Modelos estadísticos para la estimación de la capacidad de pago de personas naturales con cuota estimada con información en el sistema de registro de datos crediticios : modelo para la estimación de la capacidad de pago en una población bancarizada.

Abstract

The objective of this project is to estimate the payment capacity of banked individuals who at the time of the estimation have an estimated quota using parametric and non-parametric statistical methodologies or models (Multiple Linear Regression, Random Forest, Gradient Boosting Machine and XGBoost) with information in the credit data registry system. The training of the models was preceded by a rigorous exploration and treatment of the database consisting of 950821 records and 1172 variables; in order to capture the behavior and characteristic pattern of the population, it was found to be appropriate to divide it into three study groups and for each group the four methodologies or models indicated were formulated; The first group consists of those individuals who have an estimated quota (amount of amortization in USD of credit) less than or equal to 107 USD, the second group with those whose estimated quota is greater than 107 and less than 435 USD and the third group with those with more than 435 USD in estimated quota. To improve the predictive power of the models in the tails of the distributions (subjects with very low or very high incomes), the resampling or balancing technique was used, which effectively obtained better results compared to the base models (without resampling). Among all the models implemented, the Gradient Boosting Machine Model was chosen as the best model for its level of prediction and computational performance in the three sub-populations. The choice of variables was made using the Kolmogorov Smirnov (KS) methodology for quantitative variables and the Value of Information (VI) for categorical variables; 18 variables were selected for the first group or sub-population (which in this work will be synonymous), 35 for the second and 28 for the third. The results obtained were acceptable, given that the predictive power for the three groups or sub populations was over 70% correct, however for group 1 and 2 the results were not as expected in the tails of the distribution, better predictions were expected, however, for group 3 a high prediction rate was achieved; This predictive power of the models was validated with a new database consisting of 84877 individuals and 1177 variables; the income predictions for group 1 and 2 were not as expected (as expected); however, for group 3 a very good level of prediction was obtained, especially for subjects with very high incomes.