Corporate stress and bank nonperforming loans: Evidence from Pakistan

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1327 August 2021 Corporate stress and bank nonperforming loans: Evidence from Pakistan Ali M. Choudhary and Anil K. Jain Please cite this paper as: Choudhary, Ali M. and Anil K. Jain (2021). “Corporate stress and bank nonperforming loans: Evidence from Pakistan,” International Finance Discussion Papers 1327. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2021.1327. NOTE: International Finance Discussion Papers (IFDPs) 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 International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Corporate stress and bank nonperforming loans: Evidence from Pakistan M.AliChoudhary* StateBankofPakistan CentreforEconomicPerformance,LondonSchoolofEconomics AnilK.Jain BoardofGovernorsoftheFederalReserveSystem August19,2021 Abstract UsingdetailedadministrativePakistanicreditregistrydata,weshowthatbankswithlow leverageratiosarebothsignificantlyslowerandlesslikelytorecognizealoanasnonperforming than other banks that lend to the same firm. Moreover, we find suggestive evidence that this lack of recognition impedes loan curing, with banks with low leverage ratios reporting significantlyhigherfinaldefaultratesthanotherbanksforthesameborrower(evenaftercontrollingfordifferencesinloanterms).Ourempiricalfindingsareconsistentwiththetheoretical predictionthatclassifyinganonperformingloanismoreexpensiveforbankswithlesscapital. JELClassification: G21,G33 Keywords: Creditmarkets,banks,corporatedebt,evergreening,nonperformingloans. *We would like to thank Ryan Banerjee, Bastian von Beschwitz, Mark Carey, Stijn Claessens, Ricardo Correa, LeonardoGambarcorta,YesolHuh,DavidJenkins,ChristopherKarlsten,LoganLewis,AnnieMcCrone,CameliaMinoiu,andseminarparticipantsattheBankforInternationalSettlements(BIS),theGraduateInstituteofInternational andDevelopmentStudies,andtheInternationalFinanceandBankingSocietyforhelpfulcomments.AnilJainisgratefultotheBIS,whichprovidedsubstantialsupportthroughtheBISCentralBankFellowshipprogram. Thefindings andconclusionsinthispaperaresolelytheresponsibilityoftheauthorsandshouldnotbeinterpretedasreflectingthe viewsoftheBoardofGovernorsoftheFederalReserveSystem,anyotherpersonassociatedwiththeFederalReserve System,ortheStateBankofPakistan.Apreviousversionofthispaperwascalled”BankLendingto(Zombie?)firms.” 1

1 Introduction Banks play a crucial role in the efficient allocation of credit through screening and monitoring firms as well as enforcing credit contracts. However, a bank’s incentives are not necessarily aligned with either their investors or the social planner. In this paper, we present evidence that highlights one potential friction—banks with low leverage ratios inefficiently forbearing loans to potentially insolvent firms. Our evidence is consistent with a bank’s incentive to ration scarce regulatorycapitalcausingworsecreditoutcomes. Using credit registry data from Pakistan’s central bank, we show that banks with low leverage ratiosarebothsignificantlyslowerandlesslikelytorecognizealoanasoverduethanotherbanks. Specifically,usinganempiricalstrategysimilartoKhwajaandMian[2008],weisolatecharacteristicsofbanklendingbyanalyzingfirmsthatreceivemultipleloansfromdifferentbanks. Moreover, wefindsuggestiveevidencethatthislackofrecognitionimpedesloancuring,withbankswithlow leverageratiosreportingsignificantlyhigherfinaldefaultratesthanotherbanksforthesameborrower(evenaftercontrollingfordifferencesinloanterms). Ourempiricalfindingsareconsistent withthetheoreticalpredictionthatclassifyinganonperformingloanismoreexpensiveforbanks withlesscapital. Bymaskingthestatusofaloan,banksreducetherequiredloanlossprovisions, consequentlyartificiallymaintaininghigherregulatorycapitalandleverageratios(Bushmanand Williams [2015]). In turn, this potentially allows the bank to avoid raising new, costly external financingandpossiblyattractingadditionalregulatoryscrutiny. Moreover,wefollowthetime-pathoffirmborrowingandbanklending. Wefirstshowthatfollowinganoverdueloan,thenumberofbanklendersandthetotalamountofbankloansdramatically fall. Second, we show that banks with low leverage ratios did not increase their share of lending todistressedfirmsrelativetootherbanks,butrather,reducedtotallendingtothesefirms. Finally, we examine whether there are other possible theoretical explanations for slower nonperformingloanrecognitionbybankswithlowleverageratios. Weexaminethreepossibilities. First, do firms prefer to repay banks with lower leverage ratios more than other banks? Second, do banks with low leverage ratios monitor their loans less and, consequently, have higher loan defaults (Holmstrom and Tirole [1997], Allen et al. [2011], Mehran and Thakor [2011])? Third, do banks with low leverage ratios utilize superior information and efficiently forbear their loans to firms(Rajan[1992])? Wedonotfindstrongevidencetosupportanyofthesealternativeexplana- 2

tions. What are the welfare implications of delaying the recognition of nonperforming loans? Theoretically, the implications are ambiguous. On the one hand, if a borrower faces temporary liquidity shocks but is financially solvent, a lender providing additional funds and forbearing the initial loan can be both productive and welfare-improving (Tracey [2019] and Brunnermeier and Krishnamurthy [2020a]). For instance, Fukuda and Nakamura [2011] argue that Japanese banks were successfulinreviving“zombiefirms”andavoidingbankruptcy. Moreover,inresponsetothelarge financialshockstemmingfromtheCOVID-19pandemic,manyinternationalregulatorshavepromoted loan and regulatory forbearance (Financial Stability Board [2020]) with some arguing that banksshoulddomoreevergreening(Schivardietal.[2020]andBrunnermeierandKrishnamurthy [2020b]). Ontheotherhand,thereissubstantialtheoreticalandempiricalliteraturedescribingthenegative effectsofzombielending. Increasedbanklendingtoinsolventfirmsreducesbankprofitabilityand increases financial stability risks (Blattner et al. [2019]). Moreover, the effects of zombie lending extendfarbeyondthedirectbanksandfirmsinvolved. Caballeroetal.[2008],Kwonetal.[2015] structurallymodelhowzombielendinginJapancausedaseveremisallocationofcapitalbysimultaneously propping up inefficient and unproductive firms while starving new, potentially more productivefirmsofventurecapital,therebydistortingtheallocativeroleofpricesandsubsequent decisionsonemploymentandinvestment. Incomplete information on the true nature of banks’ asset quality has financial stability implications. Indeed, banks that disguise nonperforming loans may have insufficient loan loss reserves to cover losses on their loan portfolio, and in extreme cases, insufficient capital. Moreover, a loss of confidence in banks’ asset quality and a mere reduction in the credibility in banks’ reported asset quality, can substantially undermine trust in already underdeveloped financial system, as depositors,investors,andbondholderswithdrawfundingtobanks. Our paper contributes to the large literature on such “zombie lending,” also known as “evergreening.” Zombie lending has been defined variously as lending to firms with negative profits (McGowanetal.[2017],BanerjeeandHofmann[2018]),subsidizingcredit(Caballeroetal.[2008], FukudaandNakamura[2011],Kwonetal.[2015],Acharyaetal.[2019]),orlendingtofirmswith lowexpectedfuturegrowthrates(BanerjeeandHofmann[2018]). Weidentifysuggestivezombie lendingbyshowingsystemicdelayedloanrecognitionbybankswithlowleverageratiosrelative 3

tootherbanks. We present evidence that banks are motivated to zombie lend due to incentives to ration scarce bank capital (similar to Peek and Rosengren [2005], Storz et al. [2017], Caballero et al. [2008], Acharya et al. [2019], Bonfim et al. [2020]). However, there a number of additional theoretical motivations for zombie lending. Rajan [1994] theoretically and Hertzberg et al. [2010] and Tantri [2021]empiricallydemonstratehowprincipal-agentproblems,specificallycareerconcerns,canfacilitatezombielending. BrucheandLlobet[2014]showthatzombielendingcanbeanoutcomeof insolvent banks “gambling for resurrection”. Hu and Varas [2020] theoretically show how banks maycontinuetolendtounprofitablefirmsduetotheprospectoffuturemarketfinancing. Zombielendinghasbeenfoundtobepervasivewithevidencecitedinmanydifferenteconomies, inbothadvancedandemergingmarkets—forexample,the“savingsandloancrisis”intheUnited States (Kane [1989]); the Japanese banking crisis in the 1990s (Peek and Rosengren [2005], Caballero et al. [2008] and Giannetti and Simonov [2013]); European banks (Acharya et al. [2019]) andItalianbanks(Schivardietal.[2017])followingtheGreatFinancialCrisis;Indianrurallenders (Tantri [2021]); and in Argentina (Hertzberg et al. [2010]). Two recent papers have also documented the growth in the fraction of zombie firms, suggesting that zombie lending may be rising. McGowan et al. [2017] document increases in the share of zombie firms in nine advanced economies since the mid-2000s and Banerjee and Hofmann [2018] show an increase in zombie firmsinfourteenadvancedeconomiessincetheearly1980s. By analyzing the time-path of nonperforming loans (from performing, to overdue, to eventual potential default), we contribute to a nascent literature that is starting to examine in more detail thedynamicsofnonperformingloanformationandresolution(LaevenandValencia[2013,2018], Arietal.[2019]). Tosomeextent,weareanalyzingthedynamicsofbanklendingtozombiefirms, inthewaythatBanerjeeetal.[2020]analyzethepathofzombiefirms. Therestofthepaperisorganizedasfollows: Section(2)outlinesthePakistanicreditregistrydata andthedatausedinourpaper. Section(3)presentsevidencethatbankswithlowleverageratios delayed the recognition of bad loans consistent with the theoretical prediction that classifying a nonperformingloanismoreexpensiveforbankswithlesscapital. Section(4)providesrobustness testsforalternativepossibletheoreticalexplanations. Finally,Section(5)concludes. 4

2 Data WeuseadministrativedataontheuniverseofallPakistanicorporateloansfromPakistan’scentral bank, the State Bank of Pakistan. Pakistan’s credit registry contains the universe of all corporate loans from all officially designated financial institutions in Pakistan, including loans from public banks,privatebanks,Islamicbanks,andnon-bankfinancialinstitutionssuchastrustlendersand leasing companies.1 The data includes key information on firm loans, including information on thelender,theloansize,whethertheloanissecured,andtheperformancestatusoftheloan. The creditregistryalsocontainsinformationoninterestratesandthematuritydatesofloans,butdata on these variables is sometimes missing for some firms. This dataset has been used in numerous papers including Khwaja and Mian [2005], Khwaja and Mian [2008], Choudhary and Limodio [2017],andChoudharyandJain[2019]. Our credit registry dataset stretches, which from 2007:Q1 to 2012:Q4, contains 58,206 firms and 94,483differentbank-firmrelationships. Moreover,ofthe107financialinstitutionsinthedataset, only 29 institutions report capital and leverage metrics to the central bank—these institutions are the focus of our study. Similar to Khwaja and Mian [2008] and Choudhary and Jain [2019], since firms may have multiple loans at the same bank, for each firm we aggregate all of its loans at a specific bank to create measures of that firm’s total debt at that bank. Because part of our paper’s aim is to analyze how a firm transitions from having a loan overdue more than 90 days to potentially loan default, and how banks subsequently respond to these different events, we exclude all firms that had a nonperforming loan at the start of our dataset, since we cannot track whentheirfirstloanwentoverdue. Ourpaperfocusesonanalyzingnonperformingloans,soweexploitthethreedifferentdefinitions inthecreditregistry. First,theleastsevere,theloanisoverduemorethan90daysbutlessthan365 days. Second,theloanisoverduemorethan365daysbuthasnotdefaulted. Finally,loandefault, where we code a loan as defaulted if any of the following three events are reported to the credit registry: aloaniswrittenoff,theloanisrestructured,orthebankinitiateslitigationtorecoverthe loan. 1Foreaseofexposition,werefertoallfinancialinstitutionsas“banks”unlessotherwisestated. 5

3 Results 3.1 Recognitionofnonperformingloansbybankswithlowleverageratios Thekeyresultfromthissectionisthatsomebanks—specifically,bankswithlowerleverageratios— weremorelikelytodelaytherecognitionofnonperformingloans. Toidentifythisresultweutilize a Khwaja and Mian [2008] strategy. Specifically, we compare loan outcomes for a firm that borrowedfrommultiplebanks. Ourempiricalfindingsareconsistentwiththetheoreticalprediction that classifying a nonperforming loan is more expensive for banks with less capital. By masking thestatusofaloan,banksreducetherequiredloanlossprovisions,consequentlyartificiallymaintaininghigherregulatorycapitalandleverageratios(BushmanandWilliams[2015]). Inturn,this potentially allows the bank to avoid raising new, costly external financing and attracting additionalregulatoryscrutiny. Westartbyexaminingiftherearesystemicdifferencesacrossbanksinthedesignationofnonperformingloans. Specifically,forthosefirmsthathadanonperformingloanduringourdataset,we examine if banks with less capital (lower capital ratios or lower leverage ratios) were less likely to be the first bank to designate this loan as nonperforming. Since banks with lower capital ratiosmay lendtodifferent firms(thathave differingratesof havingnonperformingloans), inthis test, werestrictourattentiontothosefirmsthatborrowfrommultiplebanks—thatis, werelyon within-firmvariation. Werunthefollowingcross-sectionalregression: Firstoverduebank(NPL) = β ×Measureofbankcapital +β ×Controls +(cid:101) (1) b,f 1 b 2 b,f b,f where “First overdue bank (NPL) ” is a dummy variable equal to 1 if bank b was the first bank b,f to designate a loan from firm f as nonperforming in the period 2007:Q2 to 2012:Q4 (the length of our dataset).2 “Measure of bank capital ” is a measure of bank capital; for our preferred reb gressions it is a dummy variable for whether bank b’s leverage ratio is in the bottom quartile in 2007. Additionally, we investigate differences in loan outcomes using different nonperforming loandefinitions: overduemorethan90days,overduemorethan365days,andloandefault. 2Asdescribedearlier,weomitallfirmsthathadanoverdueloaninthefirstperiodofourdataset(2007:Q1)because weareunabletodeterminewhentheloanfirstbecameoverdue. 6

The key result from table (1) is that for those firms that borrowed from multiple banks, banks with low leverage ratios were relatively slower to designate a loan as overdue than other banks. In column 1, for firms that borrowed from multiple banks, we see that banks with low leverage ratioswereover5percentagepointslesslikelytobethefirstbanktodesignatealoanasoverdue more than 90 days than other banks. In column 2, we see a similar pattern, with banks with low leverage ratios more than 6 percentage points less likely to be the first bank to designate a loan as overdue more than 365 days. Interestingly, when it comes to designating a loan as defaulted, banks with low leverage ratios were equally, or more, likely to be the first lender to designate a loanasdefaulted—afindingweexpoundinthenextsetoftables. Table 1: Differences in the first lender to designate a loan as nonperforming: Banks with low leverageratios (1) (2) (3) Firstoverduebank(90+) Firstoverduebank(365+) Firstdefaultbank Lowleveragebank -0.056∗∗∗ -0.077∗∗∗ 0.014 (0.020) (0.025) (0.027) Observations 5602 3161 2382 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 2161 1114 728 FirmFEs Yes Yes Yes Dep.variablemean 0.27 0.28 0.27 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexamineswhetherbankswithlowleverageratioswereslowertodesignatealoanasoverduerelativetoother banksforthesamefirm. “Firstoverduebank(NPL) ”isadummyvariableequalto1ifbank b wasthefirstbank b,f todesignatealoanfromfirm f asnonperformingintheperiod2007:Q2to2012:Q4(thelengthofourdataset),where anonperformingloanisdefinedinthreeways(overduemorethan90daysincolumn1,overduemorethan365days incolumn2,andloandefaultincolumn3). “Lowleveragebank”isadummyvariableequalto1ifbankb’sleverage ratiowasinthebottomquartileofourdatasetin2007(equaltojustbelow7percent).Werestrictattentiontofirmswith multiplebankloansandhaveatleastoneoverdueloanduringoursample. Standarderrorsareclusteredatthefirm level. Intable(2),weexaminewhetherbankswithlowleverageratiosarelesslikelytodesignatealoan as nonperforming relative to other banks that also lend to the same firm. To do so, we restrict attention to firms with active lending relationships from multiple banks and use the following regression: Nonperformingloan = β ×Lowleveragebank +α +(cid:101) (2) b,f 1 b f b,f 7

where “Nonperforming loan ” is a dummy variable for whether the loan from bank b to firm b,f f becomes nonperforming during the length of our dataset (2007:Q2 to 2012:Q4) and α is a firm f fixedeffect. The results in table (2) show that banks with lower capital ratios were relatively less likely to declare a loan as overdue more than 90 days (column 1) or declare a loan as overdue more than 365days(column2)thanotherbanks. However,bankswithlowleverageratiosweresignificantly morelikelytodeclareadefaultedloanthanotherbanks(column3). Loansbylowleveragebanks were 130 basis points more likely to default than loans from other banks to the same firm—a default rate that is more than 20 percent higher than other banks (the mean rate of loan defaults onloanstoborrowerswithmultiplebankswasaround6percent). Theresultsintable(2)areconsistentwiththeexplanationthatbankswithlowleverageratiosare delayingtherecognitionofoverdueloans(negativeandstatisticallysignificantresultsincolumns 1and2)and,inturn,leadingtoworseloanoutcomes(higherratesofloandefaultsincolumn3);in otherwords,banksarewillingtochooseshort-termgainforpotentiallylargerlong-termpain. For instance,bankswithhigherleverageratiosmayhavetakenmoreimmediatecorrectiveactionsto recoveroverdueloansthanbankswithlowleverageratios,causingrelativelyhigherloandefaults for banks with lower leverage ratios. Additionally, since banks with low leverage ratios may be delaying the recognition of loan defaults (as well as overdue loans), we may be underestimating theextentofevergreeningbybankswithlowleverageratios;consequently,theresultincolumn3 wouldbebiaseddownwards. 8

Table2: Differencesinnonperformingloanratesacrossbanksforthesamefirm (1) (2) (3) Loanoverdue(90+) Loanoverdue(365+) Loandefaulted Lowleveragebank -0.038∗∗∗ -0.012∗ 0.013∗∗ (0.0082) (0.0065) (0.0059) Observations 8215 8215 8215 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 2726 2726 2726 FirmFEs Yes Yes Yes Dep.variablemean 0.17 0.10 0.060 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexamineswhethernonperformingloanratesforthesamefirmthatborrowedfrommultiplebanksvaried across banks. “Loan overdue (90+) ” is a dummy variable for whether the loan from bank b to firm f becomes b,f overduemorethan90daysduringthelengthofourdataset(2007:Q2to2012:Q4). Similarly,“Loanoverdue(365+) ” b,f and“Loandefault ”aredefinedforaloanoverduemorethan365daysandaloandefault,respectively.“Lowleverage b,f bank”isadummyvariableequalto1ifbankb’sleverageratiowasinthebottomquartileofourdatasetin2007(equal tojustbelow7percent). Werestrictattentiontofirmswithmultiplebankloans. Standarderrorsareclusteredatthe firmlevel. Taking the results in tables (1) and (2) together, we show that banks with low leverage ratios are bothslowerandlesslikelytodesignatealoanasoverduethanotherbanksthatlendtothesame firm. These results suggest that banks with low leverage ratios are evergreening some of their loans. However, there are other plausible theoretical explanations that are consistent with these results,andinsection(4)weexplorethreealternativeexplanations. First,dofirmsprefertorepay banks with lower leverage ratios more than other banks? Second, do banks with low leverage ratiosmonitortheirloanslessandconsequentlycausehigherloandefaults? Third,dobankswith lowleverageratiosefficientlyforbeartheirloanstofirms? 3.2 Relativeexposureoflowleveragebankstofirmswithoverdueloans In section (3.1), we found suggestive evidence that banks with low leverage ratios mask nonperformingloans. Oneadditionalpredictionfromthebankevergreeningliteratureisthatbanksroll over a firm’s existing debt into new larger performing loans. We investigate this possibility in threeways. First,arebankswithlowleverageratiosmorelikelytokeeplendingtofirmswithan overdueloanatadifferentbank? Second,dobankswithlowleverageratiosincreasetheirrelative share of total lending to a firm with an overdue loan at a different bank? Finally, are banks with 9

lowleverageratiosmorelikelytostartanewbank-firmrelationshipwithafirmthatrecentlyhad anoverdueloanatadifferentbank? Overall,wefindthatbankswithlowleverageratiosdidnot materially increase their exposure to firms that recently had a nonperforming loan at a different bank. Thisbodeswellfortheallocationforcredit,asbanks(onaverage)tendtoreducetheircredit exposuretofinanciallyvulnerablefirms. Tostart,weexaminewhetherbankswithlowleverageratiosweremorelikelytokeeplendingto a firm that has a nonperforming loan at a different bank. To do so, we examine the set of banks that still lend to a firm four quarters after the firm’s first nonperforming loan at a different bank. Specifically,weconductthefollowingregression: Activeloan b,f,t+4 = β 1 ×Lowleveragebank b +α f +α t+4 +(cid:101) b,f,t+4 (3) where“ActiveLoan ”isadummyvariableequaltooneifbank b hasanactiveloantofirm b,f,t+4 f four quarters following the firm’s first nonperforming loan at a different bank. As before, we include a firm fixed effect, α , to ensure we’re only estimating the effect from firms with multif ple lender and we also include a time fixed effect (α t+4 ) to account for any aggregate changes in lendingpatternsovertime. Banks with low leverage ratios were generally as likely as other banks to keep lending to firms with a nonperforming loan, as shown in table (3). Across all the different definitions of nonperforming loan, four quarters following the firm’s first nonperforming loan, low leverage banks were roughly equally likely to stop lending (on average, between 10 and 13 percent of banks stopped lending to firms following a nonperforming loan). The only regression that is different atastatisticallysignificantlevel(andonlyatthe10percentlevel)isthatbankswithlowleverage ratiosseemslightlylesslikelytokeeplendingtofirmsoneyearafteranoverdueloanatadifferent bank. 10

Table 3: Changes in firm’s lending relationships following a nonperforming loan at a different bank (1) (2) (3) ActiveLoan ActiveLoan ActiveLoan Lowleveragebank -0.013 -0.031∗ 0.025 (0.014) (0.017) (0.020) Observations 2717 1501 1184 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 738 368 280 Eventtype Overdue90+ Overdue365+ Default FirmFEs Yes Yes Yes TimeFEs Yes Yes Yes Dep. variablemean 0.87 0.88 0.90 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 This table examines whether a firm was relatively more likely to continue a lending relationship at a bank with a lowleverageratiorelativetootherbankswithinfourquartersofanonperformingloanatadifferentbank. “Active loan b,f,t+4 ”isadummyvariableequaltooneifbankbhasanactiveloantofirm f fourquartersfollowingthefirm’s firstnonperformingloanatadifferentbank. “Lowleveragebank”isadummyvariableequalto1ifbankb’sleverage ratiowasinthebottomquartileofourdatasetin2007(equaltojustbelow7percent).Thisregressionrestrictsattention toonlyfirmswithmultiplebankloansandonlyobservationsfourquartersafterthefirm’sfirstnonperformingloanata differentbank,whereanonperformingloanisdefinedinthreeways(overduemorethan90daysincolumn1,overdue more than 365 days in column 2, and loan default in column 3). To ensure we can observe changes four quarters afterthefirstnonperformingloan,werestrictourobservationstothosefirmsthathadtheirfirstnonperformingloan between2007:Q2and2011:Q4.Standarderrorsareclusteredatthefirmlevel. Second,weexaminewhethertherelativeshareoflendingbylowcapitalbanksincreasedtofirms withanonperformingloan. Toexaminethisquestionwecreateanewvariable,“ChangeinDebt share ”,whichisdefinedasthefollowing: b,f,t Changeindebtshare = Debtshare −Debtshare (4) b,f,t b,f,t b,f,t−4 where“Debtshare ”isdefinedas3: b,f,t 3Note that the summation of a ”debt share” across all banks for a particular firm will not necessarily add to one becausesomefirmsmayhavezerototaldebtattimet,inthatcase,wehavedefinedthebank’s“debtshare”tothat firmtobezero. 11

Debtoffirm f tobankbattimet  ifTotaldebtoffirm f attimet > 0 Debtshare = Totaldebtoffirm f attimet (5) b,f,t 0 ifTotaldebtoffirm f attimet = 0 Therefore,“Changeindebtshare ”,measuresthechangeinbankb’sshareoflendingtofirm f at b,f,t timetrelativetothebank’sshareoflendingfourquarterspreviously. Inourregression,werestrict attention to the set of banks that were lending at the time of the firm’s first nonperforming loan and examine whether banks with low leverage ratios relatively increased their share of lending fourquarterslater. Thespecificregressionwerunis(notethatsincewearearegressingachange inafirm’sdebtfromaspecificbankovertime,theinclusionofatimefixedeffectisnotneededfor thisregression): Changeindebtshare = β ×Lowleveragebank +α +(cid:101) (6) b,f,t 1 b f b,f,t Table(4)showsthatbankswithlowleverageratiosdidnotincreasetheirshareoflendingfollowing a firm’s first nonperforming loan, suggesting that banks with low leverage ratios did not roll overthefirm’sdebtsintonewloansrelativelymorethanotherbanks. 12

Table 4: Changes in the bank’s share of a firm’s total credit following a nonperforming loan at a differentbank (1) (2) (3) Changeindebtshare Changeindebtshare Changeindebtshare Lowleveragebank -0.0026 -0.0022 -0.0029 (0.0064) (0.0070) (0.0057) Observations 2717 1501 1184 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 738 368 280 Eventtype Overdue90+ Overdue365+ Default FirmFEs Yes Yes Yes Dep.variablemean -0.011 -0.012 -0.0039 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexamineswhethertherelativeshareoflendingbylowcapitalbanksincreasedtofirmswithanonperforming loanatadifferentbank. “Changeindebtshare ”measuresthechangeinbankb’sshareoflendingtofirm f attime b,f,t trelativetothebank’sshareoflendingfourquarterspreviouslyandisformallydefinedinequations(4)and(5).“Low leveragebank”isadummyvariableequalto1ifbankb’sleverageratiowasinthebottomquartileofourdatasetin 2007(equaltojustbelow7percent).Thisregressionrestrictsattentiontoonlyfirmswithmultiplebankloansandonly observationsfourquartersafterthefirm’sfirstnonperformingloanatadifferentbank,whereanonperformingloan isdefinedinthreeways(overduemorethan90daysincolumn1,overduemorethan365daysincolumn2,andloan defaultincolumn3). Toensurewecanobservechangesfourquartersafterthefirstnonperformingloan,werestrict ourobservationstothosefirmsthathadtheirfirstnonperformingloanbetween2007:Q2and2011:Q4.Standarderrors areclusteredatthefirmlevel. Finally, we examine whether banks with low leverage ratios were relatively more likely to start new lending relationships with firms with an overdue loan at a different bank. Firms that are overdue on their loans may try to repay their loans by taking new loans at different banks. To testthispossibility,weexaminewhetherbankswithlowleverageratiosstartrelativelymorenew lendingrelationshipswithafirmwithanonperformingloan.4 Werunthefollowingregression: Newloan = β ×Lowleveragebank +α +α +(cid:101) (7) b,f,t+4 1 b f t b,f,t+4 where “New Loan ” is a dummy variable equal to one if bank b started a new banking relab,f,t+4 4Forthisregressionwecreateadummyvariable(“newloan”)forallnewpossiblebank-firmrelationships;hence, thenumberofobservationsintable(5)aresignificantlylargerthanforallotherregressions. Moreover,toincreasethe powerofourtestsandbecausewearecomparingwhetherlowleveragebanksandotherbanksweremorelikelyto startnewlendingrelationshipstothesamefirm,weincludebothfirmswithonlyonelenderandfirmswithmultiple lenders(whereasintables(1)to(4),werestrictedattentiontoonlythosefirmswithmultiplelenders). 13

tionshipwithfirm f withinfourquartersofthefirm’sfirstnonperformingloaneventatadifferent bank,andα andα arefirmanddatefixedeffectsrespectively. f t Theresultsintable(5)showweakevidencethatbankswithlowleverageratiosweremorelikely tostartnewlendingrelationshipswithfirmsthatrecentlyhadanonperformingloan. Bankswith lowleverageratioswere13basispointsmorelikelytostartanewbankingrelationshipwithafirm thathadanoverdueloanmorethan90daysthanotherbanks(column1),buttherewasnosizable or statistically significant effect for loans overdue more than 365 days or loan defaults (columns 2 and 3). Moreover, even though banks with low leverage ratios were more likely to start a new relationship with a firm with an overdue loan at another bank, the effect is economically very small. Forinstance,wefoundthat12percentofbanks,oneyearafterthefirm’sfirstoverdueloan, stoppedlendingtothatfirm(themeanof“activeloan”intable(3)column2). 14

Table 5: Likelihood of forming a new bank lending relationship following an overdue loan at a differentbank (1) (2) (3) NewLoan NewLoan NewLoan Lowleveragebank 0.0013∗∗∗ 0.00038 0.00029 (0.00025) (0.00026) (0.00045) Constant 0.0031∗∗∗ 0.0026∗∗∗ 0.0033∗∗∗ (0.00011) (0.00012) (0.00020) Observations 248535 167404 75746 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 8966 6040 2765 Eventtype Overdue90+ Overdue365+ Default FirmFEs Yes Yes Yes TimeFEs Yes Yes Yes Dep.variablemean 0.0037 0.0027 0.0034 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexamineswhetherbankswithlowleverageratioswererelativelymorelikelytostartnewlendingrelationshipswithfirmswithanoverdueloanatadifferentbank.“NewLoan b,f,t+4 ”isadummyvariableequaltooneifbank bstartedanewbankingrelationshipwithfirm f withinfourquartersofthefirm’sfirstnonperformingloaneventata differentbank,whereanonperformingloanisdefinedinthreeways(overduemorethan90daysincolumn1,overdue morethan365daysincolumn2,andloandefaultincolumn3).“Lowleveragebank”isadummyvariableequalto1if bankb’sleverageratiowasinthebottomquartileofourdatasetin2007(equaltojustbelow7percent).Forthisregressionwecreateadummyvariable(“newloan”)forallnewpossiblebank-firmrelationshipsandincludebothfirmswith onlyonelenderandfirmswithmultiplelenders;hencethenumberofobservationsinthistablearesignificantlylarger thanforallotherregressions. Toensurewecanobservechangesfourquartersafterthefirstnonperformingloan,we restrictourobservationstothosefirmsthathadtheirfirstnonperformingloanbetween2007:Q2and2011:Q4.Standard errorsareclusteredatthefirmlevel. 3.3 Dynamiceffectsonfirm-bankrelationshipsfollowinganonperformingloan Insections(3.1)and(3.2), weestablishedthatbankswithlowleverageratiosweremorelikelyto delaytherecognitionofnonperformingloansbutdidnotincreasetheiroverallexposuretofirms with nonperforming loans. In this section, we examine the time-path of firm credit characteristics (total loan size outstanding and the number of lenders) before and after a firm’s first loan is classified as nonperforming at any bank. To increase the power of our results, and since we are interestedinhowfirmsrespond,weincludeallfirmsinthecreditregistrywithanonperforming loan during our sample period (that is, in contrast to tables (1) to (4), we also include firms that borrowfromonlyasinglelender). Thissection’smainresultisthatafirm’stotaldebtandafirm’s 15

numberoflendersdramaticallyfallsfollowingthefirm’sfirstnonperformingloanatanybank. To examine how a firm’s total debt changes over time, we define a new normalized variable, ”IndexedDebt ”, whichmeasuresfirm f’stotaldebtinquarter t relativetothefirm’stotaldebt f,t inthequarterinwhichthefirm’sfirstloanbecomesnonperforming.5 Specifically, Firm f’stotalfirmdebtattimet IndexedDebt = (8) f,t Firm f’stotalfirmdebtatoccurrenceoffirstnonperformingloan Furthermore, since we define a nonperforming loan in three ways (overdue more than 90 days, overduemorethan365days,orloandefault),weanalyzehowindexeddebtchangesinresponse toeachofthesethreedifferentevents. Thespecificregressionwerunis6: IndexedDebt = β ×Quartersuntilfirstloannonperforming f,t B f,t (9) +β ×Quarterssincefirstloannonperforming +α +(cid:101) A f,t f f,t where α is a firm fixed effect. In this regression, we estimate how the firm’s path of total debt f changesintheeightquartersbefore(β )andeightquartersafter(β )thefirm’sfirstloanbecomes B A nonperforming.7 The results are presented in table (6). Column 1 assesses how the debt changes with the first occurrence of a loan being overdue more than 90 days; similarly, columns 2 and 3, assess how the debt changes but for the first occurrence of a loan being overdue more than 365 daysandaloandefaulting,respectively. Table (6) shows two main results. First, a firm’s total debt was relatively steady before the first occurrence of a loan being overdue more than 90 days (statistically and economically insignifi- 5Weusethevariableindexeddebtbecausewewanttobothmeasurerelativechangesincredit(therefore,notusing absolutevalues)andbeabletoaccountforthefirms’totaldebtbeingzero(thereby,excludingtheuseoftakinglogsof firmdebt).Sincethefirmmusthavesomedebtatthepointatwhichtheloanbecomesoverdue,usingthatloanamount seemsanappropriatedenominatorfortheindex. 6“Quartersuntilfirstloannonperforming”isthenumberofquartersbeforethefirm’sfirstloanbecomesoverdue andiszeroforthequartersafterthefirstoccurrenceofanonperformingloan.Symmetrically,“Quarterssincefirstloan nonperforming”isthenumberofquartersafterthefirm’sfirstloanbecomesoverdueandiszeroforthequartersbefore thefirstoccurrenceofanonperformingloan. 7Toensurewehaveabalancedpanelweincludeonlythosefirmsforwhichtheirfirstnonperformingloanoccurred between2009:Q1and2010:Q4. Thisrestrictionensuresthatwehavebothobservationsforthefirms’totaldebtforthe eightquartersbeforeandafterthefirstloanbecomesnonperforming. 16

cant coefficient on the variable “quarters until first overdue loan more than 90 days”). Second, followingaloangoingoverduemorethan90days,thefirms’totaldebtstartedtodramaticallyfall (negative coefficient on the variable “quarters since first overdue loan more than 90 days”). For thefirmsthatsubsequentlyhaveeitheraloanoverduemorethan365daysoraloandefault, this reductionincreditcontinueswithafirm’stotaldebtfallingbothbeforealoangoesoverduemore than 365 days and before the first loan default (as observed by the large and statistically significantcoefficientson“quartersuntilfirstoverdueloanmorethan365”and“quartersuntilfirstloan default”). The key inference from the results in table (6) is that banks do seem to take prudent actions following a firm’s loans going overdue with banks subsequently significantly reducing lending to the firm—by just over 5 percent per quarter. A key financial stability concern would be if banks systemically continued to increase lending to firms that were in financial stress, which is not the casehere. 17

Table6: Changesinafirm’sdebtbeforeandafterthefirm’sfirstnonperformingloan (1) (2) (3) IndexedDebt IndexedDebt IndexedDebt Qtrs.untilfirstoverdueloan(90+) 0.0055 (0.0041) Qtrs.sincefirstoverdueloan(90+) -0.051∗∗∗ (0.0029) Qtrs.untilfirstoverdueloan(365+) 0.038∗∗∗ (0.0046) Qtrs.sincefirstoverdueloan(365+) -0.031∗∗∗ (0.0027) Qtrs.untilfirstloandefault 0.012∗∗ (0.0049) Qtrs.sincefirstloandefault -0.021∗∗∗ (0.0031) Observations 45334 33146 15138 Observationlevel Firm-bank Firm-bank Firm-quarter Numberoffirms 3267 2385 1085 FirmFEs Yes Yes Yes Dep.variablemean 0.89 0.95 0.96 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 “Quartersuntilfirstloannonperforming”isthenumberofquartersbeforethefirm’sfirstloanbecomesoverdueand zeroforthequartersafterthefirstoccurrenceofanonperformingloan. Symmetrically,“Quarterssincefirstloannonperforming”thenumberofquartersafterthefirm’sfirstloanbecomesoverdueandiszeroforthequartersbeforethe firstoccurrenceofanonperformingloan. Toensurewehaveabalancedpanelweincludeonlythosefirmsforwhich theirfirstnonperformingloanoccurredbetween2009and2011. Thisrestrictionensuresthatwehavebothobservationsforthefirms’totaldebtfortheeightquartersbeforeandafterthefirstloanbecomesnonperforming. Anegative coefficientfor”Quartersuntil...”impliesthattotaldebtwasrisingbeforethefirstloanwasdenotedasoverdue.Anegativecoefficientfor”Quarterssince...”impliesthatthetotaldebtfellfollowingthefirstloanwasdenotedasoverdue. Standarderrorsareclusteredatthefirmlevel. In addition to examining how total debt responded to changes in a firm’s nonperforming loans, we can observe how the number of bank relationships changes before and after a loan becomes nonperforming. Todoso,werunregressionsthataresimilartothoseinequation(9),specifically: 18

Numberofbankrelationships = β ×Quartersuntilfirstloannonperforming f,t B f,t +β ×Quarterssincefirstloannonperforming (10) A f,t +α +(cid:101) f f,t where“Numberofbankrelationships ”isthenumberofactivelendingrelationshipsforfirm f f,t at time t, and α are firm fixed effects.8 “Quarters until first loan nonperforming” and “Quarters f sincefirstloannonperforming”aredefinedasinequation(9). In this regression, we estimate how the number of lenders to a firm changes in the eight quarters before and eight quarters after the firm’s first loan becomes nonperforming.9 The results are presentedintable(7). Table (7) shows two main results. First, the number of lenders was increasing before the first occurrence of a loan being overdue more than 90 days, by over 0.05 lenders per quarter. Second, followingaloangoingoverduemorethan90days,thenumberoflendersstartedtodramatically fall,byabout0.04lendersperquarter. The results presented in tables (6) and (7) together show that banks, on aggregate, significantly reducedexposuretofirmsfollowingtheirfirstnonperformingloan. Moreover,thedesignationof anonperformingloanseemstohavelargerealeffectsonthefirm’scapacitytoborrow,withasharp changeinthefirms’creditgrowthandnumberoflendingpartnersfollowingthedesignation. 8Theuseofafirmfixedeffectensuresthatwecontrolfortheaveragenumberofbankrelationshipsafirmhasover theperiod. 9Similartotable(6),weincludeonlythosefirmsforwhichtheirfirstnonperformingloanoccurredbetween2009 and2011. 19

Table 7: Changes in a firm’s total number of bank relationships before and after the firm’s first nonperformingloan (1) (2) (3) Numberofbankrel. Numberofbankrel. Numberofbankrel. Qtrs.untilfirstoverdueloan(90+) -0.053∗∗∗ (0.0028) Qtrs.sincefirstoverdueloan(90+) -0.043∗∗∗ (0.0023) Qtrs.untilfirstoverdueloan(365+) -0.017∗∗∗ (0.0028) Qtrs.sincefirstoverdueloan(365+) -0.038∗∗∗ (0.0026) Qtrs.untilfirstloandefault -0.0070 (0.0064) Qtrs.sincefirstloandefault -0.039∗∗∗ (0.0045) Observations 46564 34244 15778 Observationlevel Firm-quarter Firm-quarter Firm-quarter Numberoffirms 3326 2446 1127 FirmFEs Yes Yes Yes Dep.variablemean 1.37 1.50 2.31 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 “Quartersuntilfirstloannonperforming”isthenumberofquartersbeforethefirm’sfirstloanbecomesoverdueand is zero for the quarters after the first occurrence of a nonperforming loan. Symmetrically, “Quarters since first loan nonperforming” is the number of quarters after the firm’s first loan becomes overdue and is zero for the quarters beforethefirstoccurrenceofanonperformingloan. Toensurewehaveabalancedpanelweincludeonlythosefirms forwhichtheirfirstnonperformingloanoccurredbetween2009and2011. Thisrestrictionensuresthatwehaveboth observations for the firm’s total debt for the eight quarters before and after the first loan becomes nonperforming. Anegativecoefficientfor”Quartersuntil...”impliesthatthenumberofbank-firmrelationshipswasrisingbeforethe firstloanwasdenotedasoverdue. Anegativecoefficientfor”Quarterssince...”impliesthatthenumberofbank-firm relationshipsfellfollowingthefirstloanwasdenotedasoverdue.Standarderrorsareclusteredatthefirmlevel. 4 Alternative potential explanations Thissectionexaminesthreeotherpossibletheoriesthatwouldbeconsistentwiththeresults. First, do firms prefer to repay banks with lower leverage ratios more than other banks? Second, do banks with low leverage ratios monitor their loans less and consequently have higher loan de- 20

faults? Third,dobankswithlowleverageratiosefficientlyforbeartheirloanstofirms? Wedonot findstrongevidencetosupportanyofthesealternativeexplanations. A key possibility is that firms may value their relationships with banks with low leverage ratios more or receive more favorable loan terms from these banks. In turn, this may cause firms to strategically repay other banks first. We explore this possibility in two ways. First, we examine whether the results in section (3.1) are robust to the inclusion of loan-level controls. Second, we examinewhetherbankswithlowleverageratiosoffermorefavorableloanterms. Table(8)showsthatfirmsaremorelikelytogooverduefirstonlargerandunsecuredloans. Similarly, table (9) shows that firms are more likely to be overdue on larger and unsecured loans. However,theresultsinbothtablesshowthatbankswithlowleverageratios—evenaftercontrollingforloanlevelterms—areslowertodesignatealoanasoverdueandlesslikelytodesignatethe loan as overdue. These results suggest that differences in loan terms are not the principal cause forrelativelyloweroverdueratesforbankswithlowleverageratios. Table 8: Differences in the first lender to designate as a loan as nonperforming, including loanlevelcontrols (1) (2) (3) Firstoverduebank(90+) Firstoverduebank(365+) Firstdefaultbank Lowleveragebank -0.066∗∗∗ -0.082∗∗∗ 0.015 (0.021) (0.027) (0.028) Ln.bankloan 0.027∗∗∗ 0.036∗∗∗ 0.048∗∗∗ (0.0049) (0.0071) (0.0078) Unsecuredloan 0.0080 0.0071 0.019∗ (0.012) (0.024) (0.010) Observations 5377 3041 2286 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 2132 1107 721 FirmFEs Yes Yes Yes Dep.variablemean 0.28 0.29 0.27 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 “Firstoverduebank(NPL) ”isadummyvariableequalto1ifbankbwasthefirstbanktodesignatealoanfromfirm b,f f asnonperformingintheperiod2007:Q2to2012:Q4(thelengthofourdataset),whereanonperformingloanisdefined inthreeways(overduemorethan90daysincolumn1,overduemorethan365daysincolumn2,andloandefaultin column3). “Lowleveragebank”isadummyvariableequalto1ifbankb’sleverageratiowasinthebottomquartile ofourdatasetin2007(equaltojustbelow7percent). Loan-levelcontrolsaredefinedusingloanvaluesandsecured statusasof2007:Q1. Werestrictattentiontofirmsthathavemultiplebankloansandatleastoneoverdueloanduring oursample.Standarderrorsareclusteredatthefirmlevel. 21

Table9: Differencesinoverdueratesforthesamefirmacrossdifferentbanks,includingloan-level controls (1) (2) (3) Loanoverdue(90+) Loanoverdue(365+) Loandefaulted Lowleveragebank -0.050∗∗∗ -0.018∗∗∗ 0.011∗ (0.0086) (0.0067) (0.0062) Ln.bankloan 0.025∗∗∗ 0.018∗∗∗ 0.012∗∗∗ (0.0021) (0.0017) (0.0013) Unsecuredloan 0.0050 0.0077 0.00029 (0.0063) (0.0057) (0.0054) Observations 7552 7552 7552 Numberoffirms 2516 2516 2516 FirmFEs Yes Yes Yes Dep.variablemean 0.18 0.11 0.062 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 “Loanoverdue(90+) ”isadummyvariableforwhethertheloanfrombankbtofirm f becomesoverduemorethan90 b,f daysduringthelengthofourdataset(2007:Q2to2012:Q4). Similarly,“Loanoverdue(365+) ”and“Loandefault ” b,f b,f aredefinedforaloanoverduemorethan365daysandaloandefault,respectively. “Lowleveragebank”isadummy variableequalto1ifbank b’sleverageratiowasinthebottomquartileofourdatasetin2007(equaltojustbelow7 percent). Loan-levelcontrolsaredefinedusingloanvaluesandsecuredstatusasof2007:Q1. Werestrictattentionto firmsthathavemultiplebankloans.Standarderrorsareclusteredatthefirmlevel. In table (10), we explore differences in initial loan terms for firms with multiple loans between bankswithlowleverageratios andotherbanks. Specifically, we runthefollowingregressionfor firmswithmultipleloansinthefinalquarterofourdataset(2012:Q4):10 Outcomeofinterest = β ×Lowleveragebank +α +(cid:101) (11) b,f 1 b f b,f wheretheOutcomeofinterest arethefirm’sloansizewithbankb,thelengthofthefirm’srelab,f tionshipwithbankb,theinterestrateofthefirm’sloanswithbankb(weightedoverallthefirm’s loanswithbankb),andthemonthstomaturityforthefirm’sloanswithbankb(weightedoverall thefirm’sloanswithbankb).11 10Weusethefinalquarterofourdatasetsothatwecanexaminethelengthofthefirm’slendingrelationshipwith thatbank. 11Dataforloanmaturityandloaninterestratesaremissingforsomefirms;therefore,therearefewerobservationsin columns3and4thancolumns1and2. 22

The evidence is mixed as to whether banks with low leverage ratios offer more favorable ratios, whichispresentedintable(10). Onekeymeasure,interestrates(column3),showsthatbankswith low leverage ratios charged higher rates. Also, two key measures of the lending relationship are relativelysimilaracrossthebanks—namely,loansize(column1)andmonthstomaturity(column 4). These three results suggest that loan terms from banks with low leverage ratios were not morefavorablethanotherbanks. However,intheoppositedirection,wealsofindthatfirmshad significantly longer lending relationships to banks with lower leverage ratios (column 2). Given the long empirical and theoretical literature on the importance of relationship lending (such as Rajan[1992],PetersenandRajan[1994],BootandThakor[2000],Boot[2000]),thisresultsuggests that firms may receive greater benefits from their long-term relationships with the banks with low leverage ratios. Overall, we can neither rule out that firms received greater benefits, nor conclusively say that firms received less benefits from the lending relationships from banks with lowleverageratios.12 Table10: Comparingloantermsforthesamefirmacrossdifferentbanks (1) (2) (3) (4) Ln.bankloan LengthofRelation.(Qtrs.) InterestRate Monthstomaturity Lowleveragebank 0.055 2.17∗∗∗ 0.46∗∗ 0.28 (0.054) (0.17) (0.20) (0.21) Observations 13087 14228 6344 8256 Observationlevel Firm-bank Firm-bank Firm-bank Firm-bank Numberoffirms 4943 5012 3400 3928 FirmFEs Yes Yes Yes Yes Dep.variablemean 17.1 15.9 12.8 5.36 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexploresdifferencesinloantermsforfirmswithmultipleloansbetweenbankswithlowleverageratiosand other banks. “Ln. bank loan ” is the natural logarithm of firm f’s total loans at bank b. “Length of Relationships b,f (Qtrs.) is the number of quarters that firm f had an active loan with bank b prior to 2012:Q4. “Interest rate is b,f b,f weighted interest rate of firm f’s existing loans with bank b (weighted by the total size of each loan). “Months to maturity istheweightedmonthstomaturityoffirm f’sexistingloanswithbankb(weightedbythetotalsizeofeach b,f loan). We restrict attention to only active loans in the final quarter of our dataset (2012:Q4) so that we can examine thelengthofthefirm’slendingrelationshipwiththatbank. Dataforloanmaturityandloaninterestratesaremissing forsomefirms;therefore,therearefewerobservationsincolumns3and4thancolumns1and2. Standarderrorsare clusteredatthefirmlevel. 12Thereistheadditionalpossibilitythatthefindingthatbankswithlowleverageratioshavealongerlendingrelationshipisanoutcomeofevergreening. Specifically, ifbankswithlowleverageratiosforbeartheirloansmorethan otherbanks,thenthisrelationshipwillbemechanicallylonger. 23

A further theoretical possibility is that banks with low leverage ratios monitored their loans less than other banks, causing both a relatively slower designation of overdue loans, and greater defaultsforthesebanks(HolmstromandTirole[1997],Allenetal.[2011],MehranandThakor[2011]). If low leverage banks monitored their loans less, we would expect larger effects for unsecured loansbecausethesearetheloanswheremonitoringisthemostrelevantsincethebankhaslarger expected loss given default and the borrower has less incentive to repay. We find no evidence to supportthispotentialexplanation. Wetestwhetherbankswithlowleverageratioswererelativelymorelikelytohavegreaterdefault rates on unsecured loans relative to other banks (similar to the test in table (2) but concentrating ondifferencesbetweensecuredandunsecuredloansforbankswithlowleverageratios). In table (11) we present the results of the following regression on the set of firms with multiple lendingrelationships: NonperformingLoan = β ×Lowleveragebank ×UnsecuredLoan b,f 1 b b,f (12) +β ×UnsecuredLoan +β ×Lowleveragebank +α +(cid:101) 2 b,f 3 b f b,f where“UnsecuredLoan ”isadummyvariableequaltooneiffirm f hasanunsecuredloanwith b,f bankb,andallothervariablesaredefinedaspreviously.13 Starting with the results in the third column of table (11), we find that the coefficient on our variable of interest “Lowleveragebank ×UnsecuredLoan ” is both negative and statistically b b,f significant, which is inconsistent with the prediction that banks with low leverage ratios were monitoring loans less. Specifically, if banks with low leverage ratios were monitoring loans less, we would expect this coefficient to be positive—that is, banks with lower leverage ratios would haverelativelyhigherdefaultratesforunsecuredloans—becauseloansthatareunsecuredrequire the most monitoring. We include firm fixed effects and restrict attention to only those borrowers thatborrowfrommultiplebanks;therefore,thisresultisrobusttobankswithlowleverageratios lendingtoadifferentsetoffirmsasotherbanks. Turningtotheresultsinthefirstcolumnoftable(11), wefindthatthecoefficientonourvariable 13This regression is the analogue of the regressions in table (2) but includes both a dummy variable for an unsecuredloanandtheinteractionofourmeasureforlowbankleveragewiththedummyvariableforwhethertheloanis unsecured. 24

of interest “Lowleveragebank ×UnsecuredLoan ” is both positive and weakly statistically b b,f significant. Theinterpretationofthiscoefficientwithrespecttothetheoryoflowermonitoringby less capitalized banks is difficult. For instance, if banks with less capital were monitoring loans less, we could expect this coefficient to be negative, because banks may be slow to recognize the loan as overdue due to their lack of monitoring capacity. Alternatively, we may expect higher ratesofoverdueloansforthesebanksduetothelackofmonitoringofborrowerbehaviorcausing moreloanstobecomeoverdue. Table 11: Differences in nonperforming loan rates for different loan types across banks with differentleverageratios (1) (2) (3) Loanoverdue(90+) Loanoverdue(365+) Loandefaulted LowleveragebankxUnsecured 0.024∗ 0.0073 -0.029∗∗ (0.013) (0.011) (0.012) Lowleveragebank -0.041∗∗∗ -0.012∗ 0.017∗∗∗ (0.0084) (0.0066) (0.0062) Unsecuredloan -0.00017 0.0085 0.020∗∗ (0.0099) (0.0087) (0.0088) Observations 8215 8215 8215 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 2726 2726 2726 FirmFEs Yes Yes Yes Dep.variablemean 0.17 0.10 0.060 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 This table examines whether banks with low leverage monitored loans less by examining whether unsecured loans were more likely to default for banks with low leverage ratios. “Loan overdue (90+) ” is a dummy variable for b,f whethertheloanfrombankbtofirm f becomesoverduemorethan90daysduringthelengthofourdataset(2007:Q2 to2012:Q4). Similarly,“Loanoverdue(365+) ”and“Loandefault ”aredefinedforaloanoverduemorethan365 b,f b,f daysandaloandefault,respectively.‘UnsecuredLoan ’isadummyvariableequaltooneiffirm f hasanunsecured b,f loanwithbankb. “Lowleveragebank”isadummyvariableequalto1ifbankb’sleverageratiowasinthebottom quartileofourdatasetin2007(equaltojustbelow7percent).Standarderrorsareclusteredatthefirmlevel. As a final possible explanation, we analyze if banks with low leverage ratios efficiently forbear theirloans. Thatis,dothesebanksprovidenecessaryliquiditytofirmsthataresolventbutfacing either cash-flow difficulties or overcoming a temporary demand shock (Fukuda and Nakamura [2011])? Ineffect,ratherthanbanksdelayingtherecognitionofproblemloanstoprotecttheirbalancesheet,werebankswithlowleverageratiosusingtheirdiscretiontoimproveloanoutcomes? 25

Wepresenttwopiecesofevidencethatdonotsupportthisview. First, intable (2)in section (3.1), we found that banks withlow leverageratios hadhigherdefault ratesthanloansbyotherbankstothesamefirm(column3). Therefore,thisresultstronglyrefutes thesuggestionthatbankswithlowleverageratioswereinducingbetterloanoutcomes. Second,ifbankswithlowleverageratioswereeffectivelyforbearingstrictlyproductiveloans,one would anticipate that at an overdue loan at a bank with a low leverage ratio would be a strong predictor for a future loan default. This relationship follows because the bank with low leverage ratios would classify a firm’s loans as nonperforming only if the bank believed that firm was insolvent,whichinturn,wouldcausehigherfinaldefaultrates. Toexplorethisidea,weexamine directly whether overdue loans by banks with low leverage ratios are more predictive of future loandefaultsthanloansissuedbyotherbanksthatlendtothesamefirm. Specifically,weexamine theconditionalprobabilityofaloandefaultbetween2007:Q3and2012:Q4onthelikelihoodofthe loanbeingdesignatedasoverduemorethan90daysin2007:Q2. Werunregressionssimilarto: Loandefault = β ×Overdueloan(90+) ×Lowleveragebank b,f,2007:Q3−2012:Q4 B b,f,2007:Q2 b,f (13) +β ×Overdueloan(90+) +α +(cid:101) B b,f,2007:Q2 f b,f,2007:Q3−2012:Q4 where“Loandefault ”isadummyvariableforwhethertheloanfrombank b tofirm f defaults b,f between 2007:Q3 to 2012:Q4 and α is a firm fixed effect. This regression tests whether overdue f loans from banks with low leverage ratios were a better predictor of a future loan default than overdueloansfromotherbanks. Theresultsintable(12)showthatoverdueloansbybankswithlowlevelsofcapital(asmeasured by capital or leverage ratios) were less predictive of a future default by economically significant magnitudes(11percentforbankswithlowcapitalbanksand3.5percentforbankswithlowleverage ratios) for the same firm. These results strongly refute the interpretation that banks with low leverageratiosmayhavebeenefficientlyforbearingproductiveloans. 26

Table12: Predictivepowerofanoverdueloanonafutureloandefaultacrossbankswithrelatively lesscapital (1) (2) (3) Loandefaulted Loandefaulted Loandefaulted Overdueloan(90+) 0.15∗∗∗ 0.22∗∗∗ 0.20∗∗∗ (0.018) (0.034) (0.035) Lowcap.bankxOverdueloan(90+) -0.11∗∗∗ (0.038) Lowcapitalbank -0.0049 (0.0058) Lowlev.bankxOverdueloan(90+) -0.035 (0.039) Lowleveragebank 0.015∗∗ (0.0062) Observations 19332 9965 9965 Observationlevel Firm-bank Firm-bank Firm-bank Numberoffirms 5799 3326 3326 FirmFEs Yes Yes Yes Dep.variablemean 0.11 0.12 0.12 Standarderrorsinparentheses ∗ p<0.10,∗∗ p<0.05,∗∗∗ p<0.01 Thistableexamineswhetherbankswithrelativelylesscapitalwereefficientlyforbearingproductiveloansbyexaminingthepredictivepowerofaloanoverduemorethan90daysonafutureloandefault. “Loandefault ”isadummy b,f variableforwhethertheloanfrombankbtofirm f defaultedintheperiod2007:Q3to2012:Q4.“Overdue(90+) ”isa b,f dummyvariableforwhethertheloanfrombankbtofirm f wasoverduemorethan90daysin2007:Q2.“Lowleverage bank”isadummyvariableequalto1ifbankb’sleverageratiowasinthebottomquartileofourdatasetin2007(equal tojustbelow7percent). “Lowcapitalbank”isadummyvariableequalto1ifbankb’scapitalratiowasinthebottom quartileofourdatasetin2007(equaltojustbelow10percent).Thistablerestrictsattentiontoonlyfirmswithmultiple bankloansin2007:Q2.Standarderrorsareclusteredatthefirmlevel. Taken together, we do not find strong evidence that the alternative explanations can explain our results. Thissupportsourmainexplanationthatbankswithlowleverageratiosweredelayingthe recognitionoftheirnonperformingloanstomitigatethehitontheircapital. 5 Conclusion We study how banks that vary in their capital structure respond to firm distress. We provide evidencethatbankswithlowleverageratiosintentionallydelayclassifyingtheirloansasnonper- 27

forming thereby postponing the regulatory hit to their capital ratios. By masking the status of a loan,banksreducetherequiredloanlossprovisions,consequentlyartificiallymaintaininghigher regulatory capital and leverage ratios. Moreover, we find suggestive evidence that this delay in recognizing bad loans, although improving the banks’ capital position in the short-term, causes worse final loan outcomes, with evidence that the banks with a greater delay in recognizing bad loansalsohadgreaterresultantloandefaults. Somewhat surprisingly, we find that banks with low leverage ratios do not materially increase their exposure to firms that recently had a nonperforming loan at a different bank. This bodes wellfortheallocationforcredit,asbanks(onaverage)tendtoreducetheircreditrisktofinancially vulnerablefirms. Overall, our results contribute to the large literature on zombie lending and highlight the importanceofimprovingthepromptandaccuratedisclosureofbanks’nonperformingloans,especially bybanksthatmaybeundercapitalized. 28

References V. V. Acharya, T. Eisert, C. Eufinger, and C. Hirsch. Whatever it takes: The real effects of unconventionalmonetarypolicy. TheReviewofFinancialStudies,32(9):3366–3411,2019. F.Allen,E.Carletti,andR.Marquez. Creditmarketcompetitionandcapitalregulation. TheReview ofFinancialStudies,24(4):983–1018,2011. M.A.Ari,S.Chen,andM.L.Ratnovski. Thedynamicsofnon-performingloansduringbankingcrises: anewdatabase. 2019. R. Banerjee, B. Hofmann, et al. Corporate zombies: Anatomy and life cycle. Bank for International Settlements,MonetaryandEconomicDepartment,2020. R.N.BanerjeeandB.Hofmann. Theriseofzombiefirms: causesandconsequences. BISQuartely Review,2018. L. Blattner, L. Farinha, and F. Rebelo. When losses turn into loans: the cost of undercapitalized banks. 2019. D.Bonfim,G.Cerqueiro,H.Degryse,andS.Ongena. On-siteinspectingzombielending. 2020. A. W. Boot. Relationship banking: What do we know? Journal of Financial Intermediation, 9(1): 7–25,2000. ISSN1042-9573. doi: https://doi.org/10.1006/jfin.2000.0282. URLhttps://www. sciencedirect.com/science/article/pii/S1042957300902821. A.W.BootandA.V.Thakor.Canrelationshipbankingsurvivecompetition? ThejournalofFinance, 55(2):679–713,2000. M.BrucheandG.Llobet.Preventingzombielending.TheReviewofFinancialStudies,27(3):923–956, 2014. M. Brunnermeier and A. Krishnamurthy. Corporate debt overhang and credit policy. In BPEA conference,2020a. M. Brunnermeier and A. Krishnamurthy. Covid-19 sme evergreening proposal: Inverted economics. 2020b. 29

R. M. Bushman and C. D. Williams. Delayed expected loss recognition and the risk profile of banks. JournalofAccountingResearch,53(3):511–553,2015. R. J. Caballero, T. Hoshi, and A. K. Kashyap. Zombie lending and depressed restructuring in japan. AmericanEconomicReview,98(5):1943–77,2008. M. A. Choudhary and A. K. Jain. How public information affects asymmetrically informed lenders: Evidencefromacreditregistryreform. JournalofDevelopmentEconomics, page102407, 2019. M. A. Choudhary and N. Limodio. Deposit volatility, liquidity and long-term investment: Evidencefromanaturalexperimentinpakistan. 2017. Financial Stability Board. Covid-19 pandemic: Financial stability implications and policy measurestaken. 2020. S.-i. Fukuda and J.-i. Nakamura. Why did zombie firms recover in japan? The world economy, 34 (7):1124–1137,2011. M. Giannetti and A. Simonov. On the real effects of bank bailouts: Micro evidence from japan. AmericanEconomicJournal: Macroeconomics,5(1):135–67,2013. A.Hertzberg,J.M.Liberti,andD.Paravisini.Informationandincentivesinsidethefirm: Evidence fromloanofficerrotation. TheJournalofFinance,65(3):795–828,2010. B. Holmstrom and J. Tirole. Financial intermediation, loanable funds, and the real sector. the QuarterlyJournalofeconomics,112(3):663–691,1997. Y.HuandF.Varas. Atheoryofzombielending. 2020. E.J.Kane. TheS&Linsurancemess: Howdidithappen? TheUrbanInsitute,1989. A. I. Khwaja and A. Mian. Do lenders favor politically connected firms? rent provision in an emergingfinancialmarket. TheQuarterlyJournalofEconomics,120(4):1371–1411,2005. A.I.KhwajaandA.Mian.Tracingtheimpactofbankliquidityshocks: Evidencefromanemerging market. TheAmericanEconomicReview,pages1413–1442,2008. 30

H. U. Kwon, F. Narita, and M. Narita. Resource reallocation and zombie lending in japan in the 1990s. ReviewofEconomicDynamics,18(4):709–732,2015. L.LaevenandF.Valencia. Systemicbankingcrisesdatabase. IMFEconomicReview,61(2):225–270, 2013. M. L. Laeven and M. F. Valencia. Systemic banking crises revisited. International Monetary Fund, 2018. M.A.McGowan,D.Andrews,andV.Millot. Thewalkingdead? 2017. H. Mehran and A. Thakor. Bank capital and value in the cross-section. The Review of Financial Studies,24(4):1019–1067,2011. J. Peek and E. S. Rosengren. Unnatural selection: Perverse incentives and the misallocation of creditinjapan. AmericanEconomicReview,95(4):1144–1166,2005. M. A. Petersen and R. G. Rajan. The benefits of lending relationships: Evidence from small businessdata. TheJournalofFinance,49(1):3–37,1994. R. G. Rajan. Insiders and outsiders: The choice between informed and arm’s-length debt. The Journaloffinance,47(4):1367–1400,1992. R.G.Rajan. Whybankcreditpoliciesfluctuate: Atheoryandsomeevidence. theQuarterlyJournal ofeconomics,109(2):399–441,1994. F. Schivardi, E. Sette, and G. Tabellini. Credit misallocation during the european financial crisis. Bancad’Italiaworkingpaper,2017. F.Schivardi,E.Sette,andG.Tabellini. Identifyingtherealeffectsofzombielending. TheReviewof CorporateFinanceStudies,9(3):569–592,2020. M.Storz,M.Koetter,R.Setzer,andA.Westphal. Dowewantthesetwototango? onzombiefirms andstressedbanksineurope. 2017. P.Tantri. Identifyingever-greening: Evidenceusingloan-leveldata. JournalofBanking&Finance, 122:105997,2021. B.Tracey. Therealeffectsofzombielendingineurope. BankofEnglandWorkingPaper,2019. 31