Predição de risco de crédito com aprendizado de máquina supervisionado: um estudo de caso com dados desbalanceados

Abstract

This work evaluates the application of penalized Logistic Regression via LASSO in credit risk prediction, focusing on the treatment of imbalanced data, a common characteristic in financial datasets where the proportion of defaulters is usually low. A total of 36 synthetic datasets were generated through simulation, with different combinations of sample size and proportion of defaulters. The model was calibrated using cross-validation and dynamic adjustment of the cut-off (decision threshold), without employing resampling techniques such as SMOTE (Synthetic Minority Over-sampling Technique, which generates synthetic examples of the minority class) or undersampling (reducing the majority class to balance the distribution). Instead, resampling procedures with replacement from the original dataset were adopted to generate different scenarios of sample size and imbalance. The results showed robust performance in metrics such as AUC, F1 Score, and balanced accuracy, even in scenarios with severe imbalance, that is, when the proportion of defaulters is extremely low (between 1% and 5%). The LASSO penalization contributed to the automatic selection of relevant variables, while maintaining model interpretability. As a practical contribution, the study demonstrates that, with careful adjustments, it is possible to obtain statistical models that are simple, effective, and compliant with regulatory requirements, providing reliable support for credit risk management. Furthermore, the findings can be generalized to other contexts involving imbalanced data, expanding the applicability of the proposed approach.