Answers are grounded in this paper and cite their sources.
Credit card default prediction is a mature application of machine learning in consumer finance, yet published studies typically stop at discrimination metrics without connecting a model’s output to the economics of a specific intervention a bank could actually deploy. This paper builds an interpretable default-risk model using the public UCI Default of Credit Card Clients dataset (Yeh and Lien, 2009) and couples it to a cost-benefit framework built from cited, real-world figures. A logistic regression baseline achieves test ROC-AUC of 0.741, and a random forest comparison model achieves 0.780, on a cohort of 30,000 Taiwanese credit card clients with a 22.12% next-month default rate. Rather than a single flat dollar loss, this paper estimates avoided loss per client from the client’s own most recent bill amount multiplied by a published average loss-given-default rate for credit cards (61.54%; Bellotti & Crook, 2012). Against this, two realistic intervention channels are costed: an SMS reminder ($0.03) and a live-agent call ($2.70, reused from a companion study), with effectiveness drawn from a randomized field experiment on credit card debt repayment (a 4.4-percentage-point reduction in delinquency; Bursztyn, Fiorin, Gottlieb, & Kanz, 2019). Under these figures, the break-even cost per intervention for untargeted, blanket outreach is $457.45, far above both realistic channel costs, so model-based targeting adds negligible net value over contacting every client at either cost. This paper argues the finding is not a property of the models but of the ratio between typical credit card exposure sizes and low-cost communication channels, and discusses what would need to change for precision targeting to matter in this setting.
No takes yet. Share an insight, caveat, or question.
Aryan Patre (2026) studied this question.