Temporal Stability of Explainable AI in Financial Fraud Detection: A Framework for SHAP Auditing Under Drift

Abstract

Explainable artificial intelligence is increasingly used to support the review and governance of financial fraud models, yet a single explanation snapshot cannot by itself establish that a model remains interpretable, trustworthy, or reliable as transaction patterns evolve after deployment. This article develops a framework for auditing the temporal stability of feature-attribution explanations under concept drift and the severe class imbalance characteristic of fraud detection, and tests it empirically. It distinguishes local from global explanations, predictive attribution from causal explanation, and a genuine shift in explanation content from an unrelated shift in predictive performance, and it proposes a five-stage audit: a leakage-safe reference window, a documented and held-constant explainer configuration, comparison of explanation distributions across monitoring windows, triangulation against input drift and mature-label performance, and structured human review. The audit is implemented on the SHAP-guided, leakage-safe pipeline of Zuberi et al., applied to the public European credit-card dataset, with an XGBoost model and a logistic-regression baseline monitored across eight three-hour windows. Explanation change alone was an unreliable signal: global importance rankings for XGBoost fell well below their noise floor (Spearman 0.99) in windows 1, 2, 3, 4, but performance declined in only 3 of those, and a decline in window 8 (PR-AUC 0.20) occurred while the global ranking stayed stable (Spearman 0.90). Restricting the audit to confirmed fraud cases exposed that decline, with median attributions for V4 and V14 shifting by -5.2 and -2.2. The logistic baseline had more stable explanations and stronger day-two performance. Segment analysis by transaction amount showed recall gaps that aggregate figures concealed, although segment fraud counts were small.