DOES REVENUE DIVERSIFICATION STILL MATTER IN BANKING? EVIDENCE FROM A CROSS-COUNTRY ANALYSIS

Banks have been revising their business models since the ̄nancial crisis, diversifying income sources to pursue pro ̄tability and stability in a rapidly evolving environment. The e®ectiveness of this strategy is still debated. We investigate if revenue diversi ̄cation of 1250 EU and US banks improved performance or its stability between 2008 and 2016. We adopt a broad econometric approach and de ̄ne diversi ̄cation as the share of non-interest revenue and the HH index of the net operating income. We ̄nd that diversi ̄cation is not clearly associated with performance or its volatility, that bene ̄ts change remarkably over time and, where present, show signi ̄cant variability. Our results support recent evidence on the limitations of diversīcation in banking, raising potential concerns on converging supervisory practices and general calls for revenue diversity. The variability of business models and the impacts of di®erent economic and institutional environments matter.


Introduction
The evolution of business models is inherent to all¯rms. In the banking sector, this process showed an increased speed over the last decade. Leading changes includē nancial innovation and technological advances but extend to shifts in clients' behavior and lessons learnt from the¯nancial crisis and embedded in recent supervisory and regulatory responses.
Diversi¯cation appears as a natural choice in order to restore or strengthen pro¯tability in an uncertain environment. If excessive concentration could threat stability, banking diversi¯cation seems unable to be free from drawbacks: the recent academic debate underlines its limitations and the lack of unambiguous results.
Nonetheless, supervisors call for more income diversi¯cation, especially in Europe where the focus on traditional commercial banking is high. Fee-based revenues could be able to counterbalance the degrading quality of loan portfolios and smooth pro¯tability patterns. However, during recessions, an increase in volatility and covariance of income sources can o®set diversi¯cation bene¯ts and lead to a return to concentration.
The appearance and growth of FinTech companies, despite being worth a fraction of the global banking business, is both a threat and an opportunity for banks struggling with performance, with impacts on traditional (for example, interest income arising from¯nancing SME's working capital digitalizing assessment processes) or diverse sources of revenues (for example, fees and commissions generated from automated asset management advice services).
Moreover, as a response to the¯nancial crisis, regulation and supervision changed dramatically worldwide, but with timing, breadth and depth that vary widely across countries. Changes in regulatory capital absorbed by di®erent activities, in quantity and quality of scrutiny within the Supervisory Review and Evaluation Process (SREP), together with unprecedented changes in the underlying environment (above all, the EU Banking Union) are crucial drivers conditioning banks' business models, operations, pro¯tability and stability over time.
The purpose of this paper is to test the attitude of revenue diversi¯cation in enhancing bank pro¯tability or reducing its volatility. In particular, we examine the relationship between the degree of diversi¯cation (i.e. the share of non-interest income or the HH Index of the net operating income) and alternative declinations of risk-return pro¯les.
As a sample to test our hypotheses, we build a dataset of 1250 listed EU and US banks for the period between Q1-2008 and Q4-2016. In terms of methodology, we¯rstly explore the cross-sectional nature of our data through OLS regressions on mean values. Then we employ a dynamic¯xed-e®ect panel model to assess its time dimension. Finally, we run static panel regressions to compare within and between dynamics of our panel data.
Our contribution is to empirically test if diversi¯cation bene¯ts hold in a crosscountry dataset, through di®erent econometric approaches, over a recent sample period experiencing di®erent macroeconomic conditions and evolutionary trends in business models.
While an extended literature investigates the e®ects of revenue diversi¯cation on banks' risks and performance in the US (for example, De Young & Rice 2004a, 2004b, Stiroh 2004, Europe (Kohler 2014, Kholer 2015, Mergaerts & Vander Vennet 2016 and Emerging Markets (Berger et al. 2010a(Berger et al. , 2010b, there are almost no studies comparing US and EU banks for a period including both the subprime and the European sovereign crisis. Additionally, our paper di®ers also in investigating a sizeable cross-country sample including both bigger and smaller institutions, through quarterly data, in order to enhance the computation of revenues volatility. Finally, we are also able to compare big EU and US banks, as well as bigger and smaller US banks (due to the lack of small listed banks in EU).
We¯nd that EU and US banks behave signi¯cantly di®erent in terms of diver-si¯cation bene¯ts. European banks lack evidence of positive impacts of non-interest revenues on both nominal and risk-adjusted performance measures. For the US, instead, e®ects are signi¯cant in terms of performance volatility and risk-adjusted pro¯tability, but substantially di®erent between smaller and larger institutions. Finally, e®ects are stronger between banks rather than within: in other terms,¯rm-speci¯c diversi¯cation strategies and the ability to adapt to the environment seem to matter more than revenue diversity over time.
Our results convey signi¯cant implications. On one hand, the convergence of supervisory frameworks may prove unable to grasp fairly advantages and disadvantages of diversi¯cation across di®erent business models. Additionally, a generalized call for increased revenue diversity does not seem to be backed by empirical data: we provide evidence that bene¯ts are contingent on \which" diversi¯cation, as well as \where" and \when" it occurs. This paper is structured as follows. In Sec. 2, we provide the review of the literature. In Sec. 3, we describe our sample, data, variables and econometric strategy. In Sec. 4, we discuss our¯ndings and their policy implications. Finally, in Sec. 5, we provide our conclusions.

Literature Review and Hypotheses Development
A bene¯cial \portfolio-e®ect" generated by revenue diversi¯cation is common wisdom in banking management. Goddard et al. (2007) analyze the European banking system since the mid-1980s, showing that banks' response to the changing competitive environment has usually included several key strategies: diversi¯cation, product di®erentiation and consolidation. However, the empirical research on the link between diversi¯cation and risks or pro¯tability shows mixed results.
We review the leading literature according to the geographical area of the sample (North America, Europe, Emerging Markets and cross-country). Then, we examine the literature focusing on causes and consequences of income diversi¯cation.
North America. For the US, Boyd & Graham (1986), using a wide sample of large bank holding companies (BHCs) during the period 1971-1983, note that extending to non-bank activities increases the risk of failure. Demsetz & Strahan (1997), studying a sample of BHCs during the period 1980-1993,¯nd that better diversi¯cation does not translate into risk reduction. De Young & Roland (2001) test whether changes in product mix a®ect earnings volatility in 472 commercial banks between 1988 and 1995. They¯nd that switching to fee-based activities generates an increase in leverage and also in volatility of revenues and earnings. De Young & Rice (2004a), examining 4712 commercial banks between 1989 and 2001,¯nd that marginal increases in non-interest income are associated with poorer risk-return trade-o®s. In another study, De Young & Rice (2004b) analyze banks during the period 1986-2003 and¯nd that diversi¯cation gains from fee-based activities appear to be scarce: fee-income boosts bank earnings but increases their volatility. Stiroh (2004) highlights that non-interest income is typically a volatile component of income. At the bank level, a greater reliance on this revenue source is associated with lower risk-adjusted pro¯ts and higher risks. Calm es & Liu (2009)¯nd that non-interest income has driven the variance of Canadian banks' aggregate operating-income growth: by contributing to banking income volatility, market-oriented activities do not necessarily yield diversi¯cation bene¯ts. Al-Obaidan (1999), instead, analyzes a panel of US large commercial banks during the period 1985-1990, and¯nds that while diversi¯cation reduces technical e±ciency, it improves allocative and scale e±ciency, generating an overall economic gain in the industry. Shim (2019) provides evidence that increased loan diversi¯cation has a positive impact on the bank's¯nancial strength.
Europe. Focusing on European banks for the period 1996-2002, Lepetit et al. (2008) show that bank expansion into non-interest income activities generates higher risks, including insolvency. Baele et al. (2007), analyzing a panel of banks over the period 1989-2004,¯nd that a higher share of non-interest income positively a®ects banks' franchise values, but increases their systematic risk. Acharya et al. (2006), analyzing 105 Italian banks over the period 1993-1999,¯nd that diversi¯cation fails to produce greater performance and to reduce risks. Hayden et al. (2007), employing a unique data set of individual bank loan portfolios of 983 German banks for the period 1996-2002,¯nd scarce evidence of signi¯cant links between performance and diversi¯cation, that seems to be associated with reductions in bank returns, even after controlling the risk. Busch & Kick (2009), through a panel of German banks during the period 1995-2007, show that risk-adjusted returns on equity and total assets are positively associated to higher fee-income activities; however, a strong engagement in fee-generating activities goes along with higher risks. Again on Germany, Kohler (2014) investigates the impact of non-interest income on bank risks between retail and investment banks, showing that the former increases their stability if they expand their non-interest income, while the latter become riskier. Mercieca et al. (2007), using a sample of 755 European small banks for the period 1997-2003,¯nd an inverse association between non-interest income and bank performance. Kholer (2015), analyzing the impact of business models on bank stability in 15 European countries between 2002 and 2011, shows that banks are more stable and pro¯table if they increase the share of non-interest income. Mergaerts & Vander Vennet (2016), focusing on a large sample of banks from 30 European countries over the period 1998-2013,¯nd that higher levels of diversi¯cation are associated with higher pro¯tability.
Emerging Markets and cross-country. On emerging economies, Berger et al. (2010a), investigating Russian banks during the period 1999-2006, show that performance tends to be non-monotonically associated to the diversi¯cation strategy. In a second study, Berger et al. (2010b), focusing on Chinese banks during the period 1996-2006,¯nd that di®erent forms of diversi¯cation are steadily associated with lower pro¯ts and higher costs. Sanya & Wolfe (2011), studying 226 listed banks across 11 emerging economies, provide evidence that diversi¯cation across and within both interest and non-interest income sources reduces insolvency risk and improves pro¯tability. Focusing on four South Asian banking markets (Bangladesh, India, Pakistan and Sri Lanka) during the period 1998-2008, Nguyen et al. (2012) argue that banks become more stable through diversi¯cation across both interest and non-interest income activities. Comparing Islamic and conventional banks, Paltrinieri et al. (2020) show that diversi¯cation provides lower rewards for Shariah-compliant banks than conventional ones.
Considering cross-countries' studies, Roengpitya et al. (2017), investigating a panel of annual data relative to 178 banks from 34 countries for the period 2005-2015, provide evidence that commercial banking models exhibit more stable pro¯tability than trading, and banks switching to retail-funding see their return on equity (ROE) improve by 2.5% on average, relative to non-switchers. Guerry & Wallmeier (2017), assessing the e®ects of diversi¯cation on bank evaluation unveil that the diversi¯cation discount decreases over time and vanishes after the¯nancial crisis, while Kim et al. (2020)¯nd that a moderate degree of bank diversi¯cation increases bank stability, but excessive diversi¯cation has an adverse e®ect.
Causes and consequences of diversi¯cation. A related stream of literature examines causes and consequences of the e®ect that revenue diversi¯cation has on pro¯tability and risks. Chiorazzo et al. (2008), investigating Italian banks during the period 1993-2003, nd that income diversi¯cation increases risk-adjusted returns, the association is stronger at large banks, but limits to diversi¯cation gains exist as banks get larger. Studying US credit unions for the period 1993period -2004period , Goddard et al. (2008¯nd that similar diversi¯cation strategies are not appropriate for large and small credit unions. De Jonghe (2010) argues that since diversifying¯nancial activities in one \umbrella" institution does not improve the stability of the banking system,¯nancial conglomerates usually trade at a discount. According to this statement, Laeven & Levine (2007)¯nd that there is a diversi¯cation discount for¯nancial conglomerates that engage in multiple activities. Elsas et al. (2010), using a panel data from nine countries over the period 1996-2008,¯nd robust evidence against a conglomerate discount: diversi¯cation increases bank pro¯tability and the resulting market valuation.
Overall, the rich literature in this¯eld corroborates the idea that revenue di-versi¯cation is not necessarily bene¯cial, that its strength may depend on other¯rmor environment-speci¯c variables, and that resulting gains on pro¯tability and risk are far from being guaranteed. Stiroh & Rumble (2006) elegantly introduce the concept of the \dark side" of diversi¯cation by arguing that volatile patterns in noninterest income o®set the bene¯ts at the portfolio level: the (adverse) variance e®ect may counterbalance the (positive) correlation e®ect. Under this assumption, the net in°uence of revenue diversi¯cation on bank performances is ambiguous.
In line with this literature, we develop the following hypotheses to be tested. We expect that an increase in the share of non-interest income, in years characterized by market turmoil, is associated to negative performance measures (nominal and risk-adjusted) and to increases in their volatility. Instead, a revenue diversity measure able to capture di®erent potential directions of diversi¯cation and concentration strategies should show the opposite behavior: an increase of this variable should be associated with an improved performance and a reduced volatility. However, we also expect to¯nd a very weak signi¯cance of these two variables once other¯rm-level covariates are included as control variables, as well as a high degree of diversity once comparing di®erent banks within di®erent banking systems.
H1. Non-interest income is associated with a poorer nominal or risk-adjusted performance and an increase in its volatility.
H2. Diversi¯cation is associated with improvements in the nominal or risk-adjusted performance and a decrease in its volatility.
H3. Both the share of non-interest income and the level of diversi¯cation are weakly signi¯cant in explaining performance and its volatility.
H4. The ability of non-interest income and diversi¯cation to explain changes in performance and its stability varies widely across banks and banking systems.
With reference to typical¯rm characteristics, we expect a negative association with performance and a positive association with its stability when considering the quality of the loan portfolio, the weight of traditional lending activities for each bank and the level of cost e±ciency.

De¯nition of variables
Since we are interested both in the level and volatility of bank pro¯tability, we use seven di®erent dependent variables that are widely adopted in the banking literature (Table 1).
Typically, ROAE is more volatile than ROAA and is more in°uenced by bank leverage: we use both variables to cross-check from di®erent points of view the e®ects of revenue diversi¯cation on bank pro¯tability.
The expected positive portfolio-e®ect traditionally attributed to income diversication can be further investigated through volatility and risk-adjusted measures: consistently with existing literature (Stiroh & Rumble 2006, Mercieca et al. 2007, Goddard et al. 2008, we use the standard deviation and risk-adjusted versions for ROAA and ROAE and the Z-Score. According to Stiroh & Rumble (2006), we calculate two di®erent variables to account for the level of income diversi¯cation: NONsh and DIV. NONsh measures the share of net operating income represented by non-interest revenue (i.e. net trading incomes, net fees and commissions incomes, net insurance incomes, other non-interest incomes). Low levels of NONsh suggest the prevalence of traditional banking activities (borrowing and lending), typical for commercial banks. In this sense, a greater share of NONsh signals an income diversi¯cation strategy; however, this source of revenue may prevail also in other business models (for example, in corporate banking).
The second measure, DIV, accounts for this issue. This variable is built according to the Her¯ndahl-Hirschman Index approach; it measures for each bank the overall level of revenue diversi¯cation within the net operating income and it is calculated as follows (Eq. (3.1)): DIV ¼ 1 À ½ðNONshÞ 2 þ ð1 À NONshÞ 2 : ð3:1Þ By construction, DIV assumes values between 0 and 0.5; the minimum value is associated with banks that exhibit a single source of operating revenues (i.e. maximum concentration). The maximum value of the variable is reached when there is an equal contribution of interest and non-interest revenues in total operating income (i.e. maximum diversi¯cation). Consider two banks with a level of NONsh equal to 0.2 and 0.8, respectively. According to Eq. (3.1), DIV is the same for both banks and equals to 0.32. This means that, from a diversi¯cation point of view, despite these banks show the same value for DIV, their mix of revenue sources is di®erent. This is the reason why both NONsh and DIV are important for the estimation process and the interpretation of ndings. Naturally, as observed by Stiroh & Rumble (2006), these covariates are correlated; however, since DIV is a quadratic transformation of NONsh, the use of both variables in a single estimation follows a mainstream behavior in literature.
Our set of independent variables includes also other information from banks' nancials, with their expected signs disclosed also in Table 1. Firstly, we account for size using the natural logarithm of total assets (TA): through this variable we control \size e®ect" on pro¯tability and earnings volatility. To control leverage e®ects, we include the ratio between the tangible equity and total assets (Equity/TA): typically, higher levels of this variable signal a greater resilience capacity of the bank in troubled periods. Moreover, since we investigate a period characterized by a severe credit crisis, we observe the orientation towards lending through the ratio between loans and total assets (Loans/TA).
To control banks' e±ciency and the quality of the credit portfolio, we include also the cost-income ratio (Cost Income) and the ratio between the loan loss provision and loans (LLP), the latter in lagged form in order to reduce endogeneity issues. Finally, we include the level and squared value of asset growth to account for the annual (non-linear) variation in bank size (Asset growth and Asset growth 2 ).

Data
All data are obtained from the SNL Financial database, which includes a wide range of bank¯nancial information. We focus on banks from Europe and the US for the period Q1 2008-Q4 2016 (36 quarters). Since we need quarterly data to calculate pro¯t volatility for each year, we include in our sample only listed banks.
Saving banks, thrifts and mutual banks are excluded from the sample due to their peculiar asset-liability composition. All data are converted in Euro, but potential e®ects linked to exchange rates are captured using country and year dummies.

Uncertain Uncertain
Notes: This table summarizes and de¯nes the variables used in our analysis. Independent variables are end-of-year¯gures, with the exception of Asset growth, which is the arithmetical growth rate of assets between year t À 1 and year t.

S. Rossi, et al.
We drop banks that show NONsh values outside the [0;1] range. Yearly observations based only on data from one quarter are excluded (to avoid distortion on volatility measures), as well as banks with less than eight available quarters of data.
Pro¯tability measures (ROAA and ROAE) are built as the average of available quarterly data for a speci¯c year: this is in line with the end-of-year value in the database. However, quarterly data are necessary to build a measure of pro¯t volatility for each year. In our dataset, this measure corresponds to the standard deviation of ROAA and ROAE. Since a single quarterly data produces a null standard deviation, we impose a¯lter on data excluding years without at least two quarters of available data.
The remaining variables are end-of-year¯gures, whereas Asset growth is calculated as the arithmetic growth rate of assets between year t and year t À 1.
The outcome of this selection process is a dataset that includes 1250 banks, with over 95% of the sample from the US. In order to provide a more balanced view of the results, we use a cut-o® of 3 billion Euros to split US banks into \bigger banks" and \smaller banks". This threshold corresponds to the minimum value of Total Assets observed in the European bank sample. Therefore, since all the European banks are over the cut-o®, the \bigger US banks" sample is directly comparable with the European one. Unfortunately, we do not have a sample of smaller European banks, since in this area the average size of listed banks is greater than in the US. For the Notes: This table summarizes our sample by geographical area. In order to split US banks into \bigger banks" and \smaller banks", we use a cut-o® of 3 billion Euros, a threshold corresponding to the minimum value of Total Assets observed in the European sample. This allows us to compare the \bigger US banks" sample with the European one; furthermore, we can compare smaller and bigger US banks.

Does Revenue Diversi¯cation Still Matter in Banking?
sake of our analysis, this process allows us to compare large US and EU institutions, and large and small US institutions, that we consider consistent with the purpose of testing the impact of diversi¯cation for di®erent market-and¯rm-speci¯c backgrounds.
The¯nal sample consists of 60 European banks, 134 larger US banks and 1056 smaller US banks. The composition of our panel is outlined in Table 2.
Given the period of signi¯cant instability covered by our dataset, several variables show extreme values. We manage this issue through a winsorizing process (2.5% on each tail of the whole sample of available observations).
Tables 3 presents the summary statistics of our variables for three di®erent subsamples, averaged over the whole period under investigation.

2050003-10
Our preliminary result shows that European banks are less pro¯table than the corresponding US ones, have a higher exposure to non-interest revenues and are more diversi¯ed.
The behavior of our target variables (mean values) for our three subsamples (European banks, large US banks, small US banks) are provided in Fig. 1.
Performance measure (ROAA and ROAE), risk measures (standard deviation of ROAA and ROAE), risk-adjusted measures (RAROAA, RAROAE and Z-Score) and diversi¯cation measures (NONsh and DIV), respectively, show similar average trends in the period under investigation but provide some interesting changes occurred in the 2010-2011 years.
EU banks are associated with higher returns, lower standard deviation and comparable risk-adjusted measures until 2010. From 2011 and afterwards, the trend changed signi¯cantly. EU banks exhibit worse returns, higher standard deviations and worse risk-adjusted performances, with this trend remaining consistent until the end of our investigated period. In terms of diversi¯cation measures, however, the behavior of our three subsamples remains relatively stable, with EU banks increasing signi¯cantly especially in terms of NONsh.

Econometric estimations
In order to explore both the within and the between dimensions of our dataset, we employ four di®erent set-ups in our econometric estimation.  (2006), we start by calculating the mean value of each variable over the time span of our dataset: this allows us to run an OLS regression that explores the cross-sectional nature of the data.
Then, we use a dynamic¯xed-e®ect panel regression on the original data to explore the time dimension of our panel.
Finally, in order to compare the within and between dynamics of the panel data, we run two static panel regressions using, respectively, within and between estimators.

Cross-sectional analysis
Equation (3.2) reports the baseline OLS model used to estimate the cross-section e®ects of revenue diversi¯cation on banks' pro¯tability and risk-return measures.
where X i is a vector of bank-speci¯c information, c is the intercept and " i is the error term. All the variables are averaged over the whole period under investigation; country dummies are included.

Panel analysis
Equation (3.3) reports the baseline dynamic panel model used to estimate the e®ects of revenue diversi¯cation on banks' risk-return measures. We focus on these variables because they are more appropriate for evaluating bank performance, since they directly account for both the pro¯tability and the riskiness (measured as the volatility of pro¯ts). ð3:3Þ In this set-up, pro¯tability variables are averaged over each year; annual standard deviation measures for ROAA and ROAE express their volatility across the available quarters of a speci¯c year. Covariates are end-of-year¯gures. All regressions include interacted year-country dummies: in static panel estimations, the autoregressive term is omitted. Hausman tests suggest the use of¯xed-e®ect estimators against a random-e®ect speci¯cation; we also include regressions using between estimators in order to explore di®erent behaviors of within and between components of the panel.

Robustness checks
In order to check the robustness of our results, we perform several additional estimations, omitted for space constraints but available from authors upon request.
First of all, we test alternative regressions on original data, winsorizing on each subsample and using di®erent winsorizing approaches (for example, 1% on each tail). Our results are not a®ected.
We also try di®erent variable speci¯cations, in particular using the standard deviation of ROAA and ROAE calculated on the whole period instead of on a yearly basis. Again, our results do not change.
Third, since e±ciency and credit quality deterioration (namely, the cost-income ratio and loan loss provisions) are potential components of the dependent variables, we estimated all the regressions excluding these covariates. Once more, results are con¯rmed.
Because of NONsh varying signi¯cantly in our panel data analysis, we perform a full set of estimations comparing banks with a low level of NONsh with banks showing high level of this covariate, based on the sample median value of NONsh (0.30 for bigger banks and 0.17 for smaller banks). We¯nd signi¯cant di®erences arising only in Europe and only for high values, with a negative association with RAROAE, RAROAA and the Z-Score, and a positive association with ROAE, ROAA and their respective standard deviations.
Finally, since endogeneity concerns are crucial in estimations exploring pro¯tability and diversi¯cation strategies, we use two-step system GMM models to examine the robustness of our results also in this direction. Once again, coe±cients sign and statistical signi¯cance are in line with our main results presented as follows.

Cross-sectional analysis
We conduct the¯rst part of our analysis as follows: we¯rstly consider EU and bigger US banks, where comparability should be greater (Table 4), and then replicate the analysis for smaller US banks. Table 4 presents the results for European and bigger US banks in terms of both level and volatility of pro¯tability.
A general lack of statistical signi¯cance of NONsh emerges, across all regressions. DIV, on the other hand, shows a weakly signi¯cant negative e®ect on the level of ROAA for the EU sample and a statistically signi¯cant negative e®ect on pro¯ts volatility for larger US banks. In general, coe±cients deriving from pro¯t volatility estimations are in line with our hypotheses H1 (NONsh positive for volatility) and H2 (DIV negative for volatility), but statistical signi¯cance is more closely related The natural log of total assets (SIZE), the tangible equity to total assets (Equity/TA), loans to total assets (Loans/TA), the cost to income ratio (Cost income), lagged loan loss provisions (LLP), Annual growth of assets (Asset growth) and its square (Asset Growth 2 ) are the bank-speci¯c control variables. All regressions include country dummies. Robust standard errors are in parentheses. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively. with the expectations under H3 (NONsh and DIV are weakly and rarely signi¯cant) and H4 (material variability across banking systems). Performance level estimations provide mixed results.
Size is associated with negative but not statistically signi¯cant coe±cients: since we are here comparing banks of similar size, the explanatory power of this variable is reduced.
Leverage and loans share exhibit a mild e®ect on the dependent variables. It emerges that a greater loan share depressed bank pro¯tability during the period under examination: this outcome is not surprising given the speci¯c features of the recent crisis.
A clearer role is played by e±ciency and loan quality: both variables (Cost Income and LLP) are associated with highly signi¯cant coe±cients that are negative for pro¯tability levels and positive for volatility. As expected, lower levels of e±ciency and higher deterioration of credit portfolio quality depress pro¯tability and increase pro¯t volatility. This e®ect is likely to be even stronger in a period of falling margins.
Asset growth is associated with positive coe±cients in pro¯t level regressions and negative ones in pro¯t volatility estimations; the quadratic term shows an opposite sign, indicating that, as expected, a faster growth can generate more instability in the risk-return pro¯le. Table 5 shows the results for risk-adjusted performance measures. We¯nd that NONsh is constantly associated with negative coe±cients. Instead, DIV shows positive coe±cients in the US sample, while the opposite happens for the European one. Statistical signi¯cance is scattered, while it is more common in larger US banks regressions; compared with Table 4 we observe more statistically signi¯cant coe±cients. This is due to the combination of level and volatility of pro¯ts, which contribute to the computation of the dependent variables used in these estimations. Risk-adjusted measures provide a more insightful picture on pro¯tability.
Results are therefore supportive of our four main hypotheses. Overall, outcomes in Table 5 are also consistent with the literature (Stiroh & Rumble 2006): a contrast exists between diversi¯cation bene¯ts and the potential adverse e®ect linked to more volatile non-interest revenues.
Tangible equity and loans share exhibit mainly negative coe±cients. Cost income and loan loss provisions are associated with negative and strongly signi¯cant coe±cients, underlining the relevance of these two variables during the recent crisis period. Table 6 includes the results of econometric estimation for smaller US banks. Overall, results con¯rm the previous analysis and are in line with all our hypotheses. NONsh increases the volatility of banks' pro¯ts and is associated with negative coe±cients in risk-adjusted pro¯t regressions. DIV, on the contrary, reduces volatility and gives bene¯ts to risk-adjusted performance measures. In this subsample, characterized by a greater variability in the size of banks, TA shows positive and statistically signi¯cant coe±cients for risk-adjusted performance measures: larger banks hence bene¯t in terms of RAROAA, RAROAE and Z-Score.
In these estimations, negative and signi¯cant coe±cients are associated with leverage, loans share, cost-income ratios and loan loss provisions. The coe±cient for Equity/TA in the last column is not surprising: the variable enters the equation used to calculate Z-Score with a positive value.
Asset growth should promote pro¯tability, but the e®ect on its volatility may be less immediate. In our¯ndings, the coe±cients associated with the variable in riskadjusted performance regressions are not statistically signi¯cant: apparently, growth may be bene¯cial for some entities and increase volatility for others, without an easily predictable outcome. Notes: This table presents the impact of diversi¯cation on pro¯tability and stability measures using OLS estimations for the EU and larger US banks samples. Variables are averaged over the whole period. Bank pro¯tability measures are the risk-adjusted return on average assets (RAROAA), the risk-adjusted return on average equity (RAROAE) and Z-Score. Non-interest income (NONsh) and DIV are income diver-si¯cation variables. The natural log of total assets (SIZE), the tangible equity to total assets (Equity/TA), loans to total assets (Loans/TA), the cost to income ratio (Cost income), lagged loan loss provisions (LLP), Annual growth of assets (Asset growth) and its square (Asset Growth 2 ) are the bank-speci¯c control variables. All regressions include country dummies. Robust standard errors are in parentheses. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively.
Since in the econometric estimation, NONsh and DIV are considered as explanatory variables but they share a common root, it is worth examining their joint e®ect on dependent variables. More speci¯cally, when the share of non-interest revenues changes (for example, it increases its level), two di®erent e®ects occur. The¯rst one is simply a greater exposure to this source of revenues: the outcome of this event can be read directly observing the coe±cients associated to NONsh. The second one is linked to the income diversi¯cation level and requires a speci¯c explanation.
Since interest and non-interest shares sum to one, the equation for calculating DIV can be written as follows (Eq. (4.1)): ð4:1Þ Notes: This table presents the impact of diversi¯cation on pro¯tability and stability measures using OLS estimations for the smaller US banks sample. Variables are averaged over the whole period. Dependent variables of these estimations are the level and volatility of ROAA and ROAE, the risk-adjusted return on average assets (RAROAA), the risk-adjusted return on average equity (RAROAE) and Z-Score. Noninterest income (NONsh) and DIV are income diversi¯cation variables. The natural log of total assets (SIZE), the tangible equity to total assets (Equity/TA), loans to total assets (Loans/TA), the cost to income ratio (Cost income), lagged loan loss provisions (LLP), Annual growth of assets (Asset growth) and its square (Asset Growth 2 ) are the bank-speci¯c control variables. All regressions include country dummies. Robust standard errors are in parentheses. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively.

Does Revenue Diversi¯cation Still Matter in Banking?
An increase in the level of NONsh has two consequences on DIV (a linear and a quadratic one) that are usually referred as direct and indirect e®ects in the literature (Stiroh & Rumble 2006). The total or \net" in°uence of a change in NONsh on the dependent variable can be calculated as the sum of these two e®ects.
We investigate this issue for our risk-adjusted measures (RAROAA, RAROAE, Z-Score), by evaluating the partial and total e®ect of a 1% increase of the noninterest share of revenues for di®erent percentiles of NONsh for each subsample of banks. This allows us to explore the e®ect of a change in NONsh for banks that show di®erent compositions of revenues: in fact, we can expect that the bene¯ts stemming from diversi¯cation strategies change for di®erent levels of NONsh.
A very easy way to understand this statement is considering two banks that have a level of NONsh equal to 0.4 and 0.6 (40% and 60%, respectively). An increase in NONsh means a gain in diversi¯cation for the¯rst bank, but more concentration for the second one. DIV is initially equal to 0.48 for both banks; however, after the change observed in NONsh, its level is equal to 0.4838 for the¯rst bank and 0.4578 for the second.
We consider the 10th, 25th, 50th, 75th and 90th percentile of NONsh for each group of banks. Results for RAROAA are presented in Table 7.
The most noticeable result is the recurring negative sign of the net e®ect for EU banks: both direct and indirect e®ects go in this direction. This outcome is indirectly reinforced by the only positive coe±cient that can be found in the highest percentile column for the indirect e®ect: considering that for this cohort the average level of NONsh is 0.55, an increase in this variable means a concentration strategy and not a diversi¯cation one.
US banks À À À larger and smaller À À À show a common pattern of coe±cients: they are mainly negative for the direct e®ect and positive for the indirect e®ect. The total in°uence on RAROAA is mainly positive but turns to negative for the highest percentiles. This suggests that an increase in NONsh provides di®erent outcomes across diverse levels of non-interest income revenues (for instance, banks that are starting their diversi¯cation strategy or those that already express higher weights of non-interest income). Table 8 shows the direct, indirect and total e®ect of a change in NONsh on RAROAE.
Overall, results remain consistent. For European banks almost all coe±cients remain negative, while for US banks we have negative direct e®ects that are counterbalanced by positive indirect ones. The total e®ect is mainly positive for the latter sample, while statistical signi¯cance varies with the chosen percentile. Table 9 completes this analysis for the Z-Score.
Here, coe±cients for the indirect e®ect turn to positive for European banks, but the total e®ect remains negative. For bigger US banks, coe±cients are consistent with previous results. For smaller US banks, instead, positive e®ects of diversi¯cation are not su±cient to overcome the negative outcomes stemming from a greater exposure to non-interest revenues. This latter result is stable across all columns, while statistical signi¯cance is present only for higher percentiles.
Considered altogether, results from this alternative setting con¯rm again our hypotheses, in particular the expected weak signi¯cance of NONsh (H3) and the variability across banking systems and bank characteristics (H4).

Panel analysis
Our empirical analysis includes also several panel estimations. As for the previous analysis, we¯rst focus on EU and larger US banks, extending then the comparison to smaller US banks. Table 10 shows the results of dynamic estimates for the¯rst comparison.
Besides a weak autocorrelation of the dependent variables, the regressions conrm the previous analysis. NONsh and DIV are associated respectively with negative and positive coe±cients, while rarely being statistically signi¯cant, con¯rming all our hypotheses. Among the other explanatory variables, it is worth noting that cost-  Tables 5 and 6 ; indirect e®ect is calculated as the impact of diversi¯cation on the dependent variable, given a 1% increase in NONsh. Total e®ect is the sum of direct and indirect e®ect. Standard errors are reported in bold below the estimated coe±cients. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively.

Does Revenue Diversi¯cation Still Matter in Banking?
income and loan loss provisioning still show negative coe±cients, but these are statistically signi¯cant only in the US bank subsample. Asset growth has a negative e®ect on risk-return measures, while the opposite is true for the level of tangible equity. Table 11 compares within and between estimators for static panel analyses. Once more, NONsh and DIV take the expected sign (hypotheses H1 and H2). However, the statistical signi¯cance of the coe±cients is relatively low (hypothesis H3).
Ine±ciency and low loan quality adversely a®ect risk-return measures. The same is true for loans share, which exhibits signi¯cant coe±cients in RAROAA and RAROAE between-regressions for European banks. The changing sign of some coe±cients in within and between regressions (for example, those associated with asset growth) seems to indicate that individual e®ects have a larger impact in explaining the risk-adjusted pro¯tability.
This represents a relevant¯nding and is also con¯rmed by the results obtained on the smaller US banks subsample (Table 12) and once again supports our last hypothesis (H4).  Tables 5 and 6 In this case, coe±cients associated with DIV and Equity/TA change sign in within and between estimations. Since between estimators explore the cross-sectional nature of data, the last three columns are more similar to the outcome of the previous analysis on mean values (see Table 6). NONsh is still characterized by negative coe±cients and the same holds for loans share, Cost income and LLP.
All econometric estimations draw a picture in which there is not a clear-cut evidence of a relationship between income diversi¯cation and risk-adjusted performance of banks. This is especially true for the European banks, which exhibit weak statistical signi¯cance for the coe±cients associated to non-interest share of income and diversi¯cation level. Moreover, the joint e®ect of these variables, when signi¯cant, is negative. The implication of this outcome is that during the recent crisis period a greater balance of sources of revenue has not provided better risk-adjusted results for European banks.
Instead, for US banks, the opposite seems to be true. A greater exposure to noninterest income has usually a negative impact on risk-adjusted measures, but diversi¯cation e®ects are usually strong enough to counterbalance this e®ect. We¯nd evidence for this relationship in particular for smaller and less diversi¯ed banks. We see this as a signi¯cant contribution to the literature since it holds in a very di®erent economic environment. The seminal work by Stiroh & Rumble (2006) has been conducted in a period of less¯nancial turbulence compared to our analysis. In recent crisis and post-crisis years, both interest and non-interest revenues have Notes: This table presents the impact of diversi¯cation on risk-adjusted pro¯tability measures using dynamic¯xed-e®ects panel estimations for the EU and larger US banks samples. Bank and time¯xed e®ect are used. Bank pro¯tability measures are the risk-adjusted return on average assets (RAR-OAA), the risk-adjusted return on average equity (RAROAE) and Z-Score. Non-interest income (NONsh) and DIV are income diversi¯cation variables. The natural log of total assets (SIZE), the tangible equity to total assets (Equity/TA), loans to total assets (Loans/TA), the cost to income ratio (Cost income), lagged loan loss provisions (LLP), Annual growth of assets (Asset growth) and its square (Asset Growth 2 ) are the bank-speci¯c control variables. All regressions include interacted country-time dummies. Robust standard errors are in parentheses. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively.    Z-Score. Non-interest income (NONsh) and DIV are income diversi¯cation variables. The natural log of total assets (SIZE), the tangible equity to total assets (Equity/TA), loans to total assets (Loans/TA), the cost to income ratio (Cost income), lagged loan loss provisions (LLP), Annual growth of assets (Asset growth) and its square (Asset Growth 2 ) are the bank-speci¯c control variables. All regressions include interacted country-time dummies. Robust standard errors are in parentheses. ***, **, * indicate statistical signi¯cance at the 1%, 5% and 10% level, respectively.
S. Rossi, et al.  Constant À27:00*** À23:34*** À123:23 experienced a high volatility, with uncertain e®ects on the bene¯cial portfolio-e®ect usually attributed to diversi¯cation. Our results suggest that for US banks these macroeconomic di®erences have not changed the relationship between NONsh, DIV and risk-adjusted performance measures.
Given its current modest relative size, it is unlikely that this result is driven by the emergence of alternative business models or competitors from the FinTech industry. Despite it was not the aim of this paper to speci¯cally test a di®erentiated approach from regulation and supervision in the EU and US markets, this is reasonably a signi¯cant contributing factor in explaining why responses vary so widely between the two banking systems. However, results from the between estimators as well as the high variability of statistical signi¯cance of covariates in our di®erent econometric approaches suggest another conclusion. Most part of the unreliability of diversi¯cation strategies in improving performance or providing stability lies in¯rm-speci¯c factors, rather than in market-wide or bigger exogenous shocks.

Policy implications
In our view, our results lead to signi¯cant policy implications.
On one side, we con¯rmed empirically that revenue diversi¯cation in banking is a complex matter. Its ability to enhance the level and the stability of performance is limited, since they seem more closely linked to¯rm-and market-speci¯c features. Moreover, the robustness of revenues diversity is especially questionable durinḡ nancial turmoil. The impact of non-interest revenues is signi¯cant and negative on bank risk for both US (large and small) and EU institutions. At the same time, it is signi¯cant and negative for risk-adjusted performance measures for larger US banks, whereas it is signi¯cant and positive for smaller US banks only for the ROAA.
We argue that diversi¯cation augments¯rm risk, but at the same time is unable to enhance pro¯tability to balance this e®ect or is even harmful (for larger US institutions). The bene¯t for smaller US banks may be attributed to their overall lower engagement in such activities, as well as a potential larger bene¯t for entities initiating non-interest-bearing operations. Studying the impact of changes in noninterest revenues in di®erent business models (de¯ned by selected percentiles of the related distribution) strengthens this claim.
Among other variables, those that bear most signi¯cance in both markets and regardless of the size of institutions are the cost e±ciency and the quality of the credit portfolio. Despite not further investigated in this paper but incidentally arising from our results, we argue that diversi¯cation, by combining human resources, capital and expertise, shows a short-term improvement of cost e±ciency and pro¯tability but, in the long run, venturing in non-interest revenues could result in a higher volatility of earnings.
Moreover, shifting e®orts from traditional to innovative banking activities may lower the attention, at least partially, to the quality of the loan portfolio.

Does Revenue Diversi¯cation Still Matter in Banking?
Alternatively, non-interest-bearing activities may be the result of the worsening quality of the loan portfolio, without evidence of a gain in terms of risk-adjusted performance.
The overall implication of our results is a call for greater scrutiny and care, for both banks and supervisors, before assuming that diversi¯cation could provide an easy path to restore or improve performance, without a®ecting risks, at the¯rm-level and in terms of¯nancial stability. This is particularly relevant if revenue diversity is a direct response to worsening macroeconomic conditions.

Conclusion
In this paper, we investigate the impact of revenue diversi¯cation on bank pro¯tability and its volatility. We examine the relationship between the degree of diver-si¯cation and several measures of risk-return pro¯les in a cross-country analysis, including 1250 listed EU and US banks from Q1-2008 to Q4-2016.
Through OLS regressions, dynamic¯xed-e®ect panel models and static panel regressions, we¯nd that diversi¯cation is not clearly associated with the level or quality of performance, that bene¯ts change over time and, where present, show signi¯cant variability. Moreover, we¯nd a di®erent behavior in US and EU banks in terms of diversi¯cation bene¯ts.
European banks do not show a material impact of non-interest revenues on both nominal and risk-adjusted performance measures, while for US banks e®ects are signi¯cant in terms of performance volatility and risk-adjusted pro¯tability, but substantially di®erent for smaller and larger institutions. Finally, we show that¯rm-speci¯c diversi¯cation strategies matter more than the overall sectoral pursuit of revenue diversity over time.
Our results provide additional evidence on the limitations of diversi¯cation in the banking sector, supporting signi¯cant policy implications. Supervisors should be careful in expecting that more revenue diversity, especially in Europe, bears necessarily bene¯ts for the banking system. The diversity and adequacy of business models to di®erent economic environments, rather than alternative revenue sources, seems to produce greater and persistent e®ects on bank pro¯tability and volatility.
In terms of future research, it could be useful to understand the impact of speci¯c exogenous (for instance changes in regulation or supervision, growth of the FinTech sector) as well as endogenous shock (such as human capital) on pro¯tability and stability, especially in terms of¯rm-level factors that determine a di®erent response from banks.