整合类神经网路与分类回归树在建构企业危机诊断模式上之应用

Vol.5,No.4,Oct2002WEBJOURNALOFConstructionofEnterpriseDistressDiagnosisModelbyUsingtheIntegrationofNeuralNetworkandClassificationandRegressionTreesApproachChih-ChouCHIUTe-NienCHIEN:[e-mail]-NienChienInstituteofCommerceAutomationandManagementNationalTaipeiUniversityofTechnologyAbstractRecently,wehaveseenthattheoccurrencepossibilityoffinancialdistressincreasesgraduallybecauseofthedramaticallychangeofglobaleconomicenvironment.Toforecastthefinancialdistress,companiesalwaysestablishesapredictivemodeloffinancialdistressbyexpandingthesamplesandthedefinitionoffinancialdistress.However,judgingfromthedefinitionoffinancialdistress,themeaningoffinancialdistresscompanyliesonthestockcompanieslistedinTaiwanstockexchangecorporations.ThemainobjectiveofthisstudyistoexploretheperformanceofenterprisedistressdiagnosisbyintegratingtheartificialneuralnetworkswithClassificationandRegressionTrees(CART)technique.Besides,bothfinancialcapitalindicatorandintellectualcapital(IC)indicatorareincludedinthemodeltomeasuretheassetsofcompanies.Astheresultsreveal,theICindicatorsareveryimportantintheenterprisedistressdiagnosis,especiallyinhigh-techenterprise.Inadditions,wefindoutthatbothtraditionalfinancialindicatorsandICindicatorssignificantlyinfluencethediagnosticcorrectnessofenterprisedistressbyapplyingourproposedapproach.Moreover,resultsfromthepresentstudyindicatethattheproposedcombinedapproachpredictmuchaccurateandconvergemuchfasterthanthattheconventionalneuralnetworkapproach.Intheotherwords,withoutsuchgoodinitialestimatefromCART,aneuralnetworktakesalongtimetoachieveanaccurateresult.KeywordsIntellectualcapital,enterprisedistressdiagnosis,decisiontree,neuralnetworks561965CravenandShavlik,1997;Chungetal.,1999ClassificationandRegressionTrees,CARTCARTBreiman1984recursiveruleGrossmanandPoor,1996;Kuhnertetal.,2000;Sorensenetal.,2000Vol.5,No.4,Oct200257Beaver1966Foster19781990200011Beaver(1966)Altman(1968)Ohlson(1980)CoatsandFant(1993)KohandTan(1999)Ahnetal.(2000)(1990)(1996)(1999)(2000)(1983)Gilson(1989)(1992)(1993)(1996)(2000)58IntellectualCapital,I.C.JohnKennethGalbraith1969KaleekiMasoulas,19981991StewartStewartStewartStewart,1994;1997KaplanandNorton1996Bontis1998Ulrich1998Agor1997BrainSkillManagement,BSMBSMEdvinssonandMalone1997ClassificationandRegressionTrees,CARTBreiman1984recursiverule1960AIDMorganandSonquest,1969ID3Quinlan,1986CHAIDKass,1980FACTLohandVanichsetakul,1988Vol.5,No.4,Oct200259CARTOhmannetal.19961254ID3NewIdPRISMCN2C4.5ITRULENewId43%~48%CARTMessierHansen18ID3Quinlan,1988CARTpruningprocedureCART1994AIDSUPPORTCART1996CARTrefinedutilityfunctionspecificationmethodclusteranalysisCART22CARTGrossmanandPoor(1996)CARTMarkhametal.(1998)CARTJust-in-timeJITKuhnertetal.(2000)CARTMARSSorensenetal.(2000)CART(1994)AIDSUPPORTCART(1996)CARTFreeman60andSkapura1992weight1X1X2XnZjf(Zj)yjW1jW2jWnj11X1X2…XnWijZj=ÓWijXifZjyjRumelhartetal.,1985Fishetal.,1995;BerryandLinoff,1997Vellidoetal.,1999Vellidoetal.19991992199878Back-PropagationNeuralNetwork,BPN1996199920003Vol.5,No.4,Oct2002613OdomandSharda(1990)1970198265TamandKiang(1992)Logit595919CoatandFant(1993)94188Altmanetal.(1994)KohandTan(1999)165(1996)(2000)LogitCARTBPNCARTBPNBPNCARTpruningprocedurecrossvalidationCARTruleCARTCARTCARTCARTCART62heterogeneityGiniCARTmaximumtreeCARTerrorrateerrorcostCARTresubstitutionestimateCARTCARTCARTFishetal.,1995supervisedlearningforwardinputlayerhiddenlayeroutputlayer22Vol.5,No.4,Oct200263wij2localminimumFreemanandSkapura,1992Zhangetal.1998Davies1994trialanderrorgradientsteepestdescentmethodÄWij)W(WijijE¶¶h-=Δ11çlearningrateE∑-=2)(21jjATEjTjA64localminimum1FreemanandSkapura,1992Vellidoetal.,1999CARTBPNCARTBPNCARTCARTCART3llCARTCARTBPN3TaiwanEconomicJournalDataBank,TEJAVol.5,No.4,Oct200265277B136CARTBPNMSExcelSPSSSalfordSystemsCART4.0VestaQnet199870CART704624CARTvariableimportance1361CART454CART-%100.0074.4371.0564.2759.8455.235CART-1IF=48.095Class=02IF48.095Class=141365CART48.09548.0956646CART6CART87.5%4CART-6CART-121083.33%216.67%1218.33%1191.67%87.50%Horniketal.1989Zhangetal.1998BPNCARTBPN6111213141551Rumelhartetal.19860.010.10.20.3RMSErootmeansquarederror0.00013000RMSE7Node1Class=0=48.095N=46TerminalNode1Class=0N=23TerminalNode2Class=1N=23Vol.5,No.4,Oct200267TrainingRMSETestingRMSE0.010.20130.18790.10.19440.17960.20.18690.1749110.30.17770.17150.010.20130.18780.10.19550.18010.20.18320.1739120.30.17860.17960.010.20100.18750.10.19630.18110.20.18420.1762130.30.17820.17080.010.20080.18730.10.19690.18540.20.18720.1797140.30.18000.17260.010.20100.18740.10.18800.17780.20.18290.1754150.30.17790.17197{6-13-1}61310.3RMSE5{6-13-1}RMSE5RMSEBPN891.67%56-13-1BPNRMSE-688BPN-121191.67%18.33%1218.33%1191.67%91.67%CARTBPNCARTBPN761CART13141516171RMSE0.0001300097-15-11510.1RMSE6RMSE6RM