JFE机器学习预测股票走势

ContentslistsavailableatScienceDirect JournalofFinancialEconomics journalhomepage:www.elsevier.com/locate/jfec JournalofFinancialEconomics153(2024)103791 Chartingbymachines ScottMurray∗,YusenXia,HoupingXiao GeorgiaStateUniversity,RobinsonCollegeofBusiness,35BroadStreet,Atlanta,GA,30303,UnitedStates ARTICLEINFOABSTRACT Datasetlink:https://dx.doi.org/10.17632/x63r376783.2 JELclassification: G11G12 Keywords: EfficientmarkethypothesisMachinelearning DeeplearningChartingTechnicalanalysis Cross-sectionofstockreturns Wetesttheefficientmarkethypothesisbyusingmachinelearningtoforecaststockreturnsfromhistoricalperformance.Theseforecastsstronglypredictthecross-sectionoffuturestockreturns.Thepredictivepowerholdsinmostsubperiodsandisstrongamongthelargest500stocks.Theforecastingfunctionhasimportantnonlinearitiesandinteractions,isremarkablystablethroughtime,andcaptureseffectsdistinctfrommomentum,reversal,andextanttechnicalsignals.Thesefindingsquestiontheefficientmarkethypothesisandindicatethattechnicalanalysisandchartinghavemerit.Wealsodemonstratethatmachinelearningmodelsthatperformwellinoptimizationcontinuetoperformwellout-of-sample. 1.Introduction Theweakformoftheefficientmarkethypothesis(EMHhereafter)stipulatesthattheconstructionofaprofitableportfoliobasedonlyoninformationdiscernablefromplotsdepictingthehistoricalperformanceofstocks(priceplotshereafter)shouldnotbepossible.Assuch,techni-calanalysis,orcharting,shouldbeafruitlessinvestmenttechnique.Academicresearchontechnicalanalysishasbroadlysupportedthisprediction.Despitethebroaddismissaloftechnicalanalysisintheaca-demicliterature,itremainswidelyusedbyinvestmentmanagers.Thecontinuedwidespreaduseoftechnicalanalysissuggeststhatitsmeritmaynotbefullydiscoveredinacademicresearch,andthatfurtherin-vestigationiswarranted. Inthispaper,wetesttheweakformoftheEMHbyexaminingwhetherforecastsproducedbymachinelearning(MLhereafter)predictthecross-sectionoffuturestockreturns.Theforecastsarebasedonlyondatathatareeasilydiscernablefromhistoricalpriceplots,specif-icallythecumulativestockreturnsoverthepast12months.WefindstrongevidencethattheML-basedforecastshaveeconomicallyimpor- tantandhighlystatisticallysignificantpredictivepower.Thispredictivepowerprevailsinmostsubperiodsofourfocal196307-202212testpe-riod,includingthemostrecent201501-202212subperiod,andremainsstrongamongthelargest500stocks.Theforecastingfunctionisremark-ablystablethroughtimeandhighlycomplex,withsubstantialnonlinearandinteractioncomponentsthatareimportantforprediction.Finally,thepredictivepowerofourML-basedforecastsisnotexplainedbythewell-knownmomentum(JegadeeshandTitman(1993))andreversal(Jegadeesh(1990))effects,norbypreviouslystudiedtechnicalorML-basedsignals. ExecutionoftheMLprocessrequiresustomakeseveralimplemen-tationdecisions.Toalleviateconcernsrelatedtodatamining(Harveyetal.(2016))orout-of-sampleforecastingpower(McLeanandPontiff (2016),Greenetal.(2017)),weusedatafrom192701-196306,whichwerefertoasthe“optimizationperiod”,toselectourMLmodel.1Ouranalysesleadustouseaconvolutionalneuralnetworkwithlongshort-termmemoryastheMLarchitecture,mean-squarederrorasthelossfunction,aweightingschemethatassignsthesametotalweighttoob- servationsfromeachmonthandequalweighttoeachstockwithina *Correspondingauthor. E-mailaddresses:smurray19@gsu.edu(S.Murray),ysxia@gsu.edu(Y.Xia),hxiao@gsu.edu(H.Xiao). 1LoandMacKinlay(1990)andFama(1991)alsoraiseconcernsaboutfalselysignificantrelationsbetweenpredictivevariablesandthecross-sectionoffuturestockreturns.Giglioetal.(2021)proposeamethodologytoaddresstheissueofmultipletesting,raisedbyHarveyetal.(2016),inthecontextoflinearassetpricingmodels.Schwert(2003)showsthatthemagnitudesofthesizeandvalueeffectsdecreaseaftertheperiodexaminedbytheinitialresearchontheseeffects. https://doi.org/10.1016/j.jfineco.2024.103791 Received27May2022;Receivedinrevisedform12January2024;Accepted13January2024 Availableonline24January2024 0304-405X/©2024ElsevierB.V.Allrightsreserved. month,andanormalizedmeasureoffuturereturnasthedependentvariable.2OuruseofdatafromonlytheoptimizationperiodtomaketheseMLmodeldecisionsensuresthatourmaintestresultstrulyreflectout-of-samplepredictivepower. WeusetheoptimizedMLmodeltogeneratestockreturnforecastsduringthe196307-202212period,whichwerefertoasthe“testperi-od”.OurmainanalysesexaminetheperformanceofportfoliosformedbysortingstocksontheML-basedforecasts,whichwedenote𝑀𝐿𝐸𝑅.Wefindthat𝑀𝐿𝐸�isastrongpredictorofthecross-sectionoffuturestockreturns.Theaverageexcessreturnsofvalue-weighteddecileport-foliosconstructedusingbreakpointscalculatedfromonlyNYSE-listedstocksincreasefrom−0.14%permonthforthedecileoneportfolioto0.93%permonthforthedecile10portfolio.The1.08%permonthgen-eratedbytheportfoliothatislongthedecile10portfolioandshortthedecileoneportfolioisnotonlyeconomicallylarge,buthighlystatis-ticallysignificant,witha𝑡-statisticof5.51.Whileourmethodologyisdesignedtoensurethatourresultsreflectout-of-sampleperformance,thisandother𝑡-statisticsfromourassetpricingtestsfarsurpassthebenchmark𝑡-statisticsproposedbyHar