Self-Evolving AI Agents for Financial Risk Prediction Using Continual Learning and Neuro-Symbolic Reasoning

Akash Vijayrao Chaudhari; Pallavi Ashokrao Charate

Authors

Akash Vijayrao Chaudhari Senior Associate, Santander Bank, Florham Park, NJ, USA Author
Pallavi Ashokrao Charate Senior Systems Analyst, Worldpay, Cincinnati, OH, USA. Author

Keywords:

Adaptive AI, Continual Learning, Neuro-Symbolic Reasoning, Financial Risk Prediction, Explainable AI

Abstract

The financial services industry demands AI systems that are both adaptive to changing data patterns and capable of delivering transparent, explainable decisions. Traditional machine learning models used for financial risk prediction often degrade in performance due to data drift and are viewed as black boxes, raising concerns over fairness and regulatory compliance. In this paper, we propose a self-evolving AI agent that unifies continual learning techniques with neuro-symbolic reasoning to enable accurate, adaptive, and interpretable financial risk prediction. The agent employs Elastic Weight Consolidation and memory replay to update itself incrementally without catastrophic forgetting. A symbolic reasoning module encodes expert rules to provide logical overrides and explanations, ensuring compliance with domain policies. We demonstrate the system’s efficacy on credit risk prediction tasks, showing that it outperforms static and retrained models under data drift while offering consistent, rule-based justifications for its decisions. This combination of adaptability and interpretability makes our approach well-suited for high-stakes, evolving environments in financial decision-making.

References

Chaudhari, A. V., & Charate, P. A. (2024). Data Warehousing for IoT Analytics. International Research Journal of Engineering and Technology (IRJET), 11(6), 311–320.

Chaudhari, A. V., & Charate, P. A. (2025). Autonomous AI Agents for Real-Time Financial Transaction Monitoring and Anomaly Resolution Using Multi-Agent Reinforcement Learning and Explainable Causal Inferences. International Journal of Advance Research, Ideas and Innovations in Technology (IJARIIT), 11(2), 142–150.

Chaudhari, A. V. (2025). AI-powered alternative credit scoring platform. ResearchGate. https://doi.org/10.13140/RG.2.2.13191.92325

Chaudhari, A. V. (2025). Policy-driven federated cloud data warehouse for finance. ResearchGate. https://doi.org/10.13140/RG.2.2.13191.92325

Chaudhari, A. V. (2025). A cloud-native unified platform for real-time fraud detection. ResearchGate. https://doi.org/10.13140/RG.2.2.19902.80962

Lumenova AI (2025). Model Drift: Types, Causes and Early Detection. (Blog). – Explains concept drift and how models lose accuracy as data evolves, with examples of drift in credit risk.

Nwafor, C.N. et al. (2024). Enhancing transparency and fairness in automated credit decisions: an explainable novel hybrid machine learning approach. Scientific Reports, 14, 17532. – Discusses the transparency issues of black-box ML in credit risk and uses SHAP for explainability.

Kirkpatrick, J. et al. (2017). Overcoming catastrophic forgetting in neural networks. PNAS, 114(13):3521-3526. – Introduces Elastic Weight Consolidation (EWC) for continual learning to protect old knowledge.

Robins, A.V. (1995). Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146. – Early approach to continual learning using rehearsal and pseudo-rehearsal to prevent forgetting.

Singh, R. (2023). Neuro-symbolic AI: AI with reasoning. (Finextra Blog). – Outlines the integration of neural networks with symbolic reasoning and the analogy to System 1 (intuition) and System 2 (logic) thinking.

SmythOS (2023). Symbolic AI in Finance: Transforming Risk Management and Decision-Making. (Industry Article). – Describes applications of symbolic and neuro-symbolic AI in finance, noting the benefits for interpretability and compliance with domain rules.

Yeh, I-C. & Lien, C-H. (2009). The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients. Expert Systems with Applications, 36(2):2473-2480. – Introduces the UCI credit card default dataset and compares various models’ accuracy in predicting default.