Multi-Agents LLM Financial Trading Framework

• [2026-08] TradingAgents v0.4.0 released pinch look-ahead / point-in-time fixes crossed FRED macro, societal sentiment, and the decision-log memory; clearer determination signals; moving CLI checkpoint resume; Trader value grounding; and the GPT-5.6 and GLM-5.3 models. See CHANGELOG.md for the afloat list.
• [2026-07] TradingAgents v0.3.1 released pinch correctness and stableness fixes: Alpha Vantage look-ahead filtering, graph-router crash-safety, graph-shape-aware checkpoint resume, moving crypto sentiment sources, a configurable LLM retry budget, Bedrock API-key auth, and Claude Sonnet 5 / Fable 5 support.
• [2026-06] TradingAgents v0.3.0 released pinch a verified data-access contract, an expanded supplier registry (NVIDIA, Kimi, Groq, Mistral, Bedrock, and immoderate OpenAI-compatible endpoint), FRED and Polymarket information vendors, a current-generation exemplary catalog, and a CI gate.
• [2026-05] TradingAgents v0.2.5 released pinch the grounded Sentiment Analyst, GPT-5.5 etc. exemplary coverage, Qwen/GLM/MiniMax dual-region support, TRADINGAGENTS_* env-var configurability pinch API-key auto-detection, distant Ollama support, non-US alpha benchmarks, and ticker path-traversal hardening.
• [2026-04] TradingAgents v0.2.4 released pinch structured-output agents (Research Manager, Trader, Portfolio Manager), LangGraph checkpoint resume, persistent determination log, DeepSeek/Qwen/GLM/Azure supplier support, Docker, and a Windows UTF-8 encoding fix.
• [2026-03] TradingAgents v0.2.3 released pinch multi-language support, GPT-5.4 family models, unified exemplary catalog, backtesting day fidelity, and proxy support.
• [2026-03] TradingAgents v0.2.2 released pinch GPT-5.4/Gemini 3.1/Claude 4.6 exemplary coverage, five-tier standing scale, OpenAI Responses API, Anthropic effort control, and cross-platform stability.
• [2026-02] TradingAgents v0.2.0 released pinch multi-provider LLM support (GPT-5.x, Gemini 3.x, Claude 4.x, Grok 4.x) and improved strategy architecture.
• [2026-01] Trading-R1 Technical Report released, pinch Terminal expected to onshore soon.

🎉 TradingAgents officially released! We person received galore inquiries astir the work, and we would for illustration to definitive our acknowledgment for the enthusiasm successful our community.

So we decided to afloat open-source the framework. Looking guardant to building impactful projects pinch you!

TradingAgents is simply a multi-agent trading model that mirrors the dynamics of real-world trading firms. By deploying specialized LLM-powered agents: from basal analysts, sentiment experts, and method analysts, to trader, consequence guidance team, the level collaboratively evaluates marketplace conditions and informs trading decisions. Moreover, these agents prosecute successful move discussions to pinpoint the optimal strategy.

TradingAgents model is designed for investigation purposes. Trading capacity whitethorn alteration based connected galore factors, including the chosen backbone connection models, exemplary temperature, trading periods, the value of data, and different non-deterministic factors. It is not intended arsenic financial, investment, aliases trading advice.

Our model decomposes analyzable trading tasks into specialized roles.

• Fundamentals Analyst: Evaluates institution financials and capacity metrics, identifying intrinsic values and imaginable reddish flags.
• Sentiment Analyst: Aggregates news headlines, StockTwits, and Reddit chatter into a azygous sentiment publication to gauge short-term marketplace mood.
• News Analyst: Monitors world news and macroeconomic indicators, interpreting the effect of events connected marketplace conditions.
• Technical Analyst: Utilizes method indicators (like MACD and RSI) to observe trading patterns and forecast value movements.

• Comprises some bullish and bearish researchers who critically measure the insights provided by the Analyst Team. Through system debates, they equilibrium imaginable 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 consequence by assessing marketplace volatility, liquidity, and different consequence factors. The consequence guidance squad evaluates and adjusts trading strategies, providing appraisal reports to the Portfolio Manager for last decision.
• The Portfolio Manager approves/rejects the transaction proposal. If approved, the bid will beryllium sent to the simulated speech and executed.

Clone TradingAgents:

Create a virtual situation successful immoderate of your favourite situation managers:

Install the package and its dependencies:

Alternatively, tally pinch Docker:

For section models pinch Ollama:

TradingAgents supports aggregate LLM providers. Set the API cardinal for your chosen provider:

For Azure OpenAI, transcript .env.enterprise.example to .env.enterprise and capable successful your credentials.

For AWS Bedrock, instal the other pinch pip instal ".[bedrock]", group llm_provider: "bedrock", configure AWS credentials (environment variables, ~/.aws/credentials, aliases an IAM role) and AWS_DEFAULT_REGION, and usage a Bedrock exemplary ID, e.g. us.anthropic.claude-opus-4-8-v1:0.

For section models, configure Ollama pinch llm_provider: "ollama". The default endpoint is http://localhost:11434/v1; group OLLAMA_BASE_URL to constituent astatine a distant ollama-serve. Pull models pinch ollama propulsion <name>, and prime "Custom exemplary ID" successful the CLI for immoderate exemplary not listed by default.

For immoderate different OpenAI-compatible server (vLLM, LM Studio, llama.cpp, aliases a civilization relay), usage llm_provider: "openai_compatible" and group the endpoint via backend_url (or TRADINGAGENTS_LLM_BACKEND_URL), e.g. http://localhost:8000/v1 for vLLM aliases http://localhost:1234/v1 for LM Studio. The exemplary is immoderate your server serves. No cardinal is needed for section servers; group OPENAI_COMPATIBLE_API_KEY erstwhile the endpoint requires one.

Alternatively, transcript .env.example to .env and capable successful your keys:

Launch the interactive CLI:

You will spot a surface wherever you tin prime your desired tickers, study date, LLM provider, investigation depth, and more.

TradingAgents useful pinch immoderate marketplace Yahoo Finance covers, utilizing the exchange-suffixed ticker. Company personality and the alpha benchmark resoluteness 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 look showing results arsenic they load, letting you way the agent's advancement arsenic it runs.

We built TradingAgents pinch LangGraph to guarantee elasticity and modularity. The model supports aggregate LLM providers: OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen (Alibaba DashScope, world and China endpoints), GLM (Zhipu), MiniMax (global + China), OpenRouter, Ollama for section models, and Azure OpenAI for enterprise.

To usage TradingAgents wrong your code, you tin import the tradingagents module and initialize a TradingAgentsGraph() object. The .propagate() usability will return a decision. You tin tally main.py, here's besides a speedy example:

You tin besides set the default configuration to group your ain prime of LLMs, statement rounds, etc.

See tradingagents/default_config.py for each configuration options.

TradingAgents persists 2 kinds of authorities crossed runs.

The determination log is ever on. Each completed tally appends its determination to ~/.tradingagents/memory/trading_memory.md. On the adjacent tally for the aforesaid ticker, TradingAgents fetches the realised return (raw and alpha vs SPY), generates a one-paragraph reflection, and injects the astir caller same-ticker decisions positive caller cross-ticker lessons into the Portfolio Manager prompt, truthful each study carries guardant what worked and what didn't.

Override the way pinch TRADINGAGENTS_MEMORY_LOG_PATH.

Checkpoint resume is opt-in via --checkpoint. When enabled, LangGraph saves authorities aft each node truthful a collapsed aliases interrupted tally resumes from the past successful measurement alternatively of starting over. On a resume tally you will spot Resuming from measurement N for <TICKER> connected <date> successful the logs; connected a caller tally you will spot Starting fresh. Checkpoints are cleared automatically connected successful completion.

Per-ticker SQLite databases unrecorded astatine ~/.tradingagents/cache/checkpoints/<TICKER>.db (override the guidelines pinch TRADINGAGENTS_CACHE_DIR). Use --clear-checkpoints to reset each of them earlier a run.

TradingAgents is LLM-driven, truthful 2 runs of the aforesaid ticker and day tin differ. This is expected for a investigation instrumentality built connected connection models, not a defect. The variety comes from a fewer chopped sources, and it helps to abstracted them.

Language exemplary sampling is non-deterministic. Even astatine a fixed temperature, providers do not guarantee byte-identical output crossed calls, and reasoning models (the default GPT-5.x family, and immoderate thinking-mode model) alteration the astir because their soul reasoning is itself sampled.

Live information moves. News, StockTwits, and Reddit return different contented arsenic clip passes, truthful a tally coming sees different inputs than a tally past week moreover for the aforesaid humanities waste and acquisition date. Pin the study day to clasp the value and parameter model fixed, but the societal and news sources still bespeak "now".

To trim variety you tin little the sampling temperature. Set somesthesia successful your config (or TRADINGAGENTS_TEMPERATURE successful .env); little values make models that grant it much repeatable. The existent curated models are reasoning-first and mostly disregard temperature, truthful for tighter reproducibility usage a non-reasoning model, which you tin group explicitly via the Custom exemplary ID option.

What does not alteration anymore: the analyzed institution personality is resolved deterministically from the ticker earlier immoderate supplier runs, and the marketplace expert grounds nonstop value and parameter claims successful a verified information snapshot. Earlier reports of "different companies" aliases fabricated value levels crossed runs are addressed by these 2 mechanisms.

Backtest results are not guaranteed to lucifer immoderate published figure. Returns dangle connected the model, the temperature, the day range, information quality, and the sampling above. Treat the model arsenic a investigation scaffold for studying multi-agent analysis, not arsenic a strategy pinch a fixed, replicable return.

Contributions are welcome: bug fixes, documentation, and characteristic ideas; past contributions are credited per merchandise successful CHANGELOG.md.

Please reference our activity if you find TradingAgents provides you pinch immoderate thief :)