DAUNT: Ensemble Disagreement as Actionable Uncertainty for Imbalanced Fraud Detection

Detecting a few hundred fraudulent transactions among hundreds of thousands is an extreme class-imbalance problem where one miss can cost a full transaction value. Stacking heterogeneous classifiers is the standard recipe, yet under a leakage-free, precision–recall evaluation, its ranking gain over the best single model is inconsistent: sizable when the base learners are diverse, and negligible when they are redundant. The ensemble’s dependable value lies elsewhere: member disagreement is a usable, threshold-free estimate of epistemic uncertainty. We propose Daunt (Disagreement As UNcertainty for Triage), which turns this disagreement into three deployment layers over one ensemble: (i) routing the most uncertain transactions to human review by ranking them on base-learner disagreement, (ii) deciding alarms by an example-dependent rule that weighs the fraud score against the transaction amount, and (iii) attaching interpretations verified for faithfulness and stability. On two contrasting datasets, the anonymized ULB (0.17% fraud) and feature-rich IEEE-CIS (3.5% fraud), deferring the 5% most uncertain transactions raises system recall from 0.76 to 0.91 (ULB) and 0.66 to 0.75 (IEEE-CIS) while removing every automated false alarm, and the example-dependent rule recovers more fraudulent money than any global threshold. Daunt reframes the heterogeneous ensemble as a source of actionable uncertainty rather than an end in itself.