研究基于GA-SVM的上市公司财务困境预警(ijeme-v1-n2-01)

I.J.EducationandManagementEngineering2011,2,1-7PublishedOnlineAugust2011inMECS()DOI:10.5815/ijeme.2011.02.01Availableonlineat①LIYun-feia,ZHANGQianbaSchoolofEconomicsandManagement,Xi'anUniversityofTechnology,710048,Xi’an,Chinaa,bSchoolofmanagement,Xi’anPolytechnicUniversity,710048,Xi’an,ChinaAbstractBasedonfinancialmanagementandenterprisesofearly-warningtheory,thispaperconstructsafinancialdistressearly-warningmodelusingGA-SVM.First,ituseslistedcompaniesappearinginShanghaiStockExchangeandShenzhenStockExchangein2007-2009assamplebooks.DefiningSTlistedcompanieswhichhaveabnormityoffinancestatusassignatureofthelistedcompany'sfinancialcrisis.Thenitusesthedatainthefinancialstatementsknowntothepublicastheinputfeaturevectorandcombinegeneticalgorithmandsupportvectormachine.UseTakinganempiricalresearchwiththefinancialdistressearly-warningmodel.Testresultsofthedemonstrationstudyshowsthemodelhasasuperiorityinpredictingfinancialdistress.IndexTerms:FinancialDistress;Early-Warning;GeneticAlgorithm;SupportVectorMachine©2011PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheInternationalConferenceonE-BusinessSystemandEducationTechnology1.IntroductionFinancialdistressearly-warningsystemisafinancialanalysissystemwhichcanforecastfinancialdistressforenterprisesorenterprisegroupsbysettingupsomesensitivityfinancialindicatorsandobservingitschanges.Financialdistressisthemostintegrativeandsignificantrepresentationofenterprisecrisis.AscontinualperfectingofChina’sstockmarketmechanismandenterprisebankruptcysystem,financialdistresshasmadeenterprisessuffergreatlossanddirectlyaffectthesurvivalanddevelopment.Furthermore,anewfinancialcrisiswillprobablyoccuriflargenumberofenterprisesgetsintofinancialdistressatthesametime.Thus,howtoeffectivelyforecastfinancialdistressofenterpriseshasbecomeanurgentneedresearchissues.Onthecurrent,studyonfinancialcrisisearly-warninghasbecomeahotacademicresearchproblem.Intheearlystudy,regressionanalysis,mathematicalprogramming,fuzzyanalysis,k-NearestNeighbor(KNN)andexpertsystemareusedtopredictfinancialdistress.Sincethesearelinearmethodinsubstanceandtheireffectivenessdepends①ThispaperisfundedbyXi’anPolytechnicUniversitybasicresearchprojects.(09XG04).Correspondingauthor:E-mailaddress:[e-mail];[e-mail]2ResearchonFinancialDistressEarly-warningofListedCompaniesBasedonGA-SVMonrestrictiveassumptions,itsapplicabilityandeffectivenessarelimited.RecentlyalotofresearchshowsthatArtificialIntelligence(AI)methodisparticularlyoutstandingindealingwithnonlinearcomplexsystemproblems.Inparticular,ArtificialNeuralNetworks(ANN)isthemostwidespreadusedmethod.Forexample,Back（1996）[1],Zhang（1999）[2],Yang（1999）[3]andPendharkar（2005）[4]usedifferentANNmodeltopredictfinicaldistressandcomparewiththetraditionalmethod.TheresultshowsthattheeffectofANNisbetterthanthetraditionallinearmethod.AlthoughmanytheoriesandempiricalstudieshaveprovedthatANNismoreeffectivethanthetraditionalmethod,thetraditionalneuralnetworkmethodisbasedonlargetrainingset.Ifthetrainingsetissmall,itisdifficulttodeterminethestructureoftheneuralnetworkandmayhavethelocalminimumproblemwhichwillreducetheefficiency.Sotheperformanceofthismethodneedstobeimprovedwhenhavingsmalltrainingset.Supportvectormachine(SVM)isbasedonstatisticallearningtheoryandstructuralriskminimizationprinciple.Itsmainideaismappingnonlinearproblemintheoriginalspacetoalinearprobleminhigh-dimensionalfeaturespace,thenfindsthehyperplaneandmaximizesthedistancebetweenhyperplaneandsupportvectorinhigh-dimensionalfeaturespace.ThisensuresthattheclassificationperformanceandgeneralizationabilityofSVMtobesignificantlybetterthanANN.2.PrinciplesandMethods2.1.Supportvectormachine(SVM)Supportvectormachine(SVM)isbasedonstructuralriskminimizationprinciple.Inotherwords,itgetsmaximumgeneralizationthroughminimizingtheupperboundofgeneralizationerrorrisk.ThemainideaofSVMistoachievethemaximumgeneralizationabilitybyconstructingtheoptimalhyperplanewhichhasthelargestdistancebetweenthedifferenttypesofsamplesetsinsamplespaceorfeaturespace[5].Inessence,SVMisanalgorithmmethod.Whenthesampleislinearlyseparable,SVMsolvesthemaximal-marginsolutioninthesamplespace.Whenthesampleislinearlynonseparable,SVMismappingsamplesettohighdimensionalspacethroughproperkernelfunction,thusthesamplesetislinearseparableinhighdimensionalspace.Inlinearseparablecondition,let),(iiyx,ni,...,2,1,dxR,}1,1{ybethesampleset.bxwxg)(isthegeneralformoflinearclassificationfunctionind-dimensionalspace.Thehyperplaneequationis0bxw(1)Normalizetheclassificationfunctionandmaketwosamplesmeet1)(xg.Thus,theclassesmarginis2||||w.Somaximizingtheclassesmarginequalsminimizing||||wor2||||w.Ifahyperplanecancorrectlyclassifyallthesamples,itmustmeet:0]1)[(bxwyi,ni,...,2,1(2)Ahyperplanewhichmeettheaboveconditionsandlet2||||wminimizeiscalledoptimalhyperplane.Thesolutionoftheoptimalhyperplanecanbetransformedintothefollowi