Machine Learning in Default Prediction to Enhance Loan Decisions and Customer Experience

Consumers generally have a fragmented perspective of how financial establishments operate, even though those institutions have a direct viewpoint into people’s lives, due to continual data collection. Yet, algorithmic decision-making methods are increasingly common in the financial sector because of the abundance of customer data and the demand for default decisions. Nevertheless, classification strategies can either produce understandable results with a clear model structure or utilize purposeful computationally costly procedures with minimal understanding of model findings. “Black box” models do not specify how they function and typically generate higher default predictions when compared to other models, so there is a tradeoff between model precision and explainability. However, a central tenet of the banking risk management approach is to avoid “black box” models. Financial institutions must be able to offer borrowers adequate grounds for rejection, and regulators must be able to grasp the economic reasoning behind these models. Still, a large portion of algorithmic decision-making is still a “black box,” fiercely safeguarded and unreachable to the public. This research maximized credit default prediction for a financier, by predicting future payment default of each credit card customer. A credit default prediction model was created using monthly customer profiles. The supervised learning methodology used a decision tree classifier and predictions were made using XGBoost. The model had an 84% accuracy rate for making predictions, with an AUC of 0.935. Furthermore, this study offered a current assortment of peer-reviewed scholarships, as well as an algorithm that presented interpretable results for banking stakeholders.