International Evidence on the Impact of Macroprudential Policies on Bank Risk Taking and Systemic Risk

Transcription

1 International Evidence on the Impact of Macroprudential Policies on Bank Risk Taking and Systemic Risk Wenqing Gao Thomas W. Miller, Jr. Alvaro G. Taboada Mississippi State University Mississippi State University Mississippi State University January 2018 Abstract We examine the impact of the adoption of borrower-targeted and financial institution-targeted macroprudential policies on bank risk taking and systemic risk. Using a difference-in-differences (DiD) methodology and a sample of 3,357 banks from 77 countries over the period , we find that macroprudential policy instruments mitigate bank risk taking and lower systemic risk. The results are driven by borrower-targeted policies, not by those that target financial institutions assets or liabilities. The impact of borrower-targeted macroprudential policies on systemic risk is stronger in countries with more closed economies, with less concentrated banking sectors, and with tighter restrictions on bank activities. Policies adopted after the global financial crisis (GFC) are less effective in constraining risk taking.

2 1. Introduction The global financial crisis (GFC) and the systemic risk buildup that preceded it has led to a push for tougher regulations aimed at constraining excessive bank risk taking. An increasingly popular approach taken by regulators and policy makers around the world is to use macroprudential policies aimed at stabilizing the financial sector by constraining bank risk taking. These macroprudential tools have been commonly used in both emerging and developed markets, although they are more widely used in the former (Cerutti, Claessens and Laeven, 2017). The widespread use of these macroprudential policy tools has led to a new and growing literature aimed at assessing their effectiveness (e.g. Cerutti et al. 2017; Claessens, 2015). The evidence to date on the effectiveness of these policies is mixed (see e.g. Claessens, 2015; Galati and Moessner, 2013; Aiyar, Calomiris, and Wieladek, 2014) and few studies (e.g. Altunbas, Binici, and Gambacorta, 2017) examine their impact on bank risk taking. In this paper, we expand on the existing literature and shed some light on the debate about the effectiveness of macroprudential tools by examining the impact of the adoption of macroprudential policies on bank risk taking and systemic risk. By focusing on the impact of the initial adoption of borrower-targeted and financial institution-targeted macroprudential instruments, we are able to use a shock-based research design that mitigates the endogeneity concerns inherent in previous studies. 1 In addition, the staggered implementation of these policies across countries allows us to isolate their impact and understand the causal effects. Focusing on a country s first adoption of macroprudential policies has some advantages in terms of identification. This advantage comes with costs. We cannot examine the impact of changes (tightening or loosening) or the intensity of macroprudential tools on risk taking. Also, we are unable to assess 1 See e.g. Claessens, 2015 for a discussion of the identification and endogeneity challenges faced by studies about the effectiveness of macroprudential policies. 2

3 the extent to which these policy tools are binding. To the extent that these macroprudential policy tools are not binding, however, this condition should bias against finding significant results. In addition, by using a broad cross-country sample, we are able to assess how differences in ex-ante institutional quality and regulatory environment impact the effectiveness of these policies. Using a difference-in-differences (DiD) methodology and a broad sample of 3,357 banks from 77 countries from , we find evidence that the adoption of borrower-targeted macroprudential policy instruments (e.g. loan-to-value limits (LTV caps) and debt-to-income ratio limits) is associated with reductions in bank risk taking. We further find that borrower-targeted macroprudential policies lower systemic risk, especially in more closed economies, in countries with less concentrated banking sectors, and in countries with more restrictions on bank activities. We do not observe any significant impact on risk taking associated with financial institutiontargeted instruments. While constraining risk taking, the adoption of these policies does not adversely affect bank profitability and lending. Finally, we document that macroprudential tools adopted after the GFC are less effective than those adopted before the GFC. Our results are robust to a battery of tests that aim to address potential concerns with the DiD methodology. To address concerns about confounding events driving our results, we control for the effects of two major events that took place in many countries during our sample period, specifically, the adoption of International Financial Reporting Standards (IFRS), and the enactment of major takeover reforms (Lel and Miller, 2015). In addition, we examine the impact on bank risk taking in the years leading up to the adoption of macroprudential polices and find that the decline in bank risk taking and systemic risk occurs after the adoption of such policies. We also test the robustness of our results to the use of alternate control groups, including a propensity score matched (PSM) sample of banks. Finally, we introduce additional controls to account for 3

4 the impact of alternate tools that might be used to foster financial stability, including monetary and fiscal policy tools. In all cases, we find qualitatively similar results. Our study builds on Cerutti et al. (2017) who introduce an updated database on the use of macroprudential policies in 119 countries between 2000 and Cerutti et al. (2017) find that the use of macroprudential policies is associated with lower credit growth, especially household credit at the country level. Their results are more pronounced in emerging markets with more closed economies. We expand on their study by examining the effect of the adoption of macroprudential policies on risk taking behavior and systemic risk. In addition, we provide evidence at the bank level, which allows us to examine bank characteristics that could affect the effectiveness of these policies in constraining bank risk taking and systemic risk. Our study contributes to the existing literature that examines the effectiveness of macroprudential policies. These studies have assessed the impact of macroprudential policies on financial stability (see e.g. Lim et al., 2011; Bakker et al. 2012; Claessens et al., 2013), real estate prices and housing credit (Kuttner and Shim, 2016; Crowe and Dell Ariccia, 2013; Zhang and Zoli, 2016) and equity flows (e.g. Bruno and Shin, 2014; Zhang and Zoli, 2016)). The evidence to date is mixed. Some studies show that macroprudential policies are effective in reducing credit growth, the procyclicality of credit, and leverage (Lim et al., 2011), in curbing real estate booms (Crowe and Dell Ariccia, 2013), and in reducing the sensitivity of capital flows to global conditions (Bruno and Shin, 2014). Yet, other studies suggest that macroprudential policies can be circumvented, which reduces their effectiveness, and in some cases can even create negative externalities (Aiyar et al., 2014; Cizel, Houben, and Wierts, 2016). Aiyar et al. (2014) examine the relationship between counter-cyclical capital requirements and lending by both regulated and unregulated banks in UK. Their results show that regulated banks respond to tighter capital 4

5 requirements by reducing lending, while unregulated banks increase lending. Cizel et al. (2016) find evidence that credit shifts from bank-based financial intermediation towards nonbank intermediation following the adoption of macroprudential policies, reducing the policies effect on total credit. Our study sheds light on the ongoing debate about the effectiveness of macroprudential policies. By focusing on the impact on bank risk taking and systemic risk, we expand on previous studies by exploring the extent to which these policies achieve a main objective. Namely, to curtail the buildup of systemic risk in the financial system due to externalities and market failures (see e.g. De Nicolò, Favara, and Ratnovski, 2012; Brunnermeier et al., 2009). As Claessens (2015) points out, many of these studies struggle with identification and endogeneity, which are issues that can only be partially addressed with econometric techniques such as GMM. 2 A contemporaneous and closely related paper by Altunbas et al. (2018) examine the impact of macroprudential policies on bank risk taking. Altunbas et al. (2018) study the effects of macroprudential policies on bank risk taking in 61 economies over the period from 1990 to They find that tightening macroprudential policies is associated with reductions in bank risk taking. They further document that the impact is stronger for banks that are weakly capitalized, smaller, with low liquidity buffers and fewer deposits. Our study expands on Altunbas et al. (2018) and other related studies by using a DiD methodology that takes advantage of the staggered introduction (first adoption) of macroprudential policies across a broader set of countries. Focusing on the impact of a country s first adoption of macroprudential policies allows us to address endogeneity concerns and to provide more direct evidence of a causal relationship between 2 Claessens (2015) and Galati and Moessner (2013) provide thorough reviews of the literature on the impact and effectiveness of macroprudential policies. 5

6 the onset of these policies and bank risk taking. 3 Our study also adds to the findings in Altunbas et al. (2018) by assessing the impact of the staggered implementation of macroprudential policies on systemic risk. Finally, we assess how de facto and de jure country characteristics influence the effectiveness of macroprudential policies in constraining risk taking and systemic risk. 2. Data and Methodology 2.1. Data We obtain data from several sources. Data on the introduction of macroprudential policies comes from the updated database in Cerutti, et al. (2017), which uses data collected from the IMF survey: Global Macroprudential Policy Instruments (GMPI), carried out by the IMF s Monetary and Capital Department during The database provides indicators for the use of 12 macroprudential policy instruments in 119 countries from These instruments are grouped and aggregated into two categories: 1) Borrower-targeted, which are aimed at borrowers leverage and financial positions, and 2) Financial institution-targeted, which are aimed at financial institutions assets or liabilities. 5 We use the database to identify countries that introduce a borrower-related or financial institution-targeted instrument for the first time after Because the database starts in 2000, we cannot determine whether countries that have a macroprudential 3 As Claessens (2015) points out, many of the studies that assess the effectiveness of macroprudential policies struggle with identification and endogeneity, which are issues that can only be partially addressed with econometric techniques such as GMM Focusing on the first adoption of a macroprudential policy in a country, as opposed to changes (tightening or loosening) in policies allows us to use a shock-based research design as an alternate way to mitigate the endogeneity concerns inherent in prior studies. 4 While the GMPI survey covers 18 different instruments, because of data availability, the Cerutti et al. (2017) database covers only 12 of these. 5 Specifically, borrower-targeted instruments include loan-to-value caps (LTV_CAP) and debt-to-income ratios (DTI). Financial institution-targeted instruments include: time varying loan loss provisioning (DP), countercyclical capital requirements (CTC), leverage ratios (LEV), capital surcharges on systemically important financial institutions (SIFI), limits on interbank exposures (INTER), concentration limits (CONC), limits on foreign currency loans (FC), reserve requirements ratios (RR_REV), limits on domestic currency loans (CG), and levy/taxes on financial institutions (TAX). Appendix A has definitions for these and other variables use in our analyses. 6

7 policy in place in 2000 adopted the policy in 2000 or before. Because of this fact, our treatment group of countries includes only those that adopt a macroprudential policy for the first time after We obtain financial data on all commercial banks, savings banks, cooperative banks, and bank holding companies in all 119 countries from Fitch Fundamentals Financial database. 6 In addition, we collect stock price data and market index returns used to construct our market-based risk proxies from Bloomberg. Our primary measure of bank risk taking is the logarithm of z-score (Z-score). Z-score measures the distance from insolvency and is estimated as the sum of return on assets (ROA) and the ratio of equity-to-assets divided by the standard deviation of ROA (Roy, 1952; Laeven and Levine, 2009). In addition to Z-score, we use three additional variables to measure bank risk taking: NPL-to-loans, non-performing loans-to-gross loans;, Equity volatility, the annualized standard deviation of weekly stock returns, and; Tail risk, the negative of the average return on bank s stock over the 5% worst return days in a given year, following Ellul and Yerramilli (2013). Because stock price data from Bloomberg are only available for large publicly traded banks, our sample size is vastly reduced when using the market-based proxies for risk (Equity volatility and Tail risk). In addition to bank risk taking, we obtain two variables to measure systemic risk. Our main measure of systemic risk is the marginal expected shortfall (MES) measure from Acharya, et al. (2017). We compute MES as the average bank return during the worst 5% of market return days in a year. We estimate MES for all publicly traded banks in our sample with available data on stock prices from Bloomberg. Our second measure of systemic risk is R-squared, which measures 6 For countries that are covered in the Claessens and van Horen (2014) database, we restrict our sample of banks to those banks covered in that database. For other countries, we obtain data from Fitch Fundamentals on banks classified as commercial banks, savings banks, cooperative banks and bank holding companies. 7

8 the total variation of returns of a given bank explained by returns of all other banks in a given country (see e.g. Anginer, Demirgüç-Kunt and Zhu, 2014). We obtain R-squared by estimating weekly returns of a bank on the average returns of all other banks in a given country. Following other studies (Morck, Yeung, and Yu, 2000; Anginer, et al., 2014), our measure of systemic risk is the logistic transformation of R-squared: log ((R 2 i,j,t) / (1- R 2 i,j,t)).. From World Bank databases, we collect country-level data on measures that have been shown to influence bank risk taking and systemic risk (e.g. Demirgüç-Kunt and Huizinga, 2010; Engle, Jondeau, and Rockinger, 2015; Brunnermeier, et al., 2015). To control for financial development and growth we use the growth in real GDP (GDP growth) obtained from the World Development Indicators database. From the Global Financial Development Database (GFD), we obtain the total banking sector assets scaled by GDP (Bank sector assets); the non-interest income to total income (Non-interest income) to proxy for the extent of noncore banking activities; the proportion of banking assets held by the three largest banks (Concentration); stock market index returns (Market return), and stock market volatility (Volatility), which is the annualized standard deviation of weekly stock market index returns. At the bank level, we use six variables to control for bank-specific characteristics including bank size, measured as the logarithm of total assets (Size), bank noninterest income (Bank noninterest income), non-deposit short term funding (ST funding), leverage (Leverage), measured as total assets divided by total equity, return on assets (ROA), and noninterest expense scaled by total income (Noninterest expense). Finally, we obtain data on additional de facto and de jure country characteristics. We obtain two de facto measures of financial market openness. Financial openness is the sum of external assets and liabilities, scaled by GDP from the updated dataset from Lane and Milesi- Ferretti (2007)., Foreign bank ownership is foreign-owned banking assets-to-total banking sector 8

9 assets, calculated using data from Claessens and Van Horen (2014). Finally, we obtain data on two de jure measures of regulatory quality from Barth, Caprio, and Levine (2013). Restrictions on bank activities is an index measuring regulatory impediments to banks engaging in securities market activities, insurance activities, and real estate activities. Official supervisory power is an index measuring whether supervisory entities have authority to take action to prevent and correct problems. The definitions of all variables appear in Appendix A. Our initial sample consists of 4,634 banks from 117 countries. We drop banks with missing data on total assets, as well as those with negative book value of equity. We also exclude banks in countries with missing data on control variables. Our final sample consists of 3,357 banks from 77 countries, with 33,833 bank-year observations. Table 1 lists the final sample of countries in our study, as well as the first year of adoption of each type of macroprudential policy. We observe that many countries adopt (or have macroprudential policies in place) in In our empirical tests, these countries serve as our control group, along with countries that did not adopt any macroprudential policies during our sample period. Table 1 shows that 27 (19) countries adopt borrower-targeted (financial institutiontargeted) macroprudential policies for the first time during our sample period. Moreover, there are eight countries that adopt both borrower- and financial institution-targeted macroprudential policies for the first time during our sample period. These countries are Hungary, Indonesia, Latvia, Netherlands, Romania, Serbia, South Korea, and Thailand. In Table 2, we show descriptive statistics of our main bank- and country-level variables. The average (median) bank in our sample has $3.4 ($3.2) billion in assets. On average, banks derive 21% of their income from nontraditional banking activities, and about 25% of their funding comes from non-deposit short-term funding (ST funding). Average Tail risk is 4.4%, which is 9

10 similar to what Ellul and Yerramilli (2013) report for their sample of US bank holding companies (4.7%). Not surprisingly, because our sample period covers the GFC, average ROA is low, 0.88% Methodology To test the average effect of the introduction of macroprudential policies on bank risk taking and systemic risk, we use a DiD design in which our treatment group consists of banks in countries that adopt macroprudential policies between 2001 and Because our setting involves multiple treatment groups and time periods (Wooldridge, 2007), we use firm fixed effects and year fixed effects. The firm and year fixed effects identify the within-firm and within-year change in risk taking and systemic risk between treatment and benchmark banks when countries adopt these macroprudential policies. This approach implicitly takes as the benchmark group all banks from countries that did not adopt a macroprudential policy as of a particular time as well as banks in countries that have macroprudential policies in place since the start of the period (Bertrand and Mullainathan, 2003). We use various specifications of the following model: YY ii,cc,tt = αα + ββ 1 PPPPPPPP TTTTTTTTTT + ββ 2 TTTTTTTTTT + ββ 3 PPPPPPPP + ββ 4 XX cc,tt 1 + ββ 5 ZZ ii,cc,tt 1 +δδ tt + θθ ii + εε iiiiii (1) Yict refers to our set of bank risk taking and systemic risk proxies: Z-score; NPL-to-loans; Equity volatility; Tai risk; MES, and R-squared. Post is a binary variable that is equal to one starting the year a borrower- or financial institution- targeted macroprudential policy is adopted in a country and zero otherwise. For countries in the control group, we set Post equal to one after 2011 and zero otherwise. 7 Treatc,t is a binary variable that is equal to one for banks in countries that adopt a borrower- (financial institution) targeted macroprudential policy during our sample 7 Most countries implemented macroprudential policies after In unreported results, we use alternate years to define the post-period for the control group of countries and obtain similar results. 10

11 period and zero otherwise. Because the macroprudential database starts in 2000, we do not know whether countries with a macroprudential policy in place in 2000 adopt the policy in 2000 or before Consequently, our treatment group consists of countries that adopt a macroprudential policy for the first time after In Equation (1), Xc,t-1 is a vector of country-level control variables that are associated with bank risk taking and systemic risk, as described in the previous section. Country-level control variables include: 1) GDP growth; 2) Volatility; 3) Market return; 4) Noninterest income; 5) Bank sector assets, and 6) Concentration. Zi,c,t-1 is a vector of bank-level control variables, which consists of: 1) Size; 2) Bank noninterest income; 3) ST funding; 4) Leverage; 5) ROA, and 6) Noninterest expense. Finally, δt and θi are year and firm fixed effects, respectively, and εi,c,t is the error term. Our variable of interest is the coefficient on the interaction term (ββ1). If the adoption of macroprudential policies is associated with reductions in bank risk taking and systemic risk, we expect ββ1 to be positive (negative) and significant when using Z-score (all other bank risk measures). In all regression estimations, we use robust standard errors clustered by country because the adoption of macroprudential policies is a country-level decision. Our DiD approach compares changes in bank risk taking and systemic risk subsequent to the adoption of macroprudential policies with changes in bank risk taking in countries without such policies during the corresponding years (or those countries where such policies were adopted earlier). By doing so, we aim to isolate the effect of these macroprudential policies from other confounding factors that potentially affect bank risk taking and systemic risk. 8 Using this approach runs the risk of misclassifying countries that adopt macroprudential policies for the first time in However, by including these countries in the control group, we are biasing against finding significant results, as banks in these countries will have experienced the impact of the adoption of such policies. 11

12 3. Empirical analysis 3.1. Impact of macroprudential policies on bank risk taking We first assess the average effect of the adoption of macroprudential policies on bank risk taking. To do so, we estimate Equation (1) using the four bank risk taking measures as the dependent variable. Table 3 shows the results. In Panel A we report results using the adoption of borrower-targeted macroprudential policies, while Panel B shows results using the adoption of financial institution-targeted policies. The results in Panel A of Table 3 show strong evidence that the adoption of borrowertargeted macroprudential policies are effective in constraining bank risk taking. In Models (1), (3), (5), and (7), we report results from our baseline models using Z-score; NPL-to-loans; Equity volatility, and Tail risk, respectively. The coefficient on the interaction term, (Post x Treat) is statistically significant in all regression specifications. This significance is consistent with the view that the adoption of borrower-targeted macroprudential policies is associated with a reduction in bank risk taking. The results are both statistically and economically significant. Taking the coefficient in Model (1), results suggest that after the adoption of borrower-targeted macroprudential policies, Z-score increases by for banks in treated countries, which represents a 3.5% (10%) increase relative to its mean (standard deviation). 9 In contrast, banks in the control group do not experience a significant change in Z-score in the post-period (coefficient on Post is with a t-statistic of 1.46). Results using NPL-to-loans are similar, but of larger magnitude. From Model (3), banks in 9 From Model (1) of Panel A of Table 3, banks in the treatment group of countries experience a [ ] increase in Z-score after the adoption of borrower-targeted macroprudential policies. This represents a 3.47% increase relative to its mean (3.54 from Table 2) or a 10% increase relative to its standard deviation (1.23). 12

13 the treatment group of countries experience a reduction in NPL-to-loans of -0.98%, which represents a 15.5% reduction relative to its mean (6.3%). The results using the stock market-based proxies of bank risk are similar in magnitude and significance, despite the smaller sample size when using such measures. As an example, taking the coefficients in Model (7), while Tail risk increases by 0.5% in the post-period for banks in the control group of countries, Tail risk declines by 0.4% for banks in the treatment group, which represents a 9.1% reduction relative to its mean (4.4%). 10 Results are similar when using Equity volatility. A key assumption of our identification strategy (and any DiD design) is that treatment and control banks follow similar trends before the adoption of macroprudential policies. To test this parallel trends assumption, we augment Equation (1) by introducing an indicator variable, Pre, that is equal to one in years t-3, t-2 and t-1 relative to the adoption of the policy, and zero otherwise. We also include the interaction term (Pre x Treat). If bank risk taking in treatment and control group of banks follows similar patterns, we expect the coefficient on the interaction term (Pre x Treat) to be insignificant. We show results from the estimation of the augmented Equation (1) in Models (2), (4), (6), and (8) of Panel A of Table 3. The results confirm our prior findings that the adoption of borrower-targeted macroprudential policies is associated with reductions in bank risk taking. In all regression specifications we observe significant coefficients on the interaction term (Post x Treat). Furthermore, the results suggest that treatment and control group of banks followed similar patterns before the adoption of the macroprudential instruments; the coefficient on the interaction term Pre x Treat is never significant. Overall, our results show that borrower-targeted 10 From Model (7) of Panel A of Table 3, banks in the treatment group of countries experience a [ ] decrease in Tail risk after the adoption of borrower-targeted macroprudential policies. This represents a 9.1% decrease relative to its mean (0.044). 13

14 macroprudential policies are associated with reductions in bank risk taking, which falls after the adoption of such policies. In Panel B of Table 3, we replicate results from Panel A, using the adoption of financial institution-targeted macroprudential policies. To simplify the presentation, we do not report the coefficients on the control variables. We find no evidence that the adoption of financial institutiontargeted macroprudential policies significantly affect bank risk taking. The coefficient on the interaction term, (Post x Treat) is not significant in any of the regression specifications in Panel B. Several countries in our sample adopt both borrower- and financial institution-targetedpolicies during our sample period; thus, our results in Panel A could be driven by countries that adopt both types of policies. To further assess whether the reduction in bank risk taking is driven by borrower-targeted macroprudential, in Panel C we show results in which we include Post indicators and the interactions with Treat, using both borrower- (Post Borrower) and financial institution-targeted (Post FI) macroprudential policies. The results in Panel C confirm our prior findings that borrower-targeted and not financial institution-targeted policies are effective in constraining bank risk taking. The interaction term Post Borrower x Treat is statistically significant at the 5% level or better in all regression specifications, while the Post FI x Treat is insignificant. The coefficients on Post Borrower x Treat are almost identical to those in Panel A. One concern with our results thus far is that we are using all banks from the control group of countries. Thus, our results could reflect inherent differences between banks in the control and treatment group. To address this concern, we use a propensity score matching (PSM) approach to identify comparable control banks. The propensity-score-matching approach involves pairing treatment and control banks based on similar observable characteristics (Dehejia and Wahba, 14

15 2002). We implement this procedure by first estimating a logit regression to model the probability of being a sample treatment bank using all controls in our main regressions. Next, we estimate the propensity score for each bank using the predicted probabilities from the logit model. We then match each treatment bank to the control banks using the nearest neighbor matching technique (with replacement). We report the results from the logit model estimation in the internet appendix, along with tests that validate the quality of our matching procedure. Panel D of Table 3 reports results from the estimation of Equation (1) using only the PSMmatched sample. The sample size is significantly reduced, but the results continue to show that the adoption of borrower-targeted macroprudential policies is associated with a reduction in bank risk taking. The statistical significance of the results is lower, likely due to the reduction in sample size, but the economic magnitude of the results is similar to the one using the full sample. As an example, from the coefficients in Model (1), in Panel D of Table 3, relative to the PSM-matched control banks, banks in the treatment sample experience a higher increase in Z-score after the adoption of borrower-targeted macroprudential instruments, which represents 5.3% increase relative to its mean (3.49), or 16% of its standard deviation (1.15). Results using the alternate measures of bank risk taking are similar to those reported earlier, except that the coefficient on Post x Treat is insignificant when using NPL-to-loans as the dependent variable. Overall, the results in this section show support to the view that the adoption of borrowertargeted macroprudential policies can be effective in constraining bank risk taking. In contrast, we find no evidence that the adoption of financial institution-based macroprudential instruments affect bank risk taking. In the rest of the paper, we thus focus on the impact of borrower-targeted policies. 15

16 3.2. Impact of macroprudential policies on systemic risk We now turn to assess the impact of the adoption of borrower-targeted macroprudential policies on systemic risk. We use two widely used proxies of systemic risk: MES, from Acharya et al. (2017), and R-squared (Anginer et al., 2014). Our analysis parallels the one used in the prior section (Panels A and D of Table 3). Table 4 shows results from the estimation of Equation (1) using our proxies for systemic risk. The results in Panel A of Table 4 show some evidence that the adoption of borrowertargeted macroprudential policies lowers systemic risk, as measured by MES. The impact is both statistically and economically significant. Taking the coefficient in Model (1), while MES increases by 0.4% in the post-period for banks in the control group of countries, MES declines by 0.3% for banks in the treatment group, which represents a 15% reduction relative to its mean (2%). While the coefficient on Post x Treat is negative in the regressions using R-squared, the results are not statistically significant. This result may be explained in part by the smaller sample size. In Panel B, we show results using the PSM-matched sample for the borrower-targeted macroprudential policies. Results confirm our prior findings that the adoption of borrowertargeted macroprudential instruments is associated with a reduction in MES. Banks in the treatment group experience a (20% of its mean) larger reduction in MES after the adoption of borrower-targeted macroprudential instruments relative to PSM-matched sample Impact of de facto and de jure country characteristics on bank risk taking De facto Characteristics We next examine how country de facto characteristics including financial openness and the size and structure of the banking sector affect the impact of macroprudential policies. To this end, we first sort countries based on four de facto measures: 1) Bank sector assets domestic bank 16

17 assets-to-gdp; 2) Concentration assets of the three largest commercial banks as a share of total commercial banking assets; 3) Financial openness the sum of external assets and liabilities, scaled by GDP from the updated dataset from Lane and Milesi-Ferretti (2007); and 4) Foreign bank ownership- foreign-owned banking assets-to-total banking sector assets. 11 We average each measure by country over and rank countries based on the cross-country median. First, we explore the extent to which a country s financial openness affects the effectiveness of borrowertargeted instruments. In more open economies, borrowers have alternate sources of financing such as nonbanks, in addition to being able to obtain funding through cross-border activities. The imposition of tighter restrictions on borrowers, such as lower LTV ratios may thus provide an incentive for borrowers in such counties to seek alternate sources of financing. The impact of the adoption of borrower-targeted instruments may thus be less effective in economies that are more open. To examine this hypothesis, we classify countries as Open (Closed) if its average Financial openness is above the cross-country median. As a second proxy for openness, we use the ratio of foreign banking assets-to-total assets, Foreign bank ownership. In addition, we explore the role of the size and structure of a country s banking sector, using Bank sector assets and Concentration. Using each measure, we create an indicator variable Large (High concentration) that is equal to one for countries with above median values, and zero otherwise. In Panels A and B of Table 5 we show results from estimations of Equation (1) separately for economies based on the two proxies for financial openness, while Panels C and D show results using the proxies for banking sector size and concentration, respectively. The results in Panel A 11 To construct this measure, we use the updated Claessens and van Horen (2014) database and obtain data on banks assets from Fitch Fundamentals Financial data. We sum the assets of all foreign owned banks in each country, scaled by the assets of all commercial banks, savings banks, and bank holding companies with data available from Fitch Fundamentals database. 17

18 of Table 5 show some evidence that the impact of borrower-targeted macroprudential polices is stronger in more closed economies. While there is no significant difference on the impact on Z- score, the results in Panel A of Table 5 reveal that the adoption macroprudential policies significantly reduce NPL-to-loans, Equity volatility and Tail risk in more closed economies, while the impact is insignificant in open economies. For Equity volatility and Tail risk, the difference between Open and Closed economies is significant (p-value of χ 2 test is and 0.078, respectively). We find similar, albeit weaker results, using Foreign bank ownership as a proxy for openness (Panel B). The impact of macroprudential policies in constraining bank risk is stronger in countries with lower foreign ownership of banks, although the difference is only significant when using Z-score. 12 Our results are consistent with the findings in Cerutti et al. (2017) who document that the impact of macroprudential policies on credit growth is stronger in countries with more closed economies. This supports the view that borrowers in such countries may not have as many alternate funding sources, which may enhance the value of macroprudential policies that target borrower quality. In Panels C and D of Table 5, we assess how banking sector size and structure impact our results. While the results in Panel C do not show evidence that the effectiveness of macroprudential policies varies by size of the baking sector, results in Panel D show that the adoption of these policies is more effective in constraining bank risk taking in countries with less concentrated banking systems. Banks in less concentrated banking sectors have been shown to have higher levels of nontraditional banking activities, which are associated with higher systemic risk (Engle, et al., 2014; Brunnermeier, Dong, and Palia, 2015). These findings suggest that 12 The last column of Panel B of Table 5 shows the p-values from χ 2 tests of differences in the coefficient Post x Treat between High and Low foreign ownership countries. 18

19 macroprudential policies may be effective in constraining bank risk taking in countries with higher systemic risk. In the next section, we explore the impact on systemic risk De jure characteristics We next explore whether the effectiveness of macroprudential policies varies based on de jure measures of regulatory quality. To test this, we rank countries based on two measures of regulatory quality from Barth, et al. (2013): 1) Restrictions on bank activities, and 2) Official Supervisory power. We average each measure by country and classify countries as High (Low) if the value is above (below) the cross-country median. It is unclear how regulatory quality may affect the impact of the adoption of macroprudential policies. For example, through tighter restrictions on bank activities, regulators may reduce banks ability to engage in riskier banking activities, such as nontraditional bank activities. If this is the case, the implementation of additional policies aimed at constraining bank risk taking may not be as effective in such countries in which bank risk taking may already be low. In contrast, in countries with more restrictions on bank activities, if properly enforced, banks may be unable to escape the constraints imposed by new macroprudential policies, which would make them more effective in such countries. The results from estimations of Equation (1) using cuts based on de jure measures of regulatory quality are shown in Panels E and F of Table 5. The results in Panel E show that the impact of macroprudential policies is stronger in countries with tougher restrictions on bank activities. In all regression specifications the coefficient of Post x Treat is statistically significant for countries with High restrictions, but insignificant in countries with Low restrictions. The difference in the coefficient Post x Treat between countries with High and Low restrictions is statistically significant in regressions using NPL-to-loans and Equity volatility. 13 These results are 13 From Panel E of Table 5, the p-value of the F-test is (0.065) in regressions using NPL-to-loans (Equity volatility). 19

20 consistent with the view that these policies are more effective in constraining risk taking in countries in which banks may be constrained to engage mostly in traditional lending activities. In Panel F, we show results for regressions in which we classify countries base on Official supervisory power. We do not find evidence that the effectiveness of macroprudential policies differs based on Official supervisory power. While the impact of macroprudential policies on Z- score is larger in countries with Low supervisory power, none of the coefficients on Post x Treat are significantly different across groups based on the other three bank risk taking proxies. Overall, in this section we find evidence that the impact of macroprudential policies on bank risk taking is concentrated in countries with tougher restrictions on bank activities. The power of supervisors does not seem to affect the effectiveness of these policies in curbing risk taking. We next turn to assess how country characteristics affect the impact of these policies on systemic risk Impact of de facto and de jure country characteristics on systemic risk We now turn to explore the impact of borrower-targeted macroprudential policies on systemic risk and the extent to which de facto and de jure country characteristics affect their effectiveness. 14 We follow a similar approach as in the prior section examining the impact on bank risk taking. We use the two measures of systemic risk: 1) MES and 2) R-squared. In Panels A and B of Table 6 we show results from regressions of systemic risk conditional on four de facto measures country characteristics: 1) Bank sector assets; 2) Concentration; 3) Financial openness and 4) Foreign bank ownership. In Panel A we show results using the two de facto measures of openness, while Panel B reports the results for the de facto measures of bank 14 In unreported results, we examine the effectiveness of financial institution-targeted macroprudential policy instruments. We do not find evidence that the adoption of financial institution-targeted instruments affects systemic risk. 20

21 sector size and concentration. Finally, Panel C shows results using the two de jure measures of regulatory quality: Restrictions on bank activities and Official supervisory power. Results in Panel A show evidence that the impact of macroprudential policies on systemic risk is stronger in closed economies. Taking the coefficients in Model (6), banks in our treatment sample experience a 0.8% reduction in MES (or 34.8% relative to its mean of 2.3%) following the adoption of macroprudential policies in countries with low levels of foreign bank ownership. In contrast, there is no significant impact in countries with higher levels of foreign ownership of banks. The difference in the coefficient in the interaction term (Post x Treat) between High and Low foreign bank ownership is significant at the 5% level (p-value of χ 2 test is 0.027). We find similar results using our alternate proxy for systemic risk, R-squared, and using the alternate measure of financial openness. In Panel B of Table 6 we show results using the de facto measures of Bank sector assets and Concentration. We do not observe significant differences on the impact of macroprudential policies on systemic risk by size of the banking sector. However, the results show that the impact of the adoption of macroprudential policy tools on systemic risk is significant in countries with less concentrated banking sectors. Taking the coefficient in Model (6), following the adoption of macroprudential policies banks in our treatment sample experience a 0.6% reduction in MES (or 27% relative to its mean of 2.2%) in countries with low concentration. In contrast, there is no significant impact in countries with higher levels of concentration. We obtain similar results using the alternate measure of systemic risk, R-squared. The average effect on systemic risk using R- squared was insignificant in Table 4. Thus, the impact on systemic risk comes from countries with less concentrated banking sectors. 21

22 Finally, in Panel C of Table 6 we analyze the effect of de jure measures of regulatory quality. The results show strong evidence that the adoption of macroprudential policies lowers systemic risk in countries with more restrictions on bank activities. From Model (1) of Panel C of Table 6, we observe that following the adoption of macroprudential policies banks in our treatment sample experience a 0.7% reduction in MES (29.2% relative to its mean of 2.4%) in countries with High restrictions on bank activities. The impact is significantly different from that in countries with Low restrictions on bank activities (p-value of χ 2 test is 0.000). The results in Models (3) and (4) of Panel C of Table 6 reveal significant differences between countries with High and Low restrictions when using the alternate measure of systemic risk, R-squared. Overall, our results in this section suggest that the effectiveness of macroprudential policies in reducing systemic risk is impacted by both de facto and de jure country characteristics Impact of macroprudential policies on lending and profitability The results thus far provide evidence that borrower-targeted macroprudential tools are effective in constraining bank risk taking. In addition, the evidence suggests that such policies are effective in lowering systemic risk, but only in closed economies, countries with less concentrated banking sectors and those with more restrictions on bank activities. In this section, we analyze whether the reduction in bank risk taking has led to a decline in lending, which could adversely affect bank profitability. To test this hypothesis, we estimate various specifications of Equation (1) using proxies for credit growth, Loan growth change in gross loans scaled by lagged total assets and bank profitability (ROA net income divided by average assets, and ROE net income divided by average equity). We include country and bank level controls that have been show to affect bank profitability. Country-level controls include: 1) GDP growth; 2) Volatility; 3) Market return; 4) 22

23 Noninterest income; 5) Bank sector assets, and 6) Concentration. Bank-level controls include controls for size (Size), reliance on nontraditional bank activities (Bank noninterest income), reliance on deposit funding (Deposits-to-assets), Leverage, and cost efficiency (Noninterest expense). We show results from the estimation of Equation (1) using our bank performance proxies in Table 7. In Models (1) and (2) we show results using Loan growth, in Models (3) and (4) we show results using ROA, and Models (5)-(6) show results using ROE. We do not observe that the adoption of borrower-targeted macroprudential instruments adversely affects lending or bank profitability. The coefficients on the interaction term, Post x Treat are not significant in any of the regression specifications. The evidence suggests that the adoption of these policies, while constraining bank risk taking do not adversely affect bank profitability or credit growth. The effectiveness of macroprudential policies in constraining bank risk taking and systemic risk differs based on de facto and de jure country characteristics. In Panels A and B of Table 8 we explore whether de facto and de jure country characteristics affect the impact of the adoption of macroprudential policies on lending growth and bank profitability. To examine this, we expand Equation (1) using interactions with indicator variables based on the four de facto (Financial openness; Foreign bank ownership; Bank sector assets, and Concentration) and two de jure (Restrictions on bank activities, and Official supervisory power) measures. Specifically, we create an indicator, High, for each de facto (de jure) country characteristic that is equal to one if the measure is above the cross-country median. Our variable of interest is the triple interaction term (Post x Treat x High). A significant coefficient on the triple interaction term would indicate differences in lending growth (bank profitability) post adoption for banks in our treatment group from countries with high de facto (de jure) measures. 23

24 We show results using the de facto (de jure) country characteristics in Panel A (B) of Table 8. The results in Panel A of Table 8 show some evidence that the impact of macroprudential policies leads to lower profitability in countries with higher levels of foreign bank ownership. From Model (5), the triple interaction term [Post x Treat x High] is negative and statistically significant, which indicates that post adoption, treated banks in countries with High foreign bank ownership experience a decline in ROA relative to control banks. The impact is statistically significant. Treated banks in High foreign bank ownership countries experience a 0.08 decline in ROA post adoption, which represents a 9.7% decline relative to its mean (0.83 for this sample). In contrast, for treated banks in countries with Low foreign bank ownership, where the adoption of macroprudential policies are more effective in constraining bank risk taking, do not experience a significant change in ROA post adoption. 15 The impact on ROE is similar to that reported for ROA (Model (6)). The evidence suggests that banks in countries in which the adoption of macroprudential policies is ineffective in constraining bank risk taking may pursue strategies that are unprofitable (perhaps in an attempt to circumvent the implementation of tougher lending restrictions). The results in Panel A of Table 8 also show that treated banks in countries with Large banking sectors experience an increase in profitability post adoption of macroprudential policies. From Model (8), post adoption, treated banks in countries with Large banking sectors experience an increase in ROA of 0.402, a 49% increase relative to its mean (0.82 for this sample). In contrast, treated banks in countries with Small banking sectors experience a decrease in ROA post adoption 15 From Model (5) of Panel A of Table 8, post adoption, the change in ROA for treated banks is 0.29 [sum of coefficients on Post + Post x Treat] = The change is insignificant (p-value of F-test 0.691). In contrast, treated banks experience a 0.08 decrease in ROA post adoption in countries with High foreign bank ownership [sum of coefficients on Post x Treat x High + Post x Treat + Post x High + Post] = ] =-0.08 [p-value of F-test 0.034]. 24