BANK RISK-TAKING AND CAPITAL REQUIREMENTS

Transcription

1 BANK RISK-TAKING AND CAPITAL REQUIREMENTS Rebeca Anguren Gabriel Jiménez * February 2017 Abstract In this paper we empirically investigate the effect of the increase in regulatory capital requirements on bank lending and, particularly, on the banks appetite for risk. We exploit two unique datasets of loan applications to new banks and granted loans to nonfinancial firms from Spain which allows us to control for time-varying observable and unobservable firm heterogeneity and, at the same time, for time-varying observable and unobservable bank supply. Our analysis shows that the increase in bank capital requirements established as a result of the recent financial crisis implies a reduction on the supply of bank credit in the short run. We also observe that this transitory reduction is lower for riskier borrowers. These results are indicative of a change in risk-taking strategies by more affected banks. Nevertheless, concurrently to these effects, banks were hedging the greater risk assumed by reducing the amount lent for each new loan granted and requiring increased guarantees for these borrowers. JEL Codes: G01; G28; G21. Keywords: bank capital; credit supply; risk-taking; Basel III. * Corresponding author: Gabriel Jiménez, Banco de España; gabriel.jimenez@bde.es. Rebeca Anguren, Banco de España; rebeca.anguren@bde.es. This paper reflects our views and not the views of Banco de España or the Eurosystem. 1

2 1. Introduction Beginning in 2008, Spain and Western economies in general had suffered one of their worst banking crises, which led to a lasting economic recession affecting all levels of the real economy (Jiménez et al, 2017). These spillovers to the real economy were due to banks deleveraging, among other reasons. The response by the banking sector to the financial shock was the consequence of the weaknesses in the solvency and liquidity position of banks which were not prepared to absorb a shock of that dimension. As a response, authorities decided to review the prudential regulatory framework to ensure the soundness of the banking system and the stability of the flow of credit to the nonfinancial sector. Reforms strengthened multiple dimensions of the regulatory framework including the level and quality of capital requirements. This paper analyzes the effects of increased capital requirements on banks behavior by focusing on changes in their credit supply (i.e. whether a deleveraging may be expected after regulatory changes are applied) and, in particular, on their risk-taking profiles (i.e. whether a search for yield takes place as an attempt to boost profitability to accumulate capital through retained earnings and also maintain constant ROE). We exploit two complementary unique datasets. The first database contains nonfinancial firms loan applications from Spain which allows us to control for timevarying observable and unobservable demand and supply factors. Here, we perform two analysis. First, we investigate the impact of increased capital requirements for Spanish banks from December 2011 on the risk-taking effect under the extensive margin in the short run. That is, on the probability of loan applications to be approved depending on the risk of the applying firm with non-current banks (i.e. banks that are not currently lending to the firm). Second, we analyze the impact of this increase in capital 2

3 requirements on loan characteristics (amount, collateral and maturity) of granted applications. The second dataset is a credit register of all granted loans to non-financials Spanish firms. Here, we analyze the intensive margin of bank loans to firms, which is the impact on the size of each individual granted loan. Again, we will focus the analysis on the risk-taking behavior of banks but focusing the question on the effect of the introduction of Basel III. Again, this database allows us to control for observable and unobservable demand and supply factors. To address our objective of analyzing the impact of capital requirements on credit supply there are three identification challenges. The first challenge is to disentangle credit demand from credit supply. Our hypothesis states that banks would react to higher capital requirements by reducing their supply or by taking more risk. Given that the quality of borrowers depends on other factors such as the economic cycle and/or the monetary policy, it is necessary to control for the quality of the demand to isolate the supply channel. In turn, there may be a problem of selection if banks with greater weakness in their balance sheets are linked more to more vulnerable companies. For instance, less capitalized banks may have a more risky credit policy, having led them to work with poorer creditors on whom a worsening economic situation has a greater impact, further reducing their demand. The second challenge is to separate changes in credit supply derived from an exogenous shock from endogenous movements resulting from the need to increase capital in order to augment the amounts lent and risks assumed. Previous literature has tried to address both challenges. Uluc and Wiedalek (2015) use loan-level data on all new mortgages issued in UK during , controlling for loan demand effects using geographical time dummies and controlling for supply characteristics using bank controls and time-invariant bank fixed effects. These authors exploit the regulatory regime applied by the Financial Services Authority 3

4 (FSA) which resulted in exogenous shocks as banks capital requirements. These requirements were adjusted over time and individually to address risks not covered by the prudential regulatory framework. Auer and Ongena (2016) use credit register data from the Swiss National Bank (SNB) to capture loan demand through the use of business-type fixed effects based on multiple characteristics of each firm (instead of individual fixed effects). In this case the focus of the analysis is the activation of the Switzerland s countercyclical capital buffer (CCB) in 2012 which introduced an exogenous shock to capital requirements affecting differently banks depending on their share of residential RWAs to total assets. Finally, the third challenge is to be able to have a measure of enterprise risk that synthesizes the probability of future default. Our paper contributes to this literature by first using a specific database of loan level dataset of applications and granted loans, which allows us to perfectly separate demand and supply effects, and second by analyzing the impact of the exogenous shock to capital requirements resulting from the increase in capital requirements after the last financial crisis, paying particular attention to the introduction of Basel III capital requirements. Moreover, we use an industry-based scoring function as an instrument of the risk quality of a firm, which includes key financial and non-financial variables of the firms and can be seen as a sufficient statistic of the one-year likelihood of distress. Our identification exploits two different but complementary strategies. First, we study bank risk-taking using loan applications made by non-financial firms to noncurrent banks (first time loan applications) extensive margin. We employ a panel database that collects monthly information on the resolution of loan applications from 2007 to Here, the exogenous variation comes from capital increases required to Spanish banks after December of In particular, Basel III was announced in 4

5 December 2010 and reviewed in June 2011; European authorities introduced additional ad-hoc capital requirements; and Spanish authorities applied multiple measures to reinforce the stability of the banking sector (see Section 2.1 for a detailed explanation). Therefore we have multiple shocks to capital requirements starting from Our strategy relies on the saturation of the model with large sets of fixed effects to address first, the concurrent changes in credit demand and credit supply; and, second, the endogeneity issue which comes from other sources of bank-time variability which are contemporaneous with the banks behavior against tighter capital requirements (for example, a general change in credit policies of the bank due to a new business strategy). Therefore, the benchmark specification in this strategy includes firm*time fixed effects to control for observed and unobserved time-varying firm heterogeneity, capturing credit demand. This procedure analyzes the probability of getting approval of a loan request by a given firm, exploiting the fact that many firms made more than one loan application to different banks in the same month (Khwaja and Mian (2008), Jiménez et al (2014)). Moreover, given that we are interested in analyzing time-varying changes in bank capital and credit standards, our benchmark model also includes bank*time fixed effects and, in some specifications, industry*province*bank*time fixed effects, to control for other concurrent changes in the supply of credit or in the risk taking of banks, in order to reduce as much as possible the endogeneity concern related to the existence of multiple shocks to capital requirements and to better identify banks riskshifting. We also check the sensitivity of our estimates to unobserved heterogeneity assessing the change in our main coefficients when more controls such as, for example, more dummies or interest rates are included in the regression (following Altonji et al (2005) and Oster (2014)). 5

6 The second strategy completely changes the previous point of view. Under this approach we analyze the change in loan amounts for the same non-financial firm with their relationship banks before and after Basel III intensive margin exploiting the cross-sectional differences in banks capital before the introduction of the new solvency rules. We estimate a cross-sectional model that compares the gap between the capital ratio of each bank with respect to Basel III capital requirements. This gap is calculated in December 2011, well in advance of the compliance date for the new capital rules. In particular, we calculate a proxy of the leverage ratio based on the proposed capital rules, and we compare it with the actual requirement that will apply at the end of the transition period. This gap or shortage of capital is computed two years before the publication of the transposition of Basel III in the EU and more than five years before the date established for full compliance of Basel III. Under this strategy the exogeneity comes from the new capital rules. Again, demand effects are removed through the use of unobserved firm fixed effects; and bank fixed effects; and industry*province*bank fixed effects remove bank unobservable time-invariant components and account for the possible endogenous matching between firms and lenders. Our results show that, first, that there is a negative effect on the credit supply both in terms of the extensive margin (that is the willingness of banks to provide loans to a new non-financial firm) and in terms of the intensive margin (that is, the amount granted by banks for each firm). Additionally, we show that there is a risk-shifting effect in banks behavior. In particular, those banks that were increasing more their capital ratios reduced the number of applications approved with less intensity for firms with higher risk. At the same time, banks were also reducing the loan amount granted individually to riskier firms and requiring increased collateralization as a way to hedge the greater risk assumed under these transactions. We also show that the risk-taking 6

7 effect depends on monetary policy, as banks were more prone to grant loans to riskier borrowers when interest rates decreased. Given the timeframe used in these analysis ( ) the paper captures the short run response of banks to new capital requirements. These effects of capital requirements on credit supply, including the amount and composition of credit, are the result of a failure of the Modigliani-Miller Theorem and the fact that capital requirements are a binding constraint for banks. For example, if the cost of issuing capital if higher than the cost of issuing debt (e.g. due to a different tax treatment between debt and equity), then an increase in capital requirements ends up with an increase in the cost of financing for the banks which could affect their willingness to provide credit and also alter their risk-taking behavior. Thus, in the short-term, a deleveraging to reduce capital requirements and a search for yield with a view to improve profitability and keep a constant ROE could take place under these conditions. Some studies have already demonstrated these short-run effects. Regarding deleveraging, multiple studies have shown a negative impact of increased capital requirements on credit supply in the short-run, both from a micro and a macro perspective (Bridges et al, 2014, MAG, 2010, Noss and Toffano, 2014, and Uluc and Wiedalek, 2015), although none of them are able to isolate the supply channel. Regarding risk-taking the literature is scarce. For instance, Uluc and Wieladek (2015) analyze the impact of capital requirements on mortgage loan size and risk-shifting behavior for the UK banking sector, finding that riskier borrowers are not affected by the reduction in the loan size by affected banks. However, as they comment to fully separate loan supply form demand, data on loan applications and outcomes is necessary, something unavailable for them but not for us. Jiménez et al (2015) find that banks with higher provision requirements focus their credit supply to firms with a 7

8 higher ex-ante interest paid and leverage, and with higher ex-post default, which suggests that an increase in capital requirements may increase bank risk-taking and search for yield, although the authors do not analyze a true shock on capital requirements, they do not study the extensive margin using loan applications and they are not able to approach the risk of the firm through a market-used tool as we have. Auer and Ongena (2016) study compositional effects on banks credit supply derived from the application of the countercyclical capital buffer in Switzerland which was targeted to the residential real estate sector. Among other results, they find that banks shifted lending to riskier and smaller firms (and, at the same time, increased both the interest rate and their one-time commissions). However, they do not have a comprehensive database such the one we have to control for demand using firm fixed effects. Moreover, they use proxies to try to capture the credit risk of a firm, instead of using a scoring function. Additionally, from a different perspective and more focused on a different shock, Jiménez et al (2014) find that expansive monetary policy increased risk-taking in bank lending. We contribute to the literature in different ways. First, we are able to exploit a dataset that allows to controls for demand effects to identify loan supply. Second, we use this data to study compositional effects on credit supply of a capital requirement shock. Third, we have at our disposal an external measure of the credit risk of the firms used by the market. We use this scoring system as a sufficient statistic of the solvency of a firm and it allow us to check whether banks react to a new capital requirement by increasing their exposure to riskier borrowers. This is much better that the proxies for firm risk used in the empirical literature quoted above. These short-run effects, which could be accompanied by an increase in lending rates to compensate the higher cost of capital, have a negative impact on real activity. But, 8

9 importantly, apart from these short-run effects, there are relevant economic benefits of increased capital requirements in the long-run. Under the new steady state one could expect an improvement in the access to market funding by banks and a reduction in the cost of this funding given their strengthened solvency position. Additionally, increased capital requirements reinforce the banking system thus reducing bank failures and the cost of banking crises, which helps to ensure a more stable credit supply. Moreover, previously undercapitalized banks will be able to increase their credit supply after improving their solvency positions. These long-term effects counterbalance short-run effects thus resulting in positive effects for the real economy. LEI (2010) showed that the net benefits of increased capital requirements remain positive for a broad range of target capital ratios. Indeed, some studies show that the underlying assumptions could be conservative and net benefits could be even higher (Cecchettti, 2014). Jiménez et al (2015) find that procyclical capital requirements (dynamic provisioning) which reinforce banks solvency position help to mitigate credit crunch during recessions. In conclusion, although in the short-run we may find some deleveraging and risk-shifting behavior by banks, under the new steady state situation the net impact may be positive due to the improvement of the stability of the banking system. The paper is organized as follows. Section 2 describes first the empirical strategy adopted in our analysis together with the database. Under Section 3 we describe the results obtained under both the extensive and intensive margin analysis. Robustness checks are included and described in their respective subsections. Section 4 concludes. 2. Empirical Strategy and Data 2.1 Empirical Strategy 9

10 We follow two different strategies that correspond to the analysis of the two margins: extensive (which studies loan applications to non-current banks - i.e. banks with which the company does not maintain a relationship) and intensive (which studies the credit volume granted by current banks). The econometric analysis of the extensive margin, which measures the likelihood of a loan application to a non-current bank to be granted, is based on a specification of the following form: where I(loan application granted ijt ) is a dummy variable that equals one if the loan application of firm i to bank j in month t is successful and the loan is granted in the period t to t+3, and equals zero otherwise (see Jiménez et al 2012 and 2014); Scoring it-1 is a variable that measures the financial risk of a firm i through a weighted average of some financial characteristics (higher values of the variable indicate higher risk); Capital ratio jt-1 is the annual change in the capital-to-assets ratio of a bank defined as the ratio of bank equity over total assets 1 ; I t is a dummy variable that takes value one for subsequent observations to December of 2011, to try to capture the shocks to requirements faced by banks after that date (see further explanations below); η i are firm fixed effects that control for firm time-invariant observable and unobservable demand factors that impact the probability of obtaining the loan requested; η j are unobserved bank-specific time-invariant bank effects to control for observed and unobserved supply factors that impact the probability of granting the loan; η t are time fixed effects; Other 1 In fact, it is a leverage ratio. In Spain off-balance-sheet activity by banks has been almost nonexistent. 10

11 controls ijt-1 includes linear double interactions terms related to the triple interaction Capital ratio jt-1 *Scoring it-1 *I t ; and ε ijt is the error term. 2 Under this strategy we face a challenge related to the non-existence of a specific isolated shock to capital requirements but rather a set of measures over the period of time considered in our analysis. We have decided to set December 2011 as the starting point from which a series of shocks took place aimed at increasing the capital requirements of Spanish banks. First, Basel III was announced in December 2010 but then reviewed in Afterwards Basel III was transposed into EU law in Second, in 2012 there were two changes in the law in Spain that increased provisions to performing and non-performing loans to construction and real-state firms based on the exposure as of 2011M12, to be accomplished at the end of Third, during the last part of 2012 a stress test exercise was conducted by Oliver Wyman over the Spanish banking system. Fourth, the EBA issued a recommendation in 2011 to raise capital ratios to 9% CET1 by June Fifth, the European Banking Authority has been carrying out a series of periodic annual stress tests to assess the resilience of the EU banks. Therefore, in the analyzed period there were specific changes in capital requirements concurrent with other structural changes related to Basel III. This will imply a possible correlation between bank s capital decisions and bank s risk taking policies. However, we will show that the election of December 2011 as the reference for our analysis is consistent with the results derived from the data and that our results are robust when the same exercise is performed for different cutting points. Additionally, as robustness checks, under the benchmark specification we control for firm*time fixed effects (η it ) to isolate all time-varying unobserved (and observed) 2 The change in capital ratio is winsorized at the 1 st and 99 th percentile. 3 Basel III was transposed in the EU later in June 2013 but many of its requirements did not take place until

12 firm fundamentals, and for bank*time fixed effects (η jt ) and even, as a robustness exercise, for province*industry*bank*time fixed effects (η psjt ) to better identify and separate risk-taking from other supply channels and to reduce potential biases that were derived from the possible endogeneity of the bank's decision (Eq. (2)). Thus, we isolate the bank s decision comparing the supply of credit for the same firm in the same month. Standard errors are corrected for clustering at the level of the firm, bank and time and all variables are lagged one period (or one year in the case of the financial characteristics of the firms, given their annual frequency) to alleviate endogeneity concerns. 4 The equations are estimated using linear probability models (Ordinary Least Squares) because of the large set of fixed effects along several dimensions and the resulting incidental parameters problem in nonlinear maximum likelihood estimation (Abrevaya (1997)) and the difficulties interpreting the effect marginal of the interaction terms. Equations (1) and (2), therefore, try to capture the bank s risk-taking effect derived from capital requirements through three variables and their interactions. First, we have a variable that captures in a single value the financial risk of a firm applicating for the loan. This variable comes from an industry-used scoring system (see Section 2.2 for a detailed explanation). The higher the scoring, the higher the risk associated to that firm. Thus, the coefficient on Scoring it-1 is expected to be negative (α 1 <0). Secondly, we have the change in the ratio of banks' capital, capturing the effort made by the bank to improve its solvency position. The expected sign of the coefficient on Capital ratio jt-1 4 We also used second or third lags and the results obtained are similar. 12

13 may not clear. If the increase in loan supply is part of the credit policy of the bank, a mechanical positive effect is expected if the bank decides to keep its solvency ratio unaltered (α 2 >0). If, on the contrary, there was an external shock that forces banks to put more capital into the game, banks may reduce credit to avoid having to put even more equity. If this is the case a negative coefficient is expected. Under Eqs (1) and (2) we capture this second effect which describes the exogenous movement using the multiplicative term Capital ratio jt-1 *I t ; with a negative expected coefficient (α 3 <0). Finally, to identify the risk-taking effect of new capital requirements, we interact the lagged change in the capital ratio of the bank with the lagged measure of firm risk (scoring) and the time-indicator variable. This triple interaction is our variable of interest given that if banks cut credit to riskier firms (α 1 <0) and a positive coefficient is estimated under the triple interaction (α 4 >0), it would mean that the contraction in loan granting is lower for riskier companies among banks that increased capital in response to new regulatory changes. This would be evidence of risk-taking by banks to meet higher capital requirements. As a reminder, we are working with loan applications to non-current banks. It is worth noting that this approach discards any concern about the possibility that banks were keeping alive riskier firms to avoid their distress and, therefore, the need of increasing provisions to grant these loans. This analysis is complemented with the study of the lending policy of the banks through the study of the characteristics of the loans granted. Our focus now is on the size of the loan, the existence of collateral and the maturity of the loan. We therefore try to check whether more stringent capital requirements trigger a change in the loan terms by estimating the following equation: 13

14 where the following independent variables are included: bank*time(year_month), firm fixed effects and loan controls 5. Standard errors are corrected for multi-clustering at the level of the firm, bank and time. In the second part of our analysis, we also perform a complementary analysis more focused on the impact of Basel III as a way to alleviate all possible endogeneity concerns. In this case the focus is on the credit volume granted by current banks (intensive margin). The differences with the previous analysis do not rely only on the margin investigated but also in the approach followed. Here we perform a difference-indifference analysis where we compare the lending behavior of the same bank with the same firm before and after the Basel III shock, exploiting the heterogeneity of the impact of the policy among banks. In particular, we take advantage of the fact that banks are differently affected by the policy depending on their solvency situation at the end of However, the identification is not complete without controlling for other factors related to bank supply and without taking into account firm heterogeneity, which determines demand. The first point is addressed including bank fixed effects and banklevel characteristics, and the second one with the inclusion of firm fixed effects, capturing observed and unobserved bank and firm fundamentals. Moreover, the strength of the firm-bank relationship is captured with the length of the relationship. Lastly, there can be a selection bias between firms and banks whether the set of firms working with more affected banks are different in some dimension. This selection can be controlled with the inclusion of province*industry*bank fixed effects. The basic cross-sectional 5 Loan controls include the log of the size of the loan and information about the maturity and whether the loan is collateralized. When the dependent variable is among the loan controls it is removed from this subset. 14

15 specification followed to analyze the intensive margin of lending by current banks takes the following expression: where logcommitment ijt is the change of the logarithm of committed credit (drawn plus undrawn credit amount, to reduce demand effects) by bank j to firm i from period 2011Q4 to t (2014Q4 is used as the baseline); Shortage of Capital j is a continuous variable that measures the relative position of bank j in December 2011 to Basel III capital requirements (a more detailed description is provided below); η ips are province*industry fixed effects; Other controls ij includes a comprehensive set of other bank characteristics (such as the log of total assets, the ROA, the liquidity ratio, the solvency ratio, the non-performing loan ratio, the weight of construction and real estate loans over business loans or the weight of firm loans over total loans) and bank-firm relationship characteristics (such as the log of one plus the number of months with the bank or the default status of the firm with the bank); and v ijt is the error term. 6 Standard errors will be clustered at the firm, province, industry and bank level and all bank variables are set in 2011Q4 to alleviate endogeneity concerns. The equation is estimated using Ordinary Least Squares. In Eq. (4), a negative coefficient on Scoring i is expected (β 1 <0) and also a negative relationship between the shortage of capital and the change in bank supply (β 2 <0). The shortage of capital is computed two years before the introduction of Basel III in the EU and more than five years before of full compliance of Basel III by banks. We proxy the 6 The change in credit is winsorized at the 1 st and 99 th percentile. 15

16 Basel III fully loaded ratio in terms of the leverage ratio at 2011 using the fact that the leverage ratio is just the product of the Tier 1 ratio and the density ratio (the riskweighted assets divided by total assets), which denotes the average risk weight per unit of asset): Given that under Basel III banks have to meet a minimum Tier 1 capital ratio of 8.5% (plus an additional buffer for global systemic banks (G-SIBs)) and at the same time a minimum of 3% of the leverage ratio, we use the following definition: Shortage of Capital j = max(3%, (8.5%+1%I(G-SIBs) j )*density ratio j )-Capital ratio j (6) Since the leverage ratio under Basel III rules is not available for the period considered, we use Capital ratio j as a proxy of the leverage ratio which is defined as equity minus goodwill over total assets minus goodwill (this can be considered as a good approximation given that the deduction of goodwill is one of the main adjustments to capital under Basel III and the off balance sheet activity of Spanish banks has been almost non-existent). Our main coefficient of interest is β 3, which measures the relative impact on credit supply of the capital shortage at end 2011 depending on the risk of the firm. If a positive sign is found (β 3 >0), this will imply that less compliant banks (those with higher shortage to regulatory minimum requirements and therefore the ones who have to make a greater effort in the future to comply with Basel III), are those who change their lending practices towards greater risk. 16

17 Our benchmark model includes firm fixed-effects (η i ) to control for demand; and also province*industry*bank fixed effects (η ipsj ), which contains bank fixed effects, to control for alternative sources of supply variability and to better isolate the risk-taking effect: Eq. (7) exploits the fact that many firms have more than one bank relationship by including firm fixed effects as a way to control for unobserved firm characteristics (Khwaja and Mian (2008)). With this technique we are comparing the change in the size of the loan of the same firm with different banks. This is not quite restrictive as multiple bank relationships account for near 90% percent of all business loans. For the intensive margin we also perform an instrumental variable regression. The estimated equation is similar to Eq. (7) but Shortage of Capital j is replaced by Capital ratio jt, the change in the capital ratio from 2011M12 to t. Given the endogeneity of this variable we instrument it using Shortage of Capital j at the end of 2011 as instrument. Therefore, the estimated equation is now: (8) 2.2 Data 17

18 The information about loan applications to non-current banks and the amount granted by current ones is obtained from the Credit Register (CIR) of the Banco de España (which is the owner of this database as the regulator and supervisor of the Spanish banking system). This dataset contains information about all loans above 6,000 euros granted in Spain by all operating banks since 1984 on a monthly basis, which ultimately makes it a real census of banking loans. Simultaneously, since February 2002 the CIR also stores data on loan applications of firms that are not currently borrowing from the requesting bank. When the two datasets are combined it is possible to know whether a loan application is approved by the bank and accepted by the petitioner and the loan is finally granted, and also to know the future performance of the loan (see Jiménez et al 2012 and 2014 for a deeper explanation of these two databases). Information on the characteristics of the firms that are used as controls in some regressions comes from the Commercial Register. Therefore, the economic and financial information of the firms comes from the balance sheets and income statements that Spanish corporations must submit yearly, by law, to the Spanish Mercantile Registers. We are able to merge the CIR datasets with the Commercial Register datasets because both databases contain a unique identifier for each firm, allowing to incorporate new features of the companies to those already own the CIR (province and industry, for instance). Finally, we also match the CIR with a banks dataset containing the balance-sheet and income statement of the Spanish banks at a monthly frequency owned by the Banco de España in its role of banking supervisor and also biannual information reported at consolidated level related to the solvency position of the bank. Given the discrepancy in the periodicities of the databases used, we work with the oneyear lagged information for firms, while we interpolate, at monthly frequency, bank characteristics when needed. 18

19 In the paper we restrict the sample to non-financial small and medium-enterprises (SMEs) 7 (which represents more than 90% of all Spanish companies) and to commercial banks, savings banks and credit cooperatives (which represents around 95% of the total system). We have checked the robustness of the results when credit cooperatives are dropped from the sample, and results don t differ neither economically nor statistically. 8 As a way to exogeneize the financial risk of the firms we use an industry-like credit rating based on key financial and non-financial variables which are used by the private sector during the granting process of a commercial loan to SMEs. This rating function follows the spirit of the classic Z-Score model (Altman, 1968) and uses fifteen financial ratios and firm balance sheet characteristics to assign a score to each company (firm characteristics are related to financial indebtedness, solvency, liquidity, profitability, expertise, firm structure and credit history). The rating system segments each of the variables used in nine classes. Each class will have a value between one and nine, with one being assigned to the lowest risk and nine to the highest risk. The final score is just the weighted sum each of the ratings assigned to the firm characteristics analyzed. So, at the end of the process, each company is associated with a continuous measure ranging from one to nine, where higher values are related to a more likelihood of default 9. For the analysis of the extensive margin we collect monthly data from 2007M12 to 2015M12. Given that we consider January 2012 as the starting point of the regulatory 7 The definition used to characterize SMEs is the one defined in the EU recommendation 2003/ Moreover, we have selected only large-enough commercial and savings banks (above the first quartile of the distribution of total assets) and we find the same results. 9 We have checked the validity of this variable as an ex-ante measure of the credit risk of a firm analyzing whether it is a good predictor of the probability of default for one year ahead of non-defaulted firms. Results show a positive and statistically significant at the 1% coefficient with a F-statistic closes to 140, after controlling for industry, province and time dummies and multi-clustering at the level of the industry (NACE two digits) and of the province (much larger than 10, the rule of thumb for instruments in an IV context), which ensures the strength of this proxy. 19

20 shocks we have a window of five years before and five years after that date. As a robustness test, we have expanded the sample until February 2002 and the risk-taking effect is not statistically different. The sample used in the intensive margin of lending expands quarterly from 2010Q2-2015Q4. We fix the sample in 2011Q4 as the last period before the capital shocks started and then studied the change of the amount granted at the firm-bank level until the end of 2015, and also some quarters before to check whether there is a pre-trend in the data or not. 3. Results 3.1 Extensive Margin Table 1 reports the descriptive statistics of the loan application database. In the Appendix the definition of the variables can be found. Our sample consist of 149,136 loan applications over the period 2007M12 to 2015M09 made by 49,753 firms to 59 banks, including commercial banks, savings banks and credit cooperatives. To make results comparable across different specifications we have ketp constant the sample restricting it to firms and banks with more than one loan application in a moth. On average the 32% of the loan requests were accepted and the loans were granted. This figure hides a step-shaped evolution of the rate, from the level of around 45% before 2007 to the level of 35% once the crisis started. The Scoring variable takes the average value of 3.7 and its standard deviation is In Table 2 we present the estimation results of the linear probability model given in Equations (1) and (3). All variables have been demeaned to make all coefficients comparable across models (columns). The first column controls for firm, bank and time (year:month) fixed effects and the triple interaction is not included. The estimated 20

21 coefficient on Scoring equals ***. 10 The effect of the firm s risk profile on the likelihood that a loan application is granted is the one expected. In terms of its economic impact, a one standard deviation increase in the scoring decreases the average probability by 9%. In other words, the probability of a loan application to be granted is 12.3% lower for firms in the third quartile of the distribution of the scoring variable compared with those in the first quartile. On average, banks that increased its capital ratio granted loans with a higher probability. A one percentage point increase in bank capital ratio rises the average probability a loan application to be granted by 2.1%. However this effect becomes the contrary after 2011, when Spanish banks had to strengthen its capital. The economic magnitude of the change in banks capital ratio after 2011 is just the opposite: -2.2% decrease in average probability of loan application to be granted after a 1pp increase in the capital ratio. We take this result as evidence of a negative effect on bank loan supply derived from the imposition of higher capital requirements. This result coincides with previous research related to the negative impact of increased minimum capital requirements on bank lending (Bridges et. all, 2014, BIS, 2010, and Uluc and Wieladek, 2015). Column II adds the triple interaction of interest (and all the double terms associated). Interestingly, the estimated coefficient on our key variable of interest, Capital ratio jt-1 *I(TIME>=2012M1) t *Scoring it-1 is 0.016***, meaning that banks that increase capital ratios after 2011 are, on average, less likely to grant loans but less so for riskier firms. For instance, the probability to grant a loan increases by 3% for firms in the third quartile of the risk distribution compared to less risky firms in the first one. 10 As in the Tables, ***, ** and * indicates statistical significant at the 1, 5 and 10 percent level, respectively. 21

22 This result confirms our hypothesis that banks react to negative capital shocks by taking more risk, trying to reduce the impact of higher capital ratios on their profitability (return over equity) and as a way to increase capital by retaining earnings. This result coincides with some previous evidence in the literature. For example, Uluc and Wieladek (2015) use data for UK banks in and find that capital increases imply a loan contraction which affected to a lesser extent borrowers with impaired credit history, which is consistent with risk-shifting behavior. Auer and Ongena (2016) find a compositional effect after the activation of the CCB in Switzerland, which was targeted at loans secured by domestic residential properties. In particular they find that banks with higher shares of residential RWA to total assets lent more to corporations and also that banks shifted lending to riskier firms. In the next columns we will progressively saturate the model in all its possible dimensions. Column III controls for concurrent changes in loan supply or in the risktaking of banks with bank*time (year:month) fixed effects. The coefficient on the interaction of interest is not statistically different to the previous one. Column IV removes firm observed and unobserved time-varying heterogeneity adding firm*time (year:month) fixed effects but keeps only bank fixed effects (i.e. does not include timevarying fixed effects by bank). Again, results remain unaltered. Finally, the last column (our baseline) includes both sets of dummies and shows that the coefficient is 0.019***, with no significant differences. It is important to note that the value of the R 2 has increased from 39% to 52% as long as we saturated the model. At the same time, the estimated coefficient on the triple interaction has remained unchanged. This implies that our estimates are quite robust to 22

23 the presence of non-controlled unobserved heterogeneity (Altonji et al (2005) and Oster (2014)), which alleviates endogeneity concerns. Finally, to stress the sensitivity of our results to the election of the time period from which markets and regulators demanded banks to strengthen their capital, Figure 1 shows the estimated coefficients on the triple interaction for different dates. What we do is to run the baseline model but multiplying Capital ratio jt-1 *Scoring it-1 by time (year:semester) dummies. Results clearly show that just after 2011 banks started to follow a risk-taking strategy that ended in the second semester of This date coincides with the full compliance of Basel III for almost all Spanish banks and the relaxation of the Spanish risk premium. In sum, analyzing the effect on bank loan supply of more stringent capital requirements, our evidence suggest that in the short run banks react cutting the supply of credit but that this behavior is less pronounced against riskier firms. This effect is transitory and after three years disappears. Our result should be read in line with the results found in Section regarding the loan characteristics. Under this section we also find that these banks tend to lend a lower amount of credit for each loan and at the same time require higher collateralization as a way to cover the greater risk assumed in the transaction Robustness Tests Table 3 presents the results of a battery of robustness checks. In Panel A, columns I to IV use the same specification as in our baseline model (column V of Table 2) checking particular issues. The last three columns of Panel A control for the matching between firms and banks in a way to reduce as much as possible any unobserved heterogeneity that could bias the results. 23

24 Panel B takes into account that in the Spanish financial market the two largest banks are internationally active, which could make our capital ratio variable at the individual level a bad proxy of the solvency position of these banks. Therefore in this Panel we perform an analysis at the consolidated level. Are capital ratios increasing due to an increase in the capital base or a shrink in balance sheets? Our underlying assumption until now is that a positive growth in the capital ratio is due to an increase the bank s capital base. Column I of Panel A decomposes the change in capital ratio to test whether this is true or not. On the one hand, and increase in capital ratio can be the result of a growth in the numerator of the rate. On the other hand, the capital ratio can just increase as a result of a deleveraging strategy. To test what is behind the observed results we estimate a model where instead of the change in the capital ratio, the change in the log of the amount of capital and the change in the log of total assets are introduced. A positive and statistically significant coefficient on the interaction term log(capital jt-1 )*I(TIME>=2012M1) t *Scoring it-1 is found, whereas the coefficient on log(total Assets jt-1 )*I(TIME>=2012M1) t *Scoring it-1 is insignificant. This confirms that observed increases in capital ratios after 2011 are moved by a larger capital base. Did interest rates have an effect on the lending behavior of banks during the period of interest? Column II of Panel A studies whether the risk-taking pattern depends on monetary policy. We follow Jiménez et al (2014) and we multiply our variable of interest with the change in the overnight interest rate, thus having a quad. The negative coefficient is 24

25 highly statistically significant, meaning that banks are more prone to take on more risk when interest rates are low. In a context where low interest margins are stressing the income statements of banks, this can be seen as a way to alleviate this effect. Are results driven by a specific bank business model? As can be seen in the Column III of Panel A, the interaction term is again positive, statistically significant and not statistically different form the one obtained in our baseline estimation when construction and real estate are excluded from our sample. This sector suffered its own penance during the crisis and may be driving the overall results. As we have shown this is not the case. Are the results robust when controlling for the restructuration and recapitalization of the Spanish banking system? During the period considered in our study Spain needed the financial assistance of the EU to reform its financial system. The European Financial Stability Facility agreed to provide funds for the recapitalization of the Spanish banking sector. To disentangle this effect from our results in Column IV of Panel A we present the results of the baseline model where bailed-out banks or banks that acquired other after 2011M12, and got funding support for the operation, were excluded. This is a very restrictive check but although the number of observations decreased by 80%, which reduces accuracy, the magnitude of the coefficient of the triple interaction is not statistically different from our baseline (0.023*). Are results robust when controlling at the same time for time-varying unobservable demand, supply and interactions of them? 25

26 Columns V to VII of Panel A check the sensitivity of our results to the inclusion of unobservable firm-bank characteristics. There can be a selection bias whether firms and banks are not randomly attached. Given that our sample consists on firms asking for a loan to a non-current bank, this seems a minor problem. However, some banks can be more focused on a certain sector or province, with a particular risk profile. To minimize this potential problem a set of industry*province*bank fixed effects are added to our baseline model. Moreover, we also control for time-varying firm-bank unobserved heterogeneity through the set of industry*province*bank*time (year) fixed effects and even industry*province*bank*time (year:month). Column V of Panel A controls for the first set of dummies. We consider 52 different regions (50 provinces plus two autonomous cities) and 17 industries. Therefore, column V includes more than 6,200 dummies. The number of observations decrease a little bit (the 3% of the observations is lost), but is still representative of our sample. The estimated effect of the triple interaction is 0.017*** and the R 2 increases 4 percentage points. This result seems to indicate that the firm-bank selection is not an issue here. However, we saturate the model with yearly varying firm-bank fixed effects (column VI) and (year:month) varying firm-bank dummies (Column VII). The number of controls reaches over 17,000 and although the sample is reduced drastically as the number of fixed effects increases, the coefficients equal to 0.032*** and 0.021*, respectively, being both non statistically different from our baseline estimate. This means that firm-bank attachment does not affect the risk-taking channel. Moreover, the R 2 peaks at 71% without a significant variation of our coefficient of interest. This ensures an insignificant bias derived from unobservable non controlled variables. Are results robust when excluding internationally active banks? 26

27 With Panel B of Table 3 our aim is to show that our results are robust to the exclusion of the two largest banks and to the level of aggregation of banks. Basically, the capital ratio used until now as a proxy of the solvency rate of the bank is a leverage ratio at individual level (equity minus goodwill over total assets). This can be seen as a good proxy as long as the bank operates locally. This could not the case for the two largest internationally active banks in Spain. The first three columns of Panel B exclude these two major banks. The triple interaction coefficient in this sample yields an effect on the likelihood of a loan to be granted of 0.027*** (Column I), which is not statistically different from our baseline model (Column V in Table 2). After excluding these two banks, our sample is focused on banks that mainly operate in Spain and therefore we assume that RWAs at the consolidated level are a good proxy for those at the individual level (our database does not provide us RWAs at the individual level). Therefore, we perform an additional robust analysis using consolidated data for the capital ratio using RWAs. In particular, Column II uses the Tier 1 ratio (equity minus goodwill over risk-weighted assets) which is a reasonable measure at the individual level for local banks. The coefficient is again positive and statistically significant (0.012***). Finally, column III decomposes the Tier 1 ratio and that the results show that, again, the effect is driven by the change in capital (i.e. the numerator of the ratio). Do the results change when considering consolidated data for banking institutions? Columns IV to VI have the same structure as columns I to III but now the sample changes because banks are considered at the consolidated level. All results go in the same direction than before and are even more relevant. 27