A Cost-Sensitive Method for Credit Card Default Prediction

Yiqi Zhou

2025

Abstract

With an increasing number of credit cards being issued, financial institutions must predict credit card default to minimize bad debts and remain profitable. Traditional machine learning approaches typically emphasize maximizing overall accuracy but neglect the higher cost of missing actual defaulters (false negatives, FNs). This research introduces a cost-sensitive framework that involves robust data cleaning, multi-model comparisons, and threshold tuning. Based on a University of California, Irvine (UCI) dataset containing 30,000 records and 25 features, the procedure merges anomalous categories, removes duplicates, and standardizes skewed numeric columns. Benchmarking three popular algorithms—Logistic Regression, Random Forest, and eXtreme Gradient Boosting (XGBoost)—reveals that XGBoost attains the highest Area Under the Curve (AUC) and recall for credit defaulters. Building on these findings, the XGBoost decision threshold is further adjusted by assigning heavier penalties to FNs than to false positives (FPs). Experimental results indicate that without such an intervention, potential bad debt nears 877,900 (or 11.58 percent of the total), whereas the cost-sensitive approach reduces it to approximately 432,400 (or 5.64 percent), highlighting the limits of raw accuracy metrics. This paper concludes with a discussion of interpretability, advanced hyperparameter tuning, and deployment considerations in real-world finance, reflecting data up to October 2023.

Paper Citation

in Harvard Style

Zhou Y. (2025). A Cost-Sensitive Method for Credit Card Default Prediction. In Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE; ISBN 978-989-758-765-8, SciTePress, pages 516-520. DOI: 10.5220/0013701200004670

in Bibtex Style

@conference{icdse25,
author={Yiqi Zhou},
title={A Cost-Sensitive Method for Credit Card Default Prediction},
booktitle={Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE},
year={2025},
pages={516-520},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013701200004670},
isbn={978-989-758-765-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE
TI - A Cost-Sensitive Method for Credit Card Default Prediction
SN - 978-989-758-765-8
AU - Zhou Y.
PY - 2025
SP - 516
EP - 520
DO - 10.5220/0013701200004670
PB - SciTePress