HedgeAgents: A Balanced-aware Multi-agent Financial Trading System

In one sentence

Summary

The paper addresses the fragility of existing LLM-based trading agents, which suffer significant losses during market volatility, by introducing HedgeAgents, a multi-agent framework designed for robust hedging. The system mimics a hedge fund structure with a central fund manager and specialized experts for Bitcoin, stocks, and forex. These agents utilize LLMs for cognitive processing, employing a reflection-driven decision-making pipeline that integrates memory retrieval, reinforcement learning-based action selection, and self-reflection updates. The core innovation lies in three coordination mechanisms: Budget Allocation Conferences for periodic portfolio rebalancing, Experience Sharing Conferences for knowledge accumulation, and Extreme Market Conferences for emergency risk mitigation during high volatility. Experiments were conducted on a dataset spanning 2015-2023, comprising daily prices, volumes, and news for Bitcoin, DJIA components, and forex pairs. The testing period (2021-2023) evaluated performance against rule-based, reinforcement learning, and other LLM-based baselines using metrics like Total Return, Sharpe Ratio, and Maximum Drawdown. The system achieved a 70% annualized return and 400% total return over three years, significantly outperforming state-of-the-art models. Ablation studies confirmed the necessity of all three conference types, with the Extreme Market Conference crucial for limiting drawdowns. The framework also demonstrated robustness across different LLM backbones, with GPT-4 yielding the best results, while maintaining low operational costs. HedgeAgents excels in risk-adjusted returns, achieving the lowest Maximum Drawdown (14.21%) and highest diversification metrics among all tested models. It effectively navigates extreme market conditions, such as the May 2022 downturn, where baselines failed. The system generates interpretable investment experiences stored in memory, comparable to human expert insights. However, limitations include reliance on specific LLM capabilities, potential data leakage risks mitigated by temporal isolation tests, and the assumption that historical correlations hold. The system is primarily a proof-of-concept for multi-agent coordination in finance, with real-world deployment caveats regarding transaction costs and slippage not fully modeled.

Research problem

Existing LLM-based trading agents lack robustness, suffering significant losses (up to -20%) during market declines and fluctuations. The paper aims to develop a resilient multi-agent system that incorporates hedging strategies to balance risk and return, addressing the lack of risk management mechanisms in current automated trading frameworks.

Core contributions

• Introduction of HedgeAgents, a multi-agent system integrating hedging strategies with LLM-based cognitive capabilities.
• Design of a hierarchical agent structure with specialized experts (Bitcoin, Stocks, Forex) and a central fund manager.
• Development of three coordination mechanisms: Budget Allocation, Experience Sharing, and Extreme Market Conferences.
• Demonstration of superior performance across 9 financial metrics, including 400% total return and low drawdown.
• Ablation studies validating the synergistic effect of each conference module on portfolio robustness and returns.

Method

• Multi-agent architecture with specialized asset experts and a fund manager.
• LLM-driven decision making with memory retrieval (text-embedding-3-large).
• Budget Allocation Conference for portfolio weighting optimization using VaR and covariance.
• Experience Sharing Conference for consolidating investment insights into general memory.
• Extreme Market Conference for emergency risk mitigation during high volatility.
• Reflection-driven reinforcement learning loop for policy optimization.
• Tool integration for technical analysis, news processing, and risk assessment.

Results & findings

• HedgeAgents achieved 71.60% Annualized Return and 405.34% Total Return over 3 years.
• Sharpe Ratio of 2.41, outperforming all baselines.
• Maximum Drawdown of 14.21%, significantly lower than baselines (e.g., FinAgent 28.24%).
• Extreme Market Conference reduced drawdown by 71.93% when active.
• Budget Allocation Conference contributed most to return growth.
• GPT-4 backbone yielded best performance, but open-source models like Qwen-72B were competitive.
• Operational cost was low, averaging 2 cents per day over 3 years.
• System maintained positive returns during May 2022 market downturn while baselines failed.

Data & experiments

• Dataset: Daily prices, volumes, and news for Bitcoin, DJIA stocks, and forex from 2015-2023.
• Baselines: Rule-based (MV, ZMR, TSM), RL-based (SAC, DeepTrader, AlphaMix+), and LLM-based (FinGPT, FinMem, FinAgent).
• Metrics: Total Return, Annual Return Rate, Sharpe Ratio, Calmar Ratio, Sortino Ratio, Maximum Drawdown, Volatility, Entropy, Effect Number of Bets.
• Ablation: Removal of individual conference modules to assess impact on risk and return.
• Sensitivity: Testing different LLM backbones (GPT-4, Gemini, Qwen, etc.) to evaluate backbone dependency.

Limitations & caveats

• Relies on specific LLM capabilities; performance varies with backbone.
• Assumes historical correlations hold for hedging strategies.
• Does not fully model transaction costs and slippage in real-world trading.
• Potential data leakage risks mitigated but not eliminated.
• Limited to Bitcoin, Stocks, and Forex; may not generalize to other assets.
• Operational costs, while low, may scale with more complex strategies.

Curator notes

Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking; Will LLMs be Professional at Fund Investment? DeepFund: A Live Arena Perspective