Predicting credit card defaults

: a comparative analysis of machine learning models

Abstract

In a time of rising interest rates and economic difficulties, the accurate prediction of credit card defaults is essential for financial institutions. In this master thesis, different machine learning models are evaluated to improve the predictive performance compared to traditional methods. Models such as XGBoost, Random Forest, K-Nearest Neighbor, Neural Networks and Support Vector Machines are developed and compared to the Logistic Regression. Performances are evaluated based on accuracy, precision, recall and F1-score and compared to their computation time as a cost factor. The learning curves are then analyzed to identify possible over- or underfitting. Finally, the importance of each feature is reviewed to gain an understanding of which data banks should include in their analysis. The results show that all advanced machine learning models outperform the Logistic Regression, with XGBoost coming out on top. However, KNN lags behind the other models. Key features identified include payment status and closeness to credit limit. Therefore, this study underlines the potential of advanced machine learning models to improve the prediction of credit card defaults, but also highlights the need for institution specific analysis due to data variability and computational complexity.

Date of Award	17 Oct 2024
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Francesco Rotondi (Supervisor)