TradingAgents 是一个多智能体 LLM 金融交易研究框架，把复杂交易任务拆成多种角色：基本面、情绪、新闻、技术四类分析师并行产出报告，随后由看多与看空研究员展开结构化辩论，再由交易员决定方向与时点，风险管理和投资组合经理做最终把关。项目支持多家模型供应商以及不同模型层级混用，提供命令行与 Python 包两种用法、LangGraph 断点续跑、决策日志持久化，并强调时点一致性，使回测只能看到当日已公开的数据，避免前视偏差。适合做交易流程与智能体协作研究，官方明确说明不构成投资建议。
pip install .
--ticker and --date, the run's settings recorded in every report, and backtests that see only data published by each analysis date.Full release notes are in CHANGELOG.md.
Earlier news
TRADINGAGENTS_* env-var configurability with API-key auto-detection, remote Ollama support, non-US alpha benchmarks, and ticker path-traversal hardening.🚀 TradingAgents | ⚡ Installation & CLI | 🎬 Demo | 📦 Package Usage | 🤝 Contributing | 📄 Citation
🎉 TradingAgents officially released! We have received numerous inquiries about the work, and we would like to express our thanks for the enthusiasm in our community.
So we decided to fully open-source the framework. Looking forward to building impactful projects with you!
TradingAgents is a multi-agent trading framework that mirrors the dynamics of real-world trading firms. By deploying specialized LLM-powered agents: from fundamental analysts, sentiment experts, and technical analysts, to trader, risk management team, the platform collaboratively evaluates market conditions and informs trading decisions. Moreover, these agents engage in dynamic discussions to pinpoint the optimal strategy.
TradingAgents framework is designed for research purposes. Trading performance may vary based on many factors, including the chosen backbone language models, model temperature, trading periods, the quality of data, and other non-deterministic factors. It is not intended as financial, investment, or trading advice.
Our framework decomposes complex trading tasks into specialized roles.
The selected analysts work at the same time, each on its own tools, and the research debate starts once all of their reports are in.
Comprises both bullish and bearish researchers who critically assess the insights provided by the Analyst Team. Through structured debates, they balance potential gains against inherent risks.
Composes reports from the analysts and researchers to make informed trading decisions, determining the timing and magnitude of trades.
Continuously evaluates portfolio risk by assessing market volatility, liquidity, and other risk factors. The risk management team evaluates and adjusts trading strategies, providing assessment reports to the Portfolio Manager for final decision.
The Portfolio Manager approves/rejects the transaction proposal. If approved, the order will be sent to the simulated exchange and executed.
Clone TradingAgents:
git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
TradingAgents needs Python 3.11 or later. Create a virtual environment in any of your favorite environment managers:
conda create -n tradingagents python=3.13
conda activate tradingagents
Or with uv:
uv venv --python 3.13
source .venv/bin/activate
Install the package and its dependencies (uv pip install . with uv):
pip install .
Alternatively, run with Docker:
cp .env.example .env # add your API keys
docker compose run --rm tradingagents
After updating the repository, rebuild the image with docker compose build.
Results, reports, the memory log and the cache live in the tradingagents_data volume. To keep them in a folder on the host instead, create the folder and point TRADINGAGENTS_DATA_DIR at it, in .env or the shell: mkdir -p data && TRADINGAGENTS_DATA_DIR=./data docker compose run --rm tradingagents.
For local models with Ollama:
docker compose --profile ollama run --rm tradingagents-ollama
TradingAgents supports multiple LLM providers. Set the API key for your chosen provider:
export OPENAI_API_KEY=... # OpenAI (GPT)
export GOOGLE_API_KEY=... # Google (Gemini)
export ANTHROPIC_API_KEY=... # Anthropic (Claude)
export XAI_API_KEY=... # xAI (Grok)
export DEEPSEEK_API_KEY=... # DeepSeek
export DASHSCOPE_API_KEY=... # Qwen (international, dashscope-intl.aliyuncs.com)
export DASHSCOPE_CN_API_KEY=... # Qwen (China, dashscope.aliyuncs.com)
export ZHIPU_API_KEY=... # GLM via Z.AI (international)
export ZHIPU_CN_API_KEY=... # GLM via BigModel (China, open.bigmodel.cn)
export MINIMAX_API_KEY=... # MiniMax (global, api.minimax.io)
export MINIMAX_CN_API_KEY=... # MiniMax (China, api.minimaxi.com)
export OPENROUTER_API_KEY=... # OpenRouter
export MISTRAL_API_KEY=... # Mistral
export MOONSHOT_API_KEY=... # Kimi (Moonshot)
export GROQ_API_KEY=... # Groq
export NVIDIA_API_KEY=... # NVIDIA NIM
export FRED_API_KEY=... # FRED macro data (free, optional)
export ALPHA_VANTAGE_API_KEY=... # Alpha Vantage
export TYPESAFE_API_KEY=... # Jev social-post screening (optional)
For Azure OpenAI, copy .env.enterprise.example to .env.enterprise and fill in your credentials.
To run on your ChatGPT Plus or Pro plan instead of an API key, choose "Subscription" in the CLI's provider menu, or set llm_provider: "chatgpt". Sign in once with tradingagents login chatgpt: it prints an address to open in your browser, which returns to this machine on 127.0.0.1. When the browser runs on another machine, or TradingAgents runs in Docker (docker compose run --rm tradingagents login chatgpt), paste the address the browser ends up on. The sign-in is kept in ~/.tradingagents/chatgpt-auth.json, and tradingagents logout chatgpt revokes it. Runs count toward your plan's usage limits, and the plan does not accept temperature or max_tokens.
For AWS Bedrock, install the extra with pip install ".[bedrock]", set llm_provider: "bedrock", configure AWS credentials (environment variables, ~/.aws/credentials, or an IAM role) and AWS_DEFAULT_REGION, and use a Bedrock model ID, e.g. us.anthropic.claude-opus-5-5.
For local models, configure Ollama with llm_provider: "ollama". The default endpoint is http://localhost:11434/v1; set OLLAMA_BASE_URL to point at a remote ollama-serve. Pull models with ollama pull , and pick "Custom model ID" in the CLI for any model not listed by default.
For any other OpenAI-compatible server (vLLM, LM Studio, llama.cpp, or a custom relay), use llm_provider: "openai_compatible" and set the endpoint via backend_url (or TRADINGAGENTS_LLM_BACKEND_URL), e.g. http://localhost:8000/v1 for vLLM or http://localhost:1234/v1 for LM Studio. The model is whatever your server serves. No key is needed for local servers; set OPENAI_COMPATIBLE_API_KEY when the endpoint requires one.
With TYPESAFE_API_KEY set, the Sentiment Analyst screens StockTwits and Reddit posts with TypeSafe's Jev before reading them. Posts that are not about the company are dropped, and each source opens with a count of the remaining posts by stance: bullish, bearish, neutral, or unclear. Without the key, posts pass through unscreened. jev-latest moves with new releases; set TYPESAFE_DEFAULT_MODEL to a versioned ID such as jev-1.13.0 to hold it fixed across runs. To reach Jev through OpenRouter, put an OpenRouter key in TYPESAFE_API_KEY and set TYPESAFE_BASE_URL=https://openrouter.ai/api.
Alternatively, copy .env.example to .env and fill in your keys:
cp .env.example .env
Launch the interactive CLI:
tradingagents # installed command
python -m cli.main # alternative: run directly from source
You will see a screen where you can select your desired tickers, analysis date, LLM provider, research depth, and more. Your previous run's answers come back as the defaults, so pressing Enter accepts them. The TRADINGAGENTS_* variables in .env still skip their step entirely.
To run without questions, for a scheduled job or a script, answer the per-run steps with flags and the rest with TRADINGAGENTS_* variables:
export TRADINGAGENTS_LLM_PROVIDER=openai TRADINGAGENTS_QUICK_THINK_LLM=gpt-6-luna TRADINGAGENTS_DEEP_THINK_LLM=gpt-6-sol
export TRADINGAGENTS_OUTPUT_LANGUAGE=English TRADINGAGENTS_MAX_DEBATE_ROUNDS=1 TRADINGAGENTS_MAX_RISK_ROUNDS=1
tradingagents --ticker NVDA --date 2026-09-23 --analysts market,news,fundamentals --save --no-show
Each flag skips only its own question. Run without a terminal, a missing answer stops the run before it starts and names the flag or variable to set.
At start-up the interactive CLI mentions a newer release and notices from Tauric Research. It checks at most once a day, in the background, and sends nothing about your runs; set TRADINGAGENTS_NO_UPDATE_CHECK=1 to turn it off. Runs without a terminal never check.
A saved report also includes complete_report.html, the report as one page with its sections listed beside the text, for reading in a browser, on a phone or in print. Answering the save question at the prompt also asks about the page and can open it in your browser; --no-html skips it, and so does ta.save_reports(state, "NVDA", html=False) from Python.
TradingAgents works with any market Yahoo Finance covers, using the exchange-suffixed ticker. Company identity and the alpha benchmark resolve automatically per market.
US: AAPL, SPY
Hong Kong: 0700.HK · Tokyo: 7203.T · London: AZN.L
India: RELIANCE.NS, .BO · Canada: .TO · Australia: .AX
China A-shares: Shanghai .SS, Shenzhen .SZ (e.g. 600519.SS for Kweichow Moutai)
Crypto: BTC-USD, ETH-USD
An interface will appear showing results as they load, letting you track the agent's progress as it runs.
We built TradingAgents with LangGraph to ensure flexibility and modularity. The framework supports multiple LLM providers: OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen (Alibaba DashScope, international and China endpoints), GLM (Zhipu), MiniMax (global + China), OpenRouter, Ollama for local models, Azure OpenAI for enterprise, and your own ChatGPT plan.
To use TradingAgents inside your code, you can import the tradingagents module and initialize a TradingAgentsGraph() object. The .propagate() function will return a decision. You can run main.py, here's also a quick example:
from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG
ta = TradingAgentsGraph(debug=True, config=DEFAULT_CONFIG.copy())
# forward propagate
state, decision = ta.propagate("NVDA", "2026-09-01")
print(decision)
# the same report tree the CLI saves, under results_dir/reports
ta.save_reports(state, "NVDA")
You can also adjust the default configuration to set your own choice of LLMs, debate rounds, etc.
from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG
config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "openai" # e.g. openai, google, anthropic, deepseek, groq, ollama; openai_compatible covers any OpenAI-compatible endpoint (vLLM, LM Studio, llama.cpp, ...)
config["deep_think_llm"] = "gpt-6-sol" # Model for complex reasoning
config["quick_think_llm"] = "gpt-6-luna" # Model for quick tasks
config["max_debate_rounds"] = 2
ta = TradingAgentsGraph(debug=True, config=config)
_, decision = ta.propagate("NVDA", "2026-09-01")
print(decision)
The quick model serves the analysts, researchers, debaters and trader; the deep model serves the research and portfolio managers. Each can run on its own provider, for example the managers on Claude while the rest run on OpenAI:
config["deep_think_provider"] = "anthropic"
config["deep_think_llm"] = "claude-opus-5-5"
A tier on its own provider uses that provider's key and default endpoint; set quick_think_backend_url or deep_think_backend_url for a local or relay endpoint. The TRADINGAGENTS_DEEP_THINK_PROVIDER and TRADINGAGENTS_QUICK_THINK_PROVIDER variables set them for the CLI, together with the tier's model variable.
See tradingagents/default_config.py for all configuration options.
US company statements come from SEC EDGAR, which records the date every figure was filed. A run dated in the past reads the statements exactly as they stood that day: a fiscal year that has ended but has not been filed yet is not served, and a figure restated later still reads as first reported. Apple's 2008 total assets were filed as $39.6B and restated to $36.2B in 2010, so a run dated in between reads $39.6B. EDGAR needs no account or API key.
Other companies' statements come from Yahoo Finance, which dates a statement by the period it covers rather than by when it was published. A run dated today reads them; a run dated in the past is told they are withheld, since Yahoo cannot say which figures were public by then.
Insider trades are dated by when they happened, not when they were filed, so a run dated in the past is told they are withheld as well.
SEC asks callers to identify themselves and refuses requests that carry no contact address, so a default one is sent. Set your own so SEC can reach you rather than the project:
SEC_EDGAR_USER_AGENT="Your Name [email protected]"
It covers companies that file with the SEC, including foreign companies listed in the US. Anything else, such as Hong Kong or A-share listings, falls through to the next vendor in the chain. EDGAR's machine-readable filings begin in 2009, and a fourth quarter is reported as unavailable rather than derived, because filers publish it only inside the annual figure.
By default the agents do not know what you hold, so their guidance is written for a reader who applies it to their own position. Pass a portfolio to have the trader, the risk analysts and the portfolio manager work against your actual book.
from tradingagents.portfolio import PortfolioContext
portfolio = PortfolioContext.model_validate({
"cash": 25000.0,
"currency": "USD",
"positions": [{"ticker": "NVDA", "quantity": 120, "average_price": 150.0}],
})
_, decision = ta.propagate("NVDA", "2026-09-01", portfolio=portfolio)
The CLI takes the same content as a JSON file: tradingagents --portfolio my_book.json.
An empty positions list means a flat book, which is different from passing nothing. A run without a portfolio is never treated as flat.
TradingAgents persists two kinds of state across runs.
The memory log is always on. Each completed run appends its decision to ~/.tradingagents/memory/trading_memory.md. While the analysts of a later run work, TradingAgents settles every logged decision whose holding period has passed: it fetches the realised return (raw, and alpha against the instrument's regional benchmark) and generates a one-paragraph reflection. The Portfolio Manager then reads the most recent decisions for the same ticker plus recent lessons from other tickers, so each analysis carries forward what worked and what didn't. If settling fails, the run goes on and its report says so.
Override the path with TRADINGAGENTS_MEMORY_LOG_PATH.
To settle decisions without running an analysis, for a scheduled job, call ta.settle_all_pending(); it returns the decisions it settled and any it could not.
Checkpoint resume is opt-in via --checkpoint. When enabled, LangGraph saves state after each node so a crashed or interrupted run resumes from the last successful step instead of starting over. The run view says whether it resumed a saved run or started fresh. Checkpoints are cleared automatically on successful completion.
Per-ticker SQLite databases live at ~/.tradingagents/cache/checkpoints/.db (override the base with TRADINGAGENTS_CACHE_DIR). Use --clear-checkpoints to reset all of them before a run.
tradingagents --checkpoint # enable for this run
tradingagents --clear-checkpoints # reset before running
config = DEFAULT_CONFIG.copy()
config["checkpoint_enabled"] = True
ta = TradingAgentsGraph(config=config)
_, decision = ta.propagate("NVDA", "2026-09-01")
One run gives one decision, which cannot tell you whether the system decides well. run_backtest runs the same pipeline over a grid of tickers and dates, writes to a memory log of its own, and scores the decisions whose holding window has since traded.
from tradingagents.backtest import iter_grid, run_backtest, summarize
dates = iter_grid("2026-06-01", "2026-08-01", every_n_days=7)
result = run_backtest(["NVDA", "AAPL"], dates, config, selected_analysts=["market", "news"])
print(summarize(result).render())
From the CLI:
tradingagents backtest NVDA,AAPL --start 2026-06-01 --end 2026-08-01 --every 7
Each cell is scored on realized alpha against the instrument's regional benchmark, grouped by rating. A decision is entered at the first close after its date, because a run reads news published after that day's close. Your own memory log is never written to, and re-running the same grid with run_id=result.run_id skips the cells that already ran, so an interrupted sweep continues where it stopped.
TradingAgents is LLM-driven, so two runs of the same ticker and date can differ. This is expected for a research tool built on language models, not a defect. The variation comes from a few distinct sources, and it helps to separate them.
Language model sampling is non-deterministic. Even at a fixed temperature, providers do not guarantee byte-identical output across calls, and reasoning models (the default GPT-6 family, and any thinking-mode model) vary the most because their internal reasoning is itself sampled.
Live data moves. News, StockTwits, and Reddit return different content as time passes, so a run today sees different inputs than a run last week even for the same historical trade date. Pin the analysis date to hold the price and indicator window fixed, but the social and news sources still reflect "now".
To reduce variation you can lower the sampling temperature. Set temperature in your config (or TRADINGAGENTS_TEMPERATURE in .env); lower values make models that honor it more repeatable. The current curated models are reasoning-first and largely ignore temperature, so for tighter reproducibility name a non-reasoning model in your config, or in TRADINGAGENTS_DEEP_THINK_LLM and TRADINGAGENTS_QUICK_THINK_LLM. Any model ID your provider serves is accepted, whether or not the picker lists it.
config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "openai"
config["temperature"] = 0.0
# Reasoning models ignore temperature. For tighter reproducibility, name a
# non-reasoning model in deep_think_llm / quick_think_llm.
What does not vary anymore: the analyzed company identity is resolved deterministically from the ticker before any agent runs, and the market analyst grounds exact price and indicator claims in a verified data snapshot. Earlier reports of "different companies" or fabricated price levels across runs are addressed by these two mechanisms.
Backtest results are not guaranteed to match any published figure. Returns depend on the model, the temperature, the date range, data quality, and the sampling above. Treat the framework as a research scaffold for studying multi-agent analysis, not as a strategy with a fixed, replicable return.
Contributions are welcome: bug fixes, documentation, and feature ideas; past contributions are credited per release in CHANGELOG.md.
Please reference our work if you find TradingAgents provides you with some help :)
@misc{xiao2025tradingagentsmultiagentsllmfinancial,
title={TradingAgents: Multi-Agents LLM Financial Trading Framework},
author={Yijia Xiao and Edward Sun and Di Luo and Wei Wang},
year={2025},
eprint={2412.20138},
archivePrefix={arXiv},
primaryClass={q-fin.TR},
url={https://arxiv.org/abs/2412.20138},
}
内容由编辑整理，部分信息来自 GitHub 与项目官网，版权归原作者所有。发现错误？报告问题