Modelos de regressão logística clássica, Bayesiana e redes neurais para Credit Scoring

Abstract

Important advances have been achieved in the granting of credit, however, the problem of identifying good customers for the granting of credit does not provide a definitive solution. Several techniques were presented and are being developed, each presents its characteristics, advantages and disadvantages as to their discrimination power, robustness, ease of implementation and possibility of interpretation. This work presents three techniques for the classification of defaults in models of Credit Score, Classical Logistic Regression, Bayesian Logistic Regression with no prior information and Artificial Neural Networks with a few different architectures. The main objective of the study is to compare the performance of these techniques in the identification of customers default. For this, four metrics were used for comparison of models: predictive capacity, ROC Curve, Statistics of Kolmogorov Smirnov and capacity of hit models. Two data bases were used, an artificial bank and a real bank. The database was constructed artificially based on an article by Breiman that generates the explanatory variables from a multivariate normal distribution and the actual database used is a problem with Credit Score of a financial institution that operates in the retail Brazilian market more than twenty years.