The List / Agents and Multi-Agent Systems / paper
TradingAgents: Multi-Agents Large Language Models for Financial Trading
Abstract
Significant progress has been made in automated problem-solving using societies of agents powered by large language models (LLMs). In finance, efforts have largely focused on single-agent systems handling specific tasks or multi-agent frameworks independently gathering data. However, the multi-agent systems' potential to replicate real-world trading firms' collaborative dynamics remains underexplored. TradingAgents proposes a novel stock trading framework inspired by trading firms, featuring LLM-powered agents in specialized roles such as fundamental analysts, sentiment analysts, technical analysts, and traders with varied risk profiles. The framework includes Bull and Bear researcher agents assessing market conditions, a risk management team monitoring exposure, and traders synthesizing insights from debates and historical data to make informed decisions. By simulating a dynamic, collaborative trading environment, this framework aims to improve trading performance. Detailed architecture and extensive experiments reveal its superiority over baseline models, with notable improvements in cumulative returns, Sharpe ratio, and maximum drawdown, highlighting the potential of multi-agent LLM frameworks in financial trading. TradingAgents is available at https://github.com/TauricResearch/TradingAgents.
In one sentence
Summary
The paper addresses the limitation of existing LLM trading systems that lack realistic organizational modeling and suffer from inefficient natural language communication interfaces. It proposes TradingAgents, a framework inspired by professional trading firms, featuring specialized agents including fundamental, sentiment, news, and technical analysts, alongside bullish and bearish researchers, traders with varied risk profiles, and a risk management team. The system employs a hybrid communication protocol combining structured reports for clarity and natural language debates for reasoning, utilizing both quick-thinking and deep-thinking LLMs to balance efficiency and analytical depth. Experiments were conducted via backtesting on major technology stocks (Apple, Nvidia, Microsoft, Meta, Google) from January to March 2024, using a multi-modal dataset comprising historical prices, news, social media sentiment, insider transactions, financial statements, and 60 technical indicators. The framework was benchmarked against Buy-and-Hold, MACD, KDJ+RSI, ZMR, and SMA strategies. Evaluation metrics included cumulative return, annualized return, Sharpe ratio, and maximum drawdown, with agents making decisions based solely on available data to prevent look-ahead bias. TradingAgents significantly outperformed all baselines, achieving cumulative returns of 26.62% for Apple, 24.36% for Google, and 23.21% for Amazon, surpassing the best rule-based baselines by margins of 6-24%. It demonstrated superior risk-adjusted performance with higher Sharpe ratios and controlled maximum drawdowns, despite rule-based methods sometimes showing lower drawdowns. The debate mechanism and structured communication effectively mitigated context loss and hallucination, providing explainable trading decisions. However, the study is limited to a short backtesting period on large-cap tech stocks, and the reliance on API-based LLMs incurs computational costs, while real-world deployment faces challenges like transaction fees and market impact not fully captured in the simulation.
Research problem
Current LLM-based financial agents often fail to replicate the collaborative dynamics of real-world trading firms and rely on inefficient natural language communication that leads to context loss. This paper addresses the need for a structured, multi-agent framework that mimics professional organizational workflows to improve trading performance and explainability.
Core contributions
- Proposes TradingAgents, a multi-agent LLM framework simulating professional trading firm structures with specialized roles.
- Introduces a hybrid communication protocol combining structured reports and natural language debates to reduce context corruption.
- Implements a risk management team with varied risk profiles (aggressive, neutral, conservative) to optimize portfolio exposure.
- Demonstrates superior backtesting performance against rule-based baselines on major tech stocks using multi-modal data.
- Provides an open-source implementation allowing plug-and-play strategy integration and model exchangeability.
Method
- Specialized agents: Fundamental, Sentiment, News, Technical Analysts, Bull/Bear Researchers, Traders, Risk Managers, Fund Manager.
- Hybrid communication: Structured reports for data exchange, natural language for debates.
- ReAct prompting framework for reasoning and acting.
- Model selection: Quick-thinking models for retrieval, deep-thinking models for analysis and decision-making.
- Risk management: Three perspectives (risky, neutral, safe) to adjust trading plans.
- Fund Manager approval workflow for final trade execution.
Results & findings
- TradingAgents achieved 26.62% cumulative return on Apple, outperforming the best baseline by 24.57%.
- Google and Amazon returns were 24.36% and 23.21% respectively, surpassing baselines by 16.58% and 6.10%.
- Sharpe ratios were significantly higher than all baselines, indicating superior risk-adjusted performance.
- Maximum drawdown was controlled effectively, balancing high returns with risk management.
- The debate mechanism between bullish and bearish researchers improved decision robustness.
- Structured communication reduced information loss compared to pure natural language dialogue.
Data & experiments
- Backtesting simulation from January 1 to March 29, 2024, on Apple, Nvidia, Microsoft, Meta, and Google.
- Multi-modal dataset including historical prices, news, social media sentiment, insider transactions, financial statements, and 60 technical indicators.
- Benchmarks against Buy-and-Hold, MACD, KDJ+RSI, ZMR, and SMA strategies.
- Evaluation metrics: Cumulative Return, Annualized Return, Sharpe Ratio, and Maximum Drawdown.
- No look-ahead bias ensured by using only data available up to each trading day.
Limitations & caveats
- Backtesting period is short (3 months) and limited to large-cap tech stocks.
- Does not account for transaction costs, slippage, or market impact in simulations.
- Reliance on API-based LLMs incurs significant computational costs.
- Real-world deployment may face challenges not captured in the simulation.
- Performance may vary across different market conditions and asset classes.
Curator notes
Relevant to agentic investment committee workflows.