Model Drift and Recalibration in Fraud Detection Models for Financial Reconciliation
Journal of Contemporary Science and Technology Management · 30.12.2025 · 1 (04) · 510-540
DOI: 10.66592/jht58edb ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Contemporary Science and Technology Management, online erschienen am 30.12.2025. Verfasst von Carol Kpobie und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Fraud detection models trained on historical data are used against populations that evolve. In financial reconciliation, the evolution is complicated by label delay, severe class imbalance, and model-ranked investigation. This paper categorizes five drift types by cause, characteristic, and solution, and develops a monitoring portfolio for each. A constant logit shift affects probabilities, counts and thresholds but leaves the ranking unchanged, and therefore does not affect precision of top K selections. Prevalence shift multiplies precision at a constant threshold by the prior odds ratio. A recalibration ladder prioritizes base-rate correction, threshold adjustment, monotone calibration and re-training by their cost and ability to change the drifts. Model-based investigation induces a feedback loop that makes observed yield go up even as precision goes down, with the difference determined by the rate at which non-adverse items are labeled as adverse. Thus, we propose a protocol that uses an explicitly unranked audit sample. With a prevalence of 0.4 per cent, ninety random samples per period yield an expectation of 0.36 adverse items, making sample size too small. The simulation is described as a protocol without simulated results.
Zitieren
APA
Kpobie, C., Odihi, F., Osahor, D. U., & Okoli, I. E. (2025). Model Drift and Recalibration in Fraud Detection Models for Financial Reconciliation. Journal of Contemporary Science and Technology Management, 1(04), 510-540. https://doi.org/10.66592/jht58edb
Vancouver
Kpobie C, Odihi F, Osahor DU, Okoli IE. Model Drift and Recalibration in Fraud Detection Models for Financial Reconciliation. Journal of Contemporary Science and Technology Management. 2025;1(04):510-540. doi:10.66592/jht58edb
BibTeX
@article{Kpobie2025,
author = {Kpobie, Carol and Odihi, Friday and Osahor, Deborah Ujeh and Okoli, Ifeoma E.},
title = {Model Drift and Recalibration in Fraud Detection Models for Financial Reconciliation},
journal = {Journal of Contemporary Science and Technology Management},
year = {2025},
volume = {1},
number = {04},
pages = {510--540},
doi = {10.66592/jht58edb}
}
EUVIMED