How Resilient Is Mortgage Credit Supply? Evidence from the COVID-19 Pandemic

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. How Resilient Is Mortgage Credit Supply? Evidence from the COVID-19 Pandemic Andreas Fuster, Aurel Hizmo, Lauren Lambie-Hanson, James Vickery, Paul Willen 2021-048 Please cite this paper as: Fuster, Andreas, Aurel Hizmo, Lauren Lambie-Hanson, James Vickery, and Paul Willen (2021). “How Resilient Is Mortgage Credit Supply? Evidence from the COVID-19 Pandemic,” Finance and Economics Discussion Series 2021-048. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2021.048. 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.

How Resilient Is Mortgage Credit Supply? Evidence from the COVID-19 Pandemic* Andreas Fuster Aurel Hizmo Lauren Lambie-Hanson James Vickery Paul Willen July 12, 2021 Abstract WestudytheevolutionofUSmortgagecreditsupplyduringtheCOVID-19pandemic. Although the mortgage market experienced a historic boom in 2020, we show there was also a large and sustained increase in intermediation markups that limited the pass-throughoflowratestoborrowers. Markupstypicallyriseduringperiodsofpeak demand, but this historical relationship explains only part of the large increase during the pandemic. We present evidence that pandemic-related labor market frictions andoperationalbottleneckscontributedtounusuallyinelasticcreditsupply,andthat technology-based lenders, likely less constrained by these frictions, gained market share. Risingforbearanceand defaultriskdidnotsignificantly affectrateson“plainvanilla”conforming mortgages, but it did leadto higher spreadson mortgages without government guarantees and loans to the riskiest borrowers. Mortgage-backed securities purchases by the Federal Reserve also supported the flow of credit in the conformingsegment. Keywords: mortgage,credit,financialintermediation,fintech,COVID-19. JELclassification: G21,G23,G28. *Fuster: SwissNationalBankandCEPR.Hizmo: FederalReserveBoard. Lambie-HansonandVickery: FederalReserveBankofPhiladelphia. Willen: FederalReserveBankofBostonandNBER.WethankEric Hardy, Natalie Newton and Dick Oosthuizen for outstanding research assistance, Scott Frame, Philipp Schnabl,FelipeSeverino,andnumerousothercolleaguesforcomments,aswellasworkshopparticipants attheJournalofFinanceandFama-MillerCenterConferenceontheFinancialConsequencesofCOVID-19, University of Zurich, Louisiana State University, the Philadelphia Fed, Penn State University, Dartmouth (Tuck), theDallasFed, andtheCFPBResearch Conference. Opinionsexpressed inthispaperarethoseof theauthorsanddonotrepresenttheopinionsoftheFederalReserveBanksofBostonorPhiladelphia,the FederalReserveBoard,theFederalReserveSystem,ortheSwissNationalBank.

I Introduction 2020 was an extraordinary year for the US mortgage market. Lenders underwrote a record $4 trillion in new mortgages, a far higher volume than in any year since 2003.1 On the price side, the Freddie Mac Primary Mortgage Market Survey (PMMS) rate on 30-year fixed-rate mortgages (FRMs) fell below 3 percent for the first time since the series began in 1970, reaching lows of about 2.7 percent. Rocket Companies, the largest US mortgagelender,recorded$9.4billioninnetincome,upmorethan900percentfrom2019. Market participants in March of 2020 would not have expected such good news. The emergenceofCOVID-19appearedtopresagebadtimesfortheindustry. Withthespread of the virus, lockdowns, and social distancing, how would lenders close loans? Who would buy houses? How would mortgage servicers deal with commitments to pay principalandinteresttoinvestorsaswellastaxesandinsurancewhenCongresshadtoldborrowers that they could postpone payment without penalty? Yet borrowers and lenders seeminglyovercameallthesechallenges. Underlyingthesepositiveoutcomes,however,aresignsthatthemortgagemarkethas not functioned entirely normally during the pandemic. In addition to historically high lending and historically low rates, we have seen historically high spreads between mortgage rates and commonly used benchmarks. Figure 1 shows that while mortgage and Treasuryratesgenerallyco-moveclosely,thedifferencebetweenthe30-yearFRMrateand 10-year Treasury yield widened significantly during the pandemic, peaking at levels not seen since the 2008 financial crisis.2 And while mortgage lenders were highly profitable, industrypractitionersreportedtightercreditstandardsinresponsetotheheightenedrisk offorbearanceanddelinquency.3 In this paper, we study the evolution of the mortgage market during the COVID-19 1From2008through2019,annualmortgageoriginationsneverexceeded$2.4trillion,andtheyexceeded$2 trillion only three times. Lending volume was $3.7 trillion in 2003, which was the previous record. The 2003 lending boom was significantly larger than the 2020 episode in real terms or scaled by the stock of outstanding mortgages, however. Statistics are from Inside Mortgage Finance and can be seen on p. 8 of UrbanInstitute(2021). 2Theeffectivedurationofa30-yearmortgageismuchshorterthan30yearsduetoprepayment. Although widelyused,10-yearTreasuryyieldsarenotaperfectbenchmarkforreasonsweexplainbelow. 3Forexample,theMortgageBankersAssociationestimatesthatmortgagecreditavailabilitydroppedsharply attheonsetofthepandemictoitslowestlevelinsixyears,andithadnotrecoveredmuchthroughtheend of2020(MortgageBankersAssociation,2021). 1

Figure1: MortgageRatesandTreasuryYields (a)InterestRates (b)Mortgage-TreasurySpread 8 6 4 2 0 % 30−Year Mortgage Rate (PMMS) 3 10−Year Treasury Yield 2.5 2 1.5 1 01jan2000 01jan2005 01jan2010 01jan2015 01jan2020 % ,dleiY yrusaerT raeY−01 sunim etaR egagtroM raeY−03 01jan2000 01jan2005 01jan2010 01jan2015 01jan2020 Datasource: Freddie Mac 30-Year Fixed Rate Mortgage Average in theUnited States[MORTGAGE30US] and Board of Governors of the Federal Reserve System (US) 10-Year Treasury Constant Maturity Rate [DGS10],retrievedfromFRED,FederalReserveBankofSt. Louis. Sampleperiodthroughendof2020. pandemicandevaluatetheextenttowhichtheeventsofthepandemicledtoacontraction in mortgage credit supply. Our analysis combines financial market data on prices and yields for mortgage-backed securities (MBS), time-series data on mortgage interest rates, microdataonmortgagerateoffersandinterestratelocksfromtheOptimalBlueplatform, andseveralothermortgagedatasets. Wemakethefollowingfourmainpoints. 1. ExceptforaperiodoffinancialmarketturbulenceinMarch2020,thehighmortgage- Treasuryspreadismorethanaccountedforbyasustained100basispointincreasein the “primary-secondary” spread—that is, by an increase in the price of intermediationintheprimarymortgagemarket. Asimilarspikeintheintermediationmarkup is observed based on measures of gain-on-sale earned by mortgage lenders. Thus, thesituationisquitedifferentfrom2008,whenmortgagerateswereelevateddueto thehighspreadonMBSinfinancialmarkets. 2. Intermediation markups typically rise during refinancing booms due to capacity constraints(Fusteretal.,2017). Butthishistoricalrelationshipaccountsforonlypart oftheincreaseinmarkupsin2020. Inotherwords,theelasticityofmortgagesupply was abnormally low. We present evidence that operational issues and labor market frictionsrelatedtothepandemiccontributedtothislowerelasticity. Consistentwith 2

these frictions, we show that technology-based (“fintech”) lenders gained market shareamongmortgagesthataremorecomplexandlaborintensivetooriginate. 3. The mortgage rate spread widened for loans bearing the greatest credit risk for financial intermediaries. These loans span the socio-economic spectrum, including both non-guaranteed jumbo mortgages to high-income borrowers, and low-creditscore Federal Housing Administration (FHA) mortgages.4 The number of lenders offering credit in these segments of the market also fell sharply. While forbearance and default risk was not an important driver of higher markups for a typical prime mortgage,itdidleadtoacontractionofsupplyinsomeotherpartsofthemarket. 4. MBS purchases by the Federal Reserve (“quantitative easing” or QE) lowered mortgage rates and supported mortgage credit supply. We identify these effects in part usinginstitutionalfeaturesofthe“to-be-announced”forwardmarket. Wearealsoabletoruleoutseveralplausiblealternativeexplanationsforthesharprise in mortgage markups. Using geographic variation, we show that factors related to local market competition, or the direct macroeconomic and health effects of the virus, do not play an important role; instead, the rise in markups was broadly based. Using the cross sectionofmortgages,weshowthatforbearanceanddefaultriskdidnotplayasignificant role in producing the higher spreads observed in the prime conforming market, where lenders are less exposed to default than in other market segments. Finally, we show that borrowersshoppedmoreandswitchedmorefrequentlybetweenlendersduringthepandemic; along with our evidence on local competition, this speaks against the hypothesis thatthehighmarkupsreflectanincreaseinlenderpricingpower. Beyondsheddinglightonthefinancialeffectsofthepandemic,ourresultsyieldmore general lessons about the resilience of the US mortgage finance system and the frictions that limit interest rate pass-through to borrowers. A first lesson is that, despite recent improvements in information technology, industry capacity constraints still bind during periodsofpeakdemand. Althoughmortgageratesfelltorecordlowsin2020,ourresults imply that rates paid by borrowers would have been lower still under a system where rates adjust automatically with market yields during an economic downturn, as in the automatic stabilizer mortgage proposed by EberlyandKrishnamurthy (2014) (see also 4FHAmortgagesaregovernmentguaranteed,buttheseguaranteesdonotfullyinsulatemortgageintermediariesagainstdefaultrisk. SeeKimetal.(2018)andthediscussioninSectionVIformoredetails. 3

Gurenetal., 2020; Campbelletal., 2020), or in a system with a larger role for adjustablerate mortgages. That said, we also find that fintech lenders gained market share during the pandemic, particularly for complex loans, suggesting that technology indeed helps ameliorate capacity constraints and overcome operational bottlenecks, consistent with Fusteretal.(2019).5 A second lesson is that government guarantees, which play a key role in the agency mortgage market, supported intermediation and lowered rates in the face of a large macroeconomicshock,aswasalsothecaseduringtheGreatRecession(Calemetal.,2013; VickeryandWright,2013). Atthesametime,ourresultshighlightthelimitsoftheseguarantees. Because lenders are not fully indemnified against risk, especially in the FHA segment, the pandemic did result in a contraction in supply for the riskiest borrowers. We are also able to disentangle the effects of government credit guarantees and Fed QE (e.g., DiMaggioetal., 2020) by comparing conforming, jumbo, and super-conforming loans (the latter carry a government guarantee but are only partly eligible for Fed purchases). We find that both QE and government guarantees supported credit supply in the conformingsegment,althoughguaranteesappeartohavehadamorepersistenteffect. The evidence in this paper adds to a large body of research about the transmission of monetary policy and interest rate shocks through the mortgage market and the role of financial frictions, including Bergeretal. (2020), DiMaggioetal. (2017), Fusteretal. (2013), and Fusteretal. (2017); see Amrominetal. (2020) for a recent review. Our results arealsorelatedtoresearchonthegrowingpresenceofnonbankmortgageintermediaries, which now originate more than half of new loans and are more sensitive to liquidity risk (Kimetal.,2018;Jiangetal.,2020;Buchaketal.,2018). Finally, we contribute to a growing body of research about the mortgage market and consumercreditmarketsmoregenerallyduringtheCOVID-19pandemic.6 Anetal.(2021), Cherryetal. (2021), Capponi etal. (2021), and McManusandYannopoulos (2021) study mortgage forbearance during the pandemic. Agarwaletal. (2021) show that savings 5ErelandLiebersohn (2020) show that nonbank fintech lenders also gained importance in small-business lending during the pandemic, accounting for a substantial portion of Paycheck Protection Program (PPP) loans. Kwanetal.(2021)showthatbankswithstrongerinformationtechnology(IT)capabilitiesoriginated morePPPloansandattractedmoredepositsduringthepandemic. 6Brackeetal.(2020)conductastudysimilarinseveralrespectstooursfortheUKmortgagemarket,finding evidenceoftightercreditstandardsandadropinlendingaftertheonsetofthepandemic. Thisagainsuggeststhatgovernmentguarantees(whichareabsentintheUK)wereimportantinsustainingUSlending. 4

from refinancing during the pandemic are more skewed toward higher-income borrowers than in earlier refinancing waves. Work on other consumer credit markets includes Iversonetal.(2020)onbankruptcyandHorvathetal.(2020)oncreditcardborrowing. II Data II.A Optimal Blue The Optimal Blue platform connects mortgage lenders to whole loan investors, allowing lenderstosearchforpricinginformation,initiateratelocks,andsellmortgages.7 Lenders are typically nonbank mortgage companies, but smaller banks and credit unions are also represented. Investors include large banks and other institutions that retain loans purchased on the platform in portfolio or package them for securitization. More than 1,000 lendersandover200investorsareactiveontheplatform,andOptimalBlueestimatesthat the platform has been used to lock one-third of US mortgage originations in recent years. Coveragemaybeevenhigherfor2020duetoariseintheshareofnonbanklending.8 Ouranalysisusesthefollowingtwoformsofinformationproducedbytheplatform. Rate locks. Rate lock data reflect individual locks for mortgages processed by Optimal Blue, and include a comprehensive set of underwriting variables including the loanto-value (LTV) ratio, FICO score, debt-to-income (DTI) ratio, loan amount, loan program, purpose (purchase or refinance), asset and income documentation, employment, occupancystatus,propertytype,andZIPcode. Thedataalsoincludelender,branch,andloan officeridentifiers. Thedatacoverabout280metropolitanareasaswellasruralareas. Rate lock data have several advantages over servicing data often used in mortgage research. First, the data include not only the mortgage note rate, but also the discount points or credits paid or received by the borrower and the duration of the lock. This is important because in practice borrowers select from a menu of rate-point combinations, and points vary through time and across loans. Second, we observe the date and time 7Optimal Blue data (as referenced throughout) are anonymized mortgage market/rates data that do not containlenderorcustomeridentitiesorcompleteratesheets. 8Among the top 50 lenders by volume, nonbanks accounted for 69 percent of originations in 2020, and 74 percentin2020:Q4,upfrom60percentin2019. Source: InsideMortgageFinance,Jan. 29,2021. 5

of the rate lock, as opposed to the closing date, which generally differs from the pricingrelevantlockdatebyweeksormonths.9 Third,thedataareavailableinnearrealtime. Mortgageoffers. Wealsousedataonthemenuofloancontractsandassociatedinterest rates offered by lenders through Optimal Blue’s “Pricing Insight” engine. The engine allows users to retrieve the real-time distribution of offers (that is, combinations of rates and net points and fees) for a loan with given characteristics in a particular local market. Theinterfaceisdesignedforlenderstocomparetheirpricingwiththatoftheirpeers. We use the engine to run daily searches in one local market (Los Angeles), twice-weekly searches in four markets, and weekly searches in 15 additional markets, collecting offer data for 100 loan types representing different combinations of FICO score, LTV ratio, loan program, purpose (purchase or cash-out refinance), occupancy (owner-occupied or investor),ratetype(30-yearfixedor5/1adjustable),andloanamount.10 Mortgageofferratesrepresentadirectmeasureofsupplythatcanbeobservedregardless of whether the offer results in a loan in equilibrium. The Insight data also allow us to see how the numberof lendersactiveindifferent segments of the marketevolvesover the pandemic. A limitation of the offer data, however, is that there is no fixed lender identifier,meaningwecannottracklendersovertimeorincludelenderfixedeffects. II.B Other Data Sources We use a variety of other data sources. For mortgage rates, in addition to Optimal Blue, we use data from the Freddie Mac Primary Mortgage Market Survey and the Mortgage Bankers Association (MBA). All MBS pricing information comes from J.P. Morgan Markets. Mortgage servicing rights valuation data are provided by SitusAMC, an independent valuation service company. For direct evidence on lender income and costs, we use the MBA Quarterly Performance Report. Data on mortgage industry employment come from the Bureau of Labor Statistics’ (BLS) Current Employment Statistics, and job postings come from Burning Glass Technologies. Data on county-level daily COVID-19 cases come from the New York Times GitHub repository. Evidence on borrower search is ob- 9Therateistypicallylockedaboutthetimewhentheborrowersubmitsaloanapplication,butitcanhappen beforeorafter. 10All searches condition on full income, asset, and employment documentation, and require that the mortgage finances a single-unit home. If a lender represents more than one investor, we observe the most competitiveinvestoroffer. FormoredetailsseeBhuttaetal.(2021),whocomparemortgageofferratesand lockratestostudytheefficiencyofborrowersearchinthemortgagemarket. 6

tained from Google Trends. We also use public Home Mortgage Disclosure Act (HMDA) dataforgeographicmarketcharacteristics(e.g.,lenderconcentration). Lastly,weuseloanlevel data from eMBS on mortgages securitized through Fannie Mae, Freddie Mac, and GinnieMaetostudyfintechlending,andBlackKnightMcDashservicingdata(“McDash data”)tostudythejumbomarketandtocross-validateseveralofourresults. III Rates and Markups for Conforming Mortgages Inthissection,wemeasuretheevolutionofmortgageratesandmarkupsin2020. Westart, inSectionIII.A,bybreakingdownthespreadbetweenmortgageratesandbenchmark10yearTreasuryyieldsintoseveralcomponents. InSectionIII.B,weusethisdecomposition to show that one component, the primary-secondary spread, a measure of intermediary markups,largelyexplainsthebehaviorofthemortgage-Treasuryspread. InSectionIII.C, wecalculateintermediarymarkupsusinggain-on-sale,analternativeandarguablymore accurate measure. Gain-on-sale paints a very similar picture: Both methods imply large and persistent increases in markups. In Section III.D, as a reality check, we look directly atintermediaryincomereportedbylenders,whichalsoshowslargegains. Our focus in this section and in Section IV is on the largest segment of the market, conventional conforming loans (which are generally securitized via Fannie Mae or Freddie Mac). These are typically “plain-vanilla” 30-year fixed rate loans to prime borrowers making a down payment of 20 percent or more. As the core of the market, conforming ratesandquantitiesarecloselytrackedbypolicymakersandmarketparticipants. III.A Decomposing the Mortgage-Treasury Spread Define r as the “primary mortgage rate” paid by the borrower, r as the yield on a 10p 10 year Treasury note, and r as the yield on a new-production MBS into which a typical s newly originated mortgage with note rate r would be securitized. We can decompose p themortgage-Treasury spreadas − − − r r = r r + r r . (1) p 10 p s s 10 Primary-secondaryspread MBSyieldspread | {z } | {z } 7

WecanfurthermoredecomposetheMBSyieldspreadasfollows: Option Option-Adjusted − ≈ − r r r r + + . (2) s 10 dur 10 (cid:18) (cid:19) Cost Spread (OAS) DurationAdjustment | {z } Duration Adjustment reflects the fact that the MBS may have a different duration from the10-yearTreasury. MBSdurationisnotknownex-ante,sinceitdependsonprepayment behavior,soinsteaditisestimatedusingamodelthatsimulatesdifferentinterestrateand prepayment paths. Option Cost measures the valueof the borrower’s prepayment option; the borrower can prepay at any time and will tend to do so especially when rates fall (in ordertorefinanceintoanewloan). SinceMBSinvestorsareshortthisoption,theyrequire compensation in terms of a higher MBS yield. Finally, Option-Adjusted Spread (OAS) is a residual that captures various factors that affect relative pricing between Treasuries and MBS, such as liquidity differences, the risk-bearing capacity of marginal investors in the twomarkets,relativebondsupply,perceivedcreditriskdifferences,andnon-interest-rate prepaymentrisk(Boyarchenko,Fuster,andLucca,2019).11 III.A.1 Measurement While headline mortgage rates and Treasury yields are readily available, decomposing the mortgage-Treasury spread into the four components in equations (1) and (2) is less straightforward. Most important, MBS yields, durations, option costs, and OAS are not directly observed, but are obtained based on MBS pricing models featuring interest rate simulationsandprepaymentprojections;werelyonmodelsfromJ.P.MorganMarkets,as notedearlier. Twofurthercomplicationsarethat(i)MBSaretradedincouponsof50basis point (bp) increments, and a mortgage with a given note rate could be securitized into different new production coupons; and (ii) MBS investors do not receive the entire note rate, as some portion of the cash flow is diverted to pay for the agency credit guarantee (g-fees) andservicing. Weaccountforthesefactorsbycalculatinganetnoterate: − − Net Note Rate = r g s. (3) p 11FortextbookdescriptionsofMBSpricing,seeFabozzi(2016)orDavidsonandLevin(2014). Notethatthe decompositionin(2)isnotexact,sincetheOASandthe“zero-volatilityspread”(whichisthesumofOAS andoptioncost)arecalculatedasanaveragespreadrelativetoeachpointontheinterestratecurvebased onwhichfuturecashflowsarediscounted,whiletheleft-hand-sidemeasureisasimplespreadatagiven pointonthatcurve. 8

Our baseline measure of r is the 30-year FRM rate from the Freddie Mac PMMS; for the p guaranteefee gwetaketheeffectiveg-feeonnewproductionMBSfromFannieMae’s10- Qdisclosures;andthe(base)servicingfeesissetto25bp,whichisthemarketconvention forMBSissuedbythegovernment-sponsoredenterprises(GSEs)FannieMaeandFreddie Mac. We then calculate the MBS yield, duration, option cost, and OAS by interpolating valuesbetweenthetwoMBScouponsoneithersideofthenetnoterate. III.B What Happened over 2020? How Does It Compare with 2007–09? Panel A of Figure 2 traces out the evolution of the mortgage-Treasury spread in 2020 in terms of changes relative to the beginning of the year, and it decomposes these changes into the four components discussed above. The mortgage-Treasury spread increased rapidly starting in late February, peaking at about 90bp above pre-pandemic levels in late March (seen in the black dots in the figure). The mortgage-Treasury spread then graduallynormalized,buteveninAugustitremainedabout50bphigherthanatthestart oftheyear. Thespread didnotreturntopre-pandemiclevelsuntilNovember2020. The decomposition in the figure shows that this rise in the mortgage-Treasury spread ismorethanentirelyaccountedforbyasharpincreaseintheprimary-secondaryspread— that is, by a higher price of intermediation in the primary mortgage market. In contrast, the three financial market components (duration, option cost, and OAS) were actually lowerthantheirpre-pandemiclevelsovermostof2020. There were, however, some temporary effects of financial market disruptions during the flight to liquidity in March 2020. OAS increased sharply in March, reflecting an amplification of risk premia in financial markets as well as deleveraging by mortgage REITs. Theoptioncostalsoincreased,reflectingtheriseininterestratevolatility. Butthesespikes were short-lived. On March 15, the Federal Reserve resumed its MBS quantitative easing program, and on March 23 it announced that it would purchase agency MBS in the volume needed to support market functioning. These actions were associated with a rapid normalization of OAS, which then fell below pre-pandemic levels. The duration adjustment also became negative after February 2020, reflecting a drop in the model-implied durationofnewlyoriginatedmortgages(from4.7yearsinJanuaryto3.2yearsbyJuly). Forcomparison,panelBofFigure2presentsthesamedecompositionofthemortgage- 9

TreasuryspreadovertheJune2007–May2009financialcrisisperiod,whenthespreadwas similarly elevated. The figure shows that the sources of the elevated spread are very different between the two episodes. The primary-secondary spread barely moved during the financial crisis, and it is the spread on MBS in secondary financial markets, and particularly a rise in OAS, that accounts for the unusually high gap between mortgage rates andTreasuryyields,atleastuntilearly2009,whenfinancialmarketsstarttonormalize. Given these facts, our focus in this paper is on understanding the high markup in the primary mortgage market during the pandemic. But first we consider two alternative waysofmeasuringthemarkup: gain-on-saleandlenderprofitmargins. III.C Measuring Intermediation Markups: Gain-on-Sale The primary-secondary spread examined above is appealing in the context of decomposing different factors affecting mortgage rates actually paid by borrowers. It is also frequently followed by policymakers and market commentators, thanks to its apparent simplicity. However, it has a few conceptual shortcomings, as discussed in detail in Fusteretal.(2013)andFuster,Lo,andWillen(2017). Mostimportant,theprimary-secondaryspreaddoesnotaccuratelycapturehowbanks andnonbankmortgagelendersthinkaboutthemargintheyearnfromoriginatingamortgage that is securitized in the MBS market. Rather than “earning” the mortgage rate and “paying” the MBS yield over the life of the loan, originators earn most of their margin upfront, as the difference between the amount they receive from selling the loan in the secondary market and the amount paid to the borrower. The only relevant “flow” income over the life of the loan is the servicing strip (equal to 25bp per year in case of GSE loans, and 44bp for FHA loans). This can be earned by either the mortgage originator or another firm that acquires the right (and obligation) to service the loan. The value of the mortgageservicingright(MSR)isusuallycapitalizedandincorporatedinthecalculation ofthelendermargin,whichisoftenreferred toasthe“gain-on-sale.” Formally,consideramortgagewithinitialbalanceS . Assumethattimeiscontinuous 0 and the loan has constant prepayment hazard λ . Using the notation from above, the p valueoftheloanis ∞ V = e −rstS r − g+λ dt. t p p Z 0 (cid:0) (cid:1) 10

AssumingthatthelendercanselltheloanintheMBSmarketforV,thelender’sprofit,or “gain-on-sale,” on the loan is V − S . With a constant prepayment hazard, S = S e −(λp )t, 0 t 0 andgainonsalecanbeexpressed as V − S r − r − g 0 p s = . (4) S r +λ 0 s p The logic here is exactly as described above: The higher the prepayment hazard λ , the p − > lessprofitabletheloan(assuming r g r ). p s Equation (4) omits two components that are important in actual gain-on-sale calculations. First,lenderstypicallyvaluetheservicingflowsseparatelyfromtherestofthecash − flow. Thenetservicingincomeflowiss(1 c ),wherec isthecostofservicingasashare s s of servicing income. Second, the lender typically collects points and fees at origination, which we denote as F, so the initial cash outlay is not S but S − F.12 Taking those two 0 0 factorsintoaccount,wecanwritegain-on-saleas r − r − g − s s(1 − c ) F p s s π = + + . (5) r +λ r +λ S s p s p 0 Gain-on-Sale |{z} SecondaryMarket MortgageServicing Pointsand | Inc{ozme } |Righ{t(zMSR}) |F{ezes} − (1 c )/(r +λ ) is known as the servicing multiple. Similarly, 1/(r +λ ) for the first s s p s p term can be seen as a valuation multiple used to value the “interest-only (IO)” strip providedbythenumerator(payingaconstantfractionoftheunpaidprincipalbalanceofthe loanwhiletheloanisactiveandnothingwhentheloanprepays). Our approach to measuring gain-on-sale follows Fuster,Lo,andWillen (2017). To compute the value of secondary market income, we assume that the lender sells the loan intheTBAmarket.13 Specifically,westartwithr − g − sandinterpolatebetweenthetwo p − − nearestMBScoupons(forexample,ifr g s = 3.25,wetaketheequal-weightedaverp agepriceofa3.0couponanda3.5coupon). TovaluetheMSR,weusevaluationmultiples provided by SitusAMC, an independent valuation service company. These multiples are based on transaction values of brokered bulk MSR deals, surveys of market participants, 12Fisnetofloan-levelpriceadjustments(LLPAs)paidtotheGSEwhensellingtheloan. 13“TBA”standsfor“tobeannounced,”whichistheMBSmarketwherenewagencymortgagesaretypically forward-sold. SeeVickeryandWright(2013)foradditionalinstitutionaldetails. 11

and a pricing model, and capture the net value of servicing, that is, the value of the incomeflowminusexpectedcosts.14 Themultiplesareprovidedtousatmonthlyfrequency at the coupon-by-loan type (GSE versus FHA) level.15 All components are expressed per $100loanamount,asisstandardintheindustry. Figure 3 compares our estimates for intermediation markups based on the simple primary-secondary spread measure from Section III.A.1 and the gain-on-sale methodology from this section. Panel A shows the primary-secondary spread based on three different mortgage rate series, the PMMS survey rate (our baseline), the Mortgage Bankers Association (MBA) weekly application survey, and the Optimal Blue Insight data.16 Despite some differences across the three series, the primary-secondary spread increases significantlyattheonsetofthepandemicregardlessofwhichmortgagerateisused.17 Panel B of Figure 3 shows time series of our estimates of gain-on-sale in the conformingmarket,basedonthesamethreemortgagerateseries. Inallcases,gain-on-salejumps byabout250bpfrommid-FebruarytoearlyApril2020,andthenstayshighovertherest of the year. Based on the PMMS and MBA series, it is on average roughly 200 bp higher than in the second half of 2019; with the Insight series, the increase is around 175 bp. If wecomparepanelsAandBofFigure3,weseesimilarpatterns,withthemaindifference beingthattheprimary-secondaryspread(panelA)revertsmoretowarditspre-pandemic levelattheendof2020thangain-on-sale(panel B). Gain-on-sale shows just how profitable mortgage lending was in the last three quartersof2020. Originationsovertheperiodaveragedmorethan$1trillionaquarter. Multiplyingourquarterlygain-on-salemeasuresbyoriginationvolumesimpliestotalindustry 14Fusteretal. (2013) calculate a measure of π called “OPUC,” where they explicitly consider the choice of coupon into which a loan is securitized but for the most part assume constant multiples over time. In InternetAppendixA.2,weshowthatresultsusingOPUCarequalitativelysimilar. 15InternetAppendixFigureA.1showstheevolutionofthesemultiplesforconventionalconformingandFHA segments. Multiples for given fixed coupons decreased over time, reflecting the drop in market interest rates. However,theinterpolatedmultipleatthetypicalcouponratefornewloansdecreasedonlymodestly beforefullyrevertingoverthecourseof2020. 16The PMMS and MBA surveys ask lenders for both rates and discount rates offered. Since the discount points change each week, and there is a rate-point trade-off, the published survey rates are not directly comparableweektoweek. Toaddressthis,weusethepoint-ratetrade-offsestimatedintheOptimalBlue offer (Insight) data for 2017 through 2020 to adjust survey rates to a “constant discount points” rate. For previousyearswhenOptimalBluedataarenotavailable,weestimatethesetrade-offsusingMBSprices. 17InternetAppendixFigureA.2showsthatthepatternsarealsoqualitativelysimilarwhenweusetheCurrent Coupon(CC)tomeasureforMBSyields. TheCCyieldisnotasreliableinthistimeperiod,however,since everyMBScoupontradedfaraboveparin2020. SeeFusteretal.(2013)forarelateddiscussion. 12

gain-on-sale of about $162 billion, or $54 billion a quarter. Had gain-on-sale remained at its pre-pandemic level of 2.5 percent, the industry would have earned only $82 billion overthesethreequarters—thatis,halfasmuchasitdid. III.D Lender Profit Margins Directevidenceonintermediaries’loanproductionprofitsalsoindicatesrecord-highmargins. We obtain data from the Mortgage Bankers Association Quarterly Performance Report, which has been collecting quarterly data from roughly 340 mortgage banks on revenues and expenses since 2008:Q3.18 Panel A of Figure 4 plots the time series of average total net production income (in basis points of loan volume) reported by the firms in the sample—meaning the margin between loan production revenues and expenses. We note thatmortgagebanksdonotnecessarilycapturetheentireintermediationmarkup,asthey oftenactascorrespondentlendersforlargerbanks(“aggregators”)thatthenselltheloans inthesecondarymarketandalsocapturepartofthemarkup. Income in Q2 through Q4 of 2020 is extraordinary compared with all other quarters in the data. Net production income is about 60 bp in 2020:Q1, already at the higher end of incomes over the previous few years, but it was more than three times as high, 203 bp, in Q3. Furthermore, this higher net margin was earned on loan production that was almost twice as high. In panel B, we multiply the net production income by the average reported origination volume per firm, which was $1.335 billion in Q3 and $1.472 billion in Q4, versus $728 million in Q1. The resulting total net production income therefore skyrocketedfrombelow$5millioninQ1toanaverageof$21.4millionovertheremaining quartersof2020,withapeak ofmorethan$27millioninQ3. These data provide further confirmation that the price of intermediation in the mortgage market was historically high in 2020. They also show that high measured gain-onsale at least partly translated into high lender profit margins rather than being absorbed by expenses not accounted for in our gain-on-sale calculations.19 We now study the role ofdifferenteconomicfactorsinaccountingforthesestrikinglyhighmarkups. 18Note that the sample of reporting firms changes over time, and we have access to only aggregate series, not firm-level series. Also, these statistics are not specific to the conventional conforming segment; they combinealloriginations. 19Forinstance,Fusteretal.(2013)discusspipelinehedgingcosts,whichareproportionaltoimpliedinterest volatility. However,whilethisvolatilityspikedearlyinthecrisis,ithassincebeenatorbelowitslevelof JanuaryandFebruary2020. 13

IV Capacity Constraints We first examine whether industry capacity constraints can explain the high intermediationmarkupsdocumentedabove. Individualloanofficersandunderwritersarelimitedin how quickly they can process mortgage applications, and lenders face significant adjustment costs in hiring and training new workers, particularly in the short run. Fusteretal. (2017) find evidence that these constraints lead to an increase in mortgage markups duringperiodsofpeakdemand—inotherwords,givenlimitedcapacity,lenders“steponthe brakes”byincreasingthemarginschargedtoborrowers.20 IV.A Is 2020 Consistent With Historical Patterns? Figure5studieswhetherthehistoricalrelationshipbetweenintermediationmarkupsand capacity utilization can account for the high markups in 2020. The figure plots mortgage demand against two measures of intermediation markups we have used already: the primary-secondary spread and the gain-on-sale (π). Our preferred proxy for demand is the difference between the weighted average coupon (WAC) on the stock of mortgages and the 10-year Treasury yield. This measure is highly correlated with applications for (refinance) loans and is arguably exogenous to current mortgage supply shocks, because itdoesnotrelyoncurrentmortgagerates.21 WealsousetheMBAmortgageapplications indexasanalternativedirectmeasureofdemand. The figure shows that the price of intermediation during the pandemic significantly exceeds what would be predicted simply based on the level of demand. Red squares in Figure 5 are from 2020, with the number indicating the month of the observation. Blue circlesarefrom2012through2019. Thereisaclearpositivehistoricalrelationshipbetween markups and mortgage demand. But from about April 2020 onward, markups are well abovethelineofbestfit,regardlessofwhichdemandormarkupmeasureisused. 20Consistentwiththepresenceofcapacityconstraints,Fusteretal.(2017)showthatloanprocessingtimesare strongly positively correlated with mortgage application volumes. SharpeandSherlund (2016) also find evidenceofcapacityconstraints, showingthatrefinancingwavesinducesubstitutionawayfromcomplex mortgagesthataremorelaborintensivetoprocessandunderwrite. 21In contrast, using the difference between WAC and the current mortgage rate may produce bias in the estimated markup relationship, because mortgage rates reflect the intersection of demand and supply − (Fusteretal.,2017).TheWAC 10-yearTreasuryspreadishighlycorrelatedwithanestimateofthepercentageofactivemortgageborrowerswhoarerefinancecandidates(fromLambie-Hanson,2020),measuredin eithermonthlylevels(ρ=0.77)ormonthlychanges(ρ=0.70)fromJanuary2008throughOctober2020. 14

Table1quantifiesthe“excess”markupbyregressingtheprimary-secondaryspreador gain-on-saleoneitherofthemortgagedemandproxiesandtimedummiescorresponding to different phases of the pandemic. In all cases, the pandemic dummies indicate large, statistically significant excess markups from March 2020 onward. The historical relation between markups and demand can account for only about 20 to 40 percent of the rise in intermediationmarkups,dependingonthespecification.22,23 Column 1 shows that the primary-secondary spread is 77 bp higher than expected in March through May, and 86 bp higher in June through August (that is, it increased an additional 9 bp). By comparison, the one-standard-deviation residual from this regression excluding the COVID-19 period is only 8.9 bp (last row of table). Thus, one could consider the COVID-19 period roughly a “9 sigma” event that persists for well over six months. Results are similar in column 2, although the regression fit is slightly weaker. The excess intermediation markup is also highly significant when we use gain-on-sale as the dependent variable (columns 3 and 4). Gain-on-sale is $1.05 to $1.30 higher than expected overMarchthroughMay,risingto$1.40to$1.60overJunethroughAugust.24 IV.B Operational Constraints in Expanding Capacity We now document several pieces of evidence suggesting that operational issues related to the pandemic made it more difficult for lenders to expand capacity in 2020. This providesaplausibleexplanationforwhythemortgagesupplycurvewasunusuallyinelastic relativetopreviouslendingbooms. 22On the high end of this range, the coefficient in column 4 would project a maximum increase in gain-onsaleof109bp,giventhattheMBAapplicationsindexincreasedfromanaverageof449inJanuary2020to amaximumof1002. Thiscompareswithanactualincreaseingain-on-saleof279bp. 23The pandemic dummies from March through August are comparatively smaller for gain-on-sale than for theprimary-secondaryspread,consideringthatthelatterisaflowmeasurewhiletheformermeasuresthe total gain earned by intermediaries. This reflects (i) a flattening of the relationship between gain-on-sale andtheprimary-secondaryspreadduringthepandemic,and(ii)thefactthatgain-on-saleisbelowtheline ofbestfitatthestartof2020,whiletheprimary-secondaryspreadisclosetoitsfittedvalue. 24Ourestimatescouldoverstategain-on-saleiftheSitusAMCservicingmultiplesfailtoproperlyreflectthe risks for servicers of elevated forbearance and non-payment. But in Section V.A we show there is no evidenceofarisinginterestratepremiumforhigh-riskconventionalconformingloansduringthepandemic, which speaks against this hypothesis. We also have no reason to doubt that the SitusAMC valuations internalize the risks of forbearance; in fact, given secondary market illiquidity for MSRs, transactions may understategoing-concernvaluesiftheymainlyreflectforcedsalesatfire-saleprices. Finally,evenifweset theservicingmultipletozero—averyextremeassumption—thiswouldreducegain-on-salebyonlyabout $1,stillnotbringingitfullyinlinewithhistoricalpatterns. 15

IV.B.1 LaborMarketFrictions Anecdotalevidencesuggeststheremote-workenvironmentcreatedsignificantchallenges inhiring,training,andonboardingnewloanofficers,processors,andothermortgageemployees. One leading industry participant told us that it has been difficult to train new employees and to monitor their work, and that as a result lenders prioritized hiring experienced professionals (generally from competitors) who were well known and trusted and did not require much training or supervision. Consistent with this message, an executive at large nonbank lender Mr. Cooper explained their limited expansion in capacity as follows: “It reflects the fact that we add capacity at a deliberate pace with an eye on the long term. And additionally, when the crisis hit and we shifted to work-from-home status, that slowed ourhiringandouronboarding”(Ivey,2020). Thediversionofstaffandmanagerialattention toprocessthewaveofforbearanceapplicationsappearstohavebeenanadditionaldrain ontheeffectivelaborsupply,atleastearlyinthepandemic(BerryandKline,2020). Disruptions to the licensing of mortgage loan officers also held back growth in labor supply. Nonbank loan officers must be licensed through the Nationwide Multistate Licensing System (NMLS), a process including background checks, fingerprinting, an examination, and ongoing education.25 During the early months of the pandemic, half of the fingerprinting locations were shuttered, and about 10 percent remained closed as of December 2020. Testing sites also closed in the early months, although the NMLS began offeringaremotetestingoptioninSeptember. Thesefactorslimitedtheinflowofdenovo loan officers and also affected employees switching from banks to nonbanks or changing theirstateofemployment.26 Thisqualitativeevidencesuggeststhatthesupplyoflaborwasrelativelyinelasticduring the pandemic, which would imply unusually slow hiring compared with the level of labor demand and also a sharp increase in the price of labor. We now present evidence supportingthesepredictions. 25Anewlicenseisrequirednotjustfordenovoloanofficers,butalsoforexperiencedloanofficerstransitioningfrombankstononbanksormovingacrossstates. Inthelattertwocases,theloanofficerispermittedto temporarilyoperatefor120daysbeforeobtaininganewlicense(MortgageBankersAssociation,2019). We weretold,however,thatinsomecaseslicensingdelaysexceededthis120daygraceperiod. 26In an email communication, one industry participant told us, “Testing centers closed at times and have had limitedavailability. ThishaspreventednewLOs[loanofficers]fromjoiningtheindustryrightaway. Wehaveseen delaysofanywherefromafewweekstoafewmonthsdependingontherestrictionsinthearea.” 16

PanelAofFigure6showsthatlendersaddednearly50,000mortgageloanofficersand underwriters (MLOs) after the emergence of COVID-19 in March 2020. This took MLO employment to a post-financial crisis high of 360,000, although employment remained 120,000 MLOs short of its all-time peak in 2004. Data from Burning Glass in panel A also show a sharp rise in MLO job postings during the COVID-19 period.27 Postings fluctuated around 1 vacancy per 100 employees prior to COVID-19, but jumped to more than2per100duringthepandemic. To test whether MLO hiring was low relative to the number of postings, we estimate thefollowingregression: − log(MLO t+1 ) log(MLO t ) = α+β 1 p t +β 2 p t−1 +β 2 p t−2 +ε t , wherethedependentvariableisthelogchangeinMLOemploymentbetweenmid-month t and mid-month t+1 (from the BLS Current Employment Statistics, Establishment Survey) and p are monthly MLO postings per employee from Burning Glass. The regression is estimated on data from March 2012 through February 2020. Coefficients are shown in InternetAppendixTableA.1andimplyacorrelationofabout0.5betweenactualemploymentgrowthandemploymentgrowthpredictedjustbycurrentandlaggedjobpostings. PanelBofFigure6presentspredictionsfromthisregressionmodelthatillustratehow anomalous the pandemic period was for the hiring-vacancies relationship. Dots labeled “COVID” are out-of-sample predictions for March 2020 through December 2020. Actual employment growth fell short of the predicted value in all 10 months. Overall, we estimate that monthly employment growth averaged about 1.5 percentage points less than we would expect. In panel A, the line labeled “Counterfactual Employment” quantifies the cumulative effect of this hiring gap. Actual MLO employment at the end of 2020 was 60,000 lower than what would be predicted based on the historical relationship between postingsandemploymentgrowth. 27Burning Glass Technologies is an employment analytics and software company that crawls more than 40,000 online job boards and company websites daily, scraping and cleaning information on job postings. We focus on job postings with the O*NET job code 13-2072, which includes professionals who evaluate, authorize,orrecommendapprovalofcommercial,realestate,orcreditloans,includingmortgageloanofficers and agents, collection analysts, loan servicing officers, loan underwriters, and payday loan officers. We further capture ”mortgage loan officers” as postings in this category that have the words ”mortgage,” ”mtg,””home,”or”residential”inthejobtitle. 17

Turningtothecostoflabor,itisdifficulttoobservethecompensationpackagesoffered to loan originators, which often involve a combination of a signing bonus, fixed salary, andloanvolumeincentives. Nevertheless,availabledatafromtheMortgageBankersAssociation, analyzed in InternetAppendix B.2, are consistent with a significant increase in labor costs during the pandemic. Labor costs per dollar of loan volume typically fall significantlyduringlendingboomsduetoscaleeconomies;butweshowlaborcostsdeclined only slightly during the 2020 boom. A conservative estimate is that personnel expenses scaled by loan volume were unusually high by at least 25 bp by 2020:Q4, equivalent to morethan10percentofaveragepersonnel costsmeasuredoverthepreviousfiveyears. IV.B.2 OperationalIssuesinMortgageOrigination The pandemic also made the process of originatinga mortgage more complex and uncertain. Obtainingdocumentationofborroweremploymentandincomewasoftenchallengingbecausemanybusinesseswerephysicallyclosed,andthehighrateofjoblossoftenrequiredcheckingemploymentstatusmultipletimesbeforeclosing(BerryandKline,2020). County recorder offices were closed or operating on limited schedules, making it hard to complete title searches and causing delays in recording mortgages, which in turn raised concernsaboutthepotentialforfraud(Hughes,2020). Otherstepsintheprocess,suchas property appraisals and obtaining notarized signings of closing documents, were more cumbersome due to social distancing. For example, Michael Fratantoni of the Mortgage Bankers Association noted that “there are steps in the process that are still physical, like the actual mortgage note. If it’s a paper note, someone still has to show up at the office and collect thoseandsenditontothenextstepintheprocess.” (BerryandKline,2020) IV.C Technology and Capacity Constraints The mortgage industry has undergone significant technological change in recent years. Didtechnologyplayaroleinexpandinglendingcapacityandmitigatingoperationaland labor market frictions during the pandemic? For instance, information technology can enabletheoriginationprocesstobelesslaborintensiveandmoreautomatedandthereby more scalable, consistent with Fusteretal.’s (2019) finding that technology-based mortgage lenders adjust processing capacity more elastically.28 Technology-based lenders are 28Rocket Mortgage CEO Jay Farner emphasized this scalability in the context of the 2020 refinancing wave inthefirm’s2020:Q3earningscall,asreportedbyindustrynewsletterHousingWire: “FarneralsosaidRocket 18

alsolikelytorelyondigitizedrecordsratherthan paperdocumentsand beabletotransitiontheirworkforcemoreeasilytoremotework. Itisalsopossiblethatborrowersplaced more value on a high-quality online platform during the pandemic, given the challenges offace-to-face interactionsandthetransferofphysicaldocuments.29 We use loan-level data from eMBS and the classification of “fintech” lenders from Fusteretal. (2019) to assess whether technology-based lenders expanded lending more rapidly during the pandemic.30 We investigate two questions. First, did fintech lenders increase their overall market share? Second, did they gain market share particularly for mortgages that are complex and time intensive to underwrite? This second test is motivatedbySharpeandSherlund(2016),whofindthatcapacityconstraintsinducelendersto ration credit by prioritizing simpler and less labor-intensive mortgage applications. Following SharpeandSherlund, we use low-credit-scores and purchase mortgages as two measuresofthesemorecomplex-to-underwriteloans. We estimate linear probability models in which the dependent variable is a dummy equal to 1 if the originator is a fintech lender in the Fusteretal. (2019) classification. The right-hand-side variable of interest is a pandemic dummy equal to 1 for mortgage applications underwritten and processed during the pandemic, which, given the time to close a loan, we proxy by an origination date after March 2020. The model is estimated using loan-leveleMBSdataonagencymortgageoriginations,combiningdatafromFannieMae, FreddieMac,andGinnieMae. ThesampleperiodisJanuary2019throughDecember2020. LikeFusteretal.(2019),ourmainspecificationsexcludebankloans,toisolatedifferences between fintech lenders and other nonbank lenders with a similar funding model and commonregulatoryenvironment. ResultsinTable2indicateanincreaseinthefintechmarketshareduringthepandemic, wasabletoscalemoreaggressivelythancompetitorsduetoitstechstackandbusinessmodel,whichhesaidisfarmore efficientbecauseitdoesn’trequiredisproportionatelyhighheadcounts.” (Kleimann,2020). 29Forexample,RocketMortgageCEOFarnerclaimedduringthecompany’sthirdquarterearningscallthat “[d]emandforacompletelydigitalexperiencehasneverbeenstrongerandRocketisdelivering.” (Kleimann,2020). 30TheFusteretal.(2019)classificationidentifiessixmortgagelendersthatofferedafullyonlineapplication asoftheendof2016: Rocket(QuickenLoans),BetterMortgage,GuaranteedRate,LoanDepot,Movement Mortgage,andSupremeLending(EverettFinancial).Onlinelendingplatformsweremuchmoreubiquitous in2020thanin2016,butthissetoflendershasremainednearthetechnologyfrontier. Relyingonthislist also limits the influence of look-back biases or subjective judgments about which lenders are technology leaderstoday,becausetheFusteretal.(2019)classificationwassetwellbeforethestartofthepandemic. It does,however,meanwemisssometechnology-basedlendersthathaveemergedmorerecently. 19

concentrated among complex loans. Columns 1 through 3 focus on purchase mortgages, whicharetypicallymorecomplicatedandlaborintensivetounderwritethanrefinancings. Incolumn1,whichincludesnocontrols(thatis,asimpledifferenceinmeans)thefintech share was 2.8 percentage points higher during the pandemic than in the pre-pandemic period (defined as loans closed from January 2019 through March 2020). Column 2 controls for state fixed effects (the most granular geographic information in these data) as wellasloanandborrowercharacteristicssuchasquadraticfunctionsofcreditscore,debtto-income, and loan-to-value (see table notes for full list of controls). The coefficient is somewhat smaller, at 2.1 percentage points, but remains statistically and economically significant. Quantitatively,theincreaseisaboutone-fifthofthesamplemeanof11.1. Col- × < umn 3 includes a pandemic (credit score 680) interaction term, which indicates that the rise in fintech share is significantly larger for low-credit-score mortgages. The total effectfortheseloansisabout4.1percentagepoints. The next three columns focus on refinancings, the sector where fintech lenders historically have specialized. Interestingly, the overall fintech share of refinancings was little changed during the pandemic, although again the fintech share did rise for low-creditscore borrowers, as shown by the positive interaction term in column 6. The final three columnsarebasedonapooledsampleincludingbothpurchasemortgagesandrefinances. The increase in the fintech lenders’ overall share reflects their gains in the purchase marketaswellasacompositional shiftinlendingtowardrefinancings. In InternetAppendix C, we trace out the timing of these changes by replacing the pandemic dummy with monthly time dummies and plotting the resulting coefficients. Thepurchase-mortgagegraphisparticularlystriking: Thefintechmarketsharewasvery stable at about 7 percent prior to the pandemic, but then it jumped sharply by about 3 percentage points as the pandemic took hold, before declining later in 2020 as conditionsstartedtonormalize.31 Asanadditionalrobustnesstest,were-estimatetheanalysis in Table 2 while retaining bank lenders in the sample. The results (also presented in InternetAppendixC)aresimilartoourmainestimates. These results support the view that technology and automation helped the mortgage industry manage the capacity constraints and operational challenges associated with the 31Weinvestigatethecompositionofthisincreaseinthefintechshare,findingthatitisbroadlybasedacross thesetofsixfirmsclassifiedastechnology-basedlenders. 20

pandemic. In particular, the fintech market share increased most markedly for purchase mortgages and loans to low-credit-score borrowers, exactly those loans which are typicallyde-prioritizedduringperiodswhencapacityconstraintsbind(SharpeandSherlund, 2016). V Alternative Explanations: Risk, Health and Competition We now consider several other plausible drivers of the rising price of intermediation for conforming mortgages, related to default and forbearance risk, the direct health and economic effects of the virus, and lender market power. The evidence we present below suggests that none of these factors played an important role in accounting for the high markupsin2020. V.A Forbearance and Default Risk The pandemic led to a surge in mortgage forbearance and non-payment, reflecting high unemployment and a CARES Act provision requiring servicers to provide up to a year of forbearance on federally backed mortgages.32 Non-payment is costly for mortgage lenders and servicers, even in the conforming market where loans are securitized and carryacreditguarantee. First,thereisliquidityriskbecausetheservicerisrequiredtoadvancemortgagepayments,taxes,andinsuranceeveniftheborrowerstopspaying. These advances will eventually be reimbursed, but funding them in the interim may be expensiveorinfeasible.33 Second,thereis“pipeline”riskthattheloanbecomesnonperforming just after origination, before it can be sold.34 Third, servicing a delinquent loan is much morelaborintensiveandcostly. 32AtthepeakinJune2020,8.6percentofallmortgageswereinforbearance(source: MortgageBankersAssociation). Thetotalmortgagenon-paymentratemorethandoubledduringtheearlypartofthepandemic: from3.2percentinJanuary2020to7.8percentbyMay2020(BlackKnightFinancialServices,2020). 33Thiswasacauseofgreatconcernintheindustryearlyinthepandemic.OnApril21,2020,itwasannounced that servicing advances for conforming mortgages securitized by Fannie Mae and Freddie Mac would be cappedatfourmonthsofprincipalandinterest(FederalHousingFinanceAgency,2020).Thiscapdoesnot applytotaxes,insurance,andotherpayments. 34Suchloanstypicallywouldbesoldprivatelyatasignificantdiscount. TheGSEsandFHAtooksomesteps to limit pipeline risk during the pandemic, but these steps only partially protected lenders (for example, the GSEs agreed in April 2020 to purchase loans in forbearance, but only at a 500–700bp discount: See https://bankingjournal.aba.com/2020/04/fhfa-to-purchase-qualified-loans-in-forbearance.) 21

We use the cross section of mortgage rates to study whether this heightened default risk led to a risk premium that increased mortgage rates for borrowers. Our approach uses the fact that the rate of non-payment was much higher for low-credit-score borrowers (see InternetAppendix E for evidence based on McDash servicing data). If rising default risk is being priced into conforming mortgage rates, we should see an increase in theinterestratespreadbetweenlow-versushigh-scoreconformingmortgages. WefirstestimatethisdefaultriskspreadusingOptimalBlueInsightdataonmortgage offers from lenders. Our data include offer rates for otherwise identical mortgages with a FICO score of both 680 and 750. We measure the rate premium for a FICO 680 loan by estimatingtheregression: × rate = α +β (FICO = 680)+ε , (6) imt mt t i imt whererate istheinterestrateonanofferiinCBSAmduringweekt, FICO = 680is imt i × a dummy for a FICO score of 680 rather than 750, and α are CBSA week fixed effects. mt Other loan characteristics are held fixed (e.g., DTI=36, LTV=80 and balance of $300K). β t thentracesouttheevolutionoftheFICO680-750interestratespread. WealsocomputeananalogousspreadusingOptimalBlueratelocksdata,estimating: rate = α +δ +β × FICObin +Γ X +ε , (7) ilmt mt lt t i ilmt ilmt whererate istheinterestrateonlockiissuedbylenderlinCBSAmduringweekt; ilmt × × FICObin isasetoffiveFICObindummies;α andδ areCBSA weekandlender month i mt lt fixed effects; and X is a set of loan and borrower controls including log loan amount, ilmt DTIandDTIsquared,anddummiesforthelockperiodandpropertytype. Forloanswith discountpointsorcredits,weconvertthelockratetoazero-pointsequivalentinterestrate usingthemarketrate-pointtrade-offmeasuredeachweekinOptimalBlueInsight.35 35Borrowers pay discount points to buy down the interest rate on their mortgage. Conversely, mortgages withnetcreditshavehigherinterestrates. Theshapeofthetrade-offbetweenratesandnetpointsmoves overtimeandhasbecomeflattersincetheonsetofthepandemic. OurOptimalBlueInsightdatainclude daily offered interest rates for loans with different numbers of net points (+2, +1, 0, –1, and –2). We use this pricing grid to adjust the interest rate on each loan to an equivalent zero-point loan. We then use this adjusted rate as our dependent variable. An alternative approach is to include additional controls interactingpointsandcreditswithtimedummies;thisgeneratessimilarresults. 22

Estimates of β from these two models are reported in the top two panels of Figure 7. t Thereisnoevidenceofapersistentincreaseintheinterest-ratespreadonlow-FICOmortgages, for either mortgage offers or rate locks. The interest rate spread does spike temporarily in March 2020, associated with the onset of the pandemic and dislocation in the MBS market. But by April, the interest rate spread returns to a narrow band around or evenslightlybelowthepre-pandemiclevelsofabout40bp. Dataonquantitiespaintasimilarpicture. PanelCofFigure7plotsthenumberofconforming lenders offering mortgages to borrowers with different FICO scores, calculated byaveragingthenumberofrateoffersintheOptimalBlueInsightdataacrossthe20Case- Shillermetropolitanstatisticalareas(MSAs). Thenumberoflendersdropstemporarilyin March, and more so for lower FICO loans, but it is subsequently quite stable, and there is no evidence of significant rationing to low-credit-score borrowers. Panel D plots the percentage of conforming rate locks below two FICO thresholds (680 and 640). For purchase mortgages (the two solid lines), there is no evidence of a drop in the percentage of low-FICOratelocks. Thereisadropforrefinances,butthisistypicalduringarefinancing boom because high-credit-score borrowers have a greater propensity to refinance when ratesfall(e.g.,Keysetal.,2016) This evidencesuggests that heightened default risk isnot an important reason for the high markups on prime conforming mortgages during the pandemic. Default risk was more important for rates in the high-risk FHA market, however, as we show in Section VI. V.B Macroeconomic and Health Shocks Although we find little role for individual default risk, perhaps intermediation markups incorporate a more general premium due to the macroeconomic and health shocks associated with the virus. In such a case, we might nevertheless expect to see heterogeneity, but primarily across locations, depending on how severely the locality is affected by the spreadofthepandemicanddropineconomicactivity. In Table 3, we examine variation in locked rates across the 100 most populous metro areas. We regress loan-level interest rates (adjusted for points) from 1.1 million conventionalconformingloanslockedfromNovember2019throughAugust2020onmetroarea 23

fixed effects, loan characteristics (e.g., loan-to-value ratio, credit score) interacted with weeklocked,anddifferentproxiesforhowseverelyametroareawasaffectedbyCOVID- 19,intermsofboththeeconomicandhealthimpacts. The first variable we try is COVID-19 cases per 1,000 metro area residents in the preceding calendar month. Since this variable equals 0 everywhere prior to February 2020, the preceding months provide our “pre-period.” In fact, higher local case numbers are associated with slightly lower mortgage interest rates (column 1). The effect is extremely small, however—a one-standard-deviation rise in case rate is associated with only a 0.3 bpdropinmortgagerates. Resultsaresimilarifweinsteaduseadummyformetroareas inthetopquartileofcasespercapita(column2). Next we study the economic shock. Job losses, measured as the standardized yearover-yearchangeinthelocalunemploymentrate,arepositivelycorrelatedwithmortgage rates,althoughagaintheeffectisnoteconomicallymeaningful. Aone-standard-deviation increase in the unemployment rate is associated with a 1.2 bp increase in mortgage rates. The aggregate increase in unemployment (from 3.5 percent to a peak of 14.8 percent nationally)wouldcorrespondto2.7standarddeviations,orarateimpactofabout3bp.36 V.C Competition Table3alsoexamineswhetherimperfectcompetitionmayhelpexplaintheriseinmarkups, motivatedbyScharfsteinandSunderam(2016),whofindlowermortgageratepass-through in concentrated markets. Using two measures of market concentration interacted with a COVID-19 dummy (= 1 from March 14, 2020, onward), we find no evidence that concentration reduced pass-through of lower rates during this episode. In column 4, the concentration measure is the share of mortgages originated in 2019 by the metro area’s top four lenders, while in column 5 it is the local Herfindahl-Hirschman Index (HHI). Both measures are derived from 2019 HMDA data and standardized to a mean of zero and a standard deviation of one. Neither interaction term is statistically significant, and theconfidenceboundsexcludeanyeconomicallysignificanteffect.37 36Our sample in Table 3 runs through August 2020. We choose to focus on the early part of the pandemic, when the primary-secondary spread was at its largest and the unemployment rate was high, to gauge an upperboundontheimpactoftheeconomicshock. Theunemploymenteffectisevensmallerifweextend thesamplethroughtheendof2020. 37Asshownincolumn6, theresultsofTable3remainsimilarifweconsideramultivariatespecificationincludinginteractionsforthevirusspread,unemployment,andconcentrationtogether.InInternetAppendix 24

We also consider the possibility that the pandemic itself reduced competition and thereby led to higher markups. Detailed updated measures of local concentration are not yet available, but national data provide no support for this hypothesis—in fact, the market share of the largest lenders if anything decreased in Q2 and Q3 of 2020, as we documentinInternetAppendixFigureA.11basedondatafromInsideMortgageFinance. Inshort,theriseinmortgagespreadsduringthepandemicwasanationalphenomenon, and is not significantly related to the degree of virus spread, the depth of the economic crisis, or local market concentration. Consistent with these findings, InternetAppendix Figure A.12 shows, based on Optimal Blue data, that although mortgage rates do vary across metro areas, the differences are generally small, and from early 2019 through late 2020,therankorderingofmetroareasisquitepersistent. V.D Shopping Even if there are many lenders, each originator may enjoy market power if borrowers do not search extensively (e.g., Wolinsky, 1986). It is also possible that borrowers search less actively during refinancing booms, because they do not need to scour the marketplace to beat their current rate, which could increase lenders’ bargaining power (see Bhuttaetal. 2021 for some direct evidence on this point). However, mortgage search activity appears to have been extremely high during the pandemic. Google Trends data in panel A of Figure 8 indicate that web searches for refinance-related terms spiked in March and remained elevated thereafter. As shown in panel B of Figure 8, borrower search activity was also unusually high relative to what would be predicted based on the level of refinanceincentives,measuredbytheWAC-10-yearspread asinourearlieranalysis. Also speaking against a “low search intensity” hypothesis, servicer retention—the shareofborrowersrefinancingthroughtheirexistingservicer—decreasedduringthepandemic. In2020:Q3,only18percentofborrowerswhorefinancedremainedwiththeirpreviousservicer,thelowestrateforatleast15years(BlackKnightFinancialServices,2020). TableA.4,wealsoestimatethesamemodelsusinganalternativedatasetofconventionalconformingloans, theMcDashdata. Againwefindlittlerelationshipbetweeninterestratesandeithervirusspreadormarket concentration,andtheeffectofunemploymentisalsoinsignificant(withanegativepointestimate). 25

VI Credit Supply in Riskier Market Segments So far our analysis has focused on prime conforming mortgages. We now turn to other segments—theFHAmarketandthejumbomarket—thatpresentadditionalriskandcomplexityforintermediaries. Studyingtheseriskiersegmentsandcomparingthemwiththe conformingmarketshedslightontwoquestions: (1)Didthespikeinforbearanceanddefaultriskreducecreditsupplyfortheriskiestborrowersandthosenoteligibleforgovernmentguarantees;and(2)didFederalReservepolicyinterventionthroughtheresumption ofMBSquantitativeeasingsupportmortgagecreditsupply? VI.A Default Risk and Credit Guarantees VI.A.1 TheFHAMarket We previously showed that the rise in forbearance and non-payment risk had little effect on prime conforming mortgage rates. But now we revisit this question for the FHA market, where mortgage intermediaries are significantly more exposed to default risk. This is because FHA borrowers default at much higher rates, and because institutional featuresoftheFHAmarketshiftmorerisktothelenderandservicer(Kimetal.,2018).38 For instance,servicersareobligatedtoadvancepaymentsonpast-duemortgagesuntiltermination or modification, and they face significant delays before being reimbursed.39 The liquidityriskofservicingadvancesisparticularlyacuteforthenonbankmortgagecompaniesthatnowdominateFHAlending,becausethesenonbanksarefinancedbyshort-term debtandcannotissueinsureddepositsoraccessgovernmentliquiditybackstops.40 First we estimate cross-sectional differences in interest rates between FHA borrowers with lower versus higher default risk. We proceed using the same methodology as in Section V.A, first analyzing interest rates on mortgage offers and rate locks, and then 38FHAborrowersaretypicallylower-incomeandoftenfirst-timehomebuyers. InJune2020,15.7percentof FHAloanswere60ormoredayspastdue,comparedwithonly6.7percentofconventionalloans. (Source: 2020:Q2MBAdelinquencysurvey.) 39TheFHAservicerisalsonotfullycompensatedforforeclosurecosts.Tozer(2019)estimatesthatuncompensated costs are about $10,000 per FHA claim. To limit liquidity outflows, the FHA determined that loans thatre-performafterexitingforbearancecanbemadecurrentbyissuingapartialclaim,reimbursingtheservicerforprincipalandinterestadvancesduringforbearance. TheFHAalsocreatedatemporaryliquidity facilityforservicers. 40UrbanInstitute (2021) calculates that in December 2019, 90 percent of Ginnie Mae mortgages were originatedbynonbanks,andthenonbanksharehadincreasedto93percentbyDecember2020. 26

studying quantities. For interest rates, we focus on the rate difference between borrowers with a credit score of 640 versus 680. A credit score of 680 is typical for FHA loans (roughlyatthe60thpercentileinthefirstquarterof2020),while640isroughlyatthe20th percentile.41 Results in Figure 9 show a clear increase in the interest-rate premium for the riskiest FHA borrowers. Panel A traces the evolution of the FICO 640-680 interest rate spread in the offer data. The spread was stable at about 20 bp before the pandemic, but it rose sharply beginning in April 2020, coincident with the post-CARES Act surge in forbearance. The spread peaked at about 70 bp in June, 50 bp above pre-pandemic levels. The spread then decreased and displayed some volatility, but it remained elevated through to December 2020. The low-FICO spread also increased in the locks data (panel B of Figure9),althoughthepeakincreasewasonlyabout25bp,comparedwith50bpintheoffer data.42 Dataonquantitiesarealsoconsistentwithadeclineincreditsupplytohigh-riskFHA borrowers. PanelCofFigure9showsthatthenumberoflendersofferinganyFHAmortgages drops by one-quarter during the market volatility in March 2020. In the highest risk segment (640 FICO) the number of lenders then falls further in April, to half the prepandemic level, even though lenders re-enter the market for lower risk loans. Lenders graduallyreturntotheFICO640segmentlaterin2020. Similarly,PanelDshowsthatthe share of FHA purchase rate locks to borrowers with a FICO score lower than 640 drops from 30 percent before the pandemic to less than 15 percent in April, coincident with the spike in rate spreads and drop in the number of lenders.43 Notably there is no similar declineforFHArefinances. ForanFHAissuer,refinancinganexistingcustomerdoesnot presentadditionalrisk,becausetheissuerisalreadyresponsibleforadvancingpayments 41InternetAppendix Figure A.16 displays the credit score distribution of FHA originations over the 2019– 2020 period. Other assumptions for the offer data are that we study 30-year purchase-money FRMs with zeropoints,LTV=95to97percent(whichisbyfarthemostcommonintheFHAsegment),andDTIof36 percent. 42This difference between offers and equilibrium outcomes may reflect heterogeneity in supply responses across lenders—e.g., a subset of risk-averse lenders post very high rates for low-FICO loans, but these lenders comprise only a small share of locks because their uncompetitive rates mean that few borrowers choosethem. Weseeasimilarpatterninthejumboresultsinthefollowingsection. 43As the figure shows, the fraction of loans to lower-than-680 FICO borrowers also falls but by a smaller amount. We also find similar patterns for quantities using McDash origination data—see Figure A.16 in the InternetAppendix. The timing of the drop in low-credit-score lending is less sharp in McDash than OptimalBlue,however,whichweinterpretasbeingduetothevariabletimelagbetweenthelockdateand originationdate(recallingthatMcDashreportsonlythelatter). 27

iftheborrowerdefaults.44 So far we have focused on variation in risk within the FHA market. Section A.4 in the InternetAppendix instead compares markups on FHA loans with less risky prime conforming loans. We indeed find that both the primary-secondary spread and gain-onsale increased by larger amounts for FHA mortgages, consistent with an amplification of theriskpremiumassociated withFHAlending. VI.A.2 TheJumboMarket Jumbo mortgages are large loans exceeding the conforming limits for agency securitization. Although jumbo borrowers typically have high incomes and credit scores, jumbos are relatively risky for lenders and investors (usually banks) because they do not carry government-backed credit guarantees. Comparing jumbo and conforming loans thereforeshedslightonwhethertheseguaranteesstabilizedcreditsupplyinthepandemic. We use Optimal Blue data to estimate the interest rate spread on jumbo mortgages comparedwithsmallerbutotherwiseidenticalconformingmortgages,followingourearliermethodology. Resultsarepresented inFigure10. Panel A plots the jumbo-conforming offer rate spread for a mortgage with LTV of 80 and credit score of 750. Prior to the pandemic, jumbos carried an interest rate premium of 10-25 bp. This spread then increases sharply, to 80-100 bp from April through August 2020, before declining to about 50 bp by the end of 2020. The estimated spread based on rate locks (Panel B) follows a similar pattern, from 0-10 bp to a peak of 40-50 bp from AprilthroughAugust,thenadeclinetoabout30bpbyDecember.45 PanelBofFigure10plotstheinterestratepremiumon“superconforming”(alsoknown as“conforming jumbo”)mortgages. These areloansthatexceed thenationalconforming loan limit but are eligible for agency securitization because they are located in a county 44Infact,refinancingmayevenreduceriskbyloweringtheborrower’smonthlypayment(FusterandWillen, 2017).Furthermore,alargemajorityofFHArefinancesoccurunderastreamlinedrefinancingprogramthat waives many requirements, including the need to conduct a property appraisal. Thus, operational issues relatedtothepandemicwoulddisproportionatelyaffecttheFHApurchasemarket. 45We also see a rise in the jumbo-conforming spread in time-series rate data from Mortgage News Daily and theMortgageBankersAssociation—seeFigureA.18intheInternetAppendix. Insightdataalsoindicatea largeincreaseinthejumbo-conformingspreadfor5/1ARMs. 28

with a higher local limit (VickeryandWright, 2013).46 Superconforming loans are somewhat less liquid and are also less likely to be purchased by the Federal Reserve in its quantitative easing program, as we discuss in Section VI.B. Including superconforming × time dummies in the rate locks regression model, we find that the superconformingconforming spread does indeed rise early in the pandemic, from 10-20 bp to 30-35 bp, coincident with the period of significant Fed MBS purchases. The spike in the spread is onlytemporary,however;itfallsbacktonear-normal levelsbymid-June. We also find a reduction in the volume of jumbo lending. The number of lenders offering jumbos to FICO 750 borrowers on the Optimal Blue platform drops by more than half in the early stages of the pandemic (Panel C of Figure 10), before slowly recovering to about 20 percent below pre-pandemic levels by December 2020. For lower FICO borrowers (680 and 640) the number of active lenders collapses almost to zero. By contrast, the number of lenders in the conforming market remains fairly steady, as shown earlier in Figure 7. The volume of jumbo locks as a fraction of all rate locks declines by about half,bothforpurchaseloansandrefinancings(panel D),beforerecovering. This evidence is consistent with a negative supply response, but an important caveat is that Optimal Blue primarily reflects mortgages originated by nonbanks. Banks have a strongpresenceinthejumbomarket,andseverallargebanksstoppedpurchasingjumbos fromnonbanksduringthepandemic(Eisen,2020). Apotentialconcernisthatthedropin quantityweobservesimplyreflectssubstitutioninlendingfromnonbankstobanks. For a more representative picture of the jumbo market, we turn to McDash loan-level data, which have good coverage of both banks and nonbanks. We estimate linear probability models in which the dependent variable equals 1 for a jumbo loan, using McDash loans within a 10 percent window around the applicable conforming loan limit (the national limit, or the relevant higher local limit for mortgages in high-cost counties). The keyexplanatory variableisapandemicdummyequalto1fromApril2020onward. Results in panel A of Table 4 show a significant decline in the fraction of jumbo loans, of about 7 to 8 percentage points compared with a sample average of 16 percentage 46FannieMaeandFreddieMacmaynotpurchaseorsecuritizemortgagesexceedingtherelevantconforming loanlimit. Thenationalconforminglimit($510,400in2020forasingle-familyhome)appliesinmostcounties, but the limit is higher in counties with high home prices, up to $765,600 in 2020. Superconforming mortgagesareloansforamountsbetweenthenationallimitandthesehigherlocallimits. 29

points.47 These results are fairly similar across purchase mortgages and refinancings and arealsorobusttocontrollingforloancharacteristics. To sum up, the results in this section indicate that heightened forbearance and nonpaymentriskdidhavesomechillingeffectsonmortgagecreditsupplyoutsidetheprime conforming market. Those affected include low-income FHA borrowers as well as highincomejumboborrowersineligibleforgovernmentguarantees.48 VI.B Quantitative Easing TheFedrapidlyincreasedthesizeofitsMBSportfolioduringtheearlymonthsofthecrisis,from$1.37trillionatthestartofMarch2020to$1.90trillionbythestartofJuly(source: New York Federal Reserve Bank). Since the Fed purchases only agency MBS (excluding jumbos), our evidence above on the jumbo market may in part reflect the beneficial effects of Fed QE on mortgage credit supply. Understanding the effects of QE during the pandemicisalsoofindependentinterest. Theearliertime-seriesevidenceinFigure2suggeststhattheresumptionofFedQEin mid-March did reduce mortgage financing costs. OAS drops by about 40 bp right after QEbegins,andoptioncostdeclinesduetoadropininterestratevolatility. Thesechanges are almost exactly offset by a rise in the primary-secondary spread, however, leaving the mortgage-Treasuryspread almostunchanged. Adifferentwaytomeasuretheseeffectsthatdoesnotjustrelyontime-seriesvariation is to study superconforming mortgages exceeding the national conforming limit, as we alreadydidinSectionVI.A.2. Superconformingmortgagesarelesslikelytobepurchased by the Fed because the Fed obtains agency MBS pools through the “to-be-announced” or TBA market, and pools comprising more than 10 percent of superconforming loans are not TBA eligible (VickeryandWright, 2013; HuhandKim, 2020). Matching eMBS loanlevel data to security-level data on the Fed’s MBS holdings, we confirm that the proba- 47In InternetAppendix G we trace out the dynamics of the jumbo share by replacing the single pandemic dummywithavectoroftimedummies.Thesecoefficientsshowasustaineddownwardtrendinthefraction ofjumboloansbeginninginearly2020,whichbottomsoutinthethirdquarter. 48InInternetAppendixHwealsostudylendinginthesmallnon-qualified-mortgage(“non-QM”)partofthe jumbomarket(thenon-QMsegmentisstudiedindetailinDeFuscoetal.,2019). Wefindnon-QMlending dropsinlinewithjumbolendingasawholeduringthepandemic,butthereisnoobviousevidenceofan amplificationofthelegalriskassociatedwithnon-QMloans. 30

bility of an agency mortgage ending up in a pool in the Fed’s MBS portfolio does indeed dropsharplyjustabovethenationalconforminglimit,decliningbyabout40percent. Using this fact, panel B of Table 4 restricts our McDash sample just to high-cost counties and studies relative shifts in lending around both the national conforming loan limit (columns 1 through 3) and the higher local limit (columns 4 through 6). This allows us to disentangletheeffectsofFedQEfromcreditguarantees. TheresultsindicatethatbothQE and guarantees appear to have promoted credit supply—lending drops in relative terms just above both the national and local limits during the pandemic. The effects around thelocallimit(theupperboundtoqualifyforgovernmentguarantees)arequantitatively largerrelativetothesamplemean,however,suggestingthatcreditguaranteeshadarelativelylargereffectinbolsteringcreditsupply. Figure A.20 in the InternetAppendix further explores the dynamics of these effects, finding that the loan volume effects around the national limit (which reflect only QE) are significantintheearlystagesofthepandemic,buttheyarelesspersistentthantheeffects at the local limit. This is consistent with our earlier interest rate estimates in Figure 10, which show that the superconforming-conforming spread is elevated only through June 2020,whilethejumbo-conformingspread remainshighthroughout2020. We conclude that Fed QE lowered mortgage rates and expanded credit supply during the pandemic, but that the direct effects of QE may have faded after a few months. Ourestimatesdonotreflectthefullgeneral-equilibriumeffectsofQE,however,andthus shouldbeinterpretedaccordingly. VI.C Summing Up Ourresultsindicatethatgovernment-backedcreditguaranteessupportedmortgagelending during the pandemic, evidenced by a relative increase in rates and decline in quantities in the jumbo market, which does not feature guarantees. Our analysis of the FHA market,however,alsoshowsthatpublicguaranteesarenotalwayssufficienttofullyinsulatelendersagainstdefaultrisk—wefindevidenceofacontractioninsupplyforhigh-risk FHA borrowers despite the presence of guarantees. Federal Reserve quantitative easing alsosupported creditsupply,particularlyintheearlystagesofthepandemic. 31

VII Conclusion The mortgage market experienced a historic boom in 2020, with record origination volumesandlenderprofits. Butwhilemanyborrowersbenefitedfromrecord-lowmortgage rates, the evidence presented in this paper suggests that intermediation frictions limited thepass-throughoflowerratestohouseholds. Thislimitedpass-throughisonlypartially explainedbythehistoricalrelationshipbetweenmarkupsandmortgagedemand. Several pieces of evidence indicate that operational frictions reduced the elasticity of mortgage supply; in other words, hiring new workers and expanding capacity was more difficult than usual, keeping lender markups high. At the same time, there was a shift in the underwritingofcomplexloanstowardmoreelastic“fintech”lenders. As in the 2008 financial crisis, government credit guarantees have played an importantroleinstabilizingcreditsupply,reflectedinalargerincreaseinmortgageinterestrate spreadsinthejumbosegmentofthemarket,wheresuchguaranteesareabsent. However, our results also highlight the fact that such guarantees have not fully insulated lenders from risk, and that this residual exposure has been priced into mortgage interest rates at leastforsomeborrowers. Ourresultsaresuggestiveofsomesurprising“barbell”distributional effects of the pandemic, with low-income FHA borrowers and high-income jumbo borrowersexperiencingadeclineincreditavailabilityrelativetomiddle-classborrowers. Our results show that during the last three quarters of 2020, intermediaries collected 4 to 5 percent of a mortgage’s balance for the service of linking borrowers with savers, or about $160 billion in total. One surprising question is whether for rate-and-term refinances, it makes sense for intermediaries to collect anything. For an individual lender making a loan to a new borrower, careful underwriting makes sense, as the lender is taking on a new risk. But from the standpoint of the economy as a whole, rate-and-term refinancesuniformlylowertheoveralllikelihoodofdefault,evenifnonewunderwriting isdone. Bindingcapacityconstraintsduringlendingboomsmeanthatunderwritingrate-andtermrefinancesstarvesresourcesfromoriginatingpurchaseloansandcash-outrefinances, for which underwriting is more neccessary from the standpoint of the economy as a whole. In 2020, operational frictions related to the pandemic only amplified these capacity constraints and further increased the price of intermediation. Our findings, there- 32

fore, reinforce arguments for both streamlined refinances and automatically refinancing mortgageproducts. Muchlikeduringthefinancialcrisis,suchproductswouldhavesubstantially strengthened the transmission of low interest rates to households during the 2020pandemic. 33

References AGARWAL, S., S. CHOMSISENGPHET, H. KIEFER, L. C. KIEFER, AND P. C. MEDINA (2021): “Inequality During the COVID-19 Pandemic: The Case of Savings from MortgageRefinancing,”WorkingPaper. AMROMIN, G., N. BHUTTA, AND B. J. KEYS (2020): “Refinancing, Monetary Policy, and theCreditCycle,”AnnualReviewofFinancialEconomics,12,67–93. AN, X., L. CORDELL, L. GENG, AND K. LEE(2021): “InequalityintheTimeofCOVID-19: EvidencefromMortgageDelinquencyandForbearance,”WorkingPaper21-09,Federal ReserveBankofPhiladelphia. BERGER, D. W., K. MILBRADT, F. TOURRE, AND J. VAVRA (2020): “Mortgage PrepaymentandPath-DependentEffectsofMonetaryPolicy,”WorkingPaper25157,National BureauofEconomicResearch. BERRY, K. AND A. KLINE (2020): “Virus Scare Creates Perfect Storm for Mortgage Lenders,”AmericanBanker,March19. BHUTTA, N., A. FUSTER, AND A. HIZMO (2021): “Paying Too Much? Borrower SophisticationandOverpaymentintheUSMortgageMarket,”DiscussionPaper14924,CEPR. BLACK KNIGHT FINANCIAL SERVICES (2020): “MortgageMonitor: October2020Report,” Tech.rep. BOYARCHENKO, N., A. FUSTER, AND D. O. LUCCA (2019): “Understanding Mortgage Spreads,”ReviewofFinancialStudies,32,3799–3850. BRACKE, P., K. CROXSON, D. FAKHRI, P. SURICO, AND T. VALLETTI (2020): “Mortgage MarketDisruptions,”OccasionalPaper57,FinancialConductAuthority. BUCHAK, G., G. MATVOS, T. PISKORSKI, AND A. SERU (2018): “Fintech, Regulatory Arbitrage,andtheRiseofShadowBanks,”JournalofFinancialEconomics,130,453–483. CALEM, P., F. COVAS, AND J. WU (2013): “The Impact of the 2007 Liquidity Shock on BankJumboMortgageLending,”JournalofMoney,CreditandBanking,45,59–91. CAMPBELL, J. Y., N. CLARA, AND J. F. COCCO (2020): “Structuring Mortgages for MacroeconomicStability,”JournalofFinance,forthcoming. 34

CAPPONI, A., R. JIA, AND D. A. RIOS (2021): “The Effects of Mortgage Forbearance on Refinancing: EvidencefromtheCOVID-19Period,”CovidEconomics,68,1–46. CHERRY, S. F., E. X. JIANG, G. MATVOS, T. PISKORSKI, AND A. SERU (2021): “Government and Private Household Debt Relief during COVID-19,” Working Paper 28357, NationalBureauofEconomicResearch. DAVIDSON, A. AND A. LEVIN (2014): Mortgage Valuation Models: Embedded Options, Risk, andUncertainty,Oxford UniversityPress. DEFUSCO, A. A., S. JOHNSON, AND J. MONDRAGON (2019): “Regulating Household Leverage,”ReviewofEconomicStudies,87,914–958. DI MAGGIO, M., A. KERMANI, B. J. KEYS, T. PISKORSKI, R. RAMCHARAN, A. SERU, AND V. YAO (2017): “Interest Rate Pass-Through: Mortgage Rates, Household Consumption,andVoluntaryDeleveraging,”AmericanEconomicReview,107,3550–88. DI MAGGIO, M., A. KERMANI, AND C. J. PALMER (2020): “How Quantitative Easing Works: Evidence on the Refinancing Channel,” Review of Economic Studies, 87, 1498– 1528. EBERLY, J. AND A. KRISHNAMURTHY (2014): “EfficientCreditPoliciesinaHousingDebt Crisis,”BrookingsPapersonEconomicActivity,45,73–136. EISEN, B. (2020): “Wells Fargo Curtails Jumbo Loans Amid Market Turmoil,” Wall Street Journal,April4. EREL, I. AND J. LIEBERSOHN (2020): “Does FinTech Substitute for Banks? Evidence from thePaycheckProtectionProgram,”WorkingPaper27659,NationalBureauofEconomic Research. FABOZZI, F. J., ed. (2016): The Handbook of Mortgage-Backed Securities, Oxford University Press,7thed. FEDERAL HOUSING FINANCE AGENCY(2020): “FHFAAddressesServicerLiquidityConcerns, Announces Four Month Advance Obligation Limit for Loans in Forbearance,” Pressrelease,April21. 35

FUSTER, A., L. GOODMAN, D. LUCCA, L. MADAR, L. MOLLOY, AND P. WILLEN (2013): “TheRisingGapbetweenPrimaryandSecondaryMortgageRates,”FederalReserveBank ofNewYorkEconomicPolicyReview,19,17–39. FUSTER, A., S. H. LO, AND P. S. WILLEN (2017): “The Time-Varying Price of Financial Intermediation in the Mortgage Market,” Staff Report 805, Federal Reserve Bank of NewYork. FUSTER, A., M. PLOSSER, P. SCHNABL, AND J. VICKERY (2019): “TheRoleofTechnology inMortgageLending,”ReviewofFinancialStudies,32,1854–1899. FUSTER, A. AND P. S. WILLEN (2017): “Payment Size, Negative Equity, and Mortgage Default,”AmericanEconomicJournal: EconomicPolicy,9,167–91. GUREN, A. M., A. KRISHNAMURTHY, AND T. J. MCQUADE (2020): “MortgageDesignin anEquilibriumModeloftheHousingMarket,”JournalofFinance,forthcoming. HORVATH, A., B. KAY, AND C. WIX (2020): “TheCOVID-19ShockandConsumerCredit: EvidencefromCreditCardData,”SSRNElectronicJournal. HUGHES, C. (2020): “Coronavirus Complicates Home Buying and Mortgage Refinancing,”NewYorkTimes,March23. HUH, Y. AND Y. S. KIM (2020): “The Real Effects of Secondary Market Trading Structure: EvidencefromtheMortgageMarket,”WorkingPaper,FederalReserveBoard. IVERSON, B., J.WANG, J.YANG, ANDR.KLUENDER(2020): “BankruptcyandtheCOVID- 19Crisis,”WorkingPaper. IVEY, B.(2020): “Lender/ServicerRetentionRatesFallEvenasRefisRise,”InsideMortgage Finance,2020:35,September11. JIANG, E., G. MATVOS, T. PISKORSKI, AND A. SERU (2020): “Banking without Deposits: Evidence from Shadow Bank Call Reports,” Working Paper 26903, National Bureau of EconomicResearch. KEYS, B. J., D. G. POPE, AND J. C. POPE(2016): “FailuretoRefinance,”JournalofFinancial Economics,122,482–499. 36

KIM, Y. S., S. M. LAUFER, K. PENCE, R. STANTON, AND N. WALLACE (2018): “Liquidity CrisesintheMortgageMarket,”BrookingsPapersonEconomicActivity, 49,347–428. KLEIMANN, J. (2020): “On Earnings Debut, Rocket Reveals Plans to Dominate the U.S. MortgageMarket,”HousingWire,September3. KWAN, A., C. LIN, V. PURSIAINEN, AND M. TAI (2021): “Stress Testing Banks’ Digital Capabilities: Evidence From the COVID-19 Pandemic,” Working Paper, University of HongKongandUniversityofSt.Gallen. LAMBIE-HANSON, L. (2020): “Residential Mortgage Refinancing During the COVID-19 Pandemic,”Specialreport,FederalReserveBankofPhiladelphia. MCMANUS, D. AND E. YANNOPOULOS (2021): “COVID-19 Mortgage Forbearances: DriversandPaymentBehavior,”JournalofStructuredFinance,27. MORTGAGE BANKERS ASSOCIATION (2019): “Implementation of MBO Transitional Authority,”Issuebrief. ———(2021): “MortgageCreditAvailabilityDecreasedinDecember,”DataRelease. SCHARFSTEIN, D. AND A. SUNDERAM (2016): “Market Power in Mortgage Lending and theTransmissionofMonetaryPolicy,”WorkingPaper,HarvardUniversity. SHARPE, S. A. AND S. M. SHERLUND (2016): “Crowding Out Effects of Refinancing on NewPurchaseMortgages,”ReviewofIndustrialOrganization,48,209–239. TOZER, T. (2019): “A Primer and Perspective on Ginnie Mae,” Milken Institute Discussion Paper. URBAN INSTITUTE (2021): “Housing Finance at a Glance: A Monthly Chartbook,” February. VICKERY, J. AND J. WRIGHT (2013): “TBA Trading and Liquidity in the Agency MBS Market,”FederalReserveBankofNewYorkEconomicPolicyReview,19,1–18. WOLINSKY, A. (1986): “True Monopolistic Competition as a Result of Imperfect Information,”QuarterlyJournalofEconomics,101,493–511. 37

Figure2: AccountingfortheRiseintheMortgage-TreasurySpreadDuringTwoCrises A.COVID-19Pandemic(2020) 150 100 50 0 −50 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 − − − − − − − − − − − − 0 1 − J a n 0 1 − F e b 0 1 − M ar 0 1 − A pr 0 1 − M a y 0 1 − J u n 0 1 − J ul 0 1 − A u g 0 1 − S e p 0 1 − O ct 0 1 − N o v 0 1 − D e c Benchmark T−yield − 10y OAS Option cost Primary−secondary spread B.GlobalFinancialCrisis(2007-2009) 200 100 0 −100 Jul−07 Oct−07 Jan−08 Apr−08 Jul−08 Oct−08 Jan−09 Apr−09 Jul−09 Benchmark T−yield − 10y OAS Option cost Primary−secondary spread Blackdotsshowthechangeinspreadbetweenheadlinemortgagerate(fromPMMS)and10-yearTreasury yieldsrelativetothebeginningofeachepisode,inbasispoints. Decompositionbasedonthemethodology describedinSectionIII.A.Datasources: FreddieMacPMMS,OptimalBlue,J.P.MorganMarkets. 38

Figure3: IntermediationMarkupsintheConformingMarket A.Primary-SecondarySpread 280 260 240 220 200 180 160 140 120 100 80 )spb( daerpS yradnoceS−yramirP 01jan2019 01 mar2019 01 may2019 01jul2019 01sep2019 01nov2019 01jan2020 01 mar2020 01 may2020 01jul2020 01sep2020 01nov2020 01jan2021 F M O d d er A B mit p M ei B l a a ul M P c s nI e M gi S h t B.Gain-On-Sale 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 )%( elaS no niaG 01jan2019 01 mar2019 01 may2019 01jul2019 01sep2019 01nov2019 01jan2020 01 mar2020 01 may2020 01jul2020 01sep2020 01nov2020 01jan2021 F M O d d er A B mit p M ei B l a a ul M P c s nI e M gi S h t Primary-secondary spread (panel A) and gain-on-sale (panel B) measured based on the methodologies described in Sections III.A.1 and III.C, respectively. Data sources: Freddie Mac PMMS, Optimal Blue, J.P. MorganMarkets,MBA(viaHaverAnalytics). Notes: VerticallinerepresentsthedeclarationofanationalstateofemergencyinMarch13th,2020. 39

Figure4: Lenders’ReportedNetProductionIncome A.Inbasispoints 200 150 100 50 0 )spb( emocnI noitcudorP teN 1 1 1 1 1 1 1 1 1 1 1 1 1 q q q q q q q q q q q q q 9 0 1 2 3 4 5 6 7 8 9 0 1 0 1 1 1 1 1 1 1 1 1 1 2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 B.Multiplied byaverageoriginatedloanvolume 25 20 15 10 5 0 DSU noilliM ,mriF rep emocnI noitcudorP teN .gvA 1 1 1 1 1 1 1 1 1 1 1 1 1 q q q q q q q q q q q q q 9 0 1 2 3 4 5 6 7 8 9 0 1 0 1 1 1 1 1 1 1 1 1 1 2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 Datasource:MortgageBankersAssociationQuarterlyPerformanceReport(QPR).Sampleperiod:2008:Q3- 2020:Q4. SeriesinpanelAis“TotalNetProductionIncome, BasisPoints, SimpleAverage”fromTableB2 oftheQPR.ForpanelB,thisseriesismultipliedby“Avg. TotalLoansOriginated($000s)”fromthesame table. 40

Figure5: IntermediationMarkupsandtheDemandforRefinancing A.Prim.-Sec. Spreadvs. ProxyforRefiIncentives B.Prim.-Sec. Spreadvs. ApplicationVolume 275 2 2 2 5 5 0 9 683 88 4 7 75 7 4 5 5 7 8744 4 6 665 200 10 19 9 9 0 175 1 1112 1 2221 2 1111 1 1 1 1100 33 10 3 150 2 125 11 1 2212 100 75 )spb( daerpS yradnoceS−yramirP 275 2012−2019 2020 2 2 2 5 5 0 7 4 85 94 8 45 5 4 4 7 7 8 7 8 763 6 5 66 200 9 9 11900 175 1 1 2 211 1 3 1 11 0 11 2 20 2 1 11 11 1 0 3 3 150 2 125 1 11 2 21 2 100 75 1 1.5 2 2.5 3 3.5 WAC−10 Year Treasury Yield (%) )spb( daerpS yradnoceS−yramirP 2012−2019 2020 200 400 600 800 1000 1200 MBA Application Index C.Gain-on-Salevs. ProxyforRefiIncentives D.Gain-on-Salevs. vs. Application Volume 5.5 98 8 448 4.5 5 1 1 11 2 12 22 1 2 1 111 1 1 1 1 1 0 0 10 10 1 9 99 0 6 6 8 366 7 7 5 7 5 5 5 7 7 4 44 4 3 3.5 3 3 3 2 1 1 1 2 12 2.5 2 2 1.5 )%( elaS no niaG 2012−2019 5.5 9 8 44 8 8 2020 4. 5 5 1 1 2 2 11 9 5 7 4 9 8 5 4 1 4 6 11 1 5 1 5 1 0 9 1 1 20 6 2 0 0 2 1 1 1 1 61 7 1 0 7 7 76 3 4 3 3.5 3 3 3 2 1 1 1 2 2 1 2.5 2 2 1.5 1 1.5 2 2.5 3 3.5 WAC−10 Year Treasury Yield (%) )%( elaS no niaG 2012−2019 2020 200 400 600 800 1000 1200 MBA Mortgage Application Index Notes: numbersnexttoredsquaresdenotethecalendarmonthin2020. Thetrendlineandthe90%confidenceintervalsareestimates usingdata2012-2019. SpreadsandgainonsalecomputedbasedonthemethodologydescribedinSectionIII.C.Datasources: Freddie MacPMMS;J.P.MorganMarkets;SitusAMC;MortgageBankersAssociation(viaHaverAnalytics). 41

Figure6: MortgageLoanOfficerJobPostingsandEmploymentGrowth A.MortgageLoanOfficerPostingsandEmployment 400 350 300 250 2012 2013 2014 2015 2016 2017 2018 2019 2020 sdnasuohtnI 2.5 2.0 1.5 1.0 0.5 seeyolpmE001reP Counterfactual employment ActualEmployment Postingsper100 Employees B.PredictedversusActualEmploymentGrowth 0.04 0.03 0.02 0.01 0.00 -0.01 -0.02 -0.03 -0.03 -0.02 -0.01 0.00 0.01 0.02 0.03 0.04 PredictedEmploymentChange egnahCtnemyolpmElautcA b bc b b bc Pre-C bc O bc bc bc V bc bc bc bc bc I bc bc bc bc bc D bc bcbcbc bcbc bc bc bcbc bc bc bc bc bc bc bc bc bc bcbc bcbc bc bc bc bcbc bc bc bc bcbc bc bcbc bc bcbc bc bc bc bc bc bc bc bcbc bc bc bc bcbc bc bc bcbc bc bc bcbc bc bc bc bc bc bc bcbcbc b b b b b b COVID b bc bcbc bc bc bc bc bc bc bc PanelAshowsmortgageemploymentandpostingsperemployeefromtheBLSEstablishmentSurveyand BurningGlassTechnologies. PanelBcomparespredictedandactualvaluesfromtheregression logMLO t+1 − logMLO t = α+β 1 p t +β 2 p t−1 +β 2 p t−2 +ε t where MLO and p correspondtomortgageemploymentandpostingsperemployeeinPanelA.Baseline t t regression is run on data from 3/2012 to 2/2020. Dots labeled “COVID” are out-of-sample predictions for 3/2020 to 12/2020. Counterfactual employment in Panel A uses the predicted growth rates from the regressiontocomputeemploymentintheCOVIDperiod. 42

Figure7: CreditSupplyintheConformingMarket A.Offerratespread: FICO680vs750 B.Locksratespread: FICO680vs740+ .8 .7 .6 .5 .4 .3 .2 )%( daerpS .8 .7 .6 .5 .4 .3 .2 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 )%( daerpS 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 C.Numberoflenderspostingoffers D.Distributionofratelocksbycreditscore 160 140 120 100 80 60 40 20 0 sredneL .oN .1 .08 .06 .04 .02 FICO=640 FICO=680 FICO=750 0 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 erahS Purchase, FICO<680 Refi, FICO<680 Purchase, FICO<640 Refi, FICO<640 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 Notes:MeasuresofcreditsupplycomputedusingthemethodologydescribedinSectionV.A.Verticallinerepresentsthedeclarationofa nationalstateofemergencyinMarch13th,2020.Datasource:OptimalBlue 43

Figure8: WebSearchesforMortgageRefinancing A.Googletrendssearchindexforrefinancerelatedterms 400 300 200 100 0 xednI hcraeS 2012m1 2013m1 2014m1 2015m1 2016m1 2017m1 2018m1 2019m1 2020m1 2021m1 B.Googlesearchesvs. therefinanceincentive 400 3 300 200 7 4 68 5 9 2 12 10 11 100 1 0 xednI hcraeS elgooG 2012−2019 2020 1 1.5 2 2.5 3 3.5 WAC−10 Year Treasury Yield Notes: TheGooglesearchindexisconstructedbyaveragingthesearchesforthefollowingterms: mortgagerate,mortgagerefinance, mortgagelender,refinancecost,mortgagecost.Datasource:GoogleTrends. 44

Figure9: CreditSupplyintheFHAMarket A.Offerratespread: FICO640vs680 B.Ratelockspread: FICO640vs680 .8 .6 .4 .2 0 )%( daerpS .8 .6 .4 .2 0 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 )%( daerpS FICO=[620,639) FICO=[640,659) 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 C.Numberoflenderspostingoffers D.Distributionofratelocksbycreditscore 160 140 120 100 80 60 40 20 0 sredneL .oN .8 .7 .6 .5 .4 .3 .2 FICO=640 .1 FICO=680 FICO=750 0 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 erahS Purchase, FICO<680 Refi, FICO<680 Purchase, FICO<640 Refi, FICO<640 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 Notes:MeasuresofcreditsupplycomputedusingthemethodologydescribedinSectionV.A.Verticallinerepresentsthedeclarationofa nationalstateofemergencyinMarch13th,2020.Datasource:OptimalBlue 45

Figure10: CreditSupplyintheJumboMarket A.Offerratejumbo-conformingspread B.Ratelockspreads 1.2 1 .8 .6 .4 .2 0 −.2 −.4 −.6 )%( daerpS .7 .6 .5 .4 .3 .2 .1 0 −.1 01oct201 0 9 1nov201 0 9 1dec201 0 9 1jan202 0 0 1feb20 0 2 1 0 mar202 0 0 1apr20 0 2 1 0 may202 0 0 1jun2020 01jul202 0 0 1aug202 0 0 1sep202 0 0 1oct202 0 0 1nov202 0 0 1dec202 0 0 1jan2021 )%( daerpS Jumbo−Conforming Spread Super Conforming to Conforming Spread 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 C.Numberoflenderspostingoffers D.Shareofjumboratelocks 160 140 120 100 80 60 40 20 0 sredneL .oN .25 FICO=640 FICO=680 FICO=750 .2 .15 .1 .05 0 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 erahS Purchase Refinance 01oct201 0 9 1nov201 0 9 1dec2019 01jan2020 01feb202 0 0 1mar2020 01apr202 0 0 1may2020 01jun2020 01jul202 0 0 1aug202 0 0 1sep2020 01oct202 0 0 1nov202 0 0 1dec2020 01jan2021 Notes:MeasuresofcreditsupplycomputedusingthemethodologydescribedinSectionV.A.Verticallinerepresentsthedeclarationofa nationalstateofemergencyinMarch13th,2020.Datasource:OptimalBlue 46

Table1: IntermediationMarkupsandMortgageDemand (1) (2) (3) (4) Primary-Secondary Gain-on-Sale Spread(bp) ($per$100facevalue) ∗∗∗ ∗∗∗ Refiincentive(WAC-10YearTreasury) 21.71 0.655 (1.996) (0.0636) ∗∗∗ ∗∗∗ MBAApplicationsIndex 0.0718 0.00192 (0.0109) (0.000369) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ March-Maydummy 76.81 82.58 1.053 1.300 (10.45) (13.20) (0.270) (0.353) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ June-Augustdummy 85.77 90.42 1.383 1.598 (5.268) (5.642) (0.154) (0.178) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ September-Octoberdummy 52.19 55.34 1.285 1.451 (7.214) (8.218) (0.189) (0.223) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ November-Decemberdummy 37.67 33.81 1.317 1.285 (3.042) (4.679) (0.114) (0.165) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Constant 70.88 87.98 1.484 2.100 (4.621) (5.368) (0.188) (0.216) N 466 466 466 466 MeanY 125.28 125.28 3.04 3.04 R2 0.91 0.86 0.78 0.67 RMSE 9.91 12.26 0.39 0.47 RMSE(nodummies,-Feb. 2020) 8.93 11.29 0.38 0.47 Datasources: MortgageBankersAssociation(MBA),J.P.MorganMarkets,andFreddieMac30-YearFixed Rate Mortgage Average in the United States [MORTGAGE30US] and Board of Governors of the Federal ReserveSystem(US)10-YearTreasuryConstantMaturityRate[DGS10],bothretrievedfromFRED,Federal Reserve Bank of St. Louis. MBA applications index is a three-week backward-looking moving average. Modelsincludemonth-of-yeardummies(coefficientsnotdisplayed). Newey-Weststandarderrors(6lags) inparentheses. * ∼ p <0.10,* p <0.05,** p <0.01,*** p <0.001. 47

Table2: ChangesinFintechMortgageLendingDuringthePandemic Dependentvariable=100ifmortgageoriginatorisafintechlender,zerootherwise (1) (2) (3) (4) (5) (6) (7) (8) (9) PurchaseMortgages Refinancings AllLoans ∗∗∗ ∗∗∗ ∗∗∗ ∗ ∗∗∗ ∗∗ Pandemic 2.77 2.11 1.39 -0.22 -0.24 -0.83 3.80 0.76 0.06 (0.27) (0.28) (0.27) (0.52) (0.26) (0.33) (0.24) (0.24) (0.24) Pandemic× 2.69 ∗∗∗ 4.10 ∗∗∗ 3.48 ∗∗∗ FICO<680 (0.27) (0.41) (0.28) Numobs. 5284401 5284401 5284401 7054698 7054698 7054698 12339099 12339099 12339099 Meanofdep.var. 11.09 11.09 11.09 24.91 24.91 24.91 19.00 19.00 19.00 Loancontrols N Y Y N Y Y N Y Y FICO<680dummy N N Y N N Y N N Y Notes: FintechlendersareidentifiedbasedontheclassificationinFusteretal.(2019). Pandemicdummy= 1 if estimated origination date is April 2020 or later. Loan controls include: state dummies; government agency dummies (Fannie Mae, Freddie Mac, FHA, VA, RHS and Section 184 Indian home); dummies for refinancing andcash-outrefinancing; debt-to-income(DTI),DTI2, loan-to-value (LTV),LTV2, creditscore, creditscore2,loanprincipal,log(loanprincipal),numberofborrowers,andafirst-timehomebuyerdummy. Specifications including a FICO<680 interaction term also include an uninteracted FICO<680 dummy in thecontrols. SampleperiodisJanuary2019toDecember2020. Linearprobabilitymodelestimatedusing eMBS loan-level data on mortgages securitized by Fannie Mae, Freddie Mac and Ginnie Mae. Sample includesnonbankmortgagesonly. Standarderrorsclusteredbystate. ∼ p < 0.10,* p < 0.05,** p < 0.01, *** p <0.001. 48

Table3: MetroAreaDifferencesinConventionalConformingMortgageRates(%) (1) (2) (3) (4) (5) (6) Virusspread: ∗∗∗ COVIDcasespercapita -0.0032 (0.0008) ∗∗∗ ∗∗∗ Dummy: MSAintopquartile -0.0178 -0.0183 (0.0051) (0.0051) Unemployment: ∗ ∗ Year-over-yearchangeinU.R. 0.0112 0.0118 (0.0054) (0.0051) Marketconcentration: × COVID top4share 0.0025 0.0001 (0.0042) (0.0041) × COVID HHI 0.0048 (0.0042) GeographicFE yes yes yes yes yes yes × Loancontrols weekFE yes yes yes yes yes yes N 1,130,086 1,130,086 1,130,086 1,130,086 1,130,086 1,130,086 Data Sources: Optimal Blue locks data, New York Times COVID data, Bureau of Labor Statistics unemployment data, and Home Mortgage Disclosure Act data. Notes: Sample includes loans locked from November 2019 to August 2020 in the 100 largest CBSAs. Standard errors in parentheses, clustered at CBSA level. COVID cases per 1,000 are lagged one month. The top 4 lenders’ market share, HHI, and year-over-yearunemploymentareallstandardizedwithmeanof0andstandarddeviationof1. Mortgage interestratesareadjustedforpointspaidbyborrower(creditsreceivedfromlender). Additionalcontrols include CBSA-level fixed effects, as well as lock week interacted with loan characteristics: binned FICO score, binned loan-to-value ratio, interest rate type (fixed-rate, 5/1 ARM, 7/1 ARM, or 10/1 ARM), and loanpurpose(purchasevs. refinance). ∼ p <0.10,* p <0.05,** p <0.01,*** p <0.001. 49

Table4: ChangesinJumboLendingDuringthePandemic A.Changeinjumboshare DependentVariable=100ifmortgagebalanceexceedsconforminglimit (1) (2) (3) (4) (5) (6) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Pandemic -8.027 -6.691 -8.990 -6.582 -8.614 -6.756 (0.166) (0.149) (0.204) (0.165) (0.144) (0.117) N 270292 267645 299819 297252 589233 586663 MeanY 15.06 15.09 14.40 14.44 15.13 15.14 Originationtype Purchase Purchase Refinance Refinance All All Loancontrols N Y N Y N Y B.Changeinjumboandsuper-conformingshare,high-costareasonly DependentVariable=100ifmortgageisabovenationalorlocalconformingloanlimit > > nationalCLL localCLL (1) (2) (3) (4) (5) (6) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Pandemic -5.175 -6.487 -7.825 -8.604 -12.68 -11.14 (0.319) (0.191) (0.179) (0.293) (0.372) (0.273) N 152005 325164 492240 99478 133618 242839 MeanY 35.59 27.02 29.76 20.61 19.38 20.55 Originationtype Purchase Refinance All Purchase Refinance All Loancontrols Y Y Y Y Y Y Notes: Linear probability model, estimated using McDash loan-level data. Sample includes loans within 10% either side of the conforming loan limit (CLL) applicable to each loan (either the national or county-levellimit,whicheverisapplicable). InpanelB,sampleisrestrictedto“high-cost”countieswhere thecounty-levelCLLexceedsthenationalCLL.Pandemicdummy=1iforiginationdateisApril2020or later. Loan controls include: zip code dummies, debt-to-income (DTI), DTI2, loan-to-value (LTV), LTV2, creditscore,creditscore2,appraisalamount,andlog(appraisalamount). SampleperiodisJanuary2019to October 2020. ∼ p < 0.10, ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001. Robust standard errors in parentheses clusteredatzipcodelevel. 50