Estudo comparativo entre metodologias de aprendizado de máquina e híbridas aplicadas a risco de crédito

Abstract

For banks and companies that offer credit operations, having relationships with high-risk customers increases the probability of default, the need for capital allocation and the exposure to financial losses. Thus, there is interest in improving credit risk evaluations; and the current Big Data scenario enhances the interest in artificial intelligence methodologies, since their accuracy increases as the volume of data also increases. This dissertation aims to compare quantitative methodologies applicable to credit risk management and to conclude whether the techniques based on artificial intelligence present better performance than traditional techniques. The study includes the Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting and Hybrid Models methodologies, for both individuals and enterprises. The study concludes that Gradient Boosting is the champion methodology in the comparation made through the performance metrics AUC, KS and Hit rate. Keywords: Credit Risk. Artificial Intelligence. Machine Learning. SVM. Boosting. Hybrid Models.