Determinants of Recent CRE Distress: Implications for the Banking Sector

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Determinants of Recent CRE Distress: Implications for the Banking Sector David Glancy, Robert Kurtzman 2024-072 Please cite this paper as: Glancy, David, and Robert Kurtzman (2024). “Determinants of Recent CRE Distress: Implications for the Banking Sector,” Finance and Economics Discussion Series 2024-072. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2024.072. 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.

Determinants of Recent CRE Distress: Implications for the Banking Sector * David Glancy†1and Robert Kurtzman‡1 1Federal Reserve Board August 27, 2024 Abstract RisinginterestratesandstructuralshiftsinthedemandforspacehavestrainedCREmarkets and prompted concern about contagion to the largest CRE debt holder: banks. We use confidential loan-level data on bank CRE portfolios to examine banks’ exposure to at-risk CRE loans. We investigate (1) what loan characteristics are associated with delinquency and (2)towhatextenttheportfoliocompositionofmajorCRElendersdeterminestheirexposureto losses. HigherLTVs,largerpropertysizes,andgreaterlocalremoteworktendenciesareallassociatedwithincreaseddelinquencyrisk,particularlyforofficeloans. Weuseseveralmachine learningalgorithmstodemonstratethatvariationinexposuretotheseriskfactorscanaccount for most of the performance disparity across different types of CRE lenders. The headline result is that small banks’ comparatively modest delinquency rates mostly reflect observable portfoliocharacteristics—predominantlytheirlowholdingsoflarge-sizedofficeloans—rather thanunobservedfactorslikeextensionormodificationtendencies. Keywords: commercialrealestate,banks,CMBS JELClassification: G21,G23,R33 *We thank Joe Nichols and seminar participants at the Federal Reserve Board R&S workshop for helpful comments. Theviewsexpressedinthispaperaresolelythoseoftheauthorsanddonotnecessarilyreflecttheopinionsof theFederalReserveBoardoranyoneintheFederalReserveSystem. †PrincipalEconomist,DivisionofMonetaryAffairs,FederalReserveBoard,david.p.glancy@frb.gov. ‡PrincipalEconomist,DivisionofResearch&Statistics,FederalReserveBoard,robert.j.kurtzman@frb.gov.

1. INTRODUCTION Higher interest rates and shifts in where people work and shop have created significant stress in pockets of the commercial real estate (CRE) market (Board of Governors of the Federal Reserve System, 2023). As CRE is the largest loan category on banks’ books, these developments have caused concern about CRE loan losses exacerbating other recent banking sector strains (Acharya etal.,2023).1 Analyzingsucheffectsiscomplicatedbyalackofdetaileddataonbanks’CREloan holdings. Due to this data limitation, researchers have mostly relied on aggregate bank portfolio data(Faria-eCastroandJordan-Wood,2023,2024),ordatafromCREsegmentswithmorepublic reporting(Guptaetal.,2022;Jiangetal.,2023;GlancyandWang,2023)toassessrisksposedtothe bankingsector. However,banksserveaselectedsegmentoftheCREmarket(Glancyetal.,2022a), and recent CRE stresses have been highly uneven (Marsh and Pandolfo, 2024), so extrapolating acrossdifferentpartsoftheCREmarketcanbedifficult. Consistent with this segmentation, Figure 1 shows that loan performance differs markedly across different types of CRE lenders. Though nonperforming loan (NPL) rates for bank and commercial mortgage-back securities (CMBS) loans were comparable during the Global Financial Crisis (GFC)of2007–09,periodsofstrainsincethenhavebeenmoreconfinedtoCMBSmarkets;CMBS delinquencies (red) rose moderately in 2016 and spiked at the onset of the COVID-19 pandemic, whiledelinquenciesatlarge(blue)andsmall(green)banksremainedunder2percentduringthese episodes.2 In2023,delinquencyratesforCMBSandlargebanksrosebysimilaramounts,buttheperformance ofCREloansatsmallerbanksremainedstrong. CREstrainsappeartobehighlyconcentrated: the rise in bank CRE NPLs is driven by nonowner-occupied, nonresidential properties at large banks 1AsofMarch6,2024,totaloutstandingCREloanswereestimatedtobeabout$3.0trillion,comparedto$2.8,$2.6 and$1.9trillionforcommercialandindustrial,residentialrealestate,andconsumerloans,respectively,accordingto datafromtheH.8releaseoftheBoardoftheGovernorsoftheFederalReserveSystem. 2Even during the GFC, NPL rates for loans secured by existing properties rose less at banks than for CMBS. However,bankNPLrateswerepulledupbyanear20percentdelinquencyrateforconstructionandlanddevelopment loans (a segment not served by CMBS). See Appendix Figure B.1 for a decomposition of bank NPL rates by CRE subcategory. 2

Figure1: NonperformingLoanRatesoverTime SlX MRNU3a8RaLCN<hHR Nh` j3 +K#b SzX # NGchQq3ah1Szz$N # NGchmN03ah1Szz$N 4X fX :X lX zX lzz4 lzSS lzS: lzSe lzlz lzlk Notes: ThefigureplotsCREnonperformingloan(NPL)ratesovertimefornon-agencyCMBSloans(red) andforCREloansheldbybankswithmore(blue)andless(green)than$100billioninassets. NPLratesare loansthatare30daysormorepastdueornonaccrual,plottedasashareofaggregateoutstandingbalances. Calculationsexcludedefeasedandrealestateownedloans. Sources: CallReports,Morningstar,andauthors’calculations. (seeFigureB.2),andwithinthissegment,theincreaseisalmostwhollyattributabletoofficeloans (see Figure B.3). This heterogeneity highlights that CRE cannot be viewed as a single strained asset class. Understanding the implications of the recent stress requires a detailed understanding of where loan performance is deteriorating and which lenders are exposed to those troubled segments. In this paper, we compile data from a variety of sources to analyze why CRE loan performance differs across lenders. The analysis proceeds in two steps. First, we use loan-level panel data on CRE loan performance to investigate what factors are associated with higher delinquency rates and why loan performance differs between banks and CMBS. To accomplish this, we combine, harmonize, and analyze data from CMBS filings and confidential data from large banks’ stress tests. We find that office loans held in CMBS pools are nearly 4 percentage points more likely to go delinquent than those held by banks. This effect is driven nearly entirely by bank loans being secured by smaller properties, having lower loan-to-value (LTV) ratios, and being more likely to have recourse. For other property types, CMBS also underperform bank loans, but the difference 3

Figure2: CREExposurebyBankSize SzzX +`2hjRh cc3jchVXW +`2h/3ICN\n3N,w kzX HR, IhK3 N HR, IhK3 N h 4zX lzX h fzX SzX :zX 9X lzX zX zX 1SzzhLN 1Sh$N 1Szh$N 1Szzh$N 1Shja 1SzzhLN 1Sh$N 1Szh$N 1Szzh$N 1Shja # NGh cc3jchV0RII ac.hIR<hc, I3W # NGh cc3jchVIR<hc, I3W (a)CREConcentrations (b)CREDelinquency Notes: The figure plots CRE as a share of assets (left) and CRE delinquency (right) by bank size. CRE includes nonfarm nonresidential, multifamily, and construction and land development loans. Blue dots report CRE shares or delinquency rates at individual banks, while the black line plots an estimate of these variablesforbanksofagivensize(thekernel-weightedlocalmean). Thescaleforthey-axischangesinthe rightpanelafter10%toimprovethevisibilityofdifferencesinperformanceofbankswithintypicalranges. Sources: CallReportsandauthors’calculations. issmallerandnotasclearlyattributabletootherobservablecharacteristics.3 In the second part of the paper, we investigate potential drivers of the strong performance at small banks. Studying the performance of CRE loans at large banks (i.e., banks with over $100 billion in assets) is useful as (1) that is where performance has deteriorated more and (2) data availability enablesmore-detailedanalysis. However,itissmallerbanksthatarehighlyexposedtoCREloans. Indeed, Figure 2 shows that banks with between $1 billion and $10 billion in assets tend to have the highest concentrations in CRE (left), but also had quite low CRE delinquency rates as of end of 2023 (right). These results indicate that the deterioration in loan performance experienced to dateisunlikelytocausesignificantproblemsforthebankingsector. However,iftroublesforCRE loans at CMBS and large banks portend similar problems at smaller banks, more significant stress may be looming. To understand such risks, we need to evaluate why CRE loan performance at smallbankshasremainedcomparativelyresilient. While detailed loan-level panel data are generally not available for small banks, some broad char- 3Weprovidesomesuggestiveevidencethatagreatertendencytomodifyloansmaycontributetotheseunobserved differences. 4

acteristicspertainingto loansizes,property types,andlocations areavailabledueto thepublicreporting of mortgage liens. We can therefore estimate models of loan performance using data from large banks and investigate the degree to which these observable factors can explain the strong performance at small banks. We do so using a variety of models in order to provide predictions frombotheasilyinterpretedmodelswhereitisclearwhatvariablesdriveperformancedifferences (OLSandlow-complexitytrees)andmoreflexibleonesthatmaybetterreflectcomplexpatternsin thedata(K-nearestneighborsandrandomforests). Regardlessofthemodelchosen,wefindthatdifferencesinperformanceofCREloansatlargeand smallbankscanmostlybeattributedtodifferencesinloanssizes,andinthetypesandlocationsof the properties securing the loans. For the models that allow us to meaningfully decompose differences in loan performance between large and small banks, we find that by far the biggest difference is that small banks are less exposed to large-sized office loans, which performed particularly poorly in 2023. This factor alone accounts for about half of the 2-percentage-point-difference in NPLrates. Overall, these findings demonstrate that the resiliency of small banks’ CRE portfolios can be rationalized with property-level observables. If NPLs at small banks were low because their high concentrationsinCREloanspromptedthemtodelaylossrecognitionby“evergreening,”suchdifferences would not be picked up by our model. Instead, it appears that small banks’ NPLs are mostly lower because their loans are safer on observed dimensions. Therefore, the primary risk to smallbankscomeslessfromtheCREstrainsexperiencedthrough2023,butmorefromtheriskof stress spreading to other parts of the CRE market—for example due to an economic downturn or higher-for-longerinterestratesfurtherpressuringvaluations. The rest of the paper proceeds as follows. Section 2 reviews an array of factors identified in the literature that could contribute to differences in loan performance across CRE lenders. Section 3 identifies what factors are associated with delinquency at large banks and CMBS. Section 4 estimatesamodelofloanperformanceusingonlyloancharacteristicsthatareobservableforsmall 5

banks’ CRE loans, and examines the extent to which those factors can account for small banks’ lowerdelinquencyrates. Section5concludes. 2. LITERATUREONDIFFERENCESACROSSCRELENDERS Toguideouranalysisofthefactorspotentiallycontributingtothedifferencesinloanperformance across lenders, we first review the literature pertinent to this question. We split this discussion into two parts, separately discussing the factors that could mitigate or amplify loss for bank CRE loans. 2.1. FactorsThatCanSupportBankLoanPerformance Modification Ability The first factor supporting bank loan performance is that banks are more able to modify distressed loans to preempt foreclosure (Black et al., 2017, 2020). Indeed, Glancy et al. (2022b) present evidence that forbearance policies supported bank CRE performance at the onsetofthepandemic. BankregulatorsreleasedapolicystatementonCREloanworkoutsin2023 reaffirming that “prudent CRE loan accommodations and workouts are often in the best interest of the financial institution and the borrower”, thus demonstrating that support for bank workouts is more than a COVID-era phenomenon.4 In contrast, CMBS face limitations to modifying loans due to IRS policies, stipulations in pooling and servicing agreements, and conflicts of interest across disparate investors (Wong, 2018; Flynn Jr. et al., 2023). While the nature of workouts will presumably be different in the current environment than at the onset of the pandemic (e.g., extensionsratherthanforbearance),banks’flexibilityinthisregardmayenablethemtopreventor morequicklyresolveloanstress. 4The guidance is available on the Federal Reserve’s website at https://www.federalreserve.gov/supervisionreg/ srletters/SR2305a1.pdf. 6

RecourseandLowerLeverage AsecondfactorsupportingtheperformanceofbankCREloans isthatborrowerstypicallyhavemoreequityatstake. First,bankloanstypicallyhavelowerLTVs, which gives more room for property values to fall before the lender starts taking losses (Glancy etal.,2022a). Second,bankCREloansarepredominantlyrecourseloans,meaningthatborrowers have assets besides the subject property at stake if they default. Glancy et al. (2023) show that mostCREloansfromlargeU.S.bankshaverecourse. Theauthorsdemonstratethattheserecourse loans receive more favorable underwriting terms and were less likely to require forbearance at the onset of the pandemic, both consistent with recourse supporting loan performance. Unlike bank loans, CMBS loans are generally non-recourse (barring springing guarantees for particular “bad acts”). Assuch,CMBSborrowersmaybemorewillingtowalkawayfromapropertyevenifthey havetheabilitytoservicethedebtonit.5 DifferencesinLoanSize/Location Athirdfactorpotentiallycontributingtodifferencesinloan performanceisthatbankstendtolendagainstsmallerpropertiesthanCMBS(GhentandValkanov, 2016; Glancy et al., 2022a).6 While loans against larger properties are not considered inherently riskier, they have underperformed in the current environment. Part of this effect likely reflects location;highpricedpropertiesaredisproportionatelylocatedincentralbusinessdistricts(CBDs). Amid the shift to working from home, CBDs have experienced declines in commuting activity (Monte et al., 2023) and commercial rents (Rosenthal et al., 2022), and a greater deterioration in property occupancy and income following lease expirations (Glancy and Wang, 2023). Relative to CMBS, office loans at banks, and especially small banks, are more likely to be in suburban markets,whichhavebeenlessaffectedbytheshifttoremotework.7 Evenwithinagivenlocation, largerofficeslikelyhaveatenantmixplacingthematagreaterriskofdepartures(e.g.,morespace 5Recourse might be particularly valuable to lenders in the current environment given that stress so far has been largely confined to the office sector. If sponsors or guarantors have other assets of value—for example, houses, equities, or less-troubled CRE properties—owners may be more willing and able to maintain payments on loans securedbytroubledproperties. 6By virtue of their diversified customer base, CMBS are able to fund larger loans that would generate too much concentrationriskforabalancesheetlender. 7Glancy and Wang (2023) show that small and regional banks have a lower exposure to office loans in CBDs or marketswithagreaterincreaseinremotework. 7

accountedforbymulti-locationtechcompanies). Better Screening The last factor we highlight supporting the performance of bank loans is that bankstypicallyretainthecreditriskfortheloanstheyorigination,andthuscanhavebetterscreening incentives. Work on the CMBS market before the GFC provides evidence that adverse selection(Anetal.,2011)andmoralhazard(Ashcraftetal.,2019)reducethequalityofloansinCMBS pools. In response, as part of the Dodd-Frank Act, regulators implemented risk retention requirements to reduce such agency problems (Flynn Jr et al., 2020). Nonetheless, Griffin and Priest (2023)presentevidencethatagencyconflictsbetweenoriginatorsandCMBSinvestorspersistand maycreateunderwritingweaknessesthatarerevealedduringtimesofstress. 2.2. FactorsThatCanHinderBankLoanPerformance Specializationin(Historically)RiskierLoans Whilegreaterrenegotiationflexibilitymaysupporttheperformanceofanygivenbankloan,thiseffectmaygivebanksanadvantageinfinancing riskier properties where such flexibility is more valuable (Black et al., 2017, 2020). One clear dimensionalongwhichriskdiffersisthatCMBSlendagainstincome-producingproperties,whereas banks originate more bridge and construction loans. The performance of bridge and construction loans can be more dependent on the particular business model of the borrower and thus subject banks to the risk that properties are more difficult to lease-up than expected or fail to earn returns sufficient to recoup the alteration costs. Consistent with these loans being riskier, construction and land development loans are subject to higher FDIC fee assessments, receive higher capital requirements when LTV limits are not met (Glancy and Kurtzman, 2022), and disproportionately contributedtobankfailuresduringtheGFC(Friendetal.,2013).8 8For more details on these loans receiving higher FDIC fee assessments, see FDIC Law, Regulations, Related Acts, Title 12 Chapter III Subchapter B Appendix C to Subpart A to Part 327 available at https://www.ecfr.gov/ current/title-12/chapter-III/subchapter-B/part-327. 8

More Financially Constrained Borrowers A final difference between bank and CMBS loans isthatbanksmaycatertomorefinanciallyconstrainedborrowers. RoughlyhalfofCMBSlending goes to public or institutional buyers, compared to only 10 percent of bank lending (Glancy et al., 2022b). Theselargersponsorswithmorediversifiedfundingsourcescanpotentiallymaintainloan payments in the face of a disruption to property cash flows (assuming they find it optimal to do so). SimilarselectioneffectsoccuroutsidetheCREmarket,wherebanksdisproportionatelyserve smaller,younger,andriskierfirms,whilemoreestablishedfirmsutilizemarketfinancing(Petersen andRajan,1994;BoltonandFreixas,2000). 3. CRELOANPERFORMANCEATLARGEBANKSANDCMBS This section uses loan-level bank and CMBS data to investigate the loan and property characteristicsaffectingCREloanperformancein2023. 3.1. Methodology As the last section discusses, there are a host of factors that may contribute to differences in the performance CRE loans across lenders. We now empirically examine how much these factors matter in the current environment. The data on bank loans come from FR Y-14Q filings (the data underlying bank stress tests), which provide loan-quarter information on loans with committed balances over $1 million from banks with over $100 billion in assets. The CMBS data come from Morningstar, which compiles loan-month data from CMBS disclosures.9 We classify lenders by who holds the loan, rather than who originates it, so bank-originated loans in CMBS pools are considered CMBS loans. More information on how we clean and harmonize these data are in SectionA.1. Though banks below the Y-14 reporting threshold tend to have higher concentrations in CRE, an advantageoftheY-14sampleisthatitcoversthegroupofbanksforwhichCREloanperformance 9WeexcludeagencydealsfromtheCMBSdata,aswellasdefeasedorrealestateownedloans. Forbankloans,we onlyincludefirst-lien loans againstalready-constructedpropertiestobetteralign thesamplewiththatof theCMBS market. 9

has materially deteriorated. This data is thus useful for examining the factors causing CRE loans togodelinquent. WhilewecannotanalyzetheperformanceofCREloansatsmallerbanksinsuch detail,wecanassesstheperformanceofY-14loanswithcharacteristicsresemblingthoseofsmall banks (i.e., smaller loans in smaller markets) to explore the implications for smaller banks. We conductthisexerciseinSection4. Toinvestigatehowobservablecharacteristicsrelatetoloanperformance,weestimatelinearregressionsoftheform: 100×Delinquent =β CMBS +β Maturing +β Office +γ ′X +ε, i,23 1 i 2 i,23 3 i i,23 i forthesampleofbankorCMBSloansthatwereoutstandingasoftheendof2022. Delinquent is i,23 inindicatorforwhetherloaniisdelinquentasofthelastobservationin2023,definedhereasbeing past due, performing beyond its maturity date, or liquidated.10 The main independent variables of interest are whether loan i is in a CMBS pool, whether the loan was scheduled to mature in 2023, and whether the loan is secured by an office property. The coefficient on CMBS reflects thedifferenceindelinquencyratesforCMBSloanscomparedtothosefromlargedomesticbanks, controlling for the two characteristics most strongly related to loan performance: whether they matured (capturing difficulty refinancing in an environment of falling valuations and tight lending standards) and whether the loan is secured by an office property (the property type accounting for mostoftheriseindelinquenciesover2023). X is a set of controls that includes other property type dummies.11 In some specifications, we i,23 layer in additional controls to assess whether they can account for some of the difference in delinquency between large banks and CMBS. These additional controls include LTV, property size, a recourseindicator,and geographic characteristics(whetherthepropertyisin aCBDorcitywhere 10Thequarterofobservationsiseither2023:Q4iftheloanisactiveasoftheendoftheyearorthequartertheloan paid-offorwasliquidatedotherwise. 11These variables are included in every specification but coefficients not displayed. Multifamily is the omitted category. Hotel and retail are the next-most-likely loans to go delinquent after office, while industrial delinquency ratesarenotsignificantlydifferentfromthoseformultifamilyloans. 10

more jobs can be done remotely). Observing how β changes as extra controls are added provides 1 information on the extent to which these characteristics can account for differences in loan performance. In the most expansive specifications, we also include controls for the occupancy of the property and debt yield of the loan.12 These controls thus assess the extent to which differences aredrivenbyunobservedriskstopropertyperformancethatmaterializeasadeteriorationinfinancial performance. Finally, in robustness exercises we conduct similar analysis predicting whether loans receive extensions to assess whether banks are providing more accommodation to stressed borrowers. SummarystatisticsofthemainvariablesofinterestareshowninAppendixTableB.1,withdataon large bank and CMBS loans shown in columns (1) and (2), respectively. As previously discussed, banks on average make smaller, lower-LTV loans, that often have recourse. The CMBS sample has an average delinquency rate about 3 percentage points higher than for large banks. Much of this difference is due to CMBS liquidating more loans in 2023 and having more loans that are performing after their maturity date; the difference is only one percentage point when using a narrower definition that only counts past-due or nonaccrual loans as delinquent (which is more aligned with the NPL measures from the Call Report). Columns (3) and (4) present the same data, but weighted by loan size, thus making the averages more reflective of aggregate portfolio shares. Broad patterns pertaining to differences between bank and CMBS portfolios are similar, but delinquency rates are higher, reflecting the fact that performance has deteriorated more for largerloans. 12Debt yield is the ratio of net operating income to the outstanding loan balance and thus reflects the ability of a property’scashflowstopayofftheloan. Sinceincomeandoccupancywouldnotgetupdatedintheeventthataloan paysoffin2023,wemeasurethesefinancialvariablesasofayearprior. 11

3.2. Results ThemainestimatesarereportedinTable1. Column(1)presentsestimatesfromthemostparsimonious specification, which only includes the CMBS, maturing loan, and property type indicators. The results show that CMBS loans are about 1.7 percentage points more likely to become delinquent, office loans are about 3.3 percentage points more likely to become delinquent (relative to multifamily loans), and loans that mature are about 12.2 percentage points more likely to become delinquent. Thislasteffecthighlightsthesignificantdifficultyborrowersfaceinrefinancingtopay offballoonloans;borrowerswhoareabletoremaincurrentoverthelifeoftheloanarefrequently failingtopaytheloanoffasitcomesdue. To investigate why CMBS loans have higher delinquency rates, we incorporate different factors discussed in the previous section as potentially contributing to differences in loan performance across lenders. Column (2) adds in the at-origination LTV of the loan, the logarithm of the atorigination property value, and an indicator for whether the loan has recourse. The size and LTV controls account for the fact that CMBS provide higher leverage, on average, and tend to lend against larger properties. Since CMBS loans are essentially entirely non-recourse, the effect of recourseisidentifiedoffofdifferencesintheperformanceofbank-heldrecourseandnon-recourse loans. Therefore, the coefficient on CMBS is now estimated off of the difference in the performancebetweenCMBSloansandnon-recoursebankloans. The findings show that the characteristics associated with bank loans—lower leverage, smaller properties, and recourse—are also associated with stronger loan performance. A one standard deviation higher LTV (0.16) or property size (1.16) are associated with delinquency rates that are roughly 80 and 100 basis points higher, respectively.13 The effect of recourse is smaller, with recourse loans having a delinquency rate that is about 40 basis points lower than similar nonrecourseloans. AddingthesethreeadditionalcontrolsreducesthecoefficientonCMBSfrom1.65 13Variation in property values reflect differences in both size (i.e., square footage) and valuation (i.e., price per square foot). Most of the effects on delinquency are attributable to differences in size, so we refer to the variable ln(ValueatOrig.) as“size”forthesakeofbrevity. 12

to 0.45, indicating that CMBS’ tendency to make larger, higher LTV, non-recourse loans accounts forabouttwo-thirdsoftheirinferiorperformance. Column(3)addsingeographiccharacteristicsofthepropertysecuringtheloan: whetherthepropertyisinaCBDandtheshareofjobsinthemetropolitanstatisticalarea(MSA)thatareidentified as being able to be done at home by Dingel and Neiman (2020).14 The results indicate that delinquency rates are about 2 percentage points higher in CBDs, but only modestly higher for cities moreexposedtoapotentialshifttoremotework;aonestandarddeviationincreaseintheteleworkable share (0.046) increases delinquency by about 25 basis points. Adding the extra geographic controlscausestheestimatedcoefficientonCMBStoincreaseabit;thoughbanksgenerallydoless lendinginadverselyaffectedmarkets,thisismostlyduetosizedifferenceswhichwereaccounted for by the specification in column (2). Adding these variables also only reduces the coefficient estimateonln(ValueatOrig.) veryslightly,indicatingthattheeffectsofpropertysizearenotdriven bythesegeographicriskfactors. Infact,mostestimatesarelittlechangedwiththeinclusionofZIP code fixed effects (not shown), so the risks associated with larger loans appears to predominantly reflectnon-geographicfactors. Column (4) adds in two variables pertaining to the property’s financial situation: the occupancy and an indicator for whether the property’s debt yield (net income as a share of the loan balance) is under 8 percent. These controls thus account for whether differences in performance are driven by variation in occupancy or cash flow risks across lenders. Low occupancy or debt yields are highly predictive of delinquency: a 10 percentage point drop in occupancy raises the probability ofdelinquencyby1.5percentagepoints,andalowdebtyieldraisestheprobabilityofdelinquency byabout1.9percentagepoints. Whiletheoverperformanceofbankloanscanbepartiallyexplainedbyobservablecharacteristics, there remains a large unobserved component. CMBS loans are about 0.9 percentage points more likely to go delinquent than bank loans with similar underwriting terms and even similar financial 14WeidentifyapropertyasbeinginaCBDifitsZIPcodebelongstoasubmarketthatCBREdefinesasbeingina CBD. 13

Table1: LoanPerformancebyLenderType 100×Delinquent i,23 FullSample Offices (1) (2) (3) (4) (5) (6) (7) (8) CMBS 1.65** 0.45 0.69* 0.88** 3.36** -2.17+ -1.69 -1.01 (0.26) (0.36) (0.33) (0.33) (0.67) (1.19) (1.18) (1.16) Maturing 12.23** 11.82** 11.77** 11.36** 19.56** 18.45** 18.16** 16.83** (0.96) (0.95) (0.95) (0.95) (1.88) (1.86) (1.85) (1.80) Office 3.37** 2.59** 2.48** 2.19** (0.33) (0.30) (0.30) (0.30) LTVatOrig. 4.87** 5.53** 3.92** 13.23** 15.22** 12.42** (0.72) (0.83) (0.81) (1.88) (1.98) (1.89) ln(ValueatOrig.) 0.83** 0.73** 0.50** 1.97** 1.44** 1.04** (0.11) (0.10) (0.10) (0.30) (0.31) (0.29) Recourse -0.40 -0.26 -0.30 -3.61** -3.24** -2.99** (0.25) (0.23) (0.23) (1.06) (1.05) (1.02) CBD 2.31** 1.87** 5.15** 3.92** (0.46) (0.39) (1.06) (1.04) TeleworkableShare 7.37** 5.24** 14.11** 11.77* (1.84) (1.82) (5.16) (5.00) Occupancy -14.72** -20.32** (1.61) (2.53) DebtYield<.08 1.88** 5.79** (0.45) (0.99) R2 0.056 0.061 0.064 0.081 0.072 0.093 0.100 0.134 a Observations 57,799 57,799 57,799 56,873 7,652 7,652 7,652 7,505 OtherPropertyFixedEffects? ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation: 100×Delinquent =β CMBS +β Maturing +β Office +γ′X +ε, i,23 1 i 2 i,23 3 i i,23 i whereDelinquent isanindicatorforwhetherloaniisdelinquentasofthelastobservationin2023(2023:Q4ifthe i,23 loan is active as of the end of the year, or the quarter the loan was paid-off or liquidated otherwise). Loans that are liquidatedorperformingbeyondtheirmaturitydatecountasdelinquent. Themainindependentvariablesofinterest are whether loan i is in a CMBS pool, whether the loan was scheduled to mature in 2023, and whether the loan is secured by an office property. Fixed effects for other property types are included but not displayed (multifamily is the omitted category). Column (2) adds controls for whether the loan has recourse, the at-origination LTV, and the logarithmofthepropertyvalueatorigination. Column(3)addscontrolsforwhetherthepropertyisinaCBDandthe shareofthecity’semploymentthatcanbedoneathome(DingelandNeiman,2020). Column(4)addscontrolsforthe occupancyandanindicatorforwhetherthedebtyieldisunder8%(bothasofayearpreviously). Columns(5)to(8) repeatthesameanalysisbutrestrictthesampletoofficeproperties. Standarderrors,inparentheses,areclusteredby bank-originationyearforbankloansandCMBSdealforCMBSloans. +,∗,∗∗ indicatesignificanceat10%,5%,and 1%,respectively. Sources: Y-14QH.2Schedule,Morningstar,andauthors’calculations. performance. One explanation for this result is that banks are more able to renegotiate stressed loans. Some suggestive evidence of this can be found in Appendix Table B.2, which shows that bank loans were about 5 percentage points more likely to receive extensions compared to CMBS 14

loans. Without a counterfactual saying what would have happened to these loans if they did not receive an extension, it is difficult to say precisely how banks’ willingness to extend loans affected performance differences. However, if some of these extended loans would have otherwise gonedelinquent,suchaccommodationcouldhavecontributedtothecomparativestrengthofbank loans. Columns (5) to (8) repeat the same analysis but restrict the sample to office properties. Column (5), which only controls for loan maturity, shows that CMBS’ underperformance is even more pronouncedforofficeloans,withCMBSloanshavingadelinquencyratethatisabout3.4percentage points higher than for bank loans. Additionally, defaults at maturity are even more prevalent for office loans, with a 2023 maturity raising the probability of delinquency nearly 20 percentage points. Column (6) shows that the underperformance of CMBS office loans can be entirely attributed to differences in loan sizes, LTVs, and the use of recourse. Office loans have delinquency rates that are 3.6 percentage points lower when they have recourse, 1.3 percentage points higher when the LTV is 10 percentage points higher, and about 2.3 percentage points higher for properties with an at-originationpropertyvaluethatisonestandarddeviationhigher. Oncethesefactorsarecontrolled for, CMBS loans are found to have delinquency rates that are slightly lower than bank loans. Put differently, CMBS office loans perform somewhat worse than nonrecourse bank loans of similar size and LTV. However, CMBS have higher delinquency rates overall due to differences in these threecharacteristics. Column (7) adds in geographic controls, which have a much larger effect on office loans than for CRE properties in general: the effect of being in a CBD and telework exposure are about doubledforofficeloansrelativetotheoverallCREloanpool. Thesecontrolsreducethecoefficient on ln(ValueatOrig.) somewhat, but most of the effect of size is not driven by these geographic characteristics. Finally, column (8) adds in the controls for occupancy and debt yield. Weaker financial perfor- 15

Figure3: CREDelinquencyRatesbyLoanSize 12 10 8 6 4 2 0 etaR ycneuqnileD Banks CMBS 12 10 8 6 4 2 0 2.5 5 10 20 40 80 160 320 Property Value, millions (a)FullSample etaR ycneuqnileD Banks CMBS 2.5 5 10 20 40 80 160 320 Property Value, millions (b)OfficeLoans Notes: The figure plots binscatter estimates of 2023 CRE loan delinquency rates by the logarithm of the property value at origination. The red and blue dots are the delinquency rates for CMBS and Y-14 bank loans at different deciles of their respective property size distribution, respectively. The left panel presents delinquency rates pooling across property types, while the right panel restricts the sample to office properties. Sources: Y-14QH.2Schedule,Morningstar,andauthors’calculations. mancearemorepredictiveofdelinquencywithintheofficesector. Includingthesecontrolsreduces thepredictedeffectsofsizeandthegeographicriskfactors,indicatingthatthehigherdelinquency rates for loans with those characteristics are partly attributable to worse declines in occupancy or income. The fact that those variables remain significant even in the presence of the additional financialcontrolslikelyreflectsthefactthatpropertyperformanceisslowmoving. Ownersoflarger propertiesinmoreadverselyaffectedmarketsmaygodelinquentinanticipationoffutureproperty strains,especiallyifthoseanticipatedstrainsleavethemunabletorefinanceuponmaturity. To recap the findings, many of the drivers of CRE delinquencies are as would be predicted given thenatureofthestress. Thelargestdriverisloanmaturity,reflectingproblemsrefinancingballoon payments amid tighter credit conditions, lower valuations, and higher interest rates. Additionally, delinquency is higher for properties most exposed to the decline in demand for space—namely, officepropertiesandpropertiesinCBDsormarketsmoreexposedtoremotework. Finally,theanalysispointstoafewproximatecausesofdelinquencythatarepotentiallyimportant forunderstandingthecreditoutlookforbanks. First,largebanks’CREloansmodestlyoutperform 16

CMBSloanswithsimilarterms,financialperformance,andgeographiccharacteristics. Thisresult suggests that unobserved characteristics of banks loans—such as a greater ability to modify loan terms,orborrowers’concernsaboutdamagingexistingbankrelationships—aresupportingoverall loanperformance. Second, loans securing large properties have higher delinquency rates, suggesting that they have unobserved risk characteristics—for example, a worse income outlook, more difficulty making up operating shortfalls, or sponsors more willing to strategically default—that are weighing on loan performance. The magnitude of this effect can be most clearly seen in Figure 3, which presents binscatter estimates of how loan delinquency varies by property size and lender type. The delinquency rate for bank CRE loans is modest for properties with an at-origination value under $20 million, even for office loans. Delinquency is instead concentrated in large CMBS-funded properties, or large bank-funded offices. The next section explores the extent to which such differences acrosspropertysizescanexplainthestrongerloanperformanceatsmallerbanks. 4. IMPLICATIONSFORSMALLBANKS ThissectionexplorestheimplicationsofthesefindingsfortheperformanceofCREloansatsmall banks. To the extent that small banks lend against even smaller and less urban properties, the patterns documented in Section 3 may contribute to their comparatively strong loan performance. Totestthis, wefirstcompiledatafrom varioussourcesonthecomposition ofCREloanportfolios across different types of lenders. We then estimate a set of models of bank loan performance using only the characteristics that are observable forsmall banks’ CRE loan portfolios. Lastly, we examine the fitted delinquency rates to assess the extent to which those factors can account for smallbanks’lowerdelinquencyrates. 17

4.1. CompositionofCREPortfolios The first step in understanding risk factors for small banks’ CRE portfolios is to compile data on thecompositionoftheirCREloanholdings. TheprimarydatasourceweuseisMSCIRealCapital Analytics (RCA). RCA sources data from both public records and industry contracts to provide detailedinformationonCREtransactions. The main drawback to the RCA data is that it only covers properties above $2.5 million dollars in value. Though this sample covers the majority of non-owner-occupied CRE lending, the omission of smaller transactions is potentially problematic given the strong association between property size and loan performance. To mitigate this potential bias, we supplement the RCA data with data onopencommercialmortgageliensinthepublicrecordsprovidedbyCoreLogic. Specifically,we use RCA data for loans with original balances over $2.5 million (as such transactions should be reliablyincludedinRCA)andCoreLogicforloansbelowthisthreshold. To maintain consistency with the previous analysis, we continue to focus on non-owner-occupied CREloanssecuredbyexistingproperties(i.e.,weexcludeconstructionloansandowner-occupied CRE loans). This sample covers the CRE loans that have exhibited the most stress through 2023 (see Appendix Figure B.2). More information on how we construct the data, including details lender name matching, sample selection, and how we impute whether RCA loans are still outstandingareinSectionA.2. After harmonizing, name matching, and combining these data sources, we have a cross-section of at-origination characteristics for what should be the near universe of outstanding commercial mortgages and identifiers for the type of lender that originated the loan. For each of these loans i from one of the lender types j (small bank, large bank, CMBS, and unclassified), we have a set of observable loan or property characteristics X that includes the loan size, origination date, i,j property type, and property location. While this set of variables does not include all of the factors studied in Section 3, it includes most of the major variables affecting performance that are likely 18

to differ notably across lenders. The most significant change to the specification is adding loan size, ln(BalanceatOrig.) in place of the measures of LTV and property value. Since we do not reliably have information on property values for refinances in CoreLogic data (appraised values are available in Y-14Q and RCA data but not in public records), we replace the measures of LTV and property value with this single variable which reflects both leverage and property size and is morereliablyavailableinthedata. The weighted average portfolio characteristics for large and small banks based on this data are shown in the last two columns of Table B.1. The averages for large banks are broadly in line with theweightedaveragesoftheY-14datashownincolumn(3),suggestingthatthedataweconstruct fromRCA/CoreLogicsuccessfullymatchesbanks’trueportfoliocomposition. Thestatisticsshows that loans at large banks tend to be larger in size, slightly more likely to be secured by office propertiesandmuchmorelikelytobelocatedinaCBD. 4.2. BankDelinquencyModels The second step to understanding differences in loan performance is to estimate the probability of delinquencyasafunctionoftheloancharacteristicsthatareobservableforthesmallbanksample. We use the Y-14 data to estimate a delinquency function Dˆ (X ) = E (Delinquent |X , j = m i,j m i,j i,j LargeBank) where m indexes the model used to estimating Dˆ(·). In order to be able to benchmark the fitted delinquency rates against those observed in the data, we use a sample of loans and definitionofdelinquencyalignedwiththenon-performingloanratesavailableintheCallReports. Namely, the sample considered is the pool of loans with outstanding balances as of 2023:Q4 and delinquencyismeasuredbywhetheraloanis30+dayspastdueornonaccrual. In choosing a model to estimate the probability of delinquency, we face a trade-off between interpretability and flexibility; simpler models are better for understanding why performance differs acrosslenders,whereasmorecomplexonesmayimprovethefitandthusbetterassesstheextentto whichthevariablesconsideredcanjointlyaccountfordifferencesinperformance(withlessclarity 19

as to which particular variables matter). To provide a mix of of these benefits, we estimate four modelswithanincreasingdegreeofflexibility. 1. Linear probability model (OLS): Specification includes property type fixed effects, with the main other variables—CBD, Teleworkable share, and ln(Balance at Orig.)—also interactedwiththeofficeindicator. 2. Decision tree: Sequentially splits sample by feature values so as to achieve the best model improvement with each split. We set a high threshold for splitting to reduce the number of leafs and simplify interpretation. The algorithm thus splits the feature space into a small number of regions, and the predicted delinquency rate is the share of Y-14 loans within the regionthataredelinquent. 3. K-nearest neighbors (KNN): Finds the observations in the Y-14 data with the most similar featurevalues,andthepredicteddelinquencyrateistheaveragefortheK-closestY-14loans, weightingbythesimilarityoftheX vector. 4. Random forest: Repeatedly takes bootstrapped samples of the training set and fits shallow trees to the samples. The estimated delinquency rate is an average of the prediction of the treesacrossthesamples. FortheOLSestimator,theprimarydecisionisthespecification. MotivatedbytheresultsinTable1, whichshowthatthemainriskfactorsconsidereddisproportionatelyaffectofficeloanperformance, we choose a mostly linear model, with an additional interactions of Size, CBD and Teleworkable Sharewiththeofficeindicator. The latter three models estimate delinquency nonparametrically, and account for interactions and nonlinearitieswithoutusspecifyingthem. Instead,theprimarydecisioniswithregardtohyperparameters. Foreachmodel,wesearchoveraparametergridandusestratified5-foldcrossvalidation tochoosetheparametersthatproducethelowestmean-squarederrorintheleft-outdata. MoredetailontheseestimatorsandhyperparametertuningareinSectionA.3. 20

The coefficient estimates for the linear model are in column (1) of Table 2. Variables pertaining to loan size and telework ability are demeaned so the coefficient on office shows the higher delinquency rate for an office property outside of a CBD with an average level for other risk factors. Overall,theresultsofferfewsurprisesrelativetowhatwasfoundpreviously. Higherloanbalances are associated with higher delinquency, particularly for offices, consistent with previous results showing higher delinquency rates for larger properties and higher LTV loans. We also see higher delinquencyratesinCBDsandcitieswheremorejobscanbedoneremotely,particularlyforoffice properties. Column (2) presents estimates from the same specification but with the broader sample and definition of delinquency used in the previous analysis. The main findings generally hold, but the property-specificinterceptsareabithigher. Namely,thepredicteddelinquencyratesaresomewhat highersincetheybetteraccountformaturity-defaults,butthedriversofdifferencesindelinquency arenotmeaningfullydifferent. Columns(3)supplementsthepredictionsofdelinquencywithapredictionoftheyear-aheadprobabilityofdefaultbasedonbanks’internalriskratings. Ifsmallbanksareonlyoutperforminglarger banks because they have loans that are expected to deteriorate later (e.g., if stress were to start in CBDs before spreading to other markets), these forward-looking measures would allow us to pickupwhethersmallerbankshavecharacteristicsassociatedwithanexpectedfuturedeterioration in performance. For the most part, the factors associated with delinquency tend to be associated with expected future delinquency with a broadly similar intensity. If anything, estimated effects in column (3) tend to be a bit higher than estimates in columns (1) and (2), indicating that banks expect the risk factors associated with delinquency so far to be associated with somewhat further deteriorationinperformancegoingforward. The delinquency rate estimates for the decision tree estimator are visualized in Figure B.4. The first split in the tree is by whether the loan is secured by an office, the second is by whether the 21

Table2: BankDelinquencyModel 100×Delinquent Year-ahead (CallDefinition) (Inc. maturitydefault PD(%) andliquidation) (1) (2) (3) ln(BalanceatOrig.) 0.16** 0.25** 0.37** (0.04) (0.07) (0.06) CBD 0.34 0.37 1.30** (0.22) (0.25) (0.38) TeleworkableShare 4.17* 2.83 11.16** (1.78) (2.11) (2.43) Office 1.21** 1.28** 1.71** (0.23) (0.24) (0.28) ×ln(BalanceatOrig.) 2.19** 2.36** 2.90** (0.40) (0.42) (0.45) ×CBD 2.55* 3.35** 2.58* (1.12) (1.23) (1.20) ×TeleworkableShare 17.11** 15.19** 18.04** (5.36) (5.82) (6.28) Retail 0.55** 1.00** 0.74** (0.16) (0.23) (0.22) Industrial 0.03 0.24 -0.21 (0.14) (0.22) (0.19) Hotel 1.68** 2.46** 2.95** (0.56) (0.71) (0.80) Intercept 0.36** 0.51** 1.27** (0.08) (0.12) (0.19) R2 0.027 0.026 0.063 a Observations 46,925 46,925 39,419 Notes: Thistablepresentsestimatesfromtheequation: 100×Delinquent =α +β′(Office ×X)+γ′X +ε, i,23 p(i) i i i i where Delinquent is a delinquency measure as of 2023:Q4, α is a fixed effect for i’s property type, and X is i,23 p(i) i a set of risk factors that are observable both in Y-14 and RCA/CoreLogic data: the logarithm of the at-origination loanbalance, anindicatorforwhetherthepropertyisinaCBD,andtheshareofjobsini’sMSAthatareidentified as being able to be done at home by Dingel and Neiman (2020). Column (1) predicts delinquency for the sample of loans that are on the balance sheet as of the end of 2023, column (2) presents estimates using the measure of delinquency from Section 3 (which includes liquidated and performing ballooned loans as delinquent and paid-off loansasperforming),andcolumns(3)presentsequivalentanalysispredictingthereportedyear-aheadProbabilityof Default. Standarderrors, inparentheses, areclusteredbybank-originationyear. +,∗,∗∗ indicatesignificanceat10%, 5%,and1%,respectively. Sources: Y-14QH.2Scheduleandauthors’calculations. 22

office loan is large-sized (an at-origination balance over $23.3 million), and the third is based on whether the large office loans are in a high-telework-eligible market (Teleworkable Share> 0.44). The delinquency rates in the decision tree estimator are 0.5% for non-office loans, 1.4% for small officeloans,8.4%forlargeofficeloansinlow-teleworkmarkets,and23%forlargeofficeloansin high-teleworkmarkets. The KNN and random forest estimators do not have as clear a correspondence between features and the probability of delinquency. However, the fitted delinquency rates are highly correlated with the first two estimators, suggesting that common drivers are at play. Predictions from the treeestimatorhavecorrelationsof0.63withtheKNNpredictionsand0.88withtherandomforest predictions (see Figure B.5). This result indicates the broad categorization in the tree estimates— where the primary division is between large office loans and everything else—drives much of the variationinthemorecomplexestimates. 4.3. Cross-bankDifferencesinFittedNPLs Whatdotheseestimatesimplyfordifferencesinloanperformanceacrosslenders? Wenowusethe delinquency models discussed in Section 4.2 to generate predicted delinquency rates for different typesoflendersusingthecross-lenderportfoliodatadiscussedinSection4.1. Theobjectofinterestistheexpecteddelinquencyrate: FittedDelinquency j =∑ω i,j Dˆ m (X i,j ),whereω i,j istheshare i|j of j’sloanportfolioaccountedforbyloani. Thisestimateshowstheextenttowhichdifferencesin the performance of loans across lenders can be attributed to broad differences in the composition oftheirCREportfolios. IfthedriversofCREperformanceatsmallandlargebanksaresimilar,but small banks just perform better because loans are safer on the modeled dimensions, then the fitted delinquency rates should match those observed in the data. If small banks’ CRE loans perform betterforotherreasons(e.g.,duetorelationships,betterunderwriting,oragreatertendencytoevergreen),thentheirstrongerperformancewouldbeforunobservedreasons,andfitteddelinquency ratesacrossbanksofdifferentsizeswouldnotdifferbymuch. 23

Figure 4 plots observed NPLs by lender type based on Call Report data (the blue bars), and the fitted NPLs based on the four delinquency models (the bars below the blue bars). For the types of loans included in the previous analysis (multifamily and non-owner-occupied, nonfarm, nonresidential loans), small banks have NPL rates of about 0.6 percent, whereas large banks have NPL rates of around 2.6 percent. This 2-percentage-point differential is well explained by then differencesintheobservableloanandpropertycharacteristicsincludedinthedelinquencymodels. Across the four models, the fitted delinquency rate for small banks ranges from 0.99 percent in the random forest model to 1.21 percent in the OLS model. In other words, about 1.4 to 1.6 of the 2-percentage-point difference in NPLs across large and small banks—70 to 80 percent of the gap—can be attributed to differences in the composition of their CRE portfolio along a relatively small number of dimensions (loan size, property type, and geographic exposure to the shift to remotework). For large banks, the fitted delinquency rates align closely with the observed ones: fitted delinquency rates range from 2.19 percent to 2.63 percent, relative to an observed rate of 2.61 percent. BoththeOLS-andrandomforest-fitteddelinquencyratesarewithin2basispointsoftheobserved one. While this result is not particularly surprising given that the estimates are fit to large bank data,itdoesincreaseconfidenceinthemethodology. Itwouldbepossibleforthefitteddelinquency rates to deviate from the actual ones due to sampling problems with the RCA/CoreLogic data or residuals that correlated with loan size (since the portfolio aggregations are weighted).15 That the predictionsalignwiththeobserveddatasuggeststhatthesearenotmajorproblems.16 WhatdrivesthesedifferencesinfittedNPLs? WhiletheKNNandrandomforestmodelsarecom- 15Thefirstissuewouldappearifweover-orunder-sampledloansinawaythatcorrelatedwithloanperformance (e.g.,ifsmallerloanswereundersampledduetoreportingissues).TableB.1alsoindicatesthatthisisnotaproblem,as portfoliosatlargebanksintheRCA/CoreLogicdatamatchthoseintheY-14data.Thesecondissuecouldappearifthe waythemodelaccountsforsizeeffectsismisspecified. Forexample,thedecisiontreeestimatesonlyreflectwhether office loans are above a particular size threshold, whereas Figure 3 indicates that the effects of size are continuous. Thetreethereforeislikelytounderestimatedelinquencyfortheverylargeloans(whichgetthehigherweightinthe portfolio aggregations), which may be why it produces estimates that are lower than the other models, which allow sizeeffectstobemorecontinuous. 16Thereisalsosomeevidenceinthecross-sectionthatthemodelworksasintended.FigureB.6showsthatthefitted delinquencyrateatthebanklevelincreasesroughlyone-for-onewithbanks’realizednon-performingloanrates. 24

Figure4: Realizedvs. ExpectedNonperformingLoanRates ,jn I zY9O QHb SYlS ia33 bL II FMM # NGc SYzk ` N0RL7Ra3cj SYS4 zYOO lYfS lYfS H a<3 # NGc lYSO lY:O lYfk S l k /3ICN\n3N,wh` j3hVXW Notes: The top bars shows the 2023:Q4 nonperforming loan rate for banks with under (top set) and over (bottomset)$100billioninassets. Theremainingbarsgiveweightedaveragedelinquencyratesforloansin small and large banks’ portfolios based on a linear probability (red), a regression tree (green), a K-nearest neighbors(yellow),andarandomforest(purple)model. Sources: Y-14QH.2Schedule,MSCIRCA,CoreLogic,andauthors’calculations. plex enough to make it difficult to attribute differences in predictions to particular features, the OLS-andtree-basedpredictionscanbedecomposedtoclarifywhyloansatlargeandsmallbanks appeartobeperformingdifferently. Figure5presentsawaterfallchartshowingthevariousfactors contributing to differences in loan performance between large and small banks. The first red bar shows the residual, or the unexplained amount by which small banks were overperforming large banks. WiththeOLSmodel,smallbanksareexpectedtohaveanonperformingloanrateof1.2%, but only had a nonperforming loan rate of about 0.6%, meaning about 60 basis points of their overperformance is driven by factors not in the model. The other bars show how much individual variables in the regression in Table 2 contribute to differences in delinquency rates.17 While small banks appear to benefit from having smaller loans in general and fewer loans in troubled areas (CBDs or areas with more teleworkable jobs), by far the biggest component explaining their overperformance is the size-by-office interaction; the property size-by-office interaction effect accounts for almost 1 percentage point of the 1.6 percentage point difference in fitted NPL rates. 17Specifically, if β is the vector of regression coefficients and X is the balance-weighted-average vector of loan j characteristicsforlendertype j,thenthedifferenceinfitteddelinquencyratesbetweenlender jand j′isβ′(X j −X j′ ). Eachbarshowsaparticularβ k (X j,k −X j′,k ),wherekindexesvariablesintheregressionspecification. 25

Figure5: OLSNPLDecomposition k ZYzeX ZYSkX AYzeX zX AYzeX ZYzkX lYfSX ZYO9X l ZYzfX ZYl4X ZYzkX H a<3h# NG MTHchVXW ZYfSX S Y9OX bL IIh# NG MTHchVXW z w,N3n\NCI3/hGN #h3<a HhNRhj,3882 m N h 3 u UI C N 3 0 bC y 3 + # / i 3I 3 s R b a G C y 3h uh Q88C + , 3 # /h uh Q88C s , 3 Ra Gh uh Q88C , 3 ` 3j CI B N 0 n cjaC I ? Rj 3I Q88C , 3 h 3 3I i Notes: Blue bars show the delinquency rates for small (left) and large (right) banks. The bars in between show how much each variable from the regression in column (1) of Table 2 contributes to the difference (exceptforthefirstbarwhichgivestheunexplainedcomponent). Sources: Y-14QH.2Schedule,MSCIRCA,CoreLogic,andauthors’calculations. Most of the other effects work in the same direction, but nothing else contributes more than 0.28 percentagepointstothedifferences. Thoseothervariableshaveeithertoosmallaneffectondelinquency or not large enough differences across bank sizes to contribute as notably to differences in predicted NPL rates. In short, the OLS estimates indicate that about half of small banks’ superior CREperformancecanbeexplainedbythembeinglessexposedtolargeofficeloans. We get a similar picture when we decompose the differences in fitted delinquency rates using the tree estimates. Table 3 provides the predicted delinquency rate for different loan segments (column 1) and the share of large and small banks’ CRE portfolios in those segments (columns 2 and 3, respectively). The results show that while smaller banks have a modestly smaller office exposure (roughly 17.5%, relative to 20% for large banks), the bigger differentiator is the size of the office loans. Roughly 15% of CRE lending by large banks is against large office loans (offices 26

Table3: TreeDecomposition (1) (2) (3) Pr(Delinquent) LargeBank SmallBank Share(%) Share(%) Non-office 0.54% 79.73 82.53 SmallOffice 1.39% 5.62 13.58 LargeOffice,LowTelework 8.41% 11.54 3.42 LargeOffice,HighTelework 22.89% 3.11 0.46 Weightedaveragedelinquency 2.19% 1.03% Notes: Eachrowgivesthedelinquencyrateforaparticularleafinthetreemodel(column1),ortheshareoflargeor smallbanks’CREportfolioscomposedofloansinthatleaf(columns2and3, respectively). Thefitteddelinquency rate for a particular lender is the average of (1), weighted by the portfolio shares in (2) or (3), reported in the last row. Small/Largeofficeloansaredefinedbyanat-originationloanbalanceabove/below$23.3millionandLow/High Teleworkcitieshaveateleworkeligibleshareofemploymentbelow/above44.4%. Sources: Y-14QH.2Schedule,MSCIRCA,CoreLogic,andauthors’calculations. with an at-origination balance above $23.3 million), compared to only about 4% of CRE lending at small bank loans. As small office loans have a delinquency rate of 1.4% in the Y-14 data, while largeoneshavedelinquenciesabove8%(andmuchaboveinthecaseofhigh-teleworkareas),this composition produces large differences in the fitted nonperforming loan rates of large and small banks’CREportfolios. 5. CONCLUSION Rising interest rates and shifts in the demand for space have impaired the performance of many CRE properties. As banks are large holders of CRE loans, these developments have generated concernaboutCREexposureexacerbatingotherbankingsectorstrains. Using a combination of different sources, we shed light on the factors affecting loan performance across different types of lenders. CRE loans held by large banks were less likely to go delinquent in 2023 than those held by CMBS. For office loans, these differences can be accounted for by banksmakingsmallerloans,whereborrowershavemoreskininthegame(fromeitherrecourseor property equity). These factors only partially explain the differences in performance of non-office loans, suggesting that other factors, such as banks’ willingness to work out stressed loans, also 27

contribute. Lower NPLs at small banks are estimated to predominantly reflect small banks’ lower exposure to at-risk office loans (i.e., loans secured by larger office properties). This result indicates that smallbanks’comparativelylownonperformingloanrateismostlyduetothecompositionoftheir loan portfolio rather than, for example, “extend and pretend” activities delaying the realization of delinquency. Though large banks are more exposed to adversely affected segments of the CRE markets, they generally have high capital levels and low CRE concentrations that should enable themtoweatherthoselosses,allelseequal. All told, these findings suggest that strains would need to expand to other segments of the CRE market to cause systematic problems for the banking sector. Of course, just because small banks’ CRE loans have characteristics that have insulated them so far, this does not mean that they will be immune to future stresses. Some features of bank loans—modification ability, lower leverage, and lower at-risk office exposure—are likely to continue supporting performance in the future. However, there are some segments that have performed reasonably well that could reasonably deteriorateinthefuture. First,thereasonthatsmallerloanshaveoverperformedsomuchisnotclear, which makes it hard to trust that the effect will persist to the same extent. If small office loans start to perform more similarly to large ones, delinquency at small banks would increase notably. Second, while multifamily delinquency rates are low, they are rising amid elevated interest rates, operating expenses and competition from new supply. Third, this analysis has focused predominantly on CRE loans secured by income-producing properties, whereas construction loans were a primary drivers of stress in past crises. While the delinquency rate on bank construction loans has been modest so far, difficulties leasing new space or obtaining stabilized financing could create strains in the future as more projects exit construction into a challenging environment. Thus, while this study provides a description of why small bank CRE loan performance has held up so far, and gives some reason for optimism about the outlook going forward, the situation warrants monitoring,especiallyiftheCREmarketstartsshowingsignsofbroaderstrains. 28

References ACHARYA, V. V., M. P. RICHARDSON, K. L. SCHOENHOLTZ, B. TUCKMAN, R. BERNER, S. G. CECCHETTI, S. KIM, S. KIM, T. PHILIPPON, S. G. RYAN, ET AL. (2023): “SVB and beyond: Thebankingstressof2023,”AvailableatSSRN4513276. AN, X., Y. DENG, AND S. A. GABRIEL(2011): “Asymmetricinformation,adverseselection,and thepricingofCMBS,”JournalofFinancialEconomics,100,304–325. ASHCRAFT, A. B., K. GOORIAH, AND A. KERMANI (2019): “Does Skin-in-the-Game affect SecurityPerformance?” JournalofFinancialEconomics,134,333–354. 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 EstateFinanceandEconomics,55,1–31. ——— (2020): “Safe Collateral, Arm’s-Length Credit: Evidence from the Commercial Real EstateMortgageMarket,”TheReviewofFinancialStudies,33,5173–5211. BOARD OF GOVERNORS OF THE FEDERAL RESERVE SYSTEM (2023): “Financial Stability Report.” 16–17. url=https://www.federalreserve.gov/publications/files/financial–stability–report– 20230508.pdf. BOLTON, P. AND X. FREIXAS (2000): “Equity, Bonds, and Bank Debt: Capital Structure and Financial Market Equilibrium Under Asymmetric Information,” Journal of Political Economy, 108,324–351. DINGEL, J. I. AND B. NEIMAN(2020): “Howmanyjobscanbedoneathome?” JournalofPublic Economics,189,104235. FARIA-E CASTRO, M. AND S. JORDAN-WOOD (2023): ““Stress Testing” Banks on Commercial RealEstate,”EconomicSynopses,1–3. 29

——— (2024): “Commercial Real Estate Exposure and Bank Stock Returns,” Tech. rep., Federal ReserveBankofSt.Louis. FLYNN JR., S. J., A. GHENT, AND A. TCHISTYI (2023): “The Imitation Game: How EncouragingRenegotiationMakesGoodBorrowersBad,”Workingpaper. FLYNN JR, S. J., A. C. GHENT, AND A. TCHISTYI (2020): “InformationalefficiencyinsecuritizationafterDodd-Frank,”TheReviewofFinancialStudies,33,5131–5172. FRIEND, K., H. GLENOS, AND J. B. NICHOLS (2013): “An Analysis of the Impact of the Commercial Real Estate Concentration Guidance,” Tech. rep., Office of the Comptroller of the Currency. GHENT, A. C. AND R. VALKANOV(2016): “Comparingsecuritizedandbalancesheetloans: Size matters,”ManagementScience,62,2784–2803. GLANCY, D., J. R. KRAINER, R. J. KURTZMAN, AND J. B. NICHOLS (2022a): “Intermediary Segmentation in the Commercial Real Estate Market,” Journal of Money, Credit and Banking, 54,2029–2080. GLANCY, D. AND R. KURTZMAN (2022): “How Do Capital Requirements Affect Loan Rates? Evidence from High Volatility Commercial Real Estate,” Review of Corporate Finance Studies, 11,88–127. GLANCY, D., R. KURTZMAN, L. LOEWENSTEIN, AND J. NICHOLS (2023): “Recourse as shadowequity: Evidencefromcommercialrealestateloans,”RealEstateEconomics,51,1108– 1136. GLANCY, D. AND J. C. WANG (2023): “Lease Expirations and CRE Property Performance,” FederalReserveBankofBostonResearchDepartmentWorkingPapersNo.23-10. GLANCY, D. P., R. J. KURTZMAN, AND L. LOEWENSTEIN (2022b): “Loan modifications and thecommercialrealestatemarket,”FEDSWorkingPaper2022-050. 30

GRIFFIN, J. M. AND A. PRIEST (2023): “Is COVID Revealing a Virus in CMBS 2.0?” The JournalofFinance,78,2233–2276. GUPTA, A., V. MITTAL, AND S. VAN NIEUWERBURGH (2022): “Work from Home and the OfficeRealEstateApocalypse,”NBERWorkingPaper30526. JIANG, E. X., G. MATVOS, T. PISKORSKI, AND A. SERU (2023): “Monetary Tightening, CommercialRealEstateDistress,andUSBankFragility,”NBERWorkingPaper31970. MARSH, W. B. AND J. PANDOLFO (2024): “Banks’CommercialRealEstateRisksAreUneven,” EconomicBulletin,1–4. MONTE, F., C. PORCHER, AND E. ROSSI-HANSBERG (2023): “Remote Work and City Structure,”AmericanEconomicReview,113,939–981. PETERSEN, M. A. AND R. G. RAJAN (1994): “The benefits of lending relationships: Evidence fromsmallbusinessdata,”TheJournalofFinance,49,3–37. ROSENTHAL, S. S., W. C. STRANGE, AND J. A. URREGO(2022): “JUEinsight: Arecitycenters losingtheirappeal? Commercialrealestate,urbanspatialstructure,andCOVID-19,”Journalof UrbanEconomics,127,103381. WONG, M. (2018): “CMBS and Conflicts of Interest: Evidence from Ownership Changes for servicers,”TheJournalofFinance,73,2425–2458. 31

A. DATAAPPENDIX A.1. LargeBankandCMBSPanelData Loan level panel data are available for CRE loans held by banks with over $100 billion in assets andforCREloansheldinCMBSpools,butfromdifferentdatasources. Herewedescribeinmore detailhowweprocessandharmonizethesedatasources. BanksservesomeCRE-segmentsthatCMBSdonot. Tofocusonareaswherebothlendersoverlap, werestrictthebanksampletofirstlien,non-owneroccupied,non-constructionCREloans. Datais reported as of quarter end, so analysis in Section 3 uses the sample of loans that were outstanding asoftheendof2022(inordertolookattheeffectsofupcomingloanmaturities),whereasanalysis in Section 4 uses the sample of loans were still on balance sheet as of the end of 2023 (to align withCallReportreporting). Reporting of loan balances and collateral values for bank loans can be skewed by loan participations and cross-collateralization. For participations, the reported balance is prorated to the size of the bank’s participation. When predicting delinquency we are interested in the size of the loan itselfratherthantheindividualbank’sexposuretotheloan,sowescaletheloansizeupbydividing the reported balance by the share of the loan held by the bank to back out the borrower’s actual balance. To avoid double counting loans reported by multiple Y-14 banks, we only include observationswherethebankistheonesellingtheparticipationinterest. Cross-collateralizedloanshave collateral double-counted as applying to multiple loans (i.e., property values reflect the aggregate value of a collateral pool, but loan balances only reflect the balance on an individual loan). We therefore adjust property values and LTVs by prorating the portion of the collateral attributable to agivenloan. For CMBS data, we exclude agency CMBS loans and defeased loans. Similar to bank loan participations, CMBS loans are often split over multiple pari-passu pieces. To back out appropriate loan sizes, we compute outstanding values by summing balances across such loans, and to avoid 32

double counting, we drop duplicated observations from the sample. Data is updated monthly, but notnecessarilyatmonth-end. Theendof2022andendof2023sampleisasofthelastobservation reported for those years (i.e., the data from the December data update). When analyzing maturity outomes,weomitasmallnumberofloansthatmaturebetweentheirDecemberreportingdateand yearend. Somenon-corepropertytypeshaveinconsistentidentifiersacrossthebankandCMBSdata,sowe restrict our attention to core loan categories: multifamily, office, retail, industrial and hotels. In both data sets, we omit a small number of observations with missing information on geography, property values or loan balances. We also omit observations with LTVs that are not between 0 and 0.99 in order to reduce the effect of potential reporting errors. The Teleworkable share is missing for loans against properties outside of cities. To avoid systematically dropping these observations,wesettheTeleworkablesharetothevalueforthe10thpercentileofthenon-missing sample (about 0.32), under the assumption that these more rural locations are towards the bottom of the telework-exposure distribution. When we add a dummy variable to the OLS specification for whether the Teleworkable share is missing, the estimates are near 0, which validates that loan performance is indeed similar for observations without data and observations around that portion ofthedistribution. A.2. CREOriginationData While detailed, panel data on CRE loans are not generally available for small banks, some information on loan terms at origination are available. This subsection describes how we use RCA and CoreLogic data to form data approximating the composition of at-origination characteristics ofoutstandingloansacrosslendertypes. We use data from RCA to provide information on CRE loan portfolios for loans with an atorigination balance of $2.5 million or more. While the data does not specify whether mortgages arestilloutstanding,wecanreasonablyinferwhethertheyarebasedonthepresenceofsubsequent 33

transactions. Specifically, we drop loans against properties that are later refinanced or sold, unless thesaleismarkedasinvolvingtheassumptionofexistingdebt. Wealsodroploanswithamaturity datebefore2023incasethoseloanspaid-offwithoutmortgagefinancingortherefinancedoesnot appear in the data. Loans without a maturity date listed are assumed to have a 10-year loan term. We restrict the sample to loans that finance an already-built investment property to remove types of loans that are not typically part of CMBS deals (i.e., we exclude owner-occupied properties or properties purchased for construction or redevelopment). Data is reported at the property level withloanbalancesallocatedacrossthepropertiescoveredinadeal. Toaggregatetothedeallevel, wesumloanbalanceswithinaparticularlender-dealidcombination. Propertytypesandlocations pertaintothelargestpropertyinthetransaction(byprice). We use data from CoreLogic to provide information on CRE loans with an at-origination value under$2.5million. Asthesamplecoverscommercialmortgageswithopenliens,thereisnoneedto impute whether loans are still outstanding. We omit loans flagged as construction loans or owneroccupied loans. Data is reported at the parcel level, with mortgage information repeated when multiple parcels are covered by the lien. To avoid double-counting, we only keep the mortgage information associated with the largest property (the parcel with the highest assessed value).18 Necessary data is sometimes missing, most commonly because a generic “commercial” property type is identified rather than the specific type (e.g., “office”, “retail”, etc). When computing the portfolio aggregations, we assume that missing observations are reflective of other loans under $2.5 million and scale up the aggregation weights for the observations where data is available to accountfortheseloans. BothCoreLogicandRCAreportlendernamesratherthanrssdcodes. Toidentifythetypeofentity making the loan, we fuzzy name match the lender names in each data set to National Information Center (NIC) institution data. After cleaning to standardize punctuation and other common words (e.g., replacing “national association” with “na” and terms such as “bk”, “bancshares”, etc. with 18Weassumetwoparcelsarereportingthesamemortgagewhentheyhavethesamemortgagetransactionid,lender, loanamount,mortgagedateandmortgagepurposerecorded. 34

”bank”), we match lender names based on the cosine similarity of TFIDF vectors (identifying lenders with similar character bigrams). Sincesome smallbanks haveduplicate namesbut tendto have a limited geographic footprint, we disambiguate banks with similar names by giving priority to name matches in a county where the bank has a branch (using Summary of Deposits data). Once a mortgage is matched to a bank, we assess whether they are a large or small bank based on whether the regulatory high holder (if applicable) has more than $100 billion in assets as of 2023:Q4 Call Report/Y-9C data.19 CMBS loans, which are frequently originated by banks, are identified by whether RCA codes the lender group as CMBS. Any mortgage in CoreLogic that linkstoabankisassumedtobeabankportfolioloansinceCMBSgenerallydonotoriginateloans inthatsizerange(seeGlancyetal.,2022a). A.3. MachineLearningEstimatorsandHyperparameterTuning ThethreemachinelearningestimatorsareestimatedinPython. Thedecisiontree,K-nearestneighbors and random forest models are estimated using the DecisionTreeRegressor, KNeighborsRegressor, and RandomForestRegressor classes, respectively, of the scikit-learn package. Hyperparametersaresettothevaluesthatprovidethebestperformance(theestimateddefaultprobabilities withthelowestmeansquarederror)using5-foldcross-validation,stratifyingbytheoutcomevariable. Anyparametersnotdiscussedherearesettodefaultvalues. For the decision tree estimates, the parameter considered is the “min impurity decrease”, which determines how much the model needs to improve in order to generate an additional split to the feature space. We search for the best estimator over a grid [1,2,...10]×10−5 and find it to be 3×10−5. This results in a tree with a simple structure with only four terminal nodes and each additionalsplitoccurringinthenodewiththehighestprobabilityofdefault. Thetreethusidentifies a hierarchy of compounding risk factors: office loans underperform other loans, large loan sizes compound the risk of office loans, and high telework exposure compounds risks to large office 19Loans from foreign banks that have an intermediate holding company subject to stress tests are identified as belonging to a large bank (e.g., loans marked as provided by RBC are attributed to RBC US and counted under the largebankcategory). Loansfromotherforeignbanksareexcludedfromtheanalysis. 35

loans. FortheKNNestimates,theparametersconsideredarethenumberofneighbors(K=[100,200,...500]) and the intensity with which we downweight neighbors that are further away (weights decay exponentially in the Euclidean distance between the X-vectors at a rate in [0,5,...25]). We find the optimal parameter values on that grid are 500 neighbors, with weights declining at a rate of 15. Though the number of neighbors is at the top of the grid (suggesting that there could be a benefit to including more neighbors), the gradient from increasing K is flat so we keep K at 500. Since thisdistanceandtheselectionofneighborsissensitivetothescaleofthefeatures,weusemin-max scalingforanycontinuousfeature. Finally, for the random forest estimates, we search over a parameter grid of the number of trees in theforest(“n estimators”=[100,200,...500])andtheminimumimpuritydecreaseforasplit: (“min impurity decrease”=[1,2,...10]×10−5). B. ADDITIONALTABLESANDFIGURES FigureB.1: NonperformingLoanRateDecomposition SlX lzX MRNU3a8RaLCN<hHR Nh+RLURcCjCRN MRNU3a8RaLCN<hHR Nh` j3c +RNcjan,jCRNh N0hH N0h/3q3IRUL3Nj # NGh+H/ SzX MRN8 aLhMRNa3cC03NjC I # NGhM7M` KnIjC8 LCIw S9X # NGhK7 4X +K#b fX SzX :X 9X lX zX zX lzz4 lzSS lzS: lzSe lzlz lzlk lzz4 lzSS lzS: lzSe lzlz lzlk (a)DecompositionofBankNonperformingloans (b)NonperformingLoansbyLoanType Notes: The left panel decomposes the bank nonperforming loan rate into delinquencies for construction and land development (blue), nonfarm nonresidential (red), and multifamily (green) loans. The right panel presents the equivalent nonperforming loan rates for each bank CRE loan category as well as the overall delinquency rate for CMBS loans. Nonperforming loans are loans that are 30 days or more past due or nonaccrual,plottedasashareofaggregateoutstandingbalances. CMBScalculationsexcludedefeasedand REOloans. Sources: CallReports,Morningstar,andauthors’calculations. 36

FigureB.2: NonperformingLoanRatesbyBankSize SlX SlX MRNU3a8RaLCN<hHR Nh` j3 MRNU3a8RaLCN<hHR Nh` j3 Qq3ah1Szz$NhCNh cc3jc Qq3ah1Szz$NhCNh cc3jc SzX mN03ah1Szz$NhCNh cc3jc SzX mN03ah1Szz$NhCNh cc3jc 4X 4X fX fX :X :X lX lX zX zX lzz4 lzSS lzS: lzSe lzlz lzlk lzz4 lzSS lzS: lzSe lzlz lzlk (a)Nonowner-occupiedNFNRDelinquency (b)OtherCREDelinquency SzzX SzzX MRNRsN3ahQ,,nUC30 Qj@3ah+`2 h +`2h/3ICN\n3N,w h /3ICN\n3N,w 4zX HR, IhK3 N 4zX HR, IhK3 N h h fzX fzX h h :zX :zX h h lzX lzX SzX SzX zX zX 1SzzhLN 1Sh$N 1Szh$N 1Szzh$N 1Shja 1SzzhLN 1Sh$N 1Szh$N 1Szzh$N 1Shja # NGh cc3jchVIR<hc, I3W # NGh cc3jchVIR<hc, I3W (c)Nonowner-occupiedNFNRDelinquency (d)OtherCREDelinquency Notes: The figure plots CRE nonperforming loan rates for nonowner-occupied nonfarm nonresidential (NFNR) loans (left panels) and other CRE loans (right panels). Nonperforming loans are loans that are 30 days or more past due or nonaccrual, plotted as a share of aggregate outstanding balances. Other CRE includesmultifamily,constructionandlanddevelopment,andowner-occupiedNFNRloans. Thetoppanels plot NPLs over time for banks above (blue) and below (green) $100 billion in assets. The bottom panels plot 2023:Q4 NPLs against bank size. Blue dots report individual banks’ NPLs, while the black line plots an estimate of the average for banks of that size (the kernel-weighted local mean). The scale on the y-axis expandsafter20%topreventoutlierresponsesfromobscuringvariationwithinnormalbounds. Sources: CallReportsandauthors’calculations. 37

FigureB.3: NonperformingLoansByPropertyType fX SlX Qj@3a Q88C,3 BN0ncjaC I ?Rj3I 9X ?Rj3I SzX `3j CI `3j CI Qj@3a KnjC8 LCIw KnjC8 LCIw :X Q88C,3 4X BN0ncjaC I kX fX lX :X SX lX zX zX lzSS lzS: lzSe lzlz lzlk lzSS lzS: lzSe lzlz lzlk (a)NPLDecompositionbyPropertyType (b)NPLRatesbyPropertyType Notes: The left panel decomposes the bank nonperforming loan rate for income producing properties at largebanksbypropertytype,withtheshadedregionshowingthecontributionofaparticularpropertytype. Therightpanelpresentsthenonperformingloanratesforeachpropertytype. Sources: Y-14Q,andauthors’calculations. FigureB.4: DecisionTreeDelinquencyEstimates Office <= 0.5 samples = 46925 value = 0.008 ln(Bal at Orig) <= 16.963 samples = 42100 samples = 4825 value = 0.005 value = 0.032 Teleworkable Share <= 0.444 samples = 3957 samples = 868 value = 0.014 value = 0.112 samples = 702 samples = 166 value = 0.084 value = 0.229 Notes: Decision tree estimated default probabilities. Each node defines the split that occurs at the node (if there is one), the number of observations in the training data in that node (samples) and the share of those observations that are delinquent (value). Nodes to the left correspond to feature values under the splitting thresholdandnodestotherightareabovethethreshold. Sources: Y-14QH.2Schedule. 38

FigureB.5: CorrelationAcrossModelPredictions KNN Tree Rand Forest OLS NNK eerT tseroF dnaR SLO 1.00 1 0.75 0.50 0.63 1 0.25 0.00 0.65 0.88 1 0.25 0.50 0.51 0.38 0.43 1 0.75 1.00 Notes: CorrelationsoffitteddelinquencyratesintheRCA/CoreLogicsample. Sources: Y-14QH.2Schedule,MSCIRCA,CoreLogic,andauthors’calculations. 39

FigureB.6: Cross-sectionalRelationshipBetweenFittedDelinquencyRatesandNPLs k l S z AS WXVhw,N3n\NCI3/hI nj, HR, IhK3 N HCN3 ah7Cj :9h03<a33hICN3 z S l 7Cjj30h/3ICN\n3N,whV# NGhH3q3IW Notes: Figure plots the relationship between the weighted average fitted delinquency rate for the loans matched to a particular bank in the RCA/CoreLogic data and that bank’s non-performing loan rate as of 2023:Q4. The red line shows the relationship based on a linear regression and the blue line based on local meansmoothing,bothweightingbybanks’volumeofoutstandingCREloans. Thedashedlinegivesthe45 degreeline,ortherelationshipwewouldgetwithaperfectmodel. Weaverageacrossthefourdelinquency modelstoproducethebank-specificfitted-deliquencyrate. Sources: Y-14QH.2Schedule,CoreLogic,MSCIRCA,andauthors’calculations. 40

TableB.1: AverageCREloancharacteristicsAcrossSamples Data Y-14 Morningstar Y-14 Morningstar RCA/CoreLogic Sample Large CMBS Large CMBS Large Small Banks Banks Banks Banks (1) (2) (3) (4) (5) (6) Delinquent 1.43 4.51 3.77 7.12 i,23 Delinquent(Calldefn.) 0.99 1.99 3.23 3.57 2.61 0.59 ln(ValueatOrig.) 15.84 16.79 17.62 19.22 LTVatOrig. 0.53 0.61 0.58 0.61 Recourse 0.59 0.00 0.45 0.00 Occupancy 0.94 0.89 0.90 0.87 DebtYield<.08 0.31 0.20 0.39 0.26 ln(BalanceatOrig.) 15.16 16.22 17.10 18.66 17.24 15.47 CBD 0.08 0.08 0.16 0.22 0.15 0.07 TeleworkableShare 0.39 0.37 0.39 0.37 0.39 0.37 Office 0.11 0.18 0.23 0.32 0.20 0.17 Retail 0.17 0.37 0.14 0.26 0.13 0.25 Industrial 0.08 0.06 0.12 0.10 0.17 0.20 Hotel 0.03 0.15 0.07 0.20 0.05 0.10 N 42267 15532 42267 15532 133954 209810 Weighted ✓ ✓ ✓ ✓ Notes: Columns(1)and(2)presentaverageloancharacteristicsforthesampleofloansfromlargebanksandCMBS thatwereoutstandingasoftheendof2022(thesamplestudiedinSection3). Columns(3)and(4)providethesame information, but weighting by the at-origination loan balance. Columns (5) and (6) present weighted average loan characteristicsforthesampleofoutstandingloansatlargeandsmallbanks,respectively,basedontheRCA/CoreLogic data. The measure of delinquency for the RCA/CoreLogic sample is the aggregate delinquency rate for nonowneroccupiednonfarmnonresidentialandmultifamilyCREloansfromCallReportssinceinformationonloanperformance isgenerallyunavailableintheoriginaldata. Sources: Y-14QH.2Schedule,Morningstar,CoreLogic,MSCIRCA,CallReports,andauthors’calculations. 41

TableB.2: LoanModificationsbyLenderType 100×Extension i,23 FullSample Offices (1) (2) (3) (4) (5) (6) (7) (8) CMBS -5.37** -6.40** -6.21** -6.26** -6.26** -9.74** -9.41** -9.42** (0.50) (0.55) (0.53) (0.53) (0.70) (1.12) (1.08) (1.08) Maturing 33.20** 32.27** 31.87** 31.60** 37.85** 36.59** 34.85** 34.70** (2.18) (2.14) (2.15) (2.18) (2.32) (2.26) (2.26) (2.25) Office 3.41** 1.43** 1.40** 1.33** (0.47) (0.40) (0.38) (0.37) LTVatOrig. 4.77** 4.60** 4.21** 9.43** 10.14** 9.81** (0.66) (0.68) (0.68) (2.15) (2.23) (2.24) ln(ValueatOrig.) 2.10** 1.89** 1.80** 2.58** 2.38** 2.35** (0.19) (0.18) (0.18) (0.32) (0.32) (0.32) Recourse 1.36** 1.47** 1.43** -0.13 -0.07 0.00 (0.26) (0.27) (0.26) (1.05) (1.02) (1.03) CBD 0.95** 0.91** 1.36 1.14 (0.28) (0.27) (0.83) (0.84) TeleworkableShare -1.94 -2.66 -3.64 -4.33 (1.91) (1.93) (4.92) (4.98) Occupancy -4.80** -4.01* (1.03) (1.92) DebtYield<.08 0.42+ 0.95 (0.23) (0.71) R2 0.231 0.246 0.241 0.243 0.256 0.275 0.262 0.264 a Observations 59,179 59,095 54,326 53,616 7,849 7,829 7,189 7,064 OtherPropertyFixedEffects? ✓ ✓ ✓ ✓ Notes: Thistablepresentsestimatesfromtheequation: 100×Extension =β CMBS +β Maturing +β Office +γ′X +ε, i,23 1 i 2 i,23 3 i i,23 i whereExtension isanindicatorforwhetherloanireceivedanextension(i.e.,haditsmaturitydatepushedout)in i,23 thelastyear. ThemainindependentvariablesofinterestarewhetherloaniisinaCMBSpool,whethertheloanwas scheduledtomaturein2023, andwhethertheloanissecuredbyanofficeproperty. Fixedeffectsforotherproperty types are included but not displayed (multifamily is the omitted category). Column (2) adds controls for the loan’s at-originationLTV,whethertheloanhasrecourse,andforthelogarithmofthepropertyvalueatorigination. Column (3)addscontrolsforwhetherthepropertyisinaCBDandtheshareofthecity’semploymentthatcanbedoneathome (DingelandNeiman,2020). Column(4)addscontrolsfortheoccupancyandanindicatorforwhetherthedebtyield isunder8%(bothasofayearpreviously). Columns(5)to(8)repeatthesameanalysisbutrestrictthesampletooffice properties. Standarderrors,inparentheses,areclusteredbybank-originationyearforbankloansandCMBSdealfor CMBSloans. +,∗,∗∗indicatesignificanceat10%,5%,and1%,respectively. Sources: Y-14QH.2Schedule,Morningstar,andauthors’calculations. 42