JSTOR

A. Data Sources

Data for the industrial, asset, and geographic decompositions of the portfolios of Italian banks in our study are taken from the regulatory reports submitted by these banks to the Bank of Italy, the Italian Bankers’ Association, and the Interbank Deposit Protection Fund of Italy. The latter is the Italian equivalent of the U.S. Federal Deposit Insurance Corporation. Our sample starts with a database of 105 primarily commercial banks that reported their asset portfolio and other data during the entire 1993–99 period. The sample period starts in 1993 since the banking law of August 27, 1993 (Consolidating Act), marked a regime shift in the Italian banking structure. It revolutionized the Italian banking system by encouraging the emergence of full‐service financial institutions in that it eliminated the distinction between specialized lending institutions (medium‐ and long‐term credit) and retail banks (short‐term credit).

A complete list of all banks and those that are publicly traded during our sample period is shown in Appendix table A1, along with the average size of each bank over the sample period. These 105 banks constitute over 80 percent of the total banking assets of Italy. These data are aggregated at the level of the bank holding company, wherever applicable. A few of the banks in our sample undertook acquisitions of other banks. The data set, however, does not provide any details as to which these acquiring banks were and which banks they acquired. Furthermore, the data set does not include foreign bank operations in Italy. Over our sample period, the foreign bank penetration of the Italian banking market was weak largely because of the prohibition on foreign banks from accepting deposits of Italian residents.

In terms of size, eight of these 105 banks are “very large” (as defined by the Bank of Italy), seven are “large,” 15 are “medium,” and the remaining 75 are “small.” In terms of geographical scope of banking activities, nine of these banks are “national,” 18 are “regional,” 13 are “intraregional,” 10 are “local,” and the remaining 55 are “provincial.” Finally, 34 of these banks are publicly traded, 62 of them were state‐owned at the beginning of 1993,4 and 70 of them were not members of a consortium or a bank holding group. Whenever our analysis employs measures of performance based on stock market data, we are constrained to focus on the publicly traded sample only. In Section III.D, we also examine separately the robustness of our analysis to state ownership and membership in a consortium or a bank holding company.

While there are natural differences between the banking sectors of any two countries, there are several dimensions along which the Italian banking system is similar to that in the United States: (1) In contrast to other banking systems in continental Europe, Italy has a large number of banks (about 850 at the beginning of our sample), giving rise to a less concentrated banking system like that of the United States. (2) The branching restrictions on banks in Italy were removed in 1990 as they were in the United States in the mid‐1990s. (3) There has been a wave of consolidation in the banking system in 1990s mirroring that in the United States. (4) The banking system comprises a few very large banks and a large number of medium‐ to small‐sized banks as in the United States. However, Italy differs from the United States in that some of its banks are state‐owned, although state ownership has been steadily declining over the past decade following the Amato‐Carli law.5

These stylized facts and the use of Italian banking data to address other important economic issues such as the benefit of relationship banking (Degatriache, Garella, and Guiso 2000) and the effect of bank mergers on loan contracts (Sapienza 2002) lead us to believe that our results would generalize to banking sectors of other countries, including the United States.6

For each bank in our sample, data are available to calculate the following portfolio decompositions:

The financial statement variables and capital structure variables are obtained from the Bank of Italy and Bankscope data bases. Stock market data items for the 34 banks that are publicly traded were taken from the Datastream and Milan Stock Exchange information bases on Italian banks. A few banks had to be discarded from the sample because of missing values of relevant variables, for example, doubtful and nonperforming loans.

B. Construction of Herfindahl Indices

We measure focus (diversification) by employing a Hirschman Herfindahl index (HHI) measure. HHI is the sum of the squares of exposures as a fraction of total exposure under a given classification. In our case, we construct two different kinds of HHIs, which consist of industrial and household sector HHIs, more simply referred to as industrial sector HHIs (I‐HHI) and broad asset sector HHIs (A‐HHI).

Since we have data only for the top five industry exposures for each bank, our measure of I‐HHI for a bank is also based on these five top industries in which loans were made by that bank. As stressed before, we would like to employ, if possible, the exposure to all industries while calculating I‐HHI for a bank. Unfortunately, our data provide only the top five exposures, ranked by their amounts. For most banks in our sample, the top five exposures cover over 70%–80% of the total size of the loan portfolio. The sixth exposure in our computation of I‐HHI considers the remaining portion of the industrial loan portfolio. For this sixth exposure, we employed two conventions: (1) where the sixth exposure is treated as a separate “hypothetical” industry and (2) where the sixth exposure is treated as being equally divided among the remaining 18 industries. Our results turned out to be insensitive to this choice, as is to be expected given that the top five exposures constitute, on average, a large proportion (over 70%) of the total exposure of a bank. Hence, we report results with I‐HHI computed using the sixth exposure as a hypothetical industry. Thus, if the proportional exposures to six industries are X₁, X₂, X₃, X₄, X₅, and X₆, respectively, then I‐HHI equals , where . Note that the HHI has a maximum of one when all loans are made to a single industry.

The A‐HHI is the sum of the squared exposures (measured as a fraction) in the form of sovereign loans, other governmental loans, nonfinancial sector loans, financial sector loans, household sector loans, and other loans.

C. Balance Sheet and Stock Market Variables

We employ the following (annual) variables obtained from the balance sheet and stock market data for the banks in our sample over the period 1993–99:

Return measures:

In addition, we also seek to establish the robustness of our results with the following measures of unexpected losses:

Table 1 presents the univariate statistics (mean, median, standard deviation, minimum, and maximum) for these variables and for Herfindahl indices for all the banks over the sample period 1993–99. Note that the mean (median) bank’s size is about US$12 billion ($3 billion) or 20 trillion (5 trillion) Italian lira, the mean (median) capital ratio is 8.732% (8.113%), and the mean (median) ratio of doubtful and nonperforming loans to assets is 5.234 (3.199).9 The average industrial and asset sectoral focus measures (I‐HHI and A‐HHI) are low, suggesting a significant degree of diversification in these areas.

Table 1 Univariate Descriptive Statistics: 105 Italian Banks, 1993–99

Open New Window

Table 2 completes the descriptive statistics by presenting the correlation matrix among these variables. As table 2 illustrates, the measures of focus, I‐HHI and A‐HHI, are not highly correlated, the correlation being 0.26. This suggests the possibility that the effects of these different diversification measures on bank risk‐return performance may be different. Further, there is significant variation in all the variables we employ, and the correlations suggest a relationship between return measures (ROA, ROE, and SR) and the balance sheet control variables (SIZE, EQRATIO, BRRATIO, EMPRATIO).

Table 2 Bivariate Descriptive Statistics: Correlation Coefficients for 105 Italian Banks, 1993–99

Open New Window

Panel A of table 3 presents the year‐by‐year quintiles of the focus measures.10 What is important for our tests is that the focus measures not only exhibit variability in the overall sample and through time, but also do so for individual banks through time. For our data, we find that the time‐series standard deviation of I‐HHI (A‐HHI), averaged across all banks, is about 0.016 (0.051), which is about half as large as the overall sample standard deviation of I‐HHI (A‐HHI), which is 0.038 (0.099). This implies that there is time‐series variability in the focus measures at the level of an individual bank that is comparable to the variability in the focus measures in the cross section. We explore this issue in some more detail later.

Table 3 Quantiles Values: 105 Italian Banks, 1993–99

Open New Window

Finally, panel B of table 3 contains the year‐by‐year quintiles of various risk measures. As is clear from the table, 1993–99 represents a turbulent period for Italian banks: losses measured as doubtful loans to assets ratio (DOUBT) reached values above 10% for about 10% of the sample in each year, and maximum values ranged from 15% to 45%. Overall, the latter half of the sample period appears to have more stable values of DOUBT. Doubtful loans trended upward between 1993 and 1996 as a result of the lingering effects of the 1992–93 crisis, reflecting in part the increased fragility of state‐owned enterprises, rising risk from exporting companies, and problems affecting the construction industry and the service sector. With the exception of the period of the Russian and Asian crises, the doubtful loans to assets ratio stabilized after 1997. In further evidence, new allowances to loan loss provisions, an ex ante measure of risk in contrast to realized doubtful loans, also followed a similar pattern over the sample period (see BNP Paribas 2001). Other risk measures, including overall stock return volatility (STDRET) and idiosyncratic stock return volatility (IDIOSYNCRATIC), exhibit similar behavior, demonstrating the high levels of riskiness of many banks in the sample. Our sample period thus provides potential insights regarding countries with banking systems subject to similarly stressful periods.

A. The Effects of Focus on Bank Return at Different Levels of Bank Risk

Before examining the relationship between bank returns and focus, at different levels of bank risk, we first consider the following linear regression to understand the average relationship between bank returns and focus: We wish to test whether diversification is better for bank returns (“don’t put all your eggs in one basket”) or, by implication, whether focus (increased HHI) is harmful to bank returns:

As noted earlier, Return_t is proxied by two variables: (i) ROA and (ii) SR. Throughout the paper, regressions are run by pooling observations across all banks and across all years and including time dummies (TIME_s) for 1995–99 as well as bank‐specific fixed effects (except when their inclusion in the specification would lead to a multicollinearity problem).

The bank‐specific fixed effects help us control for bank characteristics not captured in our specifications (assuming that they do not change dramatically over time). Furthermore, for bank fixed effects to sufficiently control for the fact that we are using pooled time‐series data for each bank, we require that enough banks switch between diversification and focus. As observed earlier, the time‐series standard deviation of focus measures for an individual bank, averaged across all banks, is about half as large as the overall sample standard deviation of focus measures. Furthermore, if we focus attention on the extremes, then only one bank features in the top 10 focused banks in all years, and only two banks feature in the top 10 diversified banks in all years, when focus and diversification are measured using A‐HHI. The corresponding numbers when measurement is done using I‐HHI are zero and three, respectively. The numbers are virtually the same if one were to compare these deciles in 1993 and 1999, suggesting that the composition of these deciles in 1993 and 1999 is essentially different. These statistics confirm that there is sufficient time‐series variation in an individual bank’s industrial and asset sector diversification.

The time dummies help us control, among other things, for the possible effect of change in macroeconomic conditions. Ideally, we would also like to isolate the linkage between diversification and performance that is specific to the bank’s own activities such as its expertise in screening and monitoring from a possible mechanical linkage arising from a response of the bank’s loan portfolio composition to the demand for loans in different industries. To be specific, if an industry a bank is lending to does relatively well compared to other industries, the bank may optimally lend greater credit to this industry and appear focused as well as performing better at the same time. A possible control for this would be the relative performance of the industries over time, proxied, for example, by the MSCI for Italian industries. Unfortunately, the industry classification of loans employed in our data does not map nicely into the one employed by MSCI data.

In addition, we employ the following control variables for each bank: log of its size LN(SIZE), its equity to assets ratio EQRATIO, its branch to assets ratio BRRATIO, and its employment expense to assets ratio EMPRATIO, all measures in year t. Note that since we use log of SIZE as the explanatory variable and simultaneously employ time fixed effects, the measurement of SIZE in U.S. dollars or Italian lira does not affect the coefficient on LN(SIZE): fluctuations in the dollar‐lira exchange rate from the beginning of one year to the next affect only the coefficients on time fixed effects. Finally, we adjust returns for risk by employing the risk measure DOUBT_t−1, the ratio of its doubtful and nonperforming loans to assets, also as an explanatory variable.

We then test whether, in contrast to the specification in equation (1), the return‐focus relationship depends on the level of bank risk. The return‐focus relationship may in fact depend in a nonlinear way on bank risk (see, e.g., Winton 1999). From traditional portfolio theory, we know that diversification increases the central tendency of the distribution of a loan portfolio. However, when debt is risky and the central tendency of distribution is low relative to the level of debt, diversification can in fact increase the probability of bank insolvency. This would occur, for example, if the downside risk of bank loans is substantial. For the sake of illustration, figure 1 plots the cumulative probability function for two normal distributions with different standard deviations and with a common mean of zero. Suppose that these distributions (suitably scaled) correspond to two possible distributions for realization on bank loans. Suppose further that the level of debt varies along the x‐axis.

(22 KB)

Fig. 1.— The effect of diversification (focus) on the probability of failure. The figure plots the cumulative probability function, Prob( ), for two normal distributions with different standard deviations and with a common mean of zero. The thick line denoted as “less diversified” has a standard deviation of 1.0, whereas the dashed line denoted as “more diversified” has a lower standard deviation of 0.5. For the sake of illustration, z is treated as the distribution of bank returns and x as the level of bank debt (under a suitable scale). If the level of debt x is to the left of the central tendency of zero, e.g., at , then a decrease in standard deviation, by reducing the likelihood of events in the left tail of the distribution (the “default” states), reduces the probability of default. However, if the level of debt x is to the right of zero, e.g., at , then a decrease in standard deviation, by reducing the likelihood of events in the right tail of the distribution (the “no‐default” states), in fact increases the probability of default.

Open New Window

If the level of debt is to the left of zero (under a suitable scale), for example, at , then a decrease in standard deviation, by reducing the likelihood of events in the left tail of the distribution (the “insolvency” states), reduces the probability of default. However, if the level of debt is to the right of zero, for example, at , then a decrease in standard deviation, by reducing the likelihood of events in the right tail of the distribution (the “no‐default” states), in fact increases the probability of default. The left‐skewed nature of a typical loan portfolio’s return distribution implies that the level of debt, in fact, may not need to be too high for this effect to arise. Thus there may be an inverted ‐shaped relationship between return and diversification as the level of risk increases from low to high. And, by implication, the relationship between return and focus may be a ‐shaped function of the level of risk.

An additional impact reinforcing the ‐shaped (nonlinear) hypothesis is the conflict of interest between bank owners and bank creditors. Specifically, an increase in the probability of insolvency reduces the incentives of bank owners to monitor their loans. If the loan portfolio has high downside risk (i.e., a high probability of asset returns falling below deposits, making the bank insolvent), then an improvement in loan monitoring and, in turn, in loan quality produces greater benefits to the creditors than to the bank owners. Since the cost of monitoring is borne by the bank owners (the residual claimants), it follows that if the loan portfolio has high downside risk, then an increase in diversification leads to weaker incentives for bank owners to monitor loans. This, in turn, leads to lower bank returns.

To try to capture the implied ‐shaped (nonlinear) nature of the return‐focus relationship as a function of bank risk, we modify equation (1) by introducing interaction terms between the focus measures and our measure of risk, the nonperforming and doubtful loans (RISK) as well as risk squared (RISK²). That is, where is a vector representing the non–risk control variables stated above. Under this specification, the effect of focus on returns is quadratic in risk. For example, for industrial focus, I‐HHI (where we have suppressed the bank‐specific index i), If, for example, the effect of a bank’s focus on its returns is ‐shaped in risk, then

As stated above, we employ different measures of bank RISK in the regression above: the ratio of doubtful and nonperforming loans to assets, DOUBT_t−1, the standard deviation of DOUBT, STDOUBT, and loan loss provisions to assets ratio, PROVISION_t−1. While DOUBT is a measure of realized losses, STDOUBT and PROVISION are potentially more attractive as ex ante measures of unexpected and expected bank insolvency risk, respectively. Note that there is only one value of STDOUBT for a bank over the entire period. Hence, the time index in RISK_t−1 is not relevant when risk is proxied by STDOUBT. In general, the risk measures we employ are based on discretionary actions of bank owners. To eliminate any bias arising from this, we also employ for the publicly traded sample two stock return–based measures of unexpected bank risk: the total stock return volatility of a bank, STDRET, and its idiosyncratic volatility, IDIOSYNCRATIC.

Table 4 presents the results for regressions of bank returns on focus specified in equations (1) and (3) with ROA as the bank return and DOUBT, STDOUBT, and PROVISION as the risk measures.11 Overall, the view that focus reduces bank returns (and thus diversification increases bank returns) is rejected for both measures of loan portfolio focus: industrial and household focus (I‐HHI) and broad asset sector focus (A‐HHI), as reflected by the positive and statistically significant (mostly at the 5% confidence level) coefficients on these measures in columns 1 and 2. The inclusion of control variables in column 2 significantly enhances the explanatory power of equation (1). The control variables for a bank’s capital ratio and the risk of its loans (doubtful and nonperforming loans to assets ratio) are strongly significant in their effect on ROA.

Table 4 Test for Nonmonotonicity in Effect of Focus on Bank Returns on Assets: 105 Italian Banks, 1993–99

Open New Window

Columns 3–5 of table 4 test whether the return‐focus relationship is nonlinear in the level of bank risk, thus linking the cross‐sectional effect of focus on returns to the level of bank risk (see eq. [3]). Interestingly, these results provide support for a ‐shaped relationship between focus and returns as a function of the level of risk of the bank. The coefficients on the interaction terms, and , are negative and positive, respectively, and are statistically significant (in some cases at the 5% and in all but one case at the 10% levels). This holds for both measures of focus, I‐HHI and A‐HHI, and for all three measures of bank risk, DOUBT, STDOUBT, and PROVISION. Computation of F‐statistics to test the statistical significance of the linear and quadratic terms, separately and jointly, revealed that the coefficients on these terms are statistically significant (at a 99% confidence level) in contributing to the explanatory power of the regressions in columns 3–5 of table 4. As noted before, this ‐shaped relationship is consistent with a weakening of bank monitoring incentives, on an increase in diversification, in the case in which the downside or insolvency risk of the bank is high.

In table 5, we repeat these tests with SR as the bank return measure. In addition, we employ stock return–based measures of bank risk. The sample size is smaller for the stock return–based measures of bank returns since only 34 out of our 105 banks are publicly traded. The control variables for a bank’s capital ratio and the risk of its loans, which were strongly significant in their effect on ROA, have a less significant impact on the bank’s SR. The coefficients on I‐HHI and A‐HHI in columns 1 and 2, corresponding to estimation (1), are strongly significant.

Table 5 Test for Nonmonotonicity in Effect of Focus on Bank Stock Returns: 105 Italian Banks, 1993–99

Open New Window

Overall, the ‐shaped relationship finds some support with SR as the measure of bank return. Most coefficients on the linear and quadratic interaction terms, and , are significant or marginally significant, but a few are insignificant. The ‐shaped relationship fares relatively better when bank risk is proxied by DOUBT, STDOUBT, PROVISION, or IDIOSYNCRATIC, compared to STDRET as the proxy for bank risk. In terms of signs, all coefficients have the correct signs except the linear terms with STDRET as the risk measure, which are found to be positive. Note, however, that a positive sign of the linear coefficients provides even further evidence that the effect of diversification on bank returns is not positive. Moreover, once we control for endogeneity of focus measures, the coefficients always take correct signs and are statistically significant. However, before proceeding to this endogeneity correction, we discuss the magnitude of the effects documented so far, in particular of the ‐shaped relationship between focus and returns as the level of bank risk changes.

To understand the economic significance of the ‐shaped relationship, we plot the marginal effect d(Return)/d(Focus) for different values of RISK for both measures of focus, I‐HHI and A‐HHI, and for different measures of Return and RISK. In figures 2a and 3a, we employ ROA as the return measure and employ DOUBT and STDOUBT as the risk measures, respectively. In figures 2b, 3b, 4a, and 4b, we employ SR as the return measure and employ DOUBT, STDOUBT, STDRET, and IDIOSYNCRATIC as the risk measures, respectively. In all plots, the marginal effect is plotted for both I‐HHI (thick line) and A‐HHI (dotted line). The range of the RISK proxy is measured over the spectrum covered by that proxy for the Italian banks in our sample over the period 1993–99 (panel B of table 3).

(36 KB)

Fig. 2.— Economic significance of the relationship between bank return and bank focus, which is nonlinear as a function of bank risk. a, Nonmonotonicity in effect of focus on ROA as a function of DOUBT. b, Nonmonotonicity in effect of focus on SR as a function of DOUBT. The figure plots the marginal effect d(Return)/d(Focus) as specified in eq. (4), the underlying specification for which is eq. (3). The marginal effect is plotted for both focus measures, I‐HHI and A‐HHI. The coefficients used to plot the relationships are obtained from table 4 and table 5. For each figure, the range of respective risk variable is taken to be between 0% and an upper bound that covers the minimum and the maximum values over our sample period (see panel B of table 3). The definitions of all variables and also a description of how they are computed are presented in Sec. II.

Open New Window

(49 KB)

Fig. 3.— Economic significance of the relationship between bank return and bank focus, which is nonlinear as a function of bank risk. a, Nonmonotonicity in effect of focus on ROA as a function of STDDOUBT. b, Nonmonotonicity in effect of focus on SR as a function of STDDOUBT. See also the legend to fig. 2.

Open New Window

(46 KB)

Fig. 4.— Economic significance of the relationship between bank return and bank focus, which is nonlinear as a function of bank risk. a, Nonmonotonicity in effect of focus on SR as a function of STDDRET. b, Nonmonotonicity in effect of focus on SR as a function of IDIOSYNCRATIC. See also the legend to fig. 2.

Open New Window

Consider figures 2a and 3a. These graphs are based on estimated coefficients from table 4, columns 3 and 4, respectively. As can be seen in these plots, d(ROA)/d(I‐HHI) is close to zero at low risk values, is small and negative at moderate risk levels (5–10% for DOUBT and 2–14% for STDOUBT), and is positive and sharply rising at high risk levels. The spectrum of high risk levels in which the effect is positive and sharply rising consists of the highest risk decile (about 10% of the sample in each year) in the case of DOUBT and the highest quartile (about 25% of the sample) in the case of STDOUBT.12

A natural question to ask is whether these observations are outliers that should be ignored. In fact, it turns out that these observations cannot be treated as mere outliers and discarded for banking systems under stress. As mentioned earlier, the 1990s were a particularly difficult period for many Italian banks and industries. We examined, for example, the sets of banks in each year with a DOUBT ratio in the top 10% of DOUBT ratios across all banks in that year. Importantly, we found that many banks experienced fluctuations in their DOUBT values from being very low to very high. This is captured in the high values of STDOUBT, the standard deviation of DOUBT, in table 1 and panel B of table 3. However, different banks experienced these fluctuations at different points during the sample period. Eliminating observations with high DOUBT values would thus amount to retaining only those data points for each bank that correspond to low or moderate values of DOUBT. Moreover, if the top 10% observations of DOUBT were omitted in each year, this would correspond to omission of over 25 banks (about one‐fourth of our sample size of 105 banks) across different years.13 That is, banks with the highest values of DOUBT in any given year are not necessarily the same banks with the highest values of DOUBT in other sample years.

Thus it appears that diversification across industries and asset sectors is not particularly beneficial for the bank returns and may in fact be especially costly for high‐risk banks. For example, for a bank with DOUBT of 25% in the previous year, the effect of increasing industrial focus from 0.16 (approximately equally exposed to six industries) to 0.20 (approximately equally exposed to five industries) is to increase its next year’s ROA by approximately 0.80%. Note that such a bank lies in the highest DOUBT decile. A similar increase in focus for a bank with a standard deviation of DOUBT of 20% results in an increase in its return of approximately 0.40%. Such a bank lies in the seventy‐fifth–ninetieth percentile region of STDOUBT in our sample. Given that the mean ROA is 0.93% with a standard deviation of 0.85% in our sample (table 1), these effects are clearly economically important.14 A similar conclusion is drawn from figures 2b, 3b, 4a, and 4b, where SR is employed as the return measure and the risk measures employed are DOUBT, STDOUBT, STDRET, and IDIOSYNCRATIC, respectively.

B. Endogeneity of Focus Measures

In our tests so far, we employed focus measures with a lag; that is, we considered the effect of Focus_t−1 on Return_t. This helps to partially address the endogeneity issue. Arguably, it is appropriate for ROA_t, since any monitoring‐related effects of focus may get captured in book returns only with a lag. However, this is less justifiable in the case of stock returns since they will reflect contemporaneous information as to the expected effects of any focus changes (assuming these changes are publicly observable). Hence, it is important to consider the effects of Focus_t on Return_t. Furthermore, the fact that diversified banks seem to be either underperforming or certainly not dominating the focused banks begs the question as to why some banks are undertaking performance‐destroying (or value‐destroying) diversification. These questions pertain to the issue of the endogeneity of bank focus measures: Specifically, if a bank has some latent characteristic that induces it to be focused and simultaneously results in greater bank returns, then estimations of equation (3) will likely produce biased estimates.15 We address this endogeneity issue next.

To account for the possible endogeneity of focus measures, we estimate a simultaneous equations system in which Return_t and Focus_t are both treated as variables to be explained and the error terms of the two equations in the system are allowed to be correlated with each other. This is essentially a seemingly unrelated regression (SUR) approach (see Theil 1971; Johnson 1972; Maddala 1977). In order to prevent the system from growing too large in terms of the number of coefficients to be estimated and, in turn, to retain statistical power in the estimation, we alternately treat one of the two focus measures, I‐HHI_t and A‐HHI_t, as endogenous in year t and the other as its exogenous value in year . When treating I‐HHI_t (A‐HHI_t) as endogenous, we employ A‐HHI_t−1 (I‐HHI_t−1) as an explanatory variable only for Return_t and not for I‐HHI_t (A‐HHI_t). This ensures that the order conditions for identifying the system are satisfied.

For the endogenous determination of Focus_t, we considered a number of independent explanatory variables as instruments:

The ex ante rationale for the use of these instruments is as follows. A large body of banking literature has shown a positive relationship between diversification and size. The standard arguments are based either on the finiteness of good projects or on diminishing returns to scale within an industry, and on the risk avoidance induced by relatively high franchise values of large banks. National banks and money center banks may have greater size and scope by definition and thus intrinsically may be more diversified. Private banks, state‐owned banks, and consortium banks may have an objective function, and, in turn, a focus or diversifying strategy, that differs from their publicly owned counterparts. For example, private banks may face less corporate governance scrutiny than public banks, state‐owned banks may be forced to lend to certain sectors or industries to fulfill state objectives (see Sapienza 2002), and consortium banks may be following a collective focus or diversifying strategy conceived at the level of the consortium. Banks with a high deposit to assets ratio may not be well diversified on the liability side and perhaps rely significantly on “core” deposits. The need to focus or diversify for these banks will differ from that of banks well diversified on the liability side, for example, those with greater access to the purchased funds market. Finally, average focus across all banks in a given year potentially captures macroeconomic conditions and the regulatory environment, not fully captured through other instruments.

Table 6 provides a summary of the bank characteristics and the instruments for the focused and the diversified subsample. In each year t from 1994 to 1999, banks were divided into two groups, focused and diversified, on the basis of whether their HHI measure in year t is below or above the median. Using univariate analysis, the table shows the industrially focused group of banks to have a higher return on assets and on equity, a greater asset sector focus (higher A‐HHI), a larger size, a smaller ratio of employees to assets, and a smaller doubtful loans to assets ratio. Publicly traded banks, banks that are not state‐owned, and banks that are part of a consortium group are more likely to be industrially focused than diversified. While the overall pattern is somewhat similar for an asset sector–focused group of banks, this group of banks is smaller in size compared to an asset sector–diversified group and is less likely to be from the set of national banks. Overall, the table suggests that the set of instruments identified above should have explanatory power for the endogeneous focus or diversification choice of banks.

Table 6 Properties of Focused and Diversified Banks

Open New Window

The simultaneous system of equations resulting from the choice of these instruments is presented below when I‐HHI_t is treated as endogenous (other specifications we estimate will be described later): where is a vector representing the control variables (LN(SIZE), EQRATIO, BRRATIO, EMPRATIO, and STATE‐OWNED DUMMY), is a vector representing the instrumental variables (NATIONAL DUMMY, PRIVATE DUMMY, DEPOSIT TO ASSET RATIO_t−1, GROUP DUMMY, and AVERAGE I‐HHI_t), and the error terms and may be correlated. Note that LN(SIZE) is included in the control variables and thus serves as a potential instrument for the focus measures. Furthermore, in contrast to the specifications examined in tables 4 and 5, STATE‐OWNED DUMMY is also included as a control variable for explaining returns. This is to allow for a possible direct effect of state ownership on bank returns due to inefficiencies, such as higher overheads, looser expense controls, and wasteful bureaucracy, that are more likely to plague state‐owned banks. Time dummies and bank‐specific fixed effects are included in determining both Return_t and I‐HHI_t (except when their inclusion leads to a multicollinearity problem). Under the specification of equations (6) and (7), the effect of focus on returns continues to remain quadratic in risk. Formally, and where we have suppressed the bank‐specific index i.

The results are reported in table 7 (for ROA in panel A and for SR in panel B). In panel A, estimated coefficients are reported for ROA_t and I‐HHI_t in columns 1 and 2, with risk measures being DOUBT and STDOUBT, respectively. Columns 3 and 4 report the estimated coefficients for ROA_t and A‐HHI_t. In panel B, the risk measures are STDRET and IDIOSYNCRATIC. Results with other risk measures are not reported for considerations of space. Examining the coefficients on the linear and quadratic interaction terms between focus and risk, we find that the results corrected for the endogeneity of focus provide even stronger and more consistent evidence in support of the ‐shaped relationship. Indeed, all coefficients have the correct sign and are statistically significant at the 10% confidence level or better. The implied marginal effects of focus on return, as risk is varied, are plotted in figures 5a, 5b, 6a, and 6b. These correspond to results in columns 1 and 2 of panel A of table 7 and columns 1 and 2 of panel B, respectively, where industrial focus I‐HHI_t is treated as endogenous, and are the counterparts of figures 2a, 3a, 4a, and 4b, respectively. The marginal effects when A‐HHI_t is treated as endogenous are not plotted for considerations of space.

Table 7 Simultaneous (SUR) Estimation of Effect of Focus on Bank Return Treating Focus as Endogenous Variable

Open New Window

(53 KB)

Fig. 5.— Economic significance of the relationship between bank return and bank focus, which is nonlinear as a function of bank risk. a, Nonmonotonicity in endogeneity‐corrected effect of focus on ROA as a function of DOUBT. b, Nonmononotonicity in endogeneity‐corrected effect of focus on ROA as a function of STDDOUBT. The figure plots the marginal effect d(Return)/d(Focus) as specified in eq. (4), the underlying specification for which is the simultaneous system of eqq. (6) and (7). The marginal effect is thus corrected for the endogeneity of focus measures, as described in Sec. III.B. In each plot, the marginal effect is plotted for both focus measures, I‐HHI and A‐HHI. The coefficients used to plot the relationships are obtained from panel A of table 7, cols. 1 and 2. For each figure, the range of respective risk variable is taken to be between 0% and an upper bound that covers the minimum and the maximum values over our sample period (see panel B of table 3). The definitions of all variables and also a description of how they are computed are presented in Sec. II.

Open New Window

(49 KB)

Fig. 6.— Economic significance of the relationship between bank return and bank focus, which is nonlinear as a function of bank risk. a, Nonmonotonicity in endogeneity‐corrected effect of focus on SR as a function of STDRET. b, Nonmonotonicity in endogeneity‐corrected effect of focus on SR as a function of IDIOSYNCRATIC. The coefficients used to plot the relationships are obtained from panel B of table 7, cols. 1 and 2. Also see the the legend to fig. 5.

Open New Window

Most notably, all the marginal effects are ‐shaped. In particular, the statistical significance of the effect with SR as the return measure and STDRET as the risk measures, which were relatively weak earlier (col. 6 of table 5 and fig. 4a), are now stronger and the coefficients have the expected signs. Similarly, the positive effect with SR as the return measure and IDIOSYNCRATIC as the risk measure, which spanned only a small range of risk values (col. 7 of table 5 and fig. 4b), is now uniformly positive after the endogeneity correction.

It is also of interest to examine the estimated coefficients in the endogenous determination of focus measures. The effects overall are similar for both focus measures, I‐HHI (cols. 1 and 2) and A‐HHI (cols. 3 and 4). Large banks and national or money center banks are more diversified as reflected by the negative sign on LN(SIZE) and NATIONAL DUMMY in the focus regressions. Interestingly, private banks are more diversified than public banks, an effect that is quite strong statistically. All else being equal, state‐owned banks are more focused, consistent with Sapienza's (2002) conclusion that these banks have an objective that is geared toward supporting specific industries, often at subsidized rates. The deposit to assets ratio and average focus of all banks in the given year do not seem to have any incremental effect, whereas being part of a consortium has a statistically insignificant effect for industrial focus but a negative effect on asset sector focus.

Interestingly, the effect of past losses or risk (DOUBT_t−1, STDOUBT, STDRET_t−1, and IDIOSYNCRATIC_t−1) on focus is always negative and significant. This implies that, all else being equal, banks that are overall risky or have recently experienced higher losses or increases in their stock return volatility choose to focus less, that is, diversify more. This lends support to the need for the endogeneity correction we have employed: If banks that choose to diversify (focus) are precisely the ones that are loss‐making (profit‐making) or risky (safe), then a negative relationship between return and diversification arises even in the absence of any direct causal effect of diversification on return. In other words, the negative relationship between return and diversification may be “spurious” in that it simply reflects which banks select to diversify and which banks choose to focus. The results in table 7 show convincingly, however, that even though this selection problem is present in our sample, it is not solely responsible for the relationship between diversification and return. The empirical relationship confirmed in table 7 confirms that some of the diversified banks, especially the riskier ones, might benefit from choosing to increase their focus instead.

Overall, our results lend empirical support for diversification (focus) having a small benefit (cost) at low bank risk levels, and in fact, hurting (helping) bank returns at very high risk levels. We find this to hold for both industrial and asset sectoral focus, for return on bank assets as well as stock returns of banks, and for a variety of accounting and stock return–based measures for unexpected and expected bank risk. It is important to note, however, that examining bank returns is only one side of the trade‐off between return and risk. Next, we examine the other side of the trade‐off, the effect of the decision to focus (diversify) on bank loan risk.

C. The Relationship between Focus and Risk: The Effect of Expanding the Loan Portfolio in New Areas

In order to study the effect of focus (diversification) on bank monitoring effectiveness and, in turn, on the quality of bank loan portfolios as banks expand the scope of their loan portfolios, we consider first the risk of bank loans as a dependent variable in the regression where, as before, are the non–risk control variables augmented to include past returns (ROA_t−1 or SR_t−1), and risk is proxied by the variable DOUBT_t, STDRET_t, or IDIOSYNCRATIC_t. If an increase in focus (increase in HHI) reduces the risk of bank loan portfolios, then

There are at least three reasons why this might arise. First, banks may lack the monitoring expertise in lending to a new sector when learning costs are present. Second, when the loan sector to which banks migrate is already being supplied with credit by other banks, new bank entrants may be subject to adverse selection and a “winner’s curse” effect.16 This suggests that diversification could lower returns on bank loans and increase the risk of failure to a greater degree when the sectors into which the bank expands are subject to greater competition. Third, diversification can cause a bank to grow in size, subjecting it to agency‐based scale inefficiencies.

Consequently, entering into “new” loan sectors may adversely affect bank loan portfolio quality (increase risk). Note that we use the qualifier “newer” for those industries in which previous exposures of the bank have been relatively small or nonexistent (rather than being newer in the sense of technological or productive aspects of the industry such as dot.com firms). To test this relationship, we construct two variables, NEW_t and FRACNEW_t, defined as follows:

Finally, we also introduce an additional variable, COMP_t, that measures the extent of competition a bank faces in lending for its top five industries, defined as follows.

To test whether the hypothesis concerning deterioration of the quality of a loan portfolio (increase in bank risk) occurs upon entry into “newer” industries (i.e., reduced focus), we modify regression (10) along two dimensions. First, we introduce NEW_t, FRACNEW_t, and COMP_t−1 as explanatory variables for RISK_t. Second, we introduce interaction terms between these three variables and the two focus measures I‐HHI_t−1 and A‐HHI_t−1.19 The resulting specification is

Consider the marginal effect of NEW_t on RISK_t. We obtain where we have suppressed the bank‐specific index i. The null hypothesis is that d(RISK_t)/d(NEW_t) is positive and is increasing in bank diversification or decreasing in bank focus. The reason is that for a well‐diversified bank, the effect of entry into new industries is primarily one of spreading its monitoring resources more widely. By contrast, for a focused bank, the effect of entry into new industries is beneficial from a traditional diversification standpoint and is also less harmful from the standpoint of a deterioration in monitoring quality since, even with an additional industry, the bank remains relatively specialized. That is, the constant term ν₁₀ is positive and the interaction term coefficients ν₁₁ and ν₁₂ are negative. The hypothesis with respect to the marginal effect of FRACNEW_t and COMP_t−1 on RISK_t take similar forms yielding the overall hypotheses20

Table 8 presents empirical evidence on how the decision to focus or diversify endogenously affects the risk of bank loan portfolios by reporting the results of tests of equations (10)–(12) above. The first three columns in table 8 correspond to the entire sample in which the risk measure employed is doubtful and the nonperforming loans to assets ratio DOUBT_t, and the last six columns correspond to the publicly traded sample in which the risk measures employed are stock return volatility STDRET_t and its idiosyncratic component IDIOSYNCRATIC_t. In each panel of three columns, the first two columns correspond to the test of hypothesis (11) and the third column corresponds to the test of hypothesis (14).

Table 8 Test for Effect of Focus on Bank Loan Risk

Open New Window

From columns 1 and 2 in each panel of table 8, we observe that the effect of both industrial and asset sectoral focus on bank risk is negative and statistically significant. The effect is also economically significant. For example, for risk measure DOUBT_t, the effect of increasing a bank’s industrial focus from 0.16 (approximately equally exposed to six industries) to 0.33 (approximately equally exposed to three industries) in year is to decrease the bank’s year t doubtful and nonperforming loans to assets ratio by approximately 0.51%. Note that the average DOUBT value in the sample period is 5.23% with a standard deviation of 5.63%. The effect is of similar magnitude for stock return–based volatility measures.

Furthermore, the above effect persists even after we control for endogeneity of the focus measures. In table 9, we consider a simultaneous equations estimation of RISK_t and Focus_t. The focus specification we test for the presence of endogeneity is identical to that of Section III.B, and, as can be seen, the coefficients on both focus terms, I‐HHI and A‐HHI, are always negative and statistically significant.

Table 9 Simultaneous (SUR) Estimation of Effect of Focus on Bank Loan Risk Treating Focus Measures as Endogenous Variables

Open New Window

Finally, column 3 in table 8 reveals that when a bank enters “new” industrial sectors, loan risk increases at a rate that is increasing in the extent of diversification of the bank. That is, the direct coefficient on NEW_t is always positive (though only marginally significant), and the coefficient on interaction terms between NEW_t and the two focus measures is negative and significant. For highly diversified banks (low I‐HHI and A‐HHI), the effect of moving into new industries is to increase bank risk (e.g., increase in DOUBT of 0.5% at the lowest values of I‐HHI and A‐HHI in the sample). For moderate diversification, the effect is close to zero (e.g., at average values of I‐HHI and A‐HHI in the sample). Finally, for highly focused banks, moving into new industries in fact reduces bank risk. The variable FRACNEW_t, the fraction of bank loan portfolio in the newer industries, has no substantial effect on bank risk.

Stronger than the effect of entry into newer industries is the effect of competition that a bank faces in lending (in the five largest industries it has loan exposures to). The direct coefficient on COMP_t−1 is positive and significant. This suggests that banks facing greater competition have riskier portfolios. This could be due either to the negative effect of competition on profits, which in turn provides risk‐shifting incentives (see Allen and Gale 2000), or the effect of market competition on charter values (see Keeley 1990). In terms of economic magnitudes, consider the simple example of two banks that are otherwise identical but one is a leader in one of its top five industries, holding an 80% share. The other bank is a relatively smaller loan player in this same industry, which does, however, belong to its own top five industries in terms of exposure amounts, holding the remaining 20% share of the market. The difference in competition faced by these two banks contributes to the difference in their doubtful loans to assets ratio of , where 2.3% is the estimated coefficient on COMP_t−1 in column 3 of the DOUBT panel in table 8.

Furthermore, the risk‐increasing effects of competition are greater the more diversified banks are. The coefficients on the interaction terms between COMP_t−1 and focus measures, I‐HHI_t−1 and A‐HHI_t−1, are both negative and statistically significant. In other words, an increase in focus, that is, a decrease in diversification, reduces risk more when the competition that the bank faces in its loan sectors is higher. This interaction effect is, however, economically small compared to the direct effects of focus measures on bank risk and the direct effect of competition on bank risk as well as the interaction effect of focus measures and entry into newer industries.

These results provide some evidence suggesting that the quality of monitoring by banks is poorer in newer industries and that banks face greater adverse selection when they expand into industries that have been previously penetrated by their competitors. This also suggests that if banks take the effect of lending competition into account and are value‐maximizing, then they should choose to diversify (if at all) into industries with lower penetration by other banks, as proposed by Boot and Thakor (2000). In a recent paper, Hauswald and Marquez (2002) also demonstrate that bank incentives to concentrate informational resources are increasing in the degree of adverse selection they face in the market, which in turn would be greater if banks expand by lending more to industries in which (lending) competition is strong.21

D. Additional Robustness Tests and Results

State‐owned versus private banks.—Sapienza (2002) finds that the objective functions of state‐owned Italian banks differ from those of private Italian banks. State‐owned banks charge lower interest rates than privately owned banks to similar or identical firms, even if the company is able to borrow more from privately owned banks. Further, she finds that state‐owned banks mostly favor firms located in depressed areas and large firms. This makes it plausible that a part of the inefficiency arising from diversification may simply be due to the presence of state‐owned banks in our sample. As a check, we employed the same classification of state‐owned and private Italian banks employed by Sapienza (2002) and reexamined our hypotheses for the private (not state‐owned) bank sample. On the basis of the available classification at the beginning of 1993, 34 banks in our sample were privately owned. The qualitative nature and the significance of our results remain unaffected by restricting our analysis to this smaller sample: (i) both focus measures improve bank returns on average, and the effect of focus on returns is ‐shaped as a function of bank risk; (ii) both focus measures reduce bank risk.22

National versus intraregional and local banks.—The measure of focus and diversification employed in our paper concerns the asset side of the bank balance sheet; that is, it is based on a bank’s loan exposures to different industries and sectors. The effect of changes in focus or diversification might affect money center banks differently since they do not rely as heavily on core (local) deposits. To check for links between asset‐side focus and performance while controlling for the liability structure of banks, we employed the classification of banks in our sample into national banks and nonnational (i.e., intraregional or local) banks. Eight out of nine national banks in our sample were also identified as money center banks. Estimation of the effects of focus (diversification) on return (tables 4 and 5) and on risk (table 8) separately for the sample of national banks and the rest of the banks produced qualitatively similar patterns for both samples. This confirms that our results are not driven by the presence of the large, national banks.23

Consortium banks.—Another feature of certain Italian banks in our sample reflects the fact that they are “part of a bank group or a consortium.” Since bank strategy to focus or diversify might be determined at a consortium‐wide level, it might be deemed as more appropriate to measure return and risk of such banks also at a consortium‐wide level. Consequently, we estimated the effects of focus (diversification) on return (tables 4 and 5) and risk (table 8) separately for the subsample of banks that are not a part of a bank group or consortium. There were 70 such banks in our sample. While the overall pattern remains qualitatively unaffected, we find that in fact, the harmful effect of diversification on risk is actually more pronounced.24

Large banks.—One final concern could be that a few big and well‐diversified banks in our sample, especially Banco di Napoli, Banco di Sicilia, and Banca Popolare di Novara, had large negative shocks to their profits in the first half of our sample period and their bad loans increased dramatically, primarily for regulatory and institutional reasons. We do include bank size as a control variable, but to be absolutely sure that these observations are not driving our results, we reran our regressions two different ways: (1) by excluding these three banks from the sample altogether and (2) by introducing an interaction term between the size variable and time dummies. The exclusion of these three banks hardly affects the quantitative or the qualitative nature of our results, and the interaction between size and time effects is almost always insignificant.25

E. Overall Effects of Diversification on Bank Performance

Combining the empirical findings of tables 3–8 regarding the effects of diversification (focus) on bank returns and bank loan portfolio risk, we summarize our results in terms of their implications for the benefits of loan portfolio diversification as follows:

We conclude the following for our sample of banks:

Crucially, a robust finding that emerges from our results is that the conventional wisdom of not putting all of one’s eggs in a single basket cannot be applied uniformly to all banks. That is, diversification per se is no guarantee of superior performance or greater bank safety and soundness, which is a major goal of regulatory policy.