Modelling Default Risk of Credit Institutions due to the Loss of their Capital

This paper analyzes license revocations from Russian banks due to a shortage of capital for the period from January 2007 to December 2018. On the basis of publicly available macroeconomic statistics and disclosed intrabank information, a predictive model of the likelihood of similar reviews in the future is built. Initially, modeling is performed using standard binary choice models such as logistic regression, random forest model, and ROC curve analysis. Using these methods allows you to rank the significance of factors for the predictive power of the model, as well as qualitatively interpret the results. However, this study shows that standard models rather well predict the likelihood of bank failures of banks that do not falsify the reports provided, but worse, they cope with this task for banks with inadequate reporting. In this connection, more specific methods are subsequently introduced, such as the classification of the training sample and the trait recognition model. The classification of the test sample with the subsequent application of individual models to each class led to a decrease in the overall accuracy of the model, but an increase in the predictive power of the model on a subsample of banks with fictitious reporting. The transition from standard interpretable variables to their derivatives gave a positive result in the predictive power of the models. Thanks to the use of the recognition model, both the overall growth of predictive ability both within the classes of banks with false and true reporting and the significant reduction of the gap in the predictive power of the model between the declared classes of banks is achieved, which allows equally well to predict the default of both banks with true reporting and with fake one.