Opaqueness and Bank Risk Taking (

This paper investigates the relationship between opaquene ss and bank risk taking. Using a sample of 199 banks from 38 countries over the period J anuary 1996 to December 2006, I analyze whether more opaque banks are riski r than less opaque banks. I find suggestive evidence that commonly used proxies for bank opaqueness are significantly related to bank risk taking as measure d by the Merton PD and the bank-individual Z-score, even after accounting for potential simultaneity between risk taking and opaqueness. More opaque banks seem t o engage more in risk taking than less opaque banks. This result provides sup port to the common view that bank opaqueness is problematic and that transpare ncy among financial institutions should be increased.


Introduction
Are more opaque banks riskier than less opaque banks?Understanding the role of opaqueness for bank risk taking is interesting and important for practitioners and researchers alike. 1 For instance, international bank regulators are currently considering the incorporation of bank opaqueness into capital regulation, a subject that is at the center of lively debate among international regulators and researchers (Franke & Krahnen, 2008, e.g.).Along the same lines, various high-profile policymakers view the inherent opaqueness of the banking sector to be problematic and call for a reduction of this opaqueness.For instance, in the U.S. and in Europe, one way to reduce the opaqueness and the associated uncertainty among market participants is to conduct stress tests. 2 The prevailing view in the public debate, thus, establishes a direct link between opaqueness and bank risk taking.However, up to now, it is not at all clear whether banks that are more opaque are in fact also riskier. 3 In this paper I specifically analyze the relationship between opaqueness and bank risk taking.While several works have investigated the drivers of bank risk taking, to the best of my knowledge, there is no study, yet, that explicitly analyzes whether and how opaqueness is related to bank risk taking.This work tries to fill this gap in the literature and aims at providing new insights into the role opaqueness plays for the behavior of banks.
In fact, it was shown by some authors that a higher degree of opaqueness makes it more difficult to assess the risk of banks.Morgan (2002) and Flannery et al. (2004) show empirically that banks differ in their degree of opaqueness.The main reasons underlying these differences are that the balance sheet structures of banks differ and that banks engage to different extents in rather opaque off-balance sheet activities.Such activities are difficult to assess for bank outsiders.
Specifically, Morgan (2002) shows that rating agencies disagree more often over bond issues by U.S. banks than over bond issues of firms from other industries, and that this stems from the inherent opaqueness of the banking industry.He finds that bank opaqueness is higher the higher the cash holdings and deposits, the more loans and leases, the more trading assets, and the more other assets a bank has.The share of such seemingly opaque assets on banks' total assets clearly differs between different banks both within a country as well as between countries.Additionally, Flannery et al. (2004) argue that banks having a more opaque asset structure are more difficult to assess by equity analysts.Iannotta (2006) provides evidence on the European banking sector in line with Morgan (2002).Haggard & Howe (2013) show that banks have less firm-specific information in their equity returns and interpret this as consistent with the view that banks are more opaque than firms from other industries.Finally, Bannier et al. (2010) show that rating agencies seem to be particularly conservative when assigning banks unsolicited ratings, and that this is to some extent driven by opaqueness.
While these studies are mainly concerned with the analysis of the opaqueness of the banking industry vis-à-vis other industries, in this paper I analyze the relationship between opaqueness and banks' risk taking behavior.Hence, I contribute to these studies by analyzing possible consequences of opaqueness on bank behavior.
Up to this point it is, however, not clear whether and how opaqueness influences a bank's risk taking behavior.According to the ongoing debate, the general consensus seems to be that more opaque banks are riskier, that increased transparency in the banking sector is likely to reduce this risk, and, in times of crises, may help to re-establish confidence in the system.One reason for the assumed positive relationship between opaqueness and risk taking is, for instance, that more opaque banks may be better able to hide their risky activities than less opaque banks.In other words, the more opaque a bank, the more difficult it would be to detect (excessive) risk taking by, for instance, regulators, shareholders or bondholders.
By analyzing behavioral implications of bank opaqueness on risk taking, this study is also related and contributes to the vast literature on bank risk taking.Numerous theoretical and empirical papers have analyzed what influences bank risk taking.For instance, several authors have investigated how competition interacts with bank risk taking (Marcus, 1984, Keeley, 1990, Hellmann et al., 2000, Boyd & De Nicoló, 2005, Jiménez et al., 2007, Martinez-Miera & Repullo, 2008, e.g.).The results in this literature are mixed.While some authors argue that more competition induces banks to engage in more risk taking by eroding their franchise values (Marcus, 1984, Keeley, 1990), others come to the opposite result (Repullo, 2004, Boyd & De Nicoló, 2005).Other studies have analyzed how capital requirements impact bank risk taking (Furlong & Keeley, 1989, Rochet, 1992, Repullo, 2004, Barth et al., 2004, Behr et al., 2010, e.g.).Again, the results are not conclusive.For instance, Furlong & Keeley (1989) and Rochet (1992) argue that higher capital requirements reduce risk taking incentives.On the other hand, Barth et al. (2004) find no robust relationship between bank risk and capital requirement stringency and Behr et al. (2010) argue that the effectiveness of capital requirements depends crucially on the prevailing market structures that banks operate in.González (2005) shows that regulation that restricts banks' activities gives them incentives to take on more risk.Finally, Laeven & Levine (2009) and Fahlenbrach & Stulz (2011) investigate the relationship between banks' corporate governance and their risk taking behavior.These studies show that risk taking incentives depend crucially on the ownership structure and the way bank managers are compensated.
However, none of the existing studies has analyzed whether and how opaqueness is related to bank risk taking.Therefore, this study also makes a contribution to this literature by exploring whether banks' risk taking behavior is influenced by differing degrees of opaqueness.Detecting whether and how opaqueness is related to bank risk taking is a novel result in the literature on bank risk taking.
I use an international sample of 199 banks from 38 countries in the time period January 1996 to December 2006 to investigate whether the risk taking behavior of banks is related to their degree of opaqueness.I employ two risk measures that are widely used in the literature, the mainly market-based probability of default (henceforth Merton PD) and for robustness tests the purely accounting-based bank-individual z-score, and explore whether these risk measures vary between banks with different degrees of opaqueness.
The opaqueness proxies I employ are rating-induced opaqueness proxies, i.e., the occurrence of a split rating and the absolute notch differences between ratings assigned by two different rating agencies (Morgan, 2002, Drucker & Puri, 2009, Bannier et al., 2010, e.g.), and opinion-related opaqueness proxies, i.e., equity analyst coverage and the forecast dispersion of earnings per share (EPS) estimates (Flannery et al., 2004, Bongaerts et al., 2013, e.g.).In all regressions I further control for a set of variables that might influence a bank's risk taking behavior.
My results provide suggestive evidence that opaqueness is related to bank risk taking.Specifically, I find that the absolute difference measured in notches between the ratings by Standard and Poor's (S&P) and Fitch has a significant relationship with bank risk taking.According to this opaqueness proxy, more opaque banks engage more in risk taking.This relationship remains when I control for potential simultaneity between opaqueness and risk taking.Furthermore, I find that banks with lower analyst coverage, i.e. a higher degree of opaqueness, have lower bank z-scores, that is, engage more in risk taking.
The remainder of the paper is organized as follows.In section two I describe the variables used for the regression analyses and provide some descriptive statistics.Section three outlines the methodology, and section four contains the main empirical results as well as a series of robustness tests.I conclude in section five.

Variables and sample construction
I use a sample of international banks in the time period January 1996 to December 2006 to analyze the relationship between opaqueness and bank risk taking. 4The dataset is constructed by merging data from several databases.First, I explore S&P's credit ratings database accessed through Wharton Research Data Services (WRDS) and identify all banks in the database with rating information in the sample period because I need the rating information for the construction of my main opaqueness proxies.I start in 1996 because S&P started only in January 1996 to indicate whether a rating is solicited or unsolicited.Several works have shown that unsolicited ratings are hardly comparable to solicited ratings (Poon, 2003, Poon et al., 2009, Bannier et al., 2010, e.g.).Thus, I exclude all observations with unsolicited ratings from the data set.This first sample selection step yields 1,046 banks from 84 countries with rating data.
Next, I compute proxies for risk taking, the dependent variables in the regressions.For the baseline regressions I employ a mainly market-based estimate of risk taking and for tests of robustness a purely accounting-based estimate of risk taking.The computation of the market-based estimate relies on the seminal works by Black & Scholes (1973) and Merton (1974).According to Merton (1974), a firm's equity can be viewed as a call option on the value of the firm's assets.The call option expires at the debt's maturity, T .If, at T , the value of assets is below the face value of liabilities (i.e., the strike price), the call option is left unexercised and the bankrupt firm is turned over to its debt holders.The probability of default in this model is hence equal to the probability that the market value of assets is less than the face value of liabilities at maturity.I henceforth refer to this probability as the Merton PD (variable name Merton PD) and use it as the risk taking proxy in the baseline regressions.This variable was, for instance, also used in Bannier et al. (2010) as a measure for bank risk taking.
The input parameters needed for computing the Merton PD are the face value of liabilities, which I proxy by computing the difference between total assets and total shareholders' equity in U.S. dollars extracted from Compustat (Items 129 and 215), the market value of equity in U.S. dollars extracted from Datastream,5 the volatility of the market value of equity over the whole fiscal year, the maturity of debt, which is assumed to be one year, and the risk-free rate of interest, which I proxy by using a one-year market yield for U.S. Treasury Notes obtained from the website of the Federal Reserve Board.The Merton PD is then calculated in three steps.First, I simultaneously estimate the asset value and asset volatility.Then, I use the estimated asset value to determine the expected return.Finally, using asset value, asset volatility, expected return, time to maturity and total book value of liabilities, I calculate the Merton PD.The data for the computation of the Merton PD are available for 199 banks from 38 countries, yielding 1,135 firm-year observations in the period January 1996 to December 2006.This is the final sample used for the analyses.
Second, to test for robustness I employ a purely accounting-based proxy for risk taking, the so-called bank z-score (Bank z-score).The z-score mea-sures an individual bank's default risk and combines accounting measures of profitability, leverage and volatility.Specifically, z indicates the number of standard deviations that a bank's return on assets has to drop below its expected value before equity is depleted and the bank is insolvent (Roy, 1952, Hannan & Hanweck, 1988, Laeven & Levine, 2009, e.g.).It is defined as: where ROA i is the return on average assets and CAR i the capital-asset ratio of bank i.The data for computing the z-score are extracted from Compustat.In line with Laeven & Levine (2009), I calculate the standard deviation of asset returns σ(ROA i ) of each bank over at least four years.
For instance, I compute the z-scores as of year-end 1997 by making use of the volatility of return on assets over the four years 1994-1997, and so on.6I use several proxies for bank opaqueness that were established in the literature.First, I measure bank opaqueness by the occurrence of a split rating between S&P and Fitch.Split ratings as a proxy for opaqueness were first introduced by Morgan (2002) who showed that bonds issued by U.S. banks (and insurance companies) exhibit significantly more split ratings than bonds issued by firms from other business sectors.He concluded that the higher likelihood of having a split rating is due to the higher degree of opaqueness associated with banks, which makes them more difficult to rate.Several other works have used split ratings as a proxy for opaqueness since then (Drucker & Puri, 2009, Livingston et al., 2007, Bannier et al., 2010, Bongaerts et al., 2013, e.g.).I use Fitch data because Fitch is the most important rating agency in the financial services industry outside the U.S. and because data availability is better than in the case of Moody's.Fitch rating data were extracted from Bankscope. 7econd, also following Morgan (2002) and Bannier et al. (2010), I use the absolute notch difference between the ratings of S&P and Fitch as a measure of the degree of opaqueness, that is, the higher the notch difference the higher the degree of opaqueness.The use of notch differences allows for the incorporation of more information about the degree of opaqueness than revealed by the occurrence of a split rating alone.
I translate both S&P's and Fitch's rating scales into numerical rating scales by using a mapping from AAA = 1, AA+ = 2, . . ., CC to C = 18. 8Then I create the dummy variable Split-rating, which equals one if the S&P and Fitch ratings for the same bank at the end of the year differ, and zero otherwise.The absolute notch differences (Notch difference) are calculated by subtracting the numerical Fitch rating from the corresponding S&P rating and using the absolute value of the difference.In the very few cases in which the absolute notch difference was larger than five notches, I set these to five to eliminate extreme outliers.This does not affect results, though.
Third, I use the number of equity analysts covering a specific bank (Analyst coverage) as a proxy for bank opaqueness.The fewer analysts covering a bank, the higher should be the opaqueness.Fourth, I use the forecast dispersion of EPS estimates (Forecast dispersion) as a proxy for bank opaqueness.Both variables were, for instance, used by Flannery et al. (2004), Livingston et al. (2007), and Bongaerts et al. (2013) as opaqueness proxies.Analyst data were extracted from the Institutional Brokers Estimate System (IBES) accessed through WRDS.I standardize the forecast dispersion by dividing it by the mean EPS forecast to account for currency differences and to eliminate the influence of differing share prices.
I also use several bank and country-specific control variables.First, I control for bank size (Size) measured as the natural log of total assets in million U.S. dollars extracted from Compustat (Item 129).Larger banks may have comparative advantages in raising capital and diversifying, which can influence a bank's risk taking behavior (Laeven & Levine, 2009, Behr et al., 2010).Additionally, they may also have easier access to derivatives markets, which can influence their risk taking.Ex ante, it is, however, not clear whether larger banks engage more in risk taking or less.Second, I control for profitability differences by using the return on equity (Return on equity) in percent.This variable is computed as net income divided by total shareholders' equity.Data are extracted from Compustat (Item 440 and item 215).I exclude extreme values by winsorizing on the 5% and 95% level.9A higher profitability may imply higher franchise values, which can limit a bank's risk taking behavior (Keeley, 1990, Jiménez et al., 2007, e.g.).
Furthermore, I control for country-specific differences which might influence bank risk taking by including S&P's country rating (Country rating) taken from the S&P website, and the annual, country-specific GDP growth in percent (GDP growth) extracted from the World Development Indicator Database accessible on the World Bank website.I also control for market structure and competition by using the fraction of assets held by the three largest banks in each country (Bank concentration). 10There is a vast literature analyzing the role of market structure and competition on bank risk taking.While the traditional view was that higher competition erodes banks' franchise values, which, in turn, leads to higher risk taking (Marcus, 1984, Keeley, 1990, Hellmann et al., 2000, e.g.), this view was recently challenged by Boyd & De Nicoló (2005) who argue that increased competition reduces banks' risk taking.Hence, while the direction of the influence of market structure and competition on risk taking is not clear, there is general consensus that both matter for bank risk taking.The bank concentration ratios I use as controls in the regressions are given in percent, they are taken from Beck et al.'s database. 11 Finally, I control for capital regulation requirements by including the regulatory capital ratio in the regressions.This variable was, for instance, also used in Laeven & Levine (2009).Regulatory capital ratios are taken from an international World Bank survey covering different aspects of bank regulation that was conducted in 1999, 2002, and 2006.The survey questions as well as the respective answers can be found in Barth et al. (2004) and on the related World Bank website.To match the capital ratios with the other data I use the following mapping procedure.

Descriptive analysis
Table 1 shows the mean, standard deviation, 1 and 99 percentile values for all variables described above.First, it can be seen that the risk proxies vary considerably between banks.

Table 1 Descriptive statistics of main variables
The table shows the mean, standard deviation (StD), 1, 25, 50, 75 and 99 percentile values for the main variables used in the multivariate regression analyses.The dependent variables, i.e. risk proxies, used in the regressions are the Merton PD and the bank z-score.The opaqueness proxies are: Split-rating is a dummy variable that takes the value one if the ratings by S&P and Fitch for the same bank differ, and zero otherwise; Notch difference is the absolute difference measured in notches between the S&P and the Fitch rating, Analyst coverage is the number of equity analysts covering the bank, and Forecast dispersion is the standardized standard deviation of the EPS estimates.The control variables are: Size is measured in total assets, Return on equity is computed as net income divided by total shareholder's equity, Country rating is S&P's numerical country rating, GDP growth is the annual growth rate of the country's GDP, Bank concentration is the fraction of assets held by the three largest banks in a country, and Capital requirements is the regulatory capital ratio.The variable construction and the different data sources used are described in detail in Table A A1 in the appendix contains all variable descriptions and the various data sources the variables were extracted from.
13 Another important driver of bank risk taking may be the existence of an explicit deposit insurance scheme.The role of deposit insurance schemes for bank risk taking was first analyzed by Merton (1977), and later, e.g., by Bhattacharya & Thakor (1993) and Demirgüc-Kunt & Kane (2002).I collected deposit insurance data from Demirgüc-Kunt et al. (2008), but the variable shows very little variation and almost all sample countries had a deposit insurance system in place in the sample period.For this reason, it cannot be used in my regression analyses.
Second, roughly 66% of all observations with ratings by S&P and Fitch exhibit split ratings.This figure is in line with Morgan (2002) who reports a split rating ratio for bonds issued by U.S. banks and rated by S&P and Moody's of 62.9% and with Bannier et al. (2010) who report a split rating ratio for international banks rated by S&P and Fitch of 65.3%.The mean notch difference is 0.85 notches, which is again closely in line with an average notch difference of 0.83 reported by Morgan (2002).On average, the banks in the sample are followed by 15 analysts with a rather large variation reaching from 1 to 42 analysts.The average size of the balance sheet of the sample banks is 152 billion U.S. dollars.Again, this variable shows a considerable variation because the sample includes some very small banks (minimum total assets of 263 million U.S. dollars) and some very large ones (maximum of roughly 2,000 billion U.S. dollars).The average country rating lies between AA and AA-.Bank concentration is on average 58% and varies considerably from a minimum of 22.08 to 100%.Finally, the average regulatory capital ratio was 8.15%, with a low of 8% in most of the sample countries and a high of 12% in Singapore.The sample contains 1,135 observations covering 199 banks from 38 countries.
Table 2 contains the correlation analysis between the two dependent variables Merton PD and Bank z-score and the independent variables.The opaqueness proxies are in most of the cases significantly (5% level or higher) correlated to the risk proxies.First, the occurrence of a split rating between S&P and Fitch is significantly positively correlated to the Merton PD and significantly negatively correlated to the bank z-score.As a split rating indicates a higher degree of opaqueness, this result suggests that more opaque banks engage more in risk taking.The same can be observed for the absolute notch differences between the S&P and Fitch ratings.Analyst coverage is significantly negatively correlated to the Merton PD, suggesting that banks followed by more analysts, i.e. with a lower degree of opaqueness, engage less in risk taking.This result does not hold, however, for the bank z-score.The forecast dispersion of EPS estimates is not significantly correlated to either of the risk proxies.

Table 2 Correlation matrix
This table shows correlations between the main variables used in the regression analyses.Z-Score is the bank-individual z-score, ∆ Notch denotes the rating differences between S&P and Fitch in notches, Coverage is the analyst coverage, Dispersion is the standardized standard deviation of EPS estimates, RoE is return on equity, and Concentration is the fraction of assets held by the three largest banks in a country.All variable explanations are as in Table 1.The table also shows significant correlations between the other control variables and the risk proxies, suggesting the importance of these controls for the regression analyses.For instance, more profitable banks have lower Merton PDs and higher bank z-scores, suggesting that they engage less in risk taking.This observation is in line with the franchise value paradigm advocated by some researchers (Keeley, 1990, Behr et al., 2010, e.g.).Banks from countries with worse (higher numerical) ratings engage more in risk taking given the positive (negative) correlation with the Merton PD (bank z-score).The economic development of a country also seems to play a role for the risk taking behavior of banks as the Merton PD is smaller when GDP growth is higher.Furthermore, banks from countries with a higher bank concentration have lower Merton PDs.On the other hand, the regulatory capital ratio is not significantly correlated to the risk taking proxies.

Variable name
Finally, Table 2 displays several significant correlations among the control variables.For instance, size is positively correlated to analyst coverage indicating that larger banks are followed by more analysts.Larger banks also seem to be more profitable, are headquartered in countries with better country ratings, and come from countries with a higher asset concentration.Also, banks from countries with a higher asset concentration seem to be more profitable, most likely reflecting the fact that they are able to exert more market power.
All in all, the correlation analysis suggests that opaqueness might indeed be related to bank risk taking.However, the significant correlations between the other control variables and the risk proxies and among the control variables stress the importance of multivariate regression analyses to disentangle the relationship between opaqueness and bank risk taking.

Methodology
Given the cross-sectional as well as time series dimension of the dataset, I analyze the relationship between opaqueness and bank risk taking by estimating an OLS regression including firm fixed effects of the following form: where Risktaking i,t is either the Merton PD or, in the robustness tests section, the bank z-score of bank i at time t; β 1 is the coefficient of the opaqueness proxy; Opaqueness i,t is one of the four opaqueness proxies Split rating, Notch difference, Analyst coverage, and Forecast dispersion corresponding to bank i at time t; and γ is a vector of coefficients of the control variable vector X i,t .The control variables are as described above.I use different control variable combinations to avoid multicollinearity problems indicated by the correlation analysis.The model includes time and firm fixed effects indicated by the vectors τ t and θ i respectively.Results for these are omitted from the tables to save space.Finally, ǫ i,t denotes the error term of the regression.Standard errors are computed using the Huber/White sandwich estimator.

Empirical Results
I first present the results using the rating-induced opaqueness proxies Split-rating and Notch difference, and then the results for the regressions using the opinion-related variables Analyst coverage and Forecast dispersion.In the first set of results I employ the Merton PD as the risk proxy, for subsequent robustness tests I employ the bank z-score.Further, this section provides results of instrumental variable regressions to account for potential simultaneity bias between opaqueness and bank risk taking.

Rating-induced opaqueness
Table 3 contains the results of the estimation of regression (1) using the two rating-induced opaqueness proxies Split-rating and Notch difference.Rating-induced opaqueness proxies were used in several works including Morgan (2002), Livingston et al. (2007), Drucker & Puri (2009), and Bannier et al. (2010).The table shows the results for the Merton PD as dependent variable.In all regressions, the sample size is smaller than the initial sample size of 1,135 observations of 199 banks because of the non-availability of Fitch rating data for some banks.
In models (I) and (II) I use Split-rating as the main explanatory variable.In both models, Split-rating is not significant indicating that banks, which are more opaque measured by the occurrence of a split rating between S&P and Fitch, are not riskier than banks that receive the same rating by both rating agencies.The table further shows that Return on equity is significantly negative in both models and Country rating is significantly negative in model (II).This indicates that more profitable banks are less risky (have smaller Merton PDs), which is in line with the view that more profitable banks have higher franchise values, and that higher franchise values mitigate risk taking (Keeley, 1990, Jiménez et al., 2007, Behr et al., 2010, e.g.).
The results further indicate that banks from countries with worse country ratings have lower Merton PDs.In model (II) I additionally include the regulatory capital ratio, captured by the variable Capital requirements, in the regressions.The coefficient is, however, not significant.14

Table 3 Rating-induced opaqueness and bank risk taking
This table shows regression results with the Merton PD as dependent variable.All independent variables are as described in Table 1.The two main explanatory variables are Split-rating and Notch difference.In model (V) and (VI) I exclude banks from countries with only one bank in the sample.All regressions are fixed effects panel regressions and additionally include year dummies.Standard errors in parentheses are computed using the Huber/White sandwich estimator.

Independent
No In models (III) and (IV) I use Notch difference as the main explanatory variable.The table shows a positive and on the 5% level significant relationship between opaqueness measured by the absolute notch difference between the S&P and Fitch rating and bank risk taking.This indicates that banks over which rating agencies disagree more strongly engage more in risk taking.I further analyze this finding by creating dummy variables indicating one, two, and three notches difference between S&P and Fitch to account for a potential non-linearity of the relationship.The results of this analysis, which are not displayed here, show that the one-notch dummy is not significant, the two-notch dummy is significantly positive on the 10% level, and the three-notch dummy is positive and significant on the 5% level.This confirms that the higher the degree of opaqueness, the more banks engage in risk taking.The results for Return on equity and Country rating are confirmed in both models.
Finally, in models (V) and (VI) I exclude observations of banks from countries with only one bank in the sample in order to avoid that the results are driven by banks from such countries.As the table displays, this does not alter the results for the opaqueness proxies.However, the capital ratio in model (VI) is now weakly significantly negative, suggesting that banks from countries with a higher capital requirement engage in less risk taking.This is in line with the findings of Laeven & Levine (2009).15

Opinion-related opaqueness
Besides rating-induced opaqueness proxies, the use of opinion-related opaqueness proxies seems to be well-established in the literature.For instance, equity analyst data as a measure for opaqueness are used by Flannery et al. (2004), Livingston et al. (2007), Bannier et al. (2010), andBongaerts et al. (2013).Table 4 shows the results of estimations of regression model (1) using Analyst coverage and Forecast dispersion as main explanatory variables.The dependent variable in the regressions is the Merton PD.The sample size is again reduced because of the non-availability of analyst data for some banks in IBES.In models (I) and (II) I use Analyst coverage as opaqueness proxy.The coefficient is not significant in either of the two specifications.The Return on equity variable has the same sign as before, but is not significant anymore.Country rating is significantly negative in model (I), GDP growth displays a negative and highly significant relationship with the Merton PD in both models, indicating that banks from countries with a higher GDP growth engage in less risk taking.Furthermore, in model (II) size is negative and weakly significant and the capital ratio is highly significant and displays a negative relationship with bank risk taking.As above, this suggests that banks from countries with a higher capital ratio engage in less risk taking.
In models (III) and (IV) the opaqueness proxy is Forecast dispersion, the standardized standard deviation of EPS estimates, computed by dividing the absolute forecast dispersion by the mean EPS forecast to account for currency and share price differences.Again, I cannot detect any significant relationship with the Merton PD.The results for the control variables confirm the findings in model (I) and (II).
In models (V) and (VI), all results are again confirmed for the opaqueness proxies when observations from countries with only one bank in the sample are excluded, results for the controls are as before.
These findings suggest that opinion-related opaqueness does not seem to be related to bank risk taking measured by the Merton PD.

Robustness Alternative risk taking proxy
To test the robustness of the findings so far, I first substitute the mainly market-based risk taking proxy with a purely accounting-based risk taking proxy, the bank-individual z-score.Other than that, all control variables and the regression set-up remain as described above.Table 5 displays the results.The table shows that neither the occurrence of a split rating between S&P and Fitch (models (I) and (II)) nor the absolute notch difference between the S&P and the Fitch rating (models (III) and (IV)) are significantly related to the bank z-score, the latter result contradicting the results from the Merton PD regressions.As in Laeven & Levine (2009), the size variable is weakly significant in all four models, indicating that larger banks engage less in risk taking.Return on equity is positive and highly significant in all four models, as before suggesting that more profitable banks engage less in risk taking.The previous findings for Country rating and GDP growth, on the other hand, are not confirmed by the bank z-score regressions.The coefficient of Capital requirements is statistically and economically highly significant, underlining that banks operating in countries with a higher capital ratio engage less in risk taking.
Models (V) and (VI) use Analyst coverage as the opaqueness proxy.Now I find that banks followed by more analysts, i.e. banks with a lower degree of opaqueness, have higher z-scores.This provides additional suggestive evidence that opaqueness is related to bank risk taking.The results for the other control variables are largely as before.
Finally, in models (VII) and (VIII) I regress the bank z-score on Forecast dispersion and the additional control variables.Again, I cannot detect any significance of this opaqueness proxy.Morgan (2002) argues that banks with a worse rating, i.e. riskier banks, are more opaque.However, it is not fully clear whether risk drives opaqueness or whether opaqueness drives risk.Throughout the paper, I have assumed that the causality goes from opaqueness to risk taking, i.e., more opaque banks engage in more risk taking because, for instance, it is easier for them to hide the risks and more difficult to detect, monitor and discipline the risk taking behavior of the bank managers.This seems to be the prevailing view among researchers and practitioners.However, it may also be that risk taking and the level of opaqueness are simultaneously determined by bank managers.For example, they could engage in more risk taking off the balance sheet, thus, increasing both risk taking and opaqueness.

Instrumental variable regressions
In this section I account for potential simultaneity by instrumenting the opaqueness proxies used in my analyses and using a two-step GMM estimator.To qualify as an instrument, I need a variable that is exogenous to bank risk taking, but has an impact on the level of opaqueness.Variables like the rigidity of accounting standards, which might drive the level of opaqueness, may also influence the risk taking by banks.Hence, they do not qualify as instruments.I therefore decided to use the opaqueness proxies employed in the previous analyses lagged by one year as instruments because it is not plausible to argue that last year's level of opaqueness has a direct influence on this year's risk taking, but it is plausible to assume that the level of opaqueness in recent years influences the current year's level of opaqueness.I conduct tests to show that these instruments, albeit not being ideal, fulfill the conditions of relevance and validity of an instrument.
Table 6 contains the second-stage results of these GMM regressions. 16 The table also shows the coefficient and standard error of the instrument from the first stage regression and the F-test of excluded instruments, confirming that the instruments employed are both relevant and valid.
Model (I) and (II) use the Merton PD as dependent variable and Notch difference as main explanatory variable to test the robustness of the result from Table 3, column 3 and 4. The sign of the coefficient for Notch difference is positive as before, however, the coefficient is only weakly significant (10% level).The results for the other control variables confirm the findings from Table 3. Specifically, Return on Equity is positive and highly significant in both models, Country Rating and GDP Growth are both negative and highly significant in both models.Size and Bank Concentration do not display a significant relationship, while Capital Requirements is negative and highly significant. 16First stage results are not displayed here, but are available upon request.

Table 6 Instrumental variable regressions
This table shows second-stage instrumental variable regression results with the Merton PD as dependent variable in model (I) and (II) and the bank z-score as dependent variable in models (III) and (IV).All independent variables are as described in Table 1.The main explanatory variables are Split-rating and Analyst coverage.All regressions are fixed effects panel regressions, and additionally include year dummies.Split-rating and Analyst coverage are instrumented using the respective variable lagged by one year as instrument.Standard errors in parentheses are computed using the Huber/White sandwich estimator.

Independent
Merton In model (III) and (IV) I use the bank z-score as dependent variable and Analyst coverage as main explanatory variable to test the robustness of the result from Table 5, column 5 and 6.The coefficient is again positive, but only weakly significant in model (III) and not significant in model (IV), although in the latter case the p-value is only slightly above the 10% level (10.54%).Of the control variables, only Bank Concentration is negative and significant in both models (III) and (IV), while Capital Requirements is significantly negative in model (IV).These results are in line with the findings from the earlier analyses.
All in all, the results from the instrumental variable regressions are somewhat weaker than before, but lend support to the finding that, even after controlling for potential simultaneity, there seems to be a significant relationship between opaqueness and bank risk taking.Livingston et al. (2007) suggest the use of the market-to-book ratio as an alternative opaqueness proxy.They argue that the market-to-book ratio measures a firm's growth opportunities and that firms with larger growth opportunities tend to be younger firms from newer industries, making them more opaque and harder to value.In unreported regressions, I used this variable as the main explanatory variable and regressed the Merton PD and the bank z-score on it and the additional controls.The market-to-book ratio was not significant in any of the specifications.

Other tests
Furthermore, as in Laeven & Levine (2009), I also included a liquidity variable in all regressions reported in Tables 3 to 6.I computed this ratio by dividing total liquid assets by liquid liabilities.Liquid assets were calculated by adding up cash and dues from banks (Compustat item 23), short-term investments (item 28), trading/dealing account securities (item 54), and accounts receivable/debtors (item 81), as a proxy for liquid liabilities I used total short-term borrowings (item 154).The inclusion of the liquidity ratio did not change any of the results reported in Tables 3 to 6. 17Finally, in all regressions from Table 3 to Table 6 I  Adding this control variable to the regressions does not alter the results reported in Tables 3 to 6.As the variable itself is not significant in any of the regressions, it is not reported along with the other control variables.

Conclusion
Recently, the inherent opaqueness associated with banks across the globe has spurred a lively debate among international regulators and researchers concerning the incorporation of bank opaqueness into capital regulation frameworks.While it seems to be general consensus that bank opaqueness is problematic, to the best of my knowledge, so far there is no empirical evidence that shows that this has any impact on the risk taking behavior of banks.In this paper I analyze the relationship between opaqueness and bank risk taking and provide for the first time empirical evidence regarding this relationship.I use two common proxies for bank risk taking, the Merton PD and the bank-individual z-score, and several opaqueness proxies on the individual bank level.The opaqueness proxies I employ are well-established in the literature and have been used in several other studies.
The empirical results in this paper provide suggestive, but not conclusive, evidence that opaqueness is positively related to bank risk taking, i.e., that more opaque banks engage more in risk taking.This is a novel result.My research contributes both to the literature on bank opaqueness and the vast literature on bank risk taking.Given the only tentative nature of my results, however, more research on the impact of bank opaqueness on banks' behavior seems warranted and bank regulators should carefully evaluate the incorporation of bank opaqueness into regulatory frameworks as a response to the recent financial crisis.
It should be noted, though, that the findings in this study may be subject to two caveats.First, if the variables used as opaqueness proxies in this study do not properly reflect the degree of opaqueness associated with banks, then my conclusions about the relationship between opaqueness and bank risk taking are not valid.However, given the lack of alternative opaqueness measures and the widespread use of the measures I employ in my analyses among researchers, the approach I chose to explore the relationship between opaqueness and bank risk taking seems to be justified.Future research should, nevertheless, focus also on the development of new, potentially better opaqueness measures to help deepen the understanding of the role opaqueness plays for the behavior of banks.Furthermore, future studies may also benefit from the use of alternative proxies for risk taking, such as value at risk measures or regulatory measures proposed by the Basel Committee on Banking Supervision.
Second, the potential simultaneity of risk taking and opaqueness may be problematic.In this paper, I have assumed that the causality goes from opaqueness to risk taking.It is, however, possible that banks choose both risk taking and opaqueness simultaneously.I have tried to account for this by running instrumental variable regressions using first lags of the opaqueness proxies as instruments.I acknowledge that these instruments, albeit fulfilling standard tests of relevance and validity, are not ideal instruments and future work in this area should also focus on the use of different, more sophisticated instruments to be better able to disentangle the causal effects opaqueness has on the risk taking behavior of banks.
This table shows the number of banks per country included in the analyses as well as the average Merton PD and the average bank z-score for all banks in the respective country.
For the years 1996 till 2000, I use the capital ratios included in the 2001 survey, for the years 2001 till 2003, I use the capital ratios of the 2002 survey, and for the years 2004 till 2006, I use the capital ratios of the 2005 survey.12All regressions further include year dummies to account for time effects not captured in the other control variables.13 include countryspecific opaqueness indices as additional control variables.These indices, which were first used by Bannier et al. (2010), are computed using the same World Bank survey data described above.For the computation, I select eight survey questions associated with bank opaqueness and specify the answers as zero/one values.I then run a principal component analysis to compute the country-specific opaqueness indices.Since survey results are not available for every year in the observation period, I use the following mapping: For the time period 1996-2000 I use the information from 1999, for the period 2001-2003 I use information from 2002, and for the period 2004-2005 I use the information from 2006.
.1 in the Appendix.

Table 4
Opinion-related opaqueness and bank risk takingThis table shows regression results with the Merton PD as dependent variable.All independent variables are as described in Table1.The two main explanatory variables are Analyst coverage and Forecast dispersion.In model (V) and (VI) I exclude banks from countries with only one bank in the sample.All regressions are fixed effects panel regressions, and additionally include year dummies.Standard errors in parentheses are computed using the Huber/White sandwich estimator.

Table 5
Bank z-score regressionsThis table shows regression results with the bank z-score as dependent variable.All independent variables are as described in Table1.The main explanatory variables are Split-rating, Notch difference, Analyst coverage, and Forecast dispersion.All regressions are fixed effects panel regressions, and additionally include year dummies.Standard errors in parentheses are computed using the Huber/White sandwich estimator.

Table A .2
Country distribution of banks and risk measures Rev. Bras.Financ ¸as (Online), Rio de Janeiro,Vol.10, No. 4, December 2012