Introduction.

The purpose of this research is to analyze the credit risk of a financial institution not supervised by the Financial Superintendence of Colombia around a scoring model that allows determining the default of clients corresponding to their consumer portfolio.

Objective.

Confront the forecasting power of two scoring models obtained through logistic regression and neural network.

Materials and methods.

The models are developed based on a sample of 43,086 obligations corresponding to a consumer portfolio, using the statistical techniques of logistic regression and neural network. The first is framed in the group of generalized linear models, which use a logit function and are useful for modeling probabilities related to an event based on other variables; while the second consists in computational models whose objective is to solve problems using relationships already stipulated, employing for this purpose a base sample of the process that is based on the success of the self-learning resulting from training.

Results.

For both models, an accuracy of 71 % in the training base and 72 % in the testing base is achieved. However, despite obtaining similar results, the logistic regression yielded the lowest rate of bad in the acceptance zone.

Conclusion.

The two techniques used are suitable for the study and prediction of the probability of default of a client corresponding to a consumer portfolio; the foregoing is supported by the high index of predictive effectiveness in both models.

        · abstract in Spanish | Portuguese     · text in Spanish     · Spanish ( pdf )