A Survey on Multi Agent Framework for Asset Allocation and Portfolio Optimization

Abstract

The rapid evolution of global financial markets, driven by increased data availability, market volatility, and the participation of diverse investors, has intensified the complexity of asset allocation and portfolio optimization. Traditional portfolio management techniques, while well established, rely on static assumptions and centralized decision-making, limiting their effectiveness in dynamic and non-stationary market environments. Recent advances in Artificial Intelligence (AI), particularly Reinforcement Learning (RL) and Multi-Agent Systems (MAS), have opened new avenues for adaptive, decentralized, and intelligent financial decision-making. This review paper presents a systematic survey of Multi-Agent Artificial Intelligence frameworks for dynamic asset allocation and portfolio optimization. It synthesizes and analyzes recent research spanning multi-agent reinforcement learning, deep reinforcement learning, large language model (LLM)-based agents, and hybrid AI architectures applied to financial markets. The review examines agent cooperation and competition mechanisms, learning paradigms, reward formulations, and risk-aware optimization strategies employed across existing studies. Additionally, emerging trends such as hierarchical multi-agent learning, role-based agent specialization, and LLM-driven financial reasoning are discussed. By critically evaluating the strengths, limitations, and practical challenges of current methodologies, this paper identifies key research gaps related to scalability, interpretability, transaction cost modeling, and real-time deployment. The survey aims to provide a comprehensive understanding of the current landscape of multi-agent AI in portfolio management and to serve as a foundation for future research toward robust, adaptive, and intelligent financial systems.

Introduction

The text reviews the growing role of Artificial Intelligence (AI) in asset allocation and portfolio optimization amid increasing financial market complexity, volatility, and data intensity. Traditional portfolio management methods, such as mean–variance optimization and rule-based strategies, are limited by simplifying assumptions, poor adaptability to non-linear dynamics, and susceptibility to human bias. As a result, AI-driven approaches have emerged to support more adaptive, data-driven investment decisions.

Early research primarily focused on deep learning models—such as LSTM, RNN, CNN, ensemble methods, and Transformers—to improve stock price prediction. While these models effectively capture temporal and complex market patterns, they are largely limited to forecasting and do not support dynamic portfolio rebalancing or sequential decision-making.

Reinforcement Learning (RL) marked a shift toward interactive and adaptive portfolio management by enabling agents to learn optimal trading and allocation strategies through market interaction. Although RL-based approaches improved adaptability and profitability, most were single-agent systems and often relied on unrealistic assumptions, such as ignoring transaction costs, liquidity constraints, and risk factors.

To address scalability and realism, recent studies have adopted Multi-Agent Reinforcement Learning (MARL), which better reflects decentralized financial markets with multiple interacting decision-makers. MARL frameworks—both cooperative and competitive—demonstrate improved diversification, robustness, and risk-adjusted returns. However, many focus on short-term trading or require significant computational resources.

More recently, Large Language Models (LLMs) and agent-based systems have been integrated into financial decision-making to enhance interpretability, reasoning, sentiment analysis, and explainability. Hybrid systems combining MARL, deep learning, and LLM-based agents show promise but still face challenges related to computational efficiency, real-time deployment, and precise portfolio weight optimization.

Conclusion

This review paper presented a comprehensive analysis of recent advancements in Artificial Intelligence–based approaches for dynamic asset allocation and portfolio optimization, with particular emphasis on reinforcement learning, multi-agent systems, and agent-oriented financial frameworks. The surveyed literature demonstrates a clear transition from traditional deep learning models focused on stock price prediction toward more adaptive, decision-driven methodologies capable of operating in highly volatile and non-stationary financial environments. Early deep learning approaches such as LSTM, CNN, ensemble models, and Transformer-based architectures significantly improved forecasting accuracy by capturing complex temporal dependencies in financial time-series data. However, these models were primarily limited to predictive tasks and lacked the ability to perform sequential decision-making or adaptive portfolio rebalancing. Reinforcement learning methods addressed this limitation by enabling agents to directly learn optimal trading and allocation strategies through interaction with market environments, thereby improving adaptability and profitability. More recent studies employing multi-agent reinforcement learning (MARL) have shown that decentralized, cooperative, and competitive agent interactions can enhance robustness, scalability, and risk-adjusted returns. Additionally, the integration of Large Language Models (LLMs) into agent-based financial systems has introduced new capabilities in interpretability, strategic reasoning, and role-based specialization. Despite these promising developments, existing approaches often rely on simplified market assumptions, incur high computational costs, or lack unified frameworks that simultaneously address scalability, explainability, and real-world trading constraints. Overall, this review highlights that while AI-driven financial systems have achieved substantial progress, there remains a need for holistic and practical solutions that can operate reliably in real-world financial markets.

References

[1] B. Jaswanth and J. Kaushik, “Stacked LSTM: A Deep Learning Model to Predict Stock Market,” Proceedings of the 4th International Conference on Advances in Computing, Communication Control and Networking (ICAC3N), IEEE, 2022, pp. 17–22. [2] K. R. Reddy, B. T. Kumar, V. R. Ganesh, P. Swetha, and P. K. Sarangi, “Stock Market Prediction Using Recurrent Neural Network,” IEEE International Conference on Current Development in Engineering and Technology (CCET), 2022, pp. 1–6. [3] R. Abivarshini and K. France, “Stock Market Price Prediction Using Deep Learning,” Proceedings of the 3rd International Conference on Inventive Computing and Informatics (ICICI), IEEE, 2025, pp. 243–247. [4] J. Y. Lee and S. J. Yoo, “Stock Price Prediction Using Transformer and Time2Vec,” International Conference on Artificial Intelligence in Information and Communication (ICAIIC), IEEE, 2025, pp. 687–692. [5] A. M. Pattanayak, B. Sahoo, and A. Swetapadma, “A Deep Reinforcement Learning Technique for Stock Price Prediction and Trading Decision,” International Conference on Cognitive Robotics and Intelligent Systems (ICC-ROBINS), IEEE, 2025, pp. 266–271. [6] P. Govindasamy, G. V. Radhakrishnan, and U. Shankar, “High-Frequency Stock Market Price Prediction Using Blockchain and Deep Learning,” International Conference on Advances in Computer Science, Electrical, Electronics and Communication Technologies (CE2CT), IEEE, 2025, pp. 259–264. [7] L. Busoniu, R. Babuska, and B. De Schutter, “A Survey on Multi-Agent Reinforcement Learning and Its Applications,” IEEE Transactions on Systems, Man, and Cybernetics, Part C, vol. 38, no. 2, pp. 156–172, 2008. [8] Z. Jiang, D. Xu, and J. Liang, “MAPS: Multi-Agent Reinforcement Learning-Based Portfolio Management System,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 31, no. 1, pp. 1–8, 2017. [9] Y. Li, J. Zheng, and Y. Yang, “Multi-Agent Reinforcement Learning for High-Frequency Trading Strategy Optimization,” IEEE Access, vol. 9, pp. 145602–145615, 2021. [10] X. Liu, Z. Wang, and H. Chen, “TradingAgents: Multi-Agents LLM Financial Trading Framework,” arXiv preprint arXiv:2412.20138, 2024. [11] A. R. Hajaghaie and R. K. Thulasiram, “AI Agents in Finance and Fintech: A Scientific Review of Agent-Based Systems, Applications, and Future Horizons,” IEEE Symposium on Computational Intelligence for Financial Engineering and Economics (CIFEr), 2025. [12] A. R. Hajaghaie and R. K. Thulasiram, “Leveraging Large Language Models and Retrieval-Augmented Generation for Enhanced Multi-Asset Portfolio Construction,” IEEE Symposium on Computational Intelligence for Financial Engineering and Economics (CIFEr), 2025.

Copyright

Copyright © 2026 Dr. D. S. Zingade, Aditya Pawar, Khushal Sharma, Anish Raut. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.