TradingAgents

• Running a multi-agent analysis of a stock or crypto pair on a specific date and reading each agent's report
• Backtesting the agent pipeline over a grid of tickers and dates to study decision quality
• Comparing how different LLM providers or debate settings change trading decisions
• Passing a portfolio file so the agents reason about an existing position rather than a flat book
• Using the codebase as a starting point for custom multi-agent financial research

TradingAgents applies role-based multi-agent reasoning to a concrete domain. The analyst, debate, trader, and risk structure is easy to inspect, and recent releases added point-in-time SEC fundamentals and grid backtesting, which reduce look-ahead bias when you evaluate it on past dates. It is still a research scaffold. Decisions vary from run to run, social and news data are not point-in-time, and there is no live execution. Treat it as a tool for studying and extending multi-agent financial analysis, not as a strategy to trade on.

About TradingAgents

TradingAgents runs a LangGraph pipeline modeled on a trading firm. Four analysts gather information in parallel: fundamentals, sentiment (news headlines, StockTwits, and Reddit), news and macro, and technical indicators such as MACD and RSI. Bullish and bearish researchers then debate the findings. A trader agent proposes a decision, a risk management team assesses it, and a portfolio manager approves or rejects it, with approved orders sent to a simulated exchange. You run it from an interactive CLI (`tradingagents`), through Docker, or in Python by creating a `TradingAgentsGraph` and calling `propagate(ticker, date)`. It works with any market Yahoo Finance covers, including US, Hong Kong, Tokyo, London, India, and China A-share tickers and crypto pairs. Other data vendors include Alpha Vantage, FRED macro data, Polymarket, and, since v0.5.0, SEC EDGAR fundamentals served as filed, so a historical run sees statements as they stood on that date. The framework supports many LLM providers, including OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, GLM, MiniMax, Kimi, Mistral, Groq, NVIDIA NIM, OpenRouter, Azure OpenAI, AWS Bedrock, Ollama, and any OpenAI-compatible server, and it uses separate deep-thinking and quick-thinking models (GPT-6 Sol and GPT-6 Luna by default). Recent releases added portfolio-aware runs that reason against your actual holdings, a decision log that reflects on realized returns in later runs, checkpoint resume, and grid backtesting scored on realized alpha against each market's benchmark. The project says plainly that it is designed for research and is not financial, investment, or trading advice. Runs are non-deterministic, news and social inputs reflect the present even for historical dates, and backtest results are not guaranteed to match published figures. It is most useful to people who want to study or extend multi-agent financial reasoning and are comfortable managing Python environments and API keys.