Comparative Analysis of Fraud Detection Methods in Banking Using Machine Learning Techniques

Fraud detection in banking requires algorithms that balance classification performance, computational efficiency, and regulatory interpretability, criteria that are rarely benchmarked together. We present a comprehensive evaluation of nine machine learning approaches (Logistic Regression, Decision Tree, Random Forest, SVM, SGD, XGBoost, CatBoost, LightGBM, and MLP) across three datasets differing in size, imbalance severity, and feature type (synthetic, real-world PCA-anonymized, and large-scale simulated). Our protocol addresses four methodological gaps prevalent in the literature: (1) SMOTE applied strictly within cross-validation folds to prevent data leakage; (2) Systematic reporting of confidence intervals for all metrics; (3) Systematic inference latency profiling; and (4) SHAP-based interpretability analysis aligned with regulatory requirements. SHAP analysis provides model-agnostic feature attributions aligned with regulatory explainability requirements. Results show gradient boosting methods achieving superior fraud detection (CatBoost: F1 = 0.86 on real-world data) with sub-6ms inference, while SVM is disqualified for production use due to O(n2) latency scaling. This study provides reproducible baselines, with full hyperparameter specifications, to support algorithm selection in operational fraud prevention systems.