Neural network forecasting for seasonal and trend time series (1)

《Neural network forecasting for seasonal and trend time series (1)》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、EuropeanJournalofOperationalResearch160(2005)501–514www.elsevier.com/locate/dswComputing,ArtificialIntelligenceandInformationTechnologyNeuralnetworkforecastingforseasonalandtrendtimeseriesa,*bG.PeterZhang,MinQiaDepartmentofManagement,J.MackRobinsonCollegeofBusiness,GeorgiaStateUniversity,35Bro

2、adStreet,NW,Atlanta,GA30303,USAbDepartmentofEconomics,CollegeofBusinessAdministration,KentStateUniversity,Kent,OH44242,USAReceived19October2001;accepted8August2003Availableonline18November2003AbstractNeuralnetworkshavebeenwidelyusedasapromisingmethodfortimeseriesforecasting.However,limitedem-

3、piricalstudiesonseasonaltimeseriesforecastingwithneuralnetworksyieldmixedresults.Whilesomefindthatneuralnetworksareabletomodelseasonalitydirectlyandpriordeseasonalizationisnotnecessary,othersconcludejusttheopposite.Inthispaper,weinvestigatetheissueofhowtoeffectivelymodeltimeserieswithbothseason

4、alandtrendpatterns.Inparticular,westudytheeffectivenessofdatapreprocessing,includingdeseasonalizationanddetrending,onneuralnetworkmodelingandforecastingperformance.BothsimulationandrealdataareexaminedandresultsarecomparedtothoseobtainedfromtheBox–Jenkinsseasonalautoregressiveintegratedmovingav

5、eragemodels.Wefindthatneuralnetworksarenotabletocaptureseasonalortrendvariationseffectivelywiththeunpreprocessedrawdataandeitherdetrendingordeseasonalizationcandramaticallyreduceforecastingerrors.Moreover,acombineddetr-endinganddeseasonalizationisfoundtobethemosteffectivedatapreprocessingapproac

6、h.Ó2003ElsevierB.V.Allrightsreserved.Keywords:Neuralnetworks;Box–Jenkinsmethod;Seasonality;Timeseries;Forecasting1.IntroductioncompaniedwiththeseasonalvariationsandcanhaveasignificantimpactonvariousforecastingManybusinessandeconomictimeseriesexhibitmethods.Atimeserieswithtrendisconsideredtosea

7、sonalandtrendvariations.Seasonalityisabenonstationaryandoftenneedstobemadesta-periodicandrecurrentpatterncausedbyfactorstionarybeforemostmodelingandforecastingsuchasweather,holidays,repeatingpromotions,processestakeplace.Accurateforecastingof