Aplicações de algoritmos de machine learning para previsão de inadimplência em concessão de crédito

• Autor(a) principal: Santos, Thiago de Jesus dos
• Data de Publicação: 2022
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFS
• Texto Completo: http://ri.ufs.br/jspui/handle/riufs/17494
• topic: Ciências atuariais; Ensino de ciências atuariais; Aprendizado de máquina; Crédito; Inadimplência; Balanceamento dos dados; Random forest; Machine learning; Credit; Default; Data balancing; Random forest; OUTROS::CIENCIAS ATUARIAIS

Due to the great computational advances, the development of models in the credit area in order to classify, measure the probability of implementation and other ways has been improved by machine learning techniques. In this context, the present study aimed to develop predictive models using machine learning techniques, in order to identify customers who are more likely to default on their debts to US financial institutions linked to a US Small Business Administration entity (SBA). Through descriptive analysis, an imbalance was observed in the data distribution of the response variable, referring to the credit to the implementation by the institutions, since 82% of the customers paid on time, while 18% became defaulter. Therefore, in this study it was proposed that Logistic Regression, Naive Bayes, Decision Tree and Random Forest (RF) methods generated models trained in three situations: 1) Real scenario (unbalanced); 2) Undersampling Scenario and 3) Oversampling Scenario. The results found indicate that the application of balancing techniques cause a reduction in accuracy and sensitivity in the part of the models, with an even larger increase in specificity of all adjustments. In addition, Random Forest obtained the best evaluation measurements among all methods used, regardless of the proposed scenario for the training set. Finally, using the “area under the curve” as an evaluation metric (AUC), the model (M12) generated by the RF model on a sampling technique resulted in the best performance in the generalization process.