Beyond Discrimination: Calibration- and Cost-Aware Evaluation of Consumer Credit Default Models

DOI

10.2991/978-94-6239-787-3_45How to use a DOI?

Keywords

Credit Default Prediction; Probability Calibration; Cost-Sensitive Learning; Model Evaluation; Financial Risk Management

Abstract

Consumer-credit default models are routinely selected on discriminatory power, usually the area under the ROC curve (AUC). Yet a lender acts on a threshold, whose quality depends on whether predicted probabilities are calibrated and on the asymmetric cost of approving a defaulter versus rejecting a good applicant. We argue, and show empirically, that discrimination alone is an inadequate basis for model choice in credit scoring. Using a controlled simulation whose documented data-generating process uniquely exposes each applicant's true default probability, we compare five learners on 30,000 applicants with a 20.5% default rate. Their AUC values are nearly identical (a 0.015 spread; the logistic scorecard and gradient boosting differ by 0.0008), yet their calibration differs by more than a factor of thirty (expected calibration error 0.007 to 0.238). Because the cost-optimal threshold for calibrated scores equals the cost ratio, the naive 0.5 threshold inflates expected cost by almost a half. Post-hoc calibration of the interpretable scorecard cuts its calibration error by over 90% and its Brier score by 31% at negligible AUC cost, restoring an economically meaningful threshold. Every finding is reproduced on the real UCI Default of Credit Card Clients benchmark (30,000 borrowers): models within 0.006 AUC differ in calibration error by a factor of seventeen, and the naive threshold inflates cost by a third. Credit-model evaluation should therefore report calibration and expected cost alongside discrimination, and post-hoc calibration is a practical safeguard for scorecards.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Zhanpeng Wang
PY  - 2026
DA  - 2026/09/29
TI  - Beyond Discrimination: Calibration- and Cost-Aware Evaluation of Consumer Credit Default Models
BT  - Proceedings of the 2026 4th International Conference on Management Innovation and Economy Development (MIED 2026)
PB  - Atlantis Press
SP  - 462
EP  - 468
SN  - 2352-5428
UR  - https://doi.org/10.2991/978-94-6239-787-3_45
DO  - 10.2991/978-94-6239-787-3_45
ID  - Wang2026
ER  -