1/40Predictingmultilateraltradecreditrisks:

comparisonsofLogitandFuzzyLogicmodelsusingROCcurveanalysis

多邊貿易信用風險評估：使用ROC曲線分析

比較Logit與FuzzyLogic模型2/40Abstract本篇主要是應用

ROC(receiveroperatingcharacteristic)曲線分析比較Logit與FuzzyLogic

模型的performance。3/401.IntroductionDefaultproblemsoftradecreditapplicantsfallbasicallywithinthescopeofdiscriminationandclassificationproblems(Johnson&Wichern,1998).Altman(1968)andBeaver(1966)[Altman,1968,Beaver,1966]contributesignificantlytotheclassificationandpredictionofbusinesscreditworthinessbyadoptingaunivariatemodelandamultiplediscriminantanalysis(MDA),respectively.4/40然而，信度與效度需建立在大規(guī)模的限制與假設上。ToavoidthedefectsofMDA,Ohlson(1980)proposesthelogisticregression(Logit),whichisfoundtobethemostaccurateofthetraditionalstatisticalmethodsforpredictionandclassificationproblems.5/40貿易風險評估決策主要決定于人主觀性的意見及判斷。而這些價值與觀感通常會造成不準確或模糊不清的語意，也無法直接由傳統(tǒng)的數學方式表達呈現。Inlightofthis,sincethe1980sscholarshavetriedadoptingartificialintelligencetechniquestofindbetterclassificationtools.6/40FuzzyLogic(FL),formulatedbyZadeh(1965),providesasystematicwayofhandlingvagueandimpreciseinformationoninputdata,theireffectsonthesystem,andtheoutput.ManyotherresearchersalsoreportthattheFL-basedmodelisfoundtohavegoodpredictiveability(Dourra&Siy,2002;Lia,Mukaidonob,&Turksenc,2002[DourraandSiy,2002,Liaetal.,2002]).7/40Thereceiveroperatingcharacteristic(ROC)curveanalysisisoftenusedinbiomedicalandpsychophysicalapplicationstosummarizethediscriminatoryaccuracyofadiagnostictestaswellastocomparetheperformanceofdifferentmodelsforbinaryoutcomes(Lloyd,1998;Marzban,1998;Pepe,2000[Lloyd,1998,Marzban,1998,Pepe,2000]).8/402.FuzzyinferencemechanismGenerally,theinferenceprocessproceedsinthreesteps:fuzzification,inference,anddefuzzification.Inthefuzzificationstep,therealinputistranslatedfromnumericalvaluetofuzzymembershipvaluesineachoftheinputfuzzysets.9/4010/40Inthefuzzyinferencestep,‘if-then’rulesthatdefinethesystembehaviorareevaluated.Inthedefuzzificationstep,theweightedaverageoftheoutputfuzzymembershipresultistranslatedbacktoanumericalvalue.11/4012/40ROCcurveanalysis13/4014/4015/4016/4017/4018/4019/4020/4021/40D：違約N：未違約TrainingsetTPR：truepositiverateFPR：falsepositiverateTNR：truenegativerateFNR：falsenegativerate22/40AUCprob.(N>D)TheareaundertheROCcurve(c-index,c-statistic)Score23/403.Researchdesign24/40VariableselectionThisstudyemploysliteraturesurveyandexpertinterviewmethodstoinvestigatethekeypredictivevariables.25/4026/4027/4028/4029/404.ResultsThecostsofFNRcanincludeprincipal,interest,collectionfees,andlegalfees.AnFPRincludesthecostsofforegonebusinesssales.30/40Differentcutoffshavebeenusedinpriorstudiesformeasuringclassificationaccuracy.Forthehighestoverallpredictionaccuracypurpose,weusetheoptimalcutoffthatisdeterminedbythelargestAUC.（最佳的門坎值決定于最大的AUC）Cutoff門坎值31/4032/4033/40TheresultshowsthattheLogitmodelprovidesagoodfittothedataandtheestimateofthevariables'parametersinthemodelismeaningful.34/40FLmodelresults35/40TestingsummaryonthetrainingsetTestingsummaryonthetestingset36/40ROCcurvesformodelresults(trainingset)37/40ROCcurvesformodelresults(testingset)38/40ComparisonsofLogitandFLmodels(trainingset)39/40ComparisonsofLogitandFLmodels(testingset)40/405.ConclusionsTheFLmodelsexceedtheLogitmodelsintermsofoverallclassificationaccuracyandpredictionaccuracy.However,byincorporatingmeasurementintheformofROCcurves,wealsofindthatLogitoutperformsFLinclassifyingnon-defaul