Statistical Analysis of Default Risk of Russian Banks Using Machine Learning Methods

Sustainable development of banking institutions has traditionally been considered as a key factor of economic growth and stability of any country. In order to improve the financial environment of the banking system, the default risk exposures of banks should be identified and managed effectively. To address this problem, financial supervisory authority attempts to build an early warning system (EWS) with the use of probability of default (PD) model. Bank failure prediction models are formulated as a classification problem with binary choice, i.e. two classes representing bankrupt and non-bankrupt banks, in a multidimensional space defined by a set of financial explanatory factors. These models are usually divided into statistical and machine learning techniques. Although statistical methods are interpretable and straightforward, their validity depends on whether certain strict assumptions are met by the relationships in the data. If one of the assumptions is violated, the results might be biased and overly optimistic, thus usefulness of traditional econometric models is limited. On the contrary, machine learning methods are able to deal with this issue, and therefore can perform more effectively than typically employed statistical techniques in the context of bankruptcy prediction accuracy. Improvement of the latter is of crucial importance since it allows to preclude more bank defaults (due to the loss of solvency) that are more ominous for country’s economy than the failure of any other financial institutions because of possible systemic banking crisis. While there has been a considerable number of empirical studies comparing statistical and intelligent techniques on forecasting bank financial failures for Western economies, no research to date have drawn such model comparison for the Russian banking industry. The aim of this study is to enhance PD modelling approach for Russian banks by executing and comparing traditional statistical methods and cutting-edge machine learning algorithms. It is hoped that this research will make an important contribution to the field of bankruptcy prediction since PD model can be utilized by three major parties in Russia, namely government, banks and creditors.