Modeling of Probability of Business Loans Prepayment

This research focuses on the problem of business loans prepayment. Basing on the real data provided by a large Russian bank, a model for prepayment prediction was built with the use of information on the borrower, loan characteristics and macroeconomic indicators. In the paper such algorithms as logistic regression, logistiс regression with L1-regularization, support vector machine, Random Forest and CatBoost were applied. According to the results obtained, the highest performance was achieved by CatBoost (ROC-AUC=0.876 on the test set), the lowest – by logistic regression (ROC-AUC=0.698). Loan term, interest rate, deal amount and type were claimed to be the most important factors for prepayment prediction; at the same time, macroeconomic conditions turned out to be non-significant. With regard to financial characteristics of a borrower, the most important factors are coefficients indicating company’s debt burden level.