editor), Torben Andersen, Bruce Hansen, Stanislav Khrapov, Nicola Loperfido, Dilip K. Patro, Akhtar Siddique, Scott Hendry, Glen Keenleyside, seminar participants at Joint Statistical Meetings 2011, Midwest Econometric Group meeting 2011, the Bank of Canada, Wayne State University (mathematics department), OCC, and the MAF 2012 conference for many useful comments. The remaining errors are ours. An earlier version of this paper has been circulated and presented at various seminars and conferences under the title Regime Switchingin the Conditional Skewness of S&P500 Returns. iii AbstractThis paper addresses an existing gap in the developing literature on conditional skewness. We develop a simple procedure to evaluate parametric conditional skewness models. This procedure is based on regressing the realized skewness measures on modelimplied conditional skewness values. We find that an asymmetric GARCHtype specification on shape parameters with a skewed generalized error distribution provides the best insample fit for the data, aswell as reasonable predictions of the realized skewness measure. Our empirical findings imply significant asymmetry with respect to positive and negative news in both conditional asymmetry and kurtosis processes.JEL classification: C22, C51, G12, G15Bank classification: Econometric and statistical methodsRésuméLes auteurs élaborent une procédure dévaluation simple des modèles paramétriques dasymétrie conditionnelle en vue de combler une lacune de la littérature sur le sujetCette procdure est baséesur la régression de lasymétrie réalisée sur lasymétrie conditionnelle. Les auteurs constatent quune spécification de type GARCHasymétrique pour les paramètres de forme, couplée à une distribution derreurs généralisée asymétrique, offre le meilleur ajustement statistique en échantillonainsi quune prévisibilité satisfaisante de la mesure de lasymétrie réaliséeIls notent une importante asymétrie dans leffet des bonnes et des mauvaises nouvelles sur le plan tant de la dynamique de lasymétrie conditionnelle que de celle de laplatissement conditionnelClassification JEL : C22, C51, G12, G15Classification de la Banque :Méthodes économétriques et statistiques 1IntroductionTraditionalmodellingofnancialtimeseriescriticallyreliesontheassumptionofconditionalnormalityofreturns.Thisassumptionimpliesthatconditionalskewnessandexcesskurtosisshouldbeequaltozero.However,empiricalevidenceisinsharpcontrasttothisassertion.Unconditionally,thesemomentsprovenottobezero.Moreover,similartothersttwoconditionalmoments,highermomentsdemonstrateconsiderabletimevariationasnotedbyBekaertetal.(1998)andGhyselsetal.(2011).Thus,explicitmodellingofconditionalhighermomentsthatallowsfortimevariation,isnecessarytoavoidmodelmisspecication.SincethepioneeringworkofHansen(1994),anumberofresearchershaveproposedparametricmodelsforconditionalskewness.Examplesofstudiesontheeconomicimportanceofconditionalskew-nessinnancialassetreturns,itseconometricmodellinganditsempiricalapplicationsincludeHarveyandSiddique(1999,2000),Chenetal.(2001),BrannasandNordman(2003),JondeauandRockinger(2003),Patton(2004),Leonetal.(2005),LanneandSaikkonen(2007),GrigolettoandLisi(2009),Wilhelmsson(2009),Ghyselsetal.(2011),DurhamandPark(2013),Conradetal.(2013),andFeunouetal.(2013).Thereexistanumberofstudiesthatfocusonconditionalkurtosis,amongthemBrooksetal.(2005)andGuidolinandTimmermann(2008).Inthispaperwefocusonconditionalskewness,orconditionalasymmetry.Theexistingresearchdoesnottelluswhichparametricconditionalskewnessmodelprovidesabettertforthedata.AsnotedbyKimandWhite(2004),thisispartiallyduetotheextremesensitivityoftraditionalskewnessandkurtosismeasurestooutliers.1Theyproposeseveralrobustmeasuresforskewnessandkurtosis.Inpreviouswork(Feunouetal.2013),wefoundthatconditionalasymmetryinreturnsisrelatedtothe\relativesemi-variance,"denedastheupsidevarianceminusthedownsidevariance.2Weshowedthatmodellingdownsideriskispossiblewhenameasureofskewnessisexplicitlyincorporatedinthemodel.WeusedPearson's(1895)\modeskewness"asthemeasureofchoice.Pearson'smodeskewnessismorerobusttooutliersthantraditionalskewnessmeasures.BuildinguponandexpandingonsuggestionsinKimandWhite(2004)andFeunouetal.(2013),thispaperllstheexistinggapintheliteratureregardingmodeladequacyforparametricconditionalskewnessmodels.Ourobjectiveisthreefold.First,weestablishthroughapropositionthattherelativesemi-variance, 1Bytraditional,wemeanstandardizedthirdandfourthmomentsofarandomvariable.2Wedenethe\upsidevariance"asthevarianceofthereturnsconditionalupontheirrealizationaboveacertainthreshold.Theirvarianceconditionalupontheirrealizationbelowthesamethresholdiscalled\downsidevariance."Basedonthesetwodenitions,wedenethedierencebetweenupsidevarianceanddownsidevarianceasthe\relativesemi-variance."1 dividedbythetotalvariance,isameasureofskewnessthatsatisesthepropertiesproposedbyGroen-eveldandMeeden(1984)foranyreasonableskewnessmeasure.Second,wedevelopanintuitiveandeasy-to-implementmethodfornon-parametricmeasurementofrealizedasymmetry.Third,basedonthismeasureofrealizedasymmetry,wecantestanyparametricmodelforconditionalskewnessandprovideamethodofhowtomodeltheconditionalskewness.Wetestanumberofparametricmodelsofconditionalasymmetrywithvariousfunctionalanddistributionalassumptions.ThetestingprocedureisbasedonMincerandZarnowitz(1969)regressionsandissimilarinspirittothemethodologydevel-opedbyChernov(2007)toclosetherealized-impliedvolatilitypredictiveregressiongap.Wendthatinadditiontoallowingtimevariationintheconditionalasymmetry,weneedtoallowfora\leverageeect,"butalsoforasymmetry-in-asymmetrytoobtainthebestcharacterizationoftheconditionalskewnessdynamics.3ThemostsuccessfulcharacterizationoftheconditionalasymmetrysharesseveralfunctionalfeatureswiththecelebratedexponentialGARCH(EGARCH)modelofNelson(1991).Ghyselsetal.(2011)proposeamethodologyformodellingandestimatingtheconditionalskewnessbasedonamixeddatasampling(MIDAS)methodofvolatilityestimationintroducedandextensivelystudiedbyGhyselsetal.(2005,2006,2007),andBowley's(1920)measureofskewness.OurworkdiersfromGhyselsetal.intwoimportantdimensions.First,weareinterestedinassessingtheadequacyofdierentmodelsofconditionalasymmetry,whileGhyselsetal.focusonasinglemodel.Second,Ghyselsetal.buildtheirmodelofconditionalasymmetrybasedonBowley's(1920)robustcoecientofskewness.Thismeasureisconstructedusingtheinter-quantilerangesoftheseriesinvestigated,whileourmeasureisbasedonthedierencebetweenupsideanddownsidesemi-variances.DurhamandPark(2013)studythecontributionofconditionalskewnessinacontinuous-timeframe-work.Fortractability,theyassumesimpledynamicsbasedonasingleLevyprocessfortheconditionalskewnessintheirestimatedmodels.Wemodeltheconditionalasymmetryindiscretetimeandassumemuchricherdynamics.DurhamandParkestablishthecostofignoringconditionalhighermomentsinmodellingreturnsdynamics.Thus,theirsimplemodelisadequatetomotivatetheirwork.Theeconomicrelevanceofconditionalasymmetryhasbeenestablishedinseveralassetpricingstudies.AspointedoutbyChristoersenetal.(2006),conditionallynon-symmetricreturninnovationsarecriticallyimportant,sinceinoptionpricing,forexample,heteroskedasticityandtheleverageeectalonedonotsucetoexplaintheoptionsmirk.Inthisstudy,wetrytondthemodelthatbestcharacterizesthedynamicsoftheconditionalasymmetryinS&P500returnsinarichsetofmodels.JondeauandRockinger(2003)characterizethemaximalrangeofskewnessandkurtosisforwhich 3Wendthattheevidenceforasymmetry-in-asymmetryitselfisrelativelyweak.Butthe exibilityoeredbyseparatingthecontributionsofpositiveandnegativeshocksimprovesthemodel'sperformancesignicantly.2 adensityexists.TheyclaimthatthegeneralizedStudent-tdistributionspansalargedomaininthismaximalset,andusethisdistributiontomodelinnovationsofaGARCH-typemodelwithconditionalparameters.Theyndtimedependencyoftheasymmetryparameter,butaconstantdegree-of-freedomparameterintheseriestheystudy.Theyprovideevidencethatskewnessisstronglypersistent,butkurtosisismuchlessso.Whilein uencedbyJondeauandRockinger(2003),ourstudydiersfromtheirworkintwoimportantdimensions.First,westudyalargernumberofmodelsanddistributionsthanJondeauandRockinger(2003),andwethereforeconsiderourstudytobemorecomprehensive.Second,sincewecomparenon-nestedmodels,werelyonMincerandZarnowitz's(1969)methodologytoinvestigatetheadequacyofmodels.Therestofthepaperproceedsasfollows.Insection2,weprovidethetheoreticalbackgroundforourstudy.Wediscusstheimplicationsofvariousdistributionalassumptionsinsection3.Section4describesthedierentparametricmodelspecicationsfortheconditionalskewnessthatwetestinourempiricalanalysis.Wereportourempiricalndingsinsection5.Section6concludes.2TheoreticalBackgroundConventionalasymptotictheoryineconometricstypicallyleadstolimitingdistributionsforeconomicvariablesthatareconditionallyGaussianassamplesizeincreases.ExamplesofsuchworkincludeBollerslevetal.(1994)andDavidson(1994).Thus,conditionalskewnessshouldconvergetozeroassamplesizeincreases.However,asGhyselsetal.(2011),Brooksetal.(2005),andJondeauandRockinger(2003)show,conditionalskewnessformanynancialtimeseriesdoesnotvanishinlargesamplesorthroughsamplingathigherfrequencies.Inwhatfollows,weextendtheirndingsanddevelopatestingframeworktocomparedierentparametricmodelsofconditionalskewness.Ingeneral,commonparametricdistributionsconsideredinempiricalworktocharacterizethedis-tributionoflogarithmicreturnsareunimodalandsatisfythefollowingconditions:Var[rjrm]Var[rjrm],Skew[r]-278;0Var[rjrm]=Var[rjrm],Skew[r]=0Var[rjrm]Var[rjrm],Skew[r]0;(1)wheremisasuitablychosenthreshold.Afewstudiesintheliteratureusedistributionsthatexplicitlyallowforskewnessinreturns;inparticular,theskewedgeneralizedStudent-tdistributionpopularizedbyHansen(1994),theskewedgeneralizederrordistributionofNelson(1991),andthebinormaldistributionappliedtonancialdatainourpreviouswork(Feunouetal.2013)allsatisfyequation(1).Weintroduceanewmeasureofasymmetry,calledtherelativesemi-variance(RSV),denedbythedierencebetween3 theupsidevarianceandthedownsidevariance,whereupsideanddownsidearerelativetoacut-opointequaltothemode.Proposition2.1Lettherandomvariablexfollowaunimodaldistributionwithmodem.Denotetheupsidevarianceas2u=Var[xjxm]andthedownsidevarianceas2d=Var[xjxm].Westandardizetherelativesemi-varianceRSV2u2dbydividingitbythetotalvariancetoobtainascale-invariantanddimensionlessmeasureforskewnessdenedas (x)=2u2d 2;(2)where2=Var[x]isthetotalvariance.Thedistributionisright-skewedif2u2d,andleft-skewedif2u2d.Theproposedskewnessmeasureiscoherent;thatis,itsatisesthethreepropertiesproposedbyGroeneveldandMeeden(1984)thatanyreasonableskewnessmeasureshouldsatisfy.4Thesepropertiesare:(P1)foranya-278;0andb, (x)= (ax+b);(P2)ifxissymmetricallydistributed,then (x)=0;(P3) (x)= (x).Proof:Notethat,foranya-278;0andb,themodeofax+bisequaltoam+b.Besidesthetotalvariance,theupsidevarianceandthedownsidevarianceofax+baregivenbyVar[ax+b]=Var[ax]=a2Var[x]=a22;Var[ax+bjax+bam+b]=Var[ax+bjxm]=Var[axjxm]=a2Var[xjxm]=a22uVar[ax+bjax+bam+b]=Var[ax+bjxm]=Var[axjzm]=a2Var[xjxm]=a22d:Thus,theskewnessofax+bisgivenby (ax+b)=Var[ax+bjax+bam+b]Var[ax+bjax+bam+b] Var[ax+b]=a22ua22d a22=2u2d 2= (x):The (x)skewnessmeasurethussatises(P1).5Todemonstratethat (x)satisesthesecondproperty,supposethatxissymmetricandunimodal;thenweknowthatthemodeisequaltothemean.Asaresult,xmissymmetricandunimodalwithmeanzero.Consequently,xmanditsopposite,mx,havethesamedistribution.Theupsidevarianceofxmisequalto2uandthedownsidevarianceofxmisequalto2d.However,2d=Var[xmjxm0]=Var[mxjxm0]=Var[mxjmx]TJ/;ø 9;.962; Tf; 22.;1 ; Td; [00;0]: 4Suitabilityof (x)asaskewnessmeasurecriticallydependsonthemeasureofvolatilityusedinmodellingthereturnsprocess.Laterinthepaper,weshowthatourresultsarebasedontheEngleandNg(1993)NGARCHvolatilitymodel.Basedonempiricalresults,wearguethatNGARCHisaperfectlyadequatevolatilitymeasureand,thus,ourconditionalskewnessmeasuresarewellspecied.5Thisresultmeansthatrelativesemi-variance,2u2d,satises(P1)uptoamultiplicativeconstant.4 So,2disalsotheupsidevarianceofmx.Sincexmandmxhavethesamedistribution,then2u=2dandinconsequence (x)=0.Thisshowsthatourmeasureofskewnesssatises(P2).Todemonstratethat (x)satises(P3),notethatthemodeofxissimplym.Theupsidevarianceofxisthusthedownsidevarianceofx:Var[xjxm]=Var[xjxm]=Var[xjxm]=2d:Similarly,wecanshowthatthedownsidevarianceofxisequalto2u,theupsidevarianceofx.Ontheotherhand,xandxhavethesametotalvariance2.Consequently,wehave (x)=Var[xjxm]Var[xjxm] Var[x]=2d2u 2=2u2d 2= (x):Ourskewnessmeasurethussatises(P3).2.1BuildingarealizedskewnessmeasureBasedonourdiscussioninsection2,wepositthatmodellingconditionalskewnessorasymmetryisequivalenttomodellingrelativesemi-variance,RSVt.Theliteratureonmodellingandmeasuringvolatilityinnanceandeconomicsisvast.Itsucestosaythat,foryearsnow,usingrealizedvari-ancefollowingthemethodologyofAndersenetal.(2001,2003)isthestandardmethodformeasuringvolatilityinnancialtimeseries.Asanexample,Chernov's(2007)studyontheadequacyofoption-impliedvolatilityinforecastingfuturevolatilitycruciallydependsonthismethodology.Wemodifythisstandardmethodologyintheliteraturetobuildanon-parametricanddistribution-freemeasureforconditionalasymmetryinreturns.Inarecentpaper,Neuberger(2012)discussesasomewhatsimilarmeasureofrealizedskewness.Weconstructourmeasuresfollowingthecommonpracticeintherealizedvarianceliteraturebysummingupnelysampledsquared-returnrealizationsoveraxedtimeinterval,RVt=ntPj=1r2j;t;wheretherearenthigh-frequencyreturnsinperiodt,rj;tisthejthhigh-frequencyreturninperiodt.WethenconstructtherealizeddownsideandupsidevarianceseriesasRVdt=nt 2ndtntXj=1r2j;tI(rj;tmt)andRVut=nt 2nutntXj=1r2j;tI(rj;tmt);(3)wherendtandnutare,respectively,thenumberofhigh-frequencyreturnsbelowandabovetheconditionalmodeofreturnmtinperiodt,andwhereI()denotesanindicatorfunction.Thus,themeasureforrealizedrelativesemi-varianceissimplydenedasRRSVt=RVutRVdt;(4)5 which,dividedbyrealizedvariance,willdenerealizedskewnessaccordingtoourproposedmeasureforskewnessintroducedinProposition2.1.Realizedvolatilitywillrefertothesquarerootofrealizedvariance.Itisawell-knownfactthatEt[RVt+1]=2t,where2t=Vart[rt+1]istheconditionalvariance,andwherert=ntPj=1rj;tisthereturnofperiodt.Thisisawell-establishedresultbasedonCorollary1inAndersenetal.(2003).WeestablishthatEt[RRSVt+1]=2u;t2d;t,andthefollowingpropositionanditsproofshowtheveracityofourassertion.Proposition2.2Lettheunidimensionalcontinuous-priceprocessfPtgTt=0,whereT0,bedenedonacompleteprobabilityspace( ;F;P).LetfFgt2[0;T]Fbeaninformationltration,denedasafamilyofincreasingP-completeandright-continuous-elds.InformationsetFtincludesassetpricesandrelevantstatevariablesthroughtimet.LetRVut+1andRVdt+1bedenedasaboveforthispriceprocess.ThenEtRVut+1RVdt+12u;t2d;t:Proof:Seetheappendix.Thus,asimpletestingprocedureconsistsofregressingRVut+1RVdt+1on2u;t2d;t,orRVut+1RVdt+1=0+12u;t2d;t+"t+1:(5)FollowingthestandardMincerandZarnowitz(1969)methodology,weviewthemodelwiththe0closesttozero,the1closesttooneandthehighestregressionR2asthebettermodelforconditionalskewness.6WehavefollowedBarndor-Nielsenetal.(2010)closelyinourtreatmentofsourcesofconditionalskewness.Thisimpliesthatintheproofoftheaboveproposition,andfollowingBarndor-Nielsenetal.(2010),wehaveshutdownthe\instantaneous"or\high-frequencyleverageeect"toderivethedesiredresults.Inpractice,wehaveshownthatourrealizedrelativesemi-variancesharesmanyfeatureswiththerealizedsemi-variancesstudiedbyBarndor-Nielsenetal.(2010).Wehaveobserved,buthavenotstudiedthisissueindepth,thatindatasampledathighenoughfrequency,conditionalskewnessisdrivenpurelybyjumps.Thisisareasonableassumptionatfrequen-ciessuchas5-minute,15-minute,half-hour,hourly,orevendailysampleddata,sincetheinstantaneousleverageeectisclearlyweakatsuchhighfrequencies.However,asthesamplingfrequencyisloweredto,forexample,monthlyorquarterlyperiods,thisassertionlosespower.Theleverageeectisanimportantcontributortoconditionalskewnessinlowersamplingfrequencies. 6Inpractice,weputlessweightontherstcondition,with0statisticallyindistinguishablefrom0.AsChernov(2007)documentsthisissueinaparallelliterature,pinningdownthecorrectfunctionalformofthestatisticalrelationshipbetweenlatentvariablesisdicult.Thus,wefocusonthemorerobustandtheoreticallymoreimportantrelationshipbetweenourskewnessmeasuresthroughslopeparameters.Thatsaid,wereportresultsforjointtestsfor0=0;1=1inourdiscussionofempiricalndings.6 Figure1showsthepathsfordailyrealizedrelativesemi-varianceandrealizedvarianceforS&P500returnsinthe1980{2010period.Bothseriesareconstructedusing15-minutereturns.Inordertoprovideamoretractablepictureofthebehaviorofrealizedrelativesemi-variance,wethinouttheplotanduseevery22nddatapointinthegure.Westudythecorrelationbetweenrealizedrelativesemi-varianceandrealizedvarianceandvolatility.Wendthatthecorrelationbetweenrealizedvarianceandrelativesemi-varianceis-0.5076,thecorrelationbetweenrealizedvarianceandskewnessis-0.0206,thecorrelationbetweenrealizedvolatilityandrelativesemi-varianceis-0.4475,andnally,thecorrelationbetweenrealizedvolatilityandskewnessis-0.0538.Thus,RRSVandRVarenegativelycorrelated.Itisimmediatelyclearthatspikesinrealizedrelativesemi-variancetypicallyleadtosignicantjumpsinrealizedvariance.Itisalsoclearthatthereisevidenceofclusteringvisibleinrelativesemi-variance,particularlyinthersthalfofthesamplingperiod.Therangeofrealizedrelativesemi-varianceiscomparabletothatreportedinourpreviouswork(Feunouetal.2013).3ModelSpecicationfortheConditionalSkewnessInmodellingthersttwoconditionalmoments,itisnotnecessarytotakeastanceontheparametricdistributionofthereturns.Unlikethersttwoconditionalmomentswherenoassumptionontheparametricdistributionisrequired,modellinghighermomentsrequiresaspecicationofaparametricdistribution.Two exiblefamiliesofdistributionsattractalotofattentionintheliterature.TheyaretheskewedgeneralizedStudent-t(GST)andtheskewedgeneralizederrordistribution(SGED).Wealsostudythebinormaldistribution,whichwerecentlyintroducedtothenanceliterature(Feunouetal.2013).Withoutlossofgenerality,westandardizethesedistributionsbyxingtheirmeantobeequaltozero,andtheirvariancetobeequaltoone.Standardizedreturnsaredenotedbyz.3.1TheskewedgeneralizedStudent-tdistributionHansen(1994)popularizedtheskewedGSTdistribution.ItsdensityisdenedbyfGST(z)=8]TJ ; -1;.93; Td; [00;]TJ ; -1;.93; Td; [00;]TJ ; -1;.93; Td; [00;:bc1+1 2bz+a 12(+1)=2ifza=bbc1+1 2bz+a 1+2(+1)=2ifza=bwhereistheskewnessparameter,representsdegreesoffreedomanda4c2 1;b21+32a2;c((+1)=2) p (2)(=2):Thisdensityisdenedfor21and11:GSTdensitynestsalargesetofconventionaldensities.Forexample,if=0,Hansen'sGSTdistributionreducestothetraditionalStudent-t7 distribution.WerecallthatthetraditionalStudent-tdistributionisnotskewed.Inaddition,if=1,theStudent-tdistributioncollapsestoanormaldensity.Sincecontrolsskewness,ifispositive,theprobabilitymassconcentratesintherighttail.Ifitisnegative,theprobabilitymassisinthelefttail.ItiswellknownthatthetraditionalStudent-tdistributionwithdegreesoffreedomadmitsallmomentsuptotheth.Therefore,giventherestriction2,Hansen'sdistributioniswelldenedanditssecondmomentexists.ThethirdandfourthmomentsofthisdistributionaredenedasEz3=m33am2+2a3=b3;Ez4=m44am3+6a2m23a4=b4;withm2=1+32;m3=16c1+2(2)2 (1)(3)if3;andm4=32 4(1+102+54)if4:Themodeofthedistributionisa=b.Thus,therelativesemi-varianceisVar[zjza=b]Var[zjza=b]=4 b2"14c2(2)2 (1)2#:(6)Wendthatusingrelativesemi-varianceasameasureofskewnessinthecontextofgeneralizedStudent-tdistributionadds exibilitytotheanalysis.NotethatforthethirdmomenttoexistforarandomvariablewithskewedGSTdistribution,mustbegreaterthan3.However,weneedonly2forrelativesemi-variancetoexist,whichisthesameconditionfortheexistenceofthesecondmoment.Thus,itispossibletostudyasymmetryevenwhenthethirdmomentdoesnotexist.3.2TheskewedgeneralizederrordistributionTheprobabilitydensityfunctionfortheSGEDisfSGED(z)=Cexpjz+j [1+sign(z+)]:WedeneC= 21 1;=1 1 23 1 2S()1;=2AS()1;S()=p 1+324A22;andA=2 1 1 23 1 2,where()isthegammafunction.Scalingparametersandaresubjectto0and11.Thisdensityfunctionnestsalargesetofconventionaldensities.Forexample,when=0;wehavethegeneralizederrordistribution,asinNelson(1991).When=0and=2;wehavethestandardnormaldistribution;when=0and=1;wehavethedoubleexponentialdistribution;andwhen=0and=1,wehavetheuniformdistributionontheintervalp 3;p 3:Theparametercontrolstheheightandthetailsofthedensityfunction,andtheskewnessparametercontrolstherateofdescentofthedensityaroundthemode().Thethirdandthefourthmoment8 aredenedasEz3=A333;Ez4=A44A3+62+34;whereA3=41+2(4=)(1=)13andA4=1+102+54(5=)(1=)14:Themodeofthisdistributionis.Asaresult,wendthattherelativesemi-varianceisVar[zjz]Var[zjz]=41A2 S()2:(7)3.3ThebinormaldistributionThebinormaldistributionwasintroducedbyGibbonsandMylroie(1973).Itisananalyticallytractabledistributionthataccommodatesempiricallyplausiblevaluesofskewnessandkurtosis,andneststhefa-miliarnormaldistribution.7Inourpreviouswork(Feunouetal.2013),weshowedthataGARCHmodelbasedonthebinormaldistribution,whichwecallBiN-GARCH,isquitesuccessfulincharac-terizingtheelusiverisk-returntrade-ointheU.S.andinternationalindexreturns.OurBiN-GARCHmodelexplicitlylinksthemarketpriceofrisktoconditionalskewness.Theconditionaldensityfunctionofastandardizedbinormaldistribution(SBin),orbinormaldis-tributionwithzeromeanunitvariance,andPearsonmodeskewness,isgivenbyfSBin(z)=Aexp 1 2z+ d2!I(z)+Aexp 1 2z+ u2!I(z);whered=p /8+p 1(3/81)2andu=p /8+p 1(3/81)2,andwhereA=p 2//(d+u).If=0,thend=u=1,andthisdistributioncollapsestothefamiliarstandardnormaldistribution.Wendthatistheconditionalmode,anduptoamultiplicativeconstant,2dand2uareinterpretedasdownsidevarianceandupsidevariancewithrespecttothemode,respectively.Specically,Var[zjz]=12 2dandVar[zjz]=12 2u:(8)Weconsiderthispropertytobethemostimportantcharacteristicofthebinormaldistribution,givenourobjectives.Theexistenceandpositivityofthequantitiesdanduimposeaboundontheparameter,givenbyjj1.p /211:3236.Finally,itistrivialtoshowthattherelativesemi-variance 7SeeBangertetal.(1986),KimberandJeynes(1987),andTothandSzentimrey(1990),amongothers,forexamplesofusingthebinormaldistributionindatamodelling,statisticalanalysisandrobustnessstudies.9 forthestandardizedbinormaldistributionisVar[zjz]Var[zjz]=p 212 p 1(3/81)2:(9)3.4Theskewness-kurtosisboundaryLet3=E(z3t)and4=E(z4t)denotethenon-centeredthirdandfourthmomentsofarandomvariablefztg1t=0.Foranydistributiononztwith(1;1)support,wehave2341with4-278;0(seeWidder1946,p.134,Theorem12.a;JondeauandRockinger2003).Thisrelationconrmsthat,foragivenlevelofkurtosis,onlyaniterangeofskewnessmaybespanned.Thisisknownastheskewness-kurtosisboundary,anditensurestheexistenceofadensity.Thus,therealchallengewhilemodellingthedistributionofztistogetcloseenoughtothisskewness-kurtosisboundary.Figure2showstheskewness-kurtosisboundaryforHansen'sskewedgeneralizedStudent-t,SGED,andbinormaldistributionsagainstthetheoreticalboundarydiscussedinWidder(1946)andJondeauandRockinger(2003).ItisclearfromthisgurethattheSGEDspansalargerareaofthetheoreticalskewness-kurtosisboundarythandoestheskewedgeneralizedStudent-tdistribution.Thus,weexpectmodelsbasedontheSGEDtooutperformmodelsbasedontheGSTdistribution.Thesharplimitonpermissibleskewnesslevelsinthebinormaldistributionseriouslylimitsitsabilitytospanthetheoreticalskewness-kurtosisboundary.Figures3and4demonstratethecontributionofskewnessandpeakednessparameters,and,re-spectively,togeneratingskewnessandkurtosisinskewedgeneralizedStudent-tdistributionandSGED,respectively.Notethatthepatternsfortheskewnesssurfaceinbothdistributionsareverysimilar.Thedierenceliesinthepermissiblevaluesforandthelevelofskewnessgeneratedbycomparablecombinationsofskewnessandpeakednessparameters.Thatis,theskewedgeneralizedStudent-tseemstogeneratelargervaluesforskewnessincomparisonwithSGED.Thepatternofthekurtosissurface,however,isdierentforthesedistributions.TheskewedgeneralizedStudent-tdistributionsdemonstrateamoreexplosivepatternaswemovetowardthecornersoftheadmissiblesetfor.ThebehaviorofthekurtosissurfaceforalladmissiblevaluesofismoresubduedforSGED.Bothdistributionsshowmildevidenceofasymmetryinkurtosisforlowervaluesof.4ModelSpecicationInthissection,weintroducethefunctionalformsofthemodelsthatwettothedatafortestingpurposes.Werstdiscusshowwemakeourmodelscomparable.Wewanttoestimatetheparameters10 ofinterestwithoutimposingrestrictionsonourestimationprocedure;however,wealsowanttopreservethetheoreticalboundsimposedontheshapeparameters.Weusesign-preservingtransformations.Inparticular,followingHansen(1994)andNelson(1991),becauseofthedierentrestrictionsontheshapeparameters,wemapthetransformedparameterstobeestimatedintothetrueparameters,usingalogisticmappingfortheskewnessandanexponentialmappingforthepeakedness.Thisstepallowsustoestimatethetransformedparameters~and~asfreevalues,andthenrecovertheoriginalparameters.ForthegeneralizedStudent-tdistribution,weusethemappings=1+2 1+exp~and=2+exp(~):(10)Thesetransformationsareintuitive.RecallthatskewedgeneralizedStudent-tdistributionrequiresthatjj1.Transformingfollowingequation(10)ensuresthattheseboundsarepreserved,regardlessoftheestimatedvalueof~.Similarly,GSTrequiresthat21.Equation(10)preservestheselimitsfor,regardlessoftheestimatedvalueof~.Fortheskewedgeneralizederrordistribution,wehave=1+2 1+exp~and=exp(~);(11)soastomaintaintherestrictionsjj1and]TJ/;ø 9;.962; Tf; 10.;Ԗ ; Td; [00;0.Finally,forthestandardizedbinormaldistribution,wehave=1 p 21241+2 1+exp~35:(12)Notethatthebinormaldistributiondoesnothaveadistinctpeakednessparameter.Thetransformationinequation(12)istomaintaintheboundonPearsonmodeskewness,jj1.p /211:3236,forthebinormaldistributiondiscussedimmediatelyafterequation(8).Weestimateandcompareresultsfromuptoninespecicationsforskewnessandpeakednessfactors,acrossthethreedistributionsdiscussedinsection3.Atotalof24modelsttothedata.Inthemostbasicmodel,bothofthesefactorsareconstantparameterstobeestimated.Werelaxthisspecicationandallowfortimevariationandfunctionalcomplexityintheseprocesses.Weassumethatthecondi-tionalvarianceprocessforreturnsfollowsanEngleandNg(1993)NGARCHspecication.Thus,theconditionalvarianceprocessfollows2t+1=0+12t(zt+1)2+22t:(13)Thischoiceoffunctionalformfortheconditionalvarianceallowsfora\leverageeect"inreturns.Weassumethatreturnsfollowrt+1=+tzt+1,wherezt+1jItGST(t;t),zt+1jItSGED(t;t),11 orzt+1jItBiN(t),andwhereItdenotestheinformationsetuptotimet.8Themostbasicmodelthatwestudyassumesconstant~and~.WecallthismodelM0.WerelaxtheassumptionoftimeinvariancefortheskewnessprocessinModel1,butmaintainthatthepeakednessprocessisstillaparametertobeestimated.InmodelM1,weassumethattheskewnessprocessfollowsasymmetricARCH(1)process,~t+1=0+1zt+1:(14)Theassumptionofsymmetryheremeansthatthearrivalofgoodorbadnewsimpactstheskewnessprocesswiththesamemagnitude.Asymmetryinvolatilityisawell-documentedfeatureofnancialdata.Severalstudiesinthe(G)ARCHliteratureaddressthisissue,whichleadstotheleverageeectinnancialdata.Amongthesestudies,wenoteNelson(1991),Glostenetal.(1993),andEngleandNg(1993).JondeauandRockinger(2003)arguethatasymmetryinARCHforskewnessandkurtosisrequiresinvestigation.Hence,westudythisissueinmodelM2,whereweassumeanAsym-ARCHstructureintheskewnessprocess,~,butassumeconstant~.Inthismodel,~follows~t+1=0+1;+zt+1I(zt+10)+1;zt+1I(zt+10);(15)whereI()denotesanindicatorfunction.WeallowforthericherGARCH(1,1)dynamicsintheskewnessprocessinmodelM3.Inthismodel,westillassumeconstant~.~follows~t+1=0+1zt+1+2~t:(16)Thismodel,exceptfordistributionalassumptionsandnormalization,isthesamemodelstudiedbyHarveyandSiddique(1999).Tostudythepotentialexistenceofanasymmetry-in-asymmetryorleverageeectintheconditionalskewness,weassumethattheskewnessprocessfollowsanasymmetricGARCHforminmodelM4.Inthismodel,~isstillassumedtobeconstant.Asym-GARCHin~impliesthefollowingfunctionalform:~t+1=0+1;+zt+1I(zt+1]TJ/;ø 9;.962; Tf; 10.;Ԗ ; Td; [00;0)+1;zt+1I(zt+10)+2~t:(17)Intheremainingfourmodels,werelaxtheassumptionofconstant~.InmodelM5,weassumea 8Tokeepournotationconsistentwiththerealizedvolatilityliterature,ourtimingconventiondiersslightlyfromthefamiliarGARCHnotation.Throughoutthepaper,thesubscripttonanyvariablemeansthatitisobservedexactlyattimet.InthetraditionalGARCHnotation,thesubscripttintheconditionalvariancemeansthatitisthevarianceofthetimetreturns.Hence,thevarianceisobservedattimet1.12 symmetricARCHstructureforboth~and~:~t+1=0+1zt+1;(18)~t+1= 0+ 1zt+1:InmodelM6,westudytheimpactofgoodandbadnewsonbothskewnessandpeakednessprocessesbyassuminganasymmetricARCHinboth~and~:~t+1=0+1;+zt+1I(zt+10)+1;zt+1I(zt+10);(19)~t+1= 0+ 1;+zt+1I(zt+1]TJ/;ø 9;.962; Tf; 10.;Ԗ ; Td; [00;0)+ 1;zt+1I(zt+10):ModelM7investigatestheimplicationsofassumingaGARCHspecicationforbothskewness,~,andpeakedness,~,processes:~t+1=0+1zt+1+2~t;(20)~t+1= 0+ 1zt+1+ 2~t:AswithM3,thisformulationisverysimilartothemodelinHarveyandSiddique(1999).Finally,westudytheimplicationsofassuminganasymmetricGARCHfunctionalformforboth~and~inmodelM8:~t+1=0+1;+zt+1I(zt+1]TJ/;ø 9;.962; Tf; 10.;Ԗ ; Td; [00;0)+1;zt+1I(zt+10)+2~t(21)~t+1= 0+ 1;+zt+1I(zt+1]TJ/;ø 9;.962; Tf; 10.;Ԗ ; Td; [00;0)+ 1;zt+1I(zt+10)+ 2~t:5EstimationResults5.1DataWeusedailyStandardandPoor's500(S&P500)indexexcessreturnsfromThomsonReutersDatas-tream.ThedataseriesstartsinJanuary1980andendsinSeptember2010.Table1reportssummarystatisticsofthedata.InPanelA,wereportannualizedreturnmeansandstandarddeviations,inpercentages,incolumnstwoandthree.Wereportunconditionalskewnessincol-umnfour.Weobservenegativeunconditionalskewnessformarketreturns.Thevalueofunconditionalskewnessisnotsmallrelativetotheaveragedailyreturns.Ourdataseemtobehighlyleptokurtotic,sincetheseriesdemonstratessignicantunconditionalexcesskurtosis,asseenincolumnve.There-portedp-valuesofJarqueandBera's(1980)normalitytestimplyasignicantdeparturefromnormalityinthedata.Ourproxyfortherisk-freerateistheyieldofthe3-monthconstant-maturityU.S.Treasurybill,whichweobtainedfromtheFederalReserveBankofSt.LouisFREDIIdatabank.Thecrashof13 October1987,theAsiancrisisof1997,theRussiandefaultof1998andthe2007{09GreatRecessionarerepresentedinthedata.OurintradaydataseriescomesfromOlsenFinancialTechnologiesandistheirlongestavailableone-minutecloselevelS&P500indexpriceseries.ThedataspantheperiodfromFebruary1986toSeptember2010.Toreducethemarketmicrostructureeectinourempiricalresults,weconstructintradayreturnsata15-minutefrequency.TheseresultsappearinPanelBofTable1.5.2DiscussionoftheresultsWereporttwosetsofresultsforeachdistribution.First,wereportestimationresultsforallmodelsstudiedforeachdistribution.WethenreporttheresultsfromrunningtheMincerandZarnowitz(1969)regressionsofrealizedskewnessmeasureonparametricskewnessresults.Wereportestimationresultsforthefullsample(1980{2009)andthreesubsamples,spanning1980{89,1990{99,and2000-09.Formodelselection,wemainlyrelyonempiricalndingsbasedonourfull-sampleestimates.Estimatedresultsbasedonsubsamplesaregenerallyfordemonstrationoftimevariationorrobustnessandplayasecondaryrole.Ineachcase,foreachdistributionandforeachsampleperiod,werstidentifythemodelthatbestcharacterizesthereturns.Ourstrategytoidentifysuchamodelistocomputelikelihoodratio(LR)teststatistics.Amodelisconsideredviablewhen(i)theLRtestrejectsequalgoodnessoftbetweenabaselinenormallydistributedmodelandthemodelwithconditionalskewness,and(ii)theLRtestrejectsequalgoodnessoftbetweeneachmodelandthemodelprecedingit.9Thatis,wecontrolforoverparameterizationofmodelsbycomparingmodelssequentially.TheLRtestisnotapplicablefornon-nestedmodels.Insuchcases,wedonotcomputeLRteststatistics.IftheLRtestdoesnotdierentiatebetweentwomodels'toritisinapplicable,thenwerstlookattheBayesianinformationcriteria(BIC)ofmodels.Ifthisisstillnothelpful,wepreferthelessparameterizedmodeloverthemoreparameterizedmodelwiththesamegoodnessoftandsimilarBIC.Basedontheestimatedparametersofthepreferredparametricmodel,welteroutthemodel-impliedrelativesemi-varianceprocess.Notethat,inourstudy,modelM8nestsallothermodels.Wethenregressrealizedrelativesemi-varianceontomodel-impliedrelativesemi-variance.Next,weturntomodelevaluation.Ourcriteriaforthesuccessofaparametricmodelofskewness(relativesemi-variance)here{indescendingorderofimportance{are:(i)aslopeparameterthatisstatistically 9Thismeansthat,inourtables,thereportedLRteststatisticsarecomputedbyusinglog-likelihoodvaluesfrommodelMxandmodelMx1.Thatis,theLRstatisticformodelM2isbasedonlog-likelihoodvaluesformodelsM1andM2.Whennecessary,weindicatethatwehaveusedlog-likelihoodvaluesfromnon-sequentialbutnestedmodelstoconstructateststatistic.14 indistinguishablefromunityatreasonablecondencelevels,(ii)thehighestpossibleR2giventhatthepreviousconditionismet,andnally(iii)whethertheslopeandinterceptcoecientsoftheMincer-Zarnowitzregressioninquestionarejointlyequaltooneandzero,respectively,giventhattheprevioustwoconditionsaremet.Thelastcriterionre ectsthedicultyofpinningdownthecorrectfunctionalformofalatentvariable,inourcaseconditionalskewness.Thus,weputlessweightoninterceptestimatesthatarestatisticallynotdierentfromzero.Chernov(2007)documentsthisissueforrealizedandimpliedvolatilityliterature.Tables2and3reporttheestimationresultsformodelsinsection4fortheGSTdistribution.Notethatthebaselinemodel(representedasNinbothtables)isrejectedinfavorofallalternativemodelsentertainedinourstudy,basedonlikelihoodratioteststhatarenotreportedtosavespace.ModelsM0{M8deliverabettertforthedata.Thus,therstcriterionismet.FormodelM0,wendthatestimatedpeakednessparametersarestatisticallydierentfromzeroatthe5percentlevelacrossallsamplesstudied.However,theskewnessparameterisstatisticallysignicantinthefullsampleandthe2000{09subsample.Weconcludethatthereisverystrongevidenceinfavoroftheexistenceofexcesskurtosisandconvincingevidencesupportingskewness,whichgetsstrongerasweusethe21st-centurydata.10ModelsM1{M4showasomewhatsimilarpattern.Whilethesizeoftheestimatedparametersmaydieracrossthesemodelsandacrosssamples,theyareallsignicantlydierentfromzeroatthe5percentlevel.Estimatedautoregressiveparameters,whethersymmetricorasymmetric(1+or1),andtheGARCH-likeparameter2aregenerallystatisticallydierentfromzero.Inaddition,weobserveevidenceofasymmetry-in-asymmetryinourresults.Thatis,positivereturnsincreasetheskewness{or,inotherwords,pushtheskewnesstowardpositivevalues,andnegativereturnsdecreasetheskewness{orpushskewnesstowardnegativevalues.Thus,positiveandnegativenewshavedierentandoppositeimpactsonskewness.Basedonthesizeoftheestimatedparameters,theimpactofpositivenewsislargerthanthatofnegativenews.ModelsM5toM8inTables2and3reporttheimpactoftimevariationontheconditionalkurtosis.Ingeneral,theGARCH-typeparameter 2andtheARCH-typeparameter 1;+arestatisticallysignif-icantatthe5percentlevel.Weobserveevidenceofanasymmetricimpactingoodandbadnewsforconditionalkurtosismodels.Theestimated 1;parameterisstatisticallydierentfromzerointhefullsample,butthisisnottrueingeneralforsubsampleresults.Inaddition,themagnitudeofestimated 1;+isgenerallylargerthanthatof 1;parameters.Thus,itseemsthat,inthiscase,goodnewsis 10PleasenotethatinTables2,3,5,6,and8,weabusethenotationtosavespace.Thatis,1+and 1+representboththecoecientsfortheARCH-liketermsandthecoecientsforpositiveinnovationshocks.15 moreimportantthanbadnewsforconditionalkurtosisdynamics.Thelikelihoodteststatistics(henceforthLRstatistics)showthatthemodelsstudieddieronhowwelltheytthedata.Overall,modelM8,whichallowsforasymmetricGARCH-typedynamicsforboth~tand~t,performsthebestbasedonLRtests.Thatis,basedonLRtestscarriedout,M8providesabettertthanmodelsM0toM7inthefullsampleandthe1990{99and2000{09subsamplesforreturns.Table4reportstheresultsfromrunningtheMincer-Zarnowitzpredictiveregressionsintroducedinequation(5),wheretheright-hand-sidevariablesareimpliedskewnessmeasuresfromoneofthemodelsinequations(14)-(21)andtheleft-hand-sidevariableistherealizedskewnessmeasureintroducedinequation(4).TheRRSVmeasureisbasedonhigh-frequencyinformation,whileimpliedskewnessmea-suresarebasedondailydata.NotethatmodelM1providesthehighestR2infull-sampleestimation,buttheestimated1issignicantlydierentfromunity.However,thebestmodelshouldbetheonewiththehighestR2amongallspecicationswith1notsignicantlydierentfromunity.OurresultssuggestthatmodelsM4andM8bothhave1notstatisticallydierentfromone,andthehighestR2of19percentinfull-sampleestimation.Boththesespecicationshavetime-varyingconditionalskew-nessdynamicsfeaturingasymmetry-in-asymmetry,butM8nestsM4throughtheconditionalkurtosisdynamics.Forallmodels,thehypothesis1=1isnotrejectedinthesecondsubsampleexceptforspecicationsM0andM7,whilethejointhypothesis0=0and1=1isnotrejectedinthethirdsubsampleexceptforspecicationsM0,M4andM6.OurpreferredmodelwouldthenbeM8basedonfull-sampleandsubsampleMincer-Zarnowitzpredictiveregressionresults,addedtothefactthatM8issignicantlyfavoredoverM4giventheLRtestsreportedinTables2and3.Tables5and6reporttheestimationresultsformodelsM0{M8whenerrorsareSGED.SimilartowhatisseenintheresultsfortheGSTcase,modelsM0{M8performbetterthanthebaselinemodelNacrosstheboard.ModelsM0{M4demonstratestatisticallysignicantpeakednessparameterestimatesacrossmodelsandsamples.Thestatisticalevidenceinfavorofasymmetry-in-asymmetryisquitestronginthefullsample.Statisticalevidencesupportingasymmetry-in-asymmetryismoremixedinthesubsamples.OurestimationresultsformodelsM5{M8implythat,again,evidenceinsupportofasymmetry-in-asymmetryintheconditionalskewnessisquitestrongforthefullsampleandinthesubsamples.Inaddition,thereissignicantsupportforasymmetryintheconditionalkurtosis,wheregoodnewsreducesfat-tailedness.Thisisnotsurprising.Arrivalof\goodnews"shouldreducemarketuncertainty,andhenceconditionalkurtosis.BasedontheLRstatistics,againmodelM8isthepreferredmodelinthe16 fullsampleandacrossthesubsamples.Itappearsasifallowingforarich,asymmetricparameterizationaddstomodel exibilityandhenceitsperformance.ItisworthnotingthatmodelM8sharesmanyfeatureswiththecelebratedNelson(1991)EGARCHmodel.Wendthatinthe1990{99and2000{09subsamples,modelsM6andM8areindistinguishableusingLRtests.BasedonBIC,modelM6ispreferredtoM8inthesesubsamples.However,asmentionedearlier,webaseourjudgmentmainlyonfull-sampleresults,wheretheLRtestclearlypicksM8overM6andM7.Table7reportstheresultsforMincer-Zarnowitzpredictiveregressionswhereimpliedrelativesemi-varianceestimatesarebasedonmodelswithSGEDerrors.Inthefullsample,modelsM0{M2,M5andM8allhaveslopeparametersthatarestatisticallyindistinguishablefromone.Jointtestsofbothslopeandinterceptparametersbeingdierentfromoneandzero,respectively,cannotberejectedfortheabovemodelsatthe5percentcondencelevelforthefullsample,andmodelsM1andM5providethehighestR2of24percent.ThesespecicationshaveARCH-typetime-varyingconditionalskewnessdynamicsincommon,butM5nestsM1throughtheconditionalkurtosisdynamics.However,neitherofthesetwomodelsispreferredinanyofthesubsamples,sincethehypothesis1=1isrejectedatconventionallevelsofcondence.Again,asmentionedearlier,webaseourjudgmentmainlyonfull-sampleresults,andmodelM1wouldbethepreferredspecication,sinceitismoreparsimoniousandslightlyfavoredoverM5basedonLRtestsandBICvaluesinTables5and6.However,notethatthejointhypothesis0=0and1=1isnotrejectedformodelM8,neitherinthefullsamplewithanR2of14percentnorinthesecondandthirdsubsampleswithR2sof7percentand16percent,respectively.Intherstsubsample,M8isalsothepreferredmodel,withanR2of22percentamongallspecicationswherethehypothesis1=1isnotrejected.ThenM8wouldbethebestmodel,shouldweextendourmodel-selectioncriteriatosubsampleregressionresultsandtheneedforttingtheexcess-returnsdataaswell.WereporttheestimationresultsforthebinormaldistributionmodelinTable8.Sincethestandard-izedbinormaldistributionisaone-parameterdistribution,itdoesnotallowforindependentestimationofconditionalkurtosis.Thus,weestimateonlymodelsM0{M4forthecaseofstandardizedbinormaldistributederrors.Similartothepreviouscases,wendsignicantsupportingevidenceforasymmetry-in-asymmetry.Almostallestimatedparametersarestatisticallydierentfromzero.BasedontheLRstatistics,M3isthepreferredmodelforthefullsample.ItallowsforsymmetricGARCH-typedynamicsintheconditionalskewnessmeasure.Inthe1980{89subsample,M3remainsthemodelofchoice.Inthe1990{99subsample,M4istiedwithM2basedontheLRtest,butM2ispreferredtoM4basedonBIC.Inthe2000{09subsample,nomodelperformsbetterthanM0,whichallowsforconstantand.17 Table9reportsmodelevaluationresultsthroughMincer-Zarnowitzpredictiveregressions,formodelsM0{M4withstandardizedbinormallydistributederrors.Acrossthefullsample,modelsM1,M2andM4haveestimated1parametersthatarestatisticallyindistinguishablefromunity.M1andM2havethehighestR2of25percentforthefullsample.Onceweconsidersubsamples,modelsM1andM4demonstratesimilarperformance;bothmodelsdonotrejectthehypothesis1=1inallsubsamplesexcept1980{89.However,basedonourresultsreportedinTable8,weknowthatmodelM1ispreferredovermodelM4inttingtheexcess-returnsdata.Thus,basedonMincer-Zarnowitzpredictiveregressionconsiderations,wechooseM1,whichallowsforasymmetricARCH-typedynamicsintheconditionalskewness.Note,however,thatalthoughmodelM3hasanestimated1parameterof0.9262thatisstatisticallydierentfromunityinthefullsample,thisvalueremainsquiteclosetoone.Furthermore,1parameterestimatesarenotstatisticallydierentfromoneformodelM3inallsubsamples.Giventheseobservations,onceweaccommodatetheneedforttingtheexcess-returnsdataaswell,wemaybetemptedtochooseM3overM1basedonLRtestingandsimilarpredictiveresults.11Comparingourresultsforthefullsampleacrossthethreedistributionalassumptions,weobservethefollowing:(i)GST-basedestimationresultsimplythatModel8,whichallowsforasymmetryinbothconditionalskewnessandkurtosis,performsbetterthanalternativemodelsstudiedincharac-terizingexcess-returnsdynamicsandinMincer-Zarnowitzpredictiveregressionsofparametricontonon-parametricmeasuresofrelativedownsidesemi-variances.(ii)SGED-basedestimationresultsim-plythatifwewantbothreasonablecharacterizationofexcess-returnsdynamicsandstrongpredictiveresultsformeasuringRRSV,thenourchoiceisM8,whichallowsforasymmetryinbothconditionalskewnessandkurtosis.However,ifweareinterestedonlyinpredictivepowerfortheRRSV,thenM1isthebetterchoice.(iii)Binormaldistribution-basedestimationresults,similartotheSGEDcase,leadustochoosetheparsimoniousM1forpredictivepurposesonly.Oncewedecidetocharacterizeexcess-returnsdynamicsaswell,thenwechoosethemore exibleM3,whichallowsforsymmetricGARCH(1,1)-typedynamicsintheskewnessprocess.Wederiveourtheoreticalframeworkfromtheliteratureonrealizedvolatility.Thus,inourtestingprocedures,wefollowthatliteraturecloselyinthatweareconcernedwithin-samplepredictabilityandnotout-of-sampleforecastability.Thus,itappearsthat,strictlyspeaking,modelM8withSGEDerrorsprovidesestimated1sthatarestatisticallyclosetounity,deliversreasonableR2sandprovidesverygoodin-samplecharacterizationofexcess-returnsdynamics.WeconcludethatamodelthatallowsforasymmetricGARCH(1,1)-typedynamicsandaSGEDistheoverallpreferredmodel.Inaddition,this 1122criticalvaluesat5percentand1percentcondencelevelsare5.99and9.21,respectively,whilethecomputedLRteststatisticisover11.18 modelprovidesstrongstatisticalsupportforasymmetry-in-asymmetryandasymmetryinconditionalkurtosis.Inquiteafewcases,wecouldnotrejectastatisticallysignicantpositivedierencebetweeninterceptparameters0and0.Whiletheupwardbiasobservedacrossmodelsanddistributionsmayseemundesirableatrstglance,wekeepinmindthatthevastliteratureontheinformationcontentofimpliedvolatilityfacesthesametypeofproblem.Chernov(2007)addressesthebiasissuebyaddingasuitablepredictortotheright-handsideofthepredictiveregression.Itispossiblethat,byaddressingthepossibilityofamissingvariable,predictiveregressionsmayimprove.However,sinceM8coupledwithanSGEDworkswellandislargelyfreeofthisbias,itisquitepossiblethatthisbiasisduetodistributionalassumptions.Wewilladdressthisinterestingissueinfutureresearch.6ConclusionsInthispaper,weaddressanimportantgapintherecentandgrowingliteratureonconditionalskew-ness.Inrecentyears,anincreasingnumberofstudieshaveaddressedtheimportanceofmodellinghighermomentsandhowtheirdynamicsaectnancialrisk-managementprocedures.Weproposeamethodologytoassessthemodeladequacyofthegrowingnumberofproposedparametricmeasuresoftheconditionalskewness.Weshowthat,theoretically,conditionalskewnessmatterswhenaseriesdemonstratessignicantrelativesemi-variance.Wethenproposeasimpleandintuitivenon-parametricmeasureofrelativesemi-variance,whichwecallrealizedrelativesemi-variance,inthespiritofthesuccessfulliteratureonrealizedvariance.Weshowthat,undermildregularityconditions,parametricmodelsofconditionalasymmetryapproximatetherealizedrelativesemi-variance.Wealsoshowthatourproposedmeasureandmethodologyhaveclosetiestotheliteratureonrealizedvariance,ontheonehand,andtotherecentdevelopmentsonrealizedsemi-varianceontheother.Wethenstudyseveralparametricmodelsofconditionalskewness.Sincetheperformanceoftheparametricmodelsofskewnesscruciallydependsonthedistributionalassumptions,westudyarangeofrelevantdistributionalmodels,includingtheskewedgeneralizedStudent-tofHansen(1994)andJondeauandRockinger(2003),skewedGED,andbinormaldistributioninourpreviouswork(Feunouetal.2013).Wendstatisticallysignicantevidenceinsupportofasymmetry-in-asymmetry,ortheso-calledleverageeectintheconditionalskewness,andstrongsupportforasymmetric(G)ARCH-typedynamicsforbothskewnessandpeakednessprocessesintheconditionaldistribution.Weconcludethat,basedonourresults,thebestmodelforconditionalskewnessisonethatallowsforasymmetric19 GARCH(1,1)dynamicsandadmitsskewedGEDerrors,appliedtobothskewnessandpeakednessprocessesintheconditionaldistributionofreturns.Thismodelsharesmanyfeatureswiththewell-knownEGARCHmodelofNelson(1991).20 ReferencesAndersen,T.G.,Bollerslev,T.,Diebold,F.X.,Labys,P.,2001.Thedistributionofrealizedexchangeratevolatility.JournaloftheAmericanStatisticalAssociation96,42{55.Andersen,T.G.,Bollerslev,T.,Diebold,F.X.,Labys,P.,2003.Modelingandforecastingrealizedvolatility.Econometrica71,579{625.Bangert,U.,Goodhew,P.J.,Jeynes,C.,Wilson,I.H.,1986.LowEnergy(2-5keV)ArgonDamageinSilicon.JournalofPhysicsD:AppliedPhysics19,589{603.Barndor-Nielsen,O.E.,Kinnebrock,S.,Shephard,N.,2010.VolatilityandTimeSeriesEconometrics:EssaysinHonorofRobertF.Engle.OxfordUniversityPress,Ch.Measuringdownsiderisk:realisedsemivariance,pp.117{136.Bekaert,G.,Erb,C.,Harvey,C.R.,Viskanta,T.,1998.DistributionalCharacteristicsofEmergingMarketReturnsandAssetAllocation.JournalofPortfolioManagement,102{116.Bollerslev,T.,Engle,R.F.,Nelson,D.B.,1994.ARCHModels.Vol.4ofTheHandbookofEconomet-rics.North-Holland,Amsterdam,pp.2959{3038.Bowley,A.,1920.ElementsofStatistics.CharlesScribnersSons,NewYork,NY.Brannas,K.,Nordman,N.,2003.Conditionalskewnessmodellingforstockreturns.AppliedEconomicsLetters10(11),725{728.Brooks,C.,Burke,S.P.,Heravi,S.,Persand,G.,2005.AutoregressiveConditionalKurtosis.JournalofFinancialEconometrics3(3),399{421.Chen,J.,Hong,H.,Stein,J.,2001.Forecastingcrashes:tradingvolume,pastreturns,andconditionalskewnessinstockprices.JournalofFinancialEconomics61,345{381.Chernov,M.,2007.OntheRoleofRiskPremiainVolatilityForecasting.JournalofBusiness&Eco-nomicStatistics25(4),381{411.Christoersen,P.,Heston,S.,Jacobs,K.,2006.Optionvaluationwithconditionalskewness.JournalofEconometrics131(12),253{284.Conrad,J.,Dittmar,R.,Ghysels,E.,2013.Ex-AnteSkewnessandExpectedStockReturns.JournalofFinance68(1),85{124.21 Davidson,J.,1994.StochasticLimitTheory,AnIntroductionforEconometricians,1stEdition.Ad-vancedTextsinEconometrics.OxfordUniversityPress,Oxford,U.K.Durham,G.B.,Park,Y.-H.,2013.BeyondStochasticVolatilityandJumpsinReturnsandVolatility.JournalofBusinessandEconomicStatistics31(1),107{121.Engle,R.F.,Ng,V.K.,1993.MeasuringandTestingtheImpactofNewsonVolatility.JournalofFinance48(5),1749{1778.Feunou,B.,Jahan-Parvar,M.R.,Tedongap,R.,2013.ModelingMarketDownsideVolatility.ReviewofFinance17(1),443{481.Ghysels,E.,Plazzi,A.,Valkanov,R.,2011.ConditionalSkewnessofStockMarketReturnsinDevelopedandEmergingMarketsanditsEconomicFundamentals.WorkingPaper,Kenan-FlaglerBusinessSchool-UNC,andRadySchoolofBusiness-UCSD.Ghysels,E.,Santa-Clara,P.,Valkanov,R.,2005.ThereisaRisk-ReturnTrade-OAfterAll.JournalofFinancialEconomics76,509{548.Ghysels,E.,Santa-Clara,P.,Valkanov,R.,2006.Predictingvolatility:gettingthemostoutofreturndatasampledatdierentfrequencies.JournalofEconometrics131,59{95.Ghysels,E.,Sinko,A.,Valkanov,R.,2007.MIDASregressions:furtherresultsandnewdirections.EconometricReviews26(1),53{90.Gibbons,J.F.,Mylroie,S.,1973.EstimationofImpurityProlesinIon-ImplantedAmorphousTargetsUsingJoinedHalf-GaussianDistributions.AppliedPhysicsLetters22,568{569.Glosten,L.R.,Jagannathan,R.,Runkle,D.E.,1993.OntheRelationBetweenExpectedValueandtheVolatilityoftheNominalExcessReturnonStocks.JournalofFinance48(5),1779{1801.Grigoletto,M.,Lisi,F.,2009.Lookingforskewnessinnancialtimeseries.EconometricsJournal12(2),310{323.Groeneveld,R.,Meeden,G.,1984.Measuringskewnessandkurtosis.TheStatistician33,391{399.Guidolin,M.,Timmermann,A.,2008.InternationalAssetAllocationunderRegimeSwitching,SkewandKurtosisPreferences.ReviewofFinancialStudies21,889{935.22 Hansen,B.E.,1994.AutoregressiveConditionalDensityEstimation.InternationalEconomicReview35(3),705{730.Harvey,C.R.,Siddique,A.,1999.AutoregressiveConditionalSkewness.JournalofFinancialandQuantitativeAnalysis34,465{488.Harvey,C.R.,Siddique,A.,2000.ConditionalSkewnessinAssetPricingTests.JournalofFinance55,1263{1295.Jarque,C.M.,Bera,A.K.,1980.EcientTestsforNormality,HomoscedasticityandSerialIndepen-denceofRegressionResiduals.EconomicsLetters6(3),255{259.Jondeau,E.,Rockinger,M.,2003.ConditionalVolatility,Skewness,andKurtosis:Existence,Persis-tence,andComovements.JournalofEconomicDynamicsandControl27,1699{1737.Kim,T.-H.,White,H.,2004.OnMoreRobustEstimationofSkewnessandKurtosis.FinanceResearchLetters1(1),56{73.Kimber,A.C.,Jeynes,C.,1987.AnApplicationoftheTruncatedTwo-PieceNormalDistributiontotheMeasurementofDepthsofArsenicImplantsinSilicon.JournaloftheRoyalStatisticalSociety:SeriesB36,352{357.Lanne,M.,Saikkonen,P.,2007.Modelingconditionalskewnessinstockreturns.EuropeanJournalofFinance13(8),691{704.Leon,A.,Rubio,G.,Serna,G.,2005.Autoregressiveconditionalvolatility,skewnessandkurtosis.TheQuarterlyReviewofEconomicsandFinance45,599{618.Mincer,J.,Zarnowitz,V.,1969.JacobMincer(editor),EconomicForecastsandExpectations.NationalBureauofEconomicResearch,NewYork,NY,Ch.TheEvaluationofEconomicForecasts.Nelson,D.B.,1991.ConditionalHeteroskedasticityinAssetReturns:ANewApproach.Econometrica59,347{370.Neuberger,A.,2012.Realizedskewness.ReviewofFinancialStudies25(11),3423{3455.Patton,A.J.,2004.Ontheout-of-sampleimportanceofskewnessandasymmetricdependenceforassetallocation.JournalofFinancialEconometrics2(1),130{168.23 Pearson,K.,1895.Contributionstothemathematicaltheoryofevolution.II.Skewvariationinhomo-geneousmaterial.PhilosophicalTransactionsoftheRoyalSocietyofLondon.A186,343{414.Protter,P.,1992.StochasticIntegrationandDierentialEquations.Springer-Verlag,Berlin.Toth,Z.,Szentimrey,T.,1990.TheBinormalDistribution:ADistributionforRepresentingAsymmet-ricalbutNormal-LikeWeatherElements.JournalofClimate3(1),128{137.Widder,D.,1946.TheLaplaceTransform.PrincetonUniversityPress,Princeton,NJ.Wilhelmsson,A.,2009.Valueatriskwithtimevaryingvariance,skewnessandkurtosis{theNIG-ACDmodel.EconometricsJournal12(1),82{104.24 Appendix:ProofofProposition(2.2)Theset-upissimilartothatofAndersenetal.(2003)(henceforth,ABDL(2003)).Proposition1ofABDL(2003)permitsauniquecanonicaldecompositionofthelogarithmicassetpriceprocessp=(p(t))t2[0;T],p(t)=p(0)+A(t)+M(t);whereAisanitevariationandpredictablemeancomponent,andMisalocalmartingale.Letr(t;h)=p(t)p(th)denotethecontinuouslycompoundedreturnover[th;t].Thecumulativereturnsprocessfromt=0onward,r=(r(t))t2[0;T],isthenr(t)r(t;t)=p(t)p(0)=A(t)+M(t):Furtherassumethatthemeanprocess,fA(s)A(t)gs2[t;t+h],conditionaloninformationattimetisapredeterminedfunctionover[t;t+h]:AnimmediateresultofthisassumptionistheCorollary1,andhenceequation(6)ofABDL(2003),whichestablishesthatVar(r(t+h;h)jFt)=E[r;r]t+h[r;r]tjFt:Inotherwords,theconditionalvarianceequalstheconditionalexpectationofthequadraticvariationofthereturnsprocess.Denotethemodeofthedistributionofr(t+h;h)conditionalonFtbym(t;h).Leti(t+h;h)=1[r(t+h;h)m(t;h)]denoteanindicatorrandomprocessthattakes1ifthereturnsbetween[t;t+h]aregreaterthanorequaltotheconditionalmodem(t;h):Sincethemeanprocess,fA(s)A(t)gs2[t;t+h],conditionaloninformationattimet,isapredeter-minedfunctionover[t;t+h];withoutlossofgenerality,weomitAfromthispointonwards.25 Wethushave2u(t;h)Var(r(t+h;h)jFt;i(t+h;h)=1)=Var(M(t+h)M(t)jFt;i(t+h;h)=1)=Var(M(t+h)jFt;i(t+h;h)=1)=EM(t+h)2jFt;i(t+h;h)=1fE(M(t+h)jFt;i(t+h;h)=1)g2=E[M;M]t+hjFt;i(t+h;h)=1fE(M(t+h)jFt;i(t+h;h)=1)g2=E[M;M]t+hi(t+h;h) (t;h)jFtEM(t+h)i(t+h;h) (t;h)jFt2;where(t;h)Pr[i(t+h;h)=1jFt]:UsingCorollary3inChapterII.6ofProtter(1992),whereitisshownthatEM(t+h)2=E[M;M]t+h,wehave2d(t;h)Var(r(t+h;h)jFt;i(t+h;h)=1)=E[M;M]t+h(1i(t+h;h)) 1(t;h)jFtEM(t+h)(1i(t+h;h)) 1(t;h)jFt2;and2u(t;h)2d(t;h)=E[M;M]t+hi(t+h;h) (t;h)[M;M]t+h(1i(t+h;h)) 1(t;h)jFt+EM(t+h)(1i(t+h;h)) 1(t;h)jFt2EM(t+h)i(t+h;h) (t;h)jFt22u(t;h)2d(t;h)=1 (t;h)(1(t;h))E[M;M]t+h(i(t+h;h)(t;h))jFt1 (t;h)(1(t;h))E(M(t+h)(i(t+h;h)(t;h))jFt)EM(t+h)(t;h)(t;h)i(t+h;h)+i(t+h;h)i(t+h;h)(t;h) (t;h)(1(t;h))jFt:SinceE[i(t+h;h)jFt]=(t;h),wehave2u(t;h)2d(t;h)=1 (t;h)(1(t;h))E[M;M]t+h[M;M]t(i(t+h;h)(t;h))jFt1 (t;h)(1(t;h))E((M(t+h)M(t))(i(t+h;h)(t;h))jFt)EM(t+h)(12(t;h)) (t;h)(1(t;h))(i(t+h;h)(t;h))+2jFt26 2u(t;h)2d(t;h)=1 (t;h)(1(t;h))E[M;M]t+h[M;M]t(i(t+h;h)(t;h))jFt1 (t;h)(1(t;h))E((M(t+h)M(t))(i(t+h;h)(t;h))jFt)(12(t;h)) (t;h)(1(t;h))E((M(t+h)M(t))(i(t+h;h)(t;h))jFt)+2M(t)2u(t;h)2d(t;h)=1 (t;h)(1(t;h))E[r;r]t+h[r;r]t(i(t+h;h)(t;h))jFt1 (t;h)(1(t;h))E(r(t+h;h)(i(t+h;h)(t;h))jFt)(12(t;h)) (t;h)(1(t;h))E(r(t+h;h)(i(t+h;h)(t;h))jFt)+2M(t)((t;h)(1(t;h)))2u(t;h)2d(t;h)=E[r;r]t+h[r;r]t(i(t+h;h)(t;h))jFtE(r(t+h;h)(i(t+h;h)(t;h))jFt)(12(t;h)) (t;h)(1(t;h))E(r(t+h;h)(i(t+h;h)(t;h))jFt)+2M(t)((t;h)(1(t;h)))2u(t;h)2d(t;h)=cov[r;r]t+h[r;r]t;i(t+h;h)jFtcov(r(t+h;h);i(t+h;h)jFt)(12(t;h)) (t;h)(1(t;h))cov(r(t+h;h);i(t+h;h)jFt)+2M(t):Hence,twocomponentsdrivetheconditionalskewness.Therstcomponentistheconditionalcovariancebetweenreturnsandtheindicatorvariablethatshowswhetherthemarkethasanupwardordownwardmovement.Thesecondcomponentistheconditionalcovariancebetweenthequadraticvari-ationofreturnsandthesameindicatorvariable.Fromthispointon,wefocusonthelattercomponent,since,asweshowlaterinthisappendix,itisdrivenbyjumpsinthereturns.Theformercomponentistheskewnessinducedbytheinstantaneouscorrelationbetweenlog-returnsandvolatility:2u(t;h)2d(t;h)1 (t;h)(1(t;h))cov[r;r]t+h[r;r]t;i(t+h;h)jFt:27 Lety(t+h;h)y(1)(t+h;h);y(2)(t+h;h)0[r;r]t+h[r;r]t;i(t+h;h)0:FromCorollary1ofABDL(2003),wehavecov[r;r]t+h[r;r]t;i(t+h;h)jFt=Ehy(1);y(2)it+hhy(1);y(2)itjFt:Inotherwords,theconditionalcovariancebetweenquadraticreturnvariationandtheupsidein-dicatorequalstheconditionalexpectationofthequadraticcovariationbetweenthequadraticreturnvariationandtheupsideindicator.Proposition2ofABDL(2003)providesaconsistentestimatorofy(1);y(2)t:Recallthatproposition:Foranincreasingsequenceofrandompartitionsof[0;T],0=m;0m;1;suchthatsupj1(m;j+1m;j)!0andsupj1m;j!Tform!1withprobabilityone,wehavelimm!18:Xj1hy(1)(t^m;j)y(1)(t^m;j1)ihy(2)(t^m;j)y(2)(t^m;j1)i9=;!hy(1);y(2)it;andthuswehavelimm!18:jt+hXj=jt[r;r]m;j[r;r]m;j1(i(m;j)i(m;j1))9=;!hy(1);y(2)it+hhy(1);y(2)it;wherejtinfm;jt(j),jt+hsupm;jt+h(j)andi(t)i(t;t)=1[r(t)(t;0)]:Thus,Pjt+hj=jt[r;r]m;j[r;r]m;j1(i(m;j)i(m;j1))isameasureofrealizedskewness.Wenowsimplifythederivedrealizedskewnessexpression[r;r]m;j[r;r]m;j1(p(m;j)p(m;j1))2;andwecanshowthati(m;j)i(m;j1)=1[p(m;j)p(m;j1)0]1[p(m;j)p(m;j1)0];sothatthemeasureofrealizedskewnessbecomesjt+hXj=jt(p(m;j)p(m;j1))21[p(m;j)p(m;j1)0]1[p(m;j)p(m;j1)0]:Thus,usingBarndor-Nielsenetal.'s(2010)(henceforth,BKS(2010))notation,wehaveshownthattherealizedskewnessisthedierencebetweentheupsiderealizedsemi-variance(RS+)andthedownsiderealizedsemi-variance(RS):28 RS+(t+h;h)=jt+hXj=jt(p(m;j)p(m;j1))21p(m;j)p(m;j1)0RS(t+h;h)=jt+hXj=jt(p(m;j)p(m;j1))21p(m;j)p(m;j1)0:Torecap,wehave2u(t;h)2d(t;h)1 (t;h)(1(t;h))ERS+(t+h;h)RS(t+h;h)jFt:FurtherusingBKS(2010)results,wecanprovidemoreinsightintothesourcesofconditionalskewness.BKS(2010)showthatRS+(t+h;h)RS(t+h;h)!t+hXt(ps)2[1ps01ps0];wherepsp(s)p(s)isthejumpattimes:Hencepositiveskewnessisdrivenbythefactthatpositivejumpamplitudesarehigherthannegativeamplitudes,whereasnegativeskewnessisdrivenbythefactthatnegativejumpamplitudesarehigherthanpositiveamplitudes.29 Figure1:Realizedvolatilityandrealizedrelativesemi-varianceforS&P500returns,1986{2009 Inthisgureweplotrealizedvarianceandrealizedrelativesemi-varianceforS&P500returns,sampledat15-minutefrequency.Toprovideamoreinformativeillustration,weplotdailyrealizedvolatilitiesandtheendofthemonthrealizedrelativesemi-variance(every22ndobservationisplotted)values.Figure2:Skewness-KurtosisBoundaryforSkewedGeneralizedStudent-t,SkewedGEDandBinormalDistributions Thisguredepictstheskewness-kurtosisboundariesforHansen(1994)skewedgeneralizedStudent-tdistributionandskewedGEDagainstthetheoreticalboundaryimpliedbyTheorem12.a,page134ofWidder(1946).Thetheoreticalboundaryisobtainedfrom2341with4-322;0,where3=E(z3t)and4=E(z4t)arethenon-centeredthirdandfourthmomentsofarandomvariablefztg1t=0.30 1986 1990 1993 1997 2001 2006 2009 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 Daily Realized Volatility From 15-minute Returns 1986 1991 1995 1999 2004 2009 -6 -4 -2 0 2 4 6 Daily Realized Skewness From 15 minutes Returns 0 5 10 15 20 25 30 35 -6 -4 -2 0 2 4 6 Kurtosis FrontierBinormal Distribution Skewed Generalized Error Distribution Figure3:ImpactofSkewnessandPeakednessParametersonSkewnessandKurtosisSurfacesforSkewedGeneralizedStudent-tDistribution Thisgureshowsthecontributionoftheskewness,,andpeakednessparameter,,ingeneratingskewness(toppanel)andkurtosis(bottompanel)surfacesfortheskewedgeneralizedStudent-tdistributionofHansen(1994).Therangeofparametersischosensuchthattheydonotviolatetheskewness-kurtosisboundary.Figure4:ImpactofSkewnessandPeakednessParametersonSkewnessandKurtosisSurfacesforSkewedGeneralizedErrorDistribution Thisgureshowsthecontributionoftheskewness,,andpeakednessparameter,,ingeneratingskewness(toppanel)andkurtosis(bottompanel)surfacesfortheskewedgeneralizederrordistribution(SGED).Therangeofparametersischosensuchthattheydonotviolatetheskewness-kurtosisboundary.31 1 0.5 0 0.5 1 3 3.5 4 4.5 5 30 20 10 0 10 20 30 and 1 0.5 0 0.5 1 4 4.2 4.4 4.6 4.8 5 0 50 100 150 200 250 and 1 0.5 0 0.5 1 0 1 2 3 4 5 5 0 5 and 1 0.5 0 0.5 1 0 1 2 3 4 5 0 5 10 15 20 25 30 35 40 and Table1:SummaryStatisticsoftheData PanelA:DescriptiveStatistics,ExcessReturns ReturnSeries Mean(%)Std.Dev.(%)SkewnessKurtosisJ-Bp-Value S&P500 3.4821.94-1.2431.870.01 PanelB:DescriptiveStatistics,Relativedownsidesemi-variance RSV(%)Std.Dev.(%)SkewnessKurtosisJ-Bp-Value S&P500(NP) 1.848.711.2116.360.01 Thetoppanelofthistablereportssummarystatisticsofexcessreturns.Calculationofthereturnsisbasedonsubtractingthedaily3-monthU.S.Treasurybillratefromthelogdierenceofthemarkettotalreturnindex.Meanofexcessreturnsandstandarddeviationsarereportedasannualizedpercentages.Excesskurtosisvaluesarereported.Thecolumntitled\J-Bp-Value"reportsp-valuesofJarqueandBera(1980)testsofnormalityinpercentages.Thebottompanelreportsthecomputedstatisticsoftheobservedrelativesemi-variance(RSV)inthedata.ReportedRSVisbasedonthemeandierencebetweenltereddownsideandupsidesemi-variances.TherstcolumnreportstheannualizedRSV,andthesecondcolumnisthestandarddeviationofthisquantity.Duetoavailabilityofhigh-frequencydataforS&P500returns,wereportthenon-parametricestimateforrelativedownsidesemi-variancefortheUnitedStatesdenotedasS&P500(NP),basedon15-minutereturns.ThesampleperiodisJanuary1980toSeptember2010.Source:ThomsonReutersDatastreamandFREDIIdatabankattheFederalReserveBankofSt.Louis.32 Table2:EstimationResultsfortheConditionalSkewnessDynamicswithGeneralizedStudent-tDistribution(1) NM0M1M2M3M4M5M6M7M8 1980-2009 0-0.1047*-0.1050*-0.1147*-0.0327*-0.0594*-0.1109*-0.1246*-0.0324*-0.0655*(0.0298)(0.0296)(0.0428)(0.0136)(0.0282)(0.0297)(0.0435)(0.0128)(0.0286)1+0.1070*0.1222*0.1080*0.1511*0.1144*0.1211*0.1208*0.1584*(0.0253)(0.0544)(0.0224)(0.0459)(0.0260)(0.0486)(0.0220)(0.0443)10.0981*0.0814*0.0840*0.0716*(0.0385)(0.0312)(0.0426)(0.0353)20.6593*0.6590*0.6694*0.6643*(0.0955)(0.0915)(0.0823)(0.0902) 0-1.4635*-1.4464*-1.4450*-1.4660*-1.4663*-1.4762*-1.0870*-0.7289*-0.4997*(0.1159)(0.1364)(0.1368)(0.1365)(0.1361)(0.1567)(0.1594)(0.1737)(0.1947) 1+-0.5805*-1.0077*-0.6848*-0.9800*(0.1898)(0.1191)(0.1283)(0.1092) 10.2368*0.1909*(0.0943)(0.0762) 20.5119*0.4309*(0.1097)(0.1158) LogL24639.424837.724846.024846.124852.424853.124849.124867.824861.624878.7LRStat396.60*16.60*0.20NA1.40NA37.40*NA34.20*BIC-3.2494-3.2732-3.2731-3.2720-3.2728-3.2717-3.2724-3.2725-3.2716-3.2715 1980-1989 0-0.0157-0.0129-0.0307-0.0022-0.0133-0.0110-0.0678-0.0018-0.0326(0.0546)(0.0517)(0.0732)(0.0168)(0.0407)(0.0512)(0.0731)(0.0165)(0.0493)1+0.08530.11200.0864*0.10560.07830.14460.0760*0.1282(0.0444)(0.0902)(0.0341)(0.0728)(0.0462)(0.0947)(0.0357)(0.0845)10.06700.07670.00730.0481(0.0706)(0.0469)(0.0646)(0.0598)20.6741*0.6702*0.6777*0.6617*(0.1566)(0.1557)(0.1681)(0.1757) 0-1.8280*-1.7839*-1.7848*-1.7891*-1.7887*-1.7515*-1.5049*-0.0407-0.6140(0.2053)(0.2007)(0.2006)(0.2012)(0.2011)(0.2128)(0.2418)(0.0219)(0.4645) 1+0.0591-0.6017*0.0510-0.5835*(0.0941)(0.1869)(0.0294)(0.1680) 10.21160.1201(0.1321)(0.1017) 20.9779*0.5330*(0.0121)(0.2576) LogL8190.18308.48310.18310.28312.48312.58310.38313.88314.28316.4LRStat236.60*3.400.20NA0.20NA7.00*NA4.40BIC-3.2268-3.2674-3.2650-3.2619-3.2628-3.2597-3.2620-3.2571-3.2573-3.2519 Thistablereportstheestimatedparametersof~and~asinequation(10),andwhereinnovationsfollowaskewedGSTprocessasinHansen(1994).Nrepresentsthebaselinenormallydistributederrorscase.ModelsM0toM8representequations(14)to(21),respectively.Tosavespace,wereportestimatedvaluesfor1and1+onthesamerow.Estimatedstandarderrorsarereportedinparentheses,belowtheestimatedparameters.\LogL"representsthecomputedloglikelihoodfunction.Wereportlikelihoodratioteststatistics,\LRStat,"withrespecttotheprecedingmodel.\NA"impliesthatthemodelsarenon-nested,andthustheLRteststatisticisnotcomputed.Allmodelsarepreferredtothebenchmarknormalmodel,basedonthelikelihoodratiotest.Thus,wedonotreporttheseteststatistics.*representsrejectionofthenullhypothesisinquestionatthe5%condencelevel.\BIC"representsBayesianinformationcriteriafortheestimatedmodels.33 Table3:EstimationResultsfortheConditionalSkewnessDynamicswithGeneralizedStudent-tDistribution(2) NM0M1M2M3M4M5M6M7M8 1990-1999 0-0.0643-0.0582-0.1636*-0.0090-0.0788-0.0527-0.1666*-0.0022-0.1037(0.0515)(0.0516)(0.0769)(0.0156)(0.0538)(0.0508)(0.0770)(0.0154)(0.0533)1+0.1503*0.3335*0.1488*0.2739*0.1572*0.3567*0.1614*0.3524*(0.0457)(0.1101)(0.0377)(0.0982)(0.0415)(0.1040)(0.0339)(0.0957)10.05960.09490.06860.0946(0.0683)(0.0509)(0.0696)(0.0528)20.7029*0.6455*0.6945*0.5963*(0.1141)(0.1341)(0.0990)(0.1210) 0-1.5922*-1.5691*-1.5289*-1.6366*-1.6053*-1.6505*-1.1899*-0.6799*-0.1532(0.2022)(0.2221)(0.2297)(0.2164)(0.2221)(0.2274)(0.2842)(0.1876)(0.2080) 1+-0.7164*-1.1082*-0.7516*-1.1332*(0.2075)(0.2077)(0.1571)(0.1588) 10.20260.1884(0.2201)(0.1723) 20.6061*0.6387*(0.0957)(0.0920) LogL8628.38687.38692.88694.68697.68698.88696.68702.78706.68714.3LRStat118.00*11.00*3.60NA2.40NA12.20*NA15.40*BIC-3.3989-3.4161-3.4151-3.4127-3.4139-3.4113-3.4135-3.4097-3.4113-3.4081 2000-2009 0-0.2392*-0.2404*-0.1073-0.1069-0.1131-0.2593*-0.0932-0.0897-0.1127(0.0530)(0.0530)(0.0835)(0.0972)(0.1023)(0.0530)(0.0805)(0.0877)(0.1090)1+0.0762-0.09450.0818-0.09770.0359-0.16220.0570-0.1522(0.0521)(0.0983)(0.0509)(0.1025)(0.0491)(0.0904)(0.0422)(0.0935)10.2391*0.2392*0.2398*0.2475*(0.0952)(0.0959)(0.0820)(0.1006)20.5504-0.03060.6375-0.1292(0.3974)(0.2909)(0.3412)(0.2886) 0-0.2152*-0.2614*-0.3293*-0.2537*-0.3307*0.3538*0.2914*-0.0378-0.4276(0.0546)(0.0560)(0.0741)(0.0625)(0.0713)(0.1140)(0.1363)(0.2236)(0.4257) 1+-2.0265*-1.9913*-1.9479*-1.7840*(0.2790)(0.2322)(0.2298)(0.2403) 10.7292*-4.0525(0.1026)(3.5040) 20.4158*0.3005(0.0875)(0.1588) LogL7866.47890.57891.77893.87892.17893.97910.27916.27916.57919.5LRStat48.20*2.404.20NA3.60NA12.00*NA6.00*BIC-3.1109-3.1143-3.1116-3.1094-3.1087-3.1063-3.1159-3.1121-3.1122-3.1072 Thistablereportstheestimationresultsforthe1990{99and2000{09subsamples.RefertothenotestoTable2formoreinformation.34 Table4:EvaluatingtheConditionalRelativeDownsideVariancewithGeneralizedStudent-tDistribution M0M1M2M3M4M5M6M7M8 1980-2009 0-3.016E-06-1.16E-05*-1.16E-05*-1.81E-05*-1.72E-05*-1.19E-05*-1.33E-05*-1.96E-05*-1.91E-05*(2.229E-06)(1.86E-06)(1.87E-06)(2.73E-06)(1.94E-06)(1.87E-06)(1.85E-06)(1.94E-06)(1.93E-06)12.4336*1.5751*1.5727*0.9441*1.02501.5274*1.4075*0.8826*0.9819(0.0839)(0.0338)(0.0345)(0.0251)(0.0277)(0.0338)(0.0304)(0.0244)(0.0497) R20.12260.26550.25680.19050.18580.25380.26310.17880.1877JT292.89*309.64*295.44*48.692*79.044*284.57*231.66*125.66*98.71* 1980-1989 0-1.37E-06-8.69E-07-8.90E-08-9.59E-06-6.36E-06-3.08E-07-2.74E-06-1.09E-05-7.90E-06(1.17E-05)(7.96E-06)(8.07E-06)(9.40E-06)(6.64E-06)(5.70E-06)(5.53E-06)(6.72E-06)(6.50E-06)12.0431*1.7049*1.7341*0.8763*1.00701.6928*1.4798*0.8189*0.9870*(0.2120)(0.0534)(0.0556)(0.0414)(0.0468)(0.1086)(0.0450)(0.0402)(0.0395) R20.08680.51060.49880.31450.32150.05430.52540.29800.3460JT24.22*174.24*174.30*9.974*0.938162.75*113.90*22.091*1.585 1990-1999 01.88E-06-1.33E-05*-1.32E-05*-1.34E-05*-1.32E-05*-1.35E-05*-1.38E-05*-1.47E-05*-1.47E-05*(4.33E-06)(3.20E-06)(3.21E-06)(3.05E-06)(2.17E-06)(2.26E-06)(2.27E-06)(2.12E-06)(2.14E-06)13.2425*1.12761.13821.02241.04221.08471.03660.9304*1.0414(0.3154)(0.1386)(0.1387)(0.1015)(0.1061)(0.1355)(0.1352)(0.0964)(0.1005) R20.04030.02560.02600.03880.03690.02480.02280.03570.0318JT50.740*18.136*17.94*19.27*37.18*36.18*37.25*48.48*47.01* 2000-2009 0-1.55E-06-4.05E-06-8.24E-06-7.45E-06-8.15E-06*-3.63E-07-9.52E-06*-2.20E-06-7.14E-06*(5.05E-06)(4.88E-06)(4.72E-06)(4.88E-06)(3.33E-06)(3.53E-06)(3.37E-06)(3.51E-06)(3.37E-06)11.2052*1.08080.95790.94890.96371.07910.8153*1.01920.9872(0.0501)(0.0426)(0.0366)(0.0394)(0.0367)(0.0428)(0.0326)(0.0410)(0.0362) R20.18800.20460.21510.18820.21650.20230.20060.19800.2113JT16.91*4.2964.3794.0126.693*3.42240.18*0.6114.613 ThistablereportsMincer-ZarnowitzregressionresultsfromRVdt+1(h)RVut+1(h)=0+1(Vart[rt+1jrt+1mt]Vart[rt+1jrt+1-295;mt]),whereerrortermsintheparametricmodelfollowaskewedStudent-tdistribution.*indicatesrejectionofthenullhypothesisthati=i;i=0;1atthe5%condencelevelorbetter.JTrepresentsthevalueoftheteststatisticbuiltunderthenullhypothesisthat0and1arejointlyequalto0and1,respectively.*andydenotetherejectionofthisnullhypothesisatthe5and10%condencelevels,respectively.Thecriticalvaluesforthistestare5.991and4.605,respectively,basedon2df=2.35 Table5:EstimationResultsfortheConditionalSkewnessDynamicswithSkewedGED(1) NM0M1M2M3M4M5M6M7M8 1980-2009 0-0.1347*-0.1317*-0.1355*-0.0450*-0.0764*-0.1316*-0.1370*-0.0449*-0.0997*(0.0293)(0.0270)(0.0346)(0.0150)(0.0257)(0.0268)(0.0277)(0.0150)(0.0273)1+0.1218*0.1281*0.1234*0.1717*0.1221*0.1244*0.1236*0.1944*(0.0211)(0.0402)(0.0186)(0.0393)(0.0185)(0.0293)(0.0151)(0.0305)10.1188*0.0948*0.1195*0.0751*(0.0314)(0.0323)(0.0129)(0.0274)20.6347*0.6226*0.6355*0.5666*(0.0830)(0.0714)(0.0799)(0.0936) 00.3199*0.3215*0.3216*0.3174*0.3168*0.3221*0.4620*0.3184*0.3730*(0.0070)(0.0071)(0.0071)(0.0071)(0.0070)(0.0225)(0.0229)(0.0573)(0.0454) 1+0.0034-0.2414*-0.0024-0.2345*(0.0186)(0.0373)(0.0216)(0.0386) 10.1280*0.1268*(0.0177)(0.0216) 2-0.00530.2492*(0.1669)(0.1053) LogL24639.424824.024838.724838.724848.124848.924838.724859.024848.124869.7LRStat369.20*29.40*0.00NA1.60NA40.60*NA43.20*BIC-3.2494-3.2714-3.2722-3.2710-3.2722-3.2711-3.2710-3.2713-3.2698-3.2704 1980-1989 0-0.0767-0.0525-0.0193-0.0202-0.0030-0.0419-0.0958-0.0199-0.0761(0.0513)(0.0505)(0.0695)(0.0176)(0.0294)(0.0382)(0.0724)(0.0144)(0.0392)1+0.1111*0.07170.1146*0.0881*0.0856*0.1085*0.0962*0.1565*(0.0395)(0.0559)(0.0255)(0.0241)(0.0399)(0.0430)(0.0237)(0.0421)10.1424*0.1263*0.02790.0334(0.0518)(0.0279)(0.0434)(0.0506)20.5797*0.5741*0.5613*0.5535*(0.1352)(0.1435)(0.1610)(0.2190) 00.3210*0.2508*0.2507*0.2527*0.2510*0.2743*0.3873*0.13480.2560*(0.0113)(0.0090)(0.0089)(0.0090)(0.0090)(0.0377)(0.0499)(0.1313)(0.0941) 1+0.0648*-0.1384*0.0423-0.1223*(0.0290)(0.0662)(0.0259)(0.0645) 10.1563*0.1291*(0.0379)(0.0355) 20.50720.4137(0.4725)(0.2611) LogL8190.18290.88295.38295.68299.08299.28298.38303.68301.38305.8LRStat201.40*9.00*0.60NA0.40NA10.60*NA9.00*BIC-3.2268-3.2604-3.2591-3.2561-3.2575-3.2544-3.2572-3.2531-3.2522-3.2478 Thistablereportstheestimatedparametersof~and~asinequation(11),andwhereinnovationsfollowaskewedGED.Nrepresentsthebaselinenormallydistributederrorscase.ModelsM0toM8representequations(14)to(21),respectively.Tosavespace,wereportestimatedvaluesfor1and1+onthesamerow.Estimatedstandarderrorsarereportedinparentheses,belowtheestimatedparameters.\LogL"representsthecomputedloglikelihoodfunction.Wereportlikelihoodratioteststatistics,\LRStat,"withrespecttotheprecedingmodel.\NA"impliesthatthemodelsarenon-nested,andthustheLRteststatisticisnotcomputed.Allmodelsarepreferredtothebenchmarknormalmodel,basedonthelikelihoodratiotest.Thus,wedonotreporttheseteststatistics.*representsrejectionofthenullhypothesisinquestionatthe5%condencelevel.\BIC"representsBayesianinformationcriteriafortheestimatedmodels.36 Table6:EstimationResultsfortheConditionalSkewnessDynamicswithSkewedGED(2) NM0M1M2M3M4M5M6M7M8 1990-1999 0-0.0767-0.0845*-0.1917*-0.0121-0.0977*-0.0827-0.1934*-0.0123-0.1486*(0.0518)(0.0462)(0.0571)(0.0182)(0.0435)(0.0469)(0.0650)(0.0097)(0.0347)1+0.1631*0.3541*0.1614*0.3099*0.1632*0.3862*0.1607*0.4361*(0.0387)(0.0571)(0.0374)(0.0722)(0.0346)(0.0673)(0.0142)(0.0385)10.07570.1046*0.1095*0.1175*(0.0582)(0.0381)(0.0456)(0.0282)20.7051*0.5881*0.7141*0.4931*(0.1130)(0.1464)(0.0617)(0.0656) 00.3210*0.3154*0.3218*0.2896*0.2987*0.3054*0.4211*0.5355*0.2802*(0.0123)(0.0121)(0.0124)(0.0116)(0.0117)(0.0390)(0.0512)(0.0743)(0.0592) 1+-0.0412-0.2250*-0.0459*-0.2925*(0.0388)(0.0692)(0.0212)(0.0609) 10.07910.1180*(0.0624)(0.0428) 2-0.8665*0.5614*(0.0555)(0.1392) LogL8628.38684.18692.98695.48699.58701.18693.48699.58702.28710.0LRStat111.60*17.60*5.00NA3.20NA12.20*NA15.60*BIC-3.3989-3.4148-3.4151-3.4131-3.4147-3.4122-3.4123-3.4085-3.4096-3.4064 2000-2009 0-0.2635*-0.2622*-0.1409-0.1136-0.1457-0.2552*-0.1326*-0.0878-0.1361(0.0492)(0.0455)(0.0767)(0.0951)(0.0924)(0.0508)(0.0661)(0.0650)(0.1137)1+0.0798-0.06950.0857-0.07130.0536-0.1378*0.0779-0.1522*(0.0451)(0.0791)(0.0478)(0.0617)(0.0668)(0.0337)(0.0615)(0.0531)10.2265*0.2271*0.1903*0.2181*(0.0997)(0.0922)(0.0734)(0.0667)20.5586-0.02390.6498*-0.1111(0.3577)(0.2454)(0.2383)(0.1505) 00.4713*0.4646*0.4589*0.4655*0.4588*0.4678*0.6886*0.2559*0.6867*(0.0199)(0.0197)(0.0193)(0.0197)(0.0193)(0.0418)(0.0403)(0.1258)(0.0939) 1+-0.1262*-0.4590*-0.1286*-0.4716*(0.0515)(0.0624)(0.0474)(0.0640) 10.2288*0.2476*(0.0297)(0.0277) 20.44270.0302(0.2648)(0.1816) LogL7866.47893.47895.07896.97895.67896.97898.17919.57899.97919.6LRStat54.00*3.203.80NA2.60NA42.80*NA39.40*BIC-3.1109-3.1154-3.1130-3.1106-3.1101-3.1075-3.1111-3.1134-3.1056-3.1072 Thistablereportstheestimationresultsforthe1990{99and2000{09subsamples.RefertothenotestoTable5formoreinformation.37 Table7:EvaluatingtheConditionalRelativeDownsideVariancewithSkewedGED M0M1M2M3M4M5M6M7M8 1980-2009 0-2.99E-06-2.36E-06-2.44E-06-9.72E-06*-7.80E-06*-2.41E-064.17E-06*-9.80E-06*2.02E-06(3.25E-06)(2.79E-06)(2.80E-06)(2.80E-06)(2.03E-06)(1.97E-06)(2.09E-06)(1.98E-06)(2.26E-06)11.00330.97020.96140.7091*0.6974*0.96940.6680*0.7071*0.9604(0.0380)(0.0226)(0.0226)(0.0186)(0.0191)(0.0225)(0.0167)(0.0186)(0.0423) R20.10370.23520.23200.19480.18110.23590.21000.19420.1376JT0.8562.4683.698256.62*265.82*3.343399.22*273.80*0.886 1980-1989 0-1.65E-061.61E-05y1.63E-05y1.21E-066.81E-061.59E-05*3.54E-05*1.07E-063.91E-05*(1.23E-05)(8.85E-06)(8.90E-06)(9.65E-06)(6.99E-06)(6.24E-06)(6.75E-06)(6.84E-06)(1.17e-05)10.845251.2084*1.2036*0.7324*0.7736*1.2065*0.9007*0.99281.0314(0.1094)(0.0447)(0.0450)(0.0357)(0.0391)(0.0444)(0.0358)(0.0357)(0.0311) R20.05770.42850.42260.30130.28600.43060.39300.29960.2199JT2.02125.07*23.80*56.20*34.49*28.16*32.19*0.06612.18* 1990-1999 02.58E-06-7.73E-06*-7.66E-06*-8.75E-06*-7.50E-06*-7.80E-06*-4.78E-06-8.76E-06*-1.94E-06(4.61E-06)(3.58E-06)(3.58E-06)(3.31E-06)(3.43E-06)(5.05E-06)(4.13E-06)(3.31E-06)(4.59E-06)11.3910y0.7908y0.7889y0.7286*0.7109*0.7880*0.95370.7280*0.9849(0.2051)(0.1210)(0.1209)(0.097)(0.0970)(0.1210)(0.0950)(0.0970)(0.0690) R20.03530.03230.03270.04320.04130.03250.02590.04320.0244JT3.9467.636*7.622*14.87*13.74*7.820*1.57514.95*0.228 2000-2009 0-1.35E-06-1.17E-05*-1.19E-05*-1.56E-05*-1.58E-05*-1.18E-05*-6.39E-06-1.57E-05*-8.05E-06(5.24E-06)(4.89E-06)(4.90E-06)(4.85E-06)(4.92E-06)(4.89E-06)(5.13E-06)(4.85E-06)(5.34E-06)11.09320.7952*0.7859*0.6909*0.6314*0.7942*1.05120.91560.9531(0.0705)(0.0500)(0.0500)(0.0460)(0.0440)(0.0500)(0.0517)(0.0465)(0.0520) R20.16120.16540.16360.15490.14140.16530.15350.15480.1606JT1.81422.22*24.14*56.33*80.56*22.49*3.50513.89*3.085 ThistablereportsMincer-ZarnowitzregressionresultsfromRVdt+1(h)RVut+1(h)=0+1(Vart[rt+1jrt+1mt]Vart[rt+1jrt+1-295;mt]),whereerrortermsintheparametricmodelfollowaskewedStudent-tdistribution.*indicatesrejectionofthenullhypothesisthati=i;i=0;1atthe5%condencelevelorbetter.JTrepresentsthevalueoftheteststatisticbuiltunderthenullhypothesisthat0and1arejointlyequalto0and1,respectively.*andydenotetherejectionofthisnullhypothesisatthe5and10%condencelevels,respectively.Thecriticalvaluesforthistestare5.991and4.605,respectively,basedon2df=2.38 Table8:EstimationResultsfortheConditionalSkewnessDynamicswithBinormalDistribu-tion NM0M1M2M3M4 1980-2009 0-0.2435*-0.1853*-0.1989*-0.0955*-0.1127*(0.0329)(0.0299)(0.0438)(0.0237)(0.0380)1+0.2055*0.2342*0.1819*0.2143*(0.0272)(0.0729)(0.0252)(0.0607)10.1948*0.1687*(0.0375)(0.0338)20.4187*0.4128*(0.0960)(0.0953) LogL24639.424667.024692.324692.424697.824698.1LRStat55.20*50.60*0.20NA0.60BIC-3.2494-3.2530-3.2552-3.2540-3.2547-3.2536 1980-1989 0-0.2022*-0.1182*-0.0595-0.0689*0.0071(0.0543)(0.0482)(0.0766)(0.0306)(0.0543)1+0.2709*0.15560.2291*0.0922(0.0436)(0.1211)(0.0400)(0.0910)10.3230*0.2883*(0.0758)(0.0633)20.3750*0.4270*(0.1203)(0.1180) LogL8190.18197.58213.78214.28216.78218.0LRStat14.80*32.40*1.00NA2.60BIC-3.2268-3.2296-3.2330-3.2301-3.2311-3.2285 1990-1999 0-0.2211*-0.1529*-0.3460*-0.0864-0.2925*(0.0579)(0.0524)(0.0799)(0.0446)(0.0880)1+0.2079*0.6721*0.1861*0.5903*(0.0440)(0.1424)(0.0432)(0.1516)10.06260.0683(0.0712)(0.0658)20.36430.1669(0.2108)(0.1573) LogL8628.38635.78644.78649.88646.08650.4LRStat14.80*18.00*10.20*NA8.80*BIC-3.3989-3.4018-3.4023-3.4012-3.3997-3.3983 2000-2009 0-0.3209*-0.3029*-0.1736-0.1140-0.1826(0.0618)(0.0593)(0.0947)(0.0855)(0.1288)1+0.0881-0.12790.0919-0.1307(0.0640)(0.1402)(0.0556)(0.1451)10.2270*0.2279*(0.1026)(0.1034)20.5806*-0.0325(0.2780)(0.3113) LogL7866.47879.67880.67882.17881.37882.1LRStat26.40*2.003.00NA1.60BIC-3.1109-3.1162-3.1135-3.1109-3.1106-3.1078 Thistablereportstheestimatedparametersof~,where~=ln[p =21+1]andinnovationsfollowabinormaldistribution.Nrepresentsthebaselinenormallydistributederrorscase.ModelsM0toM4representequations(14)to(17),respectively.Tosavespace,wereportestimatedvaluesfor1and1+onthesamerow.Estimatedstandarderrorsarereportedinparentheses,belowtheestimatedparameters.\LogL"representsthecomputedloglikelihoodfunction.Wereportlikelihoodratioteststatistics,\LRStat,"withrespecttotheprecedingmodel.\NA"impliesthatthemodelsarenon-nested,andthustheLRteststatisticisnotcomputed.Allmodelsarepreferredtothebenchmarknormalmodel,basedonthelikelihoodratiotest.Thus,wedonotreporttheseteststatistics.*representsrejectionofthenullhypothesisinquestionatthe5%condencelevel.\BIC"representsBayesianinformationcriteriafortheestimatedmodels.39 Table9:EvaluatingtheConditionalRelativeDownsideVarianceforBinormalDistribution M0M1M2M3M4 1980-2009 0-2.30E-06-1.06E-05*-1.09E-05*-1.36E-05*-1.35E-05*(3.13E-06)(2.66E-06)(2.67E-06)(2.72E-06)(1.93E-06)11.5177*1.13171.11990.9262*0.9371(0.0502)(0.0849)(0.0752)(0.0229)(0.0435) R20.13200.25490.24740.21330.2096JT107.11*18.03*19.19*35.17*50.93* 1980-1989 05.70E-06-9.84E-06-1.05E-05-1.06E-05-1.30E-05(1.11E-05)(8.08E-06)(7.86E-06)(9.04E-06)(6.53E-06)12.0913*1.2928*1.2435*0.98600.7991*(0.1570)(0.0423)(0.0386)(0.0419)(0.0362) R20.15370.48840.51480.36200.3334JT48.58*49.38*41.57*1.48834.86* 1990-1999 09.24E-07-1.44E-05*-1.85E-05*-1.34E-05*-1.79E-05*(4.24E-06)(3.10E-06)(2.93E-06)(3.07E-06)(2.09E-06)12.2048*0.85650.5137*0.94371.1143(0.2150)(0.1029)(0.0806)(0.0975)(0.0856) R20.04010.02680.01580.03590.0182JT31.46*23.52*76.39*19.43*75.47* 2000-2009 0-1.26E-06-3.28E-06-5.06E-06-6.55E-06-4.97E-06(5.05E-06)(4.90E-06)(4.76E-06)4.88E-06(3.36E-06)11.1713*1.07190.99810.93951.0034(0.0485)(0.0424)(0.0374)(0.0385)(0.0374) R20.18880.20370.22170.19190.2231JT12.54*3.3271.1344.2702.192 ThistablereportsMincer-ZarnowitzregressionresultsfromRVdt+1(h)RVut+1(h)=0+1(Vart[rt+1jrt+1mt]Vart[rt+1jrt+1-295;mt]),whereerrortermsintheparametricmodelarebinormallydistributed.*indicatesrejectionofthenullhypothesisthati=i;i=0;1atthe5%condencelevelorbetter.JTrepresentsthevalueoftheteststatisticbuiltunderthenullhypothesisthat0and1arejointlyequalto0and1,respectively.*andydenotetherejectionofthisnullhypothesisatthe5and10%condencelevels,respectively.Thecriticalvaluesforthistestare5.991and4.605,respectively,basedon2df=2.40