Estudo comparativo de modelos de Machine learning na predição de um behaviour score

Abstract

Credit Risk Modeling is an essential technique for financial institutions in making intelligent analytical decisions, such as granting credit, financing, or loans. These institutions aim to accurately identify individuals likely to meet their obligations on time by estimating the probability of customers being compliant or delinquent.This study aims to develop a Behaviour Score used by companies to assign scores to their customers based on their transactional behavior history, using a dataset with more than 400,000 customers and over 200 covariates.To achieve this, we analyzed the dataset and trained four machine learning models, ranging from traditional methods such as Penalized Logistic Regression (Oliveira et al., 2012) and Random Forest (Breiman, 2001) to more recent approaches like eXtreme Gradient Boosting (XGBoost) (Friedman, 2001; Chen & Guestrin, 2016) and Light Gradient Boosting Machine (LightGBM) (Ke et al., 2017). When comparing the models on the test sample, we found that, in general, all models effectively distinguish between the response variable classes. However, the model that achieved the best predictive performance was XGBoost, with an area under the ROC curve of 82.03%, a Gini coefficient of 64.06%, and a KS statistic of 47.97%, among other performance metrics. On the other hand, the model with the lowest predictive performance was Penalized Logistic Regression, which obtained an area under the ROC curve of 78.46%, a Gini coefficient of 56.92%, and a KS statistic of 42.73%, among other measures. Additionally, it had the longest training time compared to the other models.