概率神经网络信用评价模型及预警研究

200555:100026788(2005)0520043206(,510632):(PNN),PNN,2000106.,PNN,100%;,69177%.,87174%.,200113,69123%.PNN.,PNNMLP(,98111%),YangPNN(74%),PNN.,PNN.:;;;:F830:AStudyonCreditScoringModelandForecastingBasedonProbabilisticNeuralNetworkPANGSu2lin(DepartmentofMathematics,JinanUniversity,Guangzhou,Guangdong510632,China)Abstract:Thearticleintroducesthemethodofprobabilisticneuralnetwork(PNN)anditsclassifyingprinciple.ItconstructsaPNNstructureforidentifiedtwopatternssamples.ThePNNstructureisusedtoseparate106listedcompaniesofourcountryin2000intotwogroups.Thesimulationsshowthat,theclassificationaccuracyrateofPNNtothetrainingsamplesisveryhighwhichisupto100%,buttheclassificationaccuracyrateofPNNtothetestingsamplesisverylowwhichisonly69177%.Therefore,theclassificationeffecttothepopulationtendstobadandtheaccuracyrateisonly87174%.Furthersimulatingresultsshowthepredictingaccuracyrateisonly69123%whenthePNNisusedtopredict13pre2distressedcompanieswhicharepublishedinadvancefromChinain2001.Therefore,PNNisnotsuitabletoidentifyanewsampleortocarryoutpredictingstudy.Theresearchalsoshowsthat,PNNisnotasgoodasMLP(tothesamedata,theclassificationaccuracyrateofthemultilayerperceptronis98111%).ButcomparewithYangsworkaboutPNNsclassification(theclassificationaccuracyrateis74%)effect,theclassificationeffectofthePNNstructuregivenbyhereisbetter.Therefore,asadiscussionofmethod,PNNstillhaveresearchvalue.Keywords:probabilisticneuralnetwork;creditscoringmodel;patternsclassification;financialpredicting:2004209216:(31906),(2004B10101033);(2004Z32D0231):,,(),,:,.1Bayes,(ProbabilisticNeuralNetwork,PNN).PNNSpecht[1]1990,,,PNN,.,,,©1995-2005TsinghuaTongfangOpticalDiscCo.,Ltd.Allrightsreserved..Bayes.BP,.BP,.Tyree[2](1995)4PNN:,4,;,,;,;.YangMarjorie[3](1999)PNN,122.122,33,89.122:,1133;1426;830.PNN,,PNN66%,74%.PNNHajmeer[4](2002)Simon[5](2001).[6,7](2003)(MLP),2000106200096,98111%79117%.MLP200113,100%.PNN[6]2000106,[6]:,,.,PNN87174%,100%,69177%.,PNN[6]200113,69123%.,PNNMLP([6],98111%,100%),PNN,,PNN,,.PNN.YangPNN([3],74%),PNN.,PNN.2[6],2000106,53STPT,200053STPT,53.[6],63,32STPT,31.43,21STPT,22.[6]4:x1=,x2=,x3=,x4=.,1:;2:.3(PNN),P(x|Gi),,.,,,([9]).,4420055©1995-2005TsinghuaTongfangOpticalDiscCo.,Ltd.Allrightsreserved..nG1,G2,,Gn,P(Gi)(i=1,2,,n,).x,P(x|Gi),Bayes,GiP(Gi|x):P(Gi|x)=P(x|Gi)P(Gi)P(x)=P(x|Gi)P(Gi)nk=1P(x|Gk)P(Gk),(1):xGi,P(Gi|x)P(Gj|x),ij;i,j=1,2,,n,(2)P(Gi),Bayes:xGi,P(x|Gi)P(Gi)P(x|Gj)P(Gj),ij;i,j=1,2,,n,(3),P(Gi)=1n,GiP(Gi|x):P(Gi|x)=P(x|Gi)P(Gi)P(x)=P(x|Gi)P(Gi)nk=1P(x|Gk)P(Gk)=P(x|Gi)nk=1P(x|Gk).(4):xGi,P(x|Gi)P(x|Gj),(5)P(x|Gj)[8]:^P(x|Gi)=1(2)n2i12exp12(x-i)T-1i-1i(x-i)(6)1PNNi,iGi,-1iGi(i=1,2,,n).Gi,Gj,PNN1[3].GiP(x|Gi),GjP(x|Gj).x,P(x|Gi)P(x|Gj),(5),xGi(i,j=1,2,,n,ij).4411PNN4PNN(1,2,1),PNN2.1,x=(x1,x2,x3,x4),4;163,63G1G2,P(G1)=32,P(G2)=31,^P(x|Gi)(i=1,2,),(6);2,2,,^P(Gi|x);4,(3).2,1^P(x|Gi),.2Gix^P(Gi|x),545©1995-2005TsinghuaTongfangOpticalDiscCo.,Ltd.Allrightsreserved.2PNN.1,xG1xG2,^P(x|G1)^P(x|G2),P(G1)P(G2),xG1xG2,^P(G1|x)^P(G2|x),:xG1.4121.1x1x2x3x4-0.7072-0.5350-93.17750.03280.59033.599717.91130.5016-1-1x10.4670.890-17.520-0.0064x20.8906.657-124.4350.0783x3-17.520-124.43594932.565-6.361x4-0.00640.0783-6.3610.0338x12.9429-0.40920.00011.5243x2-0.40920.21450.0002-0.5420x30.00010.00020.00000.0017x41.5243-0.54200.001731.4374x10.1380.3143.2430.129x20.3143.673-7.2880.525x33.243-7.288245.1491.208x40.1290.5251.2080.455x124.8533-2.4269-0.3850-3.2280x2-2.42690.59970.0506-0.1377x3-0.38500.05060.01060.0228x4-3.2280-0.13770.02283.2115PNN2.2,,10,20,0,0100%,PNN6420055©1995-2005TsinghuaTongfangOpticalDiscCo.,Ltd.Allrightsreserved.100%.,113,20,13,30123%,PNN69177%.,PNN106,13,12126%,87174%(2).2PNN12(63)0(0.00%)0(0.00%)0(0.00%)100%(43)13(30.23%)0(0.00%)13(30.23%)69.77%(106)13(12.26%)87.74%5PNN[6]200113.1343.3200113()()(%)()10000030.08-7.2-0.010.092000618-0.511.09-46.450.663600629-3.23-2.7119.930.064600813-2.96-3.1394.52-0.075000658-3.22-3.11.03-0.436000653-1.58-4.270.370.047000689-4.19-4.031.040.008000675-1.3-1.21.080.079000585-0.920.38-2.410.1910000008-1.69-0.662.55-0.0111000013-5.19-5.420.950.0212000048-0.270.12-2.090.6913000533-1.690.42-3.97-0.62,PNN134,30177%,69123%(4).4PNN(3)PNNPNN111811223()1923()1311101141111115111223()16111323()17114(30.77%)69.23%PNN13,PNN,,745©1995-2005TsinghuaTongfangOpticalDiscCo.,Ltd.Allrightsreserved..6PNN,PNN,2000106.:200053STPT,53.[6].[6]:,,.,PNN87174%,100%,69177%.,PNN,.,PNN,,PNN,,,.,200113,PNN4,30177%,69123%..,PNNMLP([6],98111%,100%).YangPNN([3],74%),PNN.,PNN.:[1]SpechtDF.Probabilisticneuralnetwork[J].NeuralNetwork,1990,3(2):109-118.[2]TyreeEricW,LongJA.Assessingfinancialdistresswithprobabilisticneuralnetwork[A].ProceedingsoftheThirdInternationalConferenceonNeuralNetworksintheCapitalMarket[C].1995.[3]YangZY,MarjorieB.PlattandHarlanD.Platt.Probabilisticneuralnetworksinbankruptcyprediction[J].JournalofBusinessResearch,1999,44:67-74.[4]HajmeerM,BasheerI.AprobabilisticneuralapproachformodelingandclassificationofbacterialgrowthPno2growthdata[J].JouralofMicrobiologicalMethods.2002,51:217-226.[5]LaurentSimonM.NazmulKarim.Prob