DOI: https://doi.org/10.36347/sjet.2026.v14i11.001

Pages: 654-686

The rapid digitalization of banking and FinTech systems has intensified the need for artificial intelligence that is accurate, secure, explainable, and regulation-compliant. Conventional deep learning models deliver strong predictive performance but remain opaque black boxes, while Large Language Models (LLMs) offer powerful unstructured-data understanding yet suffer from hallucination and weak constraint satisfaction. Neither paradigm alone satisfies the transparency and compliance demands of modern financial regulation. This paper proposes the Explainable Neuro-Symbolic Trust Layer (XNSTL), a unified framework integrating LLM-based neural learning with symbolic rule reasoning, financial ontologies, and a regulatory audit mechanism to serve four critical banking tasks: fraud detection, credit risk assessment, financial forecasting, and transaction security. The XNSTL architecture comprises four modules: (i) an LLM-based feature extraction module for unstructured financial data; (ii) a symbolic rule engine encoding AML/KYC, Basel III, and institutional credit policies; (iii) a neuro-symbolic fusion module with confidence calibration and uncertainty quantification; and (iv) an explainability and audit layer generating counterfactual explanations and tamper-proof audit trails. We formalize the four tasks, define a hybrid loss function jointly optimizing predictive accuracy, symbolic constraint satisfaction, and explanation faithfulness, and propose an evaluation protocol using public benchmarks including IEEE-CIS Fraud Detection and Lending Club datasets. The framework is designed to reduce false positives in fraud detection, improve fairness in credit decisions, strengthen forecasting robustness under regime shifts, and enable real-time, audit-ready transaction security, establishing a pathway toward trustworthy and regulation-native financial AI.