Estimating Banks’ Default Probability

The present study is devoted to modeling Russian banks failure probabilities. Regarding to Russian banks failures, the researchers have not answered the 2 questions: whether more complex financial ratios, statutory ratios, are more effective than their proxies from financial statements; whether more complex model types (neural networks, “random forest” models and so on) employment improves predictive power. The author verified the hypotheses on the Russian banks quarterly data for the period between the 3rd quarter of 2010 and the 1st quarter of 2016. The author estimated the 3 specifications of logistic regressionon short-term and mid-term forecast horizons: with proxies of statutory ratios; with statutory ratios; with polynomial elements (as a proxy for more complex model types). As a result, statutory ratiosemployment did not lead to statistically significant rise in predictive power, they increased prediction accuracy only by 2% on average. By contrast, polynomial elements employment resulted in significant increase in prediction accuracy, but only on short-term forecast horizon.