Comparative Evaluation of Post-Hoc Feature Attribution Methods on Tabular Financial Data: Faithfulness, Stability, and Computational Efficiency

The deployment of machine learning in credit scoring and fraud detection has intensified regulatory and societal demand for transparent decision-making. Post-hoc feature attribution methods such as SHAP, LIME, Integrated Gradients, and Anchors promise to explain individual predictions, yet their comparative reliability on financial tabular data remains insufficiently characterized. This study conducts a controlled empirical evaluation of four prominent attribution methods across four public financial datasets spanning credit scoring and transaction fraud detection. Three classifiers---XGBoost, Random Forest, and Multilayer Perceptron---serve as the underlying predictive functions. Explanation quality is quantified along three axes: faithfulness measured by Prediction Gap on Important features and infidelity, stability measured by max-sensitivity, and computational efficiency measured by wall-clock time per explanation. Results indicate that TreeSHAP achieves the highest faithfulness and lowest sensitivity on tree-based classifiers, while Integrated Gradients attains competitive faithfulness on neural networks. LIME exhibits the largest variance across repeated runs, raising concerns for regulatory settings that require reproducible explanations. Anchors produce the sparsest explanations at the cost of reduced faithfulness. No single method dominates all evaluation criteria simultaneously, corroborating recent theoretical predictions of an inherent trade-off among explanation desiderata. These findings provide practitioners and regulators with empirically grounded guidance for selecting attribution methods in financial applications.