Aplication of Machine Learning Methods to the of Borrowers' Credit Rating

Recently, with the growing popularity of lending, the level of overdue debts has also been growing. In this connection, commercial banks need to improve the quality of credit risk management. One way to reduce credit risk is to improve the scoring models used to analyze the credit rating of borrowers. A review of the literature and previous empirical studies made it possible to determine that at the moment there is a lack of research in terms of analyzing the effectiveness of non-parametric methods for assessing the creditworthiness of individual borrowers on real data of Russian commercial banks. The main goal of this work is to build a scoring model based on modern non-parametric methods of machine learning (random forest and gradient boosting) and their comparison with traditional methods for assessing the creditworthiness of individuals (logistic and LASSO regression). The analysis was carried out on a sample of 300 thousand individual borrowers provided by the russian division of Home Credit Bank. The most effective model turned out to be the gradient boosting method - LightGBM. The use of this method allowed to increase the Gini index value by 6% compared with traditional models (logistic and LASSO regression) while this method works 3-10 times faster than other non-parametric methods or neural networks. . Using this method, it was also possible to calculate the importance of variables in improving the separating ability of the algorithm. Thus, the most important variables were features from the credit bureau and artificially generated business features based on the client’s loan application, while socio-demographic variables had little effect on increasing the predictive power of the model.