Agentic AI-Based Decision Support Architecture for Financial Risk Management

Abstract

Financial risk management is an important task in banking, and At the institutional level, where decisions are accurate and transparent and have direct consequences on the institution. stability. Traditional machine learning models have been shown to perform well on Less ability to grasp the more intricate, non-linear financial patterns data and rarely interpret results in the domain practitioners. This paper introduces FinRisk.ai, an agentic Multi-agent collaborative - AI framework, integrating deep learning in a multi-agent collaborative setting The pipeline was used for the automated financial risk prediction. The framework contributes: (1) a sequential feature construction method which Reshapes static table credit reports into temporal windows in them. Each (2) a is the LSTM, a leakage-safe crossvalidation protocol used within each. The features’ stability during the formal definition of the explainability metrics.Stability of features within the formal definition of explainability metrics. The Index (FSI) and Explanation Fidelity Score (EFS) are included in a (3) an autonomous inter-agent communication and (4) a dedicated XAI agent. Use of a Judge Agent for protocol with conflict resolution. A Human-in-the-Loop (HITL) module enables domain experts to accept, reject, or request reevaluation of any prediction. Ethical privacy-related factors such as GDPR compliance, fairness auditing and Basel III alignment are addressed. The system is evaluated on the Both datasets were applied to the same approach.Both the German Credit Dataset and the Credit Card Fraud Detection Dataset were used in the same manner. The data was split into five nonoverlapping sections and the test was repeated three times, giving a total of 15 splits. runs). The LSTM achieves a mean accuracy of 93.7%(±0.8%) and AUC of 0.963(±0.007) on credit risk prediction, beating XGBoost by 6.5TabNet by 4.0%. The LSTM has an accuracy of 99.2A maximum accuracy of 0.987 was obtained for the AUC.