A Four-Algorithm Unsupervised Voting Ensemble for Debit Card Fraud Detection: Alert Burden Reduction and Threshold Calibration

Fraud detection in Nigerian deposit money banks is constrained by the near-total absence of labelled fraud repositories, which excludes supervised learning for most institutions and leaves unsupervised anomaly detection as the only deployable alternative. Yet no study has evaluated a heterogeneous four-algorithm voting ensemble on a card-fraud benchmark, or quantified the alert-burden consequences of consensus voting. Isolation Forest, Local Outlier Factor, One-Class Support Vector Machine and a deep Autoencoder were trained exclusively on legitimate transactions and their binary votes combined by a three-of-four majority rule. On a held-out partition of 42,559 transactions containing 71 fraud cases, the ensemble attained the highest F1-score (0.2437), the highest precision (0.1566) and the lowest false positive rate (0.0049) of all five models evaluated, reducing false alerts by 90.3% relative to the highest-recall constituent and lowering the false alarm ratio from 36.8:1 to 5.4:1. McNemar testing confirmed the F1 advantage as significant against three of the four constituents, with bootstrap confidence intervals excluding zero for all four comparisons. Both the precision advantage and the alert-burden reduction replicated on 831,123 African mobile-money transactions, where the F1 advantage was significant against all four constituents at p below 0.001. Threshold ablation across all four consensus rules showed that three-of-four is F1-optimal on the card benchmark but not on the mobile-money corpus, where unanimity achieved an F1-score of 0.3749. The optimal consensus threshold is therefore a property of the deployment data rather than of the architecture, which argues for per-deployment calibration rather than transfer of a published value.