Pretend or Amend? On Evergreening in CRE

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Pretend or Amend? On Evergreening in CRE David Glancy 2026-025 Please cite this paper as: Glancy, David (2026). “Pretend or Amend? On Evergreening in CRE,” Finance and EconomicsDiscussionSeries2026-025. Washington: BoardofGovernorsoftheFederalReserve System, https://doi.org/10.17016/FEDS.2026.025. 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.

Pretend or Amend? On Evergreening in CRE* David Glancy† FederalReserveBoard May 4, 2026 Abstract Loan modifications can either amplify or mitigate credit losses depending on the strategy lenders employ. Using detailed supervisory data and a model incorporating various frictions that could encourage modifications (liquidity constraints, foreclosure costs, and loss recognitioncosts),IassesswhybanksextendCREloans. Ifindthatextensionspredominantlyaddress temporarypaymentfrictions,bothinnormaltimesandfollowingtheSpring2023bankstress episode. Contrarytoconcernsaboutbanks“extending-and-pretending”followingthatepisode, banks increased income and principal paydown requirements for extensions, contributing to strongex-postperformanceforextendedloans. Keywords: commercialrealestate,banks,evergreening JELClassification: E44,G21,R33 *I thank Martin Kornejew (discussant), Felicia Ionescu, Sam Hughes, Robert Kurtzman, Jose-Luis Peydro, and seminarparticipantsattheCentralBankofIreland-UCD-CEPRConferenceonMacro-financeandFinancialStability Policies;UMassBoston;andtheFederalReserveBoardforhelpfulcomments. Theviewsexpressedinthispaperare solely those of the author and do not necessarily reflect the opinions of the Federal Reserve Board or anyone in the FederalReserveSystem †PrincipalEconomist,DivisionofMonetaryAffairs,FederalReserveBoard,david.p.glancy@frb.gov

1. INTRODUCTION In many models of financial intermediation, the defining feature of bank credit is greater flexibility to renegotiate loan terms (see, for example, Rajan, 1992 and Hackbarth et al., 2007). Loan modificationscanreducebanks’creditlossesbyreplacingmorecostlyresolutionmethods(Bolton and Scharfstein, 1996).1 However, this flexibility can be a double-edged sword; bank actions to hide impairment can produce credit misallocation (Peek and Rosengren, 2005; Caballero et al., 2008), financial stability risks (Bruche and Llobet, 2014), and economic sclerosis (Acharya et al., 2021). Deciphering why banks modify CRE loans has become particularly important in recent years, as CRE market strains have created a need for loss-mitigating modifications at the same time that banking sector strains potentially motivated loss-obscuring ones. While communications from regulators have emphasized the benefits of working proactively with stressed borrowers (Federal ReserveSystemetal.,2023),othershavenotedtheriskofextend-and-pretendbehaviorthatcould harmbanksdowntheroad(Jiangetal.,2025;CrosignaniandPrazad,2024). In this paper, I investigate whether extension practices are more consistent with banks restructuring loans to have favorable future repayment prospects or merely extending them to delay loss recognition. I begin by presenting a model of maturity extensions that incorporates various motivations banks might have for providing extensions. In the model, banks may extend loans to delay loss recognition (Crosignani and Prazad, 2024), avoid deadweight insolvency costs (Faria-e Castro et al., 2024), or give borrowers more time to find a suitable buyer (Sagi, 2021). In deciding the extension terms to offer, banks weigh these benefits against debt overhang costs from undercapitalizedborrowersfailingtomaintaintheproperty(Myers,1977). The model demonstrates that lenders’ motivation for providing extensions can be inferred by the relationship between a loan’s debt yield—net operating income (NOI) as a share of the loan 1Thisabilitytomitigatelossesbyrenegotiatingloantermscan,inturn,accountfortheselectionofsmaller(Hackbarth et al., 2007), riskier (Black et al., 2020), or more liquidity constrained (Glancy et al., 2025) borrowers into banks. 2

balance—and the principal paydown required for an extension. Lenders looking to delay loss recognition need to provide subsidized terms to highly stressed borrowers to motivate those borrowers to extend rather than default. Borrowers with weak incomes therefore receive forbearance fromrequiredpayments. Incontrast,borrowers’participationconstraintsarenotbindingforextensions that remedy frictions in selling or refinancing a financially viable property. Those extension terms are determined by lenders’ desired credit enhancements rather than borrowers’ willingness to extend, meaning that weak income loans require larger paydowns to mitigate repayment risks. In short, risky extensions to avoid default costs entail subsidized terms and low debt yields, while saferextensionstoremedytemporarystrainsentaileitherstringenttermsorhighdebtyields. I use detailed supervisory data on bank CRE loan holdings to test which motivation for extending loans can best explain observed extension patterns. I find that extension patterns in normal times (before the pandemic) are consistent with banks addressing temporary payment strains. Namely, the probability of principal repayment declines monotonically in debt yield, meaning the lowest incomeborrowersaremostlikelytorepayprincipal. Moreover,banksprovidematurityextensions throughout the debt yield distribution, not just income-strained loans. Both of these observations areconsistentwithliquidity-drivenextensionsinthemodel. The primary empirical work analyzes how extension patterns changed after the 2023 bank stress episoderelativetonormaltimes. Ifbankingsectorstrainscausedbankstoextend-and-pretend,we would expect to see (i) an increase in extensions, driven by highly strained borrowers, (ii) easier extension terms for those strained borrowers, and (iii) a lower share of extended loans eventually payingoff. I find that none of these three predictions hold in the data, suggesting that large banks continued to efficiently manage CRE risks following the stress episode. First, I examine outcomes of pending maturities to test whether extensions either became more common or shifted towards riskier loans following the 2023 bank stress episode. Between 2023 and 2025, banks extended a large share of loans (roughly half) as they matured. However, this behavior was not unusual; banks 3

extended a similar share of loans before COVID, and even more at the onset of the pandemic. Thus loan extensions are not merely a response to the stress, but a persistent feature of bank loan servicing. Regarding differences by risk, banks reduced extensions for low-debt-yield loans after 2022, the opposite of what the model predicts would have occurred if banking sector pressures encouraged bankstodelaylossrecognition. Bankssimilarlyreducedextensionsfornonrecourseloans,further supporting the idea that extension policies became more conservative in the face of CRE market strains. Second, I test whether banks eased extension terms after 2022. As the model demonstrates, extensions that delay loss recognition entail lenient terms to the lowest quality borrowers because more dramatic accommodation is needed to prevent default. Instead, I find that riskier firms paid forextensionsbyprovidingcreditenhancementsthatimprovebanks’futurereturnprospects. Borrowers were more likely to pay down principal, provide additional guarantees, or accept higher loan spreads for extensions after 2022. Moreover, this tightening in terms was most pronounced for loans with other risk characteristics such as low debt yields, office collateral, or nonrecourse clauses. Third,Iexaminewhetherperformancedeterioratedforextendedloans. Whilethefirstsetofresults demonstrates that the quality of extensions rose on observed margins, it is still possible that banks extended loans with unobserved factors that make repayment unlikely.2 Instead, I find that stressera extensions performed slightly better than prepandemic ones, consistent with extensions going tohigher-qualitypropertiesandhavingenhancementsthatbolsterfuturerepayment.3 While aggregate patterns are inconsistent with lenient extension policies driving pandemic-era extensions, this does not rule out such behavior for some lenders. For the final piece of analysis, 2Forexample,aloanagainstanemptyofficecouldhaveahighdebtyielduntilleasesexpireandtenantscancease payingrent(GlancyandWang,2023). Borrowersmightbewillingtopaydowntheprincipaltoextendtheloanand keepcollectingthosecashflowsevenifpendingvacanciesleavelittlehopefortheloanpayingoff. 3To be more precise, the performance of extended loans improved relative to the broader universe of maturing loans. CREmarketstressesafter2022causedperformancetodeteriorateforextendedandnon-extendedloansalike. 4

I follow Crosignani and Prazad (2024) and examine differences in extension patterns by bank capitalization. I show that banks with low capital ratios behave similarly to the broader sample. If anything, worse-capitalized banks reduced extensions relative to better-capitalized banks after 2022whiletighteningextensiontermstoasimilarextent. However,estimatesarenoisyduetothe limitednumberofbanksinthesample. 1.1. Relatedliterature This paper contributes to three strands of literature. First, it contributes to work on risks posed to the banking sector by CRE market strains. Though it is well understood that changes in interest ratesandremoteworktendenciesgeneratedsevereCREvaluationdeclines(Guptaetal.,2026),the extentofbanks’exposuretoCRElossesisupfordebate. Jiangetal.(2025)findthatpotentialCRE losses place many small banks at risk of solvency runs. However, realized bank delinquencies are lower than one might expect given the extent of valuation declines (Hinzen et al., 2025), raising the question of why bank CRE loan performance hasn’t deteriorated more. Jiang et al. (2025); Crosignani and Prazad (2024) provide evidence that extend-and-pretend behavior contributes to banks’ relatively modest delinquency rates in the face of these strains. However, Glancy and Kurtzman(2024)findthatmuchofsmallandregionalbanks’strongperformancecanbeattributed toportfoliocomposition—mostnotably,theirminimalholdingsofhigh-riskofficeloans—leaving lessroomforextend-and-pretendbehaviortoexplaindelinquencypatterns. ThispaperrelatesmostcloselytoCrosignaniandPrazad(2024),whichalsousessupervisorydata on large banks’ CRE loan holdings to analyze extend-and-pretend behavior following the pandemic. They show that worse-capitalized banks were more likely to extend loans that suffered income declines. There are two key differences in this study. First, I focus predominantly on the behavior of the sample as a whole rather than differences across banks. This approach allows me toavoidtwodifficultieswithusingcross-lendervariation: limitedstatisticalpowerduetothesmall numberofY-14reporters,andcomplicationsidentifyingaggregateeffectsduetothemissingintercept problem. Though Crosignani and Prazad (2024) provide evidence that capital considerations 5

inducedsomebankstoextendloansonthemargin—inturncrowdingoutnewlending—myresults indicate that extend-and-pretend behavior was small in aggregate. Second, I incorporate information on the terms and ex-post performance of extended loans, which I theoretically demonstrate areinformativeastobanks’underlyingmotivationforextendingloans. Second, this paper relates to a broader literature on evergreening/zombie lending. This work demonstratesthatweaklycapitalizedbanksextendcredittounderperformingfirmstoavoidwriting offexistingloans(PeekandRosengren,2005;Caballeroetal.,2008)andtheresultingdistortionin creditallocationhasnegativemacroeconomicconsequences(Acharyaetal.,2021,2022). Zombie firms are typically defined by having some combination of income strains and subsidized credit (Adalet McGowan et al., 2018; Acharya et al., 2019). This notion aligns well with extend-andpretend modifications in the model, which are characterized by low debt yields and lenient principal repayment requirements. My findings complement Favara et al. (2024), which uses similar data on commercial and industrial lending to show that large U.S. banks do not engage in zombie lending regardless of capitalization. A couple of factors might contribute to the apparent lack of zombie lending in this setting. First, the banks in the sample are generally well-capitalized, and thuslacktheseverestressesandpotentialgamblingforresurrectionincentivesthatwereinplacein episodestypicallyassociatedwithzombielending(i.e.,theJapanesefinancialcrisisandEuropean sovereigndebtcrisis). Second,thebankswestudyaresubjecttostresstestswhichshoulddampen extend-and-pretendincentivessinceprojectedlossesfromstressedextensionswouldaddtobanks’ capitalrequirements.4 Finally, this paper contributes to work analyzing the servicing of distressed CRE loans. Brown et al. (2006) shows that sales of foreclosed CRE properties occur at substantial discounts relative tofundamentalvalues,motivatinglenderstorenegotiateloans. Blacketal.(2017,2020)document that banks have an advantage in renegotiating CRE loans (relative to CMBS), and Glancy et al. (2025) provide evidence that such modifications supported loan performance at the onset of the 4Delaying loss recognition by rolling over risky loans could preserve capital by avoiding losses. However, the expectedlossesfromtheseriskyloansinasevererecessionwouldresultinahigherstresscapitalbuffer,counteracting theabilityofextensionstopreservecapitalbuffers. 6

pandemic.5 This work generally assumes that lenders set modification policies to maximize loan recoveries, but does not touch on distinguishing this motivation from extend-and-pretend considerations. The rest of the paper proceeds as follows: Section 2 presents a model of loan extensions, and derives equilibrium extension terms and maturity outcomes. Section 3 describes the data and methodology. Section4presentstheempiricalfindings. Section5concludes. 2. MODEL 2.1. Setup Toaidintheinterpretationofobservedloanextensionpatterns,IdevelopadynamicmodelofCRE maturity outcomes where borrowers and lenders negotiate extensions to navigate various market frictions. All parties are risk neutral and have a discount factor of β =(1+r) 1. The timing of − the model is as follows: At the end of a period, a nonrecourse loan with an outstanding balance D against a property with NOI N is scheduled to mature. The lender makes an offer to extend the loan for another period, choosing a principal paydown of p D as a condition of the extension. A · valueof p=1signifiesthatthelenderrejectsanextension,demandingfullrepayment,while p<0 signifiesinterestpaymentsgettingpartiallycapitalizedintotheloanbalance. Next, the borrower solicits bids on the property and receives an offer to purchase the property at a cap rate (NOI over property value) of κ.6 Borrowers can therefore sell the property for N/κ, and use the proceeds to pay back the loan and accumulated interest (1+r )D, where r is the m m mortgage rate. Rent is paid after sales occur, and thus current rents are incorporated into property values, making κ (r g)/(1+r) the cap rate that equates sale price to the present discounted ≡ − valueofcashflows,wheregisperperiodexpectedincomegrowth. 5Relatedly,Flynnetal.(2024);DincandYo¨nder(2022)analyzestrategicrenegotiationonthepartofborrowers. 6Drawing a cap rate of κ is identical to pulling a value multiple of 1/κ. I express values in terms of cap rates becauseitcomplementsthefocusondebtyieldsasthemeasureofloanrisk;debtyieldcanbeinterpretedasthecap ratebelowwhichaborrowercouldsellapropertytopayoffaloan’sprincipal.Namely,κ N/V <N/D = V >D, ≡ ⇒ whereV isthesalepriceandN/Dthedebtyield. 7

Iftheborrowerrejectsthesaleoffer,theycantheneitherdefault,andforfeittheproperty,oraccept the extension. If they extend the loan, they collect the income flow N, make the required principal and interest payments (r +p)D, and repeat the game next period with D =(1 p)D and a new m ′ − N which depends on exogenous, stochastic growth (with mean g and standard deviation σ) and ′ anendogenousmaintenancedecision. Thepossibilityoflowerdriftduetodebt-overhang-induced underinvestmentisdiscussedinthenextsection. 2.2. Payoffs I incorporate frictions into the model to account for the various motivations parties might have to extendloansatmaturity. 1. Searchfrictions: κ isstochastic,creatingtheriskthatborrowersreceiveaweakofferwhen their loan comes due. Extensions deal with this liquidity risk by giving borrowers time to shop for a better offer. I assume κ follows a Pareto distribution: G(κ κ κ) = 1 | ≥ − (κ/κ) α,whereα parameterizesmarketliquidity. TheexpectedpurchaseofferisE(N/κ)= [α/(1+α)]N/κ, meaning an expected proportional discount of 1/(1+α) if forced to sell inaparticularperiod.7 2. Liquidity Constraints: Borrowers can only raise outside funds f D to meet interest short- · falls and principal repayments, necessitating a property sale or extension to avoid maturity default. Extensionssuchthat p+r N/D> f areinfeasible,andcausedefaultifborrowers m − cannotsellandpayofftheloan. 3. Foreclosure costs: Expected recovery in foreclosure is ΛN/κ, where Λ 1. Foreclosure ≤ costs create a discontinuous drop in lender payouts at the default threshold. Consequently, extensionsmayreduceexpectedlossesbygivingtheloananopportunitytorecover.8 7WhileIdiscussthisprocessintermsofsearchforabuyer, thismechanismcouldalsocapturesearchforarefinance,withκ reflectingwhetheranewloanofferissufficienttorefinancetheoutstandingloan. 8More formally, the discontinuous drop in loan values at the default threshold (absent an extension) causes loan valuestobeconvexinN. LendersarewillingtoextendloanssinceincreasesinNraiseloanvaluesmorethandeclines reduceit. 8

4. Delayed Loss Recognition: Lenders may face a cost to realizing losses in a given period (e.g., due to equity issuance costs or lost opportunities from binding capital constraints). I incorporatethisasanadditionalcostof χ(D ΛN/κ)thatlendersfaceifborrowersdefault. − The final element that I include in the model is debt overhang problems in the form of deferred maintenance. While the aforementioned elements generate benefits to extensions, endogenizing maintenance introduces a countervailing cost to them. Borrowers are able to receive an additional cash flow νN by neglecting maintenance, but this action reduces N by a factor θ, permanently ′ lowering cash flows. ν and θ are such that the return to proper maintenance is high. However, borrowers that expect to default in the near future will not prioritize future cash flows and try to extractasmuchastheycanpriortodefault. Thismechanismimposesacosttolendersofproviding extensionsagainststressedproperties. Since returns are homogeneous of degree one in D and N, I normalize payouts by D and express allpayoutsintermsofthedebtyieldn N/D.9 Thepayoutsareasfollows: ≡ Table1: Payouts Outcome BorrowersV (n;p,κ) LendersV(n;p) b l Sell n/κ (1+r ) 1+r m m Extend&Maintain n ( − r +p)+β(1 p)E[V (n)] r +p+β(1 p)E[V(n)] m b ′ m l ′ Extend&Neglect (1 − +ν)n (r +p) − +β(1 p)E[V (n)] r +p+β(1 − p)E[V(n)] m b ′ m l ′ − − − Default 0 Λn/κ ∞ whereE[V(n)]= (cid:82) V(Zµn)dF(Z) V(µn)givesexpectedfuturevaluesasafunctionofexpected i ′ i i ≡ 0 future debt yield for i b,l ; µn=n(1+g)(1 θ1[Neglect])/(1 p) gives the expected future ∈{ } − − debt yield, which depends on income drift, principal paydowns and maintenance decisions; and F isalognormalCDFrepresentingproportionaldeviationsfromexpectedNOI. Note that lenders’ value functions omit the potential cost to loss recognition: χ(1 Λn/κ). I − excludethattermsothatthepayoutsareconsistentwiththefunctionsdeterminingthecontinuation 9Note that a value function normalized to D is V(N/D)=V˜(D,N)/D, where V˜ is the pre-normalization value function. This means that paydowns have the effect of creating a normalized continuation value V˜(D,N )/D = ′ ′ (D/D)V(N /D)=(1 p)V(n). ′ ′ ′ ′ − 9

values from extensions. The cost to loss recognition is incorporated as a one-time cost that does not affect continuation values (if loss recognition is costly in the future, there is little benefit to delayingit). 2.3. Strategiesandequilibrium I solve for a Markov perfect equilibrium where borrowers optimally select an action a (n,p,κ) ∗ ∈ Extend Neglect,Maintain ,Default,Sale subject to their liquidity constraint p+r n m { ×{ } } − ≤ f, lenders optimally set paydown requirements p (n), and value functions satisfy the Bellman ∗ equations:     (cid:18) (cid:19) V b (n)=E κ  max  0 , n (1+r m ), n (r m +p ∗ (n))+β(1 p ∗ (n))V b 1+g n  (cid:124)(cid:123)(cid:122)(cid:125) κ − − − 1 p (n)   Default (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) − ∗ (cid:125) Sell Extend-Maintain    (cid:18) (cid:19) (1+g)(1 θ)  (1+ν)n (r +p (n))+β(1 p (n))V − n  − m ∗ − ∗ b 1 p (n)  (cid:124) (cid:123)(cid:122) − ∗ (cid:125)      Extend-Neglect V(n)=max π (n,p)(1+r )+π (n,p)Λn/κ l sale m def p { (cid:16) (cid:17)(cid:111) +π (n,p) r +p+β(1 p)V (cid:0) µ (n,p)n (cid:1) ext m l ∗ − (1) where π (n,p), π (n,p), and π (n,p) are payoff, default and extension probabilities, respecsale def ext tively, and µ (n,p)n is the expected future debt yield conditional on extending. These outcomes ∗ are determined by borrower’s strategy a (n,p,κ). The probabilities are stochastic from lenders’ ∗ perspective because they do not observe κ (see Appendix B.1 for their derivation). µ (n,p) is not ∗ stochasticbecausemaintenancedecisionsdonotdependonκ. ThealgorithmtosolveforthepolicyfunctionsisoutlinedinAppendixB.2. Brieflyput,themodel 10

is solved by guessing the value functions (V and V ), and then (i) solving for borrowers’ optimal b l actionasafunctionofn,pandκ basedontheguessforV ,(ii)findingthelenders’optimal p (n) b ∗ givenborrowers’policies,(iii)updatingthevaluefunctionsbasedonthesebestresponses,and(iv) iteratinguntilvaluefunctionsconverge. 2.4. Graphicalanalysis Here I will graphically characterize how equilibrium outcomes in the model are determined. The expressionsunderlyingthisanalysisarediscussedinAppendixB.3. Starting with the borrower’s decision, there are two key functions that determine whether a borrower is willing to accept an extension, and if so, whether they choose to maintain the property after. First, there is the maximum paydown they are willing to provide on an extension, denoted P (n). This curve is upward sloping since greater cash flows increase funds available to meet b principal and interest payments and raise the likelihood borrowers can profitably sell the property in the future. Second, there is a downward sloping function, denoted M (n), giving the paydown b∗ above which expected future distress is reduced sufficiently to motivate borrowers to maintain the property(iftheyextend). Panel (a) of Figure 1 plots these curves, and shows the associated loan outcomes as a function of n and p. To start, we ignore the potential for property sales, so outcomes pertain to what happens to loans witha high enoughκ draw for asale to be undesirable. Parameter values are set tomatch estimates from other literature, as is discussed in Appendix B.4, and are presented in Table A.1. These parameters entail an expected 7.5% discount if forced to sell immediately (α = 12.3) and 24%foreclosurecosts(Λ=0.76),butnocosttolossrecognition(χ =0). The curves define three regions. Borrowers default when p>P (n) (the red region), because the b principal paydown is more than they are willing or able to provide. They maintain the property when p [M (n),P (n)], meaning paydowns are high enough that borrowers choose to maintain b∗ b ∈ the property, but not so high that they choose to default (the green region). Finally, they neglect 11

the property when p < min M (n),P (n) , meaning that lenient extension policies prevent deb∗ b { } fault but do not leave borrowers sufficiently committed to the property to maintain it (the yellow region). Figure1: EquilibriumMaturityOutcomes 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(nwodyaPlapicnirP 0.25 Regions Default Neglect 0.20 Maintain 0.15 Pb 0.10 0.05 0.00 M b∗ 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) (a)Borrowers’IndifferenceConditions )p(nwodyaPlapicnirP Regions Default Neglect Maintain Pb P l∗ Pl M b∗ (b)Lenders’OptimalityConditions Notes: Panel(a)plotsthepaydownabovewhichborrowersdefault(reddashedline,P ),andtheoneabove b which borrowers maintain the property if they extend (green line, M ). Red, yellow, and green regions b∗ showwhereborrowerswithadebtyieldofnwouldchoosetodefault,neglect,andmaintain,(respectively), given an extension offer of p and a κ draw such that paying off the loan is not optimal. Panel (b) adds the minimum paydown lenders will accept (solid blue line, P ), and lenders’ optimal paydown when not l constrained by borrowers’ default or maintenance decisions (blue dashed line, P ). The colored regions l∗ denotetheoutcomesatagivendebtyieldforlenders’optimalpaydownrate. Turningtolenders’problem,theregionsinpanel(a)determinetheparticipationconstraintslenders face in setting extension terms. The key functions defining lenders’ actions are the minimum paydown lenders are willing to accept to extend a loan, denoted P (n), and the optimal paydown l forlendersthatarenotconstrainedbyborrowers’defaultdecisions,denotedP (n). l∗ Panel (b) adds these curves to the figure. The shaded regions now pertain to outcomes that occur for the equilibrium p= p (n) rather than the p on the y-axis. There are four regions defining loan ∗ outcomes. 1. P (n) < P (n): At these low debt yields, borrowers are not willing to provide a paydown b l acceptabletolenders,resultingindefault(theredregion). 12

2. P (n) [P (n),M (n)]: borrowerswillprovidepaydownsacceptabletolenders,butnothigh b l b∗ ∈ enoughforthemtomaintaintheproperty(theyellowregion). 3. P (n) [M (n),P (n)]: borrowerswillprovidepaydownsthatsolvethedebtoverhangprobb b∗ l∗ ∈ lem,butarestilllessthanlenderswoulddesire(theleft-mostpartofthegreenregion). 4. P (n) > P (n): Borrowers’ participation constraint is not binding, so lenders can achieve b l∗ theiroptimalprincipalrepayment(therestofthegreenregion). Figure2showshowthesemechanicsmapintoobservedoutcomes(whathappensatmaturity,what paydownsoccurforextensions,andwhetherextendedloanseventuallypayoff). Panel(a)presents maturity outcomes (the colored regions) and paydown rates (the black line). It clarifies that there are two distinct debt yield regions that determine extension terms. At low debt yields, lenders are constrained by borrowers’ participation constraint, so extensions are determined by how large a paydownborrowersarewillingandabletomake. Thiscauses p tobeanincreasingfunctionofn, ∗ tracingoutthepartoftheP (n)curvethatsitsabovetheP (n)curve. Athigherdebtyields,lenders b l are not constrained by borrowers’ willingness to pay down a loan, so outcomes are determined by the size of a concession that lenders would like to mitigate risks of future property value declines. In this region, p declines in n because higher incomes reduce the need for paydowns to mitigate ∗ futuredefaultrisk(p tracestheP (n)curve).10 ∗ l∗ Panel (b) shows that p and n are informative as to future repayment prospects. Hollow red dots ∗ show the probability that a loan eventually defaults (potentially after a string of extensions), and solidreddotsshowtheprobabilitythataloanwithagivendebtyieldwoulddefaultiflenderswere unwilling to extend a loan. Extensions reduce the risk of default, but depending on the debt yield and principal paydown, sometimes future default remains highly likely.11 There is a steep drop in futuredefaultriskbeyondthepointthatborrowersbecomewillingtomaintainthepropertyandthe risk of future default becomes very low once borrowers are no longer limited in their willingness 10Iassumethatlenderscannotforceborrowerstodeferinterestpayments,so p is0forthehighdebtyieldssuch ∗ thatlenderswouldpreferborrowerstooperatewithhigherloanbalances. 11ThederivationofeventualdefaultprobabilitiesisinAppendixB.1. 13

Figure2: ObservedExtensionOutcomes 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(etaRnwodyaP 1.0 0.8 p∗ 0.6 0.4 0.2 0.0 ytilibaborPemoctuO 0.25 Outcomes(RightAxis) Default Neglect 0.20 Maintain Payoff 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) (a)EquilibriumPrincipalPaydowns )p(etaRnwodyaP 1.0 0.8 p∗ 0.6 0.4 0.2 0.0 ytilibaborPemoctuO Outcomes(RightAxis) DefaultsAverted EventualDefault NoExtensions (b)EventualLoanPerformance Notes: Panel(a)plotsequilibriumprincipalpaydowns(blackline,leftscale),andtheprobabilitythatvarious maturity outcomes occur (colored regions, right scale). Red, yellow, green, and blue regions show the probability that borrowers would choose to default, neglect, maintain, and pay off a loan (respectively), given an extension offer of p (n). Panel (b) adds information on ultimate performance; hollow red dots ∗ show the probability that a loan with a given n eventually defaults, and solid red dots show the probability that a loan would have defaulted if extensions were not available. The red area shows the decline in the probabilityofeventualdefaultduetoextensions. topaydownloans. In short, even though the future performance of extended loans is not observable in real time, the condition of the property and the terms of the extension are highly predictive of future repayment prospects. If a loan has either a high debt yield or a meaningful principal paydown, it means that the borrower is committed to the property and the loan is likely to pay off once they are able to get a competitive offer to sell or refinance. Extensions with minimal concessions at lower debt yields are the ones that are unlikely to successfully pay off without a loss in the future. In the baselinecalibrationwithnocapitalpreservationincentives,theseextensionsarestillefficientsince thepotentialtoavoidforeclosurecostsisenoughtocompensatefordebtoverhangcosts. However, a byproduct of these actions is still that extensions could mask underlying CRE market strains sincebanksareextendingsomeloansthatwouldlikelydefaultinthefuture. 14

2.5. Whatdrivesextensions? This section investigates how changes in model parameters affect equilibrium maturity and extension outcomes. I start by showing what happens when the benefit to delaying loss recognition (χ) rises from 0. I then discuss model outcomes isolating effects of specific frictions to clarify what typesofextensionsarecausedbyeachfriction. Extend-and-pretendincentives Figure3plotshowincreasing χ from0to0.01affectsmaturity outcomes. The left panel plots the shift in the P (n) curve brought about by the temporary cost l to loss recognition. The paydown that lenders require to extend loans shifts down since lenders are more motivated to prevent default, even if only temporarily. The downward shift is most pronounced at very low debt yields since those loans would post the largest losses without an extension. Fortheparametervaluesinthebaselinemodel,thistemporarycosttolossrecognitionis enoughforlenderstoalwaysbewillingtoofferlenientextensiontermstopreventdefault.12 Therightpaneladdsinformationonprincipalpaydownsandtheeventualprobabilityofrepayment. Totherightofthedashedline,outcomesareallidenticaltothebaselinecalibrationshowninFigure 2.13 New extensions start to occur for the low debt yields documented in the left panel. These extensions are associated with negative paydowns (the black line showing p (n) is below 0) and ∗ virtually no chance of loans ultimately repaying (the red dots showing the probability of eventual default are near 1). Unlike in the baseline calibration, lenders do not extend loans because the chanceofavoidingforeclosurecostsisenoughtooffsetdebtoverhangeffects;lendersbenefitfrom delayingdefaultandarewillingtoextendloanswithnorealprospectforfuturerepayment. 12The baseline model has no role for bank examiners in restricting such extensions from being made, but could easily be amended by providing a floor to p reflecting examiners’ ability to assess whether loans are “restructured with reasonable repayment terms” (Federal Reserve System et al., 2023). Such a change would prevent banks from extendingthemostseverelystressedloans. 13Thisisbecausethetermsofstressedextensionsaredeterminedbythemaximumpaydownborrowerswillaccept. This maximum paydown is unchanged because (i) it is lenders’ payoffs that change and (ii) the shift in lenders’ willingnesstoextendloansistemporaryandthusdoesn’taffectborrowers’expectationsforfutureextensionpolicies. 15

Figure3: Extend-and-PretendIncentives(χ: 0 1%) → 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(etaRnwodyaP NewExtensions Pb Pl(χ=0) Pl(χ=0.01) yradnuoBsuoiverP 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) (a)Lenders’requiredpaydownsshiftdown )p(etaRnwodyaP p∗ yradnuoBsuoiverP 1.0 0.8 0.6 0.4 0.2 0.0 ytilibaborPemoctuO Outcomes(RightAxis) Default Neglect Maintain Payoff EventualDefault (b)Observableoutcomes Notes: Thesolidblueandredlinesinpanel(a)showtheminimumpaydownacceptabletolendersandthe maximum paydown borrowers will make, respectively. The dashed blue line shows how lenders’ required paydownsshiftifχ increasesto0.01. Thehighlightedareademonstratestherangeofdebtyieldsthatnewly receiveextensionsduetothechangeinlenders’objectivefunction. Panel(b)providesoutcomesatmaturity when χ =.01. Red, yellow, green and blue regions represent the probability of default, neglect, maintain, and pay off. The black line gives required pay downs by debt yield (left axis) and red hollow dots the probabilitythataloanultimatelydefaults(rightaxis). Roles of other frictions Figure A.1 presents model outcomes isolating the effects of the three driversofextensions. Panel(a)showsoutcomeswithjustsearchfrictions,panel(b)outcomeswith just foreclosure costs, and panel (c) outcomes with just costs to loss realization. Other parameters arethesameasinthebaselinecalibration. These figures show that the various motivations to extend loans affect different debt yield regions. Panel (a) demonstrates that search frictions only cause extensions at high debt yields. When foreclosure costs are removed, lenders have no reason to avoid foreclosure, and thus do not extend loans where borrowers would engage in inefficient maintenance practices after. Panel (b) shows that foreclosure costs cause extensions for debt yields just below the point at which borrowers wouldbeabletopaytheloanoff. Abovethispoint,loansalwayspayoff(sincetherearenosearch frictions), and below it, debt overhang costs outweigh the benefit of giving time for incomes to recover. Finally, panel (c) shows that costs to loss recognition prompt lenders to extend only the 16

mosthighlystressedloans,whileallowingsomecloser-to-viableloanstodefault.14 Overall,these resultspointtoa cleardistinctionbetweensearch-related frictions andthoserelated to resolution costs. Search-related extensions go to borrowers that maintain the property, and have either high current debt yields or large paydowns (increasing future debt yields) so that the risk borne by lenders is minimal. In contrast, extensions related to costs of foreclosure or loss recognitiontendtoentaillowerdebtyields,lowprincipalrepayment,poormaintenanceincentives, and poor future repayment prospects. The primary distinguishing factor is degree, as these issues are all more pronounced for extensions to delay loss recognition; since the motivation is to delay rather than minimize losses, poor future repayment prospects do less to deter extend-and-pretend behavior. 2.6. RecapandTestablePredictions These figures provide the foundation for the empirical work in the next section. To summarize the results, motivations for extensions can be inferred by a combination of the debt yield of the loans getting extended and the terms of those extensions. Extensions due to default costs go to borrowers with low debt yields and entail minimal borrower concessions since borrowers would default if such concessions were required. In contrast, extensions to deal with property sale or refinancing frictions occur throughout the debt yield distribution, entail high principal paydowns forlower-debt-yieldloans,andhaveahighprobabilityofeventuallypayingoff. Regardingtheresponsetostress,ifstrainsrelatedtomonetarypolicytighteningandregionalbankingturmoilpromptedbankstoextend-and-pretend,wewouldexpecttoseethefollowingbasedon Figure3: 1. Moreextensionsduringthestressperiod 2. Extensionstooccuratlowerdebtyields 14When all three frictions are removed, no extensions occur at all; borrowers repay loans if possible and default otherwise. 17

3. Extensionstohavemorelenientterms(lessprincipalrepayment) 4. Extendedloanstohavelowerex-postpayoffrates 3. DATAANDMETHODOLOGY 3.1. Data ItestthepredictionsfromSection2.6usingsupervisorydatathatlargebanksreportfortheirstress tests. FR Y-14Q Schedule H.2 filings provide a loan-quarter-level panel on non-owner-occupied CRE loan holdings with committed balances over $1 million from banks with more than $100 billion in assets. I start the sample in 2016q1 since a reporting change after that quarter allows me to identify what happens to loans that exit the balance sheet (e.g., distinguish payoffs from liquidations). Thesamplerunsthrough2025:Q4. One of the key variables of interest is the maturity date, from which I derive whether loans are extended and whether a loan is scheduled to mature. Per the Y-14 reporting instructions “The maturity date is the last date upon which the funds must be repaid, inclusive of extension options that are solely at the borrower’s discretion.” Consequently, changes in this maturity date reflect extensionsprovidedbylenders,ratherthantheexerciseofexistingoptions. Themainoutcomesstudiedarewhetheramaturingloanisextended,whatthetermsofextensions are, and how loans perform following extension. I examine how these extension patterns compare during normal times (2016-2019) and after the bank stress episode (2023-2025) to assess both normalservicingpatternsandchangesduringstress. Theempiricalmethodologyvariesdepending on the outcome of interest, so I discuss those in turn. The analysis is organized around the four predictions regarding extend-and-pretend behavior in Section 2.6, but each piece of analysis also speakstoextensiondriversduringnormaltimes. 18

3.2. Predictions1&2: Howmanyandwhichloansgetextended The first two predictions say that greater extend-and-pretend incentives result in more extensions, drivenbyriskierloans. Totestthesepredictions,Iestimateregressionsalongthelinesof: 100 Extension =(β X ) 2023-on Maturing +γ LowerLevelControls +τ +ε i,t+1 ′ i,t t i,t ′ i,t b(i),t i,t × × × (2) where Extension is an indicator for whether a loan is extended in the next quarter, Maturing i,t+1 i,t is an indicator for whether it is scheduled to mature next quarter, and τ is a bank quarter-fixed b(i),t effect (only included in some specifications). 2023-on takes a value of 1 for quarters starting t 2023:Q1, and 0 for quarters before the pandemic. This period after 2022 was characterized by highinterestrates,weakCREtransactionvolumes,risingCREnonperformance,andincreasedattentiononbankCREexposuresfollowingtheSpring2023regionalbankingturmoil. Iexcludethe onset of COVID from the baseline analysis so the stress period is compared to a relatively normal environment, rather than one with elevated extensions due to the pandemic-era disruptions. See Glancyetal.(2025)foradescriptionofCREmodificationpatternsearlyinthepandemic. When X only includes an intercept, β estimates how the probability that a maturing loan is i,t extended changes during the stress period, thus testing the prediction that extensions rose at that time. LowerLevelControls includes the non-interacted variables, and thus γ can also provide i,t anestimateforthechangeinextensionfrequenciesforloanswithoutpendingmaturities(fromthe coefficienton2023-on inspecificationsthatomittime-fixedeffects). t When X is expanded to include risk characteristics, the regression tests the second hypothesis; i,t namely, that riskier loans got extended during the stress period. In this analysis, X includes i,t indicatorsforwhethertheloanhasadebtyieldunder8%,isnonrecourse,orissecuredbyasmallorlarge-sizedoffice(definedbysquarefootageunderorover250,000).15 β estimatestheextentto 15DebtyieldmaybeuninformativeifNOIismissingorhasn’tbeenupdatedinthelastyear,orifrepaymentrelies onthesuccessfulexecutionofaconstructionorrenovationplanratherthaninplacecashflows(i.e.,fornonstabilized loans). Inthesecircumstances,ImarkLowDebtYield to0andincludeanindicatortodenoteamissingdebtyield. i,t 19

which loans with these risk factors were disproportionately extended at maturity during the stress period. The prediction that extend-and-pretend incentives cause extensions to increase for low debt yield loans comes directly from the model. The remaining variables in X reflect other risk i,t factors that would have similar effects to low debt yields; nonrecourse clauses remove lenders’ claimonborrowers’otherassetsandmaythusreduceforeclosurerecoveries(Glancyetal.,2023), andofficeloanshavepoorNOIgrowthprospectsduetotheshifttoworkfromhome(Guptaetal., 2026; Glancy and Wang, 2023). These latter effects are particularly pronounced for larger offices, motivatingthedecisiontosegmentofficesbysize(GlancyandKurtzman,2024). I also conduct several variants of this type of analysis. This work includes (i) examining other loan outcomes (default or payoff) besides extensions, (ii) examining differences inextension rates quarter-by-quarter rather than pooling stress and non-stress periods, (iii) analyzing pending maturities over a one-year horizon to account for loans that pay off or get extended before their quarter ofmaturity,and(iv)estimatingtheprobabilitythatloansareextendedasaflexiblefunctionofdebt yieldratherthanusingthelowdebtyieldindicator. 3.3. Prediction3: Termsofextensions To test the prediction that extend-and-pretend incentives cause banks to ease terms (accept lower principalrepayment)onextendedloans,Iestimate: 100 ∆Term =(β X ) 2023-on Extension +γ LowerLevelControls +τ +ε i,t ′ i,t 1 t i,t ′ i,t 1 b(i),t i,t × − × × − (3) where ∆Term reflects the change in principal balance, loan rate spreads or recourse status of i,t a loan. LowerLevelControls includes lower level interactions of the variables of interest, i,t 1 − as well as a vector of variables related to changes in loan terms that might occur without extensions. These include controls for the loans’ age, size, scheduled amortization, and previous spread.16 16Sizeisthelogarithmofthepreviousloanbalance.Controlsforscheduledamortizationincludeboththescheduled 20

In the baseline analysis, X only includes a constant, so β measures how much more often i,t 1 − a given term was incorporated into extensions relative to before the pandemic. The coefficient on Extension (included in γ) measures the baseline increase in frequency with which a term is i,t changedforextendedloansrelativetosimilarnon-extendedloans,thusdistinguishingtheeffectof extensionsfromchangesintermsthatoccurforotherreasons. WhenX isexpandedtoinclude i,t 1 − the risk factors discussed in Section 3.2, β measures whether banks tightened or eased terms on extensions during the stress period for riskier loans in particular. Risk factors are measured as of the quarter before the extension so that X does not reflect changes that were a part of the i,t 1 − extension(e.g.,adebtyieldthatishighduetoaprincipalpaydown). 3.4. Prediction4: Ex-postperformanceofextendedloans Totestwhetherstress-eraextensionsperformedworsethanotherloans,Iestimate 100 PaidOff =β Extended +β Extended2023-on +τ +α +ε (4) i,t+1 1 i,t 2 i,t b(i),t m(i,t) i,t × wherethedependentvariableisanindicatorforwhethertheloanpaysoffnextquarter,Extended i,t is an indicator for whether the loan was previously extended, and Extended 2023-on is an ini,t dicator for whether the loan was extended during the stress period. α is a fixed effect for m(i,t) the number of quarters to maturity, which accounts for the fact that extension terms are typically shorter than those of new loans. τ is a bank-quarter fixed effect, which accounts for broad b(i),t changesintheprobabilityofrepaymentovertime. β captures differences in payoffs between prepandemic extensions and other loans at the same 1 bankandquarter,controllingfortimetomaturity. β captureshowpayoffratesdifferforstress-era 2 extensions relative to prepandemic ones. If banks extend loans during the stress period because paymentasashareofthepreviousloanbalancebasedonthereportedamortizationscheduleandtherealizedamortizationinthepreviousquarter. Iincludethelatterduetosomeapparentreportingerrorswithscheduledamortization causinglaggedprincipalrepaymenttobemorepredictiveoffuturerepayment. Toreducetheinfluenceofreporting errorsoroutliers,missingorextremevaluesforspreadoramortizationvariablesaresetto0,anddummiesforwhether valuesaremissing,highoutliersorlowoutliersareincluded. MostCREloanshaveminimalamortization,soresults arenotsensitivetothecontrols. 21

theywanttodelaylossrecognitionratherthanbecausetheyexpectmodificationstoenhancefuture repaymentprospects,β wouldbenegative. 2 3.5. Differencesbybankcapitalization Thelastpieceofanalysisinvestigatesdifferencesinextensionfrequenciesandtermsacrosslenders by capitalization level. This analysis tests whether banks that are closer to a capital requirement extendmoreloans,extendriskierloans,orprovidemorelenienttermsonextensions. Suchbehavior would suggest that capital constraints induce some banks to delay loss recognition in order to preservecapital. For this analysis, I supplement the CRE loan data with information on bank capital ratios from Y-9C, and stress test outcomes from public disclosures by the Federal Reserve Board. How stress tests were incorporated into capital requirements changed between the prepandemic and stress periods. Before the pandemic, stress tests were pass/fail based on banks’ estimated capital in a “severely adverse scenario.” For these years, I measure distance to capital constraints by the minimum common equity tier 1 capital (CET1) ratio in the stress tests. This variable reflects how muchheadroomabankhadinpassingtheirstresstests. During the stress period, stress test results translated into capital requirements through a stress capital buffer. For this period, I measure distance to a capital constraint as the difference between a bank’s CET1 capital ratio reported in their Y-9C, and their capital requirement inclusive of the bank-specificG-SIBsurchargeandstresscapitalbuffer. I then repeat analysis along the lines of that in Equations (2), (3) and (4), but adding an additional interaction with whether a bank’s capital buffer is below the median for the quarter (Low Capital =1). b(i),t 22

3.6. Summarystatistics SummarystatisticsofthemainvariablesofinterestareshowninTableA.2,andsummarystatistics forselectedsubsamples(prepandemicobservations,2023-onobservations,maturingloans,andextendedloans)areshowninTableA.3. Thetablesshowthatextensionratesaremodestoverall(3% per quarter), but high for maturing loans (52%). Loan terms are fairly stable over the life of CRE loans, but change much more frequently in quarters of extension. For example, the probability of theprincipalbalancedecliningbyatleast5%is8%forloansreceivingextensions,butaround1% for the sample as a whole. Likewise, loans receiving extensions are three to five times more likely tohavetheircommittedbalancerise,theirspreadchange,ortheirrecoursestatuschange. 4. WHYDOBANKSEXTENDLOANS? This section examines which potential motivations for extensions match observed empirical patterns. Section 4.1 shows that banks did not materially increase the frequency of extensions after 2022. Section 4.2 shows that stress-era extensions shifted towards safer borrowers. Section 4.3 shows that terms of stress-era extensions became stricter. Section 4.4 shows that stress-era extensions paid off at similar rates to prepandemic extensions. Finally, Section 4.5 shows that these patternsdonotdiffermateriallybybankcapitalization. 4.1. Didbanksincreaseextensions? Thefirsttestablepredictionisthatextend-and-pretendincentivespromptbankstoextendloansthey wouldn’totherwiseextend,causingmoreextensionstooccur. Iffairvaluedeclinesofbanks’fixedrate assets or increased scrutiny of CRE exposures following the 2023 regional banking turmoil causedbankstoextendloanstoobscurependingloanlosses,wewouldexpecttoseemorematurity extensionsstartingin2023. Toexaminehowextensionfrequencieshavechangedovertime,Istudyoutcomesofloansthatare slated to mature in four quarters, and assess loan outcomes as of the original quarter of maturity. 23

Theoutcomesconsideredare: 1. Paidoff: Iftheloanisdisposedofinthefollowingyear,andthedispositioncodeindicatesa voluntarypayoff. Ialsoincludeasmallnumberofloansthatweresoldwithoutacharge-off. 2. Extension: Iftheloaniscurrentandonthebalancesheetattheendofthequarterofmaturity, butthematuritydatewasextendedintothefuture. 3. Ballooned: Iftheloanisnotpastdueoninterestpayments,butthereportingdateisafterthe currentmaturitydate. 4. Delinquent: If the loan is past due in the quarter of maturity, or marked as liquidated, involuntarilypaidofforsoldatalossbythatpoint. Figure A.2 shows that the occurrence of these outcomes rises slightly in the year before maturity, and then jumps in the actual quarter of maturity. Consequently, the one-year window I consider should capture most extensions, payoffs, and delinquency associated with a pending loan maturity. Figure4plotsthecompositionofoutcomesforloanswithapendingmaturity(left)andtheshareof pendingmaturitiesthatgetextended(right)byquarterofmaturity. Thefigureshowsthatextension volumes rose after 2022, but not dramatically so. The volume of maturing loans receiving extensions was a bit under 50% for most quarters since the start of 2023. This exceeds pre-pandemic extensionrates(typical inthemid-40s)butwas wellbelowthe60 percentextensionrateobserved forloansmaturingearlyinthepandemic. Theshareofloanspayingoffatmaturity(theblueareain theleftpanel)fellfromnearonehalfbeforethepandemic,toaroundonethirdin2023,butthiswas mostly attributable to more loans missing payments rather than more loans receiving extensions. Delinquency rates started to recede and payoff rates to recover for loans maturing in late-2024 or 2025, but maturity distress remained elevated relative to prepandemic norms as of the 2025:Q4 sampleend. The appendix presents similar analysis at a more disaggregated level. Figures A.3 and A.4 dis- 24

Figure4: OutcomesofPendingCRELoanMaturities S fzX fS 9O fz 99 Y4 99X 9k 9l 9l Yf T # 2 uj c II 3 R jh N R / 0 N n 3 3 3 0 0 gHC\nC0 j30 : 9 9 z X X :9:9 9z :f :9 :e:4 :f :f :e :9 :4 9z :4 :O:O:O :9 Y: T C0hQ88 :: :::: :: :k :zX :zkO :l :l :z :S Yl k9X z kzX lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (a)MaturityOutcomes 030N3ju2h3a @b lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (b)ExtensionRateoverTime Notes: The left figure shows the share of outstanding loan balances that are paid off (blue), extended (yellow),performingpasttheirmaturitydate(orange)andpastdueorliquidated(red)bythequarterofscheduled maturity. Sample is composed of loans that are four quarters from the scheduled maturity. The right panel showstheshareofbalancesthatareextended,correspondingtotheyellowregionintheleftchart. tinguish stabilized vs. non-stabilized (construction and renovation) loans and offices vs. other propertytypes,respectively. Theresultsshowthatoutcomesofloanmaturitieswerebroadlysimilar for stabilized and non-stabilized CRE loans. Across property types, office loans became much lesslikelytopayoffatmaturity: Onlyabout20%ofofficeloanbalancespaidoffduringtheperiod of stress, compared to over 40% before the pandemic. However, this change mostly reflected a higherrateofdefaultratherthananincreaseinextensions. There is also little evidence of extensions increasing after 2022 when the analysis is expanded to include loans without pending maturities. The first column of Table A.4 presents estimates from equation (2), excluding the interactions with risk characteristics and the bank-quarter fixed effect (sothat2023-on isidentified). Itshowsthatonanunweightedbasis,maturingloanswereabout6 t percentage points less likely to be extended during the stress period and non-maturing loans were about half a percentage point less likely to get extended.17 The fourth column demonstrates that delinquency rates for maturing and non-maturing loans rose by about as much as extensions fell (4.4and0.8percentagepoints,respectively). 17ThedecompositioninFigure4codesloansasdelinquentwhentheywerebothextendedanddelinquentinorder foroutcomestosumtoone. Inregressions,IcounttheseloansasbeingextendedsinceIamnottryingtodecompose loanoutcomes. 25

Overall, these time series patterns demonstrate that CRE market stress prompted more loans to default upon maturity. However, there is no clear sign of banks increasing extensions to hide the stress. Extension rates were roughly in line with historical norms, increasing somewhat relative to before the pandemic on a weighted basis and decreasing somewhat on an unweighted basis.18 These patterns are more consistent with property value declines causing borrowers to default than bankingstressescausingbankstoincreaseextensionstodelaydefault. Another implication of the fact that recent extension rates are in line with historical norms is that the pending “wall of maturities” is also not historically atypical. One frequently cited concern about recent CRE loan extensions is that they create a build-up of near-term maturities that could strain loan performance going forward (Crosignani and Prazad, 2024). While it is true that recent extensions add to pending maturities, this situation is not unique to the stress period. Figure A.6 plotstheshareofoutstandingCREloansscheduledtomatureinthenexttwoyears,andtheportion of those loans accounted for by recently extended loans. As of 2025:Q4, about 40 percent of loan balances are within two years of their maturity, a bit under 30 percent of which (i.e., about 11 percent of total balances) come from loans extended in the last two years. The share of loans with imminent maturities has risen since the onset of the pandemic, but is only slightly above prepandemicnorms. 4.2. Didextensionsriseforriskierloans? The second testable prediction is that extend-and-pretend incentives cause banks to extend lowerquality loans. In the context of the model, the marginal extensions driven by capital preservation incentives occur at very low debt yields since those loans would produce larger losses in default. Otherfactorsthatreducepropertyvaluations(e.g.,officestrains)orrecoveryprospects(e.g.,nonrecourse clauses) would also increase potential losses and thus similarly affect extension incentives. Incontrast,extensionsduetotemporaryrepaymentdifficultiesgotohigher-debt-yieldloanswhere 18FigureA.5showsthattheprobabilityofextensiongenerallyroseformaturingloansover$20millioninsize,and fellforsmallerloans,explainingthediscrepancyinweightedandunweightedresults. 26

Figure5: MaturityExtensionsbyRiskCharacteristics fz 9z :z kz lz 3a @chNRCcN3ju2 fz Ua3A+QpB/ lzlkARN :z lz z z Yz9 YS YS9 Yl /3$jhwC3I0hVMQBgHR Nh# I N,3W (a)ExtensionsbyDebtYieldandTimePeriod 3a @chNRCcN3ju2 NRNA`3,Rnac3 `3,Rnac3 z Yz9 YS YS9 Yl /3$jhwC3I0hVMQBgHR Nh# I N,3W (b) Stress period Extensions by Debt Yield and Recourse Notes: Figure presents semi-linear regression estimates of the probability of extension as a function of a loan’sdebtyieldandbank-quarterfixedeffects. EstimatescomefromacubicB-spline,restrictedtohavea continuous second derivative, using the binsreg package of Cattaneo et al. (2024). Dots provide binscatter estimatesbyquartile. SampleincludesstabilizedloanswithrecentNOIupdatesthatarescheduledtomature in four quarters. The left panel compares loans slated to mature before the pandemic (blue) and during the stress period (red). The right panel analyzes extension rates during the period of stress for recourse (lavender)andnon-recourse(pink)loans. borrowersarenotatthemarginofdefault. Figure5demonstratesthatextensionratesdeclinedforlowdebtyieldloansafter2022,contraryto banking strains causing extend-and-pretend behavior. Each chart presents cubic spline regression estimates predicting whether loans maturing in the next year receive an extension as a function of theirdebtyield,controllingforbank-quarterfixedeffects. Theleftpanelcomparesextensionrates across time periods. Extension rates were roughly flat across debt yields before the pandemic (the blueline),butdropnotablyforincome-strainedpropertiesafter2022. Theprobabilityofextension falls from around 40–45 percent at a debt yield of 8 percent, to around 30 percent for loans with non-positivenetincome. The right panel provides more detail on the decline in extensions of income-strained loans by separatelyestimatingextensionratesforrecourseandnon-recourseloans. Thedeclineinlow-debtyield loan extensions after 2022 is almost entirely accounted for by non-recourse loans. About 40 percent of recourse loans with very low debt yields get extended, compared to only about 20 percent of non-recourse loans. Namely, when banks do extend loans against income-strained 27

properties, those loans disproportionately have guarantees so that the strained property is not their only source of recovery. Taken together, these figures show that loans typically required either strong collateral or sponsor guarantees to get extended after 2022, indicating that banks are not extendingloanswithunfavorablerecoveryprospectstodelaylossrecognition. Though these results indicate that banks had fairly stringent income requirements after 2022, FigureA.7providessomeevidenceofincreasedrisktakingwithrespecttothepropertytypesreceiving extensions. Each figure plots the share of pending CRE loan maturities that receive extensions by propertytypebeforeCOVID,from2020to2022,andafter2022. Thelargestincreaseinextensions isforindustrialproperties—thepropertytypewiththelowestdelinquencyrateafterthepandemic. However, extension rates rose for office loans following the pandemic, especially on a weighted basis(theleftpanel). Giventhatthedeteriorationinloanperformanceduringthisperiodwasconcentrated in large-sized office loans (Glancy and Kurtzman, 2024), this result could reflect banks extendingloanswithworsefutureincomeprospects. Figure6presentsregressionestimatesthatjointlyaccountforhowthesedifferentriskfactors(low debt yields, nonrecourse clauses, office collateral) relate to loan extension frequencies. Blue dots (lines) plot point estimates (95% confidence intervals) from equation (2), while red dots show equivalent estimates predicting whether loans are delinquent in the following quarter. The figure plots estimates of the β vector, reflecting how much banks changed their tendency to extend risky loans upon maturity. Other coefficient estimates, for example those pertaining to extensions of maturing loans before the period of stress or changes in extensions for non-maturing loans, are showninTableA.4. Overall,thefigurevalidatestheideathatbanksreducedmaturityextensionsforriskierloans. Low debt yield loans were about 7 percentage points less likely to receive extensions after 2022, and nonrecourse loans were 5 percentage points less likely to get extended. The one dimension where maturity extensions appear to have become riskier is that the extension rate for large office loans rose3percentagepoints(thoughthedifferenceisstatisticallyinsignificantduetothelimitednum- 28

Figure6: MaturityExtensionsbyRiskCharacteristics,RegressionEstimates AkYf9 2uj3NcCRNc K jnaCN<huhlzlkARN SY9l /38 nIj AfYOS YYYuhHRsh/3$jhvC3I0 9Y:f A:YOS YYYuhMRNa3,Rnac3 kYeO zYS4 YYYuhbL IIhQ88C,3 lYz: kYzS YYYuhH a<3hQ88C,3 eYf: ASYz: YYYuhKCccCN<h/3$jhvC3I0 SY4: AS9Yzz ASzYzz A9Yzz zYzz 9Yzz SzYzz Notes: Figurepresentsestimatesofβ fromthespecification 100 Extension =(β X ) 2023-on Maturing +γ LowerLevelControls +τ +ε × i,t+1 ′ i,t × t × i,t ′ i,t b(i),t i,t where τ is a bank quarter fixed effect. Blue dots (lines) present point estimates (95% confidence interb(i),t vals)forhowmuchparticularriskfactorsincreasedtheprobabilitythatmaturingloansreceivedextensions duringtheperiodofstressrelativetobeforethepandemic. Reddots(lines)presentequivalentestimatesof the probability that loans are delinquent as of maturity. Estimates correspond to columns (3) and (6) from TableA.4. ber of large office loans). The red dots show that these risk factors are all associated with higher maturitydefaultrisk.19 Tosummarizetheaggregateevidence,banksfrequentlyextendCREloansatmaturity. Thisbehaviorisnotconcentratedinweakerloansandnotconcentratedintheperiodofstress,consistentwith extensions predominantly addressing temporary payment difficulties. In fact, maturity extension rates declined for loans with weak incomes during the stress period, especially for nonrecourse loans, contrary to banks offering lenient extension policies to prevent highly stressed borrowers fromdefaulting. The one result potentially consistent with extend-and-pretend behavior is that banks increased office loan extensions. Whether these office extensions contain provisions that bolster future per- 19Though banks generally reduced extensions of risky loans with pending maturities, there is some evidence of extensions for riskier non-maturing loans increasing. Column (3) of Table A.4 shows that extension rates for small and large office loans without pending maturities rose 0.3 and 1.4 percentage points, respectively, starting in 2023. Therewasnosignificantdifferenceinnon-maturityextensionratesbyrecoursestatusordebtyield. 29

formance prospects is therefore of particular interest. In the next section I will pay additional attentiontoofficeloanswhenassessingthetermsofextensions. 4.3. Didbankseasetermsonextensions? Next,Ishiftattentiontocharacterizingthetermsofextensions. Oneofthedefiningcharacteristics ofevergreening/zombielendingisthatlendersprovidesubsidizedcredittoinducefirmstocontinue operating when they wouldn’t otherwise be financially viable. In the model, this materializes as a reduction in required payments that, in essence, provides borrowers with a cheap option on the property. First,Iexaminehowextensiontermschangedonaverageduringthestressperiod,thenI examinedifferencesbyborrowerrisk. Extension terms, on average Table 2 presents estimates from Equation (3). The first column predicts the percentage quarterly decline in loan balances by whether or not the loan receives an extension. After 2022, borrowers paid down 2.5 percentage points more of their loan balance to extend loans. Columns 2 and 3 show that the higher paydown rate reflects both more extensions requiringpaydownsandfewerloanswithbalancesrising. CREloanextensionsinthestressperiod wereabout5.1percentagepointsmorelikelytoentailpaydownsexceeding5%oftheloanbalance relative to before the pandemic, and were about 6.4 percentage points less likely to have balances increase(potentiallyreflectinginterestdeferrals). Additionally, column 4 shows that extensions after 2022 were 1.6 percentage points more likely to switch to having recourse. As recourse provides banks an additional means of recovery beyond the subject property, this change has a similar effect on a loan’s future repayment prospects as decreasingtheloanbalance(Glancyetal.,2023). The last three columns investigate changes in loan pricing. On average, banks increase spreads on extended loans by about 8 basis points more after 2022 compared to before the pandemic. This effectisdrivenbybanksbecomingbothmorelikelytoincreasespreads(column6)andlesslikely toreducethem(column7). 30

Table 2: Terms of Extensions Paydown 1(Paydown>5%) ∆Balance>0 GainedRecourse ∆Spread 1(∆Spread>0) 1(∆Spread<0) (1) (2) (3) (4) (5) (6) (7) Extensioni,t -2.06** 3.17** 10.93** 2.22** -0.03** 5.25** 12.14** (0.24) (0.33) (0.87) (0.56) (0.00) (0.38) (0.65) ... 2023-ont 2.54** 5.13** -6.37** 1.62* 0.08** 9.52** -6.88** × (0.28) (0.51) (1.02) (0.76) (0.01) (0.93) (0.85) R2 0.054 0.068 0.138 0.167 0.103 0.284 0.190 a Observations 1,356,810 1,356,810 1,356,810 525,523 700,524 700,524 700,524 Controls ✓ ✓ ✓ ✓ ✓ ✓ ✓ Bank-quarterFE? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 ∆Term =(β +β 2023-on )Extension +γ Controls +τ +ε i,t 1 2 t i,t ′ i,t 1 b(i),t i,t × − wherethedependentvariableisthechangeinsomeloanterm,scaledby100soestimatesareinpercentage points. This variable is the principal paid down as a share of the loan balance in column (1), an indicator for whether this pay down is at least 5% in (2), an indicator for whether the committed balance rose in column (3), an indicator for whether a previously nonrecourse loan became recourse in (4), the change in loanspreadin(5)andindicatorsforwhetherthespreadroseorfellincolumns(6)and(7),respectively. The sample changes throughout the analysis: (1)–(3) exclude loans with changes in utilization or charge-offs, (4)requirestheloantopreviouslybenonrecourse,and(5)–(7)requireloanstohavefloatingrates. Themain independentvariablesareindicatorsforwhethertheloanisextendedinquartert anditsinteractionwithan indicatorforwhethert occursduringtheperiodofCREstress(2023-on). Allspecificationsincludecontrols foraloan’sage,size,amortization,andpreviousspread;andbank-quarterfixedeffects. Standarderrors,in parentheses,areclusteredbybank-quarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 31

Overall,thesefindingsallpointinthesamedirection;namely,thatthetermsonextensionsbecame more stringent. During the period of CRE market turmoil, borrowers had to pay down more debt and pay higher interest rates to continue receiving credit. While this doesn’t necessarily rule out credit being subsidized seeing as fundamental risks to the CRE market rose, it indicates that the motivation for extensions is not merely to delay losses. The model demonstrates that property strains only cause terms to become more stringent in the region where extensions remedy search frictions, not where subsidized credit is needed to keep borrowers from defaulting. That banks amended loan terms in ways that improved the prospect for future recoveries and compensated them for the risks they were taking—and the fact that borrowers were willing to accept these termswithoutdefaulting—arepositivesignalsastotheviabilityoftheseextendedloanseventually payingoff. Paydowns by loan risk How do these extension terms vary by loan risk? The model demonstratesthattherelationshipbetweenprincipalrepaymentanddebtyieldcaninformwhyextensions occur and how much risk there is to future performance. If banks provide easier terms for weaker loans, it indicates that borrowers’ participation constraint is binding and lenders are providing lenienttermstoavoidthecostsassociatedwithdefault. The question is whether banks are willing to provide the extensions that require subsidized credit. If banks limit extensions to loans that are likely to ultimately pay off—lending to customers that arecommittedtothepropertybutfacingtemporarydifficultiesraisingfundsforloanrepayment— then CRE market strains would cause extension termsto tighten, particularlyfor stressed loans. If insteadbanksarewillingtoextend-and-pretend,we’dseeeasiertermsformore-stressedproperties topreventdefault. Forthisanalysis,Ifocuspredominantlyonwhetherornotextensionsentailthe repaymentofatleast5%oftheprincipalsincethatchangemostclearlyimprovesbanks’prospects forfuturerepayment.20 20Higher spreads don’t increase the likelihood that loans pay off, they just increase the compensation that banks receivefortakingonthatdefaultrisk. Increasesinbalancesarealsosomewhathardertointerpretthanpaydownssince they could reflect interest deferrals (which increase leverage) or the funding of property improvements to enhance collateralvalues(potentiallyreducingleverage). 32

Figure 7 presents spline estimates of the share of CRE loan extensions that entailed a principal paydown of at least 5% after 2022 (red) and before the pandemic (blue). The left panel presents findingsforallpropertytypes,andtherightrestrictsthesampletoofficeloans. The figure demonstrates that principal paydowns became more likely for low-debt-yield loans during the stress period. For the full sample, about 20% of loan extensions in the bottom quartile of the debt yield distribution had material principal paydowns during the stress period, compared tounder10%ofextensionsbeforethepandemic. Thefrequencyofpaydownsatthetopofthedebt yield distribution is lower (around 5%), and differs little across the two time periods. The pickup in principal paydowns is even more stark in the office sample, with paydown rates shifting up throughoutthedebtyielddistributionandreachingover30%forverylow-debt-yieldloans. To recap, extend-and-pretend incentives would prompt lenders to provide lenient extension terms to low-debt-yield borrowers to induce them to extend rather than default (Figure 3). In contrast, we see a negative relationship between principal paydowns and debt yield both in normal times and after the bank stress episode, even at low debt yields. These results indicate that banks predominantlyextendloansforwhichborrowersarecommittedtothepropertyandwillingtoprovide credit enhancements to retain it.21 Moreover, that banks reduced extensions and increased paydownrequirementsforlow-debt-yieldloansafter2022indicatesthatbankingsectorstrainsdidnot causebankstodelaylossrecognition. Whenlow-debt-yieldborrowerswerenotwillingtoprovide enhancements, banks appear to have been comfortable allowing them to default, as evidenced by thedeclineinextensionsandriseinmaturitydefaultsforlow-debt-yieldloansinFigure6. 21Notethatnotallpotentialcreditenhancementsareobservedinthedata. Extensionsmightalsoincludecontributionstoreservesforcapitalexpenditure, leasingcosts, orotherfutureexpenses, whichwouldhavesimilareffectsto principalcurtailment. Likewise, additionalguarantees, covenants, orcashsweeporlockboxprovisionscanenhance lenders’ repayment prospects or control rights. Namely, loans that do not receive principal paydowns may receive otherunobservedenhancements. 33

Figure7: PaydownsbyDebtYield :z kz lz Sz z cNsR/hw Th@jCsh3a @b fz Ua3A+QpB/ lzlkARN :z lz z z Yz9 YS YS9 Yl /3$jhwC3I0hVMQBgHR Nh# I N,3W (a)AllExtensions cNsR/hw Th@jCsh3a @b Ua3A+QpB/ lzlkARN z Yz9 YS YS9 Yl /3$jhwC3I0hVMQBgHR Nh# I N,3W (b)OfficeLoanExtensions Notes: The left panel presents semi-linear regression estimates of the probability an extension entails a principal paydown of at least 5% as a function of a loan’s debt yield, controlling for amortization and bank-quarter fixed effects. Estimates come from a cubic B-spline, restricted to have a continuous second derivative,usingthebinsregpackageofCattaneoetal.(2024). Dotsprovidebinscatterestimatesbyquartile. SampleincludesloanextensionswithrecentNOIupdates. Estimatesforpandemic-eraextensionsareinblue andstress-eraextensionsinred. Therightpanelpresentsthesameestimates,butwiththesamplerestricted toextensionsofofficeloans. Robustness to other risk factors and loan terms Patterns are broadly similar for other risk factors (besides low debt yields) and other extension terms (besides principal paydowns). Figure 8 extends the analysis to consider other risk factors. Each dot presents coefficient estimates for β from the specification in Equation (3), corresponding to how much a particular risk characteristic shiftedtheprobabilitythatanextendedloanreceivedapaydownduringtheperiodofstressrelative tonormaltimes. The results demonstrate that paydowns for riskier loans increased after 2022, both overall and relative to extensions for safer loans. The blue dot repeats the second column of Table 2 with the additional(uninteracted)riskfactorcontrols,showingthatextensionsafter2022were5.15percentage points more likely to have a 5% paydown compared to prepandemic extensions. The red dots presentthetriple-interactioncoefficientsfromEquation(3)whenX isexpandedtoincluderisk i,t 1 − characteristics. These results show that the increase in principal paydowns for extensions is most pronounced for riskier loans. A low debt yield increases the probability of a principal paydown by 4.1 percentage points more during the period of stress, though the difference is statistically 34

Figure8: PaydownsbyRiskCharacteristic 9YS9 2uj3NcCRNhuhlzlkARN lY:l YYYuhHRsh/3$jhvC3I0 :Yz4 YYYuhMRNa3,Rnac3 lYk: YYYuhbL IIhQ88C,3 kYff YYYuhH a<3hQ88C,3 SkYlO YYYuhKCccCN<h/3$jhvC3I0 SYlS zYzz 9Yzz SzYzz S9Yzz lzYzz 2883,jhVU3a,3Nj <3hURCNjcW # c3ICN3h2883,j #wh`CcG VNRhCNj3a ,jCRNcW +@ a ,j3aCcjC, Notes: Thefigurepresentsβ estimatesfromequation(3),pertainingtochangesinthefrequencywithwhich extensions have paydowns of at least 5% during the stress period. The blue dot presents the estimated coefficient on Extension 2023-on in the specification without the risk interactions, while the red dots i,t t × present estimates of the β vector from the fully-interacted specification. Lines present 95% confidence intervalsbasedonstandarderrorsthatareclusteredatthebank-quarterlevel. insignificant. Likewise, nonrecourse loans, small office loans, and large office loans were 2.3, 3.7 and 13.3 percentage points more likely to have paydowns during the period of stress, respectively, comparedtobeforethepandemic. TableA.5presentsresultsforotherloanterms,aswellasprovidingestimatesforothercoefficients that are not displayed in Figure 8 (e.g., the prevalence with which terms change for extensions during normal times). These results show that low debt yields also increased paydowns for extensions by about 4 percentage points during the prepandemic period, meaning that low debt yields increased the probability that stress-era extensions have paydowns by nearly 8 percentage points overall(summingtheeffectofextensionsduringnormaltimesandthechangeduringtheperiodof stress). Theotherriskfactorsappeartohavematteredlittlebeforethepandemic. Regarding other extension terms, office loan extensions were roughly 2 percentage points more likelytohaverecourseaddedasaconditionofextensionsandsawspreadsriseby8–10basispoints morethanotherloanextensions(thoughtheincreaseinrecourseisstatisticallyinsignificant). Lowdebt-yield loans were less likely to have spreads rise relative to other stress-period extensions, perhaps reflecting banks’ concern about those borrowers being able to service their debt at higher 35

spreadswhenreferenceratesweretypicallymuchhigherthanatorigination. In short, the analysis of loan terms shows that banks required greater concessions from borrowers to provide extensions during the stress period. Moreover, the increase in concessions was most pronounced for riskier loans. This tightening in terms indicates that banks did not provide lenient extension policies to prevent borrowers from defaulting as would be emblematic of extend-andpretend modifications. The tightening in terms was most pronounced for office loans, indicating thattheincreaseinextensionsforofficeloansdiscussedinSection4.2isnottheresultofsubsidized creditfrombanksthatareunwillingtorecognizelosses. 4.4. Wereextendedloanslesslikelytoultimatelypayoff? Unlike extensions to address foreclosure costs or search frictions, extensions to delay loss recognition can be desirable from a bank’s perspective even if the probability of future repayment is negligible. Indeed,Figure3showsthattheextensionsinducedbycoststolossrecognitionhavealmostnohopeofpayingoff. Thissectiontestswhetherfuturepayoffratesforstress-eraextensions deteriorated. Figure9comparestheperformanceofextendedloanstothatofotherloanswithpendingmaturities that had not been previously extended. The left panel plots the share of extended loan balances that have paid off over time since the extension. Dashed, dotted, and solid blues show payoff rates for loans extended before, during, and after COVID, respectively. Red lines plot the share of loans that are in default at that time (including loans that are delinquent, nonaccrual, or past theirmaturitydate). Irestrictthesampleforthisanalysistoextensionswherethetimetoextended maturity is within 6 quarters, and exclude time horizons pertaining to periods occurring after the 2025:Q4 sample end.22 These criteria ensure that whether a payoff occurs is always observable at the specified time horizon and that maturity outcomes are realized by the final time horizon. Namely, loans not paying off by the 6-quarter ahead time period would have either defaulted or 22For example, extensions occurring in 2025:Q1 only contribute to estimated payoff rates up to a three-quarter horizon. The6-quarteraheadpayoffratepertainstoextensionsoccurringbetween2023:Q1and2024:Q2thatmature within6quarters. 36

beenextended. The figure demonstrates that stress-era extensions were less likely to pay off following the extension. 66% of pre-COVID extensions paid off within 6 quarters of the extension, compared to only 58% of loans that were extended during the period of stress. Payoff rates for COVID-era extensionsweresimilartostress-eraextensionsat60%. Regarding default, loans extended during the period of stress are more likely to default following the extension, but extensions appear to not do much to delay the reporting of default; about 11% ofloansareindefaultthequarterafterextension,andthisdefaultedshareremainsaboutflatinthe yearandahalffollowingextension. Sixquartersaftertheextension,thedefaultrateis7percentage points higher than before the pandemic, thus accounting for almost all of the 8 percentage point declineinthepayoffrate. Thisresultmeansthattheshareofloansstilloutstandingandperforming after 6 quarters due to the receipt of another extension is about unchanged relative to before the pandemicat31%. Namely,banksarenotrollingovershort-termextensionsatahigherrate. Thekeyquestioniswhetherloanperformancedeterioratedbecauselendersstartedextendinglower quality loans, or because stresses in the market caused performance to deteriorate more broadly (e.g.,worsepropertymarketliquidityreducingpayoffsforextendedandnon-extendedloansalike). Togetatthiseffect,therightpanelplotsthepayoffratesforloanswithpending(fourquarterahead) maturitiesthathadnotbeenpreviouslyextended. Theseloanswouldnotbesubjecttotheselection effects induced by extension decisions and give a comparison group for how performance would differovertimeabsentsuchselectioneffects. The deterioration in loan payoffs during the stress period is slightly more pronounced for nonextended loans. The share of non-extended loans paying off after 6 quarters declined by 15 percentagepointsrelativetobeforethepandemic(67%to52%),comparedto8percentagepointsfor extended loans (66% to 58%). In fact, the 6-quarter-ahead payoff rate for stress-era extensions is abovethatofnon-extendedloans(58%vs. 52%). In short, Figure 9 shows that extended loans usually perform similarly to other loans, paying off 37

Figure9: PayoffRatesOverTime Ye T C0AR88hTa3A+QpB/ Yf T C0AR88h+QpB/ T C0AR88hlzlkARN Y9 Y: Yk Yl YS /38 nIjhlzlkARN /38 nIjh+QpB/ /38 nIjhTa3A+QpB/ z cN RHh8Rh3a @b Ye T C0AR88hTa3A+QpB/ Yf T C0AR88h+QpB/ Y9 T C0AR88hlzlkARN Y: Yk Yl YS /38 nIjhlzlkARN /38 nIjh+QpB/ z /38 nIjhTa3A+QpB/ S l k : 9 f [n aj3achbCN,3h2uj3NcCRN (a)ExtendedLoans cN RHh8Rh3a @b Ak Al AS z S l [n aj3achiRhK jnaCjw (b)OtherMaturingLoans Notes: Blue lines plot the share of loan balances that pay off over time before (dashed), during (dotted) and after (solid) the pandemic. Red lines plot the share of loan balances that are delinquent or liquidated atthattime. Theleftpanelplotsperformanceforloansthatwereextendedbythenumberofquarterssince extension,whiletherightplotsperformancefornon-extendedloansbythenumberofquarterstomaturity. within six quarters about two-thirds of the time, and mostly receiving another extension the other thirdofthetime. Thishighrateofpayoffandlowrateoffuturedefaultisconsistentwithextensions relievingtemporarystressratherthanpersistentpaymentdifficulties. Extendedloansactuallyoutperformed other loans after 2022, consistent with the other findings about banks becoming more selective with extensions and requiring greater principal paydowns to support future repayment prospects at that time. Namely, more stringent extension policies appear to have bolstered the performanceofextendedloansduringaperiodofbroadrepaymentdifficulties. TableA.6 presentsregression estimatesalong thesame linesas thosein Figure9. Eachregression predictswhetheraloanpaysoffinthefollowingquarterbasedonwhethertheloanwaspreviously extended,andifso,whetherthelastextensionoccurredin2023orlater. Eachspecificationincludes quarter-to-maturity fixed effects to account for differences in loan terms for extensions vs. new originations. The table shows that stress-era extensions are about 2.3 percentage points less likely to pay off than other extended loans in the specification without quarter fixed effects (column 1). However, the effect switches sign when quarter or bank-quarter fixed effects are added in columns(2)and(3). Namely,oncethebroaddeteriorationinloanpayoffsduringthestressperiod is accounted for, stress-era extensions are at least 1.4 percentage points more likely to pay off in a 38

quartercomparedtoprepandemicextensions. 4.5. Didworsecapitalizedbankshaveeasierextensionpolicies? Overall, the aggregate evidence is inconsistent with incentives to delay loss recognition driving banks to extend loans with weak repayment prospects. Extension rates changed little after 2022 and actually declined for riskier loans. Furthermore, the extensions that occurred during the stress period tended to have more stringent paydown requirements and performed favorably relative to prepandemic extensions. While these findings all suggest that extend-and-pretend was small in aggregate,itdoesnotruleoutthoseincentivesaffectingbehavioratsomebanks. Inthislastsection, I examine whether extension patterns at banks near their capital constraints are more consistent withextend-and-pretendbehavior. Extensionpoliciesbybankcapitalization Toinvestigatehowextensionpoliciesdifferbybank capitalization, I re-estimate equations (2), (3), and (4), but add an additional interaction with Low Capital , which takes a value of 1 if a bank is closer to its capital requirement than the median b(i),t foragivenquarter. Table A.7 demonstrates that low capital banks, if anything, reduced extensions even more than banks thatwere comfortablyabovetheir capitalrequirements, but differencesare small andstatistically insignificant. Well-capitalized banks reduced extension rates by 0.4 percentage points for non-maturing loans and an additional 4.3 percentage points for maturing loans during the stress period (the coefficients on 2023-on and Maturing 2023-on , respectively). These contractions t i,t t × in extensions were about 0.4 percentage points larger for banks closer to their capital constraint (thecoefficientsontheaforementionedvariablesinteractedwithLowCapital ). b(i),t Table A.8 shows that changes in the stringency of extension terms were generally similar at low andhighcapitalbanks. Theincreasesinthesharesofextensionsreceivingmaterialpaydownsand higher spreads during the period of stress were 2.7 percentage points and 4.3 percentage points greater, respectively, for low capital banks. However, loans at low capital banks were also more 39

likely to have balances rise or spreads fall when extended, indicating that the differences reflect a greater general tendency to adjust terms when extending loans rather than a difference in overall stringency. Onnet,thechangesinpaydownsandspreadswereaboutthesameforhighandlowcapital banks. The differences by bank capitalization are small relative to the broader changes during the period of stress: Low and high capital banks alike increased paydowns, spreads, and recourse requirementswhenextendingloansafter2022(thoughrecourseresultsareinsignificant). Differences by loan risk The findings so far regarding the effects of proximity to capital requirementsindicatethat(i)morecapitalconstrainedbanks,ifanything,providedfewerextensions duringtheperiodofstressand(ii)highandlowcapitalbanksincreasedpaydownrequirementsand loanratespreadsbysimilaramountsduringthestressperiod. Inotherwords,capitalizationrelates onlyweaklytoextensionpatterns. Onepessimisticinterpretationofthesefindingsisthatlowcapitalbankstriedtodiscourageextensions from high-quality borrowers (so loans would pay off and free up capital) while encouraging them from those who might default without generous extension policies. Namely, aggregate patternsmightmaskashiftinthecompositionofextensionsformoreconstrainedbanks. Toassessthispossibility,IrepeattheanalysisfromTablesA.7andA.8,butaddadditionalinteraction terms to capture how capitalization relates to extension frequencies and terms for borrowers with different risk characteristics. Every specification includes lower-level interaction terms and bank-quarter fixed effects. As the bank-quarter fixed effect controls for any broad tendency to provide extensions or change loan terms, the estimates reflect differences in how banks manage extensionsacrossloansofdifferentrisk. Figure 10 summarizes the main findings from this analysis, while the disaggregated regression estimates are reported in Tables A.9 and A.10. The left figure shows how extension policies at low capital banks changed after 2022. Red dots plot the change (relative to before the pandemic) in extension rates for loans at low capital banks that are scheduled to mature next quarter, the blue dots the change in extension rates for nonmaturing loans, and the green dots the change in 40

how often extended loans have a principal paydown of at least 5%. The right figure plots the change in extension outcomes for low capital banks relative to high capital banks. The top row presentsoverallestimates(fromregressionswithouttheriskfactorinteractions),andtheotherrows plot results from the fully-interacted specification; No Risk Factor reports the predicted extension outcomes for a recourse, non-office loan, with a debt yield over 8%, and the other rows report estimatesforaloanwithasingleriskfactor. If proximity to a capital buffer induces low capital banks to extend-and-pretend, we would expect those banks to disproportionately extend riskier loans (red and blue dots would be to the right of the zero line for loans with the listed risk factors) and provide lenient extension policies to induce borrowers to extend (green dots would fall to the left of the zero line, indicating fewer extensions withprincipalpaydowns). There is no area where this combination of easing terms and rising quantities occurred. The only evidence of low capital banks providing more extensions of risky loans is with nonmaturing office loans; non-maturity extension rates rose by 0.8 and 2.3 percentage points for small and large office loans, respectively. However, the frequency of principal paydowns rose about 8.4 and 13.3 percentagepointsfortheseloans,indicatingthatthechangewasnotpromptedbyeasierextension policies. In fact, low capital banks increased paydown requirements for loans with every assessed riskfactor;namely,thefrequencyofpaydownsatlowcapitalbanksroseby7.6and10percentage pointsforlowdebtyieldloansandnonrecourseloans,respectively,after2022. Therightpanelplotsestimatesforthechangeinextensionoutcomesatlowcapitalbanksnetofthe change at high capital banks. This analysis examines whether low capital banks eased extension policies relative to other banks even if they did not ease policies in general. If paydowns rose broadly due to greater risks, or extensions fell because more borrowers defaulted, we might still seerelativelysmallershiftsforlowcapitalbanksifthesedevelopmentswerepartiallycounteracted bysomecapital-preservation-motivatedextensions. There is little consistent evidence of extension policies becoming more stringent at one type of 41

Figure10: ExtensionPoliciesbyBankCapitalization eYfl lYez q3a <3h2883,j 2 2 u u j j 3 3 N N c c C C R R N N c c h h V V N L R N jn AL aCN < jn W aCN<W A9Y:9 AzYef q3a <3h2883,j 2 2 u u j j 3 3 N N c c C C R R N N c c h h V V N L R N jn AL aCN < jn W aCN<W A A z z Y Y 4 : S S T w0RsNc :Ykf T w0RsNc SYe: MRh`CcGh7 ,jRac AlY49 zYzz MRh`CcGh7 ,jRac zYzz lYl4 eY94 SYlk HRsh/3$jhvC3I0hHR Nc ASSY4k AzYSl HRsh/3$jhvC3I0hHR Nc AlY A S z l Y94 SzYSf fYkz MRNa3,Rnac3hHR Nc ASYO Az f YSO MRNa3,Rnac3hHR Nc AzYkk eY:: 4Y:z SY9O bL IIhQ88C,3hHR Nc AkYOO zY4: bL IIhQ88C,3hHR Nc zYe l 9 Y9e SkYlO ASSYkf H a<3hQ88C,3hHR Nc AlY:f lYle H a<3hQ88C,3hHR Nc zY S : Y 4 9S Alz ASz z Sz lz Akz Alz ASz z Sz lz +@ N<3hCNhbja3cchT3aCR0h8RahHRsh+ UCj Ih# NGchVU3a,3Nj <3hURCNjcW +@ N<3hCNhbja3cchT3aCR0h8RahHRshqcYh?C<@h+ UCj Ih# NGchVU3a,3Nj <3hURCNjcW (a)ChangeforLowCapitalBanks (b)ChangeRelativetoHighCapitalBanks Notes: The left figure plots the predicted changes in extension outcomes during the stress period for low capital banks. Outcomes considered are the predicted change in the probability that a nonmaturing loan is extended (blue), that a maturing loan is extended (red) and that an extended loan receives a principal paydownofatleast5%. Theseestimatescomefromfully-interactingthespecificationsinequations(2)and (3)withalowcapitalindicator. Theaverageeffectcomesfromthespecificationexcludinginteractionswith loan risk factors, and the other estimates provide predictions for a loan with a single risk factor (besides the ”No Risk Factor” line which pertains to loans where all of the indicators in X are 0). The right panel i,t presentsestimatesofthechangeinpoliciesatlowcapitalbanksnetofthechangeathighcapitalbanks(e.g., for paydowns, the No Risk Factor dot plots the coefficient on Low Capital 2023-on Extension , b(i),t t i,t × × whiletheothersaddthecoefficientontherelevantquadrupleinteraction). bank relative to the other. Low capital banks increased paydown requirements more than high capital banks for three of the four risk factors (large office being the exception), but reduced maturityextensionratesbylessforthreeofthefourriskfactors(lowdebtyieldbeingtheexception). Changes in non-maturity extension rates were mixed: low capital banks reduced non-maturity extensionsoflowdebtyieldandnonrecourseloansrelativetohighcapitalbanks,butincreasedoffice extensions relative to those banks. Across combinations of the four risk factors and three metrics for extension stringency, low and high capital banks are thus evenly split in whether they tightenedpoliciesforriskyloansrelativetotheother. Furthermore,thesedifferencesarealmostalways statisticallyinsignificant. 42

Differences in the performance of extended loans Table A.11 repeats the analysis from Table A.6analyzingtheex-postpayoffrateforextendedloans,butaddinganextrainteractionwithLow Capital . The specification with quarter fixed effects—which compares the performance of b(i),t extended loans relative to other loans in the same quarter—shows that the payoff rate of extended loans improved slightly during the stress period (by 0.6 percentage points) for well-capitalized banks, and an additional 2.4 percentage points for low capital banks. However, the difference betweenhighandlowcapitalbanksisstatisticallyinsignificantandmostlygoesawaywhenbankquarter fixed effects are added in column (3). Overall, there is no evidence that the probability of future repayment deteriorated for extensions made by low capital banks during the period of stress. To summarize the bank capital results, low and high capital banks alike reduced extensions and tightened extension terms, particularly for riskier loans. These results suggest that proximity to capital buffers did not induce banks to provide lenient policies to extend loans with little hope of repayment. Indeed, loans extended by low capital banks during the stress period performed comparablytoloansextendedbyhighercapitalbanks. 5. CONCLUSION This paper uses detailed supervisory data on bank CRE loans and a model with competing extension motivations (search frictions, foreclosure costs, temporary loss recognition costs) to assess why banks extend CRE loans. Before the pandemic, banks frequently extended loans secured by properties with strong cash flows and disproportionately required principal paydowns to extend weaker-income loans. These patterns are consistent with extensions helping borrowers navigate temporaryfrictionsinsellingorrefinancingpropertiestomeetmaturitypayments. Furthermore, I present four pieces of evidence against 2023 banking strains prompting banks to ease extension policies to delay loss recognition. First, extension frequencies were not materially different after 2022 relative to before the pandemic. Second, banks reduced extensions for the 43

income-strained properties where extend-and-pretend incentives would be the strongest. Third, extensiontermsbecamemorestringentduringthestressperiod. Ratherthanprovidehighlysubsidized credit, banks increasingly required equity contributions and higher spreads from borrowers toextendloans. Fourth,thepayoffprobabilityforstress-eraloanextensionsdidnotdeterioraterelativetoothermaturingloans,demonstratingthatmostextensionswereforfinanciallyviableloans. Moreover,thesefourpatternsholdevenforbanksclosertotheircapitalrequirements,inconsistent withcapitalpreservationincentivesmotivatingbankstoadoptlenientextensionpolicies. In short, observed extension patterns generally indicate that banks efficiently manage the risks associated with loan extensions. However, one significant limitation of this work is that it only covers larger banks, which tend to be less exposed to CRE loans and thus perhaps have weaker extend-and-pretendincentives. Incomplementarywork,GlancyandKurtzman(2024)demonstrate that while small banks have superior CRE loan performance, this difference is mostly attributable to loan portfolio composition, leaving little room for extend-and-pretend behavior to contribute to performance disparities. Taken together, these papers thus indicate that (i) large banks do not extend-and-pretend and (ii) small banks do not obscure delinquency relative to large banks. This lineofworkthereforeprovidessuggestiveevidenceagainstpervasiveextend-and-pretendbehavior; however,moredirectanalysisofservicingdecisionsatsmallbankswouldbevaluable. 44

References Acharya, V. V., M. Crosignani, T. Eisert, and S. Steffen (2022). Zombie lending: Theoretical, international,andhistoricalperspectives. AnnualReviewofFinancialEconomics14(1),21–38. Acharya, V. V., T.Eisert, C. Eufinger, and C. Hirsch(2019). Whatever ittakes: The realeffects of unconventionalmonetarypolicy. TheReviewofFinancialStudies32(9),3366–3411. Acharya,V.V.,S.Lenzu,andO.Wang(2021). Zombielendingandpolicytraps. Technicalreport, NationalBureauofEconomicResearch. Adalet McGowan, M., D. Andrews, and V. Millot (2018). The walking dead? Zombie firms and productivityperformanceinOECDcountries. EconomicPolicy33(96),685–736. An, X., Y. Deng, J. D. Fisher, and M. R. Hu (2016). Commercial real estate rental index: A dynamicpaneldatamodelestimation. RealEstateEconomics44(2),378–410. Black, L., J. Krainer, and J. Nichols (2017). From origination to renegotiation: A comparison of portfolio and securitized commercial real estate loans. The Journal of Real Estate Finance and Economics55(1),1–31. Black,L.,J.Krainer,andJ.Nichols(2020). Safecollateral,arm’s-lengthcredit: Evidencefromthe commercialrealestatemortgagemarket. TheReviewofFinancialStudies33(11),5173–5211. Bolton, P. and D. S. Scharfstein (1996). Optimal debt structure and the number of creditors. JournalofPoliticalEconomy104(1),1–25. Brown, D. T., B. A. Ciochetti, and T. J. Riddiough (2006). Theory and evidence on the resolution offinancialdistress. TheReviewofFinancialStudies19(4),1357–1397. Bruche, M. and G. Llobet (2014). Preventing zombie lending. The Review of Financial Studies27(3),923–56. 45

Caballero,R.J.,T.Hoshi,andA.K.Kashyap(2008). Zombielendinganddepressedrestructuring inJapan. AmericanEconomicReview98(5),1943–1977. Cattaneo, M. D., R. K. Crump, and M. H. Farrell (2024). On binscatter. American Economic Review114(5),1488–1514. Crosignani, M. and S. Prazad (2024). Extend-and-pretend in the U.S. CRE market. Federal ReserveBankofNewYorkStaffReports1130. Dinc, S. and E. Yo¨nder (2022). Strategic default and renegotiation: Evidence from commercial realestateloans. Technicalreport,WorkingPaper,RutgersUniversity. Faria-e Castro, M., P. Paul, and J. M. Sa´nchez (2024). Evergreening. Journal of Financial Economics153,103778. Favara,G.,C.Minoiu,andA.Perez-Orive(2024). ZombielendingtoUSfirms. FederalReserveSystem,OCC,FDIC,andNCUA(2023). Policystatementonprudentcommercial realestateloanaccommodationsandworkouts. Flynn, S., A. Ghent, and A. Tchistyi (2024). The imitation game: How encouraging renegotiation makesgoodborrowersbad. TheReviewofFinancialStudies37(12),3648–3709. Glancy, D., J. R. Krainer, R. J. Kurtzman, and J. B. Nichols (2022). Intermediary segmentation in thecommercialrealestatemarket. JournalofMoney,CreditandBanking54(7),2029–80. Glancy, D., R. Kurtzman, and L. Loewenstein (2025). The value of renegotiation frictions: Evidencefromcommercialrealestate. JournalofFinancialIntermediation62,101144. Glancy, D., R. Kurtzman, L. Loewenstein, and J. Nichols (2023). Recourse as shadow equity: Evidencefromcommercialrealestateloans. RealEstateEconomics51(5),1108–36. Glancy, D. and R. J. Kurtzman (2024). Determinants of recent CRE distress: Implications for the bankingsector. 46

Glancy, D. and J. C. Wang (2023). Lease expirations and CRE property performance. Federal ReserveBankofBostonResearchDepartmentWorkingPapersNo.23-10. Gupta, A., V. Mittal, and S. Van Nieuwerburgh (2026). Work from home and the office real estate apocalypse. AmericanEconomicReview116(2),674–709. Hackbarth, D., C. A. Hennessy, and H. E. Leland (2007). Can the trade-off theory explain debt structure? TheReviewofFinancialStudies20(5),1389–1428. Hinzen,F.J.,F.Severino,andS.VanNieuwerburgh(2025). Too-many-to-ignore: Regionalbanks andCRErisks. Jiang, E. X., G. Matvos, T. Piskorski, and A. Seru (2025). Monetary tightening, commercial real estatedistress,andUSbankfragility. JournalofPoliticalEconomyMacroeconomics3(4),525– 573. Myers, S. C. (1977). Determinants of corporate borrowing. Journal of Financial Economics 5(2), 147–175. Peek,J.andE.S.Rosengren(2005).Unnaturalselection: Perverseincentivesandthemisallocation ofcreditinJapan. AmericanEconomicReview95(4),1144–66. Rajan, R. G. (1992). Insiders and outsiders: The choice between informed and arm’s-length debt. TheJournalofFinance47(4),1367–1400. Sagi, J. S. (2021). Asset-level risk and return in real estate investments. The Review of Financial Studies34(8),3647–3694. 47

APPENDIX A. ADDITIONALTABLESANDFIGURES 48

FigureA.1: WhatDrivesExtensions 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(etaRnwodyaP 1.0 0.8 p∗ 0.6 0.4 0.2 0.0 ytilibaborPemoctuO Outcomes(RightAxis) Default Neglect Maintain Payoff (a)JustSearchCosts(Λ=1) 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(etaRnwodyaP 1.0 0.8 0.6 0.4 p∗ 0.2 0.0 ytilibaborPemoctuO Outcomes(RightAxis) Default Neglect Maintain Payoff (b)JustForeclosureCost(α =∞) 0.25 0.20 0.15 0.10 0.05 0.00 0.05 − 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DebtYield(n) )p(etaRnwodyaP 1.0 0.8 0.6 0.4 0.2 p∗ 0.0 ytilibaborPemoctuO Outcomes(RightAxis) Default Neglect Maintain Payoff (c)JustCapitalCosts(χ =.01,α =∞,Λ=1) Notes: Each chart presents a stacked area chart showing the probability that a loan with a given debt yield defaults(red), extendsandneglects(yellow), extendsandmaintains(green), orpaysoffatmaturity(blue). Black line plots p (n). Parameters are as in Table A.1, except only including one friction at a time. Panel ∗ (a)justhassearchcosts(Λ=1). Panel(b)justhasforeclosurecosts(α =∞). Panel(c)justhascapitalcosts (settingχ =.01andturningoffsearchandforeclosurecosts). 49

FigureA.2: LoanOutcomesbyTimetoMaturity S Y4 Yf Y: Yl z c3,N I #hN RHh8Rh3a @b T C0AQ88 2uj3N030 /38 nIj Al AS z S l k : 9 f e 4 O Sz SS Sl Sk S: S9 Sf Se S4 SO lz [n aj3achjRhK jnaCjw (a)Pre-COVID S Y4 Yf Y: Yl z c3,N I #hN RHh8Rh3a @b T C0AQ88 2uj3N030 /38 nIj Al AS z S l k : 9 f e 4 O Sz SS Sl Sk S: S9 Sf Se S4 SO lz [n aj3achjRhK jnaCjw (b)2023-on Notes: Thesefiguresreportloanoutcomesbythenumberofquarterstomaturity. Eachbarshowstheshare of outstanding loan balances that are paid off (blue), extended (yellow) and delinquent or liquidated (red). The top panel shows results for the years 2016-2019, and the bottom years from 2023-2025. Quarters to maturity is based on the previous quarter’s maturity date. For example, the 0 bar shows the outcomes for loansthatwerescheduledtomaturethatquarterasofthepreviousquarter. Idonotshowabarforballooned loanssincethatoutcomecanonlyoccurfollowingmaturity. 50

FigureA.3: OutcomesofPendingCRELoanMaturities,ByStabilization S f: f: fz fzX Y4 9f 99 99X 9k9k Yf T cjh/n3gHC\nC0 j30 9zX :O :O 9z 9z # 2 uj II 3 R N R 0 N 3 3 0 0 :9X :f:e :9 :e :9 :9 :9 :4 :f :e :9 Y: T C0hQ88 :: :: :k :: :k :: :zX :z :z :S :l :l kO Yl k4 k4 k9X z kzX lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (a)MaturityOutcomes,StabilizedProperties 030N3ju2h3a @b lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (b)ExtensionRates,StabilizedProperties S fzX 9O9494 99X 99 99 Y4 9zX 9z :4 9z 9z 9z :4:O:O :4 9S 9z 9S :4 9l Yf T # c IIR jh R / N n 3 3 0 gHC\nC0 j30 :9X :9 :e :9 :f:f :e :f 2uj3N030 :: Y: T C0hQ88 :k :k :k :zX :S :l :z kOkO k4 Yl k9X z kzX lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (c)MaturityOutcomes,Non-stabilizedProperties 030N3ju2h3a @b lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (d)ExtensionRates,Non-stabilizedProperties Notes: The left figures show the share of outstanding loan balances that are paid off (blue), extended (yellow),performingpasttheirmaturitydate(orange)andpastdueorliquidated(red)bythequarterofscheduled maturity. Therightpanelsshowtheshareofbalancesthatareextended,correspondingtotheyellowregion intheleftcharts. Thetoppanelspertaintostabilizedproperties,andthebottompanelsnon-stabilizedproperties. Loan balances and scheduled maturity dates are measured as of four quarters before the scheduled maturity. 51

FigureA.4: OutcomesofPendingCRELoanMaturities,ByPropertyType S fk fz fzX 94 94 Y4 99X 9:9:9: 9: 9: 99 9l 9S 9S9l 9S Yf T cjh/n3gHC\nC0 j30 9zX :4 :O :e :4 :O :e # IIRRN30 :9X :f :9 :9 :f 2uj3N030 Y: T C0hQ88 :zX :S :k :S :k:l :S :k :z k4 k9X Yl k: kzX z le lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (a)MaturityOutcomes,OfficeProperties 030N3ju2h3a @b lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (b)ExtensionRates,OfficeProperties S fl fzX fz9O Y4 99X 99 9: Yf T # 2 uj c II 3 R jh N R / 0 N n 3 3 3 0 0 gHC\nC0 j30 : 9 9 z X X :f:e :9 :e :f 9S :9:f :e :f :f 9z 9S 9S :4 9z :4 :f :4 Y: T C0hQ88 :k :k :: :k :l :: :: :k :l :zX Yl kOk4 k4 k9X z kzX lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (c)MaturityOutcomes,Non-officeProperties 030N3ju2h3a @b lzSe lzS4 lzSO lzlz lzlS lzll lzlk lzl: lzl9 K jnaCjwh/ j3 (d)ExtensionRates,Non-officeProperties Notes: The left figures show the share of outstanding loan balances that are paid off (blue), extended (yellow),performingpasttheirmaturitydate(orange)andpastdueorliquidated(red)bythequarterofscheduled maturity. Therightpanelsshowtheshareofbalancesthatareextended,correspondingtotheyellowregion in the left charts. The top panels pertain to office properties, and the bottom panels non-office properties. Loanbalancesandscheduledmaturitydatesaremeasuredasoffourquartersbeforethescheduledmaturity. 52

FigureA.5: MaturityExtensionRatesbyLoanSize 9z :9 :z k9 kz 3a @chNRCcN3ju2 Ua3A+QpB/ lzlkARN S lY9 9 Sz lz :z 4z Sfz HR Nh# I N,3.hLCIICRNchVIR<hc, I3W Notes: Thisfigureplotstheestimatedprobabilityofextensionasafunctionofaloan’ssizeandbank-quarter fixedeffects. Eachdotestimatestheshareofloanswithpendingmaturitiesthatgetextendedbydecileofthe outstandingloanbalance(onalogarithmicscale). Estimatesforloansslatedtomaturebeforethepandemic areinblueandduringthestressperiodinred. 53

FigureA.6: MaturitiesinNextTwoYears Y: Yk Yl YS z c3,N I $hN RHh8Rh3a @b K jnaCN<hc@ a3hVNRjha3,3NjIwh3uj3N030W K jnaCN<hc@ a3hhVa3,3Njh3uj3NcCRNcW b@ a3hR8hL jnaCjC3ch8aRLh3uj3NcCRNc lzS: lzSf lzS4 lzlz lzll lzl: lzlf Notes: This figure plots the share of outstanding CRE loans that are scheduled to mature in the next two years over time (the top of the blue area), and the portion of those loans accounted for by loans that have received extensions in the past two years (the red area), and the share of pending maturities coming from recentextensions(whitedots). FigureA.7: ExtensionsbyPropertyTypeandPeriod Yf Y9 Y: Yk Yl YS z 030N3ju2h3a @b Ta3A+QpB/ Yf lzlzAlzll lzlkARN Y9 Y: Yk Yl YS z ?Rj3I `3j CI Q88C,3 KnIjC8 LCIw BN0ncjaC I TaRU3ajwhiwU3 (a)Weighted 030N3ju2h3a @b Ta3A+QpB/ lzlzAlzll lzlkARN ?Rj3I `3j CI Q88C,3 KnIjC8 LCIw BN0ncjaC I TaRU3ajwhiwU3 (b)Unweighted Notes: Each panel plots the share of loans with pending (four quarters ahead) maturities that get extended by property type. Blue, orange, and red bars give extension rates before, during and after the pandemic, respectively. Theleftpanelweightsbyloanbalanceandtherightisunweighted. 54

Table A.1: Parameters used in baseline model Parameter Description Value r Discountrate 0.045 r Mortgagerate 0.07 m f Externalfundsavailable 0.15 g ExpectedNOIGrowth 0.01 σ StandardDeviationofNOIGrowth 0.1 α ParetoShapeParameter 12.3 Λ RecoveryFactor 0.76 θ DeclineinNOIfromNeglect 0.045 ν TemporaryNOIBoostFromNeglect 0.69 χ CosttoLossRecognition 0 Notes: r=4.5%issettomatch30-yearTreasuryyieldsintheperiodofstress,andr =7%tomatchspreads m in Glancy et al. (2022). α is based on search frictions estimated in Sagi (2021). g and σ are set to match the statistics on annual rent growth in An et al. (2016). Λ and θ are based on foreclosure costs in Brown etal.(2006). ν issetsothatdeferredmaintenanceisanequal-sizedtransferanddeadweightloss. f issetto matchthe95thpercentiledifferencebetweenloanpaymentsandnetincomeinthe2023-onextensiondata. Inthebaselinecalibration,Isetχ =0basedonFavaraetal.(2024). SeeAppendixB.4fordetails. 55

Table A.2: Select Summary Statistics Mean Std p10 p50 p90 N 2023-on 0.48 0.50 0.00 0.00 1.00 1,785,400 t Maturing 0.03 0.17 0.00 0.00 0.00 1,785,400 i,t Extension 0.03 0.16 0.00 0.00 0.00 1,683,672 i,t+1 Default 0.01 0.12 0.00 0.00 0.00 1,683,672 i,t+1 Paidoff 0.19 0.39 0.00 0.00 1.00 1,683,672 i,t+1 DebtYield 0.12 0.05 0.07 0.11 0.20 615,271 i,t LowDebtYield 0.07 0.26 0.00 0.00 0.00 1,785,400 i,t Nonrecourse 0.34 0.47 0.00 0.00 1.00 1,785,400 i,t SmallOffice 0.11 0.31 0.00 0.00 1.00 1,785,400 i,t LargeOffice 0.01 0.11 0.00 0.00 0.00 1,785,400 i,t MissingDebtYield 0.66 0.48 0.00 1.00 1.00 1,785,400 i,t Stabilized 0.70 0.46 0.00 1.00 1.00 1,785,400 i,t Paydown 0.01 0.03 0.00 0.01 0.01 1,505,890 i,t 1(Paydown>5%) 0.01 0.11 0.00 0.00 0.00 1,505,890 i,t 1(∆Balance>0) 0.04 0.19 0.00 0.00 0.00 1,505,890 i,t GainedRecourse 0.01 0.09 0.00 0.00 0.00 526,040 i,t ∆Spread 0.00 0.00 0.00 0.00 0.00 800,672 i,t ∆1(Spread>0) 0.04 0.19 0.00 0.00 0.00 800,672 i,t ∆1(Spread<0) 0.03 0.17 0.00 0.00 0.00 800,672 i,t LowCapital 0.38 0.48 0.00 0.00 1.00 1,436,889 b(i),t Notes: Excludesyears2020-2022toremainconsistentwithregressionsamples. Paidoff ismissingfor i,t+1 2025:Q4 observations and for loans that exited the sample for non-standard reasons (e.g., balances falling below the reporting threshold). Stabilized is an indicator for whether the property is not a construction i,t loan, andthevalueisreportedasis(ratherthanascompletedorasstabilized). DebtYield (netoperating i,t incomeovertheloanbalance)isbottom-andtop-codedat0and0.2,respectively. MissingDebtYield=1for nonstabilized loans or loans where NOI has not been updated in the last year. Paydown variables exclude quarters with charge-offs or changes in utilization. Paydown is bottom- and top-coded at -.5 and .5, i,t respectively. GainedRecourse ismissingforloansthatalreadyhadrecourse. Variablesmeasuringchanges i,t inspreadsaremissingforloansthatdonothavefloatingrates. Bankcapitalismissingwhenbanksfirstenter thesampleasreportingbeginsbeforethefirststresstestresultscomeout. 56

Table A.3: Select Summary Statistics 2023-on =0 2023-on =1 Maturing =1 Extension =1 t t i,t i,t (1) (2) (3) (4) 2023-on 0.00 1.00 0.43 0.37 t Maturing 0.03 0.03 1.00 0.26 i,t Extension 0.03 0.02 0.52 0.16 i,t+1 Default 0.01 0.02 0.08 0.04 i,t+1 Paidoff 0.21 0.16 0.56 0.35 i,t+1 DebtYield 0.12 0.11 0.12 0.12 i,t LowDebtYield 0.08 0.07 0.05 0.06 i,t Nonrecourse 0.29 0.39 0.26 0.28 i,t SmallOffice 0.12 0.09 0.15 0.13 i,t LargeOffice 0.01 0.01 0.03 0.03 i,t MissingDebtYield 0.63 0.68 0.75 0.70 i,t Stabilized 0.68 0.72 0.42 0.41 i,t Paydown 0.01 0.01 0.01 -0.00 i,t 1(Paydown>5%) 0.01 0.01 0.03 0.08 i,t 1(∆Balance>0) 0.02 0.07 0.06 0.13 i,t GainedRecourse 0.01 0.00 0.02 0.05 i,t ∆Spread 0.00 0.00 0.00 0.00 i,t ∆1(Spread>0) 0.03 0.05 0.05 0.12 i,t ∆1(Spread<0) 0.03 0.03 0.03 0.12 i,t LowCapital 0.35 0.42 0.46 0.48 b(i),t Observations 933,209 852,191 52,343 47,442 Notes: Each column presents the mean of the variables of interest for a subset of the data. Column (1) summarizes variables in the prepandemic period, column (2) for the stress period, (3) for loans that are scheduledtomaturenextperiod,and(4)forloansthatwereextendedinagivenquarter. 57

Table A.4: Extensions During the Stress Period 100 Extension 100 Default i,t+1 i,t+1 × × (1) (2) (3) (4) (5) (6) Maturing 2023-on -5.73** -3.61* -3.65* 4.39** 2.04* 1.52+ i,t t × (1.13) (1.65) (1.66) (0.63) (0.90) (0.90) ... LowDebtYield -7.84* -6.91* 5.72** 5.46** i,t × (3.36) (3.32) (2.07) (2.03) ... Nonrecourse -4.58* -4.91* 3.29** 3.79** i,t × (2.15) (1.97) (1.12) (1.07) ... SmallOffice 0.22 0.18 2.43* 2.04+ i,t × (1.66) (1.63) (1.08) (1.09) ... LargeOffice 3.21 3.01 8.15** 7.64** i,t × (3.29) (3.27) (2.19) (2.18) Maturing 53.18** 54.74** 53.87** 5.32** 4.75** 4.57** i,t (0.60) (0.99) (1.04) (0.31) (0.45) (0.46) ... LowDebtYield 5.34* 4.29+ 3.16** 3.35** i,t × (2.60) (2.57) (1.02) (0.99) ... Nonrecourse 1.29 1.07 -0.96 -1.54** i,t × (1.08) (1.08) (0.59) (0.57) ... SmallOffice -2.61** -2.00* 1.62** 2.44** i,t × (0.97) (0.95) (0.48) (0.52) ... LargeOffice 1.09 2.14 -3.24** -2.31** i,t × (2.21) (2.21) (0.67) (0.70) 2023-on -0.58** -0.23 0.84** -0.27+ i,t (0.21) (0.22) (0.18) (0.15) ... LowDebtYield 0.44* 0.19 1.72** 1.68** i,t × (0.21) (0.17) (0.39) (0.35) ... Nonrecourse -0.56* -0.02 1.20** 0.77** i,t × (0.28) (0.15) (0.39) (0.16) ... SmallOffice 0.25+ 0.33* 1.76** 1.60** i,t × (0.15) (0.13) (0.15) (0.14) ... LargeOffice 1.40** 1.39** 13.66** 13.61** i,t × (0.52) (0.42) (0.69) (0.71) R2 0.286 0.286 0.309 0.018 0.025 0.075 a Observations 1,683,672 1,683,672 1,683,671 1,683,672 1,683,672 1,683,671 BankFE? ✓ ✓ ✓ ✓ BankQuarterFE? ✓ ✓ Controls ✓ ✓ ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 Extension =(β X ) 2023-on Maturing +γ LowerLevelControls +τ +ε × i,t+1 ′ i,t × t × i,t ′ i,t b(i) i,t Thedependentvariableisanindicatorforwhethertheloangetsextendedinthefollowingquarterin(1)–(3) and an indicator for whether it is delinquent in (4)–(6). These outcomes are not mutually exclusive. 2023on is1forquartersstartingin2023:Q1and0beforethepandemic. Maturing isanindicatorforwhethera t i,t loanisscheduledtomaturenextquarter. X includesindicatorsforwhethertheloanhasadebtyieldunder i,t 8%, is nonrecourse, is secured by a small-sized office (under 250,000 square feet in size), is secured by a large office, and an indicator for whether debt yield is missing (for non-stabilized loans or loans with stale incomereporting). CoefficientsforuninteractedX controls,andallinteractionswiththemissingdebtyield i,t indicatorarenotdisplayed. Relativetothelistedspecification,columns(1)and(4)omittheinteractionwith X , and columns (3) and (6) add bank-quarter fixed effects. Standard errors, in parentheses, are clustered i,t bybank-quarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 58

Table A.5: Extension Terms by Risk Factors Paydown 1(Paydown>5%) ∆Balance>0 GainedRecourse ∆Spread 1(∆Spread>0) 1(∆Spread<0) (1) (2) (3) (4) (5) (6) (7) Extensioni,t 2023-oni,t 2.95** 2.42** -5.63** 1.42 0.12** 13.65** -9.59** × (0.54) (0.77) (1.97) (0.94) (0.02) (1.38) (1.61) ... LowDebtYieldi,t -1.28 4.08 -4.26 -0.95 -0.02 -8.14** -0.88 × (0.86) (2.56) (3.67) (1.32) (0.04) (2.99) (2.82) ... Nonrecoursei,t 0.61 2.34+ -3.19 0.02 1.95 -1.62 × (0.59) (1.20) (2.34) (0.02) (1.45) (1.32) ... SmallOfficei,t 0.56 3.66** -0.67 1.87 0.08** 8.12** -0.83 × (0.41) (1.14) (1.44) (1.50) (0.02) (1.47) (1.14) ... LargeOfficei,t 1.28 13.29** -12.08** 2.45 0.10** 11.52** -3.12 × (0.92) (2.89) (3.76) (1.63) (0.04) (2.36) (2.68) Extensioni,t -4.04** 2.27** 15.69** 1.69** -0.04** 6.72** 16.92** (0.48) (0.35) (1.52) (0.55) (0.01) (0.71) (1.43) ... LowDebtYieldi,t 3.81** 4.05* -2.07 -0.63 0.03 6.50** -0.98 × (0.70) (1.58) (3.30) (0.67) (0.03) (2.24) (2.51) ... Nonrecoursei,t -0.47 -0.17 4.00* 0.00 0.17 0.66 × (0.51) (0.61) (1.82) (0.01) (0.62) (0.97) ... SmallOfficei,t 0.35 -0.92+ -0.54 0.92 -0.00 -0.33 -0.72 × (0.31) (0.49) (1.02) (0.88) (0.01) (0.71) (0.85) ... LargeOfficei,t 1.73* -0.03 -1.80 -2.34* -0.03* -1.60 1.94 × (0.75) (1.40) (2.61) (0.97) (0.01) (1.03) (2.19) 2023-oni,t ... LowDebtYieldi,t 0.02 0.11 0.44 -0.14 -0.00 -3.02** -2.21** × (0.03) (0.16) (0.75) (0.13) (0.01) (0.94) (0.80) ... Nonrecoursei,t 0.00 -0.05 -0.64 0.00+ -0.19 -0.42 × (0.03) (0.16) (1.26) (0.00) (0.48) (0.43) ... SmallOfficei,t 0.12** 0.26** -0.68* 0.13 0.00* 1.11** 0.69** × (0.02) (0.10) (0.32) (0.24) (0.00) (0.26) (0.25) ... LargeOfficei,t 0.05 0.54 14.27** 0.82 0.02** 1.15* 0.17 × (0.10) (0.33) (1.77) (0.55) (0.00) (0.51) (0.39) R2 0.057 0.069 0.142 0.168 0.105 0.287 0.192 a Observations 1,356,810 1,356,810 1,356,810 525,523 700,524 700,524 700,524 Bank-QuarterFE? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Controls ✓ ✓ ✓ ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 ∆Term =(β X ) 2023-on Extension +γ LowerLevelControls +τ +ε i,t ′ i,t 1 t i,t ′ i,t 1 b(i),t i,t × − × × − wherethedependentvariableisthechangeinsomeloanterm,scaledby100soestimatesareinpercentage points. These dependent variables are as described in Table 2. The main independent variables are an indicator for whether the loan receives an extension and an indicator for whether the quarter is 2023 or later. X includes indicators for whether the loan has a debt yield under 8%, is nonrecourse, is secured i,t 1 − byasmall-sizedoffice(under250,000squarefeetinsize),issecuredbyalargeoffice,andanindicatorfor whether debt yield is missing (for non-stabilized loans or loans with stale income reporting). Coefficients foruninteractedX controls;controlsforage,size,amortization,andpreviousspread;andallinteractions i,t 1 − withthemissingdebtyieldindicatorarenotdisplayed. τ isabank-quarterfixedeffect. Standarderrors, b(i),t inparentheses,areclusteredbybank-quarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 59

Table A.6: Payoffs By Previous Extension 100 PaidOff i,t+1 × (1) (2) (3) Extended -2.09** -2.72** -2.45** i,t (0.46) (0.42) (0.36) Extended2023-on -2.25** 2.24** 1.44** i,t (0.78) (0.66) (0.51) R2 0.112 0.127 0.142 a Observations 1,681,663 1,681,663 1,681,662 Quarter-to-MaturityFE? ✓ ✓ ✓ BankFE? ✓ ✓ QuarterFE? ✓ BankQuarterFE? ✓ Notes: Thistablepresentsestimatesfromtheequation 100 PaidOff =β Extended +β Extended2023-on +τ +α +ε i,t+1 1 i,t 2 i,t b(i),t m(i,t) i,t × The dependent variable is an indicator for whether an outstanding loan paid off the following quarter. Extended is an indicator for whether the loan was previously extended, and Extended 2023-on is an i,t i,t indicator for whether the loan was extended during the period of stress. τ and α are bank-quarter b(i),t m(i,t) and quarters-to-maturity fixed effects. Relative to the specification above, columns (1) and (2) replace the bank-quarterfixedeffectswithbankfixedeffectsandbankandquarterfixedeffects,respectively. Standard errors, in parentheses, are clustered by bank-quarter. +, , indicate significance at 10%, 5%, and 1%, re- ∗ ∗∗ spectively. 60

Table A.7: Maturity Extensions By Capitalization 100 Extension 100 Default i,t+1 i,t+1 × × (1) (2) LowCapital 0.02 0.13 b(i),t (0.25) (0.12) ... Maturing 2023-on -0.40 0.83 i,t t × × (2.68) (1.52) ... Maturing 1.45 -2.75** i,t × (1.24) (0.82) ... 2023-on -0.41 -0.15 t × (0.26) (0.20) Maturing 2023-on -4.29* 4.65** i,t t × (1.80) (1.00) Maturing 51.85** 5.77** i,t (0.87) (0.35) 2023-on -0.35** 0.71** i,t (0.10) (0.10) R2 0.294 0.062 a Observations 1,412,843 1,412,843 BankFE? ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 Extension =βLowCapital 2023-on Maturing +γ LowerLevelControls +τ +ε × i,t+1 b(i),t × t × i,t ′ i,t b(i) i,t The dependent variable is an indicator for whether the loan gets extended in the following quarter in (1) and an indicator for whether it is delinquent in (2). LowCapital is 1 for banks closer than the median b(i),t to their capital buffer (determined by the banks’ minimum CET1 ratio in the severely adverse scenario in the pre-SCB era and the distance to the bank-specific CET1 capital buffer, inclusive of the SCB and G- SIB surcharge, in the SCB-era). 2023-on is 1 for quarters starting in 2023:Q1 and 0 before the pandemic. t Maturing isanindicatorforwhetheraloanisscheduledtomaturenextquarter. Standarderrors,inpareni,t theses,areclusteredbybank-quarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 61

Table A.8: Extension Terms By Capitalization Paydown 1(Paydown>5%) ∆Balance>0 GainedRecourse ∆Spread 1(∆Spread>0) 1(∆Spread<0) (1) (2) (3) (4) (5) (6) (7) LowCapital b(i),t× Extensioni,t 0.33 -0.51 -2.86+ -1.13 0.01 -0.67 -3.18* (0.52) (0.69) (1.71) (1.10) (0.01) (0.81) (1.28) ... 2023-ont 0.06 2.70* 3.80+ -0.23 0.00 4.32+ 4.46** × (0.61) (1.17) (2.25) (1.78) (0.02) (2.47) (1.68) Extensioni,t -2.22** 3.32** 12.18** 2.69** -0.03** 5.53** 13.50** (0.29) (0.42) (1.11) (0.91) (0.01) (0.43) (0.96) ... 2023-ont 2.87** 4.92** -8.80** 1.98 0.10** 9.60** -9.88** × (0.34) (0.69) (1.48) (1.35) (0.01) (1.84) (1.27) R2 0.053 0.072 0.133 0.177 0.120 0.298 0.137 a Observations 1,086,286 1,086,286 1,086,286 394,537 575,108 575,108 575,108 Bank-quarterFE? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Controls? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 ∆Term =βExtension 2023-on LowCapital +γ LowerLevelControls +τ +ε × i,t i,t × t × b(i),t ′ i,t − 1 b(i),t i,t wherethedependentvariableisthechangeinsomeloanterm,scaledby100soestimatesareinpercentage points. ThesedependentvariablesareasdescribedinTable2. Themainindependentvariablesareindicators for whether the loan receives an extension, the time period is 2023 or later, and the loan is held by a bank holding company that is closer than the median to its regulatory capital constraint. Unreported controls include interactions of the main variables of interest that do not include Extension , and controls for a i,t loan’s age, size, amortization, andprevious spread. τ is abank-quarter fixedeffect. Standard errors, in b(i),t parentheses,areclusteredbybank-quarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 62

Table A.9: Maturity Extensions By Capitalization and Risk Characteristics 100 Extension 100 Default i,t+1 i,t+1 × × (1) (2) LowCapital Maturing 2023-on 2.28 (3.94) 6.40** (1.87) b(i),t i,t t × × ... LowDebtYield -3.82 (7.21) -0.77 (4.16) i,t × ... Nonrecourse 5.49 (4.51) -4.93* (2.31) i,t × ... SmallOffice -0.46 (3.52) -3.24 (2.49) i,t × ... LargeOffice -3.31 (7.20) -4.89 (4.70) i,t × LowCapital 2023-on b(i),t t × ... LowDebtYield -0.58 (0.50) 0.54 (0.72) i,t × ... Nonrecourse -0.33 (0.37) 0.99* (0.39) i,t × ... SmallOffice 0.75* (0.34) -0.64+ (0.34) i,t × ... LargeOffice 1.51+ (0.90) 5.13** (1.58) i,t × LowCapital Maturing -0.01 (2.19) -4.40** (0.82) b(i),t i,t × ... LowDebtYield 0.66 (5.25) -0.32 (1.95) i,t × ... Nonrecourse -0.02 (2.23) 2.33+ (1.22) i,t × ... SmallOffice 1.11 (2.00) 2.95** (1.06) i,t × ... LargeOffice -2.32 (4.53) 2.85* (1.43) i,t × LowCapital b(i),t ... LowDebtYield 0.37 (0.45) 0.01 (0.18) i,t × ... Nonrecourse 0.37 (0.26) -0.30+ (0.16) i,t × ... SmallOffice -0.65* (0.29) 0.47** (0.12) i,t × ... LargeOffice -1.12+ (0.68) -0.89** (0.20) i,t × R2 0.314 0.084 a Observations 1,412,843 1,412,843 Bank-QuarterFE? ✓ ✓ LowerLevelInteractions ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 Extension =(β X )LowCapital 2023-on Maturing +γ LowerLevelControls +τ +ε × i,t+1 ′ i,t b(i),t × t × i,t ′ i,t b(i),t i,t The dependent variable is an indicator for whether the loan gets extended in the following quarter in (1) and an indicator for whether it is delinquent in (2). LowCapital is 1 for banks closer than the median b(i),t to their capital buffer (determined by the banks’ minimum CET1 ratio in the severely adverse scenario in the pre-SCB era and the distance to the bank-specific CET1 capital buffer, inclusive of the SCB and G- SIB surcharge, in the SCB-era). 2023-on is 1 for quarters starting in 2023:Q1 and 0 before the pandemic. t Maturing is an indicator for whether a loan is scheduled to mature next quarter. X includes indicators i,t i,t for whether the loan has a debt yield under 8%, is nonrecourse, is secured by a small-sized office (under 250,000squarefeetinsize), issecuredbyalargeoffice, andanindicatorforwhetherdebtyieldismissing (fornon-stabilizedloansorloanswithstaleincomereporting). Coefficientsforvariablesnotinteractedwith LowCapital aswellasallinteractionswiththemissingdebtyieldindicatorarenotdisplayed. Standard b(i),t errors, in parentheses in the right columns for each model, are clustered by bank-quarter. +, , indicate ∗ ∗∗ significanceat10%,5%,and1%,respectively. 63

Table A.10: Extension Terms By Capitalization and Risk Characteristics Paydown 1(Paydown>5%) ∆Balance>0 GainedRecourse ∆Spread 1(∆Spread>0) 1(∆Spread<0) (1) (2) (3) (4) (5) (6) (7) LowCapitalb(i),t× Extensioni,t × 2023-ont -0.48 1.74 8.49+ 0.80 -0.08+ -2.09 8.23* (1.26) (1.90) (4.73) (2.13) (0.04) (3.42) (3.41) ... × LowDebtYieldi,t -1.91 -0.52 6.83 -3.89 -0.05 3.84 5.64 (1.87) (5.58) (6.95) (2.72) (0.09) (6.55) (5.28) ... × Nonrecoursei,t 3.14* 4.56 -13.55* 0.06 1.87 0.64 (1.28) (3.06) (5.62) (0.04) (4.01) (2.85) ... × SmallOfficei,t -0.31 -0.15 1.63 -4.32 0.03 6.53+ 0.36 (0.92) (3.04) (3.22) (3.69) (0.04) (3.40) (2.44) ... × LargeOfficei,t -3.50 -13.10+ 10.39 -4.55 -0.13 -9.14+ 9.53+ (2.18) (7.11) (8.12) (4.06) (0.08) (5.51) (5.28) Extensioni,t × 2023-ont 3.59** 2.62* -8.54** 0.63 0.17** 16.51** -15.26** (0.61) (1.25) (2.63) (1.72) (0.02) (2.29) (2.05) ... × LowDebtYieldi,t -0.21 3.73 -7.37 0.49 -0.02 -10.09* -2.63 (1.03) (3.27) (5.64) (2.23) (0.05) (4.15) (3.90) ... × Nonrecoursei,t -0.23 1.23 0.46 0.02 1.67 -2.42 (0.66) (1.59) (3.05) (0.02) (2.36) (2.15) ... × SmallOfficei,t 0.63 4.19* -1.17 4.79+ 0.05* 4.96* -0.37 (0.56) (1.73) (2.08) (2.79) (0.02) (2.26) (1.51) ... × LargeOfficei,t 2.68* 22.03** -17.16** 5.27+ 0.16** 16.02** -5.34 (1.35) (4.66) (5.87) (2.86) (0.05) (4.20) (3.43) R2 0.058 0.074 0.143 0.178 0.121 0.301 0.139 a Observations 1,086,286 1,086,286 1,086,286 394,537 575,108 575,108 575,108 Bank-QuarterFE? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Controls? ✓ ✓ ✓ ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation 100 ∆Term =(β X )Extension 2023-on LowCapital +γ LowerLevelControls +τ +ε × i,t ′ i,t − 1 i,t × t × b(i),t ′ i,t − 1 b(i),t i,t wherethedependentvariableisthechangeinsomeloanterm,scaledby100soestimatesareinpercentage points. ThesedependentvariablesareasdescribedinTable2. Themainindependentvariablesareinteractionsbetweenindicatorsforwhether(i)thebankholdingtheloaniscloserthanthemediantoitsregulatory capital constraint, (ii) the loan was extended in quartert, (iii) the observation is from the stress period and (iv)ariskfactor: lowdebtyield,missingdebtyield,nonrecourse,andsmallandlargeofficeindicators. All specifications include lower level interactions of all variables; controls for a loan’s age, size, amortization andpreviousspread;andbank-quarterfixedeffects. Standarderrors,inparentheses,areclusteredbybankquarter. +, , indicatesignificanceat10%,5%,and1%,respectively. ∗ ∗∗ 64

Table A.11: Payoffs By Previous Extension and Capitalization 100 PaidOff i,t+1 × (1) (2) (3) LowCapital 0.00 0.34 b(i),t (0.70) (0.74) ...xExtended -2.42+ -2.03* -2.00* i,t (1.40) (1.02) (0.83) ...xExtended2023-on 2.84 2.36 0.52 i,t (1.92) (1.53) (1.54) Extended -1.85 -1.99** -1.79** i,t (1.43) (0.63) (0.54) Extended2023-on -0.45 0.60 1.55 i,t (2.15) (1.28) (0.98) R2 0.119 0.123 0.137 a Observations 1,410,995 1,410,995 1,410,995 Quarter-to-MaturityFE? ✓ ✓ ✓ BankFE? ✓ ✓ QuarterFE? ✓ BankQuarterFE? ✓ Notes: Thistablepresentsestimatesfromtheequation 100 PaidOff =βExtended2023-on LowCapital +γ LowerLevelControls +τ +α +ε × i,t+1 i,t × b(i),t ′ i,t b(i),t m(i,t) i,t The dependent variable is an indicator for whether the loan is paid off next quarter. Extended 2023-on i,t is an indicator for whether the loan was extended during the stress period, Low Capital is an indicator b(i),t for whether the bank is closer than the median to its capital buffer. τ and α are bank-quarter and b(i),t m(i,t) quarters-to-maturity fixed effects. Lower Level Controls includes Extended 2023-on , Extended , Low i,t i,t i,t Capital , and Extended Low Capital . Relative to the specification above, columns (1) and (2) b(i),t i,t b(i),t × replacethebank-quarterfixedeffectswithbankfixedeffectsandbankandquarterfixedeffects,respectively. Standarderrors,inparentheses,areclusteredbybank-quarter. +, , indicatesignificanceat10%,5%,and ∗ ∗∗ 1%,respectively. 65

B. MODELAPPENDIX B.1. ExpectationDerivations Expectations over Sale Offers Whether borrowers choose to extend or sell depends on the sale offerborrowersreceive. Ifnishighenoughthatitispossibleforborrowerstowanttosell,maturity outcomesarestochasticandexpectedpayoutscomefromintegratingoverκ. First, I solve for the critical κ (n) below which borrowers with a debt yield of n choose to sell. ∗ If borrowers reject a sale offer, they will either default or extend, making their outside option to selling V+ (n) max V (n),V (n),0 . The optimal κ is such that borrowers are b,ext ≡ { b,maintain b,neglect } ∗ indifferent between selling and that outside option. This occurs when n/κ (n) (1+r )=V+ , ∗ − m b,ext meaningthat n κ (n)= ∗ 1+r +V+ (n) m b,ext If κ (n) < κ, there is no chance of a sale, and the borrower either defaults or extends. That is, ∗ eitherπ (n)=1orπ (n)=1dependingonwhetheramutuallybeneficialextensionisfeasible. ext def (cid:16) (cid:17)α Ifκ (n) κ,theprobabilityofasaleisπ (n)=G(κ (n))=1 κ . ∗ ≥ sale ∗ − κ ∗ (n) The borrower’s value function in the region where sales are possible comes from integrating over potentialoffers: (cid:90) κ ∗ (n)(cid:16)n (cid:17) (cid:90) ∞ V (n)= (1+r ) g(κ)dκ+V+ (n) g(κ)dκ b κ − m b,ext κ κ ∗ α n 1 n (cid:18) κ (cid:19)α = (1+r ) + m 1+α κ − 1+α κ (n) κ (n) ∗ ∗ (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) ExpectedReturnfromForcedSale OptionValueofDefault/Extension The lender’s value function comes from applying the sale, extension and default probabilities to equation(1). 66

Probability of Eventual Default Let π , π , and π be the vectors of probabilities that borm n def rowers choose extend-maintain, extend-neglect, and default on a grid of debt yield values. Let P and P denote the transition matrices that give the probability that a borrower who chooses m n extend-maintain and extend-neglect, respectively, transitions from debt yield i to debt yield j next period. Then the probability that a borrower with a given debt yield i extends in a given period and winds upatdebtyield j isgivenby: P diag(π )P +diag(π )P m m n n ≡ andtheprobabilitythattheborrowerultimatelywindsupdefaultingisgivenby: t=T Pr(EventualDefault)= lim ∑ Pt π =(I P) 1π def − def T ∞ − → t=0 Theseequationstakethedynamicsfordebtyieldimpliedbythemodel(P andP )andthesolution n m forequilibriumdefaultandextensionprobabilitiesforagivendebtyield(π ,π andπ )andgive def m n theprobabilityofeventualdefaultafterapotentialstringofextensions. B.2. NumericalSolution 1. Makeaninitialguessforborrowers’andlenders’values,V andV,onagridofdebtyields b l n. 2. Take expectations over a lognormal distribution to calculate continuation values implied by thosevaluefunctions(solvingforV (n)andV (n)). b ′ l ′ (a) Note that n = µZn, where µ = (1 θ1[neglect])(1+g)/(1 p), and Z is a log- ′ − − normally distributed variable such that E(Z) = 1. Since the effects of paydowns, neglect,andvalueappreciationintermsofnormalizedcontinuationvaluesareisomorphic to a change in initial NOI, one can take a single expectation (for µ = 1) and use that 67

functiontofindcontinuationvaluesassociatedwithotheroutcomes. (b) Expectations are estimated by Gauss-Hermite quadrature, interpolating between grid points. For quadrature points falling off the grid, I linearly extrapolate from the last two grid points to calculate lenders’ value functions below the grid and borrowers’ value functions above the grid. Other off-grid values are assumed to stay at the value forthelastgridpoint. 3. Use V (n) and V (n) to find borrowers’ and lenders’ value functions for a given action b ′ l ′ a(n,κ) Extend Neglect,Maintain ,Default,Payoff p. ∈{ ×{ } }× 4. Solveforborrowers’optimalactionsa (n,p,κ). ∗b 5. Solveforlenders’optimal p (n),givena (n,p,κ). ∗ ∗b 6. Update value functions based on a (n,p,κ) and p (n). Integrating over the Pareto distri- ∗b ∗ bution for κ gives the ex-ante values for borrowers and lenders, V and V (see Section b l B.1). 7. Checkforconvergence,otherwisereturntostep1withtheupdatedV andV. b l B.3. CharacterizationofEquilibrium B.3.1. Borrowers’problem Iwillstartbydiscussingborrowers’optimaldecisiontodefault,neglect,ormaintaingivenaparticular debt yield and paydown requirement. I abstract from sale decisions here and just characterize which outcomes borrowers select as a function of n and p when a sale offer is not worth taking. Borrowers’ optimal selection from these three options determines the outside option to selling— V+ (n)—whichinturndeterminesthelikelihoodaloanpaysoff(seeSectionB.1). b,ext Thekeyboundariesdeterminingwhereborrowerschoosetodefaultarefoundbysettingthevalues of Extend-Maintain and Extend-Neglect in Table 1 to 0 (the value from defaulting). Call these expressionsEquations(DN )and(DM )becausetheyimplicitlydefinethelocusofpointssuchthat b b 68

borrowersareindifferentbetweendefaultandneglectanddefaultandmaintain,respectively: (cid:18) (cid:19) (1 θ)(1+g)n (1+ν)n (r +p)+β(1 p)V − =0 (DN ) m b b − − 1 p − (cid:18) (cid:19) (1+g)n n (r +p)+β(1 p)V =0 (DM ) m b b − − 1 p − Theleft-handsideoftheseexpressionsisclearlyincreasinginn. AdditionallyV ( )islowwhenn b′ · is,sohigherpaydownsreducecashflows(netofloanpayments)morethantheyincreasecontinuationvalues. Thismeansthateachexpressionisdecreasingin pforarangeoflown. Consequently, from the implicit function theorem, these expressions define upward-sloping functions giving the principalpaydownsuchthatborrowersareindifferentbetweendefaultandeachextensiontypefor a given n.23 Denote these equations defining the indifference boundaries as DM (n) and DN (n). b b The upper envelope of these two expressions gives the p that makes a borrower indifferent to default, and the minimum of this and the maximum feasible paydown given borrowers’ liquidity constraintsdeterminesthemaximumachievablepaydownthatisshowninFigure1: P (n)=min max DM (n),DN (n) , f +n r (5) b b b m { (cid:124) { (cid:123)(cid:122) } (cid:125) (cid:124) (cid:123)(cid:122)− (cid:125) } Whatborrowersare Whatborrowers willingtopaydown canpaydown To characterize P (n), lenders are constrained by borrowers’ default condition when n is low and b optimally require the largest paydown that borrowers are willing to make. That is, borrowers are indifferent to default for low-n extensions. If V (n) 0—namely there’s almost no hope b ≈ of a borrower leaving the region in which they are indifferent to default—neglect is preferred 23Theslopesofthecurvesare: ∂p (cid:12) (cid:12) (cid:12) = 1+ν+β(1 − p)µ n V b′ (µ n n) and ∂p (cid:12) (cid:12) (cid:12) = 1+β(1 − p)µ m V b′ (µ m n) ∂n(cid:12) DNb 1+β(V b (µ n n) − µ n nV b′ (µ n n)) ∂n(cid:12) DMb 1+β(V b (µ m n) − µ m nV b′ (µ m n)) forµ n = (1 − θ 1 )( p 1+g) andµ m = ( 1 1+g p ). − − 69

to maintain, and P (n) (1+ν)n r . In other words, lenders require all cash flows from the b m ≈ − propertytogotowardsloanpayments. Sincethosecashflowsarelow,thisentailsinterestpayments thatexceedpropertycashflowsgettingcapitalizedintotheloanbalance. More generally, borrowers’ willingness to pay a loan down is increasing and convex in n. A higher n means that there are both higher cash flows available to the lender (reducing the need for forbearance) and more potential for price appreciation to pull the property into the region where a sale can be profitable for the borrower (making borrowers more willing to make principal and interest payments that exceed property cash flows). The first effect is linear in n and the second convex. Sincetheliquidityconstraintislinear,theliquidityconstraintbecomesbindingatahigher n, andborrowers’ maximum paydownrises dollar fordollar with debtyield since greaterproperty cashflowsrelievethatconstraint. The second relevant margin is whether borrowers choose to maintain the property. Setting the payoutstoextend-maintainandextend-neglectequaltoeachother,weget: (cid:18) (cid:18) (cid:19) (cid:18) (cid:19)(cid:19) (1+g)n (1 θ)(1+g)n νn+β(1 p) V V − =0 (M ) − − b 1 p − b 1 p b∗ − − When the left-hand side of the equation is positive, the borrower chooses to maintain the property because the increase in continuation values is enough to compensate for the savings from underinvestment. ThisexpressionimplicitlydefinesthemodificationboundaryM (n). b∗ B.3.2. ContractsandOutcomesChosenbyLenders Whennislow,lendersareconstrainedbyborrowers’willingnesstoacceptaprincipalpaydownand require the highest paydown borrowers will offer (the P (n) boundary defined in Section B.3.1). b The pivotal boundary for lenders’ management of stressed loans is whether they are willing to provide an extension to a borrower that will not accept a significant principal paydown and lacks theincentivestomaintaintheproperty. Lenders’ decisions here amount to whether Extend-Neglect gives a higher recovery than foreclo- 70

sure. Lendersareindifferentbetweenthetwooutcomeswhen: (cid:18) (cid:19) (1 θ)(1+g) (cid:16) (cid:17) r +p+β(1 p)V − n Λn/κ χ(1 Λn/κ) =0 (P ) m − l 1 p − − − l − and will prefer to extend loans if that quantity is positive. This function implicitly defines the minimumpaydownlendersaccept. Forhighn,borrowers’participationconstraintisnotbinding. Inthisarea,lendersrequirethepaydownthatsatisfiesthefirstorderconditionmaximizingtheirvaluefunction. Takingtheexpression for lenders’ value function in equation (1), and noting that for this region of n π (n,p) = 0, def (1+g) π (n,p)=1 π (n,p)and µ (n,p)= ,wecanexpressthefirstorderconditionas ext − sale ∗ 1 p − ∂ (cid:104) π (n,p) (cid:16) p 1+β(1 p)V (cid:0) (1+g)n/(1 p) (cid:1) (cid:17)(cid:105) =0 (P ) ∂p ext − − l − l∗ where π (n,p)=(κ/κ (n)) α ext ∗ κ (n)=n/(1+n p+β(1 p)V ((1+g)n/(1 p))) ∗ b − − − are the probability of extension and critical cap rate offer below which borrowers pay back the loan, as derived in Appendix B.1. The expression for κ (n) substitutes in the value to extend- ∗ maintain forV+ (n) since this expression pertains to the optimization problem when lenders are b,ext not confined by maintenance decisions. This expression shows that lenders optimally trade off the loss of immediate principal repayment from an extension, (1-p), and the value of future loan paymentsfromtheextension: β(1 p)V (cid:0) (1+g)n/(1 p) (cid:1) . l − − 71

B.4. Calibration I set r =4.5% to match 30-year Treasury yields in the period of stress, and r =7% to match the m 2.5% loan rate spread for CRE loans in Table 1 of Glancy et al. (2022). I set α =12.3 to match the 5% discount required for immediate sales during expansions from Figure 8 in Sagi (2021).24 I set f to 0.15 to match the 95th percentile of p+r n for post-2022 extensions in the data. I set m − g=0.01andσ =0.1tomatchthestatisticsonannualrentgrowthinTable2ofAnetal.(2016),and Λ=.76tomatchthe24%deadweightforeclosurecostsinBrownetal.(2006).25 Isetthedeclinein propertyvaluesfromdeferredmaintenancetoθ =0.045basedontheadditionalannualizedcapital expendituresrequiredoflendersfollowingforeclosuretocompensateforpreviousunderinvestment by financially distressed owners from Table 7 of Brown et al. (2006).26 It is unclear how much of this value decline is lost from inefficiency vs. transferred to borrowers. I assume a 50/50 split, meaningthat.5θn/κ =νn,makingν =0.69. Inthebaselinecalibration,Iset χ =0basedonthe finding from Favara et al. (2024) that large U.S. banks do not engage in zombie lending for C&I loansregardlessofcapitalization. 24Thereisa5%discountrelativetoasellerwithatwo-yearhorizon. Iapproximatethisastheexpecteddiscount fromaforcedsalerelativetoaninvestorwiththreetimesasmanysaleopportunities(i.e.,inyears0,1,and2). The minimumfrom3drawsofaParetodistributionisParetodistributedwithashapeparameter3α. Sincetheexpected saleisproportionaltoα/(1+α),Isetα tosatisfyα/(1+α)=0.95 3α/(1+3α). × 25I use the measure of foreclosure costs that doesn’t account for lenders’ required capital expenditures due to deferredmaintenancesincesuchcostsarecapturedinthemodel. 26Lenders of foreclosed properties had capital expenditure rates of 6.4%, compared to 1.5-2% for nondistressed owners. 72