Fintech Transaction Fraud Decision: Quantifying Loss, Friction, and Detection Risk Under Drift, Adversarial Adaptation, and Operational Latency

Abstract

This article presents an engineering-oriented framework that models fraud prevention as an end-to-end reliability system and quantifies how model uncertainty, drift, and decision latency propagate into distributional outcomes relevant to operations and governance, including fraud capture rate, false decline rate, expected net loss, customer friction cost, review workload, and time-to-decision. A scenario-based quantitative study is developed for card-not-present and account-to-account style transactions across normal and disruption regimes, comparing four architectures: baseline rules with static thresholds, ML scoring without calibration or governance, calibrated ML with cost-sensitive thresholds, and a governance-optimized two-tier architecture that combines calibrated risk scoring, uncertainty-aware routing to manual review, step-up verification for medium-risk cases, and dynamic thresholding under drift detection. Results show that ungoverned model deployment can reduce fraud loss but increase friction and false declines during drift, that calibrated thresholds improve stability but remain vulnerable when review capacity saturates, and that a two-tier governed approach reduces net loss while stabilizing customer impact and operational workload, particularly during fraud surges. Three copy-ready tables and complete prompts for data-driven figures are provided for Techne submission.