CRMinconsumerbanking

CRMinconsumerbankingDr.YanningZhangDecember,2001CRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineCRM(CustomerRelationshipManagement)isallaboutunifiedviewofeachcustomerCRMOverviewProductsChannelTreatmentsCustomerCRMMarketisstillfastgrowing,evenwhenoverallmarketisslowingdown.ProjectedCRMmarketwillbeover$10Billionin2003,from$3.3Billiononlyin1999.(E-commence)CRMOverviewCRMconceptisnotanewone,butitbecameahottopicinlast5yearsAcceleratedconsolidationinbankingindustryRecentmajormergersexpandedcustomerbaseMoreproductsavailabletoconsumersTechnologyallowseasiercollectionandaccessofcustomerdataChinamarketfromERPtoCRMCRMOverviewCRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineBankCRMCharacteristicsBankingisextensiveAlmosteveryfamilyhasabankaccountBankdataisdataintensiveBankdataissensitiveandproprietaryBankdataismulti-dimensionalBankindustryishighlyregulatedPersonalizedtechnologiescantailormessagetoindividualcustomerCollectdataasmuchaspossiblefromvariouschannelsCRMStartingpoint-unifiedcustomerviewBankTelephoneBankingATM/KioskPersonalBankerOnlineBankingIn-branchTellerBankCustomerLifeCycleAcquisitionActivationRelationshipManagementTargetCustomerNewCustomerInitialCustomerHighPotentialHighValueLowValueInvoluntaryChargeoffVoluntaryPayoffProspectPrescreenUnderwritingCollection/RecoveryBankCustomerLifeCycleAcquisitionActivationRelationshipManagementTargetCustomerNewCustomerInitialCustomerHighPotentialHighValueLowValueInvoluntaryChargeoffVoluntaryPayoffProspectPrescreenUnderwritingCollection/RecoveryBureaudataDemographicdataCensusdataApplicationdataPricingdataAccountactivationdataBehaviordataMonitoringbureaudataCollectiondataDataCollectedandAnalyzedAlongtheLifeCycleCRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineTwotypesofCRMFront-endcustomerexperience–OperationalCRMBack-endcustomerunderstanding–AnalyticCRMTwoTypesofCRMAnalyticCRMOperationalCRMCustomerAnalyticCRM-BrainofcustomermanagementDataMetricsAnalysisImprove-mentCharacteristicsofAnalyticCRMExtensivedatacapturing,processinganddatasourcelinkingIn-depthanalysisandmodelingIntegrationwithoperationalCRMsystemAnalyticCRMprocessisaniterativeprocess73%U.S.companiesclaimcollectinginformationaboutindividualsDatacollectionisonlythefirststep-Importanttoagreeonwhytocollect,sincedifferentpartofcompanieshavedifferentmotivationofcollectinginformationAnalyticCRM-DataProcessBankdatawarehouse-thebasicbuildingblockAnalyticCRM-DataProcessBillingfileMasterfileTransactionfileApplicationfileBureaufileMarketingfileDWDataIntegrationDatawarehouseisthecentralplacetohave360degreeviewofcustomersCommonmistakesindatawarehouseLackoflongtermvisionNocommunicationbetweendepartmentsLackofcooperation,only37%sharedatawithotherdivisionsNosufficientqualitycontrolConflictdataentrypointtocreateunnecessaryconflictNosufficientdocumentationfordataelements,especiallyformissingvaluesAnalyticCRM-AnalysisCompaniesneedtoagreeuponthemeasurementofsuccessBanksuccessmeasurements:CampaignresponserateAnnualizedlossesROA(ReturnonAssets)ROE(ReturnonEquity)orROI(ReturnonInvestment)AnalyticCRM-MetricsCRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineKeycomponentsofCRManalysisAnalyticCRM-In-depthAnalysisDataSummaryDataExplorationModelingDataMiningDecisionDataSummaryisthefoundationInmanycases,datasummaryis>70%oftheworkCommonissuesConfirmthedataqualityVerifymissingvaluesCheckvariablecodingContinuousvariablecodingMissingvaluesSpecialvaluesDiscretevariablecodingAvoidmis-intepretnumericvaluesBlankmayhavespecialmeaningAnalyticCRM–Analysis/DataSummaryDataexplorationiscanbeextensive,inmanycases,thereareover1000variablestoexploreOLAP(OnlineAnalyticProcessing)isoftenusedinvariableexplorationCustomerprofilingisamustTherearemanystatisticalmeasurestodealwithunivariateanalysisMean(Average)VarianceorStandardDeviationMediumKurtosisEtc.Inmanycases,variabletransformationisnecessary.AnalyticCRM–Analysis/OLAPandVariableExplorationCommonvariableshapesAnalyticCRM–Analysis/VariableExplorationNormalS-ShapeLinearExponentialDecreasingConvex(increasing)SkewednormalAnalyticCRM–Analysis/ModelingConceptofmodeling–mimicthefuturebypastdata.Usethedataavailableandfindpatternsfromthepastandthenapplythepatternsinthemodel.PastPresentFuturePastDataBeginsPastDataEndsPerformancePeriodStartsPerformancePeriodEndsPerformancePeriodPredictivePeriodAnalyticCRM–Analysis/ModelingModelingStagesPerformanceDefinitionDataCollectionSamplingUnivariateAnalysisModelingModelValidationModelImplementaionAnalyticCRM–Analysis/ModelingPerformancedefinitionNeedtoconsiderproperexclusionsPerformancemeasuresiswhatoneneedstopredictPerformancedefinitionneedstobecloselytiedinwithbusinessobjectivesConsensusamongallplayersSamplingBankindustryisdatarich,however,inmanycases,thesamplesindifferentperformanceisunbalancesExamples,respondersinmarketingmodel,badsincreditriskmodelStratifiedsamplingisasolutionforunbalancedsamplesUnivariateAnalysisUseinformationvaluetheorytoexploredataSegmentation/populationsplitInmanycasesthearedifferentpatternswithinsubpopulationwhichwilldecreasetheoverallmodelvalueAnalyticCRM–Analysis/ModelingTherearemanystatisticaltoolstodevelopmodelsLogisticregressionDecisionTreesNeuralNetworkGeneticAlgorithmThereisnouniversally‘best’methodology.DifferentmethodwilltailortocertainapplicationsFraudmodelingCreditmodelingMarketresponsemodelingConsiderationstomakedesicionAnalyticCRM–Analysis/ModelingModelingTechnique–DecisionTreesGoodBadDetermine‘good’or‘bad’byaskingseriesofquestions(YorN)CARSorCHAIDAnalyticCRM–Analysis/ModelingModelingTechnique–NeuralNetworkInputsNeuralnetworksarebasedonlooselycomputermodelsofhowhumanbrainswork.Mayhavemanylayers,whichisequivalenttovariousmathematicalfunctionsInmanyways,similartologisticregressionHiddenLayersOutputNeuronsAnalyticCRM–Analysis/ModelingModelStrengthMeasuresInformationValueIV=(%good-%bad)*log(%good/%bad)GoodBadscoreInfoValueDivergence(continuousformofInformationValue)[Avg(good)-Avg(bad)]^2[Var(good)+Var(bad)]/2Divergence=AnalyticCRM–Analysis/ModelingModelStrengthMeasuresKolmogorov-SmirnovStatisticKS=max|cumu%ofgood-cumu%ofbad|GoodscoreBad100%0%Allthemodelstrengthindicatorsarepositivelycorrelated.AnalyticCRM–Analysis/ModelingPredictivestrengthisnoteverything.Otherconsiderationshavetobetakenseriously.ModelStabilityPalatabilityLegalComplianceImplemen-tationPredictiveStrengthAnalyticCRM–Analysis/ModelingCautionsinmodelingOverfitting–modelingismuchmorethanjustmathematicalproblem,thebestmathematicalmodelmaycollapsebecauseofoverfittingModelComplexityDevelopmentDataValidationDataBecauseofoverfitting,themodelperformspoorlyonvalidationdataAnalyticCRM–Analysis/ModelingModelValidationValidationisanassuranceofvalidityofthemodelTwobasicformsofvalidationOut-of-timevalidationTypicallyselectdatafromanewperiodafterthedevelopmentMostdesirablewayforthevalidationMaynotalwayspossibleIn-the-timevalidationWhenOut-of-timevalidationisnotfeasibleManyvariations:Holdoffsample,20%to50%JackknifeCross-validationBootstrapAnalyticCRM–Analysis/ModelingModelImplementaionModelspecification–makesureoperationswillunderstandthemodelImplementationscheduleIntegrationwithoperationalCRMUnifiedapproach–communicationiscriticalAnalyticCRM–Analysis/DataMiningDataminingisaprocessofexplorationandanalysislargequantitiesofdatainordertodiscovermeaningfulpatterns.DataminingactivitiesClassificationEstimationPredictionClusteringDescriptiveandvisulizationDataminingisfeasible,thankstothecomputertechnologyadvanceCRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineAnalyticCRM–Analysis/DecisionSupportAlltheCRMtoolsservethepurposeofenhancemanagementdecisionsDecisionstrategyutilizesanalyticCRMresultstooptimizebusinessdecisionsAnalyticCRM–Analysis/ModelingCaseStudyPortfoliosummaryBankofSanFranciscoCreditCardProgramstarted6yearsago950,000accountswithtotalbalanceof$5,342,000inoutstandingreceivableRecentlycharge-offrateishigh.Annualizedcharge-offrateis6.3%Bankdeterminestotightenpolicybasedonnewmodel.AnalyticCRM–Analysis/ModelingPerformancedefinitionUsing12monthsperformanceExclusionsPastPresent10,960Bads(8%BadRate)137,500Goods4/19984/20005/20005/2001PerformancePeriodPredictivePeriodAnalyticCRM–Analysis/ModelingUnivariateAnalysisUsefulvariablesAccountdataApplicationdataBureaudataTransactionaldataThirdpartydataTopvariablesUtilizationinthelast12monthsDelinquencyinthelast6monthsLengthofjobTimeofbureaufileAnalyticCRM–Analysis/ModelingExampleRankorderallvariablesTop120variableswillgotothenextstepUselogisticregressiontodevelopmodelSplitpopulationtoModelA:DelinquentcustomersModelB:AnydelinquentcustomersGainschartsAnalyticCRM–CommonMistakesLackofmanagementsupportMiscommunicationMismatchofbusinessknowledgeandtechnicalskillRushintoperformancedefinitionIgnoreoperationallimitationDonotunderstanddataTooheavilyrelyonmathematicalmodelCRMOverviewBankCRMfeaturesAnalyticCRMProcessOLAP,ModelingandDataMiningDecisionstrategiesConclusionOutlineCRMisallaboutunifiedcustomerviewAnalyticCRMisthebrainofallCRMeffortsCRMneedstobeinlinewithcustomerlifecycleDataexploration,modelinganddataminingarethemostimportantcomponentsofCRMAvoid‘Garbage-inGarbage-out’CRMhasbecloselylinktobusinesspracticesConclusion演讲完毕，谢谢观看！