TradingAgents: Multi-Agents LLM Financial Trading Framework

Reading guide

Problem and motivation

Traditional quantitative trading algorithms often struggle to synthesize unstructured textual data such as news headlines, corporate earnings reports, and social media sentiment. Meanwhile, deep learning trading models operate as black boxes, suffering from low interpretability. While Large Language Models (LLMs) offer natural language understanding and explainability, previous LLM-based financial systems typically deploy single agents or unstructured multi-agent setups. These setups suffer from two major flaws: they lack the organizational modeling of actual trading firms, and they rely on open-ended natural language chat logs that degrade through a 'telephone effect,' where critical context is lost or corrupted over long horizons.

Key ideas

• Organizational Hierarchy: TradingAgents mirrors a Wall Street investment firm divided into specialized divisions: an Analyst Team, a Researcher Team, Trader Agents, a Risk Management Team, and an ultimate Fund Manager.
• Dialectical Multi-Agent Debates: Instead of unilateral decisions, the framework runs structured multi-round debates between Bullish and Bearish researchers, as well as multi-perspective discussions among aggressive, safe, and neutral risk managers.
• Hybrid Communication Protocol: The system avoids continuous, unstructured text pools by recording agent states as standardized, structured reports and data entries, reserving natural language dialogue solely for multi-turn debates.
• Model Tiering: Fast, lower-cost models (gpt-4o, gpt-4o-mini) handle data retrieval and formatting, while deep-thinking reasoning models (o1-preview) power in-depth analysis, debate, and trade sizing.

Method (with the key equations)

All agents operate using the ReAct (Reasoning and Acting) paradigm with access to specialized external tools (e.g., Yahoo Finance, EODHD, Finnhub, Reddit). Agent actions proceed in five sequential stages:

1. Analyst Team: Four concurrent agents generate structured reports across Company Fundamentals, Social Sentiment, News/Macroeconomics, and Technical Indicators (evaluating up to 60 indicators such as RSI, MACD, Bollinger Bands, ATR, ADX).
2. Researcher Team: Bullish and Bearish researcher agents query the analyst reports and debate for $n$ rounds under the supervision of a Facilitator Agent, who selects the prevailing stance.
3. Trader Execution: The Trader agent synthesizes researcher insights and market data to generate target buy/sell/hold decisions and rationale.
4. Risk Management Deliberation: Risk-seeking ('Risky'), conservative ('Safe'), and neutral analysts debate the portfolio exposure, leverage, and stop-loss needs over $n$ rounds.
5. Fund Manager Approval: The final manager approves, modifies, or rejects trade execution based on the risk team's synthesized guidance.

Portfolio performance is measured via four key financial metrics:

• Cumulative Return (CR): $$\text{CR}=\left(\frac{V_{\text{end}}-V_{\text{start}}}{V_{\text{start}}}\right)\times 100\%$$
• Annualized Return (AR): $$\text{AR}=\left({\left(\frac{V_{\text{end}}}{V_{\text{start}}}\right)}^{\frac{1}{N}}-1\right)\times 100\%$$ where $N$ is the evaluation period in years.
• Sharpe Ratio (SR): $$\text{SR}=\frac{\bar{R}-R_f}{\sigma }$$ where $\bar{R}$ is the average portfolio return, $R_f$ is the risk-free rate, and $\sigma$ is the return volatility.
• Maximum Drawdown (MDD): $$\text{MDD}=\underset{t\in [0,T]}{\max }\left(\frac{{\text{Peak}}_t-{\text{Trough}}_t}{{\text{Peak}}_t}\right)\times 100\%$$

Results (with the most important numbers)

The backtesting simulation ran from January 1 to March 29, 2024, without look-ahead bias across major tech equities ($AAPL, $GOOGL, $AMZN) against Buy & Hold and technical strategies (MACD, KDJ+RSI, ZMR, SMA):

• AAPL: TradingAgents achieved a cumulative return (CR) of 26.62% (ARR of 30.50%, Sharpe Ratio of 8.21, and MDD of 0.91%), whereas Buy & Hold yielded -5.23% (MDD 11.90%) and the best baseline (KDJ&RSI) gained only 2.05%.
• GOOGL: TradingAgents reached a CR of 24.36% and SR of 6.39 (MDD 1.69%), outperforming Buy & Hold (7.78% CR, 13.04% MDD).
• AMZN: TradingAgents achieved 23.21% CR and 5.60 SR (MDD 2.11%), beating Buy & Hold (17.10% CR, 3.80% MDD).
• Balanced Risk Control: Despite high return generation, maximum drawdowns remained strictly controlled under 2.11% across all stocks due to active intervention by the Risk Management Team.

Limitations and open questions

• High Computation/API Cost: Single-day predictions require roughly 11 LLM calls and over 20 tool calls, which limited the backtest simulation window to 3 months.
• Extreme Sharpe Ratios: The exceptionally high Sharpe ratios (>5.60 to 8.21) reflect a testing period with virtually no sharp pullbacks for the system, highlighting the need for multi-year testing across full market cycles and bear regimes.
• Execution Friction: The framework assumes clean signal fills without modeling slippage, market impact, exchange latency, or transaction commission drag in live order books.