Predição de inadimplência no pós-concessão de crédito: aprendizado de máquina com dados do sistema de informações de crédito (SCR)

Abstract

This study presents an approach focused on default analysis and prediction in the post-credit granting period, using machine learning techniques applied to public data from the Brazilian Credit Information System (SCR). The objective is to investigate the ability of predictive models to identify patterns associated with default risk, contributing to the monitoring of credit operations after they have been granted. To conduct the study, the dataset underwent several preprocessing stages, including missing value treatment, feature transformation, reduction of high-cardinality categorical variables, and the evaluation of class balancing strategies. During the modeling phase, different classification algorithms were trained and evaluated, including Logistic Regression, Decision Tree, K-Nearest Neighbors (KNN), Random Forest, Support Vector Machine (SVM), and XGBoost, using Accuracy, F1-Score, and ROC AUC as evaluation metrics. The results showed that the representation strategy adopted for categorical variables had a greater impact on model performance than the evaluated class balancing techniques, with the Random Forest algorithm applied to the frequency-based grouping scenario for the submodalidade variable achieving the best overall performance. As a complement to the study, a Minimum Viable Product (MVP), consisting of a web interface integrated with an API, was developed to demonstrate a possible practical application of the predictive model. It is concluded that the proposed approach can contribute to post-credit granting monitoring and credit risk analysis, supporting the early identification of operations with a higher probability of default.