基于人工神经网络的虚假财务报表警示系统

上海交通大学硕士学位论文基于人工神经网络的虚假财务报表警示系统姓名：陆金国申请学位级别：硕士专业：工商管理（MBA）指导教师：徐博艺20041017NEURALNETWORKBASEDFALSEFINANCIALSTATEMENTSSPOTTINGSYSTEMABSTRACTThispaperisaboutacomputerassistedfalsefinancialstatementsoflistedcompaniesspottingsystemwithartificialneuralnetworks(ANN).Theemphasisistousedataminingtechnologyonfinancialanalysis.ANNisoneofthemostrepresentativemethodsofAI.Itsapplicationisverypopularinclassicalengineeringfieldsuchaspatternrecognition,signalprocession,controlsystemandetc.RecentlyitisalsoveryactiveinsomeWesterncountriesinsocialsciencearea,suchasbanking,insuranceandotherbusinessadministrationareas.Anartificialneuralnetworkissetupthroughalargenumberofsamples,soithasaremarkablelearningabilityduetoitscharacteristicsofself-organization,self-adaptationandself-learning.ANNsaregoodatprocessinglargevolumedata,especiallysuitablefordealingwithmulti-factorprojectswithvague,imperfectinformation.Spottingandrecognizingthefalsefinancialstatementsthroughfinancialratioanalysisistechnologicallyfeasible.Usuallytherearealotofcontradictionsanddoubtfulpointsinafalsefinancialstatement.Mostcheatingcompaniesaresufferingfromcashdeficiency.Thesecompaniesareeagertodefraudoflistingqualificationthroughmakingupprofits,ortoseekexorbitantprofitthroughhandlingstocksecondmarketpriceofcompany.Andbecausethosefalseprofitslacksupportofcorrespondingcash,theyusuallyappearinvariousformsof“softassets”suchastheaccountreceivable,inventory,constructioninprocess,prepaidexpenses,andetc.Theenormousamountinvolvedinfalsefinancialstatementsofferedtechnologicalfeasibilityofquantitativemethodsoffinancialstatementassistingdetectionsystem.Insummary,thelistedcompanywhichmadeupfalseprofitsinthefinancialstatementsusuallyhasthefollowinguncommoncharacteristics.a)Thegrowingrateofaccountreceivableismuchhigherthansalesgrowingrate.Theaccountreceivablerecyclingperiodandrateareverylowb)Accountreceivableaccountsforaverylargeproportionofthetotalassets.c)Inventorygrowsfarfasterthansale,sellingcostandaccountpayableassets.Theproportionofinventoryvaluetothetotalassetsismuchhigherthantheindustryaverageleveld)Majorproductmarginisunusuallyhigherthanindustryaveragelevel.e)Thecashflowfailstocatchupwiththepaceofsalerevenue.f)Fixedassetvaluerisessharplycomparingtototalassetvalue.g)Sellingcostdropssuddenlycomparingtosellingrevenueh)Proportionofmoneyandmonetaryequivalenttothetotalassetvalueislowerthanindustryaveragelevel.i)Totalcorporationloangrows,whiletheinterestexpensesdrops.Theinputfactorsoftheartificialneuralnetworkbasedfinancialshenanigansdetectionsystemaremainlyselectedaccordingtotheabove-mentionedcharacteristicpointsfoundinfalsefinancialstatementsoflistedcompanies.Thesysteminputfactorsaredividedintothreepartsofstaticratios,company'shistoricaldata,andindustryaverageratios.Astaticratioconsistsofthebasicdataitemsofthefinancialstatementsincurrentyear,includingtotalassets,corebusinessincome,corebusinesscost,fixedassets,inventory,monetaryfund,accountreceivable,accountpayable,constructioninprocess,operationalcashflowandprepaidexpenses,andetc.Itisusedinthestaticanalysis.Company'shistoricaldataconsistofcompany'slastthreeyears'basicdataitems,andareusedintrendanalysis.Anindustryaverageratioisusedincomparisonwithothercompanyinthesameindustry.ThedevelopingtooloftheprojectisiDataAnalyzer(iDA)softwarepackageincludedinthebookDataMining--ATutorialBasedPrimer(byRichardJ.RoigerandMichaelW.Geatz).Itisanadd-oncomponentuponMicrosoftExcel,itcouldtakethefulladvantageoftheextremelyabundantpowerfulfunctionsandconvenientinterfaceofdatamanipulationandpreprocess.Itsdemonstrationeditionisverysuitableforprototypedevelopmentofadataminingproject.Artificialneuralnetworkfalsefinancialstatementspottingsystemhasbeenapreliminarysuccessprojectalready,thoughithasonlyverylimitedtrainingandtestingdataset.Itneedsalotoftimeandeffortstomakeitabettersystem.KEYWORDS:artificialneuralnetwork,datamining,artificialintelligence,financialstatementanalysis,informationtechnology12004101722004101720041017MBA111””([6])1-1[7]MBA2MBAMBAMBA3GALIDEA(BackpropagationAlgorithm,BP)BPMBA412(3BP(4)BP4.1iDAMBA5BPMBA6MBA7‹x,t›‹x,t›MBA8MBA9MBA101283.23.4,200020002590.54MBA11MBA126620MBA133-4199720003-4199720003-4119219991226%31998[4],[4][5]12345678MBA149,,MBA153-1MBA164.1iDA4.2CRichardJ.RoigerMichaelW.Geatz(DataMining—ATutorial-BasedPrimer)TheiDataAnalyzer(iDA)InformationAcumenCorpExceliDAExceladd-on,Excel4.1.2iDA[1]iDA123ESXESX4iDA5MBA17iDA6MBA18ESXExcelYesNYNoNoYesMBA191()(23(4)561999MBA2020011999811269,iDA138BPiDABP126MBA21LearningRateEpochs(Convergence)2-10.050.510000RootMeanSquaredError,RMSMBA22MBA230.25100.00100.9720.03600.9990000.0630.944MBA2481%70%MBA25BP1“600551.SH2001,200220045iDXESX2ESX345,PDFMBA26MBA27MBA20041017307MBA20041017