TradingAgents: Multi-Agents LLM Financial Trading Framework
Yijia Xiao, Edward Sun, Di Luo, Wei Wang
q-fin.TR, cs.AI, cs.CE, cs.LG
2024-12-28
TradingAgents casts seven LLM agents as a trading firm and beats all five quant baselines on a 3-month backtest of three megacap tech stocks, posting 23%+ cumulative returns.
LLM-based trading has tried two routes. One treats the model as a single tool, say a news-sentiment scorer bolted onto a conventional quant rule. The other strings several agents together and lets them talk freely to reach a decision. The second sounds more sophisticated but suffers an old flaw: when agents converse in natural language over many turns, information degrades like a game of telephone, and early critical details get buried in chat logs. TradingAgents asks whether you can harden a real trading firm's division of labor into structure, so each agent owns a job and information does not have to survive long conversations.
The framework splits a trading firm into five stages and seven roles. The Analyst Team has four members: fundamental, sentiment, news, and technical. The sentiment analyst gets web search plus Reddit and X retrieval and a sentiment-scoring routine; the technical analyst runs code and computes sixty indicators such as MACD, RSI, and Bollinger Bands. Each produces a structured report.
Next comes the Researcher Team: one bullish, one bearish, debating in natural language for n rounds, where the facilitator agent sets n and then reads the transcript to adopt the prevailing side. This is the one place the paper keeps free-form language, because debate is about clashing arguments, not passing data.
A Trader takes the researchers' verdict and decides timing and position size. A Risk Management Team re-examines the plan through three profiles (risk-seeking, neutral, conservative). A Fund Manager approves and executes.
Two design choices stand out. First, agents communicate through structured documents and diagrams by default, borrowing from MetaGPT; natural language is reserved for debate. This targets the telephone effect directly, since reports preserve the essentials and filter noise, so information stops leaking across turns. Second, every agent follows the ReAct prompting framework (reason, then act through tools). Backbone models split by task weight: gpt-4o-mini and gpt-4o handle fast work like summarization, data retrieval, and table-to-text, while o1-preview handles reasoning-heavy work like decisions and evidence-based reports. No GPU is needed, only API credits.
The backtest runs January 1 to March 29, 2024, on Apple, Google, and Amazon (the text also names Nvidia, Microsoft, and Meta, but the main table covers three). Data covers prices, news (Bloomberg, Yahoo, EODHD, FinnHub, Reddit), social sentiment (Reddit, X), insider transactions (SEDI), and filings, with sixty technical indicators per ticker. Only data available up to each trading day is used, ruling out look-ahead bias. Baselines: Buy and Hold, MACD, KDJ and RSI, ZMR, and SMA.
Four metrics (cumulative return CR, annualized return AR, Sharpe SR, max drawdown MDD):
| Ticker | TradingAgents CR | Best baseline CR | TradingAgents SR | MDD |
| AAPL | 26.62% (AR 30.5%) | KDJ&RSI 2.05% (B&H -5.23%) | 8.21 | 0.91% |
| GOOGL | 24.36% (AR 27.58%) | B&H 7.78% | 6.39 | 1.69% |
| AMZN | 23.21% (AR 24.90%) | B&H 17.1% | 5.60 | 2.11% |
The headline is at least 23.21% cumulative and 24.90% annualized. The authors stress that on the most volatile ticker, AAPL, conventional patterns broke down while TradingAgents cleared 26%. Versus the strongest baseline per ticker, cumulative returns improved by 24.57, 16.58, and 6.10 percentage points.
Two takeaways for practitioners. As engineering, this is a template for configurable financial software: backbone models, data sources, and role counts are pluggable, swapping a model needs no retraining, and adding a macro analyst role touches nothing else. The structured-communication idea generalizes past trading to any multi-agent system losing information to long chats. The other takeaway is honesty: the authors do not pitch this as a money printer ready for production, and repeat that this is a backtest with live trading left as future work.
This section is the most informative, because the authors throw cold water first. A Sharpe ratio of 2 is usually good and 3 excellent; here SR runs 5.60 to 8.21, and the authors say outright this exceeds their expected empirical range, attributing it to very few pullbacks during the period, meaning the strategy barely lost money. They exported the full decision trace to verify the arithmetic, confirming the numbers are correct without denying they look implausible.
Three concerns remain. The sample is small: three megacap tech stocks over three months during a Q1 2024 bull run, and one bear ticker could flip the conclusion. Cost forces that scale, since each prediction needs eleven LLM calls plus twenty-plus tool calls, enough budget for only a three-month backtest, so longer histories and more tickers stay untested. The paper also does not model transaction costs or slippage; with MDD as low as 0.91%, fees and bid-ask spread could eat a meaningful share of the headline return. Until live data exists, treat the 23% as evidence the architecture works, not as a return forecast.