Comparação de Algoritmos de Aprendizado de Máquina para Predição de Pontuação de Crédito

Abstract

According to the Central Bank of Brazil, the total value of creditoperations in Brazil reached R$4.2 trillion in May 2021. Financialinstitutions must consider the risk of default associated with eachoperation. Credit analysis, which evaluates this risk, can be performedusing machine learning algorithms. These algorithms comparenew loan proposals to historical data to estimate the default probabilitybased on the proposal and proponent characteristics. Theaccuracy of the model is critical to the profitability of institutions,so choosing the right algorithm is crucial. This study comparesthe performance of machine learning algorithms on three publicdatasets in the task of credit risk estimation. The results show that astack of multiple classifiers achieved the highest accuracy at 81.41%,followed by XGBoost at 80.87% and Regressão Logística at 80.48%.