Accuracy of explanations of machine learning models for credit decisions

One of the biggest challenges for the application of machine learning (ML) models in finance is how to explain their results. In recent years, innovative interpretability techniques have appeared to assist in this task, although their usefulness is still a matter of debate within the industry. In this article we propose a novel framework to assess how accurate these techniques are. Our work is based on the generation of synthetic datasets. This allows us to define the importance of the variables, so we can calculate to what extent the explanations given by these techniques match the ground truth of our data. We perform an empirical exercise in which we apply two non-interpretable ML models (XGBoost and Deep Learning) to the synthetic datasets, , and then we explain their results using two popular interpretability techniques, SHAP and permutation Feature Importance (FI). We conclude that generating synthetic datasets shows potential as a useful approach for supervisors and practitioners who wish to assess interpretability techniques.