Maturity-Aware Dual-Timescale Replay for Delayed-Feedback Financial Fraud Screening
Fraud-screening labels often arrive after a transaction, with fraudulent outcomes confirmed later than legitimate outcomes. This class-dependent observation process distorts conventional replay under concept drift. We instantiate the adaptive recent/global prioritized replay of DriftShield in a one-step cost-sensitive fraud controller, and extend it with maturity-gated access, transaction-time recency and capped inverse confirmation-visibility weighting. A second, arrival-driven scorer supplies a faster adaptation channel; its predictions are combined with the slow replay scorer using a fixed mixture. The system is tested prequentially on an explicitly synthetic 42,000-event stream with two changes in fraud signatures and asymmetric reporting delays. In the three-seed comparison, 42 model/delay runs include three component ablations and two stress-test delays. An additional 12-seed paired evaluation reports F1 0.262 and average precision 0.236 for the dual-timescale controller, versus 0.239 and 0.223 for arrival-time replay. A fast-only scorer with the same threshold controller instead reaches F1 0.265 and average precision 0.242: the slow branch does not improve that stronger comparator in this simulator. Paired intervals and a future-label nonanticipation test delimit the reproducible benefits of maturity-aware replay without implying production deployment gains.
Authors
- Kiran Tran
- Ryan White
- David Gupta
Institutions
- Florida State University (US)
Publication Details
- Journal
- International Journal of World Economic Research
- Published
- 2026-11-18
- DOI
- https://doi.org/10.54097/ypkpr569
- Primary Topic
- Data Stream Mining Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00