Cost-Sensitive Top-K Review for Capacity-Constrained Fraud Screening
Abstract
Fraud models rank transactions by estimated risk, but review teams must decide how many cases to open under limited capacity. We formulate this downstream decision as a capacity-constrained top-<inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> policy problem and separate model fitting, validation, policy selection, and future evaluation in time, freezing every deployable queue before future labels enter the analysis. Under each external ceiling, an exact integer scan of the closed policy period determines the historical review rate. Policy-period Platt calibration and the cost cutoff determine the calibrated-threshold queue, which we compare with fixed review targets and the two capacity extremes. After outcomes mature, we measure regret against a retrospective constrained oracle, show that its workload is weakly nonincreasing in the cost-coefficient ratio, and report exact near-optimal membership, boundary diagnostics, and same-<inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> RandomK cost reduction. Experiments with logistic regression, LightGBM, and TabNet cover two IEEE-CIS temporal settings and a lower-prevalence ULB setting across five seeds. At 3% capacity, five-seed median historical-transfer regret ranged from 0.023% to 0.298%, while <inline-formula> <tex-math notation="LaTeX">$K_{\max }$ </tex-math></inline-formula> belonged to every exact future near-optimal set, so the tight ceiling compressed differences among workload rules. At 10%, the six medians ranged from 8.071% to 13.304% for full-capacity review and from 25.572% to 62.642% for no review. Across fixed 5%, historical transfer, and calibrated thresholding, the range was 0.219%–2.891%. The lower-regret candidate changed across models and temporal settings. This temporal audit separates ranking value from workload choice and helps review teams compare rules and revisit them when costs, capacity, or scored populations change.