SHAP and LIME Explainability in Financial Machine Learning: A Systematic Review of Methods, Evaluation Gaps, and Barriers to Deployment

<title>Abstract</title> <p>Financial institutions rely heavily on Artificial Intelligence (AI). They use it for high-stakes decisions like credit scoring, fraud detection, algorithmic trading, risk assessment, and insurance pricing. Model complexity often results in black-box systems. Customers, auditors, managers, and regulators cannot easily inspect how these models make decisions, which creates problems for trust, fairness, accountability, and regulatory compliance. Explainable AI (XAI) tackles these problems. It offers tools to clarify model behaviour while keeping predictive performance high.In this systematic literature review, we synthesize empirical research on post-hoc XAI in finance, specifically focusing on SHAP and LIME. We analyze the literature across four areas. We look at how techniques map to tasks, whether generated explanation outputs undergo explicit evaluation, what technical and data-related barriers block deployment, and what broader validation strategies researchers use.We followed PRISMA 2020 guidelines and software engineering systematic review protocols to search Scopus, Web of Science, and IEEE Xplore in May 2026. Our search yielded 4,168 records. After removing duplicates, we screened 3,010 records and sought 325 reports for retrieval. We assessed 312 full texts. The final corpus includes 297 peer-reviewed studies. Every included study directly applies SHAP, LIME, or both to a financial machine learning task.Classification tasks account for the vast majority of our sample (219 studies, 73.7%). The most common financial domain is credit, loan, and default modelling (140 studies, 47.1%). Grouped SHAP variants appear in 279 studies (93.9%). This massive presence results directly from TreeSHAP’s native integration into gradient-boosted tree frameworks like XGBoost, LightGBM, and CatBoost, rather than representing all XAI methods in finance.We found that only 27 studies (9.1%; roughly 5.8% to 12.4%) report evidence of explanation quality. Just 21 studies (7.1%; roughly 4.2% to 10.0%) conduct a formal evaluation. Evaluation is clearly the weakest aspect of current SHAP and LIME applications in finance. The application of post-hoc XAI in finance is expanding fast. Yet, to reliably support regulated financial decision-making, the field needs much stronger validation, clearer reporting standards, and concrete evidence of deployment viability.</p>