Provisioning, Bank Behavior and Financial Crisis: Evidence from GCC Banks

2.1 Loan Loss Provisions and Capital Management

There are several reasons why banks hold capital. First, capital levels minimize the probability of going bankrupt and consequently, the costs of failure. These include not only regulatory (e. g., restrictions on the current activities, future growth prospects, dividend payments) but also financial (e. g., it acts as a signal of creditworthiness and thereby lowers interest outgo) costs. Second, higher capital may bring advantages of future growth opportunities. For example, faced with a substantial increase in loan demand, banks with low levels of capital are more likely to lose their market share to well-capitalized ones (Jokipii and Milne 2008). Finally, higher capital levels might induce banks to screen borrowers more carefully, thereby raising the probability of repayment and in turn, lowering borrowing costs (Agenor, Alper, and de Silva 2012).

Altering the levels of capital entail adjustment costs. In addition to transaction costs, these include those related to informational asymmetries. Since the issuer is at an informational advantage as compared to potential buyers, stock issuance provides credible signal of the bank’s capital, which in turn, lowers the adjustment costs. In this context, the effect of income smoothing on regulatory capital might compel banks to condition smoothing to meeting capital requirements.

More recently, Bouvatier and Lepetit (2012) have provided an explanation for the role of capital buffers in loan loss provisioning. According to the authors, capital buffers cover for expected losses over and above those not covered by LLPs. In order to build such buffers, banks use their retained earnings to increase capitalization during economic upturns, thereby negating the impact of the business cycle on provisions.

2.2 Loan Loss Provisions and Earnings Management

The idea behind earnings management is that banks adjust their provisioning behavior in order to manage their capital ratios (Kim and Cross 1998; Beatty, Chamberlain, and Magliolo 1995). The advantage of such capital management is based on two considerations. First, it acts as a signal of high quality to outside investors because it indicates stability of the bank’s income (Barth 2001). When earnings are unusually low, banks can deliberately understate their LLPs or can release LLPs to offset operational losses; reverse is the case when earnings are unusually high. Several empirical studies have confirmed that such behavior is indeed observed in practice (Wahlen 1994; Laeven and Majnoni 2003; Liu and Ryan 2006; Bushman and Williams, 2012).

The second rationale behind LLPs is that higher provisioning when banks have low capital is suggestive of the fact that these two are substitutable buffers against potential losses. Although there are studies which provide evidence of such interlinkage (Ahmed, Takeda, and Thomas 1999; Bikker and Metzemakers 2005), there is also research that does not uncover any such discernible relationship (Bishop 1996; Collins, Shackelford, and Wahlen 1995).

Several studies have explored which bank characteristics matter for earnings management. Greenawalt and Sinkey (1988) and Bhat (1996) show that banks in the US use LLPs to smooth income. In contrast, Beatty, Chamberlain, and Magliolo (1995) find that US banks use miscellaneous gains to manage earnings. In case of Spanish banks,Perez, Salas, and Saurina (2006) uncover evidence that banks practice income smoothing, although evidence in favor of banks using LLPs to manage capital is less compelling.

Finally, Fonseca and Gonzalez (2008) show that banks in more developed financial systems are more inclined to smooth earnings. In particular, they show that income smoothing is negatively related to investor protection and accounting disclosure.

2.3 Loan Loss Provisions and Signaling

The use of LLPs as a signaling device has also attracted significant attention. Several studies find that bank managers use discretionary components of provisions to signal private information about the banks’ future prospects. By way of example, Beaver et al. (1989) contend that investors interpret an unexpected increase in LLPs as a signal of the bank’s financial strength. Yet others (Musumeci and Sinkey 1990; Griffin and Wallach 1991) find that capital markets respond positively to increases in LLPs, because it signals the growing creditworthiness of the bank. This evidence of a positive market reaction to unexpected increases in LLPs has also been echoed in several other studies (Liu, Ryan, and Wahlen 1997; Beaver and Engel 1996). This contrasts with the findings of Ahmed, Takeda, and Thomas (1999) and Anandarajan, Hasan, and McCarthy (2007) who do not find evidence in favor of signaling behavior by banks in their studies. More generally, Bushman and Williams (2012) show that market discipline over bank risk-taking is typically weaker in countries with less timely loss provisioning regimes, highlighting the adverse effects of manipulating financial statements by banks.

We build on these studies by examining the LLP policy of banks across an extended time period that includes the recent economic crisis. Besides income smoothing, we also examine the capital management and signaling hypotheses, an aspect not previously explored for GCC banks. Contextually, we also examine whether the provisioning behavior of Islamic banks differs from that of commercial banks.

5.1 Univariate Tests

Table 3 enlists the LLPs and loan loss reserves (LLRs) for these banks, classified according to ownership (Panel A), country (Panel B) and by crisis period (Panel C). From the table, it appears that Islamic banks have higher LLPs as compared to Islamic banks, although the difference is not statistically significant. LLRs of commercial banks, on the other hand, are statistically and significantly higher as compared to Islamic banks.

The country-level tabulations suggest significant differences in terms of LLRs, although the variation in terms of LLPs is not so compelling. Banks in Oman and UAE are observed to have the highest levels of LLRs, whereas banks in Bahrain have the lowest LLRs, on average, reflecting in part, the economic weaknesses in the aftermath of the crisis and the Arab spring. LLRs for banks in Bahrain, for example, average around 3.0% of assets are compared to 4.8% for Omani banks. This difference is statistically significant at the 0.01 level. In several other instances and especially for Omani banks, these differences are statistically significant.

Banks also improved their LLPs, especially during the crisis. As Table 3 reveals, LLPs were roughly 0.49% of banks’ assets during non-crisis times. During the crisis, this improved to 0.57%. The difference vis-à-vis non-crisis period is statistically significant. On the other hand, LLRs appeared to have declined during the crisis, consistent with the fact that a rise in NPLs in the region more generally has impelled banks to draw down on their reserves as provisions against delinquent loans.

The correlation matrix in Table 4 indicates that bigger and profitable banks tend to have lower levels LLRs. At the macroeconomic level, higher GDP growth prompts banks to cut back on their LLPs, suggestive of countercyclical behavior. These raw correlations however, do not take on board either bank-specific or business cycle considerations.

The univariate results discussed earlier provide evidence in support of differences across country-bank pairs in terms of their LLRs (and to an extent, LLPs as well). The subsequent correlation analysis also points in a similar direction. As well, the results are suggestive of a countercyclical approach on part of banks in their provisioning behavior. These tests however, do not consider several bank-specific variables. For instance, provisioning of large banks could be different from those which are relatively small by size. The macroeconomic environment could also be a relevant consideration. Taking these observations into account, we employ a multivariate regression framework to test the several hypotheses relating to capital management, earnings smoothing and signaling.

5.2 Testing for Capital and Earnings Management

To test for these hypotheses, the empirical specification for bank s in country j at time t is assumed to be of the following form:

where the φ s’ are the parameters to be estimated.

5.3 Dependent Variable

Following from the literature, we employ two measures of the dependent variable (Depvar), depending on whether the information on provisions has been extracted from banks’ balance sheet or the income statement.

Consistent with prior research (Ahmed, Takeda, and Thomas 1999; Laeven and Majnoni 2003; Bikker and Metzemakers 2005; Bouvatier and Lepetit 2008), the first dependent variable is the ratio of loan loss provisions to total asset (LLP).

Following Grammatikos and Saunders (1990) and Walter (1991), the alternate dependent variable is the ratio of loan loss reserves divided by bank asset (LLR). This variable represents the global amount of provision for loan losses built up by the bank.

In other words, a LLP is a “shock absorber” to offset expected future losses (Laeven and Majnoni 2003), whereas a LLR is the amount of principal that is not expected to be recovered.

5.4 Independent Variables

The evolution of provisions is linked to the past due age of non-performing loans. Therefore, it seems reasonable to expect that LLPs will be correlated across years. As a result, we include the lagged-dependent variable (LDV).

The variable Equity/Asset identifies whether bank capital plays a role in influencing provisions. It is well-acknowledged that banks try to manage regulatory capital upwards via abnormal LLPs that increase LLPs, so as to convey an impression of financial health to regulators. Provided that tier-II capital is augmented via opportunistic increases in LLPs, a negative coefficient will imply capital management.

The ratio of earnings before taxes and provisions to total assets (EBTP) tests for income smoothing (Ahmed, Takeda, and Thomas 1999). If banks have earnings lower (resp., higher) than their target value, they will tend to lower (resp., raise) LLPs to stabilize them. In other words, a positive coefficient on EBTP would lend support to this hypothesis. The interaction variable – Islamic*EBTP – is incorporated to detect the differential earnings stabilization policy for Islamic banks. Cross national research by Taktak, Zouari, and Boudriga (2010) posits a positive coefficient on this variable.

Following from the literature (Hasan and Wall 2004), we control for the non-discretionary component of LLPs by including the non-performing loan ratio (NPLs).

Change in loans can be considered as a proxy for general provisions. The influence of this variable on LLPs depends largely on the quality of incremental loans (Lobo and Yang 2001).

Finally, we control for bank size by including the logarithm of bank asset (LTA). A priori, one would expect a positive sign on this variable, based on the conjecture that bigger banks have larger business volumes and consequently, more likely to have higher LLPs (Anandarajan, Hasan, and McCarthy 2007).

At the industry level, we control for overall financial development (proxied by bank credit to GDP), with an expected positive sign (Fonseca and Gonzalez 2008). We also include a proxy for concentration (measured as the asset share of three largest banks). The effect of concentration on LLPs is not clear, a priori. On the one hand, the “concentration-fragility” view would suggest that increase in banking concentration tends to engender high banking fragility (De Nicolo et al. 2004). On the other hand, the “concentration-stability” view contends that an increase in banking concentration does not result in higher banking fragility (Beck, Demirguc Kunt, and Levine 2006a, 2006b).

At the macroeconomic level, we include real non-oil GDP growth (GDPGR). ^[2] Since repayments are higher in upturns, banks need to provide less for loan losses, entailing a positive sign on this variable. Besides, we capture the stance of monetary policy by including the real interest rate (RoI). Higher rates of interest raise funding costs for borrowers, in turn, raising the repayment burden, and thereby, compelling banks to raise provisions. As a result, the expected sign on this variable is positive. The variable ν captures bank characteristics that are unobservable and finally, ε is the error term.

We also include several dummy variables. The first dummy – Crisis – takes into account the recent financial crisis. As our prior discussion observes, banks’ NPLs have witnessed an upturn in the aftermath of the crisis, which would suggest the need for higher provisioning, with an expected positive sign on this variable. Second, we include a dummy which equals one if a bank is Islamic. To the extent that Islamic banks have been impacted relatively less by the crisis as compared to commercial banks (Akhtar Aziz 2009), it seems likely that they would need to make lower provisions. We also include country dummies to control for unobservable, country-specific factors such as the level of development, geography and institutions.

5.5 Testing for Signaling Hypothesis

In order to examine the signaling model, we include the one-year ahead change in earnings before taxes and provisions (ΔEBTP_s,t+1) in the regression equation. Besides, we rationalize the equation by excluding NPLs and change in loans from the set of independent variables. Accordingly, the revised empirical specification for bank s in country j at time t reads as follows:

where the notations are the same as earlier and ηs’ are the parameters to be estimated.

As per the signaling hypothesis, discretionary changes in LLPs are positively correlated with changes in future earnings. This would suggest that, for the signaling hypothesis to be valid, the coefficient on η₄ would be positive, consistent with previous studies (Ahmed, Takeda, and Thomas 1999; Anandarajan, Hasan, and McCarthy 2007).

6.1 Results for the Capital Management and Income Smoothing Hypotheses

The inclusion lagged dependent variable would suggest that the standard OLS estimation process is not consistent. To circumvent this problem, we resort to the Generalized Method of Moments (GMM) estimator.

The reliability of the GMM estimation procedure depends on the validity of the instruments. To do this, we present the Sargan test of over-identifying restrictions, which asymptotically follows a χ2 distribution. The higher the p-value, the more likely it is that the chosen instruments are valid.

In addition, the estimates also need to ensure the absence of serial correlation in the error terms. To do this, we present tests for first-order (AR1) and second-order (AR2) serial correlation. Statistically, a low p-value for AR1 and a high p-value for AR2 would suggest that the disturbances are not serially correlated.

The regression results are presented in Table 5. The Sargan test statistic provides strong evidence in favor of validity of the instruments. Similarly, the reported AR values suggest the absence of second-order autocorrelation in the residuals. Taken together, these tests support the fact that the model specification is appropriate.

The lagged dependent variable (LDV) is positive and significant at the 1% level. This indicates that LLPs display high levels of persistence, conforming to a priori expectations (Bouvatier and Lepetit 2008; Fonseca and Gonzalez 2008).

We first briefly discuss the bank-level controls and banking industry variables. The coefficient on bank size is negative and significant in the LLR equation (Models 1–5), but does not appear to be significant in the LLP results (Models 6–10). In other words, a 10% increase in total assets lowers LLRs by roughly 0.3%. This implies that bigger (and hence, more diversified) banks are not able to adequately reduce risk exposures as compared to smaller banks. NPLs bear a positive sign across all LLP specifications, which suggest that banks provisions are higher when their loan losses increase. On the other hand, LRR bears a uniformly negative coefficient in all models, which suggests that banks tend to draw down reserves when NPLs rise. The coefficient on loan growth is negative and strongly significant, reflective of the fact that risks tend to accumulate during economic booms (Borio et al. 2001; Lowe 2003; Bikker and Metzemakers 2005). A 10% increase in loans lowers LLRs by roughly 0.4%; LLPs likewise, decline by 0.01–0.06%.

At the banking industry level, the coefficient on concentration is positive in the LLP specification, supportive of the “concentration-stability” views of Beck et al. (2006a). The coefficient on financial development is positive In LLR specification, as expected, but negative in the LLP model, indicating that greater penetration of finance impels banks to cut back provisions. The positive sign on the interest rate variable is consistent with the “financial instability hypothesis” (Minsky 1975), which suggests that higher short-term interest rates increases borrower indebtedness and exacerbates financial fragility thereby prompting an increase in loan losses.

The first variable of interest – Equity/Asset – bears a negative association with provisions. This is in line with the capital management hypothesis which predicts higher provisioning when the capital ratio is low. The results are akin to those reported for US banks (Beatty, Chamberlain, and Magliolo 1995; Ahmed, Takeda, and Thomas 1999). The marginal effect of capital on provisioning and is –0.38, similar to the magnitudes obtained in previous studies (Bikker and Metzemakers 2005).

The second variable of import – EBTP – has a negative and significant coefficient in the LLP specification (Model 6). The elasticity of LLPs with respect to EBTP, evaluated at the mean values, equals –0.43. In other words, a proportionate increase in EBTP by 1% would induce a proportionate decline in LLPs by roughly 0.4%. These results concur with Laeven and Majnoni (2003) for Asian banks. More recently cross-country studies also find similar results for certain countries (Fonseca and Gonzalez 2008; Packer and Zhu 2012). ^[3]

The GDPGR variable is negative and significant. This confirms the existence of a strong cyclicality of provisions in the GCC banking system, indicating that provisions rise when economic growth is weak. The elasticity of LLPs with respect to GDP growth varies across countries, ranging from –0.07 for Saudi Arabia to a high of –0.16 in case of Qatar: an increase in GDP by 10% would lower LLPs by 0.07–0.16%. This evidence is consistent with pro-cyclical behavior by banks, as already reported in prior empirical research (Fonseca and Gonzalez 2008; Bikker and Metzemakers 2005; Laeven and Majnoni 2003).

6.2 Capital Management and Income Smoothing: Robustness

We extend the baseline regression to examine certain additional hypotheses. Following from the literature, we focus on the results with LLPs as the dependent variable; the results pertaining to LLRs are qualitatively similar.

First, we examine whether Islamic banks tend to exhibit higher provisions in an upswing as compared to commercial banks. In Model 2, the coefficient on GDPGR*ISLAMIC is negative and significant with a point estimate equal to 0.021. To consider the overall impact, consider a country with a GDP growth equal to 6%, the median growth in the sample. Temporarily ignoring the Islamic bank effect, this would suggest that a rise in GDP by 10% would lower LLPs by 0.01%. Taking into consideration the Islamic bank, the point estimates in Model (2) provide a magnitude roughly 0.032% points. In other words, Islamic banks appear to make lower provisioning in upswings.

In Model (3), we examine the earnings smoothing hypothesis for Islamic banks. The estimates reveal that the interactive term has a magnitude equal to 0.041, which is positive and statistically significant. In other words, Islamic banks do not appear to be resorting to earnings smoothing, contrary to extant research on this aspect (Taktak, Zouari, and Boudriga 2010).

Model 4 examines the impact of the recent economic crisis on the provisioning behavior of Islamic banks. If Islamic banks were relatively more affected by the crisis and therefore lowered their LLPs, one would expect a negative coefficient. The results appear to suggest no perceptible impact of the crisis on the Islamic banks.

The final specification examines the capital management hypothesis for Islamic banks. As observed from Model (5), the coefficient on the interactive term is positive with a point estimate equal to 0.126. The overall impact is 0.027, implying that Islamic banks are less engaged in capital management as compared to commercial banks.

In sum, the results provide support for the fact that GCC banks engage in both capital management and income smoothing.

6.3 Results for the Signaling Model

Support in favor of the signaling theory would suggest that the coefficient on ΔEBTP variable would be positive (Wahlen 1994; Anandarajan, Hasan, and McCarthy 2007). The evidence strongly rejects the signaling hypothesis: in Model 1, the coefficient on ΔEBTP is negative and strongly significant with a point estimate equal to –0.078: an increase in LLPs, at time t, is associated with a lower one-year ahead reported earnings (Table 6).

6.4 Signaling Model: Robustness

As earlier, we conduct several additional checks of the baseline results. First, we examine the signaling behavior of Islamic banks. As observed from Model (2), the coefficient on the interactive term – ΔEBTA*Islamic – is negative and significant at the 0.01 level. Taken together with the coefficient on ΔEBTA, this implies that Islamic banks engage less in signaling behavior as compared to non-Islamic banks.

Second, we test whether the signaling behavior of Islamic banks is more manifest in the upswing of the business cycle. Once again, the evidence in Model 3 refutes this hypothesis, although overall, there is evidence in support of signaling.

The final model explores whether Islamic banks had resorted to signaling behavior during the crisis. Here again, there is limited evidence in support of this claim.

Summing up, the results provide limited support to the signaling hypothesis, although there is some evidence of this fact during the upswing of the business cycle.