【行业研报】JF-可预见的不平等机器学习对信贷市场的影响_市场营销策划_重点报告20230803_d.docx

《【行业研报】JF-可预见的不平等机器学习对信贷市场的影响_市场营销策划_重点报告20230803_d.docx》由会员分享，可在线阅读，更多相关《【行业研报】JF-可预见的不平等机器学习对信贷市场的影响_市场营销策划_重点报告20230803_d.docx（49页珍藏版）》请在第一文库网上搜索。

1、TheJourna1ofFINANCETheJOUma1OfTHEAMERICANFINANCEASSoC1ATIoNTHEJOURNA1OFFINANCE-VO1.I-XXVII,NO.1FEBRUARY2023Predictab1yUnequa1?TheEffectsOfMachine1earningonCreditMarketsANDREASFUSTER,PAU1GO1DSMITH-PINKHAM,TARUNRAMADORAI,andANSGARWA1THERABSTRACTInnovationsinstatistica1techno1ogyinfunctionsinc1udingcre

2、dit-screeninghaveraisedconcernsaboutdistributiona1impactsacrosscategoriessuchasrace.Theoretica11y,distributiona1effectsofbetterstatistica1techno1ogycancomefromgreaterf1exibi1itytouncoverstructura1re1ationshipsorfromtriangu1ationofotherwiseexc1udedcharacteristics.UsingdataonU.S.mortgages,Wepredictdef

3、au1tusingtraditiona1andmachine1earningmode1s.WefindthatB1ackandHispanicborrowersaredisproportionate1y1ess1ike1ytogainfromtheintroductionofmachine1earning.Inasimp1eequi1ibriumcreditmarketmode1,machine1earningincreasesdisparityinratesbetweenandwithingoups,withthesechangesattributab1eprimari1ytogreater

4、f1exibi1ity.INRECENTYEARS,NEWPREDICTIVEstatistica1methodsandmachine1earningtechniqueshavebeenrapid1yadoptedbybusinessesseekingprofitabi1itygainsinabroadrangeofindustries.See,forexamp1e,Agrawa1,Gans,andGo1dfarb(2018).Academiceconomistsa1soincreasing1yre1yonsuchtechniques(e.g.,Be11oni,Chernozhukov,and

5、Hanscn(2014),Varian(2014),K1cinbergeta1.(2018a),Mu11ainathanandSpiess(2017),Chernozhukoveta1.(2017),AtheyandImbens(2017).DOI：10.1111jofi.130902023theAmericanFinanceAssociationThepaceatwhichthesetechno1ogieshavebeenadoptedhaspromptedconcernsthattherisksassociatedwiththeirAndreasFusterisatEPF1,SwissFi

6、nanceInstitute,andCEPR.Pau1Go1dsmith-PinkhamisatYa1eSchoo1ofManagement.TarunRAmadOraiisatImperia1Co11ege1ondonandCEPR.AnsgarWa1therisatImperia1Co11ege1ondon.WethankAmitSeru(theEditor)andthreeanonymousrefereesforthoughtfu1mments.Wea1sothanktoTobiasBerg,Phi1ippeBracke,JediphiCaba1,JohnCampbe11,Frances

7、coD,Acunto,AndrewE11u1,KrisGerardi,AndraGhent,JohanHombert,Ra1phKoijen,Andres1iberman,Gonza1oMaturana,AdairMorse,KarthikMura1idharan,Danie1Paravisini,JonathanRoth,JannSpiess,JeremyStein,Danie1Streitz,JohannesStroebe1,BorisVa11ee,StijnVanNieuwerburgh,andparticipantsatnumerousconferencesandseminarsfor

8、he1pfu1discussionsandcomments.WethankKevin1ai,1u1iu,andQingYaoforresearchassistance.FusterandGo1dsmith-Pinkhamwereemp1oyedattheFedera1ReserveBankofNewYork,whi1emuchofthisworkwascomp1eted.Theviewsexpressedarethoseoftheauthorsanddonotnecessari1yref1ectthoseoftheFedera1ReserveBankofNewYorkortheFedera1R

9、eserveSystem.Incomp1iancewithTheJourna1ofFinancedisc1osurepo1icy,wehavenonf1ictsofinteresttodisc1ose.Correspondence1TarunRamadorai,Imperia1Co11ege,ExhibitionRoad1SouthKensington,1ondonSW72AZandCEPR；e-maiVt.ramadoraiimperia1.ac.uk.usehavenotbeencarefu11yeva1uated,inc1udingthepossibi1itythatanygainsar

10、isingfrombetterstatistica1mode1ingmaynotbeeven1ydistributed.See,forexamp1e,ONei1(2016),Hardt,Price,andSrebro(2016),K1einberg,Mu11ainathan,andRaghavan(2016),andK1einbergeta1.(2018b).Inthispaper,westudythedistributiona1consequencesoftheadoptionofmachine1earningtechniquesinthedomainofhouseho1dcreditmar

11、kets.Wedosobydeve1opingbasictheoretica1frameworkstoana1yzetheseissues,conductingempirica1ana1ysisona1argeadministrativedatasetof1oansintheU.S.mortgagemarket,andundertakinganinitia1assessmentofpotentia1economicmagnitudesusingasimp1eequi1ibriummode1.Theessentia1ideaUnder1yingourpaperisthatamoresophist

12、icatedstatistica1techno1ogy(inthesenseofreducingpredictivemean-squarederrorMSE)producespredictionswithgreatervariancethanamoreprimitivetechno1ogy.Whenapp1iedtothecontextwestudy,ourinsightisthatimprovementsinpredictivetechno1ogyactasmean-preservingspreadsforpredictedoutcomes一inourapp1ication,predicte

13、ddefau1tpropensitieson1oans.Academicworkthatapp1iesmachine1earningtocreditriskmode1inginc1udesKhandani,Kim,andIx)(2010)andSirignano,Sadhwani,andGiesecke(2023).Thismeansthatsomeborrowerswi11a1waysbeconsidered1essriskybythenewtechno1ogy(winners)，whi1eotherborrowerswi11bedeemedriskier(1osers),re1ativet

14、otheirpositionunderthepreexistingtechno1ogy.Thekeyquestionthereforeishowthesewinnersand1osersaredistributedacrossimportantcategoriessuchasrace,income,orgender.Weattempttoprovidec1earerguidancetoidentifythespecificgroupsmost1ike1ytowinor1osefromthechangeintechno1ogy.Todoso,wefirstconsiderthedecisiono

15、fa1enderwhousesasing1eexogenousvariab1e(e.g.,aborrowercharacteristicsuchasincome)topredictdefau1t.Wefindthatwhowinsor1osesdependsonboththefunctiona1formofthenewtechno1ogyandthedifferencesinthedistributionofthecharacteristicsacrossgroups.Perhaps,thesimp1estwaytounderstandthispointistoconsideraneconom

16、yendowedwithaprimitivepredictiontechno1ogythatsimp1yusesthemean1eve1ofasing1echaracteristictopredictdefau1t.Inthiscase,thepredicteddefau1tratewi11bethesamefora11borrowers,regard1essoftheirparticu1arva1ueofthecharacteristic.Ifamoresophisticated1ineartechno1ogythatidentifiesthatdefau1tratesare1inear1ydecreasinginthecharacteristicbecomesavai1ab1etothiseconomy,groupswithbe1owaverageva1uesofthecharacteristicwi11c1ear1ybepena1izedfo11owin