Dynamic Cost-Sensitive Fraud Detection for Real-Time Financial Transactions: A Budget-Aware Three-Way Execution Framework

Abstract

Internet-finance platforms must decide, for every arriving transaction and within milliseconds, whether to approve it, challenge it, or decline it. Recent cost-sensitive work has shown that mapping a calibrated fraud probability and the transaction amount to a three-way execution action is far more profitable than thresholding a risk score. Deployed systems, however, operate under two constraints that this line of work does not model: the number of customers a platform may inconvenience is capped by operational capacity, and both the score distribution and the fraud prior drift while labels arrive only after a verification delay. We present DyCoDe, a dynamic cost-sensitive execution framework that addresses both. First, an amount-proportional false-decline cost generalises the constant-cost model and reshapes the decision geometry. Second, the three-way policy is augmented with an online dual price on intervention capacity, updated by a proportional-integral controller, so that the realised friction tracks a prescribed budget. Third, all adaptation—prior tracking, calibrator refitting, and scorer refresh—is gated by a calibration-residual CUSUM designed for extreme imbalance and by a validated-promotion test, so the system moves only when evidence warrants it. On 144,807 real card transactions evaluated prequentially with delayed labels, DyCoDe attains the cost of the best static three-way policy while using 43.1% less customer friction and halving hard declines, and it costs 23.0–48.0% less than a clairvoyant static threshold operating at the same friction, coming within 0.6% of the offline optimal dual price. On a controlled drifting stream it lowers cost by 13.5% over a non-adapting pipeline at equal friction while retraining 43% less often than a fixed schedule. Scoring, calibration and policy evaluation take 0.583 ms per transaction on a single core.