Machine learning models for credit analysis improvements: Predicting low-income families’ default

Abstract

The main objective of this study is to investigate the behaviour of default prediction models based on credit scoring methods and computational techniques with machine learning algorithms. The predictive capabilities of the models were compared to identify default-prediction mechanisms in the “My Home, My Life” Program (Programa “Minha Casa, Minha Vida” — PMCMV). The PMCMV is one of the largest government initiatives in the world to finance home ownership in the low-income population. Implemented by the Brazilian government, the programme has provided financing in excess of USD 84 billion and by 2016 had already contracted for the construction of over 4.5 million housing units, with 3.3 million units already delivered. The models developed in this study involve different time intervals for default prediction as well as analysis without the use of traditional discriminatory variables (gender, age, and marital status). Three measurements were used to evaluate the quality of the prediction models: area under the ROC curve, the Kolmogorov–Smirnov index, and the Brier score. The results indicated that (1) the accuracy of the models improves as the number of days overdue used to define the default variable increases; (2) the best prediction results were obtained with traditional ensemble techniques — in this case Bagging (BG), Random Forest (RF), and Boosting; and (3) there was a negative impact on all criteria when a smaller number of observations was used, especially on the type II error. It was also found that the discriminatory power of the credit risk rating system is preserved when removing discriminatory variables from the models. Applying the BG algorithm, which is the best prediction method, a default rate of 11.80% could be reduced to 2.95%, which leads to a selection that would result in 197,905 fewer delinquent contracts in the PMCMV, thus representing a savings of approximately USD 3.0 billion in credit losses.