资源描述:
《on inference for partially observed nonlinear diffusion models using the metropolis–hastings algorithm》由会员上传分享,免费在线阅读,更多相关内容在工程资料-天天文库。
1、Biometrika(2001),88,3,pp.603–621©2001BiometrikaTrustPrintedinGreatBritainOninferenceforpartiallyobservednonlineardiffusionmodelsusingtheMetropolis–HastingsalgorithmBG.O.ROBERTSDepartmentofMathematicsandStatistics,LancasterUniversity,Lancaster,LA14YF,U.K.g.o.roberts@lancaster.ac.uk
2、O.STRAMERDepartmentofStatisticsandActuarialScience,UniversityofIowa,IowaCity,Iowa52242,U.S.A.stramer@stat.uiowa.eduSInthispaper,weintroduceanewMarkovchainMonteCarloapproachtoBayesiananalysisofdiscretelyobserveddiffusionprocesses.Wetreatthepathsbetweenanytwodatapointsasmiss
3、ingdata.Assuch,weshowthat,becauseoffulldependencebetweenthemissingpathsandthevolatilityofthediffusion,therateofconvergenceofbasicalgorithmscanbearbitrarilyslowiftheamountoftheaugmentationislarge.Weofferatransformationofthediffusionwhichbreaksdowndependencybetweenthetransformedmissing
4、pathsandthevolatilityofthediffusion.WethenproposetwoefficientMarkovchainMonteCarloalgorithmstosamplefromtheposterior-distributionofthetrans-formedmissingobservationsandtheparametersofthediffusion.WeapplyourresultstoexamplesinvolvingsimulateddataandalsotoEurodollarshort-ratedata.Someke
5、ywords:Diffusionprocess;Independencesampler;MarkovchainMonteCarlo.1.IDiffusionprocessesarenaturalstatisticalmodelsformanynaturalphenomena,includ-ingfinancialtimeseries.However,inferenceiscomplicatedbythefactthatcompletedatadescribingthediffusionsamplepathisofnecessitynotava
6、ilable.Infactdatawillbeadiscretesetofobservationsoftheprocessoversomefinitetimeperiod,andthemarginallikelihoodgiventhisdiscretesetofobservationswillalmostalwaysbeunavailable.ThiscomplicatesbothclassicalandBayesianinference,thoughherewewillconcentrateonaBayesiantreatmentoftheproblem
7、.SupposeforinstancethatXisadiffusionprocesssatisfyingdX=sdB+b(t,X,h)dt,(1)tttforsomeRd-valuedparameterhandBrownianmotionB,observedatdiscretetimepointst=(t,t,...,t).Occasionally,tmaybesufficientlyfinetoallowtheEulerapproximating01NMarkovchainX~N{X+(t−t)b(t,X,h),s2(t−t)}(2)ttii−1i−1tii−
8、1ii−1i−1604G.O.RO.S