Development of Rating System for Predicting Credit Risk and Probabilities of Default of Russian Financial Institutes by Using Machine Learning Models

The paper compares the ability of various statistical methods to predict probability of default of Russian financial institutes based on the publicly available information; assesses the probability of defaults of Russian banks on the base of the best model and develops the credit rating scale for assessing creditworthiness of Russian financial institutions. These tasks are of primary importance for researchers and practitioners as the previous studies indicate that the credit risk assessment analysis differ using different performance criterions on different databases under different circumstances. The paper is aimed at filling the gap in the existing research as only very few efforts were focused on prediction of credit rating scale using artificial intelligence (AI) methods. The modelled variables are Russian banks at the year-end from 2015 till 2019 The sample included 859 observations. Large set of financial, non-financial, prudential and macroeconomic data is applicated. The set of statistical methods included logit regression (LR), classification and regression trees (CART), support vector machine methods (SVM), artificial neural network (ANN) and random forest (RF), Model Based Trees (Logistic), Lasso regression and Ensemble model. The resulting models were checked for in-sample and out-of-sample predictive fit. The model and the rating system have the significant practical importance in regulation and credit analysis in banks. Keywords: Credit default prediction, Russian financial system, Ordered logit model, Artificial intelligence methods, Credit rating system.