Abstract
Credit card fraud detection is a difficult applied machine learning problem. It combines extreme class imbalance, temporal non-stationarity, and a sharp cost gap between missed fraud and false alarms. This paper presents an end-to-end experimental framework for fraud detection on the benchmark European transaction dataset (284,807 transactions; fraud prevalence 0.173%). A strict no-data-leakage protocol is enforced throughout. The data are split chronologically into training (70%), validation (15%), and test (15%) sets, and every preprocessing step - feature scaling, SHapley Additive exPlanations (SHAP)-based feature selection, and oversampling - is fitted only on the training partition. Two domain-informed features are engineered from the raw timestamp (sinusoidal hour encoding and log-transformed amount), and SHAP analysis reduces the 33-dimensional feature space to 15 features. Six oversampling strategies - SMOTE, BorderlineSMOTE, SVMSMOTE, ADASYN, SMOTEENN, and SMOTETomek - are compared across 12 classical classifiers, 3 multilayer perceptron (MLP) architectures, a purpose-built deep neural network (FraudNet), and 7 ensemble methods, giving 85 model-sampler combinations. Decision thresholds are tuned on the validation set using the [Formula: see text]-score, and all final metrics are reported on the held-out temporal test set. To characterise temporal drift explicitly, we measure distributional shift between splits using the Population Stability Index (PSI), Kolmogorov-Smirnov (KS) tests, and Jensen-Shannon divergence on the SHAP-selected features. We also report 1000-replicate bootstrap 95% confidence intervals for the leading configurations. The MLP (128-64-32) without oversampling reaches the highest individual [Formula: see text] of 0.7722 (95% CI: [0.6712, 0.8420]). The Soft Voting ensemble attains the best Matthews Correlation Coefficient (MCC) of 0.8060 (95% CI: [0.7045, 0.8807]) and an AUC of 0.9703. LightGBM under SMOTEENN shows the largest gain from oversampling, with [Formula: see text] rising from 0.0745 to 0.7588. ADASYN consistently underperforms, and no single oversampling method dominates across all model families. The drift analysis confirms measurable but modest distributional shift between splits. Because the dataset spans only 48 hours, this shift reflects short-horizon, mostly intra-day variation rather than the long-horizon concept drift seen in production; we therefore treat the chronological protocol as a methodological lower bound and emphasise this limit throughout. Together, these findings give practical guidance for designing production-grade fraud detection systems under strict temporal and data-integrity constraints.