Anatomy of Bank Distress

The Information Content of Accounting Fundamentals

By Janko Cizel / Edward Altman / Herbert Rijken

Plan

The paper
The additive value of the Risk-Weighted Framework in the prediction of bank distress
LR-Dashboard

How reliable is bank accounting disclosure?

Framework

Accounting discretion $\Longrightarrow$ ability to misrepresent true performance

Incentives to inflate true performance, especially in the state of financial distress.

Consequence: "jamming" of the accounting signal.

In an extreme case, the accounting signal may lose its ability to discriminate between healthy and unhealthy institutions.

The way we think about this question is the following.

First, in practice, banks possess substential discretion over their reported accounting performance. As an example of this, one can think of the banks' flexibility in the domains of loan-loss provisioning, and the calculation of asset risk weights, to name a few. The accounting discretion gives banks the ability to misrepresent their true performance.

Apart from their capacity to mis-represent performance, a plausible case can be made that under certain conditions, banks also have the incentives to mis-report their true performance. Specifically, a bank close to distress may use accounting discretion to improve its reported regulatory capital ratio in order to avoid negative attention of its regulator, or to avoid a run on its funding.

Now, assuming that banks in distress are systematically more likely to use accounting discretion than their healthier peers, one obtains a classical case of so-called signal jamming: namely, that the observed accounting signal loses its ability to discriminate between healthy and unhelathy banks. In the extreme case, in which the reported accounts of the distressed and non-distressed banks are indistinguishable, the information value of the accounting fundamentals in identification of bank distress is essentially non-existent.

Main hypothesis

[Accounting Discretion] $$\propto$$ -[Distress Prediction Performance of Accounting Variables]

Empirical Approach

Distressed Bank Event Database
Drivers of bank distress.
Predictive performance of accounting variables across countries
[Accounting Discretion] $\leftrightarrow$ [Predictability of Bank Distress]

Let me give you a brief outline of our empirical approach

To set the stage, we construct a comprehensive database of bank distress events in Western Europe and the U.S. during the period 2007-12. We categorize events into two broad groups of bank resolution: (1) bank closures, corresponding to resolutions in which distressed banks cease to exist as independent entities, and (2) open-bank resolutions, in which banks are allowed to continue operating with the assistance of a government bail-out.

We analyze the drivers of bank distress by modelling the two competing groups of distressed bank resolutions in a logistic regression framework. In our benchmark specifications we test for a number of bank-specific variables, including size, regulatory capital, asset quality, liquidity, franchise, or charter value1, and funding costs.

Next, we assess the classification performance of accounting fundamentals and of accounting-based models within different countries in our sample. We also test the ability of aggregated accounting measures reported in the fiscal year before the financial crisis to anticipate the aggregate degree of financial distress during the crisis years.

In the final part of the paper, we look into the association between information content of accounting fundamentals on one hand, and measures of accounting discretion on the other. We do so by exploiting the cross-country variation in the proxies of national bank disclosure standards and the stringency of their implementation by the national regulators.

Preview of the Main Results

Drivers and symptoms of bank distress:

regulatory capital (-)
asset quality (-)
charter value (-)
funding costs (+)

Information content of accounting fundamentals

Large cross-country variation in the bank distress classification performance.
At the country level, accounting fundamentals reported in the pre-crisis years of 2006 and 2007, fail to explain the 2007-10 aggregate incidence of bank distress across countries.

[Accounting Discretion] $\leftrightarrow$ [Predictability of Bank Distress]

Better classification performance of accounting variables in countries with strong disclosure laws or with more stringent enforcement of the existing laws.

In the final part of the paper we test for the association between accounting discretion and the classification performance of accounting fundamentals and show that:

first, accounting signals of bank distress tend to be stronger in countries with strong disclosure laws or with more stringent enforcement of the existing laws.

second, accounting discretion/informativeness nexus holds when looking at the time series movements in accounting fundamentals at the level of distressed banks prior to the distress event.

Taken as a whole, the evidence in this paper supports the often heard belief that excessive flexibility in financial reporting undermines the ability of accounting signals to accurately capture the underlying financial health of banks.

Literature

Bank distress prediction during the recent financial crisis
- E.g. Betz, Oprica, Peltonen, and Sarlin (2014); Cipollini and Fiordelisi (2012); Cihak and Poghosyan (2009)
Information value of accounting disclosure
- Healy and Palepu (2001); Beyer, Cohen, Lys, and Walther (2010); Altman, Gande, and Saunders (2010)
Accounting discretion and the informativeness of accounting signals
- Beaver, Correia, and McNichols (2012)

Bank Distress During 2007-12

Table 1 provides a top-down view of the bank distress database. Several interesting observations emerge from the table.

First, bank distress has been pervasive during the crisis: in the countries under study, the assets attributed to banks in distress represented on average about 30% of the total commercial banking assets4, ranging from 5% in Lux- embourg to 87% in Greece.

Second, banks resolved via closure tend to be smaller on an individual as well as on aggregate basis, compared to banks resolved via open-bank assis- tance. The average size of a closed bank in Europe (the U.S.) is about 39 billion USD (7 billion USD), whereas the average size of a bank resolved via open-bank assistance is 190 billion USD in Europe and 45 billion USD in the U.S. In aggregate terms, bank closures represent about 15% (30%) of distressed bank assets in Europe (respectively, the U.S.).

Bank Distress Events

Bank closure:
- Charter of the bank is revoked.
- Includes: liquidations, court bankruptcies, regulatory receiverships, distressed mergers (mergers, in which the acquired entity's regulatory Tier 1 capital ratio falls below the Basel II threshold of 6% prior to the merger).
- Source:FDIC, Bankscope, SNL Financial, LexisNexis.
Open-bank resolution:
- Charter of the bank is preserved, and the institution continues to operate as the independent entity.
- Includes: government bailout (e.g. investment in bank capital), coupled with a set of measures to improve the long-term viability of the bank (e.g. reallocation of the toxic assets to a bad bank).
- Source in U.S.: bank equity infusions under the Capital Assistance Program (CAP) of TARP.
- Source in Europe: State Aid cases from the European Commission.

Sample

U.S. and 15 Western European countries
Balance sheet data from Bankscope
Bank types: (1) bank holding companies, (2) commercial banks, (3) cooperative banks, (4) mortgage banks, and (5) savings banks.
Accounting figures at the highest level of conolidation.
Selection of accounting fundamentals: representative variables from the CAMEL dimensions
- Capital adequacy, Asset quality, Management quality, Earnings, Liquidity, and Sensitivity to market risk

Drivers and Symptoms of Bank Distress

How do distressed banks differ from their non-distressed counterparts in the years leading to distress?

We begin by analyzing developments in a selection of bank accounting measures in the periods leading up to the a distress event.

We do so by estimating the following specification [show on slide], where f are dummies denoting each year to distress (0 indicating the time of event) and $\alpha_{ct}$ denotes the country-year specific effects. We estimate the specification for each selected accounting ratio individually, and plot the resulting $\phi$ coefficients together with their corresponding standard errors.

Each of the accounting variables is standardized to have a zero mean and variance of one, implying the following interpretation of a coefficient $\phi_j$ : banks experiencing distress event between j and j + 1 years in the future display on average $\phi_j$ standard deviations higher or lower value of y than their non-distressed peers, controlling for country-year specific trends.

$$ y_{ict} = \alpha_{ct} + \sum_{j = 0}^n \phi_{j} f^j_{ict} + \epsilon_{ict}$$

The most important result relates to bank capitalization: distressed banks in both Europe and the U.S. tend to be significantly under-capitalized with respect to their non-distressed peers.

The economic magnitude of the result is especially big for bank closures, with the distressed/non-distressed Tier 1 capital lag reaching 0.4 standard deviations in a year before the distress event.

In Europe, the relative Tier 1 under-capitalization of distressed banks spans the period of at least 5 years before the distress event, while in the U.S. the under-performance is particularly notable during the three years before the event. Finally, U.S. distressed banks undergoing an open-bank resolutions are on aver- age significantly better capitalized than their European counterparts in the same distress group.

Decline in Tier 1 capitalization of distressed banks coincides with deterioration in their profitability, particularly for the group of banks that are eventually closed. Deterioration in profitability, in turn, is related to increasing loan-loss provisions and impairment charges, as well as to declining interest margins and operating efficiency (measured by the fraction of non-interest expenses in bank gross revenues).

Finally, a comparison of pre-event trends in accounting fundamentals between banks in Europe and the U.S. reveals a strong deterioration in fundamentals for bank closures in the U.S., whereas no such clear time-pattern is present in Europe. A temporal deterioration in fundamentals of the closed banks in the U.S. is particularly pronounced in the case of Tier 1 capital, unreserved impaired loans, loan loss provisions, non-interest expenses, and profitability.

Multivariate Model of Bank Distress

$$ Pr(\text{Distressed}_{ict} = 1) = \text{Logit}(\alpha_{ct} + x_{ict}'\theta + \epsilon_{ict} )$$

As before, we present results separately for the two types of distress, as well as for Europe and the U.S. Estimation of specification in the equation is equivalent to the estimation of an exponential hazard model, in which a probability of distress does not depend on the banks' age.

Also as before, all explanatory variables are standardized to have a mean zero and a unit variance, so that the magnitude of the reported coefficient corresponds to the impact of one standard deviation increase in the explanatory variable on the log-odds ratio. Consequently, the absolute magnitude of the coefficient can be used to judge the relative economic importance of different variables in the specification.

As you can see in the table, the likelihood of bank closure increases with (1) the degree of Tier 1 undercapitalization, (2) asset risk (measured by the ratio of RWA to book assets), (3) the amount of unreserved loan loss impairments, (4) cost of funding, and (5) the degree of operational inefficiency, (6) a decrease in bank profitability, measured by the interest margin, and (7) a decrease in asset liquidity, though the effect of the latter is statistically insignificant.

In terms of its explanatory power, the bank closure model explains bank closures with substantially higher degree of accuracy in the U.S. (with pseudo R-squared of 40%) than in Europe (13%). In the light of the univariate dynamics reported in Figure 1 this is not surprising, because most accounting ratios reported by the distressed banks in the U.S. exhibit clear negative trends already several years prior to the distress event. If we change the forecasting horizon in the U.S. to two years in the future, the R-squared of the model drops to around 20%.

Information Content of Bank Accounting Across Countries

How informative are bank accounting fundamentals in identifying distressed banks within countries?

The informativeness of a model is measured by the area under the ROC curve.

In this exercise, we evaluate the informativeness of the multivariate bank failure models developed before in discriminating between distressed and non-distressed banks within each country in our sample.

The informativeness of a model is measured by the area under the ROC curve12 (henceforth AUC) from a classification exercise in which the model-implied predictions are used to predict bank distress within a specific country. A particularly useful interpretation of the AUC is that it is the probability that the randomly chosen distressed bank observations exhibit higher values of the predicted model score than the randomly chosen surviving observation. At one end of the spectrum, a completely uninformative classifier has the AUC of 0.5, whereas a perfectly predictive classifier has an ROC of 1.

We assess the within-country predictive performance of different model predictions, by computing AUCs for each individual country. AUCs are obtained from the non-parametric ROC estimation, using bootstrap.

The table here shows the Areas Under ROC Curve (AUC) for predictors generated by a set of bank failure models, applied to predicting different types of bank distress events (1 year prediction horizon) within a set of 15 Western European Countries and the U.S. in the period 2006-2012.

Each column in the table corresponds to the model that is used to generate bank distress predictions. Each model is a logistic regression using the same vector of covariates as the multivariate models discussed before. The models differ in the sample used to estimate the model (i.e Europe, U.S., or both) and in the event that serves as the dependent variable in the model estimation

For the sake of brevity, I show only the excerpt from the table.

The main conclusions that emerge from the exercise, and are important for our subsequent analysis, are the following:

Predictions generated by any given model display substantial cross-country variation in the accuracy of the within-country forecasts of any of the three types of bank distress events.

Some of the countries with consistently low accuracy of distress predictions include Netherlands, Portugal, Ireland, and Denmark. In these countries, the accuracy of predictions in general does not exceed the AUC of 70%, and is, in many cases, close to the uninformative benchmark of 50%

Countries with consistently high levels of accuracy include the U.S., Austria, France, and Germany. The accuracy of predictions in these countries is typically above AUC of 80%.

How informative are the pre-crisis bank accounting figures in explaining the aggregate incidence of bank distress across countries during 2007-10?

y ~ country-specific asset share of distressed banks (% of total banking assets in 2008)

Our main dependent variable of interest is the fraction of book assets attributable to banks that became distressed during the period 2008-10 relative to the total amount of banking sector book assets in the fiscal year-end of 2008

Having defined the benchmark measure of country-wide bank distress, we now analyze the extent to which the variables that explain the within-country variation also explain the cross-country variation in bank distress.

Several observations emerge from the figure.

First, reported Tier 1 and Tier 2 regulatory capital ratios (reported as a fraction of risk-weighted assets) serve as poor predictors of banking problems at the country level. If anything, banks in countries with high rates of distress in 2008-10 report on average higher levels of both forms of regulatory capital in years preceding the crisis.

Apart from the reported asset risk weights, the only bank accounting-based aggregate in 2006 that exhibits a clear relation with the ex-post bank distress in 2008-10 is the net-interest margin. Specifically, countries with banks that reported on average higher net-interest margins in 2006 experi- enced higher incidence of bank distress during 2008-10.

Accounting Discretion and the Informativeness of Bank Accounting Signals

Predictability of bank distress by accounting fundamentals varies substantially across countries.

Bank management has capacity and incentives to exercise accounting discretion in order to report inaccurate information.

Bank disclosure standards and their enforcement by regulators provide a constraint on accounting discretion and on their information revelation incentives.

Bank disclosures standards and stringness of their implementation = proxy of accounting discretion

Measurement of Bank Disclosure Standards and Their Enforcement by the Regulators

Country-specific bank disclosure quality from the database of Barth, Caprio, and Levine (2013).
- more than 50 different indices from the quadrennial World Bank surveys covering 180 countries since 1999.

For each index, higher values of the index correspond to either better disclosure standards, or a more stringent implementation of the standards by the regulator.

Test 1: Bank Disclosure Quality and Accounting Information Content in a Cross-Section of Banks

Hypothesis: Given a cross section of banks in country $c$ at time $t$, the marginal impact of an accounting fundamental, $x$, on the probability of bank distress, increases with the value of the regulatory disclosure index, $R$.

That is: performance of an accounting fundamental in classifying banks within a country is greatest in countries with stringent disclosure laws or the enforcement thereof.

Measure the information of content of an accounting fundamental, $x$, as the absolute magnitude of the marginal impact of $x$ on the probability that a bank becomes distressed 1 year in the future, within a cross section of banks in country $c$ at time $t$:
$$$$ $$ INFO_{ct}(x) =\Bigg|\Bigg| \left. \frac{\partial Pr(\text{Distressed}_{ict} = 1)}{\partial x_{ict}} \right|_{\text{c,t fixed}} \Bigg|\Bigg|$$

The hypothesis can be restated as: $$$$ $$INFO_{ct}(x)\Bigg|_{\substack{\\c\in\text{Good Disclosure Country}}} > INFO_{ct}(x)\Bigg|_{\substack{\\c\in\text{Bad Disclosure Country}}} \geq 0$$

The intuition of the measure is simple: the information value of an accounting fundamental increases with its ability to identify distress in a cross section of banks in a given country-year. In line with the discussion above we expect the informativeness of an accounting measure to be greater in countries with more stringent standards or with more vigilant implementation of the standards by the regulators. Following the previous notation, this can be stated as: [show equation].

Notice that the hypothesis does not postulate the direction of the correlation between the accounting signal and bank distress, but concerns only the magnitude of the correspondence. The implications of the hypothesis can be nuanced further, by taking into account the direction of the associations between bank distress and fundamentals, predicted by the banking theory. Theoretically, one expects to observe a negative asso- ciation between bank distress and bank capital (both Tier 1 and Tier 2), and a positive association between bank distress and RWA, unreserved impaired loans, and loan loss provisions. We expect the theoretically predicted direction of the correspondence to be stronger in countries with better disclosure laws, which implies a negative interaction term, for bank capital, and a positive interaction term for RWA, unreserved impaired loans, and loan loss provisions.

$$Pr(D_{ict} = 1) = \text{Logit}(\alpha_{ct} + \theta_1 * x_{ict} + \theta_2 * R_{ct} * x_{ict} + \epsilon_{ict} )$$

This table presents the results of the tests.

We separately estimate the specification for five accounting ratios that are often considered as the most prone to manipulation, namely: (1) Tier 1 capital ratio, (2) Tier 2 capital ratio, (3) risk-weighted assets20, (4) unreserved loan losses, and (5) loan loss provisions. Columns 1-7 report the estimates for the regressions with regulatory interactions for each of the disclosure and monitoring indices described before.

Estimates reported in different panels come from separate univariate estimations of the specification. Since our regulatory variables are standardized to lie in the range between 0 (worst disclosure quality) and 1 (best disclosures quality), the interpretation of the interaction term coefficient is straight- forward: it represents a change in the marginal contribution of the accounting fundamental on the probability of bank distress as one moves from the worst-disclosure jurisdiction to the best-disclosure jurisdiction.

Results are consistent with the disclosure-quality hypothesis for Tier 1 regulatory capi- tal ratio, unreserved loan losses and loan loss provisions. In each of the cases, an accounting signal of bank distress tends to be stronger in countries with strong disclosure laws and/or with more stringent enforcement of the existing laws. Additionally, the direction of the accounting signal in each of the three cases is consistent with the theoretical prior. Specifi- cally, Tier 1 regulatory capital ratio exhibits a negative relation with bank distress, whereas the unreserved loan losses and the loan loss provisions exhibit a positive association.

Test 2: Bank Disclosure Quality and Accounting Information Content. Within-Firm Evidence

Does a time series of an accounting signal produced by a distressed bank anticipates the bank's eventual failure?

A signal is judged as informative if its reported value immediately before the distress period is distinct from its value in the periods further from the distress event.

$$ Pr(D_{ict} = 1) = \text{Logit}(\alpha_{i} + \theta_1 * x_{ict} + \theta_2 * R_{ct} * x_{ict} + \epsilon_{ict} )$$

The above specification is estimated only on the subsample of banks that become distressed at some point in the sample. The main difference between this and the previous specification is that the latter exploits the within-firm variation to estimate the coefficients, whereas the former relies on the within-country/year variation.

The table shows the results of the within-firm estimation. The main conclusions are similar to the previous section. The informativeness of the Tier 1 capital ratio, unreserved loan losses, and loan loss provisions tends to be greater in countries with better disclosure quality. On the other hand, the disclosure-contingent reversion in the association between bank distress and Tier 2 capital is even more pronounced within a firm than in a cross section. Specifically, prior to their distress event, banks in countries with low disclosure quality tend to increase their Tier 2 capital whereas their counterparts in countries with better disclosure quality tend to decrease the reported levels of Tier 2 capital.

On the additive value of Risk-Weighted framework in measuring bank capital

\begin{align} RWCR & = \frac{\text{Tier 1 Capital}}{\text{RWA}} \\ &=\frac{\text{Tier 1 Capital}}{\text{[Risk-Insensitive Exposure Measure]}}\frac{\text{[Risk-Insensitive Exposure Measure]}}{\text{RWA}} \\ & = [\text{Leverage Ratio}] * 1/[\text{Average Risk Weight}] \end{align}

Core Equity over Total Book Assets by Country

Tier 1 Regulatory Capital Ratios by Country

Asset Risk Weights by Country

"Leverage Ratio vs Risk Weights" Analytical Dashboard

Conclusions

In-depth examination of the information content of the accounting fundamentals in prediction of bank distress during the recent crisis. $$$$

Predictions generated by accounting-based models display a substantial cross-country variation in bank distress classification performance. $$$$

Pre-crisis values of accounting fundamentals, aggregated at the country level, fail to explain the 2007-10 aggregate incidence of bank distress across countries. $$$$

Accounting signals of bank distress tend to be stronger in countries with strong disclosure laws or with more stringent enforcement of the existing laws.