A credit risk management model for a portfolio of low-income consumer loans in Mexico

Abstract

Low-income consumer lending (LICL) in Latin America has experienced a boom in recent years. This has attracted the interest of a large number of financial players eager to capture a portion in this still under-banked segment. Despite this huge market potential, credit risk management in this segment is still mainly based on the subjective expertise of credit managers, with few exceptions making use of robust statistical techniques. The result of this subjective decision process is a sub-optimal concession of loans that leaves this strategic segment without adequate financing opportunities. In this work we develop a cutting-edge probability of default (PD) model specifically designed for the LICL segment. This research is one of the first academic works that explores the applicability of four cutting-edge quantitative methods with real operation data from one of the pioneers in the low-end credit segment in Mexico: Mimoni Group. The analysis was carried out on a sample of 2,000 loans including 741 defaults and 1,259 nondefaults, spanning the 2013 to 2014 time period. We run a total of 108 models utilizing Logistic regressions, CART models, Random Trees, and Clustering over training and out-of-sample data sets. Our results not only generated powerful models in terms of statistical accuracy in out-of-sample data sets, but also provided a detail list of robust PD predictors (at 95% levels) and their dynamics in explaining the default event for two Mexican low-income customer segments. Our results demonstrate the direct applicability that robust quantitative models have in improving and complementing the lending decision process in the growing LICL segment.