QUANTILE MODELING FOR WALLET ESTIMATION

STATISTICALMETHODSINECOMMERCERESEARCHChapter:QuantileModelingforWalletEstimationClaudiaPerlichIBMT.J.WatsonResearchCenter,YorktownHeights,NY,USASaharonRossetIBMT.J.WatsonResearchCenter,YorktownHeights,NY,USAAJOHNWILEY&SONS,INC.,PUBLICATIONCHAPTER1QUANTILEMODELINGFORWALLETESTIMATION1.1INTRODUCTIONE-commerceprovidesawiderangeofnewopportunitiesforCRManddirectmarketingincludingthepersonalizationofproductsandoffers,theanalysisofcustomerbehaviorfromweblogsaswellasnewcommunicationchannelssuchasemailandpersonalizedwebcontent.Avitalcomponentoftrulypersonalizedmarketingisanunderstandingofthecustomerspurchasingpower.Thetotalamountofmoneyacustomercanspendonacertainproductcategoryisavitalpieceofinformationforplanningandmanagingsalesandmarketingefforts.Thisamountisusuallyreferredtoasthecustomer’swallet(alsocalledopportunity)forthisproductcategory.Therearemanypossibleusesforwalletes-timates,includingstraightforwardtargetingofsalesforceandmarketingactionstowardslargewalletcustomersandprospects.Inamoresophisticatedsalesandmarketingenvi-ronment,thecombinationofclassicalpropensitymodelsforaparticularproductcategorywiththeknowledgeofthewalletforthesamecategorycandirectthesalesefforts:itallowsacompanytomarketnotonlytocustomersorpotentialcustomerswithalargewallet,butalsotothosewithahighprobabilityofbuyingspecificproductsinthecategory.DRAFTJune7,2007,11:02pmDRAFT2QUANTILEMODELINGFORWALLETESTIMATIONBycombiningthecustomerwalletestimateswiththedataonhowmuchtheyspendwithaparticularseller,wecancalculatetheshare-of-walletthatthesellerhasofeachcustomerforagivenproductcategory.Thisinformationallowsthesellertotargetcustomersbasedontheirgrowthpotential,acombinationoftotalwalletandshare-of-wallet.Theclassicalapproachoftargetingcustomersthathavehistoricallygeneratedlargeamountsofrevenueforthecompany(knownaslifetimevaluemodeling,seee.g.,[18])doesnotgiveenoughimportancetocustomerswithalargewallet,butsmallshare-of-wallet,whicharetheoneswithpresumablythehighestpotentialforrevenuegrowth.Share-of-walletisalsoimportantfordetectingpartialdefectionorsilentattrition,whichoccurswhencustomersincreasetheirspendinginagivencategory,withoutincreasingtheamountpurchasedfromaparticularcompany[12].Incertainindustries,customerwalletscanbeeasilyobtainedfrompublicdata.Forexample,inthecreditcardindustry,thecardissuingcompaniescancalculatethewalletsizeandrespectiveshare-of-walletusingcreditrecordsfromthethreemajorcreditbureaus[1].Formostindustries,however,nopublicwalletinformationisavailableatthecustomerlevel.Inthiscase,therearetwoapproachesusedinpracticeforobtainingwalletestimates:1.Top-Down:startsfromapublicaggregateestimatefortheoverallindustryopportu-nityinagivencountryandsplitsthisestimateacrosstheindividualcustomersusingheuristicsbasedonthecustomercharacteristics.Forexample,ifthecustomersarecompanies,theoverallopportunitycouldbedividedamongthecompaniespropor-tionallytotheirnumberofemployees.2.Bottom-Up:estimatesthewalletdirectlyatthecustomerlevel,usingheuristicsorpredictivemodelsbasedoncustomerinformation.Acommonapproachistoobtainactualwalletinformationforarandomsubsetofcustomers/prospectsthroughprimaryresearch.Amodelisthendevelopedbasedonthisdatatopredictthewalletfortheothercustomers/prospects.Althoughcustomerwalletandshare-of-wallethavebeenrecognizedasimportantcus-tomervaluemetricsinthemarketingandservicesliteratureforanumberofyears[5,15,6,8],thereisnotmuchpublishedworkonactualwalletmodeling.Thefewreferencesavailablearelimitedtocommercialwhitepapersfrommarketinganalyticsconsultingcompanies[1],averyrecentthesisproposal[4]andtworecentconferencepresentations[22,5].ThewhitepaperfromEpsilon[1]describesataveryhighlevelthemethodologythattheyusetoestimatewalletsforbothcustomersandprospectsatagivenproductcategory.DRAFTJune7,2007,11:02pmDRAFTDEFINITIONSOFCUSTOMERWALLET3Epsilonusesabottom-upapproachwhereasurveyprovidestheself-reportedcategorywal-letforasampleofcustomers.Theself-reportedwalletisusedtodevelopamultivariatelinearregressionthatcanbeappliedtothewholemarketorcustomerbase.Theydonotdescribewhichvariablesareusedinthemodelanddonotprovideexperimentalresults.In[4,5],theauthorsproposeaMultivariateLatentFactorModelthatabankcanusetoimputeacustomers’holdingsoffinancialproductsoutsidethebank,basedonacombinationofinternallyavailabledataandsurveyinformationaboutcustomers’holdingswithcompeti-tors.In[22],theauthorsdiscussthevalueofwalletestimation,reviewmethodologies,andcompareatop-downapproachwithtwobottom-upapproaches,whereagainasurveyprovidestheself-reportedcategorywalletforasampleofcustomers.Inthispaper,weaddresstheissueofpredictingthewalletforIBMcustomersthatpurchaseInformationTechnology(IT)productssuchasservers,software,andservices.Ourmodelsarebottom-up,withexplanatoryfeaturesdrawnfromhistoricaltransactiondataaswellasfirmographicdatatakenfromexternalsourceslikeDun&Bradstreet.Wetakeapredictivemodelingapproach,andattempttominimizethedependenceofourmodelsonprimaryresearchdata.InSection1.2,weaddresstheexactmeaningofwallet,surveyingseveraldifferentdefinitionsandtheirimplicatio