Constrained machine learning models for credit default prediction: who wins, and who loses

Balancing the prediction accuracy of machine learning models with their interpretability is crucial, particularly in settings such as credit decisions, where regulatory requirements demand transparent decision-making processes. Imposing monotonicity constraints to a machine learning model is one way of enhancing interpretability, but it may affect model performance and, consequently, credit outcomes. This study investigates the consequences of applying monotonicity constraints to a machine learning model used for credit default prediction. In particular, we compare the performance of an XGBoost model with and without monotonicity restrictions using Lending Club data. We then
assess who is most affected by these constraints, by focusing on individuals whose predicted default probabilities (PD) change the most, and by analyzing changes in features importance using a statistical framework provided by Shapley regressions (Joseph 2019, 2024). This allow us to assess potential bias and fairness concerns in automated decisions. Our findings suggest that the restrictions push all individual estimations towards the mean. While this effect is not homogeneous, we conclude that in principle it would not harm people’s access to credit. Notably, introducing monotonicity constraints enhances model interpretability while having a relatively small impact in the decomposition of model predictions into Shapley values, which is reassuring since it implies that the model is learning patterns
based on a similar set of explanatory variables.

This paper is part of the Banking Supervision Policy Working Paper Series in the context of the SSM-EUI partnership on SSM Banking Supervision Learning Services. Read more.