In this note we represent the well known discrete time stochastic volatility (SV) model with a leverage effect and the SV model of Jacquier, Polson and Rossi (JPR) (2002) using Gaussian nonlinear state space forms with uncorrelated measurement and transiti
2Regardingthepriordistributions,fortheparametersφandσv,wefollowexactlythe
2priorspeci cationsofKim,ShephardandChib(1998):σv~Inverse-Gamma(2.5,0.025)
whichhasameanof0.167andastandarddeviationof0.024andφ ~Beta-distributionwithparameters20and1.5whichhasameanof0.167andastandarddeviationof0.86and0.11,whereφ =(φ+1)/2.Thecorrelationparameterρisassumedtobeuniformlydistributedwithsupportbetween-1and1andhenceiscompletely at.
Inallcaseswechooseaburn-inperiodof10,000iterationsandafollow-upperiod
2of100,000.TheMCMCsamplerisinitializedbysettingφ=0.98,σv=0.025,and
ρ= 0.5.BUGScodecanbedownloadedfrommywebsite
doc.guandang.net~jyu/research.html
4EmpiricalComparisonofA-SVModels
AsarguedinSection2,theSVmodelcorrespondingtothecontinuoustimemodelisdi erentfromtheJPRspeci cationwhichdoesnotnecessarilyimplyaleveragee ect.However,nothingsaysthecontinuoustimemodelissomeformof“truth”andhenceitisinterestingtocomparetheempiricalperformanceofthesetwoalternativespeci cations.Todothis,weemployacommonlyuseddatasetwhichcontains2023dailyreturnsofS&P500from1980to1987.ThesamedatasetisalsousedinJPR.
Sinceneitherspeci cationisnestedbyeachother,wecannotusetheclassicallikeli-hoodratiotesttocomparetheperformancesofthesetwoalternativemodels.BayesiancomparisonisoftenmadeusingtheBayesfactorwhichinvolvesthecalculationofthemarginallikelihoodofthecompetingmodels.Therearevariouswaystocalculatethemarginallikelihood.Forinstance,Kimetal.(1998)andChib,NardariandShephard(2002)haveshownhowtocomputethemarginallikelihoodattheposteriormeanusingtheapproachsuggestedbyChib(1995).However,thismarginallikelihoodapproachre-mainsacomputationallyintensivetaskandisnotaparticularlyuser-friendlytool.JPRproposeaninterestingwaytocalculatetheBayesfactorbymakinguseofthespecialstructureofthemodelsandpriors.Inthispaper,wefollowNewtonandRaftery(1994)’ssuggestionwhichusestheharmonicmeanofthesampledlikelihoodvaluesasasimu-lationconsistentestimatoroftherequiredmarginallikelihood.AlternativeBayesiancomparisoncanbemadeviainformationcriteria.Inthispaperweemploythenewlydevelopeddevianceinformationcriterion(DIC)proposedbySpiegelhalter,Best,CarlinandvanderLinde(2002).AsshowninBerg,MeyerandYu(2002),DICisaparticularlyuser-friendlyande ectivetoolforcomparingSVmodels.
InTable1wesummarizetheresultsfromestimationandmodelcomparison,includ-ingtheposteriormeans,standarddeviations(SD),95%Bayescon denceintervalsforalltheparameters,theharmonicmeanestimatesoflogmarginallikelihood,andDICforbothmodels.Althoughtheleveragee ectinbothmodelsissigni cant,itismarkedlysmallerinJPR’smodel.Thissuggeststhatiftheleveragee ectwereestimatedfromJPR’smodel,itwouldbeunderestimatedinmagnitudebyabout20%.UsingthelogmarginallikelihoodvaluesweobtaintheBayesfactoroftheconventionalA-SVmodel
MCMC Methods for Estimat... 暂无评价 11页 免费 Volatility in Financial ....Asymmetry and Leverage in Stochastic Volatility Models-An Exposition 隐藏>> Asym...
MCMC Methods for Estimat... 暂无评价 11页 免费 Volatility in Financial ....Asymmetry and Leverage in Stochastic Volatility Models-An Exposition 隐藏>> Asym...
热门文档