HedgeAgents:

A Balanced-aware Multi-agent Financial Trading System

Abstract

As automated trading gains traction in the financial market, algorithmic investment strategies are increasingly prominent. While Large Language Models (LLMs) and Agent-based models exhibit promising potential in real-time market analysis and trading decisions, they still experience a significant -20% loss when confronted with rapid declines or frequent fluctuations, impeding their practical application. Hence, there is an imperative to explore a more robust and resilient framework. This paper introduces an innovative multi-agent system, HedgeAgents, aimed at bolstering system robustness via "hedging" strategies. In this well-balanced system, an array of hedging agents has been tailored, where HedgeAgents consist of a central fund manager and multiple hedging experts specializing in various financial asset classes. These agents leverage LLMs' cognitive capabilities to make decisions and coordinate through three types of conferences. Benefiting from the powerful understanding of LLMs, our HedgeAgents attained a 70% annualized return and a 400% total return over a period of 3 years. Moreover, we have observed with delight that HedgeAgents can even formulate investment experience comparable to those of human experts.

Overall Framework

The HedgeAgents framework simulates the architecture of a real hedge fund company, aiming to optimize risk hedging for multi-asset investment portfolios.The framework comprises four agents: Bitcoin Analyst Dave, Dow Jones Analyst Bob, Forex Analyst Emily, and Hedge Fund Manager Otto. Each of the three analysts is in charge of managing a specific asset. Otto, as their supervisor, is responsible for the overall risk management of the investment portfolio and the distribution of the asset investment budget. To achieve effective collaboration, we have established three types of multi-agent coordination meetings: Budget Allocation Conference, Experience Sharing Conference and Extreme Market Conference. These conferences serve to facilitate budget allocation, experience summary, and swift emergency actions.

Our HedgeAgents comprise 3 hedging agents and 1 manager. Each agent is equipped with 23 tools and possesses 3 types of memory to execute 8 actions. Furthermore, collaboration among multiple agents can be categorized into three types of conferences: budget allocation, experience sharing, and extreme market conference.

Leaderboard on Multi-Asset Financial Dataset

Leaderboard for evaluating various methods on 9 metrics in the Multi-Asset Financial Dataset.

# Models Categories Source ARR(%) TR(%) SR CR SoR MDD(%) Vol(%) ENT ENB
1 HedgeAgents 🥇 LLM-based 👑️ Link 71.60 405.34 2.41 11.02 58.00 14.21 1.30 3.13 1.53
2 FinAgent 🥈 LLM-based 👑️ Link 53.54 261.98 1.80 4.52 39.12 28.24 1.42 2.85 1.41
3 FinMem 🥉 LLM-based 👑️ Link 47.67 221.99 1.20 4.02 25.42 32.39 2.16 1.99 1.25
4 AlphaMix+ RL-Based 🎮 Link 37.59 160.47 1.62 3.69 35.52 25.56 1.17 2.93 1.22
5 FinGPT LLM-based 👑️ Link 34.22 141.82 1.93 7.64 39.57 17.08 8.76 1.76 1.33
6 DeepTrader RL-based 🎮 Link 32.78 134.11 1.41 4.06 30.43 20.95 1.21 2.02 1.30
7 SAC RL-based 🎮 Link 24.71 93.94 1.16 3.12 23.15 21.56 1.16 1.62 1.14
8 TSM Rule-Based 🛠️ Link 19.13 69.09 0.78 1.53 18.21 39.14 1.55 1.10 1.09
9 MV Rule-Based 🛠 Link 13.03 44.39 0.71 1.25 16.14 32.04 1.13 1.09 1.02
10 ZMR Rule-Based 🛠 Link -7.25 -20.21 -0.52 -3.13 -5.15 61.52 1.98 1.55 1.11
🚨Evaluation Metrics: ARR: Annual Return Rate, TR: Total Return, SR: Sharpe Ratio, CR: Calmar Ratio, SoR: Sortino Ratio, MDD: Maximum Drawdown, Vol: Volatility, ENT: Entrop, ENB: Effect Number of Bets.

🚨 The Multi-Asset Financial Dataset comprising Bitcoin, foreign exchange, and the Dow Jones component stocks. These data were sourced from reputable financial databases, namely Yahoo Finance and the Alpaca News API. The dataset spans from January 1, 2015, to December 31, 2023, encompassing daily data points such as open, high, low, and close prices, as well as volume and adjusted close prices. Additionally, daily news updates and 60 standard technical analysis indicators are included for each asset .


Cumulative Returns Comparison of our HedgeAgents and all baselines

Visualization

To thoroughly examine our framework's cognitive processes during execution, we have employed visualizations to elucidate both single-agent and multi-agent collaborative processes. Surprisingly, our model demonstrates the capacity to valuable experiences akin to human experts.

Detailed and comprehensive workflow visualization of a single agent, taking Bitcoin Analyst Dave as a research example.

Detailed visualization of the Budget Allocation Conference.

Detailed visualization of the Experience Sharing Conference.

Detailed visualization of the Extreme Market Conference.


The following content is dedicated to the graphical representation of the collective insights and strategies derived from the investment experiences of the three subordinate agents within the HedgeAgents framework.

Agent

The following content provides an exhaustive outline of the HedgeAgents team member profiles, which are integral to our investment decision-making simulation. The profiles are articulated using XML, chosen for its flexibility and robustness in structuring and representing complex data. XML's self-descriptive nature and the ability to define custom tags make it an ideal choice for encoding the intricate details of each agent's profile. It allows for a high degree of customization and scalability, which is essential for simulating a dynamic and complex environment such as a hedge fund's investment strategy.