Predição de default de empresas: técnicas de machine learning em dados desbalanceados

• Main Author: Cordeiro, Tiago Vilas Boas
• Publication Date: 2020
• Format: Master thesis
• Language: por
• Source: Repositório Institucional do FGV (FGV Repositório Digital)
• Download full: https://hdl.handle.net/10438/29873
• topic: Machine learning; Logistic regression; Random forest; Probability of default; Risk rating models; Regressão logística; Probabilidade de default; Modelos de rating; Economia; Aprendizado do computador; Análise de regressão logística; Avaliação de riscos

Given the importance of credit risk management for the banking sector, probability of default models have become fundamental. In this context, with the advances in the volume of information from customers and the computational capacity, several techniques have been studied and applied. In this study, we used two traditional linear techniques, the Linear Discriminant Analysis and Logistic Regression, and four non-linear ensemble techniques, Bagging, Random Forest, Adaboost and Stacking, applied to a problem of probability of default on brazilian companies, using information from their financial statements. The results indicate that the transformations in the data and treatment of class imbalanced have a strong impact on the predictive power of Logistic Regression. Yet, Random Forest was the technique with the best performance regardless of the scenario and the metric used.