Recourse as Shadow Equity: Evidence from Commercial Real Estate Loans

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Recourse as Shadow Equity: Evidence from Commercial Real Estate Loans David Glancy, Robert Kurtzman, Lara Loewenstein, and Joseph Nichols 2021-079 Please cite this paper as: Glancy,David,RobertKurtzman,LaraLoewenstein,andJosephNichols(2021). “Recourse asShadowEquity: EvidencefromCommercialRealEstateLoans,” FinanceandEconomics DiscussionSeries2021-079. Washington: BoardofGovernorsoftheFederalReserveSystem, https://doi.org/10.17016/FEDS.2021.079. 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.

Recourse as Shadow Equity: ∗ Evidence from Commercial Real Estate Loans David Glancy1, Robert Kurtzman1, Lara Loewenstein2, and Joseph Nichols1 1Federal Reserve Board 2Federal Reserve Bank of Cleveland December 7, 2021 Abstract We study the role that recourse plays in the commercial real estate loan contracts of the largest U.S. banks. We find that recourse is valued by lenders and is treated as a substitute for conventional equity. At origination, recourse loans have rate spreads that are at least 20 basis points lower and loan-to-value ratios that are around 3 percentage points higher than non-recourse loans. Dynamically, recourse affects loan modification negotiations by providing additional bargaining power to the lender. Recourse loans were half as likely to receive accommodation during the COVID-19 pandemic, and the modifications that did occur entailed a relatively smaller reduction in payments. Keywords: commercial real estate, recourse, LTV JEL Classification: G21, G22, G23, R33 ∗ThankstoBenCraig,AndraGhent,BenjaminKay,NedPrescott,TimRiddiough,andtheaudiencesat the2021CREDAResearchSymposium,theFederalReserveBoardFinanceForum,andtheClevelandFed forhelpfulcommentsandsuggestions. Theviewsexpressedinthispaperaresolelythoseoftheauthors and do not necessarily reflect the opinions of the Federal Reserve Board, the Federal Reserve Bank of Cleveland,ortheFederalReserveSystem.

1. INTRODUCTION Commercial mortgages are heterogeneous contracts, with terms settled through a backand-forth negotiation between the lender and the borrower. Loans that appear risky along one dimension, such as loan-to-value (LTV) ratios, often have other characteristics to mitigate those risks.1 One of these contractual terms—recourse—can act as a type of “shadow equity,” providing lenders access to borrowers’ assets beyond the pledged collateral, thus reducing some of the risks of borrower leverage.2 Recourse can theoretically provide significant value to banks by discouraging strategic default and increasing the recovery from a liquidation (Childs et al., 1996; Ghent and Kudlyak, 2011). Even if direct recoveries from deficiency judgments are limited—due to high legal costs or low values on the claimable assets of defaulting borrowers—recourse clauses can still provide lenders leverage in negotiations with distressed borrowers. In this paper, we take advantage of detailed loan-level data on the commercial real estate (CRE) portfolios of the largest U.S. banks to perform a comprehensive analysis of the value of recourse to lenders both at loan origination and during loan modification negotiations. Unlike other CRE lenders, who overwhelmingly provide non-recourse loans, banks offer both recourse and non-recourse financing.3 This heterogeneity in bank CRE loan contracts allows us to use within-lender variation to study how recourse clauses affect loan terms and outcomes relative to otherwise similar non-recourse loans. Our analysis makes four contributions to the literature. First, we provide some basic empirical facts about the prevalence of recourse in bank CRE loan portfolios and the observable differences between loans with and without recourse. Roughly three-fourths of bank CRE loans have full or partial recourse. The most notable difference between recourse and non-recourse loans is in size; the average origination amount of a recourse loan is $9 million, compared to $43 million for a non-recourse loan. This disparity implies that only 45 percent of bank CRE loans by value have recourse. Second, we show that recourse enables borrowers to receive meaningfully lower loan rate spreads. Controlling for observable loan and property characteristics, we find that recourse loans command spreads that are 20 basis points lower than otherwise similar loans. As banks may require recourse on some loans to address unobserved risk characteristics, this estimate likely provides a lower bound of the true effects. Indeed, 1See Ambrose and Sanders (2003), Harrison et al. (2004), Titman et al. (2005), and Grovenstein et al. (2005)forexamples. 2Statelawslimitingtheuseofrecourseapplytoowner-occupiedresidentialproperties. Whiletheymay apply to small multi-unit properties where the owner resides in one unit, they do not apply to the vast majorityofcommercialrealestate. 3CMBSloansarebankruptcyremotebydesignandthereforenon-recourseoutsideof“badboy”clauses, whichtriggerrecourseintheeventofaparticularbadact(suchasfraud)onthepartoftheborrower. 2

when we instrument for recourse using the lending bank’s tendency to require recourse for observably similar loans—thus identifying off lender preferences rather than borrowerspecific underwriting—we estimate that recourse loans command spreads that are 52 basis points lower. These findings suggest that lenders place significant value on the addition of recourse to a loan contract. Third, we show that recourse substitutes for more conventional forms of equity. Using a similar approach to our analysis of rate spreads, we find that recourse is associated with LTV ratios at origination that are 2.8 percentage points higher (3.4 percentage points in the IV specification). Therefore, in addition to lowering interest costs, recourse provides property owners with a means of increasing their leverage. This higher leverage may be desirable for investors that either seek a higher return on equity or lack the liquid assets to make a down payment satisfying normal underwriting metrics. Lastly, we demonstrate that recourse enhances lenders’ bargaining power in loan modification negotiations. Historically, recourse loans are modestly less likely to receive a modification or credit rating downgrade. During the COVID-19 pandemic, credit rating downgrades for recourse and non-recourse loans increased in parallel. However, despite having similar rates of stress, recourse loans were half as likely to receive modifications as non-recourse loans. In other words, borrowers with recourse were not less likely to need a modification during the pandemic, but they were much less likely to receive one. Recourse loans that did receive modifications received ones that were less beneficial to the borrower.4 Ourpaperiscloselyrelatedtotheliteratureontheuseofrecourseinrealestatelending. Most empirical work on the use of recourse is on residential mortgage lending, most notably Ghent and Kudlyak (2011).5 The existing literature on this topic for commercial mortgages is largely theoretical. The models of Childs et al. (1996) and Lebret and Quan (2017) demonstrate that borrowers can achieve lower spreads or higher leverage by taking out recourse loans.6 To our knowledge, the only other paper that empirically studies 4Theimplicationthatrecourseprovidesthelenderwithmorebargainingpowerinloanmodification negotiationsiseconomicallyimportant. AsshownbyBlacketal.(2017),banksaremuchmorelikelythan securitizedlenderstomodifyloansinordertomitigatelosses. Renegotiationsarealsomuchmorefrequent for commercial mortgages than for residential mortgages, potentially because there is less asymmetric informationbetweenborrowersandlenders(Adelinoetal.,2013). 5GhentandKudlyak(2011)showthatmanyresidentialmortgagesaresubjecttorecourse,dependingon thestate. Exploitingthesestatedifferencesinthelegalityofrecourse,theauthorsfind,amongotherresults, thatrecourseactsasastrategicdefaultdeterrentandinducesmorelender-friendlydefaultwhendefault doesoccur. Interestingly,theauthorsfindhigherinterestratesonmortgagesinrecoursestates,whichthey leaveasapuzzle. Withthemoregranularloan-levelheterogeneityinrecoursefromourdata,weshowthat recourseisassociatedwithlowerspreads,consistentwiththeory. 6In addition, Corbae and Quintin (2015) explore the role of leverage in inducing foreclosures in the GreatRecessionanditsaftermath. Theauthorsincludeanextensionoftheirmodel,findingthatrecourse canplayanimportantroleinmitigatingforeclosuresbyreducingtheincentiveforstrategicdefault. 3

recourse in commercial mortgages is Binder and Kim (2019), who show that recourse has little ability to predict future defaults.7 More broadly, we also contribute to the literature examining the underwriting and performance of commercial mortgages. Our work shows that the joint determination of various underwriting characteristics can complicate the analysis of the effects of borrower leverage. Loans may have low LTVs to offset other unobserved risks, causing LTV to lack its expected relationship with default risks (Grovenstein et al., 2005; Ambrose and Sanders, 2003) or spreads (Titman et al., 2005). Likewise, borrowers may choose low LTVs if default is more costly (Harrison et al., 2004). Consistent with this literature, we show empirically that recourse can compensate for having a high LTV and mitigate some of the risks associated with higher leverage. The rest of the paper is structured as follows. In Section 2, we discuss the data used in our analysis. In Section 3, we review the prevalence of recourse in bank CRE portfolios and discuss observable differences between recourse and non-recourse loans. In Section 4, we analyze the effects of recourse on rate spreads and leverage for CRE loan originations. In Section 5, we investigate the relationship between recourse and loan performance. In Section 6, we conclude. 2. DATA We use supervisory data collected to support the Federal Reserve stress tests, which contain loan-level information on the commercial real estate portfolios of the largest banks in the United States. The reporting panel consists of banks with consolidated assets of $100 billion or more.8 These banks report information for all loans with a committed balance of $1 million or more. The data include construction and land development (CLD) loans as well as loans secured by non-owner-occupied income-producing properties. The data include an array of information on banks’ portfolio loans: the interest rate, committed exposure (drawn plus undrawn credit), loan balance, dates of origination and maturity, amortization (for example, interest-only versus fully amortizing), whether there is a prepayment penalty, and the interest rate variability type (fixed versus floating). It also includes information on the property securing the loan: the appraised value, the type 7Also related is work by Beyhaghi (2021), which studies third-party guarantees in commercial and industrialloans. Thoughthesettingisquitedifferent—hisstudyfocusesonnon-collateralizedlendingand includesgovernmentguarantees—manyofthefindingsarecomplementary;healsofindslowerloanrates, betterperformance,andlowercollateralizationforloanswithguarantees. 8As part of their capital assessment and stress tests, banks file regulatory forms called the Y-14Q on aquarterlybasis. ThecommercialrealestatedatacanbefoundthroughScheduleH.2. Oursamplealso includes some loans from banks with $50 to $100 billion in assets because of the lower asset threshold before2019. Thedataareatthefacilitylevel,andafacilitycanincludemultipleloanstothesameentity; nonetheless,mostfacilitieshaveonlyoneloan,sowetreatthedataasloanlevel. 4

(for example, hotel versus retail), and geographic location information at the ZIP code level, which we map to core-based statistical areas (CBSAs).9 We construct the LTV ratio using the loan balance and appraised value.10 The spread between the loan interest rate and the bank’s cost of funds is calculated using the interest rate, the dates of origination, and the term.11 The data also include loan-level risk measures. First, banks provide a standardized version of their internal borrower risk rating for each loan. Banks have their own internal risk categorizations but provide a mapping from these internal ratings to a common scale along the lines of what is used for bond ratings. Internal ratings can sometimes span multiple ratings on the common scale, so there is a minimum and maximum rating provided on the common scale. When constructing an indicator for whether a borrower is rated the equivalent of BBB+ or higher, we take the maximum rating. A subset of stress test banks are also “advanced approaches” institutions. These banks are required to report their estimates of loan probability of default and loss given default, the product of which is the expected loss of the loan. For loans from banks that do not report these variables, we impute the expected loss using the average value for the loan’s particular borrower risk rating to avoid limiting our sample. Key to our analysis, banks also provide information on whether a loan has recourse. As of September 2014, the recourse field indicates whether the loan has full, partial, or no recourse. Before that date, banks only indicated whether a loan had any form of recourse and did not distinguish between full and partial recourse. We label any loan that has partial or full recourse as having recourse.12 We exclude from our sample all loans that are missing key information or that contain outliers. In particular, we drop all loans with a negative or missing committed balance, all loans with an LTV greater than two or less than zero, all leveraged loans, all acquired loans, all loans to foreign borrowers, and all loans secured by properties 9ForloanswithaZIPcodethatdoesnotmaptoaCBSA,weassignaCBSAcodeof0. 10Forcross-collateralizedloans,banksreportthepropertyvalueasthetotalvalueofallcross-collateralized properties. For example, two cross-collateralized 80 percent LTV loans on two different $10 million propertieswouldbereportedasloansof$8millionagainst$20millionincollateral. Becausecollateralis double-countedandloanamountsarenot,weadjustpropertyvaluesandLTVstoonlyreflecttheportion ofthecollateralapplicabletothatloan. Therefore,theLTVsintheexampleloanswouldbetreatedas80 percentratherthan40percent. 11Forfloatingrateloans,weuseone-monthLIBORasthereferencerate. Forfixed-rateloans,wecompute the maturity-matched swap rate. For loans with terms under two years, we linearly impute between one-month LIBOR and the two-year swap rate. For terms above two years, we linearly impute between availableswaprates. 12Togiveasenseofwhatthedatalooklikebeforecombiningpartialandfullrecourseloans,inTable A.2intheappendixweprovideaparalleltabletoTable1,breakingoutloansbywhethertheyhavefull or partial recourse. Note that this table uses data only from 2015 on, as that is when the more detailed recoursefieldwasincorporated. 5

outside the United States. We also drop loans that have missing values for recourse, cross-collateralization, loan value, origination or maturity date, state code on the property, whether the loan is floating rate, or whether the loan is the first lien on the property. In addition, we drop loans if they are the only observation for that lender-state-year-property type combination.13 We use different samples in our at-origination and our dynamic analyses. In Section 5, in which we analyze the performance of loans over time instead of outcomes at origination, we use the full sample of loan observations between 2012 and 2020, cleaning the data as above. In Sections 3 and 4, in which we analyze data as of origination, we also exclude loans that appear in the data with a lag in order to avoid selection bias due to differential attrition.14 Specifically, we drop loans that were originated before the bank began reporting data, were originated more than two quarters before they first appear in the data, or that have an origination date that differs from the earliest origination date (to exclude modified loans). We focus our analysis on commercial loans secured by stabilized properties, as we are better able to control for key characteristics affecting risk premiums on such loans. Loans against transitional properties—those properties underlying renovation or construction projects—are often valued using an estimate of its future income instead of actual income, making the property value subject to measurement issues.15 Furthermore, the performance of loans on transitional properties is highly dependent on the business model of a particular borrower, making the loan and property controls employed in our analysis less effective at controlling for risk. 13ThislastconditionisappliedforthesakeofkeepingthesampleconsistentinOLSandIVspecifications. OLSestimatesarelittlechangedwhenincludingthesesingletonobservations. 14Loans that appear in the data with a lag—for example, loans originated before a bank started Y-14 reporting—may not be reflective of the sample of loan originations for that bank-quarter: shorter-term loansmaymature,lower-qualityloansmaydefault,loansmightprepay,orloansmaybemodifiedsothat thetermsatthetimeofreportingdonotreflectoriginationvalues. However,ourresultsarequalitatively similartothosewhenrunonthemoreexpandedsample. 15Moreformally,wedefineloansontransitionalpropertiesasanyconstructionandlanddevelopment loanoranyloanforwhichthereportedpropertyvalueisanestimateforoncethepropertyiscompletedor stabilizedasopposedtothevaluebeingreported“asis.” 6

3. THE PREVALENCE OF RECOURSE IN BANK CRE PORTFOLIOS We present summary statistics on key variables from our cleaned sample of loans at origination—with finer detail broken out by recourse status and property type—in Table 1. These statistics provide information on how often recourse is a feature of bank CRE loans and how loans with recourse differ from non-recourse loans in terms of observable characteristics. Recourse is fairly common: 78 percent of loans secured by stabilized properties have recourse.16 Origination amounts for recourse loans secured by stabilized properties are, on average, only about one-fourth the size of non-recourse loan amounts, implying that the recourse share is smaller on a value-weighted basis, standing at 45 percent.17 Other terms also differ between recourse and non-recourse loans. Recourse loans have lower spreads, lower LTVs, longer terms, and are less likely to be interest only or floating rate compared with non-recourse loans, on average. The prevalence of recourse also varies across property types. For example, 82 percent of multifamily loans have recourse, whereas around 66 percent of lodging loans have recourse. On average, the largest difference in rate spreads by recourse status is for multifamily loans, for which recourse loans carry spreads that are 31 basis points lower than those for non-recourse loans. Recourse loans secured by hotels also carry notably lower spreads, while the average spreads for recourse and non-recourse loans are within 7 basis points of one another for retail, industrial, and office CRE loans. The use of recourse also differs substantially across lenders. In the top panel of Table 2, we divide lenders into quintiles by the share of their loans that have recourse.18 The top quintile of lenders have recourse on over 90 percent of their CRE loans, while the lowest quintile of lenders have recourse on 14 percent of such loans.19 Differences in the use of recourse are less stark across states. The bottom panel of Table 2 depicts quintiles of states by recourse share. The residential mortgage literature has focused on state differences in recourse laws (Ghent and Kudlyak, 2011). While laws allowing or preventing recourse on owner-occupied residential properties do not generally apply to commercial properties, there are still legal differences across states that can make it more or less difficult to obtain a deficiency judgment. Recourse shares 16TableA.1intheappendixhassummarystatisticsforloansontransitionalproperties. Justunder70 percentofloanssecuredbytransitionalpropertieshaverecourse. 17The median loan amounts are $2.1 million and $7.3 million for recourse and non-recourse loans, respectively. 18Weusequintilessothatmultiplebanksareineachbucket,thuspreservingtheiranonymity. 19Wewillexploitthisnotableheterogeneityinbanks’useofrecourseinourIVstrategydescribedinthe nextsection,thelogicbeingthatifaloanhasrecoursebecausethelendingbankalmostalwaysrequires recourse,thentherecourseclauseislesslikelytoreflectunobservedborrowerrisks. 7

for loans secured by stabilized properties range from 56 percent to 83 percent across state quintiles, with some variation by property type. 4. DO LENDERS VALUE RECOURSE? Qualitatively, the effect of recourse is straightforward: recourse should act like additional equity and reduce losses in the event of default, offer better incentives to borrowers to avoid default, mitigate the agency problems that can worsen near default, and provide lenders with more bargaining power if loans need to be modified or liquidated. The inclusion of recourse in a loan contract should enable borrowers to either achieve more favorable loan pricing or be allowed greater risk along other dimensions, such as higher LTVs. However, the quantitative significance of such effects is uncertain. Property investors tend to specialize in particular regions or property types, meaning that the value of an investor’s other assets is likely to be highly correlated with the value of the subject property. By the time a borrower has an incentive to default, a borrower’s net worth may have declined such that recourse provides little value. Moreover, the costs and difficulties of achieving a deficiency judgment may substantially reduce the value banks place on recourse. In this section, we study the extent to which lenders value recourse. With our analysis, we put reasonable bounds on the effect of recourse on spreads (in the first subsection) and LTVs (in the second subsection). 4.1. Recourse and Interest Rate Spreads To investigate how recourse affects loan spreads, we run the following regression: r = β Recourse +β LTV +γ (cid:48) X +τ +η +ξ +ε , (1) i,b,t 1 i,b,t 2 i,b,t i,b,t t b c(i) i,b,t where r is the spread on loan i from bank b in origination year t, Recourse indicates i,b,t i,b,t whether that loan has recourse, LTV is the loan-to-value ratio, and X is a vector of i,b,t i,b,t loan-level controls. The regressions also include lender (η ), origination-year (τ), and b t state-by-CBSA fixed effects (ξ ).20 Our baseline set of controls is the natural logarithm c(i) of the loan term, the natural logarithm of the committed balance at origination, and indicators for whether the loan is interest only, has a prepayment penalty, has a floating 20Thesefixedeffectsallowustocapturebothdifferencesacrossmetropolitanareasanddifferencesin statelaws. CountiesoutsideofCBSAsaregivenaCBSAcodeofzero;thus,theirfixedeffectcorrespondsto allnon-urbancountieswithinthestate. 8

rate, is cross-collateralized, and is the first lien on the property. We also include property type fixed effects (where multifamily is the omitted property type).21 The key variables of interest are LTV and the recourse indicator. If recourse is valued by lenders as a form of shadow equity, we would expect β < 0. Likewise, if regular 1 equity (the down payment) is valued by lenders, we would expect β > 0. 2 The presence of recourse in a loan contract is—of course—not random, and the OLS estimates of β are likely to be biased. Banks may require recourse on riskier loans much 1 in the way they require lower LTVs on such loans. While we include a rich set of controls to account for differences in risk characteristics, recourse loans may have unobservable characteristics that make them riskier on average than non-recourse loans. OLS estimates of the effect recourse has on loan pricing therefore arguably provide a lower bound for the true effect. To shed light on the possible magnitude of this bias, we also estimate equation (1) usingtwo-stageleastsquares. Specifically,weinstrumentforaloan’srecoursestatususing the recourse loan share of other originations in that market (property type-state-year) from the given lender.22 Intuitively, variation in recourse due to differences in lenders’ use of recourse is less likely to reflect loan-specific risk characteristics than the differences in recourse clauses for individual loans. If anything, banks that require recourse more frequently for a particular type of loan might be more risk-averse toward such loans, resulting in recourse loans being safer along unobserved dimensions. As the IV estimate likely overstates the impact of recourse, the combination of OLS and IV estimates should provide reasonable bounds for the true effect of recourse. The estimates presented in Table 3 imply that recourse lowers loan rate spreads by between 20 and 52 basis points. The first two columns are from OLS regressions. In column (1), which omits the LTV control, we get a value of –0.204 for βˆ , with 1,OLS statistical significance at the 1 percent level. Adding LTV as a control in column (2) does not meaningfully change the estimated effect of recourse; βˆ edges down to –0.206. Although positive, the coefficient on LTV 1,OLS is small and statistically insignificant, implying that, on average, lower LTV loans do not command notably lower interest rates, likely reflecting the endogeneity of LTV choice 21InTable3,weshowthecoefficientsforindustrial,lodging,office,andretail. Thecoefficientsonother property types (for example, “mixed” and “condo”) are included in the specification but not displayed (duetospaceconstraints). 22We exclude the loan of interest when calculating the lender’s recourse share for similar loans. We includethefullsampleofloansonbothtransitionalandstabilizedpropertiestoconstructtheIV.Welook atalender’srecoursesharesbypropertytype-state-yeartocapturedifferencesinexperienceorpreferences across property types, differences in recourse laws across states, and changes in preferences over time. WealsouseasimilarapproachtoinstrumentfortheLTVatorigination,usingtheaverageLTVofloans originatedinthatmarket(propertytype-state-year)foragivenlender. 9

(Titman et al., 2005). Column (3) presents the IV estimates, where both recourse and LTV are instrumented for using the recourse share and mean LTV, respectively, within the particular property type, origination year, and state for a given lender. We find a value of –0.516 for βˆ , 1,IV with the estimate still significant at the 1 percent level. The OLS and IV estimates likely bound the true effect: OLS estimates are likely biased toward 0, reflecting higher recourse on riskier loans, and IV estimates are likely biased away from zero, reflecting differences in banks’ risk aversion manifesting in ways besides frequency in the use of recourse. Accordingly, the availability of recourse likely lowers loan rate spreads by somewhere between 20 and 52 basis points. The IV estimate for the effect of LTV on spreads is larger and is statistically significant. It implies that banks provide a 9 basis point rate discount for a 10 percent increase in LTV. This result is also consistent with βˆ being biased toward zero. 2,OLS In column (4) we add additional controls for risk based on banks’ internal risk ratings for loans to our OLS specification from column (2). We include a dummy variable for whether the internal risk rating is equivalent to that of an investment-grade credit (rated BBB or higher), and the expected loss (probability of default times loss given default).23 The interpretation of the coefficients on recourse and LTV is complicated in this specification, as recourse, LTV, and other terms are presumably a component of banks’ risk ratings and expected loss calculations. This specification thus tests whether banks offer lower spreads on loans with recourse beyond the assessed effect of these variables on risk ratings. The coefficient on recourse declines only modestly when adding these additional controls. The coefficient on LTV, however, switches signs and becomes negative, indicating that the positive relationship between LTV and loan pricing is fully captured by banks’ risk ratings. The other regression coefficients have the expected signs and are fairly similar across the four specifications. One finding worth highlighting is that cross-collateralized loans receive loan rates that are about 11 basis points lower than those on other loans. Crosscollateralization pledges properties securing other loans as collateral. As such, crosscollateralization can serve a function similar to that of recourse, but with the claim on borrowers’otherassetsinaliquidationlimitedtotheequityinanotherparticularproperty (Childs et al., 1996). This finding is thus consistent with the primary findings regarding the effects of recourse on spreads.24 23Introducingthesevariableslimitsoursamplesomewhat,becauseafewhundredobservationshavea missingriskrating. Notethataddingafullsetoffixedeffectsforcreditratinggivessimilarresults. 24We focus on recourse in this paper because it is much more widely used. Only about 5 percent of stabilizedloansarecross-collateralized. 10

4.2. Recourse as a Substitute for Conventional Equity To estimate the extent to which recourse substitutes for conventional equity, we run the regression specification described in equation (1) but with LTV as the dependent variable. All controls are as before, but with the loan rate spread included in place of LTV in some specifications. The results of these regressions are in columns (5) to (8) of Table 3. The range of estimated effects of recourse on LTV is narrower than for spreads: estimates range from 2.7 or 2.8 in OLS specifications to 3.4 in the IV specification. The OLS estimate in column (5) implies that recourse loans have LTVs that are about 2.8 percentage points higher than those of non-recourse loans. That is, borrowers who have equity at stake through a recourse clause are able to have modestly less equity at stake through their down payment. The estimate in column (6), which additionally controls for loan rate spreads, rises, but only modestly, consistent with the endogenous response of loan rate spreads biasing the estimated effect of recourse in (5) downward.25 Similar to the analysis of rate spreads, unobservable characteristics that affect recourse decisions are likely to bias our OLS estimates. If loans with riskier unobserved characteristics are more likely to require recourse and have stricter LTV limits, this pattern will cause a downward bias in our estimate of the effect of recourse on LTV. To address this bias, column (7) estimates the same specification as (6) but instruments for recourse using the frequency with which the lending bank requires recourse for other similar loans exactly as in column (3), as described in the previous subsection. The estimated effect of recourse rises, albeit less than in the spreads regression. The IV estimate implies that recourse loans receive LTVs that are 3.4 percentage points higher. Lastly, in column (8) we include loan-level risk rating controls. The estimated effect of recourse on LTVs is only modestly lower than in the other OLS specifications, with recourse loans receiving, on average, LTVs that are about 2.7 percentage points higher than those of non-recourse loans. Coefficients on other variables are in line with expectations. Loans with riskier terms or property characteristics typically have lower LTVs to compensate, and loans with high LTVs tend to receive higher spreads and worse risk ratings. The findings regarding cross-collateralization and LTV are again directionally similar to those of recourse as with the spreads results, and in this case they are also quantitatively similar as well: cross-collateralized loans receive LTVs that are about 3.2 percentage points higher. 25Asrecourseloanshavebeenshowntohavelowerspreads,andlowerspreadsareassociatedwithlower LTVs,thespecificationin(5)suffersfromomittedvariablebias. Recourseshiftsoutthesetofcontractsa bankwilloffersothatahigherLTVisavailableforagivenspread. Thefullextentofasupplyshiftisnot reflectedinourestimatedeffectonLTVduetomovementalongthecurvetoalowerspread. Controlling for spreads attempts to better identify the shift in available LTVs but is likely insufficient due to the endogeneityofloanratespreads: loanswithriskierunobservablesreceivehigherspreadsandlowerLTVs. 11

Overall, recourse loans not only receive lower interest rate spreads but also higher LTVs. Recourse may thus benefit borrowers with limited liquid assets, as it enables them to provide skin-in-the-game in a way other than through a down payment, expanding the set of available loan contracts. 5. DYNAMIC EFFECTS OF RECOURSE In Section 4, we showed that banks value recourse, charging lower spreads and allowing higher LTVs for recourse loans on average. In this section, we demonstrate one way in which recourse provides value to banks. We first document that CRE market stress during the COVID-19 period predominantly manifested itself in the form of higher modification rates and internal credit rating downgrade rates, but only modestly higher delinquency rates. We then show that while recourse loans were just as likely to receive a downgrade as non-recourse loans, they were 50 percent less likely to receive a loan modification. The modifications they did receive generally involved smaller reductions in required loan payments. We interpret these results as implying that recourse provides lenders with bargaining power in loan renegotiations. 5.1. Bank Loan Modifications Were Common during the COVID-19 Pandemic Banks tend to modify CRE loans before they become delinquent (Black et al., 2017). Given the high costs of foreclosure in commercial real estate and the limited contractual impediments to loan modifications, banks have an incentive to work with borrowers to avoid default.26 IncentivesformodificationwereparticularlypronouncedduringtheCOVID-19period: the stress was a large, unexpected shock generally outside of borrowers’ control and was perceived as largely transitory. These conditions limited the moral hazard concerns that can come with modifying troubled loans. Guidance from regulators also encouraged banks to work with borrowers. The interagency statement from bank regulators stated that “[t]he agencies view loan modification programs as positive actions that can mitigate adverse effects on borrowers due to COVID-19.”27 We identify loan modifications by comparing loan terms over time.28 Specifically, 26BanksaredifferentfromCMBSinthisregard,asCMBSservicershaveamorelimitedabilitytomodify loansduetoREMICrulesandpoolingandservicingagreements(PSAs). 27Theinteragencystatementfrombankregulatorsregardingloanmodificationscanbefoundathttps: //www.federalreserve.gov/newsevents/pressreleases/bcreg20200407a.htm. Also note thatshort-termmodificationsmadeinresponsetotheCOVID-19pandemicwerenotconsideredtroubled debtrestructuringsandthereforedidnotneedtobeaccountedforinthebank’sallowanceforloanand leaselosses. 28ThismethodissimilarinspirittothatofAdelinoetal.(2013). 12

a loan is considered modified if it switched from being amortizing to being interest only, if the committed balance rises (indicating interest payments are added to the loan balance as part of a forbearance plan), if the committed balance falls in tandem with a positive cumulative charge-off (indicating a write-off), if the maturity date is extended (outside of a pre-negotiated renewal), or if the loan enters troubled debt restructuring.29 Because we are interested in banks’ decisions regarding particular loans, we omit from the analysis a few banks that modified over 30 percent of their CRE loans in 2020:Q1, as such modifications are more likely to reflect blanket policies rather than banks’ assessments of the need to modify particular loans. We additionally assess loan performance based on whether loans receive rating downgrades or become distressed. We consider loans as distressed if they are delinquent, non-accrual, or involuntarily liquidated. We define a downgrade as a decline in the lender’s internal credit rating in a given quarter. Our estimates of quarterly modification, downgrade, and distress rates before (2012- 2019) and during (2020) the COVID period are in Table 4. CRE loans were modified at a rate of about 5.4 percent per quarter in 2020, up from a rate of 1.5 percent pre-2020. Though modifications rose for all property types, the rise was particularly pronounced for lodging loans, for which the modification rate rose to almost 12 percent per quarter in 2020, compared to roughly 2 percent pre-2020. Credit rating downgrades also rose during the COVID period: quarterly downgrade rates rose from 2.8 percent pre-2020 to over 6 percent in 2020. Loans secured against lodging properties again rose the most of all property types; lodging loans were downgraded at a rate of 22 percent per quarter in 2020, compared to just under 3 percent pre-2020. Despite the high rates of downgrades, borrowers for the most part were able to remain current on their loans. Distress rates were at much lower levels compared to modifications and downgrades, even during COVID. They rose from a quarterly rate of just under 0.4 percent pre-2020 to almost 1 percent in 2020. Even loans backed by lodging properties—which saw high rates of both downgrades and modifications—only reached distress rates of about 2.8 percent per quarter during 2020. These low distress rates stand in sharp contrast to loan performance in the CMBS market, where overall delinquency rates surpassed 10 percent in June 2020 and delinquency rates for lodging and retail-backed loans about doubled that average. 29Loans are also considered to be modified if there is a change in the origination date, which occurs whenthereisasubstantialchangeinaloan’sterms. Inaddition,wealsodenoteloansasmodifiedifthere arechangesininterestratesonfixed-rateloans(suchmodificationsarerare). 13

5.2. Recourse Loans Were Less Likely to Be Modified The first step in our analysis of the dynamic effects of recourse is to test whether recourse loans were less likely to be modified, downgraded, or become distressed during the COVID period (that is, in 2020). We run the following regression: M ×100 = β Recourse +β Recourse xCOVID i,b,t 1 i,b,t 2 i,b,t t +γ (cid:48) X +ψ (cid:48) X xCOVID +ε , (2) i,b,t i,b,t t i,b,t where M is an indicator for whether loan i from bank b is modified at time t, i,b,t Recourse indicates whether that loan has recourse, COVID is an indicator for whether i,b,t t the loan is from 2020, and X is a vector of loan-level controls and fixed effects. Each i,b,t specification includes all of the controls and fixed effects from column (1) of Table 3 as well as year-quarter fixed effects. X is expanded to include LTV and credit rating i,b,t controls in some specifications. All controls, including LTV, are the current values instead of the values at origination that were used in Section 4. Analogous specifications are run with downgrade or distress indicators as dependent variables. The coefficient of interest is the interaction between recourse and COVID (β ), which 2 we expect to be negative. We expect modification rates on recourse loans to be lower for two reasons. First, borrowers with recourse have less incentive to default when property values decline, as they have other assets at stake (Ghent and Kudlyak, 2011). Thus, even if one property does not generate enough income to service the debt on that property, the borrower may still make payments using other resources or returns from other assets rather than risk those assets. Second, banks may have less incentive to provide a modification on a recourse loan because they expect to be able to recoup any losses by filing a deficiency judgment post-liquidation.30 The coefficient on recourse may suffer from some identification problems similar to those detailed in Section 4. However, as COVID is a large, exogenous shock to CRE, whose effects are arguably orthogonal to ex-ante risk assessments, any change in the estimated effect of recourse during COVID should identify the effect of recourse on the outcome variable.31 30Recourseloanscouldalsoreceivefewermodificationsbecausetheyarelessriskyalongsomeunobserved dimensionandthereforelesslikelytoneedamodification. However,ourresultsinSection4.1implythat therecourseloanslikelyhave,ifanything,riskierunobservables. 31Recourseisendogenousinthatrecourseclausesmaybeincludedtooffsetunobservedrisks. Consequently,evenifthecausaleffectofrecourseisalessfrequentneedformodification,thiseffectmaybeoffset bytheincreasedneedfor riskierloanstobemodified. Thisbiasislikely smallduringCOVID,asitis a specificmanifestationofanadverseoutcome,andsensitivitytotheCOVIDshockisoftennotalignedwith perceivedrisksatorigination. Forexample,loansingatewaycitieswereperceivedassaferpre-COVIDbut weredisproportionatelyaffectedbythepandemic. 14

The results are presented in Table 5. The dependent variables are multiplied by 100, so the coefficients are estimates of the effect on the different performance variables in percentage points. In specifications 1, 4, and 7, we consider recourse without the addition of confounding factors such as the interest rate spread and LTV. In specifications 2, 5, and 8, we add in the LTV and interest rate spread. In specifications 3, 6, and 9, we include the indicator for whether the internal risk rating is equivalent to that of an investment-grade credit (rated BBB or higher) and the loan expected loss (probability of default times loss given default). All of these controls are also interacted with the indicator for whether the loan-quarter observation is from the COVID period, thus allowing the effect of controls to vary between the two periods. The results in columns (1) to (3) of Table 5 imply that recourse loans are modestly less likely to be modified on average but were much less likely to be modified than nonrecourse during COVID. In column (1), recourse lowers the likelihood of modification by 0.17 percent in normal times, relative to an average modification rate of 1.5 percent. This estimate implies that recourse loans received modifications about 9 times for every 10 modifications on observably similar non-recourse loans. However, during COVID, the modification rate on recourse loans was 2.4 percent lower than that on non-recourse loans, relative to an overall modification rate of 5.4 percent. This finding implies that non-recourse loans had a modification rate of almost 7.3 percent, while recourse loans had amodificationrateofjustunder4.9percent. Inotherwords, recourseloansreceivedabout 2 modifications for every 3 modifications on observably similar non-recourse loans.32 The coefficient estimates become modestly more negative in column (2), where we control for LTV and rate spreads, and in column (3), where we also add information on loan-level risk. These coefficients on the interaction between recourse and COVID imply that recourse loans were modified at about 60 percent of the rate of non-recourse loans during the COVID pandemic. In contrast to modifications, which rose significantly more for non-recourse loans during COVID, the rates of downgrades and delinquencies mostly rose in parallel for recourse and non-recourse loans. Indeed, the coefficients on the interaction between recourse and COVID are statistically insignificant in columns (4)–(9) when downgrades and delinquencies are the dependent variables. In short, recourse reduced the need for banks to modify loans. Although recourse loans received less accommodation from banks, their performance did not disproportionately suffer during COVID, suggesting 32The predicted effect of recourse during COVID comes from β +β . The modification rate for 1 2 recourse loans relative to non-recourse loans comes from solving the equations NR−R = β +β and 1 2 .78R+.22NR = 5.4 for NR and R, where .78 (.22) is the share of originations that have recourse (no recourse). The predicted effect during normal times is calculated analogously, but with the difference betweenrecourseandnon-recoursebeing β andanaverageoverallmodificationrateof1.5. 1 15

that recourse motivated borrowers to maintain payments. To better understand the timing of the results, we show the predicted effects of recourse on these outcomes on a quarter-by-quarter basis in Figure 1. For this figure, we regress each outcome variable on indicators for whether the loan has recourse, including the same controls and fixed effects as in column (2) of Table 5 (that is, controlling for LTV and interest rate spreads, but not banks’ internal risk measures). The analysis is run separately for each quarter and thus generates quarter-specific estimates of the effects of recourse on the different performance measures. For example, the top chart is a plot of the expected modification rate in a quarter if every loan had recourse (the dashed line) or if no loans had recourse (the solid line), holding all other characteristics fixed. The gap between the lines is the quarterly estimate of the effect of recourse on whether the loan is modified. Other charts perform the same exercise except that the dependent variable is one of the other outcome measures (distress or downgrades). Figure 1 makes apparent that modifications, downgrades, and distress all increased during COVID, with modification and downgrade rates peaking in 2020:Q2. The figure also clarifies the economic impact of recourse. If all loans had been non-recourse, we would have expected an overall loan modification rate of 12 percent in 2020:Q2, compared to a rate of 6 percent if all loans had been recourse. While modification rates dropped in 2020:Q3 and 2020:Q4, the relative difference between recourse and non-recourse loans remains stable, with estimated modification rates for recourse loans remaining at about half the level of non-recourse loans. Figure 1 also clarifies our identification strategy. Broadly, there is little signal in the pre-COVID period, as the coefficients on recourse across loan performance measures are typically modest and frequently switch signs. Though non-recourse loans exhibited higher distress in the early aftermath of the financial crisis, there were no notably different levels or trends for recourse versus non-recourse loans leading into the pandemic. As recourse and non-recourse loans were on similar trends, and the pandemic presented an unexpected disruption in cash flows likely to be independent of at-origination recourse decisions, the interaction between recourse and COVID should cleanly identify the effect of recourse on loan performance. Altogether, the key takeaway from our analysis is that modifications were much lower on recourse loans during COVID. The 50 percent lower modification rate on recourse loanscouldhavepotentiallybeenafunctionoftheseloanshavingbetterperformancethan non-recourse loans. Yet, downgrades and distress were not lower for loans with recourse during COVID, which effectively rules out this hypothesis. In turn, our interpretation is that borrowers with recourse had significantly less bargaining power in loan modification decisions than those without recourse. We test this hypothesis further in the next 16

subsection. 5.3. Recourse Loans Receive More Lender-Friendly Modifications To further examine whether recourse provides lenders with bargaining power in loan modification negotiations, we test whether recourse loans receive more lender-friendly modifications when they do get modified. For this exercise, we limit our sample to loans that received a modification and then run regressions described in equation (2), except with an indicator for the type of modification as the dependent variable. In Table 6, we provide information on the composition of modifications by recourse status both before (2012-2019) and during (2020) the COVID period.33 The most common modification type before 2020 was an extension. Over 50 percent of all pre-2020 modifications involved an extension. The next most common modification type was an increase in the committed balance (or forbearance), which occurred in 27 percent of modifications. However, in 2020, forbearance became the most common modification type, representing almost 50 percent of all modifications. The share of modifications involving extensions dropped to 36 percent. The rates of other modification types remained similar before and during COVID, with the exception of new troubled debt restructurings (TDR), which declined. This decline is almost certainly due to the interagency regulatory guidance that short-term modifications due to COVID did not qualify as TDRs. We evaluate the three most common types of modification in the regressions: forbearance (an increase in a committed balance), a change to interest only for previously amortizing loans, and an extension. Table 7 shows that modifications of recourse loans were less likely to include forbearance and more likely to involve a switch to interest-only amortization. Each dependent variable is multiplied by 100 so that the coefficients can be interpreted as percentage point effects on the frequency of a modification being of a particular type. The coefficient on the interaction term indicates that in 2020 recourse loans were 7 percentage points less likely to receive a forbearance and almost 8 percentage points more likely to switch to interest-only amortization when modified. These estimates are little changed by the inclusion of additional controls for loan risk. Differences in modification frequencies by recourse status are generally small during normal times and small for other types of modifications. Overall, we interpret these results as further evidence of recourse providing lenders with bargaining power in loan modification negotiations. During the COVID period, recourse loan modifications were more likely to involve loans becoming non-amortizing (interest-only payments), while non-recourse loan modifications were more likely to entail 33Modificationtypesarenotmutuallyexclusive(forexample,aloancouldbeextendedandtransitionto interestonlysimultaneously),sothesepercentagesadduptomorethan100. 17

loans becoming temporarily negatively amortizing (interest payments applied to the loan balance). In other words, recourse loans saw a smaller decline in required payments. As borrowers differed little in terms of downgrade or delinquency rates, the results indicate that borrowers differed in their bargaining power as opposed to the degree of stress due to COVID. 6. CONCLUSION We examine the value of recourse in CRE loan contracts. We show that recourse is valued by banks: the presence of recourse in the loan contract at origination reduces loan rate spreads for mortgages on stabilized properties by at least 20 basis points and acts as a substitute for traditional equity, enabling property investors to borrow at LTVs that are about 3 percentage points higher than they could otherwise. We then demonstrate one way in which recourse is valuable to banks. Recourse loans were half as likely as non-recourse loans to receive a loan modification during the COVID-19 pandemic, and the modifications that were made to recourse loans were more lender-friendly. This dynamic occurred despite recourse loans facing similar rates of downgrades and distress, implying that recourse increased lenders’ bargaining positions in loan renegotiations. References Adelino, Manuel, Kristopher Geradi, and Paul Willen (2013). “Why don’t lenders renegotiate more home mortgages? Redefaults, self-cures and securitization.” Journal of Monetary Economics, 60(7), pp. 835–853. doi:10.1016/j.jmoneco.2013.08.002. Ambrose, Brent W. and Anthony B. Sanders (2003). “Commercial mortgage-backed securities: prepayment and default.” The Journal of Real Estate Finance and Economics, 26(2), pp. 179–196. doi:10.1023/A:1022978708728. Beyhaghi, Mehdi (2021). “Third-Party Credit Guarantees and the Cost of Debt: Evidence from Corporate Loans.” Review of Finance. doi:10.1093/rof/rfab012. Binder, Kyle and Jung-Eun Kim (2019). “Recourse and default in bank portfolio CRE lending.” Working Paper. URL https://kylebinder.netlify.app/cv/. Black, Lamont, John Krainer, and Joseph Nichols (2017). “From origination to renegotiation: A comparison of portfolio and securitized commercial real estate loans.” The Journal of Real Estate Finance and Economics, 55(1), pp. 1–31. doi:10.1007/s11146-016-9548-1. 18

Childs, Paul D., Steven H. Ott, and Timothy J. Riddiough (1996). “The value of recourse and cross-default clauses in commercial mortgage contracting.” Journal of Banking & Finance, 20(3), pp. 511–536. doi:10.1016/0378-4266(95)00008-9. Corbae, Dean and Erwan Quintin (2015). “Leverage and the foreclosure crisis.” Journal of Political Economy, 123(1), pp. 1–65. doi:10.1086/677349. Ghent, Andra C. and Marianna Kudlyak (2011). “Recourse and residential mortgage default: Evidence from US states.” The Review of Financial Studies, 24(9), pp. 3139–3186. doi:10.1093/rfs/hhr055. Grovenstein, Robert, John Harding, C. F. Sirmans, Sansanee Thebpanya, and Geoffrey Turnbull(2005). “Commercialmortgageunderwriting: Howwelldolendersmanagethe risks?” Journal of Housing Economics, 14(4), pp. 355–383. doi:10.1016/j.jhe.2005.09.003. Harrison, D. M., T. G. Noordewier, and A. Yavas (2004). “Do riskier borrowers borrow more?” Real Estate Economics, 32, pp. 385–411. doi:10.1111/j.1080-8620.2004.00096.x. Lebret, Daniel and Daniel C. Quan (2017). “From delinquency to foreclosure: a model of loan workout.” Working Paper. doi:10.2139/ssrn.2978154. Titman, Sheridan, Stathis Tompaidis, and Sergey Tsyplakov (2005). “Determinants of credit spreads in commercial mortgages.” Real Estate Economics, 33(4), pp. 711–738. doi:10.1111/j.1540-6229.2005.00136.x. 19

Orig. Orig. Rate Floating Loans Share Share Term LTV Prepay IO 1stLien Value Amount Spread Rate (#) of# of$ (Years) (%) Penalty(%) (%) (%) (Mil.$) (Mil.$) (bps) (%) FullSample Overall 85,668 100 100 16 8 17 57 198 71 13 34 99 Recourse 67,139 78 45 9 5 19 56 193 75 7 32 99 NoRecourse 18,529 22 55 43 20 10 58 216 54 34 41 97 Industrial Overall 5,634 100 100 14 8 8 58 233 53 16 52 97 Recourse 4,461 79 49 9 5 8 58 234 57 11 50 98 NoRecourse 1,173 21 51 34 19 7 59 230 41 35 62 97 Lodging Overall 1,882 100 100 46 21 7 57 264 42 26 63 98 Recourse 1,243 66 37 22 12 7 57 258 37 19 66 97 NoRecourse 639 34 63 93 39 6 56 275 50 39 56 99 Multifamily Overall 50,807 100 100 10 6 23 56 174 83 8 20 99 Recourse 41,665 82 52 7 4 26 55 168 88 3 19 100 NoRecourse 9,142 18 48 27 15 12 59 199 61 29 25 96 Office Overall 9,620 100 100 38 18 7 60 228 54 26 56 98 Recourse 6,689 70 31 15 8 7 60 230 56 16 51 98 NoRecourse 2,931 30 69 89 40 6 59 223 50 49 67 98 Retail Overall 11,434 100 100 18 8 8 57 228 55 17 51 98 Recourse 8,697 76 51 9 5 8 57 226 57 12 51 98 NoRecourse 2,737 24 49 45 15 8 58 232 48 31 52 98 ummary tatistics for oans at rigination by ecourse Table 1: S S L O R . Notes: This table reports summary statistics for loans at origination for the full sample and conditional on whether the lender has recourse. We first present summary statistics for loans that are secured by all stabilized properties and then for the five largest property types in our sample. All averages are unweighted. Source: Authors’ calculations using the Y-14 H.2 Schedule. 20

Quintile 1 2 3 4 5 Across Lenders Full Sample 14 45 61 80 93 Industrial 12 53 68 85 93 Lodging 15 42 59 78 97 Multifamily 8 47 63 82 94 Office 15 45 65 81 93 Retail 18 53 68 87 97 Across States Full Sample 56 63 68 74 83 Industrial 63 73 76 80 89 Lodging 49 64 73 79 86 Multifamily 49 58 67 75 85 Office 56 68 72 79 86 Retail 64 73 78 82 88 ariation in ecourse cross enders and tates Table 2: V R A L U.S. S . Notes: This table reports unweighted recourse shares at banks split into quintiles for the full sample and for the property types shown in Table 1. Each quintile has about one-fifth of banks (top panel) or states (bottom panel). The banks and states in a given quintile can vary for each property type. Source: Authors’ calculations using the Y-14 H.2 Schedule. 21

EffectonRateSpreads EffectonLTV (percentagepoints) (percentagepoints) (1) (2) (3) (4) (5) (6) (7) (8) Recourse -0.204∗∗∗ -0.206∗∗∗ -0.516∗∗∗ -0.195∗∗∗ 2.779∗∗∗ 2.848∗∗∗ 3.437∗∗ 2.733∗∗∗ (0.0298) (0.0293) (0.140) (0.0271) (0.846) (0.856) (1.713) (0.866) LTV 0.000579 0.000933∗∗∗ -0.00112∗∗ (0.000344) (0.000356) (0.000415) InterestRateSpread 0.337 0.365 -0.623∗∗∗ (0.205) (0.226) (0.202) BorrowerRatedBBB+ -0.154∗∗∗ -8.937∗∗∗ (0.0555) (2.238) ExpectedLoss 0.0975∗∗∗ 0.973∗∗∗ (0.0123) (0.185) ln(OriginationAmount) -0.163∗∗∗ -0.165∗∗∗ -0.196∗∗∗ -0.161∗∗∗ 3.124∗∗∗ 3.179∗∗∗ 3.239∗∗∗ 2.921∗∗∗ (0.00958) (0.0101) (0.0156) (0.00907) (0.387) (0.378) (0.442) (0.298) ln(MaturityinYears) -0.267∗∗∗ -0.267∗∗∗ -0.277∗∗∗ -0.243∗∗∗ 0.127 0.216 0.242 0.512 (0.0299) (0.0298) (0.0284) (0.0260) (0.331) (0.299) (0.305) (0.316) Cross-Collateralized -0.112∗ -0.114∗ -0.111∗∗ -0.113∗ 3.296∗∗ 3.334∗∗ 3.329∗∗ 3.137∗∗ (0.0584) (0.0581) (0.0548) (0.0568) (1.502) (1.498) (1.471) (1.475) IOLoan -0.168 -0.166 -0.217 -0.168 -2.322∗∗∗ -2.265∗∗∗ -2.163∗∗∗ -2.622∗∗∗ (0.119) (0.120) (0.142) (0.114) (0.662) (0.667) (0.772) (0.877) FloatingRate 0.579 0.579 0.618 0.607 -0.0873 -0.282 -0.372 1.082∗∗ (0.385) (0.385) (0.387) (0.373) (0.164) (0.247) (0.362) (0.457) FirstLien -0.337∗∗∗ -0.344∗∗∗ -0.330∗∗∗ -0.304∗∗∗ 13.17∗∗∗ 13.29∗∗∗ 13.26∗∗∗ 12.26∗∗∗ (0.0979) (0.0978) (0.0948) (0.0990) (2.475) (2.494) (2.478) (2.510) PrepaymentPenalty 0.169∗∗ 0.168∗∗ 0.177∗∗ 0.171∗∗ 1.569∗∗∗ 1.512∗∗∗ 1.489∗∗∗ 1.318∗∗∗ (0.0697) (0.0700) (0.0714) (0.0734) (0.387) (0.400) (0.426) (0.419) Industrial 0.107 0.108 0.125∗∗ 0.106 -1.447∗∗ -1.483∗∗ -1.518∗∗ -1.605∗∗ (0.0667) (0.0667) (0.0608) (0.0658) (0.635) (0.643) (0.638) (0.650) Lodging 0.448∗∗∗ 0.453∗∗∗ 0.482∗∗∗ 0.428∗∗∗ -8.174∗∗∗ -8.325∗∗∗ -8.387∗∗∗ -8.527∗∗∗ (0.0577) (0.0571) (0.0471) (0.0604) (0.950) (0.912) (0.909) (0.943) Office 0.125∗∗ 0.126∗∗ 0.148∗∗∗ 0.119∗∗ -1.402 -1.444 -1.489∗ -1.670 (0.0551) (0.0552) (0.0511) (0.0535) (0.915) (0.916) (0.892) (1.043) Retail 0.0860 0.0879 0.108∗ 0.0889 -3.281∗∗∗ -3.310∗∗∗ -3.348∗∗∗ -3.227∗∗∗ (0.0631) (0.0627) (0.0600) (0.0609) (0.555) (0.561) (0.558) (0.564) N 85,668 85,668 85,668 84,956 85,668 85,668 85,668 84,956 R2 0.44 0.44 0.42 0.45 0.20 0.20 0.20 0.25 LenderFixedEffects Y Y Y Y Y Y Y Y Orig.YearFixedEffects Y Y Y Y Y Y Y Y CBSA×StateFixedEffects Y Y Y Y Y Y Y Y IV - - Y - - - Y ffect of ecourse on ate preads and Table 3: E R R S LTV. Notes: Columns 1–4 present coefficients from regressing loan rate spreads on an indicator for whether the loan has recourse, while columns 5–8 present coefficients from regressing LTV at origination on recourse. All specifications include controls for size, term, cross-collateralization, amortization, interest rate variability, lien priority, and prepayment penalties, as well as bank, property type, origination year, and CBSA-state fixed effects. Columns 2–4 and 6–8 additionally control for LTV or loan rate spreads, respectively. In addition, columns 4 and 8 layer in risk controls: an indicator for whether the internal risk rating is equivalent to that of an investment-grade credit (rated BBB or higher) and the loan expected loss (probability of default times loss given default). Column 3 instruments for both recourse and LTV with the average value for the recourse indicator and LTV of other loans within the same bank-state-year-property type. Column 7 instruments for recourse with the share of other loans within the same bank-state-year-property type that have recourse. Standard errors, in parentheses, are clustered at the bank level. *,**,*** indicate significance at the 10%, 5%, and 1% levels, respectively. Source: Authors’ calculations using the Y-14 H.2 Schedule. 22

Obs. Modified Downgraded Distressed (#) (%) (%) (%) FullSample Pre-COVID 513,127 1.48 2.84 0.37 COVID 72,580 5.41 6.05 0.94 Industrial Pre-COVID 51,913 1.36 2.69 0.37 COVID 8,112 3.90 2.81 0.85 Lodging Pre-COVID 24,793 2.15 2.90 0.58 COVID 3,237 11.80 22.00 2.81 Multifamily Pre-COVID 135,162 1.39 2.90 0.29 COVID 18,550 4.82 5.48 0.65 Office Pre-COVID 104,587 1.84 2.92 0.51 COVID 14,469 5.90 4.49 0.70 Retail Pre-COVID 133,009 1.30 2.90 0.32 COVID 21,059 5.23 6.92 1.09 ummary tatistics for oans that are odified owngraded or is Table 4: S S L M , D , D tressed . Notes: This table reports the percentages of loans that are modified, downgraded, or distressed across the pre-COVID (2012-2019) and COVID (2020) time periods for the full sample and for the property types listed in Table 1. Source: Authors’ calculations using the Y-14 H.2 Schedule. 23

Modified Downgraded Distressed (1) (2) (3) (4) (5) (6) (7) (8) (9) Recourse -0.167∗∗∗ -0.184∗∗∗ -0.195∗∗∗ -0.0511 -0.112∗ -0.250∗∗∗ -0.00505 -0.0202 -0.0452 (0.0636) (0.0640) (0.0638) (0.0643) (0.0645) (0.0661) (0.0392) (0.0389) (0.0344) Recourse×COVID -2.246∗∗∗ -2.424∗∗∗ -2.440∗∗∗ 0.0581 -0.0780 -0.108 0.0611 0.0481 0.0296 (0.368) (0.370) (0.369) (0.288) (0.288) (0.288) (0.172) (0.173) (0.126) LTV 0.526∗∗∗ 0.300∗∗ 1.802∗∗∗ 0.0881 0.473∗∗∗ -0.241∗∗ (0.129) (0.129) (0.138) (0.143) (0.114) (0.100) LTV×COVID 4.224∗∗∗ 3.608∗∗∗ 2.922∗∗∗ 0.219 1.593∗∗∗ -0.156 (0.565) (0.566) (0.541) (0.553) (0.428) (0.308) InterestRateSpread 0.295∗∗∗ 0.207∗∗∗ 0.114∗∗∗ -0.290∗∗∗ 0.562∗∗∗ 0.244∗∗∗ (0.0379) (0.0373) (0.0327) (0.0345) (0.0521) (0.0392) InterestRateSpread×COVID -0.480∗∗∗ -0.557∗∗∗ -0.341∗∗∗ -0.623∗∗∗ 0.191 -0.217∗∗ (0.140) (0.139) (0.120) (0.126) (0.126) (0.0854) BorrowerRatedBBB+ -0.0684 -3.572∗∗∗ 0.257∗∗∗ (0.0439) (0.0547) (0.0279) BorrowerRatedBBB+×COVID -0.939∗∗∗ -4.165∗∗∗ 0.0630 (0.226) (0.193) (0.0801) ExpectedLoss 0.248∗∗∗ 0.519∗∗∗ 0.997∗∗∗ (0.0205) (0.0297) (0.0488) ExpectedLoss×COVID 0.0501 0.395∗∗∗ 0.744∗∗∗ (0.0508) (0.0853) (0.0856) N 585,641 585,641 585,641 585,641 585,641 585,641 585,641 585,641 585,641 R2 0.06 0.06 0.07 0.03 0.03 0.05 0.02 0.03 0.19 Year-QuarterFixedEffects Y Y Y Y Y Y Y Y Y LenderFixedEffects Y Y Y Y Y Y Y Y Y Orig.YearFixedEffects Y Y Y Y Y Y Y Y Y CBSA×StateFixedEffects Y Y Y Y Y Y Y Y Y PropertyTypeFixedEffects Y Y Y Y Y Y Y Y Y Loan-levelControls Y Y Y Y Y Y Y Y Y ControlsandFEs×COVID Y Y Y Y Y Y Y Y Y ynamic egressions Table 5: D R . Notes: Each column presents coefficients from regressing whether the loan is modified (columns 1–3), downgraded (columns 4–6), or distressed (columns 7–9) on an indicator for whether the loan has recourse, recourse interacted with whether the loan-quarter observation comes from the COVID period (i.e., 2020), and different sets of controls (also interacted with whether the loan-quarter observation comes from the COVID period). Specifications 1, 4, and 7 include the controls and fixed effects shown in the first specification of Table 3, along with year-quarter fixed effects. Specifications 2, 5, and 8 layer in LTV and the interest rate spread as controls. Specifications 3, 6, and 9 layer in risk controls: an indicator for whether the internal risk rating is equivalent to that of an investment-grade credit (rated BBB or higher) and the loan expected loss (probability of default times loss given default). Standard errors are clustered by loan. *,**,*** indicate significance at the 10%, 5%, and 1% levels, respectively. Source: Authors’ calculations using the Y-14 H.2 Schedule. 24

Obs. Forbearance ToIO Extension TDR WriteDown (#) (%) (%) (%) (%) (%) FullSample Pre-COVID 7,440 26.51 17.27 54.96 2.67 1.61 COVID 3,808 49.68 17.31 36.06 0.66 1.13 Recourse Pre-COVID 4,392 22.61 15.76 59.38 3.51 2.12 COVID 2,335 48.09 12.89 40.69 0.64 1.28 NoRecourse Pre-COVID 3,048 32.12 19.46 48.59 1.48 0.89 COVID 1,473 52.21 24.30 28.72 0.68 0.88 ypes of oan odifications by ecourse Table 6: T L M R . Notes: This table reports the percentages of each modification type that occur across the pre-COVID (2012-2019) and COVID (2020) time periods for the sample of loans that are modified, across all such loans and conditional on whether the lender has recourse. “Forbearance” denotes loans for which the committed balance increases. “To IO” denotes loans switched from being amortizing to being interest only. “Extension” denotes loans for which the maturity date is extended. “TDR” denotes a troubled debt restructuring. “Write Down” denotes that the committed balance falls in tandem with a positive cumulative charge-off. The types of modifications are not mutually exclusive, so percentages may add up to more than 100. Source: Authors’ calculations using the Y-14 H.2 Schedule. 25

Forbearance ToIO Extension (1) (2) (3) (4) (5) (6) (7) (8) (9) Recourse 0.227 0.0799 0.0965 0.387 0.308 0.453 -0.297 -0.291 0.138 (1.292) (1.296) (1.291) (0.975) (0.979) (0.980) (1.305) (1.303) (1.277) Recourse×COVID -7.050∗∗∗ -7.396∗∗∗ -7.016∗∗∗ 7.846∗∗∗ 7.949∗∗∗ 7.824∗∗∗ -1.014 -0.720 -0.992 (2.202) (2.190) (2.188) (1.700) (1.703) (1.701) (2.101) (2.101) (2.085) LTV 11.37∗∗∗ 11.39∗∗∗ -3.094 -1.884 -12.08∗∗∗ -8.825∗∗∗ (2.931) (2.950) (2.046) (2.054) (2.866) (2.853) LTV×COVID 6.136 8.724∗ -0.726 -0.111 0.784 -2.330 (5.108) (5.223) (3.534) (3.600) (4.838) (4.980) InterestRateSpread 0.622 0.720 -1.534∗∗∗ -1.147∗∗ -2.389∗∗∗ -1.105∗ (0.673) (0.687) (0.469) (0.489) (0.665) (0.653) InterestRateSpread×COVID -3.993∗∗∗ -3.414∗∗∗ 1.180 0.970 4.314∗∗∗ 3.225∗∗∗ (1.129) (1.110) (0.775) (0.785) (1.098) (1.054) BorrowerRatedBBB+ -3.437∗∗∗ 3.579∗∗∗ 0.381 (1.190) (0.922) (1.208) BorrowerRatedBBB+×COVID 1.649 -0.0315 -3.025 (2.059) (1.530) (1.985) ExpectedLoss -0.294∗∗∗ -0.0963 -1.033∗∗∗ (0.0872) (0.0633) (0.102) ExpectedLoss×COVID -0.588∗∗∗ 0.105 0.653∗∗∗ (0.157) (0.115) (0.196) N 11,248 11,248 11,248 11,248 11,248 11,248 11,248 11,248 11,248 R2 0.50 0.51 0.51 0.55 0.55 0.56 0.51 0.51 0.52 Year-QuarterFixedEffects Y Y Y Y Y Y Y Y Y LenderFixedEffects Y Y Y Y Y Y Y Y Y Orig.YearFixedEffects Y Y Y Y Y Y Y Y Y CBSA×StateFixedEffects Y Y Y Y Y Y Y Y Y PropertyTypeFixedEffects Y Y Y Y Y Y Y Y Y Loan-levelControls Y Y Y Y Y Y Y Y Y ControlsandFEs×COVID Y Y Y Y Y Y Y Y Y ynamic egressions for ypes of oan odifications Table 7: D R T L M . Notes: Each column presents coefficients from regressing the type of loan modification on LTV, an indicator for whether the loan has recourse, recourse interacted with whether the observation comes from the COVID period (i.e., 2020), and controls and fixed effects (also interacted with whether the loan-quarter observations come from the COVID period). Specifications 1, 4, and 7 include the controls and fixed effects shown in the first specification of Table 3, along with year-quarter fixed effects. Specifications 2, 5, and 8 layer in LTV and the interestratespreadascontrols. Specifications3, 6, and9layerinriskcontrols: anindicator for whether the internal risk rating is equivalent to that of an investment-grade credit (rated BBB or higher) and the loan expected loss (probability of default times loss given default). “Forbearance” denotes loans for which the committed balance increases. “To IO” denotes loans switched from being amortizing to being interest only. “Extension” denotes loans for which the maturity date is extended. Standard errors are clustered by loan. *,**,*** indicate significance at the 10%, 5%, and 1% levels, respectively. Source: Authors’ calculations using the Y-14 H.2 Schedule. 26

Modification snaoL fo tnecreP 0.51 0.21 0.9 0.6 0.3 0.0 Non-recourse Recourse Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 2012: 2013: 2014: 2015: 2016: 2017: 2018: 2019: 2020: Downgrade Distress snaoL fo tnecreP 0.51 0.21 0.9 0.6 0.3 0.0 Non-recourse Recourse Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 2012: 2013: 2014: 2015: 2016: 2017: 2018: 2019: 2020: snaoL fo tnecreP 0.3 5.2 0.2 5.1 0.1 5.0 0.0 Non-recourse Recourse Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 Q1 2012: 2013: 2014: 2015: 2016: 2017: 2018: 2019: 2020: verage redicted ffects Figure 1: A P E . Notes: This figure shows predicted values from regressing whether the loan is modified (top), downgraded (bottom left), or distressed (bottom right) on whether the loan has recourse with the controls and fixed effects from specification 2 in Table 5, run on a quarter-by-quarter basis. The data from this figure are presented in Appendix Table A.3. Source: Authors’ calculations using the Y-14 H.2 Schedule. 27

A. SUPPLEMENTAL TABLES This section includes supplemental tables referenced in the text. 1

Orig. Orig. Rate Floating Loans Share Share Term LTV Prepay IO 1stLien Value Amount Spread Rate (#) of# of$ (Years) (%) Penalty(%) (%) (%) (Mil.$) (Mil.$) (bps) (%) Transitional Overall 57,035 100 100 23 12 7 64 271 28 62 73 96 Recourse 39,283 69 66 21 11 7 65 278 27 64 79 96 NoRecourse 17,752 31 34 27 13 8 63 255 30 58 60 96 ummary tatistics for oans ecured by ransitional roperties Table A.1: S S L S T P . Notes: This table presents summary statistics for loans at origination secured by transitional properties for the full sample and conditional on whether the lender has recourse. All averages are unweighted. Source: Authors’ calculations using the Y-14 H.2 Schedule. 2

Orig. Orig. Rate Floating Loans Share Share Term LTV Prepay IO 1stLien Value Amount Spread Rate (#) of# of$ (Years) (%) Penalty(%) (%) (%) (Mil.$) (Mil.$) (bps) (%) FullSample Overall 58,024 100 100 18 9 17 55 197 78 14 32 99 FullRecourse 36,629 55 27 9 4 16 55 205 73 8 36 98 PartialRecourse 15,335 23 16 12 6 23 52 182 87 7 27 100 NoRecourse 15,017 22 57 50 23 10 57 207 63 40 37 97 Industrial Overall 3,767 100 100 17 9 8 56 220 63 17 49 98 FullRecourse 3,239 67 35 8 4 8 56 223 60 11 50 96 PartialRecourse 749 16 14 19 8 9 53 220 73 13 46 99 NoRecourse 830 17 51 46 25 7 57 214 56 43 56 97 Lodging Overall 1,158 100 100 49 23 7 56 247 49 29 63 97 FullRecourse 896 58 25 18 9 7 53 238 41 17 65 94 PartialRecourse 153 10 12 41 25 6 58 256 42 27 87 97 NoRecourse 503 32 63 91 40 7 55 263 59 46 59 99 Multifamily Overall 35,125 100 100 12 6 24 53 183 89 9 19 99 FullRecourse 17,905 48 26 7 3 24 54 187 89 4 21 100 PartialRecourse 11,721 31 22 9 5 28 51 172 96 3 18 100 NoRecourse 7,730 21 52 29 16 13 58 195 65 32 22 96 Office Overall 6,412 100 100 44 20 7 58 217 62 28 55 98 FullRecourse 4,852 59 20 13 7 7 59 221 60 14 50 97 PartialRecourse 972 12 9 27 15 8 56 212 61 22 60 98 NoRecourse 2,346 29 71 110 47 6 58 214 60 58 64 98 Retail Overall 8,095 100 100 19 7 8 56 215 62 17 48 98 FullRecourse 7,002 67 38 8 4 7 56 218 58 10 50 98 PartialRecourse 1,412 14 16 16 9 8 56 210 62 18 57 98 NoRecourse 2,017 19 46 58 18 8 56 219 64 39 46 98 ull and artial ecourse ummary tatistics Table A.2: F P R S S . Notes: This table presents summary statistics for loans originated in 2015 or later across all such loans and conditional on whether the lender has recourse. We first present summary statistics for loans that are secured by all stabilized properties and then for the five largest property types in our sample. All averages are unweighted. Source: Authors’ calculations using the Y-14 H.2 Schedule. 3

ModificationRate(%) DowngradeRate(%) DistressedRate(%) Non-recourse Recourse Non-recourse Recourse Non-recourse Recourse 2012 Q1 2.64 3.12 2.48 3.00 0.49 0.18 Q2 1.98 2.06 1.90 1.86 2.07 1.07 Q3 2.49 2.06 1.89 1.40 1.47 0.83 Q4 1.75 2.40 2.07 2.06 1.08 0.47 2013 Q1 1.41 1.02 2.69 2.30 0.64 0.49 Q2 2.20 0.90 2.20 2.13 0.56 0.32 Q3 2.34 2.41 6.63 6.42 0.49 0.33 Q4 0.86 1.03 2.08 2.47 0.24 0.31 2014 Q1 0.94 1.27 1.52 1.92 0.46 0.14 Q2 0.98 1.89 2.58 1.72 0.30 0.13 Q3 0.88 1.09 6.20 5.64 0.15 0.32 Q4 1.58 0.87 3.09 2.40 0.26 0.20 2015 Q1 1.58 1.17 2.39 2.27 0.24 0.20 Q2 1.62 0.67 2.34 3.05 0.29 0.28 Q3 1.07 0.89 2.64 3.07 0.26 0.27 Q4 1.02 0.86 4.90 4.10 0.03 0.34 2016 Q1 1.50 1.23 2.05 2.29 0.24 0.29 Q2 1.21 1.36 2.74 2.57 0.17 0.17 Q3 1.42 1.35 3.13 2.63 0.15 0.26 Q4 1.76 1.14 2.09 2.59 0.13 0.25 2017 Q1 4.00 1.91 1.96 1.99 0.41 0.23 Q2 1.76 1.25 3.42 3.48 0.26 0.34 Q3 1.34 1.15 1.77 1.72 0.41 0.29 Q4 1.37 1.16 3.01 3.57 0.18 0.45 2018 Q1 2.29 1.34 2.24 2.65 0.27 0.48 Q2 0.62 1.46 3.90 4.60 0.25 0.54 Q3 0.91 1.19 1.84 3.02 0.32 0.49 Q4 1.69 0.93 3.06 3.58 0.33 0.56 2019 Q1 1.07 1.12 2.56 2.53 0.34 0.46 Q2 0.48 2.26 1.83 2.71 0.45 0.44 Q3 1.34 1.09 2.36 2.45 0.49 0.48 Q4 4.04 3.45 2.61 2.37 0.48 0.60 2020 Q1 5.36 5.50 3.98 3.61 0.72 0.55 Q2 11.96 6.18 11.01 9.39 0.90 0.79 Q3 7.90 4.76 5.68 6.66 0.98 1.05 Q4 5.61 2.27 4.40 4.75 1.10 1.46 alues for igure Table A.3: V F 1. Notes: This table presents the values for the lines plotted in Figure 1. Source: Authors’ calculations using the Y-14 H.2 Schedule. 4