International Evidence on Government Support and Risk Taking in the Banking Sector

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1086 August 2013 International evidence on government support and risk taking in the banking sector Luis Brandao-Marques Ricardo Correa Horacio Sapriza NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors. 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.

International evidence on government support and risk taking in the banking sector∗ Luis Brandao-Marques, Ricardo Correa, and Horacio Sapriza† August 27, 2013 Abstract Government support to banks through the provision of explicit or implicit guarantees affects the willingness of banks to take on risk by reducing market discipline or by increasing chartervalue. Weuseaninternationalsampleofratedbanksandfindthatgovernmentsupport isassociatedwithmorerisktakingbybanks,especiallypriorandduringthe2008-2009financialcrisis. Wealsofindthatrestrictingbanks’rangeofactivitiesamelioratesthelinkbetween government support and bank risk taking. We conclude that strengthening market discipline byreducingbankcomplexityisneededtoaddressthismoralhazardproblem. JELclassifications: G21,G28,H81 Keywords: Bankrisk,marketdiscipline,governmentsupport,bankregulation ∗WethankAllenBerger,MichaelKoetter,participantsattheAEA2013AnnualMeeting,the4thFinancialStabilityConferenceatTilburg University,theFDIC’s12thAnnualBankResearchConference,andattheIMF-ICDandFederalReserveBoardworkshops. WealsothankHal BasemanandOliviaKimforexcellentresearchassistance. †Brandao-Marques (lmarques@imf.org) is from the International Monetary Fund and Correa (ricardo.correa@frb.gov) and Sapriza (horacio.sapriza@frb.gov)arefromtheBoardofGovernorsoftheFederalReserveSystem. Theviewsexpressedhereinarethoseoftheauthorsand shouldnotattributedtotheInternationalMonetaryFund,itsExecutiveBoard,oritsmanagementorinterpretedasreflectingtheviewsoftheBoard ofGovernorsoftheFederalReserveSystemorofanyotherpersonassociatedwiththeFederalReserveSystem.

1 Introduction Bank bailouts during and after the 2007–2009 financial crisis have reignited the debate on the effect of government support on banks’ management incentives and on the distortions it causes in competition in the banking sector. Explicit and implicit government support can influence banks’ willingnesstotakeonriskthroughtwochannels: byreducingmarketdisciplineand/orbyincreasingthebanks’chartervalue. According to the market discipline hypothesis, government support of banks decreases the incentiveofoutsideinvestors(depositors,creditors,andshareholders)tomonitororinfluencebank risk taking. Risk-shifting may occur if deposit insurance is not fairly priced (Merton, 1977) or if governments provide guarantees to holders of bank debt (Flannery and Sorescu, 1996). Under the charter value hypothesis, government support decreases banks’ funding costs as both depositors and creditors demand lower rates. The decline in funding costs increases their interest margin and raisesbanks’chartervalues,whichleadstobankstakingfewerriskstoprotectfuturerents(Keeley, 1990). Thegoalofthispaperistodeterminewhichchanneldominates. Since,forthemostpart,thisis anempiricalissue,weusetwocross-countrysamplesofbankstoanswertwoquestions: Dobanks with more explicit or implicit government support take on more risk? Does bank regulation limit theeffectofgovernmentsupportonbankrisktaking? To answer these questions we define bank risk as the z-score (return on assets plus capital to assetratio,dividedbythestandarddeviationofreturnonassets)whichisameasureofdistanceto default.1 In addition, we measure government support as the difference between each bank’s deposit rating and bank financial strength rating assigned by Moody’s Investors Service (Moody’s), 1Our results are robust to measuring risk using a market based z-score (Forssbaeck, 2011), a measure of stock returnvolatility,andamoretraditionalmeasureofloanlosses-theloanlossprovisionstoassetsratio. 1

which, as we show later in this study, is able to predict actual bank bailouts.2 We test these hypothesis using a sample of bank and ratings data covering more than 50 developed and emerging countriesfortheperiods2003-2004and2009-2010. Our choice of cross-sectional estimations rather than a full panel analysis is explained by the change in accounting standards in Europe in the mid-2000s. As noted later, it would be a mistake to conduct any empirical tests without taking into account this break-in-series. We chose our firstcross-section(2003-2004)asabenchmarktocompareourresultstopreviousstudiesthathave testedsimilarhypothesesusingdataforthatperiod(e.g.Groppetal.,2011,andLaevenandLevine, 2009). The second cross-section is intended to capture the effect of government support on risk takingleadingtotheglobalfinancialcrisisof2008-2009. ThesampleoffinancialinstitutionsinourtestsiscomposedofbanksthatareratedbyMoodys or Fitch Ratings. We use this specific sample to assess the potential distortions of the governments support of banks on institutions that are large enough to issue debt in capital markets.3 Banks that do not issue this type of debt are mostly subject to distortions due to deposit insurance mispricing, which has been the subject of other studies (Demirguc-Kunt and Detragiache, 2002). Each of the two rating agencies that we consider in our analysis currently assigns ratings to about 1,000 banks globally. In these samples of banks, subsidiaries of other rated banks account for a significant share. We exclude these subsidiaries as their risk-taking should be captured in the consolidated statements of their parents. Moreover, since the ratings-based measure of support does not discriminate betweeen government support and parent support, including subsidiaries would 2Ratings-basedmeasuresofsupporthavebeenusedtoassesstheimplicitbenefitofgovernmentsupportonbank debt(SchichandLindh,2012)andequityreturns(Correaetal.,2013). 3The universe of rated banks represents a fraction of all banks in the world. For example, as of 2012, Moodys rated59U.S.bankinggroupsincludingsubsidiariesofforeignbanks(Moody’sInvestorsService,2012). Inthatsame year,theFederalDepositInsuranceCorporationreportedauniverseofabout7,000insuredbanksintheUnitedStates. However,thesampleofratedbanksrepresentedabout80percentofU.S.bankingassetsinthatsameyear. SeeTarullo (2013). 2

introducemeasurementerror. Our paper has two main contributions. First, it shows that, according to a sample of roughly 340 banks from about 50 countries, higher expected government support is associated with more risk taking. The intensity of government support is positively related to our measures of bank risk taking after controlling for several factors, including bank size and liquidity, the level of bank regulations, banks’ownership structure, the degree of market concentration in the banking sector, and country-specific macroeconomic conditions. We find that this relationship is stronger during a crisis period such as the recent global financial crisis. This result is also robust to several other checks, including the possible endogeneity of government support. Similarly, we run panel regressions using bank level fixed effects with data between 2005 and 2010. These estimations use market-based measures as the proxies for banks’ risk taking. We find that the link between governmentsupportandrisk-takingisevenstrongerinthissetofresults.4 Thus,inoursample,market discipline is the dominant factor shaping the relationship between support and risk in the banking industry. Our second key result is that the adoption of regulatory impediments for banks to engage in activities involving security markets, insurance, real estate, and ownership of non-financial firms reduces the magnitude of the moral hazard problem associated with government support. Capital supervision and regulation were not enough to fully prevent additional risk taking by banks with more government support during the crisis, but banks that faced more restrictions in terms of the activities that they were allowed to perform were less likely to take on more risk. Interestingly, Barthetal.(2006)findthat,fortheearly2000s,increasingactivityrestrictionstobanksledtomore risk taking by these institutions. However, our finding is consistent with new studies documenting the increasing complexity of banking organizations, which have not translated into significant 4Thefixedeffectregressionscontrolforpotentialunobservablebankcharacteristicsthatarestaticintheshortrun andthatmayaffectrisktaking(e.g.,amanagersappetiteforriskorthecompensationstructurewithinbanks). 3

economiesofscope(Cetorellietal.,2012). Previous studies on the impact of government support on bank risk taking have to a large extent looked at either measures of explicit support such as deposit insurance (Demirguc-Kunt and Detragiache, 2002) and state ownership (De Nicolo´ and Loukoianova, 2007) or indirect measures of implicit support such as bank size (“too-big-to-fail”; see Boyd and Runkle, 1993, and O’Hara andShaw,1990),withmixedresults. Morerecently,Forssbaeck(2011)explorestheimportanceof deposit insurance and ownership on bank risk taking, but his work differs from ours along several dimensions. For instance, his paper focuses on the period from 1995 to 2005 and, in contrast to our findings, shows that there is no support to the proposition that the market discipline channel becomes more important during crises. Dam and Koetter (2012) find support for the market discipline channel for the period between 1995 and 2006, but their study is restricted to German banks and focus on a measure of probability of support that is derived from actual bailouts. A recent strand of papers have used a natural experiment approach to control for possible reverse causality between government support and risk taking.5 However, we feel that we adequately address the issue of endogeneity by considering lagged expected (and not actual) support, by using an instrumental variables approach, and futher, by estimating panel regressions with firm fixed effects and marketmeasuresofrisktocontrolforunobservablefactorsthatareinvariantintheshort-run. Furthermore, the cited papers only use evidence from Germany. This has two drawbacks. The first is that a single-country analysis does not allow for the identification of the impact of regulatory factorsthatmayaffecttherelationbetweengovernmentsupportandrisktaking. Thesecondisthat it is difficult to draw general implications from those studies given the specificities of the German banking system, with a very large number of local savings and loans institutions and state controlledbanks(Landesbanken). Inrelatedwork,otherauthorshavefoundapositiveeffectofactual 5SeeGroppetal.(2013),Ongenaetal.(2013),Damaretal.(2012),Fischeretal.(2012),andSchnabelandKrner (2012),amongothers,whousethesameexperimentfromGermany,in2001. 4

government support on bank risk taking.6 We instead focus on expected, not actually received, government support and use a sample of banks from many countries. Our variable of government support measures the expected willingness and the ability of external agents to provide support to banks. It is not a measure of actual support (for which the endogeneity issue is clearly problematic) and it is not susceptible to the criticism of being less relevant during crisis periods, when governmentsmaynothavethefiscalspacetoprovidesupport. In contrast to the previously mentioned studies, Gropp et al. (2011) find that expected government support to a given bank induces more risk taking by the bank’s competitors.7 Unlike our study, they do not find a consistent relationship between support and risk taking by protected banks. In fact, their findings suggest that protected banks take on less risk, which is consistent withthechartervaluechannelbeingdominant. Thedifferentfindingscanbeexplainedbyouruse of a different measure of risk and of additional bank and country-specific controls. The z-score, our measure of risk taking, is a broader measure of risk since it encompasses both credit risk and marketrisk andsummarizessome ofthemeasures usedby Groppetal. (2011).8 Furthermore,our sampleexcludesbanksubsidiariesbutincludesthepost-financialcrisisperiod. Studyingandunderstandingbankrisk-takingbehaviorisimportantforavarietyofreasons. Excessiverisktakingbybanksisoftenassociatedwithbankfailuresandcostlygovernment-financed rescues. Banking crises are in turn associated with sharp recessions, large drops in asset prices, protracted recoveries and big increases in government debt (Reinhart and Rogoff, 2009). In addi- 6See,forinstance,BlackandHazelwood(2012)fortheeffectofTARPonU.S.banks. 7Governmentbail-outguaranteestoagivenbankmayincreaserisktakingbyitscompetitorsbecausetheydecrease theirchartervalue(HakenesandSchnabel,2010). 8The z-score is a widely used measure of risk, especially in cross-country banking studies (Laeven and Levine, 2009). Groppetal.(2011)usefourdifferentmeasures: theproblemloansratio(problemloansovertotalassets),the riskassetratio(riskyassetsovertotalassets),theliquidityratio(liquidassetsovershort-termliabilities),andtheequity ratio(bookcapitalovertotalassets). However,thefirsttwomeasuresaredifficulttocompareacrosscountriesdueto regulatorydifferences(EuropeanBankCoordinationInitiative,2012),whilethethirdisnotanactualmeasureofrisk taking. Finally,theequityratioisacomponentofthez-scoreandwetestitsdirectrelationtogovernmentsupport,in thisstudy,asarobustnesscheck. 5

tion, banks exposed to lower default risk seem to better insulate their loan supply from monetary policy changes and to offer more credit (Altunbas et al., 2010). For these reasons, theoretical and empiricalstudiesofbankrisktakinghavebeenusedbybothcentralbanksandregulatoryagencies to frame prudential policies.9 In the sense that our results provide an estimate of the magnitude of the moral hazard effect of government support to banks, they are also useful as an input for researchersandregulators. Our results have strikingly different policy implications from those of related papers. If the main channel through which government support affects bank’s risk-taking is by increasing the charter value of guaranteed banks then it makes sense to apply a capital surcharge on protected banks to decrease their rents and their (unprotected) competitors’ incentives to take on more risk (Gropp et al., 2011). Increased capital requirements would also reduce gambling incentives by putting more bank equity at risk.10 However, we do not find empirical evidence in favor of the charter value hypothesis. Furthermore, we provide direct evidence that pre-crisis capital requirementsdidnotweakenthelinkbetweengovernmentsupportandrisktakingbybanks.11 If, as we find, the dominating channel is ”market-discipline” and pre-crisis capital requirements proved to be ineffective in reducing moral hazard, then measures to increase the incentives bydepositorsandsubordinatedcreditorstomonitororinfluencebanks’attitudestowardsrisksare preferable. These include imposing more transparency and forcing more disclosure by bank managers, mandating periodic issuance of subordinated debt or using market information to improve 9SeeBoydandDeNicolo´ (2005)foradiscussiononpolicyresponsestoperceivedlinksbetweencompetitionand theriskofbankfailures. 10However, in as much as uniform capital requirements decrease the charter value of all banks, Hellmann et al. (2000)suggestusingdeposit-ratecontrolsasanadditionalprudentialmeasure. 11Theineffectivenessofpre-crisiscapitalrequirementsdoesnotimplythatmuchhighercapitalrequirementsand wider risk coverage (such as the ones in Basel III) could not weaken the link between government support and risk taking by banks (see, for instance, Admati and Pfleiderer, 2010). However, the quantity and quality of required capital,beforethecrisis,didnotpreventthebuild-upofexcessiveon-andoff-balancesheetleveragebybanks(Basel CommitteeonBankingSupervision,2009). 6

thequalityofsupervision(Rochet,2005). Moreover,theincreaseinbankcomplexityoverthepast decade may have decreased the effectiveness of investor monitoring, as it became more difficult for ”outsiders” to assess the level and types of risks taken by banks. Our second finding provides evidencethatinvestorsandregulatorsmaylimitrisktakingbybanks,evenforthosethathavegovernment support, if these banks’ range of activities is restricted. Thus, simple rules like those that wereincludedintheGlass-SteagallActcouldpotentiallybereconsidered(Haldane,2012). The rest of this paper is organized as follows. In Section 2, we describe our sample and detail our data sources and in Section 3 discuss our hypothesis and methods. In Section 4, we present our results on support and risk taking, as well as several robustness checks and, in Section 5, we discusshowregulationmayaffectthisresult. Section6concludes. 2 Data 2.1 Banks and Bank Risk Taking Weusethez-scoreasourmeasureofbankrisk. Thez-scoreequalsthereturnonassets(ROA) plus the capital asset ratio (CAR) of each bank divided by the banks’ standard deviation of return on assets (σ(ROA)). The z-score measures the distance to insolvency since it is the inverse of the probability that losses exceed equity (that is, prob(-ROA>CAR); see Laeven and Levine, 2009). Ahigherz-scorethereforeindicatesthatthebankislessrisky. A characteristic of the z-score is that it is highly skewed. For this reason, we use the natural logarithm of the z-score. We have data across 54 countries to calculate the z-score for 286 banks for the period 2003–2004, and for 321 banks in 2009–2010. These banks are also required to be rated by either Moody’s Investors Service (Moody’s) or Fitch Ratings (Fitch), two of the major rating agencies. As listed in Table 1, the number of banks per country varies from 1 to 30. The 7

resultsinthepaperarerobusttoexcludingcountrieswithlessthan2banks. InadditiontorestrictingoursampleofbankstothoseratedbyMoodysorFitch,weexcludeall subsidiariesbecauseourmeasureofsupportconsidersallexternalsupport,includingthatofparent companies. Theserestrictionsreduceoursamplesizebutmakeourapproachmuchmorerigorous. Specifically, we do not assume that, simply because a bank does not have a support rating, it does nothavegovernmentsupport. Doingotherwise,inouropinion,wouldinduceameasurementerror thatwouldbiasourresults,andhelpsexplainwhywefindtheoppositeofwhatGroppetal.(2011) find. The inclusion of most major international and systemically important banks in our sample increases the economic significance of our results and also explains why we find the opposite of otherempiricalstudiesdominatedbysmalllocalbankswhichenjoysupportmostlythroughdeposit insurance (see Ioannidou and Penas, 2010, as well as Demirguc-Kunt et al., 2008, and references therein). Forthesesmallerbanks,mispricingofdepositinsuranceistherelevantincentivefortheir risk taking behavior. Instead, large banks fund a significant portion of their assets with senior short and long term unsecured debt. The systemic government support that investors perceive theselargebanksarelikelytoget,createsandimplicitfundingsubsidy(Angineretal.,2013)with consequencesthathavenotbeenstudiedyet. To calculate the z-score, we compute the standard deviation of ROA using 5 year rolling windows. Then we average the z-score for the years included in our two cross-sections, 2003–2004 and 2009–2010. We focus on a cross-sectional analysis due to a change in accounting standards thataffectedalargesampleofEuropeanbanksandbanksinotherregions. Inthemid-2000s,some countriesreplacedlocalGeneralAcceptedAccountingPractices(GAAP)withInternationalFinancial Reporting Standards (IFRS) for publicly-traded banks based in these countries. The change in accounting standards had a notable impact on the way bank balance sheets are reported. For instance,underIFRSrules,derivativeassetsandliabilitiesarenotnetted,increasingthetotalvalue 8

of assets of the bank. To avoid including biases due to the change in accounting treatment we focus on periods in which banks consistently use one or the other accounting method, and focus on cross-sections of results.12 The accounting data on banks are from Bankscope, a commercial databasewithextensiveinformationonbanksacrosstheglobe. As a robustness check we use three additional measures of risk-taking. The first is a marketbasedmeasureofthez-score(Forssbaeck,2011). Itisdefinedastheratioofabanksaveragestock returnoverayearplusitsleverageratiooverthestandarddeviationofthisbanksstockreturninthe same year.13 As with the accounting based z-score, we use the natural logarithm of this measure in our estimations. The second measure is the standard deviation of a banks weekly stock return, which enters in the denominator of the market-based z-score measure described above. Although these market-based measures are more forward looking, as opposed to the accounting-based zscore, the downside of using them is that an important segment of banks which are not publicly traded are dropped from the sample.14 This is particularly important in Europe as a significant portion of its banking sector is not publicly-listed. For example, German Landesbanks, which enjoy notable implicit support from the government and took notorious risks prior to the global financialcrisis(Artetaetal.,2013),areexcludedfromthissampleastheirstocksarenotlisted. The third alternative risk measure is the ratio of loan loss provisions to average total assets. Although this measure is available for most banks in the sample, it only captures risk-taking through one dimension: banks loan portfolios. Prior to the global financial banks took risk through assets different than traditional loans, like U.S. issued mortgage backed securities. Risk taking through theseinstrumentsisnotcapturedbyloanlossprovisions. 12Theuseofapanelwiththez-scoreisthereforeimpossiblebecauseofthe2005IFRSshift, inconjunctionwith thefiveyearwindowneededtocalculateσ(ROA). 13Weuseend-of-monthstockpriceinformationfromBankscopetocalculatethebanksstockreturns. 14Theseequity-basedmeasuresarenotentirelyforwardlookingbecausetheyarebasedonhistoricalreturnsvolatilityratherthanoptions-impliedvolatilities. Thesedataareunavailableformostbanks. 9

2.2 Bank Support We measure bank support using bank-specific ratings information from Moody’s and Fitch. Since 1995, Moody’s has assigned bank financial strength ratings (BFSR) to banks in about 90 countries. According to Moody’s, BFSRs “are intended to provide investors with a measure of a bank’s intrinsic safety and soundness on an entity-specific basis” (Moody’s Investors Service, 2007). More importantly, this measure does not include any external support that a bank may receive from its parent, other institutions under a cooperative or mutual arrangement, or the government. Moody’s also assigns a bank deposit rating to the banks it rates. This is the rating agency’s opinion on a bank’s ability to repay its deposit obligations punctually. As such, they incorporate both the bank’s BFSR rating and Moody’s opinion of any external support. Since this measure includes any type of external support, including that of parent companies, not just that of governments, we exclude from the sample all bank subsidiaries. This reduces the size of the sample considerablybuteliminatesanimportantsourceofmeasurementerror. In the main specifications, the bank-specific government support measure is defined as the difference (in rating notches) between a bank’s BFSR and its long-term foreign currency deposit rating. As a robustness check, we also define support in terms of the probability of a government bailout as in Gropp et al. (2011). This amounts to assigning a default probability to each bank accordingtotheBFSR(thedefaultprobabilityintheabsenceofabailout,d)andanotheraccording tothedepositrating(thetotaldefaultprobability,takingaccountbailouts,td),usinghistoricaloneyearaheaddefaultfrequenciescollectedbyMoody’s. Thebailoutprobabilityisp = 1−td/d. Fitch Ratings provides a similar measure of the probability of support which we use as an additional robustnesscheck. Figure 1 shows the evolution of average and median government support since 1996 for all 10

banks included in the sample. Support tends to increase during periods of economic distress, as was the case during the East Asian and Russian crises of the late 1990s, and the recent financial crisis. 2.3 Control Variables Wecontrolforaseriesofcharacteristicsatthebank,industry,andcountrylevels. Forthemost part, we follow Laeven and Levine (2009). The bank-specific controls include revenue growth (measured as the growth in total revenues relative to the previous period), size (the bank’s log of total assets), and liquidity (bank’s liquid assets to liquid liabilities) and are all sourced from Bankscope. We also control for bank ownership by including a variable of cash flow rights of largeshareholders(LaevenandLevine,2009,see)anddummyvariableswhichsignalgovernment, institutional, individual, or other type of ownership (data from Capital IQ, SNL Financial and banks’websites). Allbank-specificdataisfromBankscope. At the country level, we control for per capita income, inflation, inflation variability (data fromtheWorldBankDevelopmentIndicators),thequalityofinvestorprotectionandthedegreeto whichcontractsareeffectivelyenforcedinacountry(bothfromthe2003and2009DoingBusiness ReportoftheWorldBank). The level of competition in banking markets is another factor which affects risk taking. Some studies suggest competition among banks for deposits decreases charter value and therefore leads to riskier portfolios being held by banks (for instance Keeley, 1990, and Hellmann et al., 2000). For this reason, we control for bank concentration at the industry and country level using the Hirsch-Herfindahlindex(datafromBankscope). Intermsofbankingregulations,wecontrolfortheexistenceofadepositinsuranceschemeand 11

for the level of capital requirements (measured by the minimum capital-asset ratio requirement). Data on deposit insurance comes from Demirguc-Kunt et al. (2008), the Institute International Bankers(GlobalSurveys2009and2010),theInternationalAssociationofDepositInsurers(IADI), theCentralBankofEgypt,andtheSingaporeDepositInsuranceCorporationLimited(SDIC). Finally, we use as regressors several variables which measure the intensity and breadth of regulation in the banking sector and at the country level, as defined in Barth et al.’s (2006) bank regulatory database. We use the level of capital stringency, the level of official bank supervisory power, and an index of activity restrictions (all defined in Barth et al., 2006). Capital stringency measures the regulatory approach employed to determine and verify the extent of the capital at risk at banks. The variable reflects, among other information, whether the minimum capital-asset ratio(risk-weighted)requirementisbasedonBaselguidelines,whethermarketvalueofloanlosses not realized in accounting books is deducted, or if the initial disbursement of capital can be done with borrowed funds. The official supervisory power variable measures the extent to which the regulatoryorsupervisoryauthorities havethepowertotakespecific actionstopreventandcorrect problems. This includes the right to meet with external auditors to discuss their report without the approvalofthebank,therighttoorderthebank’sdirectorsormanagementtoconstituteprovisions to cover actual or potential losses, among other rights. Activity restrictions is an index measuring regulatory limitations to banks operating in securities markets, insurance activities, real estate, and engaged in the ownership of non-financial firms. For the 2003–2004 cross-section we use information from the 2003 regulatory database, and for the 2009–2010 cross-section we use the datacompiledinthe2008versionofthedatabase. 12

2.4 Summary Statistics Table 2 provides summary statistics for the key regression variables. Statistics are based on averages for the periods 2003–2004 and 2009–2010 using annual data for our measure of risk taking (z-score). For for the other variables we use annual data for 2002 and 2008. The table indicates that there is ample variation in the bank risk taking measures and in the other relevant variables across banks in the sample periods. The table also shows a slight increase in the level of measured risk-taking (0.3 standard deviations of the z-score) and a somewhat more substantial increase in the average size of banks (0.5 standard deviations), when we compare 2003–2004 to 2009–2010. If we take previous studies at face value, these two facts in isolation are consistent with larger banks, possibly with more market power, taking on less risk. However, it is important toexplorewhetheranincreaseingovernmentsupportmayhaveledtomorerisktakingbybanks. In fact, regardless of the measure we use, the data shows a sizable increase in the average level of support from 2002 to 2008. The increase is even more significant when we look at the median level of support. The median probability of support estimated by Moody’s increases from 0% to 40%, from the first sample period to the second one, signaling a widespread increase in government support to banks. This increase is much more pronounced in Moody’s measure than inFitch’s(Figure1). INSERTFIGURE1 13

3 Hypothesis and Empirical Strategy Our first hypothesis is that bank risk taking is related to government support to the banks. The basicempiricalspecificationtotestthehypothesisisformulatedasfollows, Z = β +β ×GS +β ×X +β ×W +ε b,c,t 0 1 b,c,t−1 2 b,c,t−1 3 c,t−1 b,c,t where Z is the natural logarithm of the z-score of bank b in country c for period t, GS is b,c,t b,c,t−1 government support for bank b from country c, X is a matrix of bank level control variables, b,c,t−1 W are country-level controls, ε is the error term, and β ,β , and β are slope coefficients c,t−1 b,c 1 2 3 or vectors of coefficients. The standard errors are adjusted to control for clustering at the country level. Because we are using government support lagged by at least one period, we claim that supportcausesrisktakingbybanks. The approach just outlined may be compromised if GS is endogenous or if there are omitted variables (i.e. the possibility that cov(Z ,ε ) (cid:54)= 0). We used two approaches to deal with the b,c b,c problem. The first is to saturate the regression with many bank and country specific measures to capture as much of the error term as possible (Bitler et al., 2005, and Laeven and Levine, 2009). The second approach we consider is to use instrumental variables. In addition to the benchmark regressionabove(withoutinstruments),weinstrumenteachbank’sgovernmentsupportasfollows. For each bank n, we employ the average GS of the other n-1 banks in the country, which reflects industryandcountryfactorsexplainingGS.Theinstrument’svalidityreliesontheassumptionthat an innovation in the risk taking of any given bank does not affect government support to other banks. The interaction between national regulations and government support, and the interaction betweenbanklevelownershipandgovernmentsupport,areconsideredinthesecondhypothesis. Our 14

secondhypothesisisthatbanksupervisionandregulationaffectstheimpactofgovernmentsupport onbanks’risktakingbehavior,whichwetestusingthefollowingspecification: Z = β +β ×GS +β ×R +β ×GS ×R +β ×X +β ×W +ε b,c,t 0 1 b,c,t−1 2 c,t−1 3 b,c,t−1 c,t−1 4 b,c,t−1 5 c,t−1 b,c,t where R are country-specific regulatory standards, so that GS ×R captures the inc,t−1 b,c,t−1 c,t−1 teraction between the bank-specific government support measure and national regulations, and β 3 isthecoefficientestimateoftheinteractioneffect. 4 Results 4.1 Benchmark Regression The benchmark empirical results on the link between bank risk taking and government support are reported in Table 4. The first main finding is that larger government support is associated with greater risk taking by banks, as reflected in the negative coefficient for government support (GS) found for almost all specifications. The second important result is that the relationship betweengovernmentsupportandbankrisktakingispresentforboththe2003–2004and2009–2010 periods, but the coefficients are generally more statistically significant during the latter period. Regressions 1 and 8 control for recent bank performance (revenue growth), and show that a one standard deviation increase in government support is associated with a 4.5 percent decrease on the average z-score for the 2003–2004 period, but the relationship is not statistically significant. For 2009–2010, the government support coefficient is negative and statistically significant, and its magnitude indicates that a one standard deviation increase in government support is associated to a 6.9 percent increase in bank risk taking, relative to the average z-Score. These findings are 15

consistentwiththeviewthatincreasinggovernmentsupporttobankstendstoreducemarketdiscipline,inducingfurtherbankrisktaking. ThepositiveassociationbetweenGS andriskholdswhen controlling for bank characteristics and country-level features, and after including country fixed effects,asweshownext. Toconsiderthepossibilitythattheassociationbetweengovernmentsupportandbankrisktaking reflects other bank level differences instead of cross-bank differences in government support, the regression results shown in columns 2 and 9 control for the bank-specific characteristics of revenue growth, size, and the liquidity ratio. We have three comments on the results. First and foremost, the positive association between GS and banks’ risk-taking remains significant for the 2009–2010 period and insignificant for 2003–2004. Our results are therefore robust to the inclusion of bank-specific characteristics. Second, while revenue growth seems to capture well the charter value effect (in as much as banks with faster growth are better able to generate rents), size on its own does not seem to impact risk taking (the variable is almost never significant).15 Third, banks with higher liquidity take (significantly) more risks. Our interpretation is that liquidity is capturingabank-specificappetiteforrisk: bankswithariskierbusinessmodel(forinstance,more securities’trading)keepmoreliquidityathandincaseoflossesormargincalls. We also take into account the possibility that the link between government support and bank risk taking captures cross-country heterogeneity instead of cross-bank differences in government supportbyrunningregressionswithcountryfixedeffects(columns4and11). Alternatively,regressions in columns 5 and 12 control for several country-specific characteristics, including the level of economic development in each bank’s home country (per capita income), indicators of capital requirements,thelevelofinvestmentprotection,16 thepresenceofdepositinsurance,thedegreeto 15Oneexplanationforthelowsignificanceofgrowthasadeterminantofrisktakingisthatthetoo-big-to-faileffect and the charter value hypothesis cancel each other out. Another explanation is that larger banks are better at risk diversificationbutalsohardertomonitorbecauseofincreasedcomplexity. 16Using Djankov et al.’s (2008) revised anti-directors index or their anti-self-dealing index does not change the 16

whichthelawiseffectivelyandfairlyenforcedinacountry,andtheHerfindahlconcentrationindex for the banking system. The results yield two comments. First, for both cross-sections, the result that government support leads to riskier banks is robust to conditioning on either country controls or fixed effects. Second, of all country controls, only per capita income and inflation volatility aresignificantforbothtimeperiods. Whileanincreaseininflationvolatilityalwayscausesriskier banks, the change in the sign of the coefficient associated with income per capita reflects the fact thatadvancedeconomieswerethemostaffectedbythe2007–2009crisis.17 It is possible that our results are affected by a possible endogeneity of government support. We explicitly tackle this using an instrumental variables approach. As shown in regressions 3 and 10,theinstrumentalvariableresultsconfirmthatGS ispositivelyandsignificantlyassociatedwith bank’s risk taking, at least for the crisis period. In fact, not only does the coefficient associated withGS remainstatisticallysignificant,butitsmagnitudedoesnotchange. Bank ownership structure has been shown to be an important explanation of the level of risk taking by banks since it critically conditions the conflict over risk between bank managers and owners (Laeven and Levine, 2009). In regressions 6 and 13, in addition to the previous bank and country level controls, we control for cash-flow rights and for ownership structure (as in Laeven and Levine, 2009) by looking at the extent to which there are large shareholders in the bank and by differentiating between government, institutions, individuals and others. The positive and significantassociation betweenbank risktakingand governmentsupport isrobustto theseadditional controls. flavorofresults,whichareavailablefromtheauthorsifrequested. 17Sincethetwotimeperiodsreflecteddifferentmacroeconomicandfinancialsectorconditions,wecheckedwhether controllingforequitymarketvolatilitymadeadifference. Forthiseffect, weusedthepreviousyear’saveragedaily volatility of the banking sector stock index from Datastream for each country (when available). The (untabulated) results were unchanged. We also tried to control for financial sector soundness (which would proxy for regulatory forbearance),usingtheBankSoundnessindexfromtheGlobalCompetitivenessReportbutthisdidnotaffectresults either. 17

A final specification issue we tackle is the one pertaining to the timing of support being given andriskmaterializing. Inourbenchmarkspecificationssupportislaggedbyoneperiod(weregress the 2003–2004 and 2009–2010 z-Scores on 2002 and 2008 supports, respectively). Since investment and credit decisions (possibly affected by government support) may take longer than one year to affect results, we regress the z-Score averages on 2001 and 2007 support (using a longer lagwouldrestrictseverelyoursamplesize). Theresults,shownincolumns7and14,arebasically thesameasintheotherregressions. 4.2 Robustness We perform three robustness exercises which involve using alternative definitions for risk taking and government support, estimating panel regressions with fixed effects, and considering changes in bank valuation in the tests. Finally, we assess the predictive power of the selected Moody’s measure of government support to anticipate actual government bailouts, and compare it to Fitch’s alternative. In the first exercise, instead of the z-Score, we use the individual components of the z-score (ROA, Capital to Assets, and the standard deviation of ROA). We regress thesemeasuresonbankcontrolsandoncountrycontrols,asinthebenchmarkregressiondiscussed before.18 The results are available on Table 5 for the selected time periods: 2003–2004 and 2009–2010. The regressions show a strong and statistically significant effect of government support on ROA regardless of the time period. In the pre-crisis sample, government support was also positively 18Wealsotriedusingloanlossprovisionsasapercentageofassetsasanalternativemeasureofrisk. Thismeasure presentstwoproblems. First,thedefinitionofwhatareloanlossesandofhowmuchandwhentoprovisionforthose lossesvariesacrosscountriesbyagreatdeal. Thiscausesamisspecificationproblem. Asecondproblemwithusing loan loss provisions is that it provides a very incomplete measure of risks taken. Specifically, loan loss provisions (imperfectly) cover risks associated with loan portfolios and disregard other types of credit risks, let alone market risks which affect a broader set of assets held by banks and were more important during the recent financial crisis. Preliminaryfindingsseemtoconfirmthisandareavailableuponrequest. 18

and significantly related to the volatility of ROA. In the crisis sample, government support was negatively and significantly related to the capital to assets ratio. We interpret these findings as follows. Before the crisis, support tended to encourage riskier bets by banks which translated into more volatile returns. After the crisis, two additional interpretations arise. On the one hand, it is possible that banks took more risk by increasing leverage. On the other hand, it could also be the case that banks took more risks, which led to more losses and lower capital buffers to withstand shocks. Interestingly, in contrast to what we find for the z-Score itself, size matters for each individual component of the z-Score, particularly for the second time period. In fact, larger banks tend to be moreleveraged-“too-big-to-fail”effect-butalsotohavelessvolatilereturnsonassets-diversification effect. The combination of the two countervailing two effects in the z-Score explains why, inthebenchmarkspecification,banksizedoesnotsignificantlyimpactbankrisk. In our second robustness test we replace our accounting-based measure of the z-score for a market-based version. Moreover, we use the volatility of banks stock returns and the ratio of loan loss provisions over banks average assets as additional measures of risk. The advantage of these measures over the accounting based z-score is that they are forward looking or calculated with data for just one year. This allows us to run panel regressions with bank-level fixed effects. As noted before, the inclusion of fixed effects control for unobservable bank characteristics that may be correlated with risk taking and invariable in the short run. For example, the compensation structure or culture of risk in a bank may lead to more risk taking (Cheng et al., 2010). If these traitsareinvariantoverashortperiod,theyarecapturedbythefixedeffects. Table 6 shows the results of these panel estimations for the period between 2005 and 2010. All regressors enter the specifications contemporaneously with the exception of our government support measure. Columns 1, 3, and 5 report results using support lagged by one year, while the 19

other columns show the same specifications using support lagged by two years. As shown in the first two columns of the table, the coefficient on government support is negative and significant, whenusingthemarket-basedmeasureofz-scoreastheriskproxy. Thisconfirmstheresultsshown withtheaccounting-basedmeasureinacross-sectionalsetting. Thelagofthegovernmentsupport does not matter for the level of significance or sign of the coefficient. The next two columns show resultsfortheestimationsusingstockreturnvolatilityastheriskproxy. Inthiscase,thecoefficient on government support is positive and significant, implying that more government support leads tomorestockreturnvolatility. Lastly,thelasttwocolumnsshowtheestimatedcoefficientsforthe samespecificationusingloanlossprovisionsastheriskmeasure. Thecoefficientsforbothspecificationsarepositiveandsignificant,butonlythecoefficientontheone-yearlagsupportmeasureis statistically significant. This result confirms that risk-taking through loans was important but not theultimatemechanismavailabletobankspriortothecrisis. Asecondrobustnesstestrequiresreplacingournotches-baseddefinitionofgovernmentsupport withonewhereweassignprobabilitiesofagovernmentbailoutasinGroppetal.(2011). Wethen replicate the regressions presented in Table 4: two regressions with bank controls only, one with countyfixedeffects,andonewithcountrycontrolsforbothtimeperiods. OurfindingsareinTable 7. Most results are qualitatively the same as the ones for the benchmark regressions. During the crisis,usingourpreferredspecification(countryfixedeffects),aonestandarddeviationincreasein the probability of a bailout led to an 8 percent increase in risk (relative to the mean).19 This effect issignificantatthe1percentsignificancelevel. Weextendourrobustnesscheckbyperformingtheexerciseusingprobabilitiesofagovernment bailout derived from data collected by Fitch Ratings (the same data source used by Gropp et al., 2011, and Forssbaeck, 2011). We run the same regressions as in Table 7 and present the results in 19This would be equivalent to going from no support to a level slightly below the median level of support in the industry. 20

Table 8. The main difference in terms of results is that government support is not significant for the pre-crisis period. In fact, as in Gropp et al. (2011), we find that for that period (2003–2004), a higher probability of a government bailout is not associated with the supported bank taking on more risk.20 However, when we look at the crisis period (2009–2010), we do find strong evidence of moral hazard in government support to banks, as we had in the regressions with the Moody’sbasedmeasuresofsupport. So far, we have only implicitly considered the hypothesis of bank charter value determining the link between support and risk taking. We did this by including the degree of market concentration(measuredbytheHirsch-Herfindahlindex)asoneoftheindustry-countrycontrols. Results on columns 5 through 7 and 12 through 14 in Table 4 show that market concentration is never significant.21 Thisdoesnotmeanthatthechartervaluechannelisirrelevantsincecompetitioncan affectchartervalueinmorethanoneway(Martinez-MieraandRepullo,2010,suggestaU-shaped relationship between competition and the risk of bank failure). For instance, competition in lending markets may be negatively related to bank risk taking, as suggested by Boyd and De Nicolo´ (2005).22 The third robustness exercise is therefore to explicitly consider the charter value channel. We do this by allowing for the joint determination of bank risk and bank valuation and then testing for the link between risk and government support independent of bank value. We expand our baseline specification with bank and country controls by including Tobin’s Q as an endogenous explanatory variable. We calculate the Tobin’s Q as total assets plus market value of equity (data 20Thisisprobablydue,atleastinourstudy,tothisprobability-basedmeasureofgovernmentsupportnotshowing enoughvariationinthepre-crisissample(seeFigure1). 21WealsotriedtocapturethechartervalueeffectwithvariablesrepresentingbarrierstoentrysuchasBarthetal.’s (2006)indexofbarrierstoentryandeitherthenumberorthechangeinthenumberofbanksinthecountry(normalized byGDP).Changingthevariableshadnoimpactonourresults(availablefromtheauthorsuponrequest). 22Theirargumentisthatifthereislowcompetitionamongbanksforloanstofirms, interestrateschargedwillbe higherandthiswillforceentrepreneurstochooseriskierprojects,therebyincreasingcreditriskbornebybanks. 21

fromBankscope)minusbookvalueofequitydividedbytotalassets. Weestimatethemodelusing two-stage GMM and two excluded instruments in the first stage regression: a dummy variable for thebank’sstockbeingwidelyheld(Widely)andthenumberofbanksnormalizedbythecountry’s gross domestic product.23 We are only able to do it for the second period due to data availability. TheresultsinTable9showthatwhenitcomestoexplainingbankrisktaking(secondstageregression), our variable of government support is still significant (albeit at the ten percent level only) butbankvalueisnot. 4.3 Predictive Power of Government Support The relevance of the empirical work we present in this study relies on the adequacy of our measures of government support. In addition, we assess the measure of government support from Moody’s, which was adopted for the baseline regressions, against the alternative from Fitch’s. In Table 3 we can see that Moody’s and Fitch’s probability-based measures of support were mildly correlated before the crisis and became more correlated with the crisis. For the period before the crisiswealsoseethatMoody’smeasurewasuncorrelatedwithsizewhileFitch’swassignificantly correlatedwithbanks’totalassets.24 Thesetwofactssuggestthat,fortheperiodwhenthetwomeasureswerethemostdifferentfromeachother(beforethecrisis),Moody’smeasurewascapturing, to a larger extent than Fitch’s, other aspects of government support besides the ”too-big-to-fail” hypothesis. Amoredefinitivewayofsettlingtheissueistotestwhetherthesemeasuresareabletopredict actualbail-outs. Onewayofdoingthisistorunaprobitregressionofactualgovernmentinterven- 23WetriedusingLaevenandLevine’s(2009)excludedinstruments-shareofassets,beinglistedontheNewYork Stock Exchange, and the country having barriers to entry to the banking industry - but these proved to be weak instruments. 24Afterthecrisis,theyarebothcorrelatedwithsize,asexpected. 22

tionsinbanksonourmeasuresofsupport. Tothiseffectwedefineabinaryvariabley whichtakes it value1ifbankieitherreceivedacapitalinjectionbyitsgovernmentorwaspartiallyortotallynationalized between 2008 and 2010. We start with data on capital injections in Europe from Brei et al. (2011) and complement those with information retrieved from Laeven and Valencia (2012) and FT.com. The data include 238 banks but there is ratings information for only 137, of which roughlyonethirdwereintervened(Table10). Wepositthatthelikelihoodofabankbeingactuallyrescuedbyitsgovernmentdependsonthe amount of ex-ante government support and on how distressed the bank was prior to the crisis, as well as on other characteristics such as size, capital, and liquidity. We use loan loss provisions as a percentage of average assets as a measure of bank distress.25 Since the impact of support will certainly depend how distressed the bank was to start with, we include an interaction of support with loan loss provisions. In order to make the interpretation of the effect of interacted variables easier, we replace our main variable of support by a dummy variable which takes value 1 if the Moody’s-based support measure (in notches) was positive (support) and 0 otherwise (no support). WealsoconditionforthestateoftheeconomyusingtheaverageGDPgrowthratefor2007–2009. To address any concerns of endogeneity, all controls (except for GDP growth) correspond to 2007 values. Weestimatethefollowingmodel: y∗ = α +α GS α LLP +α GS ×LLP +Γz +u , it 0 1 it−1 2 it−1 3 it−1 it−1 it−1 it where y∗ > 0 (y = 1) if bank i was the target of a government intervention between 2008 and it it 2010. LLP is loan loss provisions as a percentage of average total assets and z a vector with the 25By including these controls we are controlling for systemic importance since size, leverage, and asset risk (e.g. loanlossprovisions)arethemaindriversofsystemicrisk(Hovakimianetal.,2012). 23

other controls mentioned above. The estimation results are in Table 11. Estimates of coefficients in probit models do not have an economic interpretation, especially when interaction terms are present. For this reason we focus on the pairwise comparison of marginal effects. This measure estimates the average predicted probabilities of having a capital injection conditional on being in eachofthetwosupportgroups(nosupportorsomesupport)andunconditionalontheothercontrol variables. Wefindthatbankswhoenjoyedsupportin2007weremorelikelytoberescuedin2008– 2010 by 30 percentage points and that the difference is highly significant. We take this as strong evidence in favor of the predictive ability of our measure of support and therefore of its economic significance. ThesameestimationusingtheequivalentFitchmeasureshowsinsignificantmarginal effects of government support (available from the authors upon request). Therefore, Moody’s measureisabetterchoiceforourbaselineempiricalspecifications. 5 Regulation and Government Support Our research is the first attempt to explore the interactive effects of national regulations and bank-specific government support on the risk taking behavior of individual banks. We use data on regulation for 2003 and for 2008 from Barth et al. (2008). These data consider regulations emphasizedbytheBaselCommitteeandthatthetheoreticalliteraturehaspinneddownasaffecting bankbehavior(LaevenandLevine,2009). Weuseanindexofregulatoryoversightofbankcapital, capital stringency, a measure of official supervisory power and a measure of activity restrictions (seeSection2.3fordetaileddefinitions). The thoeretical underpinnings of how these regulations relate to bank risk taking and government support are complex and suggest multiple effects often with opposite consequences (see Barth et al., 2004, and references therein). The impact of regulations on capital adequacy on 24

bank risk taking is, in principle, ambiguous. While sticter capital adequacy requirements increase the amount of capital at risk thereby, at least in theory, counteracting the moral hazard of government support and limited liability, they may also reduce monotoring incentives. On the other hand, broader official supervisory powers may compensate for the lack of market monitoring of banks (possibly due to perceived government support), but its effectiveness will depend on how closely the supervisor can be monitored by taxpayers and their representatives. Finally, restricting the range of bank activities may decrease the number of opportunities for banks to increase risk, make them less complex and therefore easier to monitor, and act as a limit on bank size and systemic importance, thereby reducing the number of “too-big-or-too-systemic-to-fail” institutions. Alternatively,activityrestrictionsmayhamperbanks’abilitiestodiversifyrisksanddecreasetheir chartervalue,whichwouldincreaserisktaking. It is crucial then to test empirically how these regulations and government support interact to shape banks’ willingness to take on risk. Table 12 shows the interaction of government support withthevarioustypesofbankregulationsincrosssectionregressionsforthe2003–2004and2009– 2010 periods. The regressions include all the bank and country level controls used in the previous tables. Theresultsindicatethatforthe2003–2004period,seenincolumns1to4,governmentsupport was not a significant factor for bank risk taking, and regulation did not play a significant role either. Incontrast,theperiodencompassingtherecentfinancialcrisisisassociatedwithastronger correlation between government support and risk taking. The interaction coefficient for activity restrictions and government support is positive and significant during the crisis period, indicating that limiting the scope of activities and markets where banks should be allowed to operate has limited their risk taking behavior. The magnitude of the interaction coefficient, however, suggests thatactivityrestrictionshavenotfullyoffsetthemoralhazardeffectfromgovernmentsupport. Ourfindingsareincontrast,atleastwhenitcomestoregulationswhichrestrictbankactivities, 25

withwhatwaspreviouslyfoundbyBarthetal.(2006),especiallyifweconsiderthepost-financial crisis period. In fact, while they find that activity restrictions encourage more risk taking and increase bank fragility, we find that they decrease risk taking by banks with government suport. On the other hand, Laeven and Levine (2009) find that activity restrictions encouraged risk taking in 2003-2004, while we find an insignificant effect for the same period (after controlling and interacting with support) and a significant effect (with the opposite sign) for the 2009-2010 period. Therefore, to understand the impact of these regulations on bank stability, we must pay close attention to crisis periods and to the role of governments in providing implicit or explicit bailout guarantees. 6 Conclusion Governmentsupporttobanksthroughtheprovisionofexplicitorimplicitguarantees,intheory, has an ambiguous effect on banks’ risk-taking. On the one hand, by providing support, governments can encourage banks to take more risk because of a moral hazard effect, i.e., the market discipline hypothesis. On the other hand, support can make banks more conservative because it increasestheirchartervalue,i.e.,thechartervaluehypothesis. We use two measures of government support to banks - in notches and in terms of probability of a bailout - from two sources (Moody’s and Fitch Ratings) to capture their attitudes towards risk. After controlling for bank-level and country-specific factors, we find that the intensity of government support is positively related to our measures of bank risk taking. We find that this relationship is stronger for the 2009–2010 period relative to 2003–2004. Our results are robust to endogeneity as well as to the way we measure risk taking. We conclude that the lack of market discipline, especially during the crisis, shaped the relationship between government support and 26

riskinthebankingindustry.26 Moreover,capitalrequirementsregulationandenhancedsupervisory powersfailedtocurbrisktakingduetogovernmentsupporttobanks. Our results suggest that measures to increase the incentives by depositors, small shareholders, and subordinated creditors to monitor or influence banks’ attitudes towards risks should decrease themoralhazardassociatedwithgovernmentsupporttothefinancialsystem. Thisshouldstartwith thecreationofregulatoryenvironmentswhichdonothinderprivatemonitoringofbanks,butcould ultimately lead to limits on the amount of support that governments can pledge.27 Alternatively, restricting banks’ ability to engage in activities involving security markets, insurance, real estate, andownershipofnon-financialfirmsweakensthelinkbetweengovernmentsupportandrisktaking by banks. The way through which restrictions on bank activities ameliorate the problem (either by reducing banks’ ability to engage in risky activities or by reducing banks’ complexity and therefore facilitating monitoring by outside investors and bank supervisors) will be the subject of furtherresearch. The degree to which CEO incentives are aligned with the interests of shareholders influences the amount of risk taking in non-financial firms. However, the existing evidence on banks is still inconclusive.28 An important extension to our paper is to investigate the role of bank governance variables besides ownership. For instance, large board sizes in banks may be optimal given the complexity of the banking business and the large size of many of these firms. This stands in sharpcontrasttonon-financialfirmswhereboardsizeispositivelyrelatedtofree-ridingproblems. Banks are also different from non-financials and other financial firms in that they have many out- 26In fact, one can argue that the lack of effective market discipline was one of the main triggers of the crisis. Therefore,enhancingmarketdisciplineshouldbeanimportantgoalforfinancialregulatoryreforms(Levine,2011). 27NierandBaumann(2006)foundthat,inthe1990’s,marketdisciplinemechanisms,suchasincreaseddisclosure anduninsuredfunding,wereeffectiveininducingbankstolimitdefaultriskbyincreasingcapitalbuffers. However, theeffectisreducedwhenbanksenjoyahighdegreeofsupport. 28For instance, there is some evidence that greater reliance on option compensation or cash bonuses did not have a negative impact on bank performance during 2008-09 crisis (Fahlenbrach and Stulz, 2011) but may have led to acquisitionswhichincreaseddefaultriskbyacquiringbanks(HagendorffandVallascas,2011). 27

side investors (i.e. depositors), are highly leveraged, and are possible beneficiaries of government support. Thistranslatesintoshareholders’interestsbeingoftenconducivetotoomuchrisktaking, at least from a systemic risk point of view. What the optimal bank governance structure should be, given a desired level of systemic risk, is still not totally understood and will certainly be the motivationforfutureresearch. 28

References Admati,A.R.andP.Pfleiderer(2010). Increased-LiabilityEquity: AProposaltoImproveCapital Regulation of Large Financial Institutions. Stanford University Graduate School of Business ResearchPaperNo.2043. Altunbas, Y., L. Gambacorta, and D. Marques-Ibanez (2010). Bank Risk and Monetary Policy. JournalofFinancialStability6(3),121–129. Anginer, D., V. V. Acharya, and A. J. Warburton (2013). The end of market discipline? investor expectationsofimplicitstateguarantees. Unpublishedmanuscript. Arteta, C., M. Carey, R. Correa, and J. Kotter (2013). Revenge of the steamroller: Abcp as a window on risk choices. Technical Report 2013-1076, Board of Governors of the Federal Reserve System. Barth,J.R.,G.Caprio,andR.Levine(2006). RethinkingBankRegulation. CambridgeUniversity Press. Barth, J. R., G. Caprio, and R. Levine (2008). Bank Regulations are Changing: For Better or Worse. ComparativeEconomicStudies50(4),537–563. Barth, J. R., G. C. Jr., and R. Levine (2004). Bank regulation and supervision: what works best? JournalofFinancialIntermediation13(2),205–248. BankCapitalAdequacyRegulationunder theNewBaselAccord. Basel Committee on Banking Supervision (2009, December). Strengthening the Resilience of the BankingSector. ConsultativeDocument.Basel: BankofInternationalSettlments. Bitler, M., T. Moskowitz, and A. Vissing-Jørgensen (2005). Testing Agency Theory with EntrepreneurEffortandWealth. TheJournalofFinance60(2),539–576. Black, L. K. and L. N. Hazelwood (2012). The Effect of TARP on Bank Risk-Taking. Journal of FinancialStability. Articleinpress. Boyd,J.H.andG.DeNicolo´ (2005). TheTheoryofBankRiskTakingandCompetitionRevisited. TheJournalofFinance60(3),1329–1343. Boyd,J.H.andD.E.Runkle(1993). SizeandPerformanceofBankingFirms: TestingthePredictionsofTheory. JournalofMonetaryEconomics31(1),47–67. Brei, M., L. Gambacorta, and G. von Peter (2011). Rescue Packages and Bank Lending. BIS WorkingPapers357,BankforInternationalSettlements. 29

Cetorelli, N., B. H. Mandel, and L. Mollineaux (2012). The evolution of banks and financial intermediation: Framing the analysis. Federal Reserve Bank of New York Economic Policy Review18,1–12. Cheng, I.-H., H. Hong, and J. A. Scheinkman (2010, July). Yesterday’s Heroes: Compensation andCreativeRisk-Taking. WorkingPaper16176,NationalBureauofEconomicResearch. Correa, R., K.-H. Lee, H. Sapriza, and G. Suarez (2013). Sovereign Credit Risk, Banks’ Government Support, and Bank Stock Returns Around the World. Journal of Money, Credit, and Banking. Forthcoming. Dam, L. and M. Koetter (2012). Bank Bailouts and Moral Hazard: Evidence from Germany. ReviewofFinancialStudies25(8),2343–2380. Damar,E.,R.Gropp,andA.Mordel(2012). TheEx-anteVersusEx-postEffectofPublicGuarantees. BankofCanada. De Nicolo´, G. and E. Loukoianova (2007). Bank Ownership, Market Structure and Risk. IMF WorkingPapers07/215,InternationalMonetaryFund. Demirguc-Kunt,A.andE.Detragiache(2002). DoesDepositInsuranceIncreaseBankingSystem Stability? AnEmpiricalInvestigation. JournalofMonetaryEconomics49(7),1373–1406. Demirguc-Kunt, A., E. J. Kane, and L. Laeven (2008). Determinants of Deposit-Insurance AdoptionandDesign. JournalofFinancialIntermediation17(3). pp.407-438. Djankov,S.,R.L.Porta,F.L.deSilanes,andA.Shleifer(2008). TheLawandEconomicsofSelfdealing. Journal of Financial Economics 88(3), 430 – 465. Darden - JFE Conference Volume: CapitalRaisinginEmergingEconomies. European Bank Coordination Initiative (2012). Non-Performing Loans in Central, Eastern and SoutheasternEurope. Brussels: EuropeanBankCoordinationInitiative. Fahlenbrach, R. and R. M. Stulz (2011). Bank CEO Incentives and the Credit Crisis. Journal of FinancialEconomics99(1),11–26. Fischer, M., C. Hainz, J. Rocholl, and S. Steffen (2012). Government guarantees and bank risk takingincentives. InAFA2012ChicagoMeetingsPaper. Flannery,M.andS.Sorescu(1996). EvidenceofBankMarketDisciplineinSubordinatedDebentureYields: 1983-1991. TheJournalofFinance51(4),1347–1377. Forssbaeck,J.(2011). OwnershipStructure,MarketDiscipline,andBanks’Risk-takingIncentives underDepositInsurance. JournalofBanking&Finance35(10). pp.2666-2678. 30

Gropp, R., C. Gruendl, and A. Guettler (2013). The impact of public guarantees on bank risktaking: Evidencefromanaturalexperiment. ReviewofFinance. Forthcoming. Gropp, R., H. Hakenes, and I. Schnabel (2011). Competition, Risk-shifting, and Public Bail-out Policies. ReviewofFinancialStudies24(6). pp.2084-2120. Hagendorff,J.andF.Vallascas(2011). CEOPayIncentivesandRisk-Taking: EvidencefromBank acquisitions. Journal of Corporate Finance 17(4), 1078 – 1095. Special Section: Managerial Compensation. Hakenes, H. and I. Schnabel (2010). Banks without Parachutes: Competitive Effects of GovernmentBail-outPolicies. JournalofFinancialStability6(3). pp.156-168. Haldane, A. G. (2012). The Dog and the Frisbee. Speech given at the Federal Reserve Bank of Kansas City’s 36th economic Policy symposium, “The Changing Policy Landscape”, Jackson Hole,Wyoming(August). Hellmann, T. F., K. C. Murdock, and J. E. Stiglitz (2000). Liberalization, Moral Hazard in Banking, and Prudential Regulation: Are Capital Requirements Enough? The American Economic Review90(1),pp.147–165. Hovakimian, A., E. J. Kane, and L. Laeven (2012, May). Variation in Systemic Risk at US Banks During1974-2010. WorkingPaper18043,NationalBureauofEconomicResearch. Ioannidou, V. P. and M. F. Penas (2010). Deposit insurance and bank risk-taking: Evidence from internalloanratings. JournalofFinancialIntermediation19(1),95–115. Keeley, M. C. (1990). Deposit Insurance, Risk, and Market Power in Banking. The American EconomicReview80(5),pp.1183–1200. Laeven, L. and R. Levine (2009). Bank Governance, Regulation and Risk Taking. Journal of FinancialEconomics93(2),259–275. Laeven, L. and F. Valencia (2012, June). Systemic Banking Crises Database: An Update. IMF WorkingPapers12/163,InternationalMonetaryFund. Levine, R. (2011). Regulating finance and regulators to promote growth. In Achieving Maximum Long-run Growth, Volume 26 of Proceedings of the Federal Reserve Bank of Kansas City EconomicSymposium,JacksonHole,pp.271–311. Martinez-Miera, D. and R. Repullo (2010). Does Competition Reduce the Risk of Bank Failure? ReviewofFinancialStudies23(10),3638–3664. 31

Merton,R.C.(1977). AnAnalyticDerivationoftheCostofDepositInsuranceandLoanGuarantees: An Application of Modern Option Pricing Theory. Journal of Banking and Finance 1(1), 3–11. Moody’s Investors Service (2007). Bank Financial Strength Ratings: Global Methodology. Moody’sInvestorsService,GlobalCreditResearch. ReportNumber: 102151. Moody’sInvestorsService(2012). BankingSystemOutlook: UnitedStatesofAmerica. Nier, E. and U. Baumann (2006). Market Discipline, Disclosure and Moral Hazard in Banking. JournalofFinancialIntermediation15(3),332–361. O’Hara,M.andW.Shaw(1990). DepositInsuranceandWealthEffects: TheValueofBeing”Too BigtoFail”. TheJournalofFinance45(5),1587–1600. Ongena, S., A. Popov, and G. F. Udell (2013). when the cat’s away the mice will play: Does regulationathomeaffectbankrisk-takingabroad? JournalofFinancialEconomics108(3),727 –750. Reinhart, C. M. and K. S. Rogoff (2009). The Aftermath of Financial Crises. The American EconomicReview99(2),466–472. Rochet,J.-C.(2005). PrudentialPolicy. MonetaryandEconomicStudies23(S1),93–119. Schich,S.andS.Lindh(2012). ImplicitGuaranteesforBankDebt: WhereDoWeStand? OECD Journal: FinancialMarketTrends2012(1),1–22. Schnabel,I.andT.Krner(2012). Abolishingpublicguaranteesintheabsenceofmarketdiscipline. WorkingPaper,UniversityofMainz. Tarullo, D. K. (2013). Evaluating progress in regulatory reforms to promote financial stability. SpeechatthePetersonInstituteforInternationalEconomics,Washington,D.C. 32

Figure1: GovernmentSupport,1995–2011 Thebluelinerepresentsmediangovernmentsupport(byyear)measuredbythedifferencebetweenabank’sBFSR anditslong-termforeigncurrencydepositrating,asmeasuredbyMoody’s. Theredline(rightscale)representsthe medianofthesamemeasureconvertedtoprobabilitiesofdefaultasinGroppetal.(2011). Thegreenlineisthe equivalentmeasureofprobabilityofgovernmentsupportbutusingdatafromFitchRatings. 3 11 2.5 00..88 2 00..66 1.5 00..44 1 00..22 0.5 0 00 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 year Governmentsupportinnotches,Moody’s(left-handscale) Probabilityofgovernmentsupport,Moody’s(right-handscale) Probabilityofgovernmentsupport,Fitch(right-handscale) 33

doirePdnayrtnuoCybelpmaS :1elbaT ohwsknabroseinapmocgnidlohknabotdetcirtsersielpmasehT .doirepdnayrtnuocyb,elpmasruonisknabforebmunehtswohselbatsihT .elbaliavas’ydooMmorfsgnitartisopeddnaRSFBhtiwspuorggniknabfosdaehera 0102–9002 4002–3002 0102–9002 4002–3002 6 5 tiawuK 5 5 anitnegrA 5 4 aisyalaM 9 8 ailartsuA 1 1 occoroM 9 8 airtsuA 7 7 sdnalrehteN 6 5 niarhaB 5 6 yawroN 3 3 muigleB 3 3 natsikaP 6 1 lizarB 1 1 amanaP 6 6 adanaC 2 2 ureP 2 3 elihC 5 5 senippilihP 7 7 anihC 9 9 dnaloP 2 4 aibmoloC 2 2 lagutroP 1 1 cilbupeRhcezC 2 2 rataQ 8 3 kramneD 6 2 noitaredeFnaissuR 2 4 tpygE 9 9 aibarAiduaS 1 1 dnalniF 2 2 eropagniS 9 01 ecnarF 2 1 aikavolS 12 42 ynamreG 2 1 ainevolS 8 7 eceerG 1 1 acirfAhtuoS 7 6 gnoKgnoH 21 9 niapS 1 1 yragnuH 5 3 nedewS 3 2 dnalecI 8 4 dnalreztiwS 31 9 aidnI 3 2 nawiaT 4 3 aisenodnI 8 8 dnaliahT 5 6 dnalerI 01 8 yekruT 5 5 learsI 7 4 setarimEbarAdetinU 81 02 ylatI 21 41 modgniKdetinU 91 03 napaJ 32 42 aciremAfOsetatSdetinU 3 3 nadroJ 0 1 aleuzeneV 6 5 fOcilbupeRaeroK 743 033 latoT 34

scitsitatSyrammuS :2elbaT yrtnuocro,yrtnuoc,knabrofelpmasnisriaprotces-knabro,seirtnuoc,sknabforebmunotsreferN.tseretnifoselbairavniamrofscitsitatselpmaS htworgeuneveR.AORfonoitaiveddradnatsehtybdedividoitartessA-latipaCsulpAORehtsierocS-z.ylevitcepser,selbairavcfiiceps-rotcesdna .seitilibaildiuqilotstessadiuqilfooitarsiytidiuqiL .srallod.S.UfosnoillibnistessalatotsieziS .seuneverssorgfoetarhtworglaunnaehtsi s’ydooM.RSFBs’ydooMdnagnitartisopedycnerrucngierofs’ydooMneewtebsehctonniecnereffidehtsi)sehctongnitarni(troppuss’ydooM .)1102(.latepporGnisatroppusfoseitilibaborpotni)sehctongnitarni(troppuss’ydooMfonoisrevnocehtsi)ytilibaborpni(troppus 0102–9002 4002–3002 .ved .dtS naideM naeM N .ved .dtS naideM naeM N selbairavcfiiceps-knaB 49.0 71.3 20.3 123 41.1 14.3 03.3 682 erocS-z 44.2 65.0 66.0 233 30.1 18.0 99.0 313 )%ni(stessAnonruteR 49.6 83.0 60.1 323 76.1 72.0 86.0 882 stessAnonruteR .veD .dtS 88.51 34.7 92.8 233 49.5 84.6 35.7 313 )%ni(oitaRstessAotytiuqE 97.0 00.0 11.0- 752 )desab-tekram(erocS-z 50.0 10.0 20.0 062 nruterkcotsegarevA 80.0 11.0 21.0 062 ytilitalovnruterkcotS 79.0 25.0 18.0 423 65.0 23.0 84.0 113 )%ni(snoisivorPssoLnaoL 07.4 11.0 70.0- 923 00.1 81.0 02.0 113 htworgeuneveR 0.654 3.94 0.522 233 6.042 5.23 6.821 313 )snoillib$ni(eziS 32.56 90.02 69.23 233 50.02 32.11 20.81 413 ytidiuqiL 46.2 00.2 66.1 843 06.2 00.0 30.1 133 )sehctonsgnitarni(troppuss’ydooM 63.0 04.0 24.0 843 63.0 00.0 62.0 133 )ytilibaborpni(troppuss’ydooM 73.0 16.0 55.0 502 93.0 26.0 45.0 731 )ytilibaborpni(troppushctiF selbairavcfiiceps-yrtnuoC 29.0 21.01 28.9 16 20.1 88.9 15.9 16 emocniatipacreP 33.7 59.1 13.1 26 91.6 42.3 97.4 26 noitaflnI 2.4 37.1 25.3 26 5.4 28.1 47.3 26 ytilitalovnoitaflnI 10.0 80.0 90.0 36 10.0 80.0 90.0 45 stnemeriuqerlatipaC 5.1 26.5 47.5 36 5.1 26.5 47.5 36 xedninoitcetorprotsevnI 73.0 1 48.0 36 24.0 1 87.0 36 ecnarusnitisopeD 43.35 06 94.96 36 43.35 06 94.96 36 ecrofnE 28.22 38.14 32.24 75 8.42 62.53 43 45 sthgirwoflhsaC 12.0 11.0 71.0 75 22.0 00.0 41.0 45 pihsrenwotnemnrevoG 91.0 90.0 41.0 75 51.0 00.0 1.0 45 pihsrenwolanoitutitsnI 71.0 00.0 60.0 75 31.0 00.0 50.0 45 pihsrenwolaudividnI selbairavcfiiceps-rotcesdnayrtnuoC 72.0 63.0 83.0 771 92.0 72.0 63.0 771 xednilhadnfireH 35

Table3: Correlations Correlations among main variables of interest. z-Score is the ROA plus Capital-Asset ratio divided by the standard error of ROA (in logs). Revenue growth is the annual growth rate of gross revenues. Size is the logarithm of total assets. Liquidityisratioofliquidassetstoliquidliabilities. Moody’ssupport(inratingnotches)isthedifferencein notchesbetweenMoody’sforeigncurrencydepositratingandMoody’sBFSR.Moody’ssupport(inprobability)isthe conversionofMoody’ssupport(inratingnotches)intoprobabilitiesofsupportasinGroppetal.(2011). Moody’s Moody’s Fitch Size Liquidity support support support (notches) (probability) (probability) PanelA:2003–2004 Size 1 Liquidity -0.034 1 Moody’ssupport(notches) 0.153*** -0.003 1 Moody’ssupport(probability) 0.030 0.007 0.843*** 1 Fitchsupport(probability) 0.289*** 0.030 0.439*** 0.270*** 1 PanelB:2009–2010 Size 1 Liquidity -0.019 1 Moody’ssupport(notches) 0.306*** 0.023 1 Moody’ssupport(probability) 0.413*** 0.097* 0.714*** 1 Fitchsupport(probability) 0.184*** 0.027 0.521*** 0.371*** 1 Robuststandarderrorsinbrackets: *p<0.10,**p<0.05,***p<0.01 36

Table4: BankRiskTakingandGovernmentSupport(Notches): z-Scores Dependentvariableforallcross-sectionregressionsisthenaturallogarithmofeachbank’sindividualz-Score.z-Score isROAplusCapital-AssetratiodividedbythestandarddeviationofROA.Revenuegrowthistheannualgrowthrateof grossrevenues. Sizeisthelogarithmoftotalassets. Liquidityisratioofliquidassetstoliquidliabilities. Government supportisthedifferenceinnotchesbetweenMoody’sforeigncurrencydepositratingandMoody’sBFSR.Standard errorscorrectedforcountry-levelclustering. 2003–2004 (1) (2) (3) (4) (5) (6) (7) Variables Bank Instrumental Fixed Country Ownership Support controls variables effects controls controls in2001 Governmentsupport -0.048 -0.068 0.003 -0.133*** -0.127*** -0.123*** -0.101*** [0.052] [0.051] [0.094] [0.034] [0.028] [0.031] [0.036] Revenuegrowth 0.101 0.227 0.292 0.686** 0.806*** 0.264 0.263 [0.668] [0.600] [0.607] [0.329] [0.294] [0.556] [0.578] Size 0.156** 0.127* -0.008 0.029 0.033 0.035 [0.064] [0.069] [0.053] [0.044] [0.076] [0.084] Liquidity -0.011*** -0.011*** -0.017*** -0.008** -0.008** -0.008** [0.003] [0.004] [0.003] [0.004] [0.003] [0.004] Percapitaincome 0.691*** 0.632*** 0.658*** [0.176] [0.178] [0.180] Inflation -0.002 0.003 0.004 [0.021] [0.034] [0.036] Inflationvolatility -0.131*** -0.139** -0.144** [0.041] [0.060] [0.065] Capitalrequirements 22.177** 15.804 13.367 [8.745] [11.320] [11.695] Investorprotectionindex -0.017 -0.013 0.020 [0.079] [0.090] [0.091] Depositinsurance -0.427** -0.395** -0.410** [0.167] [0.193] [0.198] Enforce 0.003** 0.004* 0.005** [0.001] [0.002] [0.002] Herfindahlindex -0.379 -0.226 0.000 [0.247] [0.339] [0.361] Cashflowrights -0.001 -0.002 [0.003] [0.003] Governmentownership 0.416 0.528 [0.312] [0.345] Institutionalownership 0.152 0.323 [0.247] [0.265] Individualownership 0.442 0.446 [0.408] [0.382] Observations 286 286 275 286 250 183 177 R-squared 0.01 0.1 0.06 0.58 0.38 0.34 0.34 Countries 54 54 44 54 49 44 44 Robuststandarderrorsinbracke3ts7: *p<0.10,**p<0.05,***p<0.01

Table4(continued). BankRiskTakingandgovernmentsupport(notches): z-Scores 2009–2010 (8) (9) (10) (11) (12) (13) (14) Variables Bank Instrumental Fixed Country Ownership Support controls variables effects controls controls in2007 Governmentsupport -0.081*** -0.080*** -0.082** -0.134*** -0.079*** -0.068** -0.046* [0.030] [0.028] [0.038] [0.037] [0.028] [0.027] [0.027] Revenuegrowth 0.017*** 0.018*** 0.018*** 0.013*** 0.018*** 0.018*** 0.018*** [0.005] [0.004] [0.004] [0.004] [0.004] [0.005] [0.005] Size -0.002 -0.004 0.004 -0.011 -0.035 -0.049 [0.043] [0.044] [0.044] [0.043] [0.049] [0.048] Liquidity -0.002** -0.002** -0.001** -0.001* -0.001** -0.001** [0.001] [0.001] [0.000] [0.001] [0.001] [0.001] Percapitaincome -0.341*** -0.366*** -0.376*** [0.112] [0.126] [0.121] Inflation -0.044* -0.043** -0.035* [0.023] [0.021] [0.020] Inflationvolatility -0.071* -0.066* -0.054 [0.041] [0.037] [0.035] Capitalrequirements -4.927 -6.712 -7.757 [8.608] [8.423] [8.262] Investorprotectionindex -0.002 -0.013 -0.008 [0.052] [0.050] [0.051] Depositinsurance -0.183 -0.168 -0.095 [0.212] [0.209] [0.200] Enforce 0.000 -0.000 -0.000 [0.002] [0.002] [0.002] Herfindahlindex -0.115 -0.150 -0.135 [0.248] [0.274] [0.284] Cashflowrights -0.005** -0.005** [0.002] [0.002] Governmentownership 0.052 0.044 [0.169] [0.169] Institutionalownership 0.359** 0.335** [0.146] [0.147] Individualownership -0.266 -0.267 [0.172] [0.170] Observations 321 320 310 320 317 305 302 R-squared 0.06 0.07 0.08 0.4 0.13 0.17 0.17 Countries 54 54 48 54 53 53 53 Robuststandarderrorsinbrackets: *p<0.10,**p<0.05,***p<0.01 38

Table5: BankRiskTakingandGovernmentSupport(Notches): z-ScoreComponents Dependentvariableforeachregressiondefinedattopofeachcolumn. Revenuegrowthistheannualgrowthrateof grossrevenues. Sizeisthelogarithmoftotalassets. Liquidityisratioofliquidassetstoliquidliabilities. Government supportisthedifferenceinnotchesbetweenMoody’sforeigncurrencydepositratingandMoody’sBFSR.Standard errorscorrectedforcountry-levelclustering. (1) (2) (3) (4) (5) (6) 2003–2004 2009–2010 Variables ROA Std. ROA Equity ROA Std. ROA Equity /Assets /Assets Governmentsupport -0.087*** 0.050** -0.075 -0.119** -0.001 -0.495*** [0.027] [0.020] [0.094] [0.054] [0.026] [0.152] Revenuegrowth 0.049 0.838 -3.334** 0.006** -0.002 0.014 [0.147] [0.515] [1.416] [0.003] [0.004] [0.027] Size -0.173** -0.027 -1.625*** -0.150* -0.170*** -1.972*** [0.076] [0.036] [0.524] [0.083] [0.045] [0.516] Liquidity 0.001 -0.001 -0.014 0.011** 0.009** 0.079*** [0.004] [0.003] [0.014] [0.004] [0.003] [0.015] Percapitaincome 0.129 -0.578*** 0.883* -0.382** 0.154 0.047 [0.143] [0.163] [0.510] [0.160] [0.101] [0.640] Inflation 0.04 -0.033 0.178 0.006 -0.002 -0.092 [0.026] [0.023] [0.119] [0.034] [0.017] [0.103] Inflationvolatility -0.002 0.204*** 0.114 0.066 0.029 0.113 [0.065] [0.048] [0.211] [0.056] [0.033] [0.232] Capitalrequirements -3.916 -7.469 79.866** -18.600* 3.39 -29.972 [14.395] [10.929] [38.889] [9.722] [6.480] [34.843] Investorprotectionindex 0.041 -0.065 0.083 -0.013 0.104* 0.467 [0.060] [0.050] [0.196] [0.127] [0.055] [0.412] Depositinsurance -0.638* 0.299 -0.794 -0.577 0.07 -1.364 [0.318] [0.211] [1.040] [0.345] [0.226] [1.447] Enforce 0.002 -0.003* -0.002 -0.002 -0.001 -0.01 [0.002] [0.002] [0.007] [0.002] [0.001] [0.009] Herfindahlindex 0.42 0.261 1.799 0.138 -0.028 0.878 [0.404] [0.338] [1.610] [0.347] [0.280] [2.148] Cashflowrights -0.002 -0.001 -0.025** -0.004 0.003 -0.01 [0.002] [0.002] [0.010] [0.003] [0.002] [0.011] Governmentownership 0.288 -0.129 1.818 -0.014 0.402 1.336 [0.230] [0.181] [1.226] [0.460] [0.269] [1.376] Institutionalownership -0.006 -0.159 2.229* 0.218 -0.035 1.659 [0.176] [0.156] [1.280] [0.177] [0.100] [1.459] Individualownership 0.739* -0.307 1.776 0.677** 0.711** 4.105* [0.418] [0.328] [1.365] [0.279] [0.308] [2.357] Observations 198 183 198 312 306 312 R-squared 0.43 0.55 0.65 0.32 0.41 0.61 Countries 45 44 45 53 53 53 Robuststandarderrorsinbrackets: *p<0.10,**p<0.05,***p<0.01 39

Table 6: Bank Risk Taking and Government Support: Panel Regressions and Market-Based Z- Scores This table presents panel regressions of bank risk taking on government support and a set of controls, as well as bank-levelfixedeffects. Dependentvariableforallpanelregressionsisthenaturallogarithmofeachbank’sindividual market-basedz-Score. Themarket-basedz-ScoreismarketROEplusCapitaltoMarketValueofEquityratiodivided by the standard deviation of market ROE. Government support is measured as probability of bailout (Gropp et al., 2011)usingdatafromMoody’s. Odd-numberedcolumnsusegovernmentsupportlaggedbyoneperiodwhileevennumbered columns use government support lagged by two years. LLP stands for Loan Loss Provisions. A lower z-score implies higher risk taking while the opposite is true for Stock Return Volatitility and LLP. Standard errors correctedforbank-levelclustering. (1) (2) (3) (4) (5) (6) VARIABLES Marketz-score Stockreturnvolatility LLPoveravg. assets Governmentsupport -0.087*** -0.097*** 0.010*** 0.008*** 0.059*** 0.024 [0.027] [0.031] [0.003] [0.003] [0.020] [0.017] Revenuegrowth 0.057* 0.070** -0.003 -0.004* -0.001 -0.002 [0.034] [0.033] [0.002] [0.002] [0.001] [0.001] Size -0.372** -0.222 0.000 -0.007 0.150 0.088 [0.150] [0.217] [0.021] [0.029] [0.094] [0.124] Liquidity 0.001* 0.001 -0.000 0.000 -0.003** -0.005** [0.001] [0.001] [0.000] [0.000] [0.002] [0.002] Percapitaincome -1.406*** -1.100* 0.056 0.088 0.559* 0.482 [0.445] [0.590] [0.054] [0.059] [0.335] [0.475] Inflation 0.004 0.006 -0.002*** -0.002*** -0.017*** -0.017*** [0.005] [0.006] [0.000] [0.000] [0.004] [0.004] Inflationvolatility 0.045** 0.064*** -0.005** -0.006*** 0.019** 0.029*** [0.019] [0.020] [0.002] [0.002] [0.009] [0.010] Herfindahlindex 0.108 0.336 -0.009 0.003 -0.403*** -0.396*** [0.302] [0.315] [0.028] [0.017] [0.142] [0.152] Cashflowrights -0.000 -0.003 0.000 0.001*** 0.005 0.005 [0.003] [0.003] [0.000] [0.000] [0.005] [0.006] Observations 1,169 1,006 1,283 1,117 1,684 1,436 R-squared 0.07 0.05 0.02 0.03 0.09 0.07 Numberofbanks 280 276 281 278 358 343 Governmentsupport(t) 1yearlag 2yearlag 1yearlag 2yearlag 1yearlag 2yearlag 40

Table7: BankRiskTakingandProbabilityofGovernmentSupportMeasuredbyMoody’s Dependent variable for all cross-section regressions is the natural logarithm of each bank’s individual z-Score. The z-ScoreisROAplusCapital-AssetratiodividedbythestandarddeviationofROA.Governmentsupportismeasured as probability of bailout (Gropp et al., 2011) using data from Moody’s. Standard errors corrected for country-level clustering. 2003–2004 2009–2010 (1) (2) (3) (4) (5) (6) (7) (8) Variables Bank Fixed Country Bank Fixed Country controls effects controls controls effects controls Governmentsupport -0.468* -0.499** -0.522** -0.494*** -0.448*** -0.434*** -0.474** -0.314** [0.245] [0.214] [0.222] [0.182] [0.160] [0.154] [0.192] [0.147] Revenuegrowth 0.127 0.252 0.620* 0.796** 0.019*** 0.019*** 0.012** 0.019*** [0.678] [0.610] [0.343] [0.314] [0.005] [0.005] [0.005] [0.004] Size 0.144** -0.01 0.031 0.004 0.039 0.006 [0.065] [0.058] [0.047] [0.045] [0.049] [0.042] Liquidity -0.011*** -0.018*** -0.009** -0.002** -0.001* -0.001 [0.003] [0.003] [0.003] [0.001] [0.000] [0.001] Percapitaincome 0.608*** -0.353*** [0.176] [0.105] Inflation 0.02 -0.023 [0.019] [0.020] Inflationvolatility -0.133*** -0.035 [0.041] [0.038] Capitalrequirements 19.277** -4.757 [9.304] [8.684] Investorprotection -0.006 0.003 [0.080] [0.056] Depositinsurance -0.518*** 0.025 [0.182] [0.188] Enforce 0.003* 0.0000 [0.001] [0.002] Herfindahlindex -0.345 -0.107 [0.253] [0.249] Observations 286 286 286 250 321 320 320 317 R-squared 0.02 0.1 0.56 0.35 0.04 0.05 0.39 0.11 Countries 54 54 54 49 54 54 54 53 Robuststandarderrorsinbrackets: *p<0.10,**p<0.05,***p<0.01 41

Table8: BankRiskTakingandProbabilityofGovernmentSupportMeasuredbyFitchRatings Dependent variable for all cross-section regressions is the natural logarithm of each bank’s individual z-Score. The z-ScoreisROAplusCapital-AssetratiodividedbythestandarddeviationofROA.Governmentsupportismeasured asprobabilityofbailout(Groppetal.,2011)usingdatafromFitchRatings.Standarderrorscorrectedforcountry-level clustering. 2003–2004 2009–2010 (1) (2) (3) (4) (5) (6) (7) (8) Variables Bank Fixed Country Bank Fixed Country controls effects controls controls effects controls Governmentsupport -0.213 -0.372* -0.260 -0.092 -0.274 -0.234 -0.458*** -0.277* [0.240] [0.198] [0.172] [0.195] [0.199] [0.174] [0.159] [0.165] Revenuegrowth -0.056 0.076 0.992 1.110* 0.024*** 0.024*** 0.010** 0.023*** [0.765] [0.714] [0.710] [0.568] [0.004] [0.004] [0.004] [0.005] Size 0.107 0.014 0.138** -0.011 0.075 -0.013 [0.073] [0.105] [0.058] [0.041] [0.054] [0.038] Liquidity -0.006** -0.005** -0.015*** -0.004 -0.001 -0.003 [0.002] [0.002] [0.004] [0.004] [0.001] [0.003] Percapitaincome 0.639*** -0.375*** [0.170] [0.128] Inflation 0.017 -0.030 [0.019] [0.034] Inflationvolatility -0.082*** -0.061 [0.024] [0.061] Capitalrequirements 26.458*** 6.139 [8.456] [8.006] Investorprotectionindex 0.002 -0.011 [0.103] [0.081] Depositinsurance -0.363** -0.084 [0.176] [0.203] Enforce 0.004* 0.000 [0.002] [0.002] Herfindahlindex 0.446 0.147 [0.390] [0.418] Observations 175 175 175 127 269 268 268 261 R-squared 0.01 0.06 0.57 0.38 0.03 0.05 0.44 0.11 Countries 43 43 43 39 50 50 50 49 Robuststandarderrorsinbrackets: *p<0.10,**p<0.05,***p<0.01 42

Table9: BankRisk,Valuation,andGovernmentSupport z-ScoreisROAplusCapital-AssetratiodividedbythestandarddeviationofROA.Tobin’sQistotalassetsplusmarket valueofequityminusbookvalueofequitydividedbytotalassets. Governmentsupportisthedifferenceinnotches between Moody’s foreign currency deposit rating and Moody’s BFSR. Widely takes value one if there is no single shareholderwithatleast25%ofthevotingsharesandzerootherwise. Numberofbanksisthenumberofbanksinthe countrydividedbythecountry’sGDPinU.S.dollars. Standarderrorscorrectedforcountry-levelclustering. Secondstage Firststage Dependentvariable z-Score Tobin’sQ Tobin’sQ -1.078 [3.258] Governmentsupport -0.062 * -0.006 * [0.036] [0.003] Revenuegrowth -0.006 0.002 [0.098] [0.004] Size 0.015 0.005 [0.047] [0.005] Liquidity -0.001 0.000 * [0.001] [0.000] Percapitaincome -0.327 * -0.011 [0.140] [0.010] Inflation -0.025 0.001 [0.018] [0.002] Inflationvolatility -0.041 0.011 * [0.050] [0.003] Capitalrequirements -5.605 0.014 [8.992] [0.921] Investorprotectionindex 0.029 0.006 * [0.060] [0.003] Depositinsurance -0.073 0.017 [0.196] [0.022] Enforce 0.001 0.000 [0.002] [0.000] Herfindahlindex -0.162 0.084 * [0.332] [0.050] Widely -0.019 [0.013] NumberofBanks -0.047 * [0.016] Observations 244 244 Hansen’sJstatisticforover-identification 1.364 Angrist-PischkemultivariateFtestofexcludedinstruments 3.84 ** Robuststandarderrorsinbrackets”: *p<0.10,**p<0.05,***p<0.01 43

Table10: GovernmentInterventionsinBanksin2008–2010inEurope,byCountry Thistableshowsthenumberofintervenedandnotintervenedbanksforwhichthereisratingsdataavailable. Y = 1 i ifbankiwasintervenedand0otherwise. DataisfromBreietal.(2011),LaevenandValencia(2012),andFT.com. Country Notintervened Intervened Total Austria 3 4 7 Belgium 1 2 3 Denmark 3 3 6 Finland 1 0 1 France 5 4 9 Germany 19 3 22 Greece 1 6 7 Iceland 0 3 3 Ireland 1 4 5 Italy 15 6 21 Netherlands 5 2 7 Norway 6 0 6 Portugal 2 0 2 Spain 9 2 11 Sweden 4 1 5 Switzerland 7 1 8 UnitedKingdom 11 3 14 Total 93 44 137 44

Table11: GovernmentSupportandInterventioninBanks This table shows the results of a probit regression where the limited dependent variable is Y = 1 if bank i was i intervenedand0otherwise. DataoninterventionsisfromBreietal.(2011),LaevenandValencia(2012),andFT.com. Sizeisthelogarithmoftotalassets. Liquidityisratioofliquidassetstoliquidliabilities. Outputgrowthistheaverage GDP growth in 2007–2009. Loan loss provisions is expressed as percentage of total average assets. Government support is 1 if the difference in notches between Moody’s foreign currency deposit rating and Moody’s BFSR is positiveandzerootherwise. Standarderrorscorrectedforcountry-levelclustering. Variables Intervention Size 0.327*** [0.112] Liquidity -0.000 [0.005] Equitytoassets 1.630 [4.933] Outputgrowth -0.141 [0.175] Governmentsupportin2007 22.694* [13.423] Loanlossprovisions 3,865.50 [2,351.445] Supportin2007XLoanlossprovisions -3,798.37 [2,370.847] Observations 123 Countries 17 Marginaleffectofgovernmentsupport 0.273*** [0.0735] Robuststandarderrorsinbrackets: ***p<0.01,**p<0.05,*p<0.1 45

noitalugeRdnanoisivrepuSknaB,troppuStnemnrevoG,gnikaTksiRknaB :21elbaT morf(noisivrepusdnanoitalugergniknablevel-yrtnuocrofselbairavdna,slortnoccfiiceps-lortnoc,cfiiceps-knabhtiwsnoissergernoitces-ssorC tessA-latipaCsulpAORehtsierocS-z .)1102,.latepporG(tuoliabfoytilibaborpsaderusaemsitroppustnemnrevoG .)8002,6002,.latehtraB .gniretsulclevel-yrtnuocrofdetcerrocsrorredradnatS.AORfonoitaiveddradnatsehtybdedividoitar 0102–9002 4002–3002 )8( )7( )6( )5( )4( )3( )2( )1( llA ytivitcA -repuslaicfifO latipaC llA ytivitcA -repuslaicfifO latipaC selbairaV snoitcirtser srewopyrosiv ycnegnirts snoitcirtser srewopyrosiv ycnegnirts **184.0- ***234.0- *373.0- 710.0- 390.0- 811.0- 833.0- 130.0 troppustnemnrevoG ]381.0[ ]280.0[ ]681.0[ ]301.0[ ]314.0[ ]512.0[ ]103.0[ ]380.0[ 220.0 000.0- 610.0- 020.0ycnegnirtslatipaC ]050.0[ ]140.0[ ]880.0[ ]780.0[ 200.0- 310.0- 830.0- 930.0latipaCxtroppuS ]020.0[ ]720.0[ ]320.0[ ]420.0[ ycnegnirts 310.0- 620.0- 340.0 620.0 yrosivrepuslaicfifO ]540.0[ ]140.0[ ]270.0[ ]960.0[ rewop 600.0 *920.0 510.0 020.0 laicfifOxtroppuS ]410.0[ ]610.0[ ]520.0[ ]420.0[ rewopyrosivrepus 730.0 430.0 160.0- 750.0snoitcirtserytivitcA ]820.0[ ]320.0[ ]480.0[ ]470.0[ ***330.0 ***530.0 400.0- 200.0 ytivitcAxtroppuS ]900.0[ ]700.0[ ]220.0[ ]320.0[ snoitcirtser 642 642 642 642 841 841 841 841 snoitavresbO 62.0 52.0 71.0 51.0 54.0 24.0 34.0 34.0 derauqs-R seY seY seY seY seY seY seY seY slortnocknaB seY seY seY seY seY seY seY seY slortnocyrtnuoC 84 84 84 84 14 14 14 14 seirtnuoC 10.0<p***,50.0<p**,01.0<p* :stekcarbnisrorredradnatstsuboR 46