Early warning models for European banks

1 Introduction

The on-going financial turmoil evident in European financial markets and the rest of the world re-emphasizes the economic importance of bank stability. Since the banking system is the main mechanism for financial stability, monitoring the appetite for risk within banks has become a central issue for regulators. The inability of central banks to foresee and respond in a timely manner to financial meltdown has contributed to a further spread of the crisis resulting in the consequent high cost of bank bailouts. Only in the period 2008–2011, 1.6 trillion Euros (equivalent to 13 % of the EU’s annual GDP) were spent to bail out European banks. The crisis has demonstrated the vulnerability of the EU banking system and highlighted the need to strengthen financial sector supervision. As has been evidenced, the early detection of a bank in distress and timely intervention help regulators to promptly address the causes and to mitigate negative effects, thus preventing a contagion effect.

This study analyses the effectiveness of the Early Warning System (EWS) based on financial market information for forecasting bank defaults during the recent financial crisis. The selection of the EDF measure over other ones is based on its two distinctive characteristics: it is grounded on corporate theory (unlike reduced models) and incorporates market information (unlike conventional structural credit risk models such as Merton’s distance-to-default indicator). It is argued in several empirical works [2, 8] as being an efficient measure of firms’ defaults which generally outperforms the basic Merton structural model and Hull and White’s reduced form models. Moreover, EDF incorporates more realistic assumptions which better adjust to reflect real-world default dynamics.

The majority of the earlier papers use data from the pre-crisis period with relatively fewer default events, while we use both pre- and post- crisis data and expect that this will improve the statistical power of our research. In this paper we study the ability of EDF and other accounting-based ratios to anticipate a financial crisis and make a timely discrimination of bank defaults. Following Otero [18], market indicators seem to value the risks not perceived in accounting terms and both of the measures are complementary. Specifically, we endeavour to answer two crucial questions: Is EDF a good measure to predict the financial distress of banks? Do CAMEL and adverse selection variables bring additional predictive power to the EDF measure?

We test the validity of EDF using the logit model in spite of intelligent techniques performing better than statistical techniques [5]. For example, Dynamic Bayesian Networks provide early-warning models which are more precise than logit methods [13]. Furthermore, López Iturriaga and Pastor Sanz [4] developed a neural network model for US banks which outperforms traditional models of bankruptcy prediction. Our selection is based on the fact that Le & Viviani [5] recognise that the use of traditional techniques still works effectively and with many intelligent techniques like K-nearest neighbour, artificial neural networks and support vector machines it is not possible to evaluate the influence of each ratio to the model.

This study proposes possible improvements to the EWS which could be useful to identify inputs to incorporate in intelligent models. In addition, the paper could help central banks to better anticipate future banking crises and take preventive action against them. For assessing a bank’s solvency, it can also be handy for bank stakeholders, depositors and creditors. Bank managers can also benefit by applying the enhancement to measure the impact of their financial decisions such as change in capital level and credit growth stability.

The paper is organized as follows: firstly, it presents a review and discussions of the related empirical studies on EWS. Secondly, the methodology section describes the statistical model and data used in the model. The next section shows the results received along with discussions followed by the final section which concludes the study.

2 Related literature

To our knowledge Gropp, Vessala and Vulpes [23] and Distinguin, Rous and Tarazi [6] are the first authors to study the EWS for EU banks based on security data. One undertaken by Gropp, Vessala and Vulpes [5] for European banks finds that the distance-to-default (DD) measure is both comprehensive and unbiased and may prove a useful leading indicator of bank fragility. By estimating logit and proportional hazard models it was found that both indicators could predict bank distress up to 18 months prior to an event, even taking into account the safety net effect. Recent regulatory changes and financial and technological innovations have changed the existing banking environment and the causes of financial distress both in US and European markets. To capture these changes, the introduction of more risk-focused indicators has been suggested to adapt risk-measuring models to the new banking environment [23]. This implication is empirically tested using a sample of 82 EU banks observed from 1991 to 2005 by Brossard, Ducrozet and Roche [20] The aforementioned study constructs DD indicators to test the predictive power of bank failure while introducing a variable detecting the adverse selection effect of rapid growth strategies in their model. The latter study confirms the robustness of DD as an early indicator of banks’ failure. DD remains significant when it is joined with CAMEL accounting indicators and after the introduction of the “Too-Big-To-Fail” effect. Another important finding is that the new indicator of the adverse selection effect improves the predictive power of the model. In the same line, the work of Auvray and Brossard [28], focuses on another factor which may have a negative effect on the reliability of share prices in predicting bank distress. The test confirms that ownership dispersion of a bank’s shareholders clearly reduces the effectiveness of distance-to-default as a predictor of bank distress and bank recovery. In contrast, when ownership is concentrated it raises the predictive power of the indicator. Empirical findings in emerging markets also are favourable towards equity market signals in predicting banks’ financial distress. Chan-Lau et al. [10] show that DD indicators can forewarn bank distress, defined as a rating downgrade to CCC or below, up to nine months ahead of the sample. Later, Chan-Lau and Sy [9] offer an alternative risk measure named Distance-to-Capital (DC) to serve as the regulatory purpose of bank supervisors in intervening well ahead of a bank’s default. Another test of the predictive power of DD for eight failed Japanese banks is proposed by Harada et al. [14]. In addition, Eichler and Sobansky [25] consider that distance-to-default based on market data works well as an indicator of banking fragility in the Eurozone. Chen and Chu [13] explain that during the financial crisis, the correlations between large- and small- sized companies in terms of EDF fluctuated severely during the financial crisis and that the large firms face a higher default risk because of the higher level of leverage. Jessen and Lando [1] also show that distance-to-default is successful in ranking firms' default probabilities, even if the underlying model assumptions are altered. Additionally, Nazeran and Dwyer [22] make a detailed analysis of EDF calibration and explain that EDF continues to be a very useful tool that incorporates the most recent information which can affect the likelihood of default. Recently, EDF has been used by Gomes, Grotteria and Wachter [11] to develop a model that explains the ability of credit indicators to forecast aggregate real outcomes.

EDF could also be included in intelligent techniques because they perform better than statistical ones [5]. Neural networks are the most popular tool in predicting defaults. Some authors like Olmeda y Fernández [7], Tam [15], López and López Iturriaga & Pastor Sanz [4] developed neural-network models that outperform traditional models of bankruptcy prediction. Other techniques like Support Vector Machines have also shown significant accuracy and in some cases they work better than neural networks [16]. What is more, Dabrowski, Beyers and Villiers [13] showed that dynamic Bayesian networks provide early-warning models which are more precise than logit methods. But Le & Viviani [5] recognise that the use of traditional techniques still works effectively and are useful when the objective is to evaluate the influence of each ratio in the model.

3 Data and empirical methodology

The sample comprises 93 European listed banks for which Moody's EDF was available. The Accounting rating data is taken from the Bankscope International Bank Database provided by Fitch/Bureau Van Dijk. Since downgrades are announced in an irregular manner and do not occur at equally-spaced intervals, we decided to use quarterly data to better match the announcement times. Finally, we cover the period 2005–2011 to consider the effect of the financial crisis.

In our analyses we apply a binary choice model, the binomial logit model. The binary model assumes that banks belong to either one of two states: in our case it is either downgraded to D+ and lower, or it is not. In binary models an increase in X corresponds to an increase in probability of outcome 1. The equation of the logit model is:

Pr ( y i = 1 | X i ) = F ( x ′ β ) = ∧ ( x ′ β ) = exp ( x ′ β ) 1 + exp ( x ′ β )(3)

The logit model uses the cumulative logistic probability distribution for the cumulative distribution function (cdf). The baseline logit model is as follows: Pr ( Default = 1 ) = F ( β 0 + β 1 EDF )(4) where F(.) is the cumulative logistic distribution and EDF is a bank’s 1-year and 5-year EDF. The signs of the regression results are interpreted as follows: a positive sign denotes an increase in the likelihood that (default event), i.e. it shows that the default probability more likely. A negative sign denotes the opposite, i.e. it shows that the default probability is less likely. It is important to interpret the sign of the coefficient but not the magnitude. The magnitude can be interpreted through the marginal effect, which in this case is equal to:

∂ F ( x ′ β ) ∂ x _ j = ∧ ( x ′ β ) [ 1 − ∧ ( x ′ β ) ] β j = exp ( x ′ β ) ( 1 + e x p ( x ′ β ) ) 2 β j(5)

To analyse whether the baseline models have more predictive power when combined with CAMEL ratios and/or variables of adverse selection effect, we introduce two groups of additional variables. In the first group we have 5 CAMEL ratios which are widely used by investors. Concerning the adverse selection effect, Brossard et al. [20] report a very significant positive impact on banks’ default probability even at the 2-years horizon. We test the impact of this factor in our dataset by applying two alternative measures: the average growth of total assets and the average growth of gross loans.

The recent European banking crisis has highlighted the need to improve early warning systems for bank fragility. Our study analyses how changes in financial environments of the banking sector have affected the predicting power of the EWS and whether as well as how the conventional models of predicting bank distress need to be modified. To discriminate defaulted banks from sound ones we use the expected default frequency (EDF) as the main explanatory variable. In line with the previous literature we examine CAMEL supervisory ratios together with EDF. Another variable tested is the adverse selection effect measured by way of two alternative ratios: past-asset growth and gross-loans growth. Our results reveal that EDF metrics combined with the four CAMEL covariates and the adverse selection variable are able to predict the defaults of European banks up to 8 quarters before an event. When comparing the final model with the one that only includes the EDF indicator, the statistical significance improves considerably, suggesting that added variables provide additional information and power to the model. Considering the adverse selection effect, both definitions contribute to the predictive power of the model. Our findings suggest that unsustainable growth strategies could lead to the subsequent deterioration of a bank’s financial state and increase the likelihood of default.

To summarize, in Early Warning Models it is beneficial to use both market-based indicators and balance sheet indicators as they mutually complement each other. It also implies that EDF provides additional information to that of balance sheet ratios but is not a complete substitute for them. The models can be applied as complementary tools to foresee a potential banking crisis and as offsite means of regular monitoring. Eight quarters’ notice of a potential bank default should permit a timely response to it, and thus improve financial stability in the banking sector as a whole. Our paper is interesting to identify variables to be included in intelligent models and to increase the accuracy of the logit models considered in this study.