Evaluación de modelos de machine learning aplicados al cálculo de pérdidas esperadas en entidades de microfinanzas.

Abstract

The project aims to compare the effect of non-parametric machine learning models against the traditional methodology of logistic regression, to estimate the expected losses of a credit portfolio of an Ecuadorian financial institution. The trained Credit Scoring models were: Logistic Regression (RGL), Random Forest (RF), and Extreme Gradient Boosting (XGB), which estimate the probabilities of being a defaulter, considering individuals with 61 or more days past due in the performance window, based on Roll-Rate analysis results. Due to the imbalance of GOOD/BAD categories, it was necessary to rebalance the sample in order to better capture the characteristics predominant in bad payers.

Considering that the project had a total of 805 variables, it was necessary to employ divergence measures to select the variables to be included in the models. The Kolmogorov-Smirnov (KS) test was used for quantitative variables, and the Value of Information (VI) test for qualitative variables, seeking diversity of information and a more accurate representation of the financial reality.

The three trained models have good performance metrics, both in discrimination power and predictive stability, as evidenced in the performance tables obtained. According to the validation results, the model with the best performance metrics turned out to be the traditional methodology, followed by XGB. For the calculation of expected losses, the basic approach according to Basel II was used, meaning the loss severity was taken as 45\%. Based on the validation results, the model that reduced provisioning was XGB, indicating that an average of 3.51\% of the portfolio should be provisioned monthly.