Intermediation Frictions in Debt Relief: Evidence from CARES Act Forbearance

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Intermediation Frictions in Debt Relief:Evidence from CARES Act Forbearance You Suk Kim, Donghoon Lee, Tess Scharlemann, and James Vickery 2022-017 Please cite this paper as: Kim, You Suk, Donghoon Lee, Tess Scharlemann, and James Vickery (2022). “Intermediation Frictions in Debt Relief:Evidence from CARES Act Forbearance,” Finance and EconomicsDiscussionSeries2022-017. Washington: BoardofGovernorsoftheFederalReserve System, https://doi.org/10.17016/FEDS.2022.017. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Intermediation Frictions in Debt Relief: ∗ Evidence from CARES Act Forbearance You Suk Kim, Donghoon Lee, Tess Scharlemann, and James Vickery† March 8, 2022 Abstract We study the role of mortgage servicers in implementing the CARES Act mortgage forbearance program during the COVID-19 pandemic. Despite universal eligibility, around one-third of the nonperforming federally-backed loans in our sample fail to enter into forbearance. The relative frequency of these “missing” forbearances variessignificantlyacrossservicersforobservablysimilarloans, withsmallservicers and nonbanks, and especially nonbanks with small liquidity buffers, having a lower propensity to provide forbearance. The incidence of forbearance-related complaints by borrowers is also higher for these servicers. We also use servicer-level variation to estimate the causal effect of forbearance on borrower outcomes. Assignment to a “high-forbearance” servicer translates to a significant increase in the probability of nonpayment, which moves essentially 1:1 with the forbearance probability. Part of thisadditionalhouseholdliquidityisusedtopaydownhigh-costcreditcarddebt. Keywords: mortgage,forbearance,debtrelief,CARESAct,COVID-19,liquidity JELclassification: G21,G23,G28 ∗Wethankseminarparticipantsatthe2022ASSAmeetings,PhiladelphiaFedConferenceonConsumer Behavior in Credit and Payments Markets, 2021 AREUEA National Meetings, and the Federal Reserve Board, as well as Meta Brown (discussant), Gene Amromin (discussant), Susan Cherry (discussant), and manyFederalReservecolleaguesandindustrypractitionersforcommentsandinformationaboutinstitutionaldetails.Opinionsexpressedinthispaperarethoseoftheauthorsanddonotrepresenttheopinionsof theFederalReserveBoard,theFederalReserveBanksofNewYorkorPhiladelphia,ortheFederalReserve System. †Kim: FederalReserveBoard(You.Kim@frb.gov); Lee: FederalReserveBankofNewYork(Donghoon. Lee@ny.frb.org); Scharlemann: Federal Reserve Board (Tess.Scharlemann@frb.gov); Vickery: Federal ReserveBankofPhiladelphia(james.vickery@phil.frb.org).

1 Introduction Financialintermediariesoftenplayanimportantroleinthetransmissionofpublicpolicy, particularlyinthecaseofdebtreliefandemergencylendingprograms. Butmisalignedincentivesorotherfrictionsmaypreventpoliciesfrombeingimplemented“ontheground” asintended.1 In this paper we study the role of a particular type of intermediary — mortgage servicers — in implementing a large new government debt relief program providing forbearance to mortgage borrowers during the COVID-19 pandemic. We find that servicers significantlyaffectforbearanceoutcomesandtheamountofliquidityultimatelyprovided to borrowers, and that variation in servicer behavior is systematically related to servicer liquidityconstraints,sizeandorganizationalform. The forbearance program, authorized by the CARES Act in March 2020, allows borrowers with federally backed mortgages to temporarily pause their mortgage payments without incurring fees, penalties or unscheduled interest and without negative effects on their credit history. The borrower simply needs to attest to a hardship related to the pandemictoqualifyforforbearance;nodocumentationofincomelossisrequired. Despite this universal eligibility, we document that a significant number of federally backed mortgage borrowers became delinquent during the pandemic without successfullyenteringintoforbearance,andthattherelativefrequencyofthese“missing”forbearances varies significantly across mortgage servicers for otherwise identical loans. Our analysis focuses on Federal Housing Administration (FHA) and Veterans Administration (VA) mortgages, the segment of the mortgage market which serves the highest-risk borrowers and which, because of institutional factors, poses the greatest liquidity risk to servicers. Usingloan-leveldatafromeMBSweestimatethattheprobabilityofnotreceiv- 1ExamplesincludeloanstobusinessesunderthePaycheckProtectionProgram(Granjaetal.,2020), mortgagemodificationsundertheHomeAffordableModificationProgram(Agarwaletal.,2017a),and streamlinedmortgagerefinancingundertheHomeAffordableRefinancingProgram(Agarwaletal.,2015). 1

ing forbearance conditional on delinquency varies between 10% and 60% controlling for loan and borrower characteristics, with a weighted interquartile range of 15 percentage points. Several pieces of evidence indicate that this variation reflects servicer behavior ratherthanunobservedloanandborrowerheterogeneity. We then investigate sources of these cross-servicer differences, and what they tell us about the role of servicer incentives and other frictions. We find that smaller servicers, nonbanks,andinparticularnonbankswithlowcashbuffersatthestartofthepandemic, are less likely to provide forbearance to borrowers. These findings indicate that scale economies and liquidity constraints, among other economic forces, were important in shaping servicer behavior. For example, mortgage nonpayment presents a source of significant liquidity risk because the servicer of an FHA or VA loan is required to finance payments to investors while borrowers are nonperforming — this liquidity risk is most significantfornonbankswithoutaccesstogovernmentbackstopsandinsureddeposits. Given that past-due FHA and VA borrowers would have universally benefited from forbearance, servicer practices that limit forbearance uptake also result in lower borrower welfare. We also find direct evidence that borrowers are less satisfied with “lowforbearance” servicers, based on the incidence of borrower complaints related to mortgageforbearancesubmittedtotheConsumerFinancialProtectionBureau(CFPB). In the second half of paper, we use variation in servicers’ forbearance practices to identify the causal effect of forbearance on borrower outcomes. We sort servicers into high (above median) and low (below median) forbearance-availability groups based on the likelihood a delinquent loan received forbearance conditional on loan and borrower characteristics. Then we compare borrower-level outcomes at high- and low-availability servicers before and after the CARES Act in a difference-in-differences framework using dynamicmortgagedatalinkedtoborrowercreditreports. Our first finding is that forbearance causes mortgage nonpayment. The probability 2

that a borrower is past-due is significantly higher for borrowers at high-forbearance servicers, by as much as 5 percentage points at the peak of the forbearance wave in May 2020. This difference in the past-due probability between high- and low-availability servicersisalmostidenticaltothedifferenceintheforbearanceratebetweenthetwogroups of servicers, implying that essentially all of the additional forbearance induced by highforbearanceservicersresultsinborrowernonpayment. Asaresult,assignmenttoahighforbearanceservicersignificantlyincreaseshouseholdcashflowsduringthepandemic. We then examine how borrowers use the additional cash made available through forbearance by examining borrowers’ non-mortgage debt accounts. We find that borrowers withbelow-mediancreditcardbalancesathigh-forbearanceservicersreducedtheircredit cardbalancesbyaround$20relativetoborrowersatlow-availabilityservicers,equivalent toatreatmenteffectof$400peradditionalforbearance. Thiscreditcardpaydownisabout one-quarteroftheaverageforbearance-drivensavingsinmortgagepaymentsforborrowers at high-forbearance servicers. In contrast, there is no paydown of credit card debt for borrowers with above-median credit card debt, who may be more liquidity constrained and therefore more likely to use the additional funds for consumption. Although borrowers at high-forbearance servicers are more likely to miss mortgage payments, their credit scores did not decrease as a result, because nonpayment during forbearance is not reportedtothecreditbureaus. Thecausaleffectofforbearanceoncreditscoresiscloseto zero. Our findings suggest that policies that reduce frictions from servicers could benefit borrowers by increasing access to forbearance and reducing variation in borrower outcomes that is unrelated to borrower fundamentals. For example, one possibility would be auto-enrolment in forbearance for borrowers drawing unemployment insurance or thosethatbecomeseriouslydelinquentafterbeingcurrentpriortothepandemic.2 2Bywayofcontrasttothemortgageforbearanceprogram,CARESactstudentloanforbearance auto-enrolledallfederalstudentloanborrowers. 3

1.1 Related literature Our paper contributes to a growing body of research on forbearance during the COVID pandemic. Related work includes Cherry et al. (2021), An et al. (2021) and Zhao et al. (2020). This literature establishes that forbearance is significantly related to borrower characteristicsandthedepthoftheeconomicshockposedbytheonsetofCOVID-19,that borrowersexperiencingincomedeclinesweremorelikelytoenterintoforbearance(Zhao etal.,2020)andthatforbearancereducedinequality(Anetal.,2021). Likeus,thesepapers also document that a significant number of delinquent borrowers did not successfully enter into forbearance. Cherry et al. (2021) also find that non-banks offer forbearance at lower rates, studying variation in outcomes across large servicers for prime mortgages securitizedthroughFannieMae. We also contribute to a broader body of research studying financial frictions and incentives facing mortgage intermediaries, much of which studies the Great Recession and itsaftermath. ForexampleAgarwaletal.(2011)andKruger(2018)findevidencethatservicers were more likely to modify mortgages retained in their own portfolios compared to loans serviced for other investors and that servicers offered HAMP modifications at differentratesduetovariationinorganizationalstructureandincentives(Agarwaletal., 2017a). Aiello (2021) finds evidence that financial constraints facing mortgage servicers significantlyreducedtheirpropensitytoworkoutdelinquentmortgagesduringtheGreat Recession. OurresearchisalsorelatedtoresearchshowingthatintermediaryeffectswereimportantfortheimplementationofothertypesofreliefprovidedduringtheCOVIDpandemic, forexampleGranjaetal.(2020)whichstudiesthePaycheckProtectionProgram. 4

2 Forbearance and the CARES Act The CARES Act was signed into law on March 27, 2020, and included significant relief for mortgage borrowers. Homeowners with federally-backed mortgages became eligible for up to 180 days of forbearance, renewable for an additional 180 days upon request.3,4 Whileinforbearance,borrowerscanskiptheirmortgagepaymentswithoutaccruingunscheduledinterest,latefeesorpenalties,orriskingforeclosure. Missedpaymentsarealso notreportedtocreditbureausandthereforedonotaffecttheborrower’screditscore.5 Eligibility under the CARES Act is very broad, extending to any agency mortgage borrower experiencing a direct or indirect financial hardship related to the pandemic. Importantly, the borrower simply needs to attest to a hardship — no documentation or other proof of income loss is required. Forbearance is not automatic however, the borrowermustrequestitfromtheirservicer. TheCARESActissilentaboutwhatshouldoccurattheendoftheforbearanceperiod. IntheweeksafterthepassageoftheAct,however,regulatorsandthemortgageagencies statedthatarangeofoptionswouldbeavailable,andborrowerswouldnotberequiredto repay missed payments in a lump sum (e.g., Freddie Mac, 2020). In April 2020, the FHA announced a National Emergency Partial Claim program, under which most borrowers thatre-performafterexitingforbearancecantransferaccumulatedmissedpaymentsinto a subordinate interest-free note which is not due until the termination of the mortgage 3TheCARESActappliesdirectlyto“agency”mortgagesbackedbyFannieMae,FreddieMac,theFHA,VA, andotherfederalagencies,whichtogethermakeupabout70%ofUSmortgagedebt. Manynonagency borrowershavestillbeenabletoobtainforbearancefromtheirservicers,althoughCherryetal.(2021)find thatforbearanceratesareabout25%loweroutsideofthefederally-backedmarket,byexaminingloanson eithersideoftheconformingloanlimit. 4TheCARESforbearanceprogramsweresubsequentlyextendedinFebruary2021. Homeownersalreadyin forbearancebecameeligibleforafurthersixmonthsofforbearance,andtheenrollmentwindowtorequest forbearancewasextendedto6/30/2021(TheWhiteHouse,2021;FederalHousingFinanceAgency,2021). 5TheCARESActpermitsaninitialforbearanceofuptosixmonthsbutservicershavemoretypically grantedforbearanceinthreemonthincrements,requiringtheborrowertorenewmorefrequently. An industrypractitionertoldusthisreflectspriorhistoricalpractice,whenforbearancehasprimarilybeen usedasashort-termdisaster-relieftool. 5

throughapropertysale,refinancingorpayoff(DepartmentofHousingandUrbanDevelopment, 2020a,b).6 Fannie Mae and Freddie Mac announced a similar payment deferral option in May (Federal Housing Finance Agency, 2020). Since missed payments do not accrueinterest,deferraleffectivelyprovidesazero-interestloantotheborrower. Despitetheseassurances,therewassignificantuncertaintyandconfusionamongborrowers and servicers about post-forbearance options, particularly early in the pandemic. Anecdotal evidence also suggests that some servicers incorrectly told borrowers that a lump-sum repayment would be expected (e.g., Wall Street Journal, 2020; Consumer FinancialProtectionBureau,2021a,b). The analysis in this paper focuses on the $2 trillion of “government” mortgages insured by the FHA and VA. This segment of the mortgage market is of particular interest because it disproportionately serves low-income and high-risk borrowers, and because FHA loans in particular have a much higher forbearance and delinquency rate than the market as a whole. It is also the segment where intermediation frictions are likely to be mostsevere,becauseFHAloanspresentsignificantadditionalriskstomortgageservicers comparedtoothertypesofloans(seesection2.3). 2.1 Forbearance trends Figure 1 traces out the evolution of forbearance and delinquency over 2020. The top panel,whichisbasedoncreditbureaudata,showsthatforbearancewasrarepriortothe pandemic but increased sharply starting in April, just after the CARES Act is enacted. The aggregate forbearance rate peaked in May at 7.3 percent, and then declined slowly 6Moreover,theFHArequiresservicerstoevaluateallborrowersforthisoption,knownasa“partialclaim”, priortotheendoftheforbearanceperiod. Loansareeligibleforthepartialclaimifi)themortgagewas currentor<30daysdelinquentasofMarch12020,ii)thepropertyisowneroccupied,andiii)the borrowerindicatestheyhavetheabilitytoresumemakingon-timepayments. Forloansnoteligiblefora partialclaim,theFHAinstructsservicerstoevaluatetheborrowerforloanmitigationoptionsinvolving loanmodification. SeeDepartmentofHousingandUrbanDevelopment(2020a)formoredetails. 6

Figure1: ForbearanceRateandDelinquencyRateOver2020 FractionofMortgagesinForbearance FractionofMortgages60+DaysPastDue Data sources: Author calculations from Federal Reserve Bank of New York (FRBNY) Consumer Credit Panel/Equifaxdata(toppanel)andBlackKnightMcDash(bottompanel). 7

overtheremainderof2020(to5.2percentasofDecember).7 Delinquency,asmeasuredby 60+ days past due (bottom panel) follows a similar shape.8 At an individual level however, not all delinquent borrowers entered forbearance, and conversely some borrowers inforbearancecontinuedmakingsomeoralloftheirscheduledmortgagepayments.9 Forbearance and delinquency is much higher in the FHA segment than the overall market. This reflects the relatively low- and middle-income FHA borrower population andthehighshareoffirst-timehomebuyers. VAmortgageshaveaforbearanceanddelinquency rate path similar to the market as a whole, while forbearance and nonpayment is relativelylowforthetypicallyprimemortgagessecuritizedbythegovernmentsponsored enterprises(GSEs)FannieMaeandFreddieMac.10 2.2 Forbearance implementation and the role of servicers The mortgage servicers that implemented the CARES Act forbearance programs on the ground vary widely in terms of size, regulation, funding, profitability and other characteristics. One might assume that servicers play a limited and passive role, given the essentially universal eligiblity for forbearance among agency borrowers and lack of documentation requirements. In practice however, borrowers and regulators report a wide range of servicer-related issues, including misinformation, processing errors, and communication difficulties, suggesting that servicer practices may indeed vary substantially. Because borrowers cannot choose their servicer, variations in servicer practice may sig- 7Otherdatasourcespaintasimilarpicture. SurveydatafromtheMortgageBanker’sAssociationindicates apeakforbearancerateof8.55%inJune2020(MortgageBankersAssociation,2020),whileBlackKnight estimatesapeakforbearancerateof8.8%,alsoinJune(BlackKnight,2020). 8Weusethetermdelinquencyasshorthandformortgagesthatarepastdue. Formallythough,aborrower whomissespaymentswhileinforbearanceisnotdelinquentontheirpaymentobligations. 9Wealsoincludeaplotofdelinquencymeasuredinsteadby30+dayspastdueintheInternetAppendix. 10Statisticsforthemarketasawholeincludethethreesegmentsshownseparately(whichtogethercomprise theagencymortgagemarket),aswellasmortgagesheldinportfoliobybanksandotherinvestorsand loanssecuritizedthroughtheprivate-labelmarket. FannieMaeandFreddieMacarecombinedinthe figurebecausetheirmortgageportfolioshavesimilarcharacteristicsandloanperformance. 8

nificantlyaffectborroweroutcomes. Consumer Financial Protection Bureau (2021a) presents systematic qualitative evidence regarding issues with servicers based on the observations of Consumer FinancialProtectionBureau(CFPB)supervisors. Thereporthighlightsthelogisticalchallenges facedbyservicers,statingthat“Manyservicersreportedoperationalconstraints,resourceburdens, and service interruptions. Many servicers also moved employees from other duties to respond to forbearance requests.” It also documents a range of deficient practices by servicers including: i.) Providingincompleteorinaccurateinformation,suchastellingconsumersthatonly delinquent borrowers qualify for forbearance, that a fee must be paid to obtain forbearance,orthatalump-sumrepaymentisrequiredattheendofforbearance; ii.) Incorrectlysendingcollectionordefaultnotices,assessingfees,orinitiatingforeclosuresforborrowersinforbearance; iii.) Changingborrowers’preauthorizedfundstransferswithouttheirconsent,orfailing toimplementtheborrowers’instructionstofreezepayments; iv.) Failuretoprocessforbearancerequestsinatimelymanner; v.) Enrollingborrowersinautomaticorunwantedforbearance; vi.) Failuretoenrollborrowersinanappropriatepost-forbearanceplan. Consumer Financial Protection Bureau (2021b) tabulates data from the CFPB’s complaintsdatabase,findingthatforbearancecomplaintsrosefromfewerthan100permonth in January and February of 2020 to a peak of over 500 in April, and a level between 300- 500 per month over the rest of 2020 and early 2021. Complaints most commonly relate to problems contacting or communicating with servicers, confusing or incomplete information about post-forbearance options, misleading or incorrect information about loan 9

balance or performance reported on the borrower’s monthly statements, and delays and denialsinputtingtheborrowerinapost-forbearancerepaymentplan.11 Media reports highlight many of the same issues. For instance Wall Street Journal (2020)describeshowthewaveofforbearancerequestsearlyinthepandemicoverwhelmed manyservicers’capacity,leadingtoextremelylongtelephoneholdtimes,non-operational servicerwebsites,andmisinformationtoborrowers. Not all borrowers experienced problems, however, and many servicers took significant steps to streamline the forbearance process, such as providing a prominent button or link on their website to a simple online application. We have also heard numerous anecdotesfrompractitionersaboutservicersthathaveengagedproactivelywithborrowers to explain forbearance and make them aware of their options (e.g., one large servicer contacts delinquent borrowers not in forbearance at a daily frequency). Taken together, thequalitativeevidencesuggeststherehasbeenawiderangeofservicerpractices,which inturncouldleadtosignificantvariationinborroweroutcomes. 2.3 The role of servicer characteristics We now discuss factors relating to financial constraints, regulation and organizational form that may lead to systematic variation in forbearance practices and outcomes across servicers. Westudytheimportanceofthesedifferentfactorsempiricallyinsection5. 1. Liquidity constraints. Mortgage servicers are required to temporarily advance scheduled payments on delinquent mortgages to investors and other parties, including 11Togiveasenseoftheissues,thefollowingarethreecomplaintsavailableinthepublicCFPBdatabase: (1) “Itriedtoreachoutto<XXX>torequestaforbearance... Unfortunately,Iwashungupontwotimes. Ispent almost3hoursonhold.”;(2)“Myinitial6monthforbearancehasbeenapproved,butI’vebeenunabletomake contactwiththeservicertoextendtheforbearance. I’vesentemails,leftvoicemessagesandtriedonlinetoextendthe forbearance. Theydonotrespond. I’mscaredandIneedhelp.”;(3)“Ihavebeentryingforoveramonthtoapplyfor a6-monthmortgageforbearanceplan(asallowedundertheFederalCaresAct)with<XXX>. Ifyougototheir websitetoapply,itdoesn’tmatterifyouareonamobiledeviceORhardwiredlaptopORdesktopcomputer,itwill notactuallyletyouapplyforaforbearance. Whenyousubmit,itsays”CRITICALERROR”.” 10

principal, interest, taxes and insurance. Servicers facing more binding liquidity constraints may therefore wish to discourage borrowers from entering forbearance, if forbearancetheninducesborrowerstopausetheirpayments.12 Servicerliquidityriskalsovariesacrossmortgages,inpartbecauserulesaboutservicingadvancesdependontheloanprogramandtheservicingagreementwiththeinvestor. FHAmortgagestypicallypresenthigherrisk,becauseborrowersaremuchmorelikelyto become delinquent, because the servicer is generally required to forward payments until loan termination or modification, and because FHA servicers face significant delays before being reimbursed for payment shortfalls (Kim et al., 2018).13 Servicing advances for GSEmortgagesaretypicallylimitedtofourmonthsofmissedpayments. 2. Regulation and legal risk. Mortgage servicers face stricter regulation and supervisoryoversightaswellashigherlegalriskinthewakeoftheGreatRecession.14 Thislegal andregulatoryriskislikelytobeparticularlysalientforlargebanks,whofacethetoughest regulatory scrutiny and who were subject to the largest post-crisis legal settlements (Buchak et al., 2018). It therefore seems plausible that legal and regulatory risk could inducetheseservicerstoadoptmore“borrower-friendly”practices,bymakingforbearance easiertoobtain.15 3. Capitalization and risk-shifting. Decisions about servicing practices involve a trade-off between risk and reward. Actions such as enabling easy access to forbearance, 12Toemphasizethispoint,itisnonpaymentratherthanforbearancepersethatcreatesaliquiditydrainonthe servicer’sresources. Althoughthetwodonotnecessarilygohand-in-hand(e.g.,asignificantnumberof borrowersinforbearancecontinuedtomaketheirmortgagepayments),welaterpresentevidencethat makingforbearanceeasiertoobtaindoesinfactcausallyleadtohighernonpayment,almostone-to-one. 13Mitigatingtheserisks,theFHAdeterminedthatCARESloansthatre-performafterexitingforbearance canbemadecurrentbyissuingapartialclaim,aswehavediscussed,reimbursingtheservicerfor principalandinterestadvancesduringforbearance. GinnieMaealsocreatedatemporaryliquidityfacility forservicers,albeitwithahighfundingrate. 14Additionalpost-crisisregulationincludesnationalmortgageservicingstandards,higherbankcapital requirementsonservicingrights,andsupervisoryoversightfromanewregulatoryagency,theConsumer FinancialProtectionBureau(CFPB).Legalriskisalsomuchmoresalient,sincebankswereforcedtopay outverylargepost-crisislegalsettlementsduetodeficientservicingpractices. 15Fusteretal.(2021)findthattighterregulatoryoversightleadstomoreconsumer-friendlyservicing practices,usingacutoffruleinwhichbanksaresubjecttoCFPBsupervisionandenforcement. 11

or investing heavily in servicing technology or staff training, are likely to be costly in the short run but may reduce the likelihood of future regulatory or legal action and also perhaps improve customer satisfaction and retention. Undercapitalized servicers may thus have weaker incentives to provide high-quality servicing, in line with the classic risk-shiftinghypothesisofJensenandMeckling(1976). 4. Size and scale. Organizational form may also be a key driver of servicer practices. For example, large servicers may enjoy scale economies (e.g., due to fixed costs) that allow them to set up more sophisticated forbearance management systems. Or conversely, small, nimble servicers may be able to adjust their practices more quickly than large organizationswithseverallayersofmanagement.16 5. Technology and operational effectiveness. Servicers vary in terms of their prior investments in technology and human capital, such as the quality of information systemsandtheservicer’swebportal,theextenttowhichservicingtasksareautomated,the quality of risk measurement systems to identify defects and fraud, and the qualifications and training of servicing staff.17 These prior investments may have improved servicers’ abilitytoquicklyandeffectivelyimplementlarge-scaleforbearance. 3 Data and summary statistics Ouranalysiscombinesloan-leveldataonmortgagecharacteristicsandperformance,FHA forbearance records, and regulatory data on the characteristics and financial condition of mortgage servicers. For the each of the two main stages of our analysis, we compile a different dataset, which we describe briefly here. Additional details on each of the data 16Inarelatedcontext,paperssuchasBergeretal.(2005)findsystematicdifferencesinlendingbehavior betweensmallandlargebanks,whichtheyinterpretasbeingduetodifferencesinorganizationalform. 17Fusteretal.(2019)findthatFHAmortgagesoriginatedbytechnology-basedlendershavelowerdefault rates,evencontrollingfordetailedloancharacteristics. Thismaybeduetodifferencesinunderwriting practices,butcouldalsoinpartreflectservicingbehavior. 12

sourcescanbefoundinAppendixSectionA. For the first stage of the analysis (Section 5), we merge loan-level performance data withservicercharacteristics. Wedrawtheidentityofeachloan’sserviceraswellasoriginationcharacteristicsandongoingpaymentperformanceforloansecuritizedintoGinnie Mae MBS from eMBS. We append information on each loan’s forbearance status and forbearancetermsfromGinnieMae’sforbearanceregister. Forindependentmortgagebanks (“nonbanks”) we draw servicer characteristics from the mortgage call report (MCR) collected by the Conference of State Bank Supervisors. For banks, we draw servicer characteristicsfromY-9Candcallreports. Additionally,becauseeMBSdataarecomprehensive, we calculate some servicer characteristics for both banks and nonbanks (such as growth through servicing transfers and measures of servicing volume) using the eMBS data. To evaluate the relationship between borrowers’ experience and servicing policy (Section 5.1), we merge in data from the CFPB complaints database at the servicer level. For each complaint, these data allow us to identify the categorical reason for the complaint (e.g., forbearance), the type of the loan the borrower has (FHA, VA, GSE, etc.), and the servicer’sname. Wesummarizethesecomplaintsattheservicerlevel. For the second stage of the analysis (Section 6), we match loan-level eMBS data with loan-level data from Black Knight McDash and the Equifax Credit Risk Insight Servicing and McDash (CRISM) dataset. From eMBS, we draw the loan’s forbearance status and a de-identified servicer id. Due to data use restrictions, we cannot merge servicer characteristics,includingnonbankstatus,intotheCRISMdata. FromtheCRISMdata,wedraw payment behavior, updated credit scores, geographic data, and additional information about the borrower’s balance sheet. For more detail on the mechanics of this match, see AppendixSectionA.1. 13

3.1 Summary statistics Table 1 presents summary statistics for the eMBS loan-level sample, which reflects the population of FHA and VA loans in Ginnie Mae securities as of February 2020. The dataset includes 10.3 million mortgages, of which about 70% are FHA loans. FHA loans have higher loan-to-value (LTV) ratios, higher debt-to-income (DTI) and lower average creditscores,reflectingthedisproportionatelylow-income,high-riskFHAborrowerpopulation. Table1: SummaryStatistics (1) (2) (3) FHA VA Total A.Loancharacteristics: CurrentUPB($000) 151,499.61 209,059.04 168,647.51 OrigLTV(%) 92.93 94.63 93.40 OrigDTI(%) 41.11 38.48 40.26 Origcreditscore 682.00 714.81 692.65 Loanage(year) 5.39 3.95 4.93 30+daysdelinquentinFeb2020 0.06 0.03 0.05 60+daysdelinquentinFeb2020 0.02 0.01 0.02 B.Forbearance&delinquency: March-November2020: Ever30+daysdelinquent 0.22 0.11 0.19 Ever60+daysdelinquent 0.15 0.08 0.13 Everpaidoff 0.17 0.30 0.21 Everinforbearance 0.17 0.08 0.14 C.Conditionalforbearance&delinquencyrates: Forbearance|delinquency(forloanscurrentinFeb2020): Everinforbearanceamongloansever30+daysDQ 0.75 0.70 0.74 Everinforbearanceamongloansever60+daysDQ 0.92 0.88 0.91 Delinquency|forbearance(forloanscurrentinFeb2020): Everin30+daysinDQamongborrowerseverinforbearance 0.85 0.84 0.85 Everin60+daysinDQamongborrowerseverinforbearance 0.71 0.72 0.72 N.Obs. 7,044,172 3,270,949 10,315,121 About 5% of loans were at least 30 days delinquent just before the onset of the pandemic. Nonpayment then increased sharply, with 19% of loans being 30 days or more 14

delinquent at some point between March and November 2020 (22% of FHA loans and 11% of VA loans). 17% of FHA loans entered forbearance at some point between March and November, compared to 8% of VA loans. 21% of loans were paid off between March andNovember,primarilyreflectingrefinancingduetolowmortgageinterestrates. Panel C of table 1 reports conditional forbearance and delinquency statistics for loans that were current as of February 2020. The table shows that 26% of FHA and VA loans that became delinquent during the pandemic did not enter into a forbearance plan. This fractionissignificantlysmaller–9%–forloansthatexperiencedseriousdelinquency(60+ days past due), but still well above zero. These facts are in some sense surprising given that any FHA or VA borrower that became distressed due to the effects of the pandemic was eligible for forbearance, and given that forbearance effectively provides a subsidy to the borrower. Conversely, 15% of borrowers remained current on their payments despite entering into forbearance. Most borrowers in forbearance skipped multiple payments however,with72%becomingatleast60dayspastdue. 4 Servicer-level variation in forbearance outcomes Wemeasurecross-servicervariationinforbearanceoutcomesbyestimatingthefollowing cross-sectionallinearprobabilitymodelusingeMBSloan-leveldata: forbearance = X β+ξ +(cid:101) . (1) i i s i The dependent variable is an indicator for whether mortgage i entered forbearance from March-November 2020, X is a set of loan controls (e.g., LTV and credit score bins) i toaccountforforbearancedemand,and ξ isavectorofservicerfixedeffects.18 s Ourbaselinemodelestimatesequation1usingthepopulationofGinnieMaeborrow- 18CoefficientestimatesonloanandborrowercontrolsarereportedintableA.1oftheInternetAppendix. 15

ersthatwerecurrentpriortotheonsetofthepandemic(January2020)butmissedatleast one payment from March to November. This set of borrowers would unambiguously benefit from forbearance, but as we have discussed, around one-third of them became delinquentwithoutsuccessfullyenteringintoaforbearanceplan. Figure 2 plots the distribution of the servicer fixed effects (ξ(cid:98)s ), showing very wide variation in forbearance outcomes across servicers for observably similar mortgages. 19 For the figure we normalize the fixed effects to show the probability that a past-due loan withsampleaveragecharacteristicsfailstoenterintoforbearance. Thelikelihoodthatthe borrower“fallsthroughthecracks”rangesfromunder10%toalmost60%. Thisvariation is not simply due to disparate outcomes among very small servicers. Weighting by loan count, the “no forbearance” probability is 38% for a servicer at the 90th percentile of the distributioncomparedtoonly12%foraserviceratthe10thpercentile. The bottom panel of figure 2 presents the same histogram conditioning on more serious nonpayment (60+ days past due). The share of “missing” forbearances is significantly smaller for this group, but in proportionate terms the cross-servicer variation is even more stark — the likelihood of not receiving forbearance is six times higher for a “low-forbearance”serviceratthe90thpercentileofthedistributioncomparedtoa“highforbearance”serviceratthe10thpercentile(18%comparedto3%). 4.1 Alternative estimates of servicer effects Wealsoestimateanalternativesetofservicerfixedeffectsbyestimatingequation1using the eMBS-CRISM matched sample. This allows us to control for a finer set of borrower and loan controls using information from borrower credit reports, including bins of borrower age, an updated credit score and information on nonmortgage debt balances.20. 19Indeed,theseestimatedservicerfixedeffectsarehighlystatisticallysignificantjointlywithF-statisticsof 435.17. 20CoefficientestimatesarereportedinsectionCoftheInternetAppendix 16

Figure2: P(noforbearance|COVIDdelinquency). Cross-servicervariationinprobabilitythatapast-due loandoesnotenteraforbearanceplan. BasedonservicerfixedeffectsestimatedusingeMBSdata. Barsare unweightedcountsofservicers. Dashedverticallinesshowweightedpercentiles(weightedbythenumber ofloansthatbecamepastduebetweenMarchandNovemberof2020.) (a)Borrowers30+dayspastdue 60 10th percentile Median 90th percentile servicer servicer servicer (weighted) (weighted)(weighted) 50 40 30 20 10 0 srecivreS fo rebmuN 0 .1 .2 .3 .4 .5 .6 Missing Forbearance (b)Borrowers60+dayspastdue 70 10th percentileMedian 90th percentile servicer servicer servicer (weighted) (weighted) (weighted) 60 50 40 30 20 10 0 srecivreS fo rebmuN 0 .1 .2 .3 .4 .5 Missing Forbearance 17

Thisapproachalsoproducesasimilarlywidedispersionofservicereffectsestimates(see figureA.2oftheInternetAppendix). We also use the eMBS-CRISM model to examine how sensitive the fixed effects are to the set of controls used, comparing specifications based on the same sample but with i) nocontrols,ii)thecontrolsavailableineMBSandiii)thefullsetofeMBS-CRISMcontrols. We find that the three resulting sets of servicer fixed effects are highly positively correlated (see figure A.3 of the Internet Appendix). In particular the credit report controls availableinCRISMmakeverylittledifferencetotheservicerfixedeffects. Within the eMBS sample, we also estimate the servicer effects three other ways aside from the two presented in figure 2: i) including all mortgages in the sample, rather than just the loans that became past due during the pandemic; ii) restricting the sample to borrowers that became delinquent early in the pandemic (February or March), prior to thepassageoftheCARESAct;andiii)includinglenderfixedeffects,sothatservicerfixed effects are identified only from mortgages where there was a transfer of servicing rights. These alternative fixed effects are strongly positively correlated with our main estimates inthetoppaneloffigure2,asshownintheInternetAppendix(figureA.4). 4.2 Servicer behavior or omitted borrower characteristics? Our interpretation is that these striking differences in forbearance outcomes are driven by variation inservicer behavior. But analternative explanationis thatthey reflectunobserved differences in forbearance demand. For instance borrowers at “high-forbearance” servicersmaybemoreliquidityconstrainedandthereforebenefitmorefromanextended payment holiday, or may be more financially literate. Our estimated fixed effects conditiononarichsetofborrowerandloancontrols,particularlyfortheeMBS-CRISMsample, 18

butofcoursedonotcontrolforallfactorsthatmayaffectforbearancedemand.21 However, three pieces of evidence suggest the servicer fixed effects are not mainly drivenbyunobservedborrowerheterogeneity: 1. Mortgages managed by high- vs low- forbearance servicers have similar ex-ante characteristics, measured in either the eMBS and eMBS-CRISM samples (see Internet Appendix tables A.3-A.5). Non-mortgage loan balances are also similar (e.g., auto, credit card and student loan balances are all within 10% comparing the two groups), and the two groups of loans also experienced similar macroeconomic conditions during the pandemic (e.g., the 12-month change in the county unemploymentratediffersbyonly0.2%). Perhapsthemaindifferenceisthatloansmanagedbylow-forbearanceservicersare somewhatyounger(4.5vs6.0yearsintheeMBS-CRISMsample). Withinagegroups howevertheappendix showsthatmortgageslookvery similaronobservables,and ourregressionsalwaysincludeafullsetofloanagedummies. 2. Thereisalmostnodifferenceinmortgageloanperformancebetweenhighandlowforbearanceservicersinthemonthsleadinguptothepandemic. Wemeasurethisby estimating a loan-level delinquency model where the dependent variable is equal to 1 if a loan current at t-1 becomes delinquent in month t.22 Differences in conditional delinquency transition probabilities for mortgages managed by high-vslow forbearance servicers are economically small, not consistently signed, and of- 21Weemphasizethatservicerforbearancepoliciespersewerenotlikelyanimportantdimensionofborrower mortgagechoicepriortothepandemic,giventhestableeconomy,risinghomeprices,theinfrequencyof forbearance,andthefactthatborrowerscannotdirectlychoosetheirservicer. Evenso,therecouldstillbe nonrandomassignmentofborrowerstoservicersinawaythatiscorrelatedwithborrowers’desiretotake advantageofforbearanceduringthepandemic. 22Measuringtransitionsintodelinquencyispreferabletomeasuringthestockofdelinquentloans,fortwo reasons: i)servicerqualitycanaffectthelengthoftimealoanremainsdelinquent,e.g.,better-quality servicersmaymakeiteasierfortheirborrowerstocureorobtainaloanmodification;ii)servicershavethe optiontopurchaseseriouslydelinquentloansoutofGinnieMaepools–suchloanswouldnolonger appearintheeMBSdataaftertheyarerepurchased. Thiscouldcreateaselectioneffectsincee.g.,since banksaremorelikelytorepurchaseloansthannonbanks. 19

ten not statistically significant – see figure A.6 and table A.10 of the Internet Appendix.23 The same is true for credit card and auto delinquencies in the eMBS- CRISM matched sample. In contrast, during the pandemic itself, borrowers assigned to high-forbearance servicers become much more likely to stop paying their mortgages,asshowninfigureA.6andasdiscussedindetailinsection6. This argues against an “omitted risk” explanation of the results, which would predicthigh-forbearanceservicersexperiencehighernon-paymentratesnotjustduring the pandemic, but also prior to it. Conversely it also speaks against the hypothesis thathigh-forbearance-servicerborrowersaremorefinanciallyliterate,whichwould be expected to result in a lower pre-COVID delinquency rate in line with Gerardi etal.(2013)andAgarwaletal.(2017b). 3. Estimated servicer fixed effects are generally insensitive to the set of borrower and loan controls used, as already discussed in section 4.1. It seems unlikely that servicerfixedeffectsaredrivenbyunobservableloanandborrowercharacteristicsbut essentiallyuncorrelatedwithobservablecharacteristics. 5 Servicer characteristics and forbearance outcomes Nowwestudyhowaservicer’s“forbearancepropensity,”asmeasuredbyitsfixedeffect, varies with servicer characteristics such as size, liquidity and organizational form. The goalofthisanalysisistounderstandwhicheconomicfactors(asdiscussedinsection2.3) aremostimportantinshapingservicerbehavior. We estimate a simple cross-sectional regression of servicer effects on characteristics 23Servicerforbearancepoliciesareunlikelytohavebeenanimportantdimensionofborrowermortgage choicepriortothepandemic,giventhestableeconomyandlowmortgagedefaultrate,theinfrequentas wellasthethelowpre-pandemicrateofforbearance,andthefactthatultimatelyborrowershavelittle choiceintheirmortgageservicer. 20

drawn from mortgage call reports (for nonbank mortgage companies), Y-9C and bank call reports (for banking organizations or nonbanks controlled by a bank), and data on totaloriginationsandservicingvolumesaggregatedfromeMBSaccount-leveldata.24 Our analysisfocusesonbanks,creditunionsandnonbankmortgagecompanies,andexcludes governmentandgovernment-sponsoredenterprisessuchasstatehousingauthoritiesand FederalHomeLoanBanks. Estimates are reported in table 2 and reveal several patterns. First, large servicers are significantly more likely to enroll their borrowers in a forbearance plan, whether size is measured by the log of servicing assets (measured using eMBS) or balance sheet size (measured using regulatory reports). As we have discussed, large servicers may enjoy scale economies because of fixed costs in setting up efficient forbearance processes (e.g., a well-designed online application form). Alternatively, large servicers may have more resourcestobettertrainservicingstaff,ormayalsobehaveinamore“borrower-friendly” waybecausetheyaremorelikelytobetargetedbyfinancialregulators. Second, organizational form matters. Nonbank mortgage companies are about 9 percentagepointslesslikelytoofferforbearancetoapast-dueborrower,whilecreditunions were about 13 percentage points more likely. The lower rate of forbearance for nonbanks is consistent with a liquidity-based mechanism. Nonbank servicers rely primarily on short-term wholesale funding and do not have access to government backstops such as the Federal Reserve discount window and Federal Home Loan Bank system advances, and at the start of the pandemic when most forbearance plans began, there were significant concerns about a nonbank liquidity crunch. By discouraging forbearance, nonbanks could induce borrowers to keep making their mortgage payments, thereby mitigating their own liquidity outflows due to contractual obligations to forward mortgage pay- 24Institutionswerematchedbynameacrossthesedifferentdatasources. Informationonfinancialstructure fromtheNationalInformationCenterandothersourceswereusedtocross-validatetheaccuracyofthe match. 21

mentsonnonperformingloans. Third, and also consistent with a liquidity-constraints channel, the level of cash balances is significantly positively correlated with servicer forbearance propensity, but only for nonbanks. Precautionary cash holdings are not important for depository institutions, which have access to backstop sources of liquidity for mortgages (e.g., through the Federal Home Loan Bank system and the discount window), and also experienced large inflows of liquidity at the start of the pandemic. But for nonbanks, not only is the overall rate of forbearance lower, but forbearance is particularly depressed for servicers with a lowlevelofexanteprecautionarycashbalances. Table 2: Regression of conditional forbearance rates on servicer characteristics: Weighted least squares, weighted by number of borrowers current in January 2020 but missed at least one payment between March and November. Column (1) is based on all servicers. Columns (2) through (4) reflect nonbank servicers only. Columns (5) through (7)reflectbankservicersonly. All Nonbankmtgcompanies Banks (1) (2) (3) (4) (5) (6) (7) Servicercharacteristics log(Servicingassets) 0.037∗∗∗ 0.031∗∗∗ 0.025∗∗∗ 0.043∗∗∗ 0.043∗∗∗ (0.007) (0.008) (0.006) (0.010) (0.010) log(Assets) 0.018∗∗∗ 0.025∗ (0.004) (0.015) Cash/assets 0.955∗∗∗ 1.083∗∗∗ -0.664 -0.942 (0.174) (0.177) (0.537) (0.699) Securities/assets 0.144 0.246∗∗∗ 0.320 0.553 (0.100) (0.085) (0.364) (0.345) Capital/assets 0.011 0.053 1.292 1.068 (0.103) (0.109) (0.794) (0.907) Servicinggrowth -0.008 0.011 -0.011 -0.035 -0.030 0.000 -0.018 (0.047) (0.055) (0.045) (0.042) (0.086) (0.089) (0.090) Servicertype Nonbankmortgagecompany -0.087∗∗∗ (0.025) Creditunion 0.131∗∗∗ (0.029) N.Obs. 152 98 98 98 45 45 45 22

5.1 Servicing quality: evidence from CFPB complaints Given that past-due Ginnie Mae borrowers would have universally benefited from entering into forbearance, at least to some degree, servicer practices that limit forbearance uptakealsoreduceborrowerwelfare. Toinvestigatefurther,westudywhetherborrowers are less satisfied with “low-forbearance” servicers, based on the frequency of mortgage forbearance-relatedcomplaintssubmittedtotheCFPBcomplaintsrepository.25 Results are presented in Table 3. Our estimates indicate that the frequency of complaints (scaled by the number of serviced mortgages) is significantly higher for lowforbearance servicers. This is direct evidence of poorer servicing quality for these firms. When we replace the servicer forbearance propensity with servicer characteristics, we findinparticularthatliquiditymatters;servicerswithlowerlevelsofprecautionarycash balanceswerethesubjectofahigherrateofcomplaints. 5.2 Summary To sum up the results of this section, we find that many nonperforming FHA and VA mortgagesentitledtoforbearanceundertheCARESActdidnotinfactenteraforbearance plan—furthermore,forbearanceoutcomesvariedsignificantlyacrossmortgageservicers forobservablyequivalentloans. Severalpiecesofevidenceindicatethattheseservicereffects reflect servicer behavior rather than unobserved loan and borrower heterogeneity. Small servicers and nonbanks were less likely to provide forbearance, particularly for nonbank servicers with low liquidity buffers at the start of the pandemic. Our results highlighthowliquidityconstraintscanleadtoadeteriorationofservicingquality,consistent with earlier evidence on foreclosures and modifications from the period of the Great 25Wemeasureforbearance-relatedcomplaintsusingasetofkeywordssimilartotheCFPB,andrestricting thesampletocomplaintsrelatedtomortgagefinancingwherethemortgageisagovernmentloan,tobe consistentwithourGinnieMaesample. 23

Table 3: CFPB complaints: Forbearance-related complaints are normalized by the number of Ginnie Mae or total agency mortgages serviced (complaints per thousand loans). Weighted least squares, weighted by number of loans serviced as of January 2020. Forbearance-related complaints are normalized by the number of Ginnie Mae or total agencymortgagesserviced(complaintsperthousandloans). Alllenders Nonbanksonly (1) (2) (3) (4) (5) (6) Servicerforbearancepropensity -0.237∗∗∗ -0.258∗∗∗ -0.706∗∗ (0.083) (0.084) (0.288) Servicercharacteristics log(Servicingassets) -0.017∗∗ -0.005 -0.042 (0.007) (0.008) (0.043) Cash/assets -0.431∗ -1.522∗∗∗ (0.236) (0.538) Securities/assets -0.377∗ -0.278 (0.191) (0.318) Capital/assets -0.021 0.182 (0.140) (0.380) Servicinggrowth 0.046 0.081 0.408∗∗ (0.049) (0.054) (0.204) Frac.govt.loansthatareFHA 0.087∗∗∗ 0.074∗∗ 0.096 (0.033) (0.036) (0.077) Frac.allloansthatareFHA -0.596∗∗ -0.533∗ (0.289) (0.288) Servicertype Nonbankmortgagecompany 0.001 0.010 -0.043 (0.020) (0.020) (0.039) Creditunion 0.071∗∗ -0.002 (0.028) (0.023) N.Obs. 129 129 129 125 92 92 24

Recession(Aiello,2021). 6 Does forbearance cause nonpayment? In this section, we use cross-servicer variation in forbearance practices (measured using the fixed effects methodology from the prior section) to estimate the causal effect of forbearanceavailabilityontheborrower’spropensitytopausemakingmortgagepayments. In the following section we also apply the same methodology to examine nonmortgage outcomessuchastotalnonmortgagedebt. For this portion of the analysis, we rely primarily on the CRISM-eMBS merge describedinSection3.26 UsagerestrictionsontheCRISMdatasetpreventusfromretaining servicerinformationinthemergedeMBS-CRISMdataset,butthemergedoesallowusto retain anonymous servicer identifiers. We use these identifiers to estimate servicer-level fixed effects using the same methodology as in the prior section. We then use the fixed effectsasasourceofplausiblyexogenousvariationinforbearanceavailabilitytotracethe effectsofforbearanceonotherborroweroutcomes. The key identification assumption underlying this approach is that servicer forbearance fixed effects are orthogonal to unobserved borrower characteristics which would affect outcomes during the pandemic (conditional on mortgage characteristics measured inCRISM).ThisisthesameidentificationassumptionrequiredforouranalysisinSection 4.2,wherewepresentevidenceofitsvalidity. 26TheeMBSdataweusedforSection4presentsomedrawbacksforthispartoftheanalysis. Most importantly,welosetheabilitytotracksomeloansbecausesomeservicersbeganpurchasingloansin forbearanceoutofGinnieMaepoolsseveralmonthsaftertheprogramwentintoplace,andthereforeexit theeMBSdatasetatthesametime. Additionally,theeMBSdataallowustoobservetheborrower’s locationonlyatthestatelevel,apotentiallysignificantdrawbackgiventhatservicersmayhavedifferent geographicexposures,andgiventhatthegeographyofthevirusdriveseconomicstress. 25

6.1 Regression specification We use a difference-in-difference approach to compare outcomes and behavior of borrowersatservicerswithmore-andless-generousforbearancepractices. Wedefine“highforbearance” servicers as those with above-median servicer fixed effects (estimated as described in Section 4, using the merged eMBS-CRISM data). We use the 6 month period preceding the March 2020 passage of the CARES Act to establish the absence of different pre-existingtrendsbetweenhigh-andlow-forbearanceservicers. Weattributedifferences in borrower outcomes and behavior after March 2020 to differences in the accessibility of forbearance. Weestimatethefollowingregression: 8 Y it = ∑ β τ S i H ×1 t=τ +Z it γ+α s +α ztτ +ε it (2) τ=−6,τ(cid:54)=0 whereY isaborroweroutcomesuchasnonpayment;SH isanindicatorvariableequalto it i 1forhigh-forbearanceservicers; Z isavectorofloanandborrowercharacteristicswhich it may affect mortgage nonpayment, including mortgage characteristics at origination, the borrower’supdatedcreditscore(measuredbytheEquifaxRiskScore)asofJanuary2020, updated principal balance, loan age, borrower age, loan type (FHA vs. VA), and several household balance sheet characteristics, measured in January 2020, including the household’s other mortgage and non-mortgage debt, and delinquency on other mortgage and non-mortgage debt; α is a vector of servicer fixed effects, which account for persistent s differences in borrower outcomes across servicers; and α is a vector of zipcode x month zt x origination month FE to account for the time-varying geographic effects of the pandemicseparatelyforloansoriginatedindifferenttimes. Weclusterstandarderrorsatthe servicerlevel. NotethatourzipxmonthxoriginationmonthFEabsorbanygeneralequilibrium effects of the program. First stage regression results can be found in Appendix 26

SectionC.2 6.2 Results First, we confirm that the path of forbearance rates is actually higher among servicers we have categorized as high-forbearance-availability servicers. Figure 3(a) plots the estimates of β from Equation 2 using a forbearance dummy as the outcome variable. The τ coefficients can be interpreted as the difference in the probability that a borrower is in a forbearanceplanatahigh-forbearanceservicervs. atlow-forbearanceservicerinagiven month,allelseequal. Figure 3(a) confirms that forbearance rates are higher at high-forbearance servicers throughout the pandemic. At the peak in April, the share of borrowers in forbearance at high-forbearanceservicerswasabout5percentagepoints(about30%)higherthanatlowavailabilityservicers. ThedifferencebeginstodiminishfromMayonwards,evenasoverallforbearanceratescontinuetorise,perhapsreflectingthatlow-forbearance-availability servicers partially “catch up” in their policies and practices. However the difference in forbearanceratesremainshighthroughoutthepandemic. Next,weexaminewhetherforbearanceavailabilitycausesmortgagenonpayment. Given that forbearance significantly reduces the cost of missing mortgage payments, it seems reasonable to expect that the nonpayment rates will rise disproportionately at servicers that make it easier for borrowers to enter forbearance. On the other hand, it is possible that high-forbearance servicers mainly encourage higher entry into forbearance among borrowers who continue to make mortgage payments or among borrowers who would not otherwise have made their payments. If so, nonpayment rates would be unaffected byforbearanceavailability. Figure 3(b) reports the monthly difference in the probability a borrower is past-due (i.e.,hasmissedatleastonepayment)athigh-forbearanceservicersrelativetolow-forbearance 27

Figure 3: Forbearance and Nonpayment. Estimates and their 95% confidence intervalsof the effect of assignmenttoa“high-forbearance”serviceronthelikelihoodofforbearanceandmissedpayments.Standard errorsareclusteredattheservicerlevel. .08 .06 .04 .02 0 -.02 2019m10 2020m1 2020m4 2020m7 2020m10 In forbearance In forb. & current (a)Inforbearance .06 .04 .02 0 -.02 2019m10 2020m1 2020m4 2020m7 2020m10 Delinquent DQ & not in forb (b)Pastdue 28

servicers. We find that the probability that a borrower is past-due is significantly higher forborrowersathigh-forbearanceservicers,byasmuchas5percentagepointsatthepeak in May 2020. Moreover, the estimates for the probability that a borrower is past-due are similartotheestimatesfortheforbearanceprobabilityinFigure3(a). Thisresultindicates that effectively all of the additional forbearance at high-forbearance servicers drives borrower nonpayment. In other words, marginal forbearance recipients at high-forbearance servicerswouldnothavemissedpaymentshadforbearancebeenmoredifficulttoaccess. Figure 3(a) confirms that this sharp increase in nonpayment is driven entirely by borrowers who are in forbearance plans. Conversely, there is also no difference in delinquency rates outside of forbearance among high-vs-low availability servicers (shown in thedarkbluelineinFigure3(b)). Additional results reported in Section I show that prepayment rates across high- and low-forbearance servicers were identical before and after the forbearance program went into place. This suggests that borrowers assigned to high-forbearance servicers were not diverted from refinancing into forbearance - an outcome that would complicate the welfareanalysisoftheprogram. Instead,onaverage,borrowersassignedtolow-forbearance servicerswhowouldhavemissedpaymentsathigh-forbearanceservicerscontinuedmakingpaymentsanddidnotrefinance. These results indicate that forbearance availability directly affects borrower decisions about whether to defer mortgage payments during the pandemic, at least for the substantial set of marginal borrowers whose forbearance outcomes are affected by servicer practices. In other words, servicer policies significantly affect household cash flows duringthepandemic. 29

7 Non-mortgage effects Our results so far show that assignment to a “high-forbearance-availability” servicer induces borrowers to obtain forbearance and also to defer their mortgage payments. This deferral puts a significant amount of additional cash in the borrower’s pocket. We now examine how borrowers use this additional liquidity, examining the rich set of informationinCRISMabouttheborrower’snon-mortgagedebtaccounts. We estimate these effects using the same methodology, but replacing the dependent variable in Equation 2 with various nonmortgage outcomes. Results are presented in Figures4and5. Figure 4 shows that forbearance availability induced some borrowers to pay down credit card balances. Borrowers with below-median credit card utilization rate at highforbearance servicers paid off around $40 relative to borrowers at low-availability servicers (Figure 4(b)). Because the forbearance rate is higher by 5 percentage points for borrowers at high-forbearance servicers27, the result implies that marginal borrowers who received forbearance as a result of assignment to a high-forbearance servicer reduced their credit card balances by about $800. This difference is about a quarter of the average forbearance-driven savings in mortgage payments of those borrowers at highforbearance servicers and is also about a quarter of the conditional mean credit card balance for borrowers with low credit card utilization. We do not find robust evidence that higher-utilization borrowers paid down credit cards, and the standard errors on these specificationsaremuchlarger(Figure4(a)). This finding shows that forbearance essentially provided a low-cost source of liquiditytohouseholds,partiallyreplacingexpensivecreditcarddebts. Householdswithlower credit card utilization before the pandemic may be less liquidity-constrained than high- 27Thedifferenceintheforbearancerateacrosshigh-andlow-forbearanceservicersdoesnotvarybycredit utilization 30

Figure 4: Effects of forbearance availability on credit card balances. Figure plots estimates and their 95%confidenceintervalsoftheeffectsofassignmenttoahigh-forbearanceserviceroncreditcarddebtfor borrowerswithabove-andbelow-medianaveragecreditcardutilizationfortheperiodfromOctober2019 to March 2020. The median average utilization is calculated for each cohort of borrowers with the same mortgageoriginationyear. Standarderrorsareclusteredattheservicerlevel. 200 Mar = 0 Avg from Apr-Jul = -8.9 (26.9) Avg from Aug-Nov = -8.9 (56.6) Sample mean = 13346 100 0 -100 -200 2019m10 2020m1 2020m4 2020m7 2020m10 (a)Creditcardbalanceamonghigh-credit-card-utilizationborrowers($) Mar = 0 20 Avg from Apr-Jul = -15.3 (11.9) Avg from Aug-Nov = -32.2** (15.7) Sample mean = 3096 0 -20 -40 -60 -80 2019m10 2020m1 2020m4 2020m7 2020m10 (b)Creditcardbalanceamonglow-credit-card-utilizationborrowers 31

Figure5:Effectsofforbearanceavailabilityonupdatedcreditscore. Figureplotsestimatesandtheir95% confidence intervals of the effects of assignment to a high-forbearance servicer on the borrower’s credit score(FICOScoreversion5),asmeasuredinCRISM.Standarderrorsareclusteredattheservicerlevel. 1 Mar = 0 Avg from Apr-Jul = 0.1 (0.2) Avg from Aug-Nov = -0.1 (0.4) Sample mean = 696 .5 0 -.5 -1 2019m10 2020m1 2020m4 2020m7 2020m10 utilization borrowers, which may explain why they were more willing to use the additional funds to pay down credit card debt rather than for consumption or to increase liquidassets. We find no evidence that borrowers used forbearance to pay down other sources of debt like auto loans, student debt, or junior liens (Table A.11). This is perhaps unsurprising, as these forms of borrowing are much cheaper than credit card debt, making them a lower priority for payoff. (Additionally, our analysis in this section relies on a relatively small absolute difference in forbearance rates across servicer-types, so we are unlikely to havesufficientpowertomeasuresmallchangesinaveragebalances.) Wealsodonotfind that households assigned to higher-forbearance servicers purchased more cars. We find no effect on the delinquency rates of non-housing debt, though the availability of other forbearanceprogramsmayhaveaffectedtheseoutcomesaswell. Figure 5 shows that although borrowers at high-forbearance servicers are more likely to miss mortgage payments, their credit scores (FICO Score version 5)28 did not decrease asaresult,becausenonpaymentduringforbearanceisnotreportedtothecreditbureaus. 28FICOisaregisteredtrademarkofFairIsaacCorporation. 32

Infact,creditscoresforthehigh-forbearancegroupofborrowersactuallyincreaseslightly (perhaps reflecting their paydown of credit card balances and/or avoiding delinquency on non-mortgage debt), although the effect is estimated with a large standard error once weincludeservicerfixedeffects. These results indicate that the CARES Act forbearance program provided a low-cost source of liquidity to mortgage borrowers, which in part allowed some borrowers to reduceotherhigher-costsourcesofborrowing. 8 Moral Hazard Forbearance may induce mortgage nonpayment through two channels. First, borrowers who experienced a negative income shock could miss mortgage payments and use the additional liquidity to smooth their consumption or avoid foreclosure. Second, borrowers who did not experience a reduction in income may miss payments simply because forbearance represented a low-cost form of borrowing. The second channel represents a formofmoralhazard,inthatitisanunintendedconsequenceoftheprogram. Therelativesizeofthesetwochannelshasimportantwelfareimplicationsforforbearance program design. If the liquidity channel dominates, then the CARES Act forbearance program reached the intended households, and the actions of “low-forbearance” servicers prevented more borrowers from benefiting from the program on the margin. If the moral hazard channel dominates, it would imply that the program design, including easy access to forbearance and the ability to defer payments until loan termination, led to poor targeting. These trade-offs are analogous to those faced by other social insurance programssuchasunemploymentinsurance(e.g.,Chetty,2008)andpersonalbankruptcy (e.g.,Indarte,2020). Our data are not particularly well-suited to estimate the extent of moral hazard, be- 33

cause we do not have access to high-frequency dynamic income and employment data for our sample. Even so, several pieces of evidence suggest that most borrowers who skipped payments in forbearance did so as a result of negative income shocks. First, Table A.12 shows that the characteristics of borrowers in forbearance are comparable between high- and low-availability servicers, suggesting that easier access did not draw observably less-risky borrowers on the margin. For example, the average non-mortgage balances and average credit scores are very similar between the two groups. Borrowers athigh-forbearanceservicerstendtobeinforbearancelongerandlesslikelyexitforbearance,butthedifferencesbetweenthetwogroupsarequantitativelysmall. Ifmoralhazard were the main channel driving nonpayments, we might instead expect high-financialliteracy borrowers at high-forbearance servicers to use the program more intensively: to missmorepaymentsandtoremaininforbearancelonger. Zhao et al. (2020), who have access to borrower income data, provide more direct evidencethatforbearanceismostlyusedbyborrowerswhoexperiencednegativeincome shocks. They document that borrowers who made use of forbearance to miss payments experienced larger declines in income, were more likely to have lost their jobs, and more likely to have received unemployment benefits than those not in forbearance. Lambie- Hanson et al. (2021) present survey evidence indicating that at least three-quarters of borrowers entering forbearance had experienced a job disruption or income loss during thepandemic. A final point is that in aggregate only a relatively small proportion of borrowers used forbearance to skip mortgage payments. In principle, many borrowers could have acted in an opportunistic manner to take advantage of the generous repayment terms offered through the forbearance program. But it is clear that the vast majority of borrowers who wereabletomaketheirmortgagepaymentsdidkeeppaying. 34

9 Conclusion Our evidence indicates that servicer policies and practices played an important role in theimplementationoftheCARESActmortgageforbearanceprogram. Despiteuniversal eligibility for forbearance among agency mortgage borrowers, a significant fraction of delinquentborrowersdidnotsuccessfullyenterintoaforbearanceprogram,andthatthe relative frequency of these “missing” forbearances varies significantly across mortgage servicers for otherwise identical loans. Forbearance outcomes are systematically related to servicer characteristics including size, liquidity and organizational form, consistent withtheroleofincentivesinshapingservicerbehavior. Usingestimatedservicer-levelvariationinforbearancepractices,wealsofindthatforbearance has significant causal effects on borrower financial outcomes. In particular, we find that assignment to a “high-forbearance” servicer translates to a significantly higher non-payment rate, without any negative effect on borrowers’ credit scores, and that part of this additional household liquidity is used to pay-down high-cost credit card debt. It doesnotappearthatassignmenttoahigh-forbearanceservicerpreventednegativehousingoutcomeslikedelinquencyoutsideofforbearance,default,orforcedsales. We emphasize that our results represent the marginal effect of forbearance among different types of servicers, and therefore do not necessarily represent the average effect of the program on its recipients as a whole. Furthermore, our results do not speak to any generalequilibriumeffectsofforbearance. Ourresultshaveimportantimplicationsforwhether,expost,servicersbenefitedfrom making the program widely available. To servicers, forbearance take-up that does not preventdelinquencyorforeclosureiscostly;theservicerdoesnotinternalizenon-housing program benefits, and unless the servicer is very large, it does not internalize general equilibriumbenefits. Ourresultssuggestthatservicerswithfewerresourcessuccessfully preservedliquiditybyrestrictingaccesstoforbearance. 35

Overall, the CARES Act mortgage forbearance program has been successful in enrolling a large number of borrowers in a short period of time, significantly mitigating the negative shock of the COVID-19 pandemic on household liquidity. The low aggregate level of nonpayment suggests that despite a high rate of induced missed payments amongforbearanceusers,theprogramwaswell-targetedtohouseholdsthatexperienced hardship. Evenso,ourresultsshowthatidiosyncraticdifferencesacrossservicersplayed a significant role in the rollout of the program and shaped household outcomes. Policymakers may wish to consider whether future debt relief programs can include features (e.g., auto-enrollment) that overcome servicer reluctance and mitigate variation in outcomesthatisunrelatedtoborrowerfundamentals,orwhethercentralizingsomeportions of the program’s operations could overcome the specific challenges faced by smaller servicers. 36

References AGARWAL, S., G. AMROMIN, I. BEN-DAVID, S. CHOMSISENGPHET, AND D. EVANOFF (2011): “The role of securitization in mortgage renegotiation,” Journal of Financial Economics,102,559–578. AGARWAL, S., G. AMROMIN, I. BEN-DAVID, S. CHOMSISENGPHET, T. PISKORSKI, AND A. SERU (2017a): “PolicyInterventioninDebtRenegotiation: EvidencefromtheHome AffordableModificationProgram,”JournalofPoliticalEconomy,125,654–712. AGARWAL, S., G. AMROMIN, S. CHOMSISENGPHET, T. LANDVOIGT, T. PISKORSKI, A. SERU, AND V. YAO (2015): “Mortgage Refinancing, Consumer Spending, and Competition: Evidence from the Home Affordable Refinancing Program,” NBER Working Papers21512,NationalBureauofEconomicResearch,Inc. AGARWAL, S., S. CHOMSISENGPHET, AND Y. ZHANG (2017b): “How does working in a financeprofessionaffectmortgagedelinquency?” JournalofBanking Finance,78,1–13. AIELLO, D. J. (2021): “Financially constrained mortgage servicers,” Journal of Financial Economics. AN, X., L. CORDELL, L. GENG, AND K. LEE(2021): “InequalityintheTimeofCOVID-19: EvidencefromMortgageDelinquencyandForbearance,”Tech.rep. BERGER, A. N., N. H. MILLER, M. A. PETERSEN, R. G. RAJAN, AND J. C. STEIN (2005): “Does function follow organizational form? Evidence from the lending practices of largeandsmallbanks,”JournalofFinancialEconomics,76,237–269. BLACK KNIGHT (2020): “Forbearances rise after three weeks of declines,” McDash Flash BlogJune26. 37

BUCHAK, G., G. MATVOS, T. PISKORSKI, AND A. SERU (2018): “Fintech,regulatoryarbitrage,andtheriseofshadowbanks,”JournalofFinancialEconomics,130,453–483. CHERRY, S. F., E. X. JIANG, G. MATVOS, T. PISKORSKI, AND A. SERU (2021): “Government and Private Household Debt Relief during COVID-19,” NBER Working Papers 28357,NationalBureauofEconomicResearch. CHETTY, R. (2008): “Moral Hazard versus Liquidity and Optimal Unemployment Insurance,”JournalofPoliticalEconomy,116,173–234. CONSUMER FINANCIAL PROTECTION BUREAU (2021a): “COVID-19 Prioritized Assessments,SpecialEdition,”SupervisoryHighlights23. ——— (2021b): “Mortgage forbearance issues described in consumer complaint,” ComplaintBulletin,April. DEPARTMENT OF HOUSING AND URBAN DEVELOPMENT (2020a): “FHA’s COVID-19 LossMitigationOptions,”MortgageeLetter2020-22,July8. ——— (2020b): “HUD Issues New CARES Act Mortgage Payment Relief for FHA Single FamilyHomeowners,”PressRelease,April1. FEDERAL HOUSING FINANCE AGENCY (2020): “FHFA Announces Payment Deferral as New Repayment Option for Homeowners in COVID-19 Forbearance Plans,” News release,May13. ——— (2021): “FHFA Extends COVID-19 Forbearance Period and Foreclosure and REO EvictionMoratoriums,”Pressrelease,February25. FREDDIE MAC (2020): “FreddieMac: LumpSumRepaymentisNotRequiredinForbearance,”Homeownershipblog,April16. 38

FUSTER, A., M. PLOSSER, P. SCHNABL, AND J. VICKERY (2019): “TheRoleofTechnology inMortgageLending,”ReviewofFinancialStudies,32,1854–1899. FUSTER, A., M. PLOSSER, AND J. VICKERY(2021): “DoesCFPBOversightCrimpCredit?” CEPRDiscussionPapers15681,C.E.P.R.DiscussionPapers. GERARDI, K., L. GOETTE, AND S. MEIER (2013): “Numerical ability predicts mortgage default,”ProceedingsoftheNationalAcademyofSciences,110,11267–11271. GRANJA, J., C. MAKRIDIS, C. YANNELIS, AND E. ZWICK (2020): “Did the Paycheck Protection Program Hit the Target?” NBER Working Papers 27095, National Bureau of EconomicResearch,Inc. INDARTE, S.(2020): “MoralHazardversusLiquidityinHouseholdBankruptcy,”Working paper. JENSEN, M. C. AND W. H. MECKLING (1976): “Theory of the firm: Managerial behavior, agencycostsandownershipstructure,”JournalofFinancialEconomics,3,305–360. KIM, Y. S., S. M. LAUFER, K. PENCE, R. STANTON, AND N. WALLACE (2018): “Liquidity CrisesintheMortgageMarket,”BrookingsPapersonEconomicActivity,49,347–428. KRUGER, S. (2018): “The effect of mortgage securitization on foreclosure and modification,”JournalofFinancialEconomics,129,586–607. LAMBIE-HANSON, L., J. VICKERY, AND T. AKANA (2021): “Recent Data on Mortgage Forbearance: Borrower Uptake and Understanding of Lender Accommodations,” ResearchBrief,FederalReserveBankofPhiladelphia. LEE, D. AND W. V. DER KLAAUW (2010): “An introduction to the FRBNY Consumer CreditPanel,”StaffReports479,FederalReserveBankofNewYork. 39

MORTGAGE BANKERS ASSOCIATION (2020): “Share of mortgage loans in forbearance increasesto8.55%,”PressReleaseJune15. THE WHITE HOUSE (2021): “Fact Sheet: Biden Administration Announces Extension of COVID-19ForbearanceandForeclosureProtectionsforHomeowners,”Release,February16. WALL STREET JOURNAL (2020): “Struggling Borrowers Want To Pause Their Mortgage Payments.ItHasn’tBeenEasy,”Tech.Rep.April16. ZHAO, C., D. FARRELL, AND F. GREIG (2020): “Did Mortgage Forbearance Reach the Right Homeowners? Income and Liquid Assets Trends for Homeowners during the COVID-19Pandemic,”PolicyBrief,J.P.MorganChaseInstitute. 40

Internet Appendix for: “Intermediation Frictions in Debt Relief: Evidence from CARES Act Forbearance” You Suk Kim, Donghoon Lee, Tess Scharlemann, and James Vickery March 8, 2022

A Datasets eMBS loan-level data. eMBS provides information on the characteristics of the population of mortgages securitized into agency MBS. The data include standard underwriting fieldssuchascreditscoreatorigination,loan-to-valueratio,loanamount,mortgagerate, and property location (state). The data set also includes dynamic information about loan performance, such as updated principal balance, nonpayment status, and crucial for our analysis, the servicer identity. Our sample consists of FHA and VA loans, which account for92%ofallloanssecuritizedintoGinnieMaeMBS. Ginnie Mae forbearance register. We measure forbearance outcomes using Ginnie Maedatalistingthemonthlyloan-levelforbearancehistoryofloanssecuritizedintoGinnie Mae MBS. The file indicates the start date of the forbearance policy, the scheduled enddate,andthenumberofmonthsofforbearancegranted. Thedatawerefirstreleased publiclyinJune2020,andwerebackfilledtothestartofthepandemicforloansthatwere inforbearanceasofJune. Theyhavesubsequentlybeenupdatedonamonthlybasis.1 Financial Call Reports. Data on servicer characteristics are drawn from quarterly regulatory filings. For bank servicers we use the bank call reports. For independent mortgage banks we use mortgage call reports (MCRs) data. MCRs are filed by financial data companies holding a license through the Nationwide Mortgage Licensing System, including all bank and nonbank agency MBS servicers. The data include balance sheet and income data and other information on business activities. Together the bank and nonbank call report datasets allow us to link servicer characteristics to forbearance and delinquencyoutcomes. Black Knight McDash and CRISM. Black Knight McDash (hereafter “McDash”) includesloancharacteristicsandperformancefortheservicingportfoliosofthelargestres- 1Onerelativelyminorreportingissueisthattheinitialreleaseoftheforbearancedataonlyincludesloans thatwereinforbearanceasofJune2020. Thus,thedatadonotallowustoobserveforbearanceamongof borrowerswhoenteredforbearanceinMarchbuthadalreadyexitedpriortoJune. 1

idential mortgage servicers in the US, covering around two-thirds of the servicing market. The Equifax Credit Risk Insight Servicing and McDash (CRISM) dataset is a match between McDash and credit bureau data on nearly 79 million individual consumers, including information on other forms of debt (e.g., credit cards, junior liens, and student loans)forprimaryborrowersandallco-borrowersontheMcDashmortgages. FRBY Consumer Credit Panel / Equifax Data (CCP). The CCP is a representative panel of the credit history of an anonymous 5% sample of the U.S. adult population (see LeeandderKlaauw(2010)fordetailsofthedataset). NarrativecodesintheCCPtogether with scheduled payment variables allow us to measure the incidence of mortgage forbearance. TheCCPdoesnotincludeloanperformancedataformortgagesinforbearance plans, since that information is not reported to credit bureaus. We use the CCP to calculate forbearance rates for the overall mortgage market (Figure 1), and to cross-validate theforbearanceinformationintheGinnieMaedata. A.1 eMBS-CRISM merge Unlike eMBS, CRISM does not include servicer identities. We are however able to merge CRISM with a vector of anonymous servicer identifiers by undertaking a fuzzy match between CRISM/McDash with eMBS loan-level data based on mortgage balance at origination,originationyear-month,mortgagerate,creditscore,whetheraloanisanFHAor VAloan,andstate.2 Thismatcheddatasetallowsustotraceouttheeffectsofservicervariationinforbearance practices on other borrower outcomes (e.g., total household debt and the performance of non-mortgage debt). It also enriches the set of available borrower-level characteristics relative to the eMBS-only dataset (e.g., since CRISM/McDash includes finer ge- 2NotethattheFederalReserve’stermsofuseagreementwithBlackKnightdoesnotpermitustoretain servicercharacteristicsinthismergeddataset. Weareabletoretainananonymizedserviceridentifier, however. 2

ographic information on the property location, and allows us to observe the borrower’s refreshed credit score just prior to the pandemic). A limitation however is that only a subset of loans can be matched, whereas in eMBS we essentially are able to observe the entireuniverseofFHAandVAmortgages. Table A.13 presents summary statistics of loan characteristics of the full eMBS data and the merged eMBS-CRISM data. As shown by the number of observations in the two columns, about 25% of loans in the eMBS data are matched to CRISM in part because of coverage of the CRISM data and our restrictive matching criteria. Although many eMBS loans are excluded in the merged data, loan characteristics are still similar between the twodifferentdatasets. B Mortgages 30+ days past due, by segment Fraction of active mortgages that are at least 30 days past due (including those that are in forbearance). AuthorcalculationsbasedonBlackKnightMcDashservicingdata. 3

FigureA.1: DelinquencyRate,30+Days 4

C Loan-level estimates C.1 eMBS sample Notes: Linear probability regression of the probability that a loan enters forbearance from March 2021 onwards, based on eMBS loan-level data. Sample is loans that are active as of January 2021. Regressions includestateandservicerfixedeffects. 5

TableA.1: First-stageregression: dependentvariable=1ifinforbearance (1) (2) EverDQsample Fullsample Everservicerchange -0.031∗∗∗ 0.001∗∗ (0.002) (0.000) Monthssincelastservicerchange 0.001∗∗∗ -0.000∗∗∗ (0.000) (0.000) First-timehomebuyer 0.035∗∗∗ 0.029∗∗∗ (0.001) (0.000) DTIatorig: 25<dti≤50 0.044∗∗∗ 0.029∗∗∗ (0.002) (0.000) dti>50 0.087∗∗∗ 0.072∗∗∗ (0.002) (0.001) Loanage(year) -0.016∗∗∗ -0.003∗∗∗ (0.000) (0.000) Loanage(year)×Loanage(year) 0.000∗∗∗ 0.000∗∗∗ (0.000) (0.000) Ln(CurrentUPB) 0.097∗∗∗ 0.021∗∗∗ (0.001) (0.000) CSatorig: 620<origcs≤680 0.016∗∗∗ -0.020∗∗∗ (0.001) (0.000) 680<origcs≤740 0.023∗∗∗ -0.067∗∗∗ (0.002) (0.001) origcs>740 0.003 -0.108∗∗∗ (0.002) (0.001) Loanpurpose: refinace 0.043∗∗∗ 0.012∗∗∗ (0.002) (0.000) LTVatorig: 80<LTV≤95 0.028∗∗∗ 0.012∗∗∗ (0.002) (0.000) 95<LTV≤100 0.034∗∗∗ 0.021∗∗∗ (0.002) (0.000) LTV>100 -0.027∗∗∗ -0.021∗∗∗ (0.002) (0.001) 30+daysdelinquentinFeb2020 -0.293∗∗∗ (0.010) FHA -0.121∗∗∗ (0.001) ServicerFE Y Y StateFE Y Y N.Obs. 1,197,226 9,764,941 Adj.R2 0.10 0.09 6

C.2 eMBS-CRISM sample Table A.2: First-stage forbearance regression: eMBS-CRISM. Dependent variable = 1 if mortgageentersforbearance. Loan-levellinearprobabilitymodelofprobabilitythatpastdue loan enters into a forbearance plan based on eMBS-CRISM matched sample. Sample is loans that are current as of January 2020 and become past-due between March and November2020. Regressionsincludegeographyandservicerfixedeffects. (1) (2) Forbearance|delinquent Forbearance|delinquent First-timehomebuyer 0.0255∗∗∗ 0.0312∗∗∗ (0.00174) (0.00170) DTIatorig: 25<dti≤50 0.0458∗∗∗ 0.0569∗∗∗ (0.00379) (0.00371) dti>50 0.0693∗∗∗ 0.0894∗∗∗ (0.00410) (0.00401) Loanage(years)×Loanage 0.000∗∗∗ 0.000 (0.000) (0.000) Ln(CurrentUPB) 0.0682∗∗∗ 0.102∗∗∗ (0.00186) (0.00143) 620<origcs≤680 -0.00672∗ 0.00590∗ (0.00267) (0.00263) 680<origcs≤740 -0.0105∗∗∗ 0.0108∗∗∗ (0.00285) (0.00278) Loanpurpose: Refinance 0.0197∗∗∗ 0.0193∗∗∗ (0.00296) (0.00290) LTVatorigination: 80<LTV≤95 0.0150∗∗∗ 0.0188∗∗∗ (0.00357) (0.00348) 95<LTV≤100 0.0243∗∗∗ 0.0284∗∗∗ (0.00358) (0.00347) FHA 0.0645∗∗∗ 0.0836∗∗∗ (0.00241) (0.00231) Borrowerage: 30<age≤45 0.0144∗∗∗ (0.00251) 45<age≤60 0.0148∗∗∗ (0.00265) age>60 -0.0207∗∗∗ (0.00308) Riskscore+ 0.000136∗∗∗ (0.00000764) Ln(Consumerdebt)+ 0.0247∗∗∗ (0.000512) Delinq.consumerdebt+ -0.0333∗∗∗ (0.00182) Otherhousingdebt+ 0.00818∗∗∗ (0.00201) Delinq.otherhousingdebt+ -0.0227∗ (0.00923) Creditutilization+ 0.00982∗∗∗ (0.00217) N 425142 429841 StateFE No No ServicerFE Yes Yes ZipcodeFE Yes Yes Standarderrorsinparentheses +MeasuredasofFebruary2020 ∗p<0.05,∗∗p<0.01,∗∗∗p<0.001 7

D Alternative measures of servicer fixed effects FigureA.2: “Missing”forbearancerate: eMBS-CRISMsample. Histogrambasedonservicerfixedeffects foreMBS-CRISMmatchedsample. 30 10th percentile 90th percentile servicer servicer (weighted) (weighted) Median servicer (weighted) 20 10 0 srecivreS fo rebmuN 0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 Missing Forbearance 8

Figure A.3: Robustness of servicer fixed effects to controls: eMBS-CRISM sample. Panel (a) shows the correlation between servicer FE estimated using borrower and servicer characteristics available only in eMBSandservicerFEestimatedusingborrowerandservicercharacteristicsavailableinCRISM.Panel(b) showsthecorrelationbetweenservicerFEestimatedwithoutcontrolsandservicerFEestimatedusingall controlsavailableintheCRISM-eMBSmerge. (a)ServicerFEEstimatedusingallcontrolsandcontrolsonlyavailableineMBS slortnoc lla gnisu EF recivreS 5. 0 5.- 1- -1 -.5 0 .5 Servicer FE using only eMBS controls (b)ServicerFEestimatedusingfullsetofeMBS-CRISMcontrolsvsnocontrols slortnoc lla gnisu EF recivreS 5. 0 5.- 1- -.8 -.6 -.4 -.2 0 .2 Servicer FE using no controls 9

D.1 Comparison of fixed effects across approaches FigureA.4: Correlationbetweenservicerfixedeffectsfromdifferentspecifications: Thesefiguresshow correlationsbetweenthebaselineservicerfixedeffectestimatesandthreealternativesetsofestimates,based on:(i)usingthesubsampleofloanswhichbecameatleast60daysdelinquent(DQ)afterMarch2020(panel a);(ii)usingthesubsampleofborrowerswhomissedatleastapaymentinFebruaryorMarch2020(panel b); using the entire sample for estimation, rather than just borrowers that became delinquent (panel c); includelenderfixedeffectsinthemodel,sothatidentificationofservicerfixedeffectsisbasedonservicing transfers(paneld). 1 .9 .8 .7 .6 .5 )raM ecnis qd yad +06( sEF recivreS .8 .6 .4 .2 0 .4 .6 .8 1 Servicer FEs (30+ day dq since Mar) (a)60+daypastduepost-March2020 )raM ro beF ni qd yad +03( sEF recivreS .4 .6 .8 1 Servicer FEs (30+ day dq since Mar) (b)MissedpaymentinFeb/Mar2020 .25 .2 .15 .1 .05 0 )elpmas lluf( sEF recivreS 1 .8 .6 .4 .4 .6 .8 1 Servicer FEs (subsample) (c)includeallloansinsample )EF rednel htiw detamitse( sEF recivreS .4 .5 .6 .7 .8 Servicer FEs (30+ day dq since Mar) (d)includelenderfixedeffects 10

E Borrower Characteristics by Servicer Type Table A.3: Borrower Characteristics across Servicers (CRISM-eMBS match) This table presentssummarystatisticsmeasuredasofFebruary2020forhigh-andlow-forbearance servicersusingthemergedeMBS-CRISMdata. Wedefine“high-forbearance”servicersas thosewithabove-medianservicerfixedeffects(estimatedasdescribedinSection4,using themergedeMBS-CRISMdata). (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 184,313.66 163,431.85 AutoLoanBalance 16,102.42 15,254.90 CreditCardBalance 8,969.34 8,709.84 12-mochangeCNTYUR(8/20) 6.02 5.80 FHA 0.67 0.70 FICOV5(updated) 693.90 703.46 LTVatorigination 93.92 94.22 Loanage(year) 4.49 6.04 N.Obs. 1,244,059 1,521,734 11

Table A.4: Borrower characteristics across servicers by origination year (CRISM-eMBS match) This table presents summary statistics measured as of February 2020 for highand low-forbearance servicers using the merged eMBS-CRISM data. We define “highforbearance” servicers as those with above-median servicer fixed effects (estimated as describedinSection4,usingthemergedeMBS-CRISMdata). (a)Originationyearupto2013 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 135,368.15 136,303.53 CurrentMortgageBalance 135,368.15 136,303.53 12-mochangeCNTYUR(8/20) 6.15 5.82 FHA 0.83 0.77 FICOV5(updated) 712.85 707.77 LTVatorigination 93.59 93.64 Loanage(year) 8.52 8.63 AutoLoanBalance 13,323.52 13,931.91 CreditCardBalance 9,135.99 8,821.26 N.Obs. 330,927 709,437 (b)Originationyearfrom2014to2017 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 184,124.12 178,095.05 CurrentMortgageBalance 184,124.12 178,095.05 12-mochangeCNTYUR(8/20) 6.06 5.81 FHA 0.68 0.68 FICOV5(updated) 695.99 701.66 LTVatorigination 94.05 94.82 Loanage(year) 4.59 4.82 AutoLoanBalance 16,511.15 16,469.54 CreditCardBalance 9,324.28 8,966.78 N.Obs. 383,878 499,469 (c)Originationyearsince2018 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 220,666.63 206,795.61 CurrentMortgageBalance 220,666.63 206,795.61 12-mochangeCNTYUR(8/20) 5.92 5.75 FHA 0.57 0.59 FICOV5(updated) 680.54 696.57 LTVatorigination 94.02 94.60 Loanage(year) 1.91 2.11 AutoLoanBalance 17,543.53 16,315.86 CreditCardBalance 8,607.69 8,046.91 N.Obs. 529,254 312,828 12

Table A.5: Borrower characteristics across servicers by origination year (eMBS) This table presents summary statistics measured as of February 2020 for high- and lowforbearance servicers using the eMBS sample. We define “high-forbearance” servicers as those with above-median servicer fixed effects (estimated as described in Section 4, usingtheeMBSdata). (a)Originationyearupto2013 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 114,464.79 118,121.23 12-mochangeCNTYUR(8/20) 5.95 5.91 FHA 0.80 0.77 Origcreditscore 699.49 705.85 OrigLTV(%) 92.62 92.67 Loanage(year) 10.26 10.11 N.Obs. 1,039,878 1,894,045 (b)Originationyearfrom2014to2017 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 178,017.12 175,246.78 12-mochangeCNTYUR(8/20) 6.09 5.87 FHA 0.69 0.61 Origcreditscore 690.85 702.36 OrigLTV(%) 93.38 93.06 Loanage(year) 4.62 4.71 N.Obs. 1,150,984 1,200,859 (c)Originationyearsince2018 (1) (2) Low-ForbearanceServicer High-ForbearanceServicer CurrentMortgageBalance 216,955.67 203,728.56 12-mochangeCNTYUR(8/20) 6.15 5.77 FHA 0.69 0.58 Origcreditscore 683.56 696.60 OrigLTV(%) 94.56 93.46 Loanage(year) 2.05 2.12 N.Obs. 1,843,638 1,326,763 13

F Alternative specifications: Role of servicer characteristics in forbearance policy TableA.6: ServicerFEs(conditionalon60+dq) All Nonbankmtgcompanies Banks (1) (2) (3) (4) (5) (6) (7) Servicercharacteristics log(Servicingassets) 0.025∗∗∗ 0.024∗∗∗ 0.017∗∗ 0.023∗∗∗ 0.024∗∗∗ (0.005) (0.007) (0.007) (0.007) (0.006) log(Assets) 0.011∗∗∗ 0.015∗ (0.004) (0.009) Cash/assets 0.622∗∗∗ 0.701∗∗∗ -0.317 -0.483 (0.142) (0.134) (0.395) (0.448) Securities/assets 0.188 0.266∗∗∗ 0.127 0.258 (0.115) (0.080) (0.189) (0.211) Capital/assets 0.007 0.033 0.793 0.671 (0.125) (0.126) (0.635) (0.653) Servicinggrowth -0.033 -0.031 -0.059 -0.076 -0.050 -0.037 -0.047 (0.046) (0.057) (0.060) (0.056) (0.080) (0.083) (0.083) Servicertype Nonbankmortgagecompany -0.021 (0.017) Creditunion 0.078∗∗∗ (0.020) N.Obs. 151 97 97 97 45 45 45 14

TableA.7: ServicerFEs(controllingforlenderFEs) All Nonbankmtgcompanies Banks (1) (2) (3) (4) (5) (6) (7) Servicercharacteristics log(Servicingassets) 0.037∗∗∗ 0.035∗∗∗ 0.030∗∗∗ 0.038∗∗∗ 0.036∗∗∗ (0.006) (0.008) (0.006) (0.009) (0.009) log(Assets) 0.022∗∗∗ 0.022∗ (0.004) (0.012) Cash/assets 0.937∗∗∗ 1.094∗∗∗ -0.288 -0.529 (0.207) (0.210) (0.450) (0.586) Securities/assets 0.120 0.238∗∗ 0.397 0.590∗ (0.108) (0.091) (0.332) (0.310) Capital/assets 0.037 0.088 0.916 0.737 (0.106) (0.114) (0.688) (0.774) Servicinggrowth 0.015 0.038 0.019 -0.009 -0.025 0.012 -0.003 (0.047) (0.057) (0.045) (0.044) (0.076) (0.082) (0.082) Servicertype Nonbankmortgagecompany -0.065∗∗∗ (0.023) Creditunion 0.140∗∗∗ (0.027) N.Obs. 152 98 98 98 45 45 45 TableA.8: ServicerFEs(conditionalondelinquencytransitioninFeb-Mar2020) All Nonbankmtgcompanies Banks (1) (2) (3) (4) (5) (6) (7) Servicercharacteristics log(Servicingassets) 0.032∗∗∗ 0.029∗∗∗ 0.017∗∗∗ 0.031∗∗ 0.035∗∗∗ (0.008) (0.010) (0.006) (0.012) (0.010) log(Assets) 0.012∗∗ 0.018 (0.006) (0.014) Cash/assets 0.997∗∗∗ 1.078∗∗∗ -0.970 -1.101 (0.277) (0.283) (0.680) (0.761) Securities/assets 0.327∗∗ 0.399∗∗∗ 0.055 0.273 (0.126) (0.107) (0.373) (0.367) Capital/assets 0.073 0.100 1.798∗ 1.521 (0.143) (0.146) (0.967) (1.003) Servicinggrowth 0.011 0.066 0.016 -0.001 -0.079 -0.064 -0.082 (0.062) (0.067) (0.043) (0.043) (0.108) (0.111) (0.110) Servicertype Nonbankmortgagecompany -0.019 (0.032) Creditunion 0.087∗∗ (0.035) N.Obs. 148 96 96 96 44 44 44 15

TableA.9: ServicerFEs(fullsampleincludingcurrentloans) All Nonbankmtgcompanies Banks (1) (2) (3) (4) (5) (6) (7) Servicercharacteristics log(Servicingassets) 0.037∗∗∗ 0.031∗∗∗ 0.025∗∗∗ 0.043∗∗∗ 0.043∗∗∗ (0.007) (0.008) (0.006) (0.010) (0.010) log(Assets) 0.018∗∗∗ 0.025∗ (0.004) (0.015) Cash/assets 0.955∗∗∗ 1.083∗∗∗ -0.664 -0.942 (0.174) (0.177) (0.537) (0.699) Securities/assets 0.144 0.246∗∗∗ 0.320 0.553 (0.100) (0.085) (0.364) (0.345) Capital/assets 0.011 0.053 1.292 1.068 (0.103) (0.109) (0.794) (0.907) Servicinggrowth -0.008 0.011 -0.011 -0.035 -0.030 0.000 -0.018 (0.047) (0.055) (0.045) (0.042) (0.086) (0.089) (0.090) Servicertype Nonbankmortgagecompany -0.087∗∗∗ (0.025) Creditunion 0.131∗∗∗ (0.029) N.Obs. 152 98 98 98 45 45 45 16

G Deferred payments Figure A.5: Deferred payments This shows the results of the coefficients from Equation 2, where the dependentvariableisameasureofthetotalborrowingthroughforbearance: thenumberofmissedpayments times the monthly mortgage payment (including taxes and insurance). This is an estimate, as we cannot directlyobservewhetherborrowersmakepartialpaymentsorcontinuetopaytaxesandinsurance. Thecoefficientscanbeinterpretedastheaveragedifferenceincumulativedeferredpaymentsamongborrowers athigh-vs. lowservicers. 200 150 100 50 0 2019m10 2020m1 2020m4 2020m7 2020m10 (a)Carriedmortgagebalance($) 17

H Pre-CARES Act delinquencies FigureA.6: New30-daydelinquencies. Differenceinthetransitionprobabilityintodelinquencyforhighforbearance vs low-forbearance servicers (estimates of coefficients from Equation 2). The y-axis indicates the fraction of newly delinquent mortgages, defined as loans that are past due in month t but current in month t-1. Includes same borrower and loan controls as our main eMBS specification (reported in table A.1). Standarderrorsareclusteredbyservicer. 60. 40. 20. 0 20.- 2019m12 2020m2 2020m4 2020m6 2020m8 2020m10 18

Table A.10: Pre-CARES-Act delinquencies This table presents estimates of regressions of various measures of delinquencies before the CARES Act on the dummy variable for high-forbearanceservicersaswellasothercontrols. TheeMBSdatafromDecember2019 and January 2020 are used for the estimates in Table (a), and the matched eMBS-CRISM data from December 2019 and January 2020 are used for the estimates in Tables (b), (c), and (d). The dependent variable for Tables (a) and (b) is the dummy variable for turning 30-day delinquent for the mortgage. The dependent variables for Tables (c) and (d) are whether a borrower is delinquent in the credit card and auto loan accounts, respectively. EMBS controls include the dummy for FHA loans, loan size, the dummy for first-time homebuyers, LTV, credit score, DTI, and the dummy for purchase loans. CRISM controls include updated credit scores and a borrower’s age. Standard errors are clustered at the servicerlevel. (1) (2) (3) (4) High-forbearanceservicer -0.0027∗∗ -0.0016∗∗∗ -0.0015∗∗∗ -0.0016∗∗∗ (0.0013) (0.0005) (0.0005) (0.0005) EMBScontrols Y Y Y StateFE Y OrigYear-MonthFE Y FHAxStatexOrigYear-MonthFE Y NonbankxFHAxStatexOrigYear-MonthFE Y Samplemean 0.013 0.013 0.013 0.013 N.Obs. 22,010,182 20,180,908 20,180,907 20,180,906 (a)New30-daydelinquencies(eMBSonly) (1) (2) (3) (4) High-forbearanceservicer -0.0032 -0.0013 -0.0009 -0.0009 (0.0031) (0.0020) (0.0017) (0.0016) EMBScontrols Y Y Y CRISMcontrols Y Y ZipcodeFE Y Y OrigYear-MonthFE Y Y FHAxZipcodexOrigYear-MonthFE Y Samplemean 0.015 0.015 0.015 0.015 N.Obs. 5,756,749 5,400,684 5,393,643 5,385,113 (b)New30-daydelinquencies(eMBS-CRISMmatch) (1) (2) (3) (4) High-forbearanceservicer -0.0089 -0.0018 0.0022∗ 0.0025∗∗ (0.0085) (0.0030) (0.0013) (0.0011) EMBScontrols Y Y Y CRISMcontrols Y Y ZipcodeFE Y Y OrigYear-MonthFE Y Y FHAxZipcodexOrigYear-MonthFE Y Samplemean 0.099 0.099 0.099 0.099 N.Obs. 5,769,271 5,400,702 5,393,661 5,385,130 (c)Creditcarddelinquencies(eMBS-CRISMmatch) (1) (2) (3) (4) High-forbearanceservicer -0.0059∗ -0.0025∗∗ -0.0013∗∗ -0.0010∗ (0.0035) (0.0012) (0.0005) (0.0005) EMBScontrols Y Y Y CRISMcontrols Y Y ZipcodeFE Y Y OrigYear-MonthFE Y Y FHAxZipcodexOrigYear-MonthFE Y Samplemean 0.031 0.031 0.031 0.031 N.Obs. 5,769,271 5,400,702 5,393,661 5,385,130 (d)Autoloandelinquencies(eMBS-CRISMmatch) 19

I Additional non-mortgage results TableA.11: Non-mortgageresultsThistablepresentsasummaryofestimatesofequation (2) for various outcome variables. Column (1) report averages of the estimates of β to 1 β and standard errors of the averages in the parenthesis. Columns (2) report averages 4 of the estimates of β to β and standard errors of the averages in the parenthesis. For 5 8 outcome variables related to auto loans, we report ”NA” under column (1) because the Equifaxdatafortheperiodcontainsanerror. Standarderrorsareclusteredattheservicer level. (1) (2) (3) (4) 2020:m4to 2020:m8to Sample N.Obs. 2020:m7 2020:m11 mean Autoloanbalance NA 10.019 15,352 35,357,290 (27.975) Otherconsumerloanbalance 0.604 5.873 4,184 35,356,333 (4.734) (10.295) Transitiontodelinquency(creditcard) 0.00017 0.00019 0.01017 33,284,225 (0.00025) (0.00030) Transitiontodelinquency(autoloan) NA 0.00001 0.00523 33,284,225 (0.00007) Transitiontodelinquency(otherconsumerloan) -0.00002 0.00000 0.00353 33,284,225 (0.00009) (0.00008) Mortgageprepayment 0.0000 -0.0005 0.0143 33,550,744 (0.0006) (0.0009) Autoloanorigination NA -0.000 0.023 35,723,355 (0.000) 20

J Characteristics of Borrowers in Forbearance TableA.12: ComparingCharacteristicsofBorrowersinForbearanceacrossServicersThis table presents summary statistics measured as of February 2020 for borrowers that were ever in forbearance with high- and low-forbearance servicers using the merged eMBS- CRISMdata. (1) (2) Low-ForbearanceServicer High-ForbearanceServicer Monthsinforbearance(asofNov2020) 4.88 5.86 Everexitedfromforebarance 0.34 0.31 CurrentMortgageBalance 197,614.49 171,090.71 AutoLoanBalance 18,313.27 17,718.46 CreditCardBalance 10,912.77 11,514.34 12-mochangeCNTYUR(8/20) 6.58 6.32 FHA 0.79 0.81 FICOV5(updated) 647.57 662.15 LTVatorigination 94.31 94.73 Loanage(year) 3.89 5.61 N.Obs. 153,010 234,236 21

K Comparison between Full and Matched Samples Table A.13: Comparison between Full and Matched Samples This table presents summary statistics measured as of February 2020 for the eMBS data and the merged eMBS- CRISMdata. (1) (2) eMBS eMBS-CRISMmatch Ever30+daysdelinquent 0.19 0.19 Everinforbearance 0.14 0.15 CurrentUPB($) 168,647.51 171,873.43 OrigLTV(%) 93.40 94.58 OrigDTI(%) 40.26 40.21 Origcreditscore 692.65 696.85 Loanage(year) 4.93 5.33 FHA 0.68 0.71 VA 0.32 0.29 N.Obs. 10,315,121 2,883,200 22