NeSy-RegReason: A Neuro-Symbolic AI Framework for Interpretable and Adaptive Regulatory Compliance in Financial Services

Abstract

Financial compliance systems must identify suspicious behaviour while also showing which legal obligations, facts, and exceptions support each decision. Purely neural models learn complex transaction patterns but usually provide post-hoc explanations that do not constitute legal reasoning; conventional rule engines are inspectable but brittle under behavioural and regulatory change. This study presents NeSy-RegReason, a neuro-symbolic framework combining Transformer and graph neural prediction, a versioned regulatory knowledge graph, OWL/SWRL/Prolog inference, rule-conflict resolution, counterfactual explanation, and an immutable audit trace. Evaluation uses author-supplied results across IEEE-CIS, PaySim, Elliptic, and a 500,000-record synthetic multi-regulation dataset. NeSy-RegReason attained 95.86% accuracy, 0.956 F1-score, 0.986 ROC-AUC, and 0.918 Matthews correlation coefficient. It achieved 98.31% compliance detection with 1.89% false positives, 0.97 explanation fidelity, 99.12% audit-trace completeness, and 98.54% conflict resolution. Wilcoxon comparisons against five principal baselines were significant (p ≤ .002; effect sizes 0.68–0.83). Ablation results show that symbolic reasoning is the most consequential component for explainability and auditability. The findings indicate that compliance performance should be evaluated as a multi-objective problem spanning prediction, rule satisfaction, explanation fidelity, trace reproducibility, and adaptation latency. Because the supplied aggregate results do not include run-level observations, code, or an external replication package, the statistical claims are reported as supplied and require independent reproduction before operational deployment.