简介:High-frequencystocktrendpredictionusingmachinelearnershasraisedsubstantialinterestinliterature.Nevertheless,thereisnogoldstandardtoselecttheinputsforthelearners.Thispaperinvestigatestheapproachofadaptiveinputselection(AIS)forthetrendpredictionofhigh-frequencystockindexpriceandcomparesitwiththecommonlyuseddeterministicinputsetting(DIS)approach.TheDISapproachisimplementedthroughcomputationoftechnicalindicatorvaluesondeterministicperiodparameters.TheAISapproachselectsthemostsuitableindicatorsandtheirparametersforthetime-varyingdatasetusingfeatureselectionmethods.Twostate-of-the-artmachinelearners,supportvectormachine(SVM)andartificialneuralnetwork(ANN),areadoptedaslearningmodels.AccuracyandF-measureofSVMandANNmodelswithboththeapproachesarecomputedbasedonthehigh-frequencydataofCSI300index.TheresultssuggestthattheAISapproachusingt-statistics,informationgainandROCmethodscanachievebetterpredictionperformancethantheDISapproach.Also,theinvestmentperformanceevaluationshowsthattheAISapproachwiththesamethreefeatureselectionmethodsprovidessignificantlyhigherreturnsthantheDISapproach.