A Novel Survival Analysis-Based Approach for Predicting Behavioral Probability of Default.

Obtaining an accurate model for predicting the probability of default is a critical requirement for financial institutions. Nowadays, COVID19 has produced high economic instability bringing borrowers' delinquencies as well as moratorium regulations. During the KYC process for applicants, it is essential to estimate the probability of default for avoiding write-offs. However, there are several borrowers who, due to the pandemic, could have lost their jobs or decreased their income, producing several borrowers' delinquencies, even write-offs by frequent delinquencies. Consequently, having a behavioral model for estimating the probability of default during the loan lifetime is vital for financial institutions. Hence, in this paper, we propose the first survival analysis-based approach for predicting the behavioral probability of default. We collected two real financial databases from different countries with different borrowers' characteristics. From our experimental results, we can conclude that Logistic Hazard provides better results than Deep Hit for predicting the behavioral probability of default. Based on our experimentation and the risk analysis experts, Deep Hit provides inconsistent results for forecasts greater than six months while considering the financial changes due to the COVID19. Otherwise, Logistic Hazard is more accurate in forecasting the behavioral probability of default for a year, and it shows results more appropriate to risk analysis experts.