Adaptive Conformal Risk Control for Payment Fraud Detection under Concept Drift and Delayed Feedback

Payment fraud engines turn risk scores into approve-or-review decisions, and the quantity an operator must keep within budget is the fraud value released to auto-approval. Two features of production systems undermine static guarantees on that quantity: the fraud mix drifts, and the chargebacks that reveal fraud arrive weeks after the decision, so a transaction without a chargeback is not yet known to be legitimate. We present an adaptive conformal risk controller that keeps an exposure-capped, amount-stratified and value-ordered conformal calibration layer as a monotone family of policies and steers one scalar, the nominal level of that layer, by integral feedback on chargebacks. The feedback is progressive: every approved transaction contributes inverse-probability-of-censoring-weighted increments at several label ages, which telescope to its true monetary loss when the dispute window closes. This yields a distribution-free certificate on the long-run released fraud value that holds under arbitrary drift and under a misspecified delay distribution, whereas naive and single-stage early feedback converge to biased levels. On five 24-month Sparkov streams with a camouflaged-fraud drift scenario and on the 183-day ULB/Worldline handbook stream, static calibration released 2.63 times the budget and monthly recalibration 1.55 times, while the proposed controller held 1.01 ± 0.02 times, close to an oracle with instantly revealed labels. When chargeback delays grew by 50% after deployment, it stayed at 1.08 times, against 1.48 times for single-stage correction.