Adaptive and Explainable Machine Learning for Temporal Drift-Aware Financial Fraud Detection and Risk Monitoring: A Review

Abstract

Financial fraud detection is a moving target: fraud tactics, legitimate consumer behavior, and the statistical relationship between transaction features and fraud risk all shift over time, so a model that is accurate on the data it was trained on can quietly degrade in production without triggering any alarm built around static accuracy alone. This review synthesizes six sources to examine how the literature is responding to that problem along three fronts named in this review's title: adaptivity, explainability, and explicit drift-awareness. A systematic survey anchors the review's broadest claim, identifying class imbalance, concept drift, adversarial behavior, and explainability as the field's persistent, interlinked deployment barriers. Two applied studies evaluate models on the same public benchmark (284,807 European credit-card transactions, 0.17% fraud rate) and are therefore directly comparable: a soft-voting ensemble combined with explicit drift quantification (via the Population Stability Index, Kolmogorov-Smirnov testing, and Jensen-Shannon divergence) reaches an AUC of 0.9703, while a SHAP-explained Random Forest on the same data reaches AUC = 0.9766. A third applied study, built and validated on five years of real transaction data from Gulf Cooperation Council banks, combines drift detection, adversarial defense, and human-in-the-loop explainability in a single deployed framework, reporting a fraud-recall improvement from 35% to 85% and a reduction in adversarial attack success from 35% to 5%, with drift recovery within 24 hours and operational latency under 150 milliseconds. A fourth study addresses adaptivity most directly, comparing instance-incremental and batch-incremental streaming learning strategies and translating technical feature importance into investigator-facing concept explanations, though without a single reported benchmark accuracy figure this review could independently verify. A documented, non-peer-reviewed industry account of fraud-model failure during the COVID-19 pandemic's abrupt shift in consumer spending is included for real-world grounding, disclosed explicitly as not being an academic source. Read together, these sources suggest that adaptivity, explainability, and drift-awareness are not competing design goals but three views of the same underlying requirement: a fraud detection system whose statistical foundations are continuously checked against the population it is currently scoring, not just the population it was trained on.