Fairness-Aware Credit Default Prediction: Addressing Educational Bias in Machine Learning Models

Machine learning algorithms are increasingly adopted in the finance industry to predict credit default risk. While more models critically achieve strong predictive performance, the potential bias regarding some sensitive demographic attributes, such as education, perpetuates in the prediction-making process. Thus, in this study, a credit default prediction model built on Extreme Gradient Boosting (XGBoost) is not only implemented to achieve solid performance, but also investigated on mitigating education-related bias through several fairness interventions. Three approaches are explored: dataset-based by massaging, model-based by reweighing, and outcome-based by threshold adjustment, each deployed at different stages of the machine learning pipeline. The primary goal is to reduce the disparity in true positive rates (TPR) across groups while maintaining high balanced accuracy. Experimenting with parameters from the classifier and the hyperparameters from the dataset and outcomes, results show that all three interventions significantly improve fairness, with the best configuration reducing the TPR gap to below 0.01 and preserving predictive performance. This work demonstrates how fairness-aware interventions can be integrated into real-world financial decision-making systems.