版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
NO.1063
JUNE2023
MeasuringtheNaturalRateofInterestafterCOVID-19
KathrynHolston|ThomasLaubach|JohnC.Williams
MeasuringtheNaturalRateofInterestafterCOVID-19
KathrynHolstonThomasLaubach,andJohnC.Williams
FederalReserveBankofNewYorkStaffReports,no.1063
June2023
JELclassification:C32,E43,E52,O40
Abstract
WemodifytheLaubach-WilliamsandHolston-Laubach-Williamsmodelsofthenaturalrateofinteresttoaccountfortime-varyingvolatilityandapersistentCOVIDsupplyshockduringthepandemic.ResultingestimatesofthenaturalrateofinterestintheUnitedStates,Canada,andtheEuroAreaattheendof2022areclosetotheirrespectivelevelsestimateddirectlybeforethepandemic;thatis,wedonotfindevidencethattheeraofhistoricallylowestimatednaturalratesofinteresthasended.Incontrast,estimatesofthenaturalrateofoutputhavedeclinedrelativetothoseprojectedbeforethepandemic.
Keywords:naturalrateofoutput,time-varyingvolatility,Kalmanfilter,trendgrowth,COVID-19pandemic
_________________
Williams:FederalReserveBankofNewYork(email:john.c.williams@).Holston:Harvard
University(email:kathrynaholston@).Thispaperbuildsuponandextendstheworkofthethreeauthorsthatwasinitiallycontainedinthenote“AdaptingtheLaubachandWilliamsandHolston,Laubach,andWilliamsModelstotheCOVID-19Pandemic,”May27,2020.TheauthorsthankparticipantsattheThomasLaubachResearchConferenceforvaluablecomments.TheyalsothankLoganCaseyforoutstandingresearchassistance.Thispaperpresentspreliminaryfindingsandisbeingdistributedtoeconomistsandotherinterestedreaderssolelytostimulatediscussionandelicitcomments.Theviewsexpressedinthispaperarethoseoftheauthor(s)anddonotnecessarilyreflectthepositionoftheFederalReserveBankofNewYork,theFederalOpenMarketCommittee,ortheFederalReserveSystem.Anyerrorsoromissionsaretheresponsibilityoftheauthor(s).Toviewtheauthors’disclosurestatements,visit
/research/staff_reports/sr1063.html
.
1
1Introduction
Thedownwardtrendinestimatesofthenaturalrateofinteresttohistoricallylowlev-elsobservedinmanycountrieshasgarneredconsiderableattentionanddebateabouttheirsourcesandconsequences(LaubachandWilliams,2016,GourinchasandRey,2019,RachelandSummers,2019).1Theeventsofthepastfewyears,includingtheCOVID-19pandemicandsubsequentpolicyactions,haverenewedthedebateoverwhetherhistoricallylownaturalratesofinterestwillpersistinthepost-pandemicera(InternationalMonetaryFund,2023,Obstfeld,2023).Answeringthisquestionusingempiricalmodelsofthenaturalrateofinteresthasbeenchallengingowingtotheunprecedentedmacroeconomicvolatilityacrosstheglobeduringthepandemic.Thispaperdevelopsandimplementsadata-drivenapproachthataddressestheextraordi-naryeffectsofthepandemicusingtheHolston,Laubach,andWilliams(HLW,2017)andLaubachandWilliams(LW,2003)modelsofthenaturalrateofinterest.Ourapproachpreservestothegreatestextentthebasicstructureandflexibilityoftheoriginalmodels,whileprovidingconsistentmodelestimatesofnaturalratesbefore,during,andafterthepandemicperiod.Italsohasbroaderapplicationtomodelsthatestimatelatentvariablesusingfrequentist(asinHLWandLW)andBayesianmethods.
TheHLWandLWmodelsapplytheKalmanfiltertotranslatemovementsinrealGDP,inflation,andshort-terminterestratesintoestimatesoftrendgrowth,thenaturalrateofoutput,andthenaturalrateofinterest.Themodel’sstructureisflexibleandincorporatestransitoryandpermanentshockstosupplyanddemandanddynamicendogenousbehaviorofinflationandoutput.However,likeothermodelsthatusetheKalmanorotherstatisticalfilters,identifyingassumptionsregardingthenatureoftheshockprocessesareimposed.Inparticular,theshocksareassumedtobeseriallyuncorrelatedanddescribedbytime-invariantGaussiandistributions.TheCOVID-19pandemicgeneratedextraordinaryswingsinmacroeconomicdatathataredramaticallyatoddswithbothoftheseassumptions.
First,theassumptionofatime-invariantGaussiandistributionfortheshockpro-cessesisclearlycontradictedbythedata.Relativetothehistoricalexperienceofthepriorhalfcentury,theCOVID-19pandemicisanextremetaileventintermsofits
1Thereisarelatedliteratureonthehistoryofrealinterestratesoveraverylongtimespan.See,forexample,Rogoff,Rossi,andSchmelzing(2022)andreferencestherein.
3
demonstratethatthesetwomodificationseffectivelyaddressthetwoeconometricissuesassociatedwiththepandemic.3Theestimationprocedureyieldsparameterestimatesconsistentwiththemodelstructure,andthepre-pandemicestimatesofthenaturalratesofoutputandinterestandthetrendgrowthrateareverysimilartothoseestimatedondataendingin2019.
ThepatternofhistoricallylowestimatesoftrendGDPgrowthandthenaturalrateofinterestexperiencedbeforethepandemicpersistaftertheCOVID-19pandemic.Inallthreeeconomies,theestimatesoftrendgrowthandthenaturalrateofinterestin2022arewithinafewtenthsofapercentagepointofthecorrespondingestimatesfor2019.Inparticular,theseestimatesprovidenoevidenceofareversalofthetrenddeclineinestimatesofthenaturalrateofinterestbasedondatathrough2022.
Inallthreeeconomies,estimatesoftheCOVID-adjustednaturalrateofoutputin2022aresignificantlylowerthanwhatthemodelpredictsbasedonpre-pandemicdata.ThesedeclinesreflectboththeestimatedeffectsofCOVID-relatedrestrictionsandpermanentnegativeshockstothenaturalrateofoutput.Accordingtothemodel,thesedeclinesinthenaturalrateofoutputarethemosteconomicallysignificantlastingeffectsoftheCOVIDera.
Thispaperisorganizedasfollows.Section2describestheHLWmodelandtheevidenceofsignificantdeparturesfromthemodel’sassumptionsbroughtonbytheCOVID-19pandemic.Section3describesthemodificationstothemodeltoaddressthepandemic-relatedeffects.Section4reportsestimationresults.Section5reportsresultsfromrobustnessexercises.Section6concludes.
2ResultsfromthePre-PandemicModel
Inthissection,weprovideashortdescriptionoftheoriginalHLWmodel.Weshowthatthedataduringthepandemicgeneratelargeoutliersthatareinconsistentwiththeassumptionsofthemodel.
ceduresforbothmodels.Specifically,consistentwiththespecificationoftheLWmodel,wenowestimatetherelationshipbetweentrendgrowthandthenaturalrateofinterestintheHLWmodel,insteadofrestrictingittounity.Wealsomakeminortechnicaladjustmentsthataligntheassump-tionsusedinthevariousstagesofthemodelestimationprocedure,describedinAppendixA1.
3InHLW(2017),themodelwasalsoestimatedusingdatafromtheUnitedKingdom.ExtendingthesampletoincludethemostrecentyearshasweakenedtheestimatedrelationshipbetweentheoutputgapandrealinterestratesintheUKdata,makingestimatesofthenaturalrateofinteresthighlyunreliable.Forthatreason,wenolongerestimatethemodelfortheUnitedKingdom.
4
2.1TheOriginalHLWModel
IntheHLWmodel,thenaturalrateofinterest,r,istherealinterestrateconsistentwithoutputequalingitsnaturalrate,y,andstableinflation.Asisstandardinthisliterature(e.g.,seeWoodford,2003),wemodeltheoutputgapandinflationdynamicsasafunctionoftherealinterestrategap,rt−r,usinganintertemporalISequationandPhillipscurverelationship,inlinewiththeNewKeynesianframework:
2
t=ay,1t−1+ay,2t−2+工(rt−j−r−j)+ϵ,t
j=1
πt=bππt−1+(1−bπ)πt−2,4+byt−1+ϵπ,t
(1)
(2)
Theoutputgapisdefinedast=100·(yt−y),whereytandyarelogarithmsofrealGDPandtheunobservednaturalrateofoutput,respectively,rtistherealshort-terminterestrate,πtdenotesconsumerpriceinflation,andπt−2,4istheaverageofthesecondtofourthlagsoftheinflationrate.4Thestochasticdisturbancesϵ,tandϵπ,taretransitoryshockstotheoutputgapandinflationequations,respectively.
WeusetheKalmanfiltertoestimatethelatentvariables,whicharethenaturalrateofoutput,itstrendgrowthrate,andaprocesscapturingotherlow-frequencydeterminantsofthenaturalrateofinterest.InkeepingwiththestandardKalmanfilterapproach,thestochasticinnovationstothemeasurementequations–theISandPhillipscurveequations–areassumedtofollowaGaussiandistributionwithstandarddeviationsσy~andσπ,respectively,andtobemutuallyandseriallyuncorrelated.
Incontrasttothetransitoryshockstotheoutputgapandinflationequations,movementsinrreflecthighlypersistent,orpermanent,shiftsintherelationshipbetweentherealshort-terminterestrateandtheoutputgap(Williams,2003).Thelawofmotionforthenaturalrateofinterestisgivenby
r=c·gt+zt(3)
wheregtisthetrendgrowthrateofthenaturalrateofoutput,andztcapturesotherdeterminantsofrWespecifythethreelatentvariablesinourstate-spacemodel
4SeeHLW(2017)Section2andAppendixAfordetailsofthemodelspecification.Wetakeasastartingpointtheopen-economyNewKeynesianmodelspecificationasinGal´ı(2008)andrelaxtwostandardrestrictionstoworkwithreduced-formISandPhillipscurveequations.
5Notethat,consistentwithLW(2003),werelaxtheassumptioninHLW(2017)ofaone-for-one
5
asfollows.Thelogarithmofthenaturalrateofoutputfollowsarandomwalkwithastochasticdrift,gt,thatitselffollowsarandomwalk,
y=y−1+gt−1+ϵy*,t
gt=gt−1+ϵg,t
(4)
(5)
andthecomponentztcapturingotherdeterminantsofr,whichisassumedtofollowarandomwalkaswell,
zt=zt−1+ϵz,t(6)
Weassumethatthedisturbancesϵy*,t,ϵg,t,andϵz,tarenormallydistributedwithstan-darddeviationsσy*,σg,andσz,respectively,andareseriallyandcontemporaneouslyuncorrelatedwithallotherdisturbances.
Equations1and2makeupthemeasurementequationsinourstate-spacemodelandcanbeexpressedas
yt=A\·xt+H\·ξt+ϵt(7)
withstochasticinnovationsϵt.Equations4,5,and6makeupthestateequationsinourstate-spacemodel,writtenas
ξt=F·ξt−1+ηt(8)
whereξtisthestatevectoroflatentvariablesandηtisthevectorofstochasticinnovations.SeeAppendixA1forthefullstate-spacerepresentationofthemodel.
2.2OutliersinEstimationwith2019:Q4ModelParameters
WenowanalyzehowtheextrememovementsinGDPandinflationduringtheCOVID-19pandemicyieldlargeoutliersinthestandardHLWmodel.Wethenshowthatesti-matesofthelatentvariablesareheavilyaffectedbythesesizableoutliers,evenwhenweconstrainthemodelparametersattheirpre-pandemicvalues,anddemonstratethatmodificationstotheHLWandLWmodelsarenecessary.
relationshipbetweentrendgrowthandthenaturalrateofinterestandestimatethisrelationship.SeeAppendixA1fordetailsonchangestotheHLW(2017)model.
6
Beforemakinganyadjustmentstothemodels,webeginbyestimatingthestan-dardHLWmodelinSection2.1withdatathrough2019:Q4,priortotheonsetoftheCOVID-19pandemic.6Forreference,theupperthreepanelsofFigure2showthefullsamplesofdata.Wefixthemodelparametersattheirestimatedvaluesandre-estimatethelatentvariablesthrough2022:Q4usingtheKalmanfilter,takingallparametervaluesasgivenfromthe2019:Q4estimatedmodel.Wealsofixtheinitialvectorofunobservedstatesanditscovariancematrixatthe2019:Q4values.7Thisexerciseisequivalenttodroppingobservationsbeginningin2020:Q1throughtheendofthesampleduringthemaximumlikelihoodestimationofmodelparameters,whileallowingtheKalmanupdatingproceduretocontinuewithoutmodificationthroughtheendofthesample.Inotherwords,wemakenomodificationstothestate-spacemodel,exceptthatthemodelcoefficientmatricesandcovariancematricesintheKalmanfilteringprocedurearefixedattheir2019:Q4values.Thiswouldbeasuit-ableapproachifwetaketheviewthatthepandemicperiodisnotinformativeforthemodelparameters,suchastheslopesoftheISandPhillipscurveequations,butisinformativeforthelatentvariables.
ThefinalstepoftheKalmanfilteringproceduretoestimatethevectorofunob-servedstatevariablesattimet(denotedasξˆt|t,andconditionalontheinformationsetattimet)isgivenbytheKalmanupdatingequation,
ξˆt|t=ξˆt|t−1+Kt·(yt−A′·xt−H′·ξˆt|t−1)
山、
尸使
one-step-aheadpredictionerror
(9)
whereξˆt|t−1istheinitialestimateofthestatevectorduringtheperiod,conditionalontheinformationsetattimet−1,andKtistheKalmangainmatrix.Thefinaltermcontainstheone-step-aheadpredictionerrors(orforecasterrors)correspondingtothemeasurementequationsinthemodel.Theseone-step-aheadpredictionerrorsaretheresidualstotheISandPhillipscurveequations,usingtheforecastofyt(thevectorofcontemporaneousendogenousvariables,thatis,theoutputgapandinflation)based
6Throughoutthepaper,weusethecurrentdatavintageatthetimeofpublication,regardlessofthesampleperiod.
7Westoretheestimatedparametervectorθfromthefinal(stage3)modelaswellasthesignal-to-noiseratiosλgandλXfromthemedianunbiasedestimationproceduresfollowingstages1and2,respectively.SeeHLW(2017)foradescriptionoftheestimationprocedureandfootnote6fortheinitializationprocessofthevectorofunobservedstates,itsconditionalexpectationξ1|0inthefirstperiod,andthecovariancematrixP1|0.
7
onthedataattimetandinformationattimet−1,correspondingtothestatevector
ξˆt|t−1:
yt−E[yt|xt,ζt−1]=yt−(A′·xt+H′·ξˆt|t−1)
(10)
InEquation9,theKalmangainmatrixdictatestheweightplacedontheone-step-aheadpredictionerrorsduringthelatentvariableestimation.AlargerKalmangainKtindicatesthatthefinalestimatesofthelatentvariablesaremoreheavilyinfluencedbythegapbetweentherealizeddataandthemodel’sprediction,relativetoξˆt|t−1,
thepriorestimateofξtconditionaloninformationintheprecedingperiod.
Inthisinitialexercise,thecoefficientmatricesH′(onthestatevectorξˆt|t−1)andA′(onthedataxt)arefixedattheir2019:Q4values.Theresultingone-step-aheadpredictionerrorsareverylargeduringmuchofthepandemicperiod.Theselargeforecasterrorstranslatedirectlytotheestimatedvectorofunobservedlatentvariables(y,gt,zt),sothatthedataduringthisperiodhaslargeeffectsonestimatesoftheselatentvariables.
BecausetheHLWmodelisanunobserved-componentsmodel,thereexistsan-othersetofmodelresidualsinadditiontotheone-step-aheadpredictionerrorsthatarecommonlyusedfordiagnostictesting.Theseauxiliaryresidualsaresmoothedestimatorsofthedisturbancestothemeasurementequations,ϵt,andtothestateequations,ηt,meaningthattheyincorporateallavailableinformationoverthefullsampleperiodandprovideadifferentinterpretationofthestochasticinnovations(HarveyandKoopman,1992;Harveyetal.1999).Theyhavetheadvantageinapplyingatestforoutliers:undertheassumptionthatthestochasticinnovationsarefromaGaussiandistribution,standardizedauxiliaryresidualsgreaterthan2(inabsolutevalue)indicateeitherthepresenceofoutliersorstructuralchange.
WeusethealgorithmfromKoopmanandDurbin(2000)toobtainthestandard-izedauxiliaryresidualstothemeasurementequations,t/e,t,andtothestateequa-tions,t/η,t.ThegoldlinesinFigure3showsthatthestandardizedauxiliaryresid-ualstothemeasurementequations,givenby
te,t
=
E[ϵt|yT,xT,ζT]
SD[ϵt|yT,xT,ζT],
(11)
indicateextremeoutlierstotheoutputgapequationinalleconomiesinoursample,underthestandardHLWmodelusing2019:Q4parametervalues.IntheUnitedStatesandCanada,standardizedauxiliaryresidualstotheISequationare9to10timesthe
8
outlierthresholdforaGaussianmodelinthesecondquarterof2020,and17timesthethresholdintheEuroArea.Whilenotasextreme,themodelresidualsalsodetectoutlierstotheinflationequationineachofthethreeeconomies,withstandardizedauxiliaryresidualstothePhillipscurveequationreachingdoubletheoutlierthresholdintheUnitedStates.
Standardizedauxiliaryresidualstotheunobservedstateequationsaregivenby
tη,t
=
E[ηt|yT,xT,ζT]
SD[ηt|yT,xT,ζT]
(12)
AsshownbythegoldlinesinFigure4,theseresidualsdemonstratethepresenceofextremeoutlierstothenaturalrateofoutputinthestandardHLWmodelacrossalloftheeconomiesinoursample.
ThisobservationisnotuniquetotheHLWmodel.Figure1showsthatGDPreal-izationsduringthepandemicareoutlierswithrespecttohistoricaldata.Macroeco-nomicfluctuationsofthismagnitudewouldresultinoutliersinanystandardmacroe-conomicmodel.Asexpected,evenwhenpandemic-eraGDPandinflationdataareexcludedduringparameterestimation,usingtheKalmanfilterwiththeseextremeoutlierspresentinthesamplesignificantlydistortsestimatesofthelatentvariablesduringthepandemicperiod.ThegoldlinesinFigure5displayslargeswingsintheestimatesofthelatentvariablesusingthemodelwith2019:Q4parametervalues.Theextremevolatilityintheseestimatesisinconflictwiththespecificationoftheselatentvariablesasreflectinglower-frequencymovements.
3COVID-adjustedModel
TheobjectiveofthispaperistoestimatethenaturalrateofinterestfollowingtheCOVID-19pandemicinawaythatisconsistentwiththeHLWmodeloutlinedinSec-tion2.1.Thelargemovementsineconomicactivityandthepersistentsupplyshocksduringthepandemicperiodviolatetwostandard,butimportant,modelassumptionsinHLW.AsweshowinSection2.2,droppingtheobservationsfromthisperiodduringthemaximumlikelihoodestimationofmodelparametersisinsufficienttoovercometheseviolatedmodelassumptions,withmodelresidualsindicatingthepresenceoflargeoutliers.Theresultingestimatesofthenaturalratesofoutputandinterestareextremelyvolatileandinconsistentwithourspecificationofrasamedium-run
9
conceptthatisdrivenbylow-frequencymovements.
ThissectiondetailstwoadjustmentstotheHLWmodelthat,takentogether,addresstheviolationstotheoriginalmodel.Thefirstistheintroductionoftime-varyingvolatilityinthemodel,whichweimplementbyallowingthevariancesofthestochasticinnovationstotheoutputgapandinflationequationstobehigherintheCOVIDerarelativetothenon-pandemicperiod.Thesecondistheintroductionofapersistent,butultimatelytemporary,COVIDsupplyshock,inadditiontothetransitoryandpermanentdemandandsupplyshocksthatarealreadypresentinthemodel.Eachofthesemodificationsaloneisinsufficienttoovercometheestimationchallengesposedbythepandemic,butinconjunctionwitheachothertheyallowforcontinuedestimationofthenaturalrateofinterest.
Importantly,becauseweareestimatinglatentvariablesthatarespecifiedasran-domwalks,simplydroppingtheobservationsduringthepandemicperiodwouldnotonlyunderstatethetrueuncertaintyassociatedwiththisperiod,butwouldalsone-cessitateinterpolatingtheselatentvariablesfromtheirpre-pandemicvalues.Instead,ratherthanimposingthatthesevariablesdonotchangeasaresultoftheCOVID-19pandemic,weareabletoletthedatainformtheestimatednaturalratesofinterestandoutputafterthepandemichasabated.Whiletheintroductionoftime-varyingvolatilityhasasimilareffectonthelatentvariablesasdroppingtheobservationsfrom2020,ourapproachprovidesmoreflexibilityinthelateryearsofthepandemicanddoesnotrequirethestrongassumptionthatseveralyearsofdatahasnoeffectonthelatentvariables.Additionally,theHLWmodelwithtime-varyingvolatilitybutwithoutseriallycorrelatedshockstosupplywouldconstrainhowthenaturalrateofoutputcouldevolveinresponsetothepandemicdata.ModelingthepersistentCOVIDsupplyshockisnecessaryinordertocapturetheeffectsofthepandemiconthenaturalrateofoutput.
3.1Time-VaryingVolatilityduringCOVID-19
COVID-19representsanextremetaileventrelativetotheassumptionofGaussiandisturbances.AsweshowinSection2.2,theresultingoutlierscontaminateestimatesofthelatentvariablesevenwhenweexcludethemfromtheestimationofmodelparameters.Thisisastatisticalproblemthatisnotuniquetoourmodelortoestimationofthenaturalrateofinterest.Wepresentastraightforwardapproachto
10
accountforthesubstantialincreaseinvolatilityduringthisperiodbyintroducingtime-varyingvolatilityinthemodelduringthewindowoftimeassociatedwiththeCOVID-19pandemic.WebuildonaninsightfromLenzaandPrimiceri(2022):ifthetimingofincreasedvolatilityisknown–asisthecasefortheCOVIDpandemic–wecanintroducetime-varyingvolatilityinthemodeldirectlybyapplyingascalefactortotheinnovationvariancesduringtheperiodofincreasedvolatility.Weapplythisinsighttoourunobserved-componentsmodelinordertoestimatethenaturalrateofinterest,butourapproachcangeneralizetoanystate-spacemodelwithlatentvariables.
Inparticular,weintroducethreenewmodelparameters,κ2020,κ2021,andκ2022.Thesearethevariancescaleparametersfor2020,2021,and2022,respectively,whichmultiplythevariancesoftheinnovationstotheoutputgapandinflationequations.Wedefinethevectorκtofvariancescaleparametersattimet,thattakesthevalues
κ2020
κ2022
κt=〈κ2021
(1
2020:Q2≤t≤2020:Q4
2021:Q1≤t≤2021:Q4
2022:Q1≤t≤2022:Q4
otherwise
(13)
Weestimatethethreevariancescaleparametersbymaximumlikelihoodtogetherwiththeothermodelparameters,withtheconstraintsκ2020≥1,κ2021≥1,andκ2022≥18.κttakesthevalueof1beforethepandemicperiodandin2023andbeyond.Section5.1considersalternativespecificationsoftime-varyingvolatility.
Thecovariancematrixofthestochasticinnovationstotheoutputgapandinflationequationsisnowtime-varyingandisgivenby
Rt=κ·R=[(κt]
(14)
withtime-varyinginnovationvariancestotheIScurveandPhillipscurveequationsof(κtσy~)2and(κtσπ)2,respectively.
8Therestrictionthatκt≥1isnecessarytoensurethatthelikelihoodestimationcannotdown-weightthevarianceofcertainobservations,whichwouldineffectallowittoplacemoreweightonfavorableobservations.Instead,theestimatedκtfactorscanonlyincreasetheinnovationvariancesduringthepandemicperiod.
11
Outsideofthepandemicperiod,theinnovationvariancesarespecifiedexactlyasinHLW(2017).Thatis,theinnovationvariancestotheoutputgapequation,σ,andinflationequation,σ,areconstantoverthesamplepriorto2020andafter2022.Duringtheyear2021,forexample,theinnovationvariancestakethevalues(κ2021·σy~)2and(κ2021·σπ)2,respectively.Thereforeκtisaratioofthestandarddeviationsofthedisturbancestothemeasurementequations(theoutputgapandinflationequations)attimetrelativetothestandarddeviationsinthenon-pandemicsample.Whenκt>1,aswefindfor2020through2022inalleconomiesinoursample,theinnovationvariancesaregreaterthaninthenon-pandemicsample.
Introducingtime-varyingvolatilitytothestochasticinnovationsviathevariancescalefactorsinκthastheeffectofdown-weightingextremeoutlierobservationsinthemaximumlikelihoodestimationofmodelparametersaswellasinestimationofthelatentvariablesviatheKalmanfilter.Whenκt>1,thediagonalcovariancematrixRtofthedisturbancestotheoutputgapandinflationequationsislargerrelativetothecasewhereκt=1,andtheresultingKalmangainissmaller.AsshowninEquation9,theKalmangaindictatestheweightplacedontheone-step-aheadpredictionerrors–thedifferencebetweenrealizedvaluesoftheoutputgapandinflationinagivenperiodandthemodel’spredictedvaluesbasedoninformationinthepriorperiod–inupdatingthefilteredestimatesofthelatentvariables.Astheinnovationvariancesinagivenperiodbecomelarge,theKalmangainshrinks,sothattheKalmanfilterplacesrelativelylittleweightonthesenewobservationsand
theestimatesofthelatentvariablesinthestatevector(thatis,y,gt,ztandthereforer)remainclosetotheestimatesfromthepriorperiod.
Inthelimitastheinnovationvariancestendtowardinfinity,theKalmangainapproacheszero,sothatnoweightisplacedonthetime-tobservationsinestimatingthestatevector.Ineffect,themodeldoesnotmakeuseoftime-tinformation,sothattheforecastofthestatevectorattimetgiventhetime-tinformationsetisunchangedfromtheforecastgiventheinformationsetattimet−1.ThislimitingcaseisequivalenttodroppingtheCOVID-19observationswhenestimatingthelatentvariables.Thesameholdsforparameterestimation:whenκtislarge,themodelforecasterrorinthisperiodisdown-weightedwhencomputingtheloglikelihoodfu
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 篮球测试活动方案
- 绿树养殖活动方案
- 基层卫生院岗位设置参考方案
- 机电设备安装调试与运行维护方案
- 黑龙江省哈尔滨市2025届高三上学期第二次月考物理试卷(含答案)
- 施工方案蛋糕
- 老凤祥十一活动策划方案
- 面包糕点营销方案
- 肇庆认证咨询iso体系报价方案
- 金秀郡营销方案
- 农村自建房全套知识培训课件
- 2025年全国出租车从业资格考试复习模拟题库(500题)含答案
- 2025年健康管理师三级专业能力考试题及答案
- 电力公司抄表系统管理办法
- 阿里巴巴销售管理制度
- 消化内镜质量改进持续措施
- 2025年公需科目一带一路考试试题及答案
- 函数表示法说课课件
- 基层应急管理培训课件
- 托管员工培训方案
- TCWAN 0142-2025 船用钛-钢过渡接头焊接工艺规范
评论
0/150
提交评论