Improving Credit Default Prediction Using Explainable AI

Despite recent improvements in machine-learning prediction methods, the methods used by most lenders to predict credit defaults have not changed. This is because most of the high-performing methods are of a black-box nature. It is a requirement that credit default prediction models be explainable. This research creates credit default prediction models using tree-based ensemble methods. It is shown that model performance can be improved by using gradient boosting methods over traditional credit default predictions models. The top performing XGBoost model is then taken and made explainable. This research proposes a model-agnostic counterfactual extraction algorithm that explains the drivers behind a particular prediction. The algorithm focuses on extracting the counterfactuals that have the fewest contrasting features. This results in counterfactuals that are easily understood by humans and can be easily translated into insights that the lay user can understand. A definite standard of explainability is defined and the counterfactual extraction algorithm results in explanations that meet this standard. Given that the explanation method is model agnostic, it can be used on any prediction model and can be deployed for a wide range of applications