Bank competition and stability: Reconciling con icting. empirical evidence

Transcription

1 Bank competition and stability: Reconciling con icting empirical evidence Thorsten Beck y Olivier De Jonghe z Glenn Schepens x 22 May 2011 Abstract Theoretical models and empirical results o er con icting evidence on the relationship between bank competition and bank stability. This paper aims to reconcile the seemingly contrasting evidence on the bank competition-bank soundness relationship. We develop a uni ed framework to assess how regulation, supervision and other institutional factors may make it more likely that the data favor one theory over the other (charter value paradigm versus risk-shifting paradigm). Cross-country heterogeneity in these factors allows us to test the assumptions and predictions of various theoretical models. We show that an increase in competition will have a larger impact on banks risk taking incentives in countries with stricter activity restrictions, more herding in revenue structure, unconcentrated banking markets and more generous deposit insurance. The authors would like to thank Fabio Castiglionesi, Hans Degryse, Claudia Girardone, Klaus Schaeck and seminar participants at HEC Paris, Ghent University, Tilburg University, Cass Business School, Roma II Tor Vergata, Université Libre de Bruxelles, Bangor Business School for interesting discussions and helpful comments. y CentER, European Banking Center, Tilburg University and CEPR. z CentER, European Banking Center, Tilburg University. x Corresponding author: glenn.schepens@ugent.be. Department of Financial Economics, Ghent University. 1

2 1 Introduction The impact of bank competition on nancial stability remains a widely debated and controversial issue, both among policymakers and academics. The belief that ercer competition among banks would lead to a more e ective banking system initiated a deregulating spiral in the late 70s and early 80s. While the deregulation of branching and activity restrictions may have resulted in more intense competition among banks, with positive repercussions, it may as well have had the unintended consequence of increasing banking sector instability. 1 Similarly, the international process of banking liberalization seemingly has gone hand in hand with an increased occurrence of systemic banking crises in the last two decades of the twentieth century, culminating in the global nancial crisis of However, there is no academic consensus on whether bank competition leads to more or less stability in the banking system. On the one hand, an increase in loan market competition leads to lower lending rates and hence lower interest margins. As banks franchise values erode, this may create incentives to gamble and may lead to a shift towards riskier activities, because of the limited liability by bank shareholders that e ectively turns bank equity into a put option on banks pro ts. On the other hand, a similar argument but in the opposite direction can be made for the bank s borrowers. If entrepreneurs are confronted with lower loan rates, they will choose safer projects and have fewer incentives for aggressive risk taking, i.e. the adverse selection and moral hazard problems will be mitigated. These two opposite e ects may help in explaining why empirical studies across di erent samples and time periods fail to nd a consensus on which e ect dominates. Moreover, comparing the results of di erent studies is complicated by the use of di erent competition and risk measures. A similar inconclusive debate as on the relationship between competition and stability has been led on the e ect of the regulatory framework on banks risk-taking incentives and ultimately bank stability. On the one hand, capital regulation and interest rate and activity restrictions 1 See among other Keeley (1990) and Jayaratne and Strahan (1998) 2

3 are seen as fostering stability (Hellmann, Murdock, and Stiglitz (2000)); on the other hand, they might lead to rent-seeking and might prevent banks from reaping necessary diversi cation and scale bene ts. The role of deposit insurance schemes has been especially controversial. While often introduced to protect small depositors lifetime savings and to prevent bank runs, they also provide perverse incentives to banks to take aggressive and excessive risks. These perverse incentives are held less in check in weak supervisory frameworks (Demirguc-Kunt and Detragiache (2002)). This paper combines the two literatures and assesses whether the relationship between competition and stability varies across markets with di erent regulatory frameworks, market structures and levels of institutional development. Speci cally, while holding the measure of bank competition and stability constant across samples, we document that support for either the competitionstability or competition-fragility view varies across countries and over time. Next, we identify and test the possible channels that may create cross-country variation in the competition-stability relationship. While we identify several country characteristics that explain the cross-country variation in the competition-stability relationship, a large amount is still unexplained. Finally, based on our results, we try to reconcile the seemingly con icting existing empirical results on the competition-stability relationship. <Insert Figure 1 around here> As way of motivation, consider the cross-country variation in the relationship between competition and stability. In our sample of banks in 62 countries, the pairwise correlation 2 between the Lerner index and the Z-score, widely used proxies for market power and bank soundness, respectively, is 0:25. 3 Figure 1, however, reveals that this full sample correlation masks a substantial 2 We refer to simple pairwise correlation coe cients in the introduction. A similar story can be made with regression based conditional relationships. However, for ease of exposition, we postpone this to later sections. 3 The Lerner index is the ratio of the di erene between price and marginal cost and the price, with higher values indicating higher market power. The Z-score is an accounting measure of bank distress. It is measured 3

4 degree of country-level heterogeneity. Each bar in Figure 1 corresponds to the country-speci c pairwise correlation between market power and bank soundness. The average pairwise correlation over the 62 countries resembles the full sample correlation. However, there is a large amount of heterogeneity in the competition-stability relationship, with correlations ranging from below -0.2 to above 0.5. In some countries, the correlation is negative and signi cant. In many others it is not statistically di erent from zero. In most countries, the number is positive and signi cant. Rather than being interested in the sign of the relationship, we are interested in the cross-sectional dispersion. Speci cally, we are interested in which country-speci c features make it more likely that competition is less harmful or more bene cial for bank soundness. Exploring the variation in the competition-stability relationship is important for academics and policy makers alike. The academic debate on the e ect of competition on bank stability has been inconclusive and by exploring factors that can explain variation in the relationship, this paper contributes to the resolution of the puzzle. Policy makers have been concerned about the e ect of deregulation and the consequent competition on bank stability but have also discussed di erent elements of the regulatory framework that have both an impact on competition and directly on stability, including deposit insurance capital regulation and activity restrictions. This paper shows a critical role for the regulatory framework in explaining the variation across countries and over time in the relationship between competition and stability and has therefore important policy repercussions. Our paper builds on a rich theoretical and empirical literature exploring the relationship between competition and stability in the banking system. On the one hand, the competitionfragility view posits that more competition among banks leads to more fragility. This charter value view of banking, as theoretically modeled by Marcus (1984) and Keeley (1990), sees as the sum of accounting pro ts and the capital to asset ratio, divided by the volatility of pro ts. As such, it indicates with how many standard deviations pro ts can fall before capital is depleted. 4

5 banks as choosing the risk of their asset portfolio. Bank owners, however, have incentives to shift risks to depositors, as in a world of limited liability they only participate in the up-side part of this risk taking. In a more competitive environment with more pressures on pro ts, banks have higher incentives to take more excessive risks, resulting in higher fragility. In systems with restricted entry and therefore limited competition, on the other hand, banks have better pro t opportunities, capital cushions and therefore fewer incentives to take aggressive risks, with positive repercussions for nancial stability. In addition, in more competitive environment, banks earn fewer informational rents from their relationship with borrowers, reducing their incentives to properly screen borrowers, again increasing the risk of fragility (Boot and Thakor (1993), Allen and Gale (2000), Allen and Gale (2004)). The competition-stability hypothesis, on the other hand, argues that more competitive banking systems result in more rather than less stability. Speci cally, Boyd and De Nicolo (2005) show that lower lending rates reduce the entrepreneurs cost of borrowing and increase the success rate of entrepreneurs investments. In addition, these rms will refrain from excessive risk-taking to protect their increased franchise value. As a consequence, banks will face lower credit risk on their loan portfolio in more competitive markets, which should lead to increased banking sector stability. However, more recent extensions of the Boyd and De Nicolo (2005) model that allow for imperfect correlation in loan defaults (Martinez- Miera and Repullo (2010); Hakenes and Schnabel (2007)) show that the relationship between competition and risk is U-shaped. Hence, the impact of an increase in competition can go either way, depending on other factors. Wagner (2010) extends the Boyd and De Nicolo (2005) model and allows for risk choices made by borrowers as well as banks. If lending rates decline due to more competition, banks have less to lose in case a borrower defaults. Hence, a bank may nd it optimal to switch to nancing riskier projects 4, which overturns the Boyd and De Nicolo (2005) 4 Other authors have also shown that more intense competition may induce banks to (i) switch to more risky, opaque borrowers (Dell Ariccia and Marquez (2004)), and (ii) acquire less information on borrowers (Hauswald and Marquez (2006)). 5

6 results. The standard response to con icting theoretical predictions is to let the data speak. Numerous authors have used di erent samples, risk measures and competition proxies to discriminate between the competition-fragility and competition-stability view. 5 Empirical studies for speci c countries many if not most for the U.S. have not come to conclusive evidence for an either stability-enhancing or stability-undermining role of competition. The cross-country literature has found that more concentrated banking systems are less likely to su er a systemic banking crisis as are more competitive banking systems (Beck, Demirguc-Kunt, and Levine (2006); Schaeck, Cihak, and Wolfe (2009)). There seems also evidence that banks in more competitive banking systems hold more capital, thus compensating for potentially higher risk they are taking (Schaeck and Cihak (2010a), Berger, Klapper, and Turk Ariss (2009)). A consequence of the recent theoretical extensions is that the predicted impact of competition on bank stability moved from a bipolar setting (good or bad per se) to a continuous approach (settings that are better or worse in relative terms). These models lead to new testable implications that exceed a mere assessment of the sign of the coe cient of bank market power. For example, by allowing loan defaults to be imperfectly correlated, the Martinez-Miera and Repullo (2010) model and the Hakenes and Schnabel (2007) model imply that the impact of competition on risk is a ected by regulatory constraints on asset diversi cation, since the latter will a ect the correlation structure of loan defaults. Our results suggest that an increase in competition will have a larger impact on banks risk taking incentives in countries with stricter activity restrictions, more herding in revenue structure and unconcentrated banking markets. These ndings are con rmed both in cross-sectional regressions as well as when we allow for additional time-series variation in the competitionstability relationship. However, we also nd that a large part of the cross-country variation 5 For an overview, see Beck (2008). 6

7 in the competition-stability relationship cannot be explained, which constitutes a challenge for further research. The remainder of the paper is structured as follows. Section 2 discusses di erent factors that might explain the variation in the competition-stability relationship documented in Figure 1. Section 3 introduces data and methodology, while section 4 presents the main results. Section 5 presents robustness, while section 6 concludes with policy implications. 2 Competition-stability relationship - theoretical considerations We argue that country-speci c features may a ect the existing empirical evidence on the relationship between competition and stability via three possible channels. First, a certain type of regulation may limit the extent to which banks can or will engage in riskier activities if their franchise values are eroded. For example, regulatory capital requirements should limit the extent to which banks can follow risk-taking incentives if banks charter value is eroded. Second, country-speci c characteristics may also a ect the adverse selection problem that banks face if they charge higher loan rates. For example, lending relationships or credit registries may reduce the likelihood that entrepreneurs will chose riskier project in response to higher loan rates. Third, institutional characteristics may a ect the proportion of systematic and idiosyncratic risk in loan defaults and may make it hence more likely that the empirical data favor one theory over the other. For example, regulatory constraints on asset, revenue or geographical diversi - cation may make it more likely that loan defaults are highly correlated and hence lead to the empirical outcome that competition is good for nancial stability. More speci cally, let denote the estimated e ect of bank market power on stability. This point estimate is in uenced by three factors: CF > 0; CS < 0; p(cf ) 2 [0; 1] where CF denotes the stability welfare 7

8 gains of a unit increase in market power (competition-fragility hypothesis), CS denotes the stability loss as a result of a unit increase in market power (competition-stability hypothesis) and p(cf ) indicates how likely it is that one theory dominates over the other. We conjecture that = p(cf ) CF + (1 p(cf )) CS. Our conjecture is that the regulatory environment, strength of supervision and the institutional framework of a country a ect CF ; CS and p(cf ). More speci cally, let x denote the speci c feature under investigation. A change in x (or two samples with di erent x) can lead to a di erent estimated impact of market power on stability via three di CS + p(cf + (1 p(cf The relative strength of each of these three channels may explain why di erent studies obtain di erent results in terms of sign and magnitude. That is, certain country-speci c features may make the assumption and prediction of a given theoretical model more realistic. In the remainder of this section, we describe which country-speci c features may play a role and why. We also introduce speci c measures that capture these di erent market-speci c characteristics. 2.1 Herding A rst important market characteristic that can in uence the relationship between competition and stability is covariation of banks behavior, also known as herding. Acharya and Yorulmazer (2007) and? show that the supervisory decision to intervene a failing bank is subject to an implicit too-many-to-fail problem: when the number of bank failures is large, the regulator nds it ex-post optimal to bail out some or all failed banks. This gives banks incentives to herd and increases the risk that many banks may fail together. Hence, herding behavior may also a ect banks incentive to increase risk-taking in response to an increase in competition. Bank activity herding is measured by a heterogeneous banking system indicator that measures whether there are substantial di erences among di erent nancial institutions within a country. It is calculated 8

9 as the within country standard deviation of the non-interest income share. If all banks in a country have a similar business model (either voluntarily or forced by regulation), the indicator will be low. When bank activities are highly correlated across banks, a rise in competition will do more damage to a banks franchise value since they do not have any other activities to fall back on. Thus, we can hypothesize that competition will have a stronger impact on bank risk behavior in more homogeneous banking systems, > 0. It is important to note in this context that we do not relate the actual activity structure of banks to the relationship between competition and stability, but rather the variation in activity structure within a market. We also look at herding in terms of risk taking behavior (systemic risk). If a majority of banks has a high risk appetite, it is very well possible that other banks feel the pressure to take on more risk due to the herding incentives described above. Therefore, we hypothesize that banks operating in an environment with a high risk taking standard, will have a stronger incentive to increase risk taking when competition changes. 2.2 Market structure A second important market characteristics that might in uence the relationship between competition and stability is the structure of the market. Martinez-Miera and Repullo (2010) and Hakenes and Schnabel (2007) show that a lower correlation of loan defaults makes it more likely that ercer competition harms stability. A bank s potential to reduce the correlation of its loan portfolio and other revenues is clearly a ected by restrictions on functional or geographical diversi cation. If x is a proxy for diversi cation, we conjecture > 0. Market concentration, measured by the Hirschmann-Her ndahl index, may also play a role in assessing the strength of the competition-stability relationship. While these measures are not good proxies for competition 6, market concentration may play an important role in 6 See Claessens and Laeven (2004). 9

10 determining the relationship between competition and stability. Fewer banks in the economy (more concentrated banking markets) make supervision more e ective and accurate. If bank supervisors have to monitor fewer banks, they may observe malpractices (risk-shifting, loan portfolio concentration) in a more timely fashion. According to Allen and Gale (2000), countries with a larger number of banks (such as the US) are, ceteris paribus, more likely to support the competition-fragility view compared to banking sectors dominated by fewer larger banks (such as Canada), < 0 where x stands for the degree of concentration. Moreover, with fewer banks in the system, entrepreneurs may behave more prudently as they have fewer outside options when they default on their loans, which raises the franchise value of the bank. On the other hand, the lower the number of banks in a country, the more they will be interconnected, which may again encourage risk-taking behavior if banks perceive themselves as too-important-to fail, > 0. The e ect of market structure on the competition-stability relationship is therefore a priori ambiguous. 2.3 Regulatory and supervisory framework A third group of country traits that in uence the relationship between competition and stability consists of regulations designed to protect bank charter values and to prevent risk-seeking behavior if charters are eroded. High capital levels reduce the moral hazard incentives to take aggressive risks. More stringent (risk-based) capital regulation may therefore limit the negative in uence that competition may have on stability < 0). Hellmann, Murdock, and Stiglitz (2000), however, show that even with capital requirements, deposit interest rate ceilings are still necessary to prevent banks from excessive risk-taking in competitive markets. Furthermore, Allen, Carletti, and Marquez (2010) show that borrowers prefer well capitalized banks, since these banks have a relatively higher incentive to monitor, which improves rm performance. They nd that franchise value and capital are substitute ways of providing banks with 10

11 monitoring incentives. Also, a recent study by Mehran and Thakor (2010) shows that there is a positive relationship between bank capital holdings and total bank value. This rise in bank value and the borrower preferences should induce a rise in bank charter value, thus lowering the banks risk appetite. These e ects allow us to hypothesize that higher capital regulation limits the negative e ect of competition on bank stability. In other words, the average impact of an increase in competition on bank fragility is larger in countries with weak capital regulation vis-à-vis strict capital requirement regimes. Another popular regulatory measure to increase the stability of banking systems is deposit insurance, as it reduces the risk of bank runs. On the other hand, generous deposit insurance schemes might increase moral hazard and thus increase risk-taking incentives in more competitive environments, > 0 (see, e.g., Demirguc-Kunt and Kane (2002)). E ective banking supervision can be important for several reasons. First, monitoring banks is both costly and di cult for both depositors and shareholders, which can lead to suboptimal bank risk behavior. Secondly, bank failures may be very costly, due to the crucial role banks play within the economic system. Taking these points into account, more e ective supervision should provide incentives to limit bank risk taking and thus could soften the e ect of competition on risk taking. On the other hand, Boot and Thakor (1993) show that supervisors may pursue self interest, which may lead to suboptimal actions. We integrate two supervisory variables into our analysis, a dummy that indicates whether there is more than one supervisor and an external governance indicator. The e ect of having multiple supervisors is not unambiguous. Kahn and Santos (2005) argue that if a single institution is responsible for di erent regulatory functions, it may not be able to su ciently monitor all banks. Also, having multiple supervisors may lead to di erent supervisory approaches, which can generate useful information which would otherwise be neglected (Llewellyn (1999)). From this point of view, having multiple supervisors should reduce banks risk taking incentives. On the other hand, a single supervisory institution may 11

12 be preferred because it reduces the chance of taking con icting policy measures. Furthermore, Llewellyn (1999) argues that a single authority could prevent gaps in the regulation that could arise when there are multiple supervisors and that having multiple supervisors could lead to supervisory arbitrage, thus relaxing the overall supervision. The empirical evidence on this topic is rather scarce. The only study that extensively focuses on the number of supervisors is Barth, Dopico, Nolle, and Wilcox (2002) who nd evidence for the supervisory arbitrage theory when there are multiple supervisors. The external governance indicator measures the strength of external auditors, nancial statement transparency and accounting practices and the existence of external ratings, thus the degree of potential market discipline. Having a wide range of private control mechanisms such as external audit and external ratings should dampen the risk incentives of a bank. 2.4 Institutional and nancial development A fourth set of country traits that can in uence the competition-stability relationship is the institutional framework and nancial system structure in which banks operate. First, we consider the contractual framework. Loan defaults can arise if a borrower is unable or unwilling to repay her loan. In the latter case, contract enforcement possibilities will be of great importance for banks. If a borrower knows that a bank will have to go through numerous procedures, wait for several weeks/months or simply has to pay large fees to enforce a contract, she will have a greater incentive to evade the loan repayment. A part of the Boyd and De Nicolo (2005) explanation of the competition-stability view relies on the fact that lower loan rates will reduce the entrepreneurs borrowing cost and thus will increase the success rate of his project, which lowers the loan default probability. However, when operating in countries with protracted contract enforcement procedures, the entrepreneur has a counteracting incentive to repay his loan, independent of his success rate. Thus, we expect that a change in competition will be more 12

13 harmful to stability when operating in a country with low credit enforcement standards. In other words, a rise in credit enforcement reduces the risk-shifting incentives of entrepreneurs, > 0. Take next the credit information sharing framework. Credit registry institutions are public or private entities which collect information on the creditworthiness of borrowers. The existence of these institutions facilitates the exchange of credit information among banks and among investors. The existence of credit registers is expected to reduce both adverse selection and moral hazard problems that are inherent on being in the lending business. Pagano and Jappelli (1993) show that information sharing lowers adverse selection problems by lowering the selection cost for lenders. Kallberg and Udell (2003) con rm these ndings when studying private information exchanges in the U.S.. They nd that private credit registry information is valuable in assessing borrower quality, after controlling for information that would be available to a single institution. Information sharing also tends to reduce moral hazard incentives through reputation e ects (see, e.g. Diamond (1989)). As borrowers realize that it will be hard to get a loan at another institution when they default on their current loan, they will have a stronger incentive to repay and they will choose safer project (Padilla and Pagano (2000), Vercammen (1995)). Furthermore, Houston, Lin, Lin, and Ma (2010) show for a sample of nearly 2400 banks in 69 countries that greater information sharing leads to higher bank pro ts and lowers bank risk. This leads us to hypothesize that countries with better information sharing systems will encounter smaller e ects on stability when competition changes, since better information systems increases a banks franchise value and will lower the entrepreneurs incentive to take more risk. Finally, we consider nancial structure and, more speci cally, competition for banks coming from nancial markets. More developed stock markets make it easier for rms to switch between bank-based and market-based funding. This could lead to an additional e ect of a change in competition on bank risk behavior. As mentioned above, Boyd and De Nicolo (2005) show that 13

14 a higher loan interest rate (due to lower competition) leads to a higher loan default probability. Martinez-Miera and Repullo (2010) add that, when loans are not perfectly correlated, higher interest rates also raise pro ts on non-defaulting loans. In countries with strong developed capital markets, however, rms will have the possibility to substitute loans with market-based funding, thus lowering the total amount of loans and bank pro t. This leads us to hypothesize that, ceteris paribus, it is more likely to nd positive e ects of competition on bank stability in countries with well developed nancial markets. 3 Data and methodology This section consists of two parts. First, we describe the sample composition and data sources. Next, we explain how we allow for a country-level variation in the estimated impact of competition on stability. In this section, we also describe how we compute the bank-speci c measures of soundness and market power. 3.1 Data sources We combine several data sources. We obtain information on banks balance sheet and income statement from Bankscope. Bankscope is a database compiled by Fitch/Bureau Van Dijck that contains information on banks around the globe, based on publicly available data-sources. Moreover, the information in the database is harmonized and provided in a global format 7 that facilitates the international comparison of banks nancial statements. Admittedly, in general this comes at the cost of losing detailed information. However, this is not an issue for the information we need in our analysis. The period of analysis is , and hence is not contaminated by the exceptional events of the global nancial crisis 8. If banks report information at the 7 As of April 2009, the global format is replaced by the Fitch Universal Format on Bankscope. 8 The time period is mainly determined by the availability of the country-speci c characteristics. In addition, it spans the period before Basel II was implemented. As such, the change in capital regulation does not a ect our 14

15 consolidated level, we delete the unconsolidated entries of the group from the sample to avoid double counting. We apply a number of selection criteria to arrive at our sample. First, we exclude countries for which we have information on less than 50 bank-year observations. Second, we limit our analysis to commercial, saving and cooperative banks. Third, we delete banks that report information for less than three consecutive years, as our risk measure is computed over rolling windows of three years. Fourth, we drop bank-year observations that do not have data available on basic variables drop. Subsequently, we winsorize all variables at the 1 percent level to mitigate the impact of outliers and to enhance robustness of the standard errors. While most of the bank-speci c variables are ratios, variables in levels (such as size) are expressed in 2007 US dollars using a GDP de ator. The bank-speci c data are linked to various country-level datasets that contain information on the regulatory framework, strength of supervision and other institutional features. More speci - cally, we employ data from the three vintages (2000, 2003 and 2007) of the Bank regulation and supervision dataset compiled by the World Bank (Barth, Caprio, and Levine (2008)). Additional information is obtained from the Heritage Foundation,the World Development Indicators and the Doing Business database. A detailed list of the variables used and the database from which they are collected can be found in Table 13. Filtering the bank-speci c database and matching it with the country-level datasets yields a sample of banks from 62 countries. The sample consists of a mix of developed and developing countries (see Table??). 3.2 Empirical framework In the literature, there are two main approaches to assessing the relationship between competition and stability: a single country or multiple country setup. In a cross-country setup, proxies measure of risk. Note also that by looking at a cross-country setting over the period , we do have other crisis episodes in the sample. 15

16 of market power at the bank- or country-level are related to bank soundness (in a linear or quadratic speci cation). The sign of the coe cient(s) then indicate whether competition helps or harms stability (or whether there is a turning point at which there is a sign reversal). These studies provide insight into the average relationship between competition and stability for the set of countries under investigation (e.g.: developing countries as in Turk Ariss (2010), developed countries as in Berger, Klapper, and Turk Ariss (2009), the European Union as in Schaeck and Cihak (2010b)), while controlling for other country-speci c factors such as macro-economic conditions, regulation and supervision. However, single country studies (such as Keeley (1990), Salas and Saurina (2003), Jimenez, Lopez, and Saurina Salas (2010), Boyd, De Nicolo, and Jalal (2006)) document a large degree of variation in the competition-stability relationship. This indicates that these other country-speci c factors may not only have a level e ect but also a slope e ect. Hence, it is not only important to control for the impact of these factors on risk but also on how they shape the competition-stability relationship. For example, activity restrictions (i.e. allowing banks to enter real estate, insurance or underwriting) may not only a ect the aggregate level of risk, but may also in uence the extent to which loan market power a ect bank risk-taking incentives. Put di erently, these variables may also determine whether it is more likely to nd support for the franchise value paradigm compared to the risk-shifting hypothesis or vice versa. This results in the following setup: Risk i;j;t = c + j Competition i;j;t 1 + j X i;j;t 1 + Z j;t + " i;j;t (1) In this setup, the indices i; j ; t stand respectively for bank, country and time. The impact of competition (as well as any other bank-speci c variable, X i;j;t ) on risk is allowed to vary at the country level. This is denoted by giving the corresponding (vector of) coe cient(s) a j subscript. The vector of bank-speci c variables, X i;t 1, characterizes a bank s business model. In particular, we include proxies for the funding structure (share of wholesale funding in total funding), asset 16

17 (loans to assets ratio) and revenue mix (share of non-interest income in total income) as well bank size (natural logarithm of total assets), credit risk (loan loss provisions to interest income) and asset growth. In addition, we include specialization dummies to allow for di erent intercepts for commercial banks, bank holding companies, saving banks and cooperatives. Summary statistics on the control variables that determine bank soundness are in the upper part of Table 1. In addition, time-varying country-speci c characteristics may also a ect bank soundness are included in the vector Z j;t. We hypothesize that j can be modelled as a function of (a subset of) these country-speci c factors. To gain insight in the potential drivers of heterogeneity in, we take a two-step approach. In a rst step, we relate bank market power to a measure of bank soundness. This relationship is assessed at the country level. More speci cally, for each country in our sample, we estimate the following equation: Risk i;t = c + Competition i;t 1 + X i;t 1 + v t + " i;t (2) Including time xed e ects and estimating this equation country by country creates many advantages. First, we allow for the maximum extent of heterogeneity in the competition-stability trade-o across countries. Second, the time dummies di er in each country regression and hence indirectly capture the level e ect of country-speci c regulation or the business cycle on bank risk. In the second step, we will explore which country-speci c variables explain the heterogeneity in the estimated j s. 3.3 Indicators of market power and bank soundness In this subsection, we describe how we measure competition, discuss the correlation with other measures of competition and introduce our indicator of bank soundness. 17

18 3.3.1 The Lerner index: measure of pricing power For our analysis, we need a measure of market power that varies at the bank level rather than a competition or concentration proxy at the country level. The Lerner index is an obvious candidate as it captures the extent to which banks can increase the marginal price beyond the marginal cost. Conditional on having an estimate of the marginal price and cost, we can construct the Lerner index for each bank and each year, as follows: Lerner i;t = P i;t MC i;t P i;t (3) where P i;t is proxied by the ratio of total operating income to total assets. As banks have the opportunity to expand their activities into non-interest generating activities, we include both interest and non-interest revenue. The marginal cost, MC i;t, is derived from a translog cost function. As Berger, Klapper, and Turk Ariss (2009), we model the total operating cost of running the bank as a function of a single, aggregate output proxy, Q i;t, and three input prices, w j i;t, with j 2 f1; 2; 3g. More speci cally, we estimate: ln C i;t = ln Q i;t + 2 (ln Q i;t ) 2 + 3X 3X j ln w j i;t + j=1 j=1 k=1 3X 3X j;k ln w j i;t ln wk i;t+ j ln w j i;t ln Q i;t+v t +" i;t We also include time dummies to capture technological progress as well as varying business cycle conditions, as well as a bank specialization dummy. Homogeneity of degree one in input 3X 3X prices is obtained by imposing the restrictions: j = 1; j = 0 and 8 k 2 f1; 2; 3g : j=1 3X j;k = 0. Marginal cost is then obtained as follows: j=1 MC i;t i;t = C i;t Q i;t 1 + 2b 2 ln Q i;t + j=1 j=1 (4) 1 2X b j ln wj i;t A (5) j=1 w 3 i;t 18

19 in which C i;t measures total operating costs (interest expenses, personnel and other administrative or operating costs), Q i;t represents a proxy for bank output or total assets for bank i at time t. The three input prices capture the price of xed assets, the price of labor and the price of borrowed funds. They are constructed as respectively the share of other operating and administrative expenses to total assets, the ratio of personnel expenses to total assets and the ratio of interest expenses to total deposits and money market funding. Following Berger, Klapper, and Turk Ariss (2009), Equation (4) is estimated separately for each country in the sample to re ect potentially di erent technologies. <Insert Table 1 around here> Table 1 presents summary statistics on the variables needed to construct the Lerner index as well as the estimated Lerner index. The average Lerner index at the country level is 10%, but varies across countries, from 3:8% in Uruguay to 22:0% in Denmark (see Table 14) The Z-Score: measure of bank soundness In our analysis, bank risk is measured using the Z-score. The Z-score measures the distance from insolvency (Roy (1952)) and is calculated as Z i;t = ROA i;t + (E=A) i;t (ROA) i;t (6) where ROA is return on assets, E=A denotes the equity to asset ratio and (ROA) is the standard deviation of return on assets. While in large parts of the literature the volatility of pro ts is computed over the full sample period, we use a three-year rolling time window for the standard deviation of ROA to allow for variation in the denominator of the Z-score. This approach avoids that the Z-scores are exclusively driven by variation in the levels of capital and pro tability (Schaeck and Cihak (2010b)). Moreover, given the unbalanced nature of our panel dataset, it avoids that the denominator is computed over di erent window lengths for di erent 19

20 banks. The Z-score can be interpreted as the number of standard deviations by which returns would have to fall from the mean to wipe out all equity in the bank (Boyd and Runkle (1993)). A higher Z-score implies a lower probability of insolvency, providing a more direct measure of soundness than, for example, simple leverage measures. Because the Z-score is highly skewed, we use the natural logarithm of Z-score to smooth out higher values 9. Table 1 shows that the average value of ln(z-score) slightly exceeds four with a standard deviation of 1:15. Our indicator of market power is signi cantly correlated with the Z-score and other indicators of competition, but also with indicators of market concentration. Table 2 presents correlations on the country-year level between the Z-score, the Lerner index if market power as well as bank and country-level indicators of competition and market structure. The Lerner index is positively and signi cantly correlated with the Z-score, consistent with Figure 1. The Lerner index is negatively and signi cantly correlated with the average market share of banks in a country and year, while there is no signi cant correlation with the number of banks. The Lerner index is higher for more concentrated banking markets, as measured by the Her ndahl index. Interestingly, there is no signi cant correlation between the Lerner index of market power and two other industry-level behavioral indicators of bank competition, the H-Statistics and the Boone index. 4 The competition-stability relationship: explaining cross- 9 Others have used the transformation ln(1+z-score) to avoid truncating the dependent variable at zero. We take the natural logarithm after winsorizing the data at the 1% level. As none of the Z-scores is lower than zero after winsorizing, this approach is similar, save for a rescaling, to the former approach and winsorizing after the transformation. 20

21 country variation 4.1 The competition-stability relationship: Cross-country heterogeneity In the Introduction, we mentioned that the full sample unconditional correlation between the Lerner index and the Z-score is positive (0:31), but that this number hides a substantial amount of cross-country variation (see Figure 1). Using a regression framework, we show that this relationship also holds when conditioning on other variables. Following common practice in the literature, we regress the bank soundness measure on the Lerner index and a wide range of control variables using our total sample. The results are presented in table 3. <Insert Table 3 around here> The rst column shows OLS estimates, whereas the second column are IV (2SLS) regression results. In the second column, we take into account that market power may be endogenous. The instruments are loan growth and lagged values of the Lerner index. We employ the panel structure of the database and control for xed heterogeneity at the country and time level by including country and time xed e ects. The standard errors are robust and clustered at the bank level. Moreover, to avoid that our results are driven by countries that are overrepresented in our sample, we weigh each variable with the inverse of the number of banks in the country. Doing so, we give equal weight to each country. We again nd a positive and signi cant e ect of a change in market power on bank stability. This result is in line with existing literature that also uses the Lerner index as a market power proxy (see, e.g. Berger, Klapper, and Turk Ariss (2009)). However, as already mentioned in the introduction, our interest is less on arguing that competition is good or bad, but rather on uncovering which country characteristics make the impact better of worse. Therefore, it is crucial to show that the regression-based methods also indicate that there is a substantial degree of heterogeneity. 21

22 <Insert Figure 2 around here> Figure 2 is very similar to Figure 1 of the introduction. The correlation between the conditional and unconditional correlation is 0:69 and highly signi cant. Both bar charts show that the unconditional and conditional correlations are positive in most countries. In Figure 2, the height of the bars shows the magnitude of the coe cient of the Lerner index when estimating Equation (2) for each country separately. The bars are sorted from low to high and the country labels are mentioned on the X-axis. The coe cients that are signi cantly di erent from zero have a lighter shade. The average of the 62 estimated coe cients equals 0.982, which resembles the full sample coe cient. Hence, on average, it seems that the franchise value paradigm dominates the risk-shifting hypothesis. In response to an increase in market power, banks will behave more prudently to protect their monopoly rents that create a larger franchise value. Or vice versa, an increase in competition increases banks appetite for risk-taking. However, there is a large amount of heterogeneity in the competition-stability relationship. The standard deviation of the coe cient across the 62 countries is A quick look at the country labels on the X-axis also reveals that it is not just a developed versus developing countries story or that regions exhibit similar behavior. In the remainder of this section, we will empirically explore what drives this high cross-country variation in the competition-stability relationship. 4.2 A binary classi cation approach Our goal is to shed light on the underlying factors that drive this heterogeneous impact. Uncovering which institutional features drive the cross-sectional variation in b j will allow us to identify which banking sectors will not be harmed (or harmed less) by more intense banking competition. To assess the impact of regulatory, supervision and institutional features on the estimated competition-stability trade-o, we rst classify the sample countries into two distinct groups ac- 22

23 cording to the median value of each such speci c feature. We perform two sorts of tests 10, which di er in their treatment of potential cross-country di erences in the impact of bank-speci c characteristics on bank risk. More speci cally, we employ a SEPARATE and POOLED approach. Both approaches di er in the amount of heterogeneity they allow in the relationship between control variables and bank risk. In the SEPARATE test methodology, we employ t-tests to compare the di erence in magnitude of the competition-stability relationship across the two groups. More speci cally, we perform t-tests on two measures. On the one hand, we look at variation in the (unconditional) correlation between the Z-score and the lagged Lerner index. On the other hand, we also estimate Equation (2) separately for each country, and then average the estimated values within each group formed based on a particular institutional characteristic. The POOLED methodology estimates Equation (2) across all banks from countries with a speci c institutional feature. We thereby impose common slopes (both on the Lerner index and the control variables) within each group. For example, we estimate equation (2) across all countries with weak activity restrictions, and then across all banks in the strong activity restriction countries. We thus estimate a single coe cient for each group of countries 11, and test whether the coe cients di er between the two groups of countries. To compare the impact of the Lerner index on bank stability across the two groups, we employ Chow tests. The POOLED approach is implemented using OLS as well as IV(2SLS). According to Oztekin and Flannery (2008), the separate and pooled methodologies each have its own merits. Averaging individual country regressions in the 10 Oztekin and Flannery (2008) perform a similar kind of analysis to examine how country-speci c features a ect the adjustment speed at which rms converge to their target leverage. 11 Larger countries with more banks may be overrepresented in a particular group. For example, the majority of banks within the sample operate in the US. On the one hand, could this lead to US-biased results, on the other hand does it make a fair comparison with the separate approach invalid. Therefore, we give equal weight to each country in the pooled approach. The weight that each individual bank observation gets is proportional to 1=n i, where n i equals the number of observations for country i. Moreover, we include time and country xed e ects and cluster the standard errors at the bank level. 23

24 SEPARATE method allows for full heterogeneity in parameter estimates and error variances. However, estimating Equation (2) for countries with few banks might also yield noisy coe cient estimates. The POOLED method assumes slope and error variance homogeneity across countries, raising the possibility that the gains from pooling would outweigh any costs imposed by ignoring the inherent heterogeneity in the slope estimates. Since we believe that neither method is superior, we document our results using both approaches. Table 5 consists of two panels. The rst panel corresponds to the separate approach, whereas the second panel contains the results of the pooled approach. Summary statistics on the countryspeci c variables are reported in Table 4. Variable de nitions and sources are reported where they occur for the rst time in the text. The country-speci c variables are obtained from various sources and may vary at a di erent level. Some variables are (almost) constant over time, other vary by country and over time. Moreover, not all variables are available for all countries or all time periods. <Insert Table 4 around here> In Table 4, the summary statistics of the country-speci c variables are categorized in four groups. The results on the SEPARATE and POOLED tests will be described in a similar order Herding As discussed above, we use two proxies of herding. The heterogeneous banking system indicator measures whether there are substantial revenue di erences among di erent nancial institutions within a country. It is calculated as the within country standard deviation of the non-interest income share. If all banks in a country have a similar business model (either voluntarily or forced by regulation), the indicator will be low. A higher value indicates that the banking system is more diverse. In heterogeneous banking systems, an increase in competition is less detrimental for bank risk compared to homogeneous banking systems. A second indicator of the too-many- 24