TradingAgents-Astock is a fork of TradingAgents that adapts its multi-agent research and debate architecture to China's A-share market. It reworks three layers of the original: data sources, analyst roles, and trading rules. It is positioned as a research and teaching implementation of the TradingAgents paper, not an investment product, and it produces no entry prices, stop-losses, position sizes, or price targets.
An analysis runs as a chain of agents. Seven analysts write reports, a bull researcher and a bear researcher debate them over several rounds, a research manager drafts an investment plan, a trader turns it into a plan under A-share constraints, aggressive, conservative, and neutral risk debaters argue it, and a portfolio manager issues the final rating and reasoning. A quick-thinking model serves the analysts, researchers, trader, and risk debaters, and a deep-thinking model serves the two managers. Individual roles can also use different models, for example giving the bull and bear sides models from different vendors. Reports are written in Chinese while the internal debate stays in English.
| Area | TradingAgents | TradingAgents-Astock |
|---|---|---|
| Data sources | Yahoo Finance, Alpha Vantage | mootdx, Eastmoney, Sina, Tonghuashun, Tencent, Cailianshe, Baidu (free, no API key) |
| Analyst roles | 4 (market, sentiment, news, fundamentals) | 7 (adds policy, hot-money tracking, lockup monitoring) |
| Trading rules | US market | A-share T+1, price limits, lot sizes, trading sessions |
| Alpha benchmark | SPY | CSI 300 |
A performance command reads past decisions from the memory log, backfills realized returns and alpha against the CSI 300, and reports direction accuracy by rating and by ticker without calling an LLM. A Streamlit web UI lets non-programmers pick a provider and model, enter a six-digit stock code, follow a 12-stage run, and export the report.
Features
- Seven analysts: market, sentiment, news, and fundamentals roles adapted to A-shares, plus policy, hot-money, and share-unlock analysts
- Bull and bear debate: two researchers argue the analyst reports for a configurable number of rounds before a research manager decides
- Risk debate: aggressive, conservative, and neutral debaters review the trader's plan before the portfolio manager's rating
- A-share rules: the trader works under T+1 settlement, daily price limits, minimum lot sizes, ST status, and trading sessions
- Free A-share data: K-lines, financial statements, the Dragon-Tiger list, unlock schedules, fund flows, and sector data from public Chinese finance sources
- Two-tier models: a quick-thinking model for most roles and a deep-thinking model for the managers, with optional per-role overrides
- Model providers: MiniMax, DeepSeek, Qwen, GLM, OpenAI, Anthropic, Google, xAI, OpenRouter, Ollama, Kimi, and any OpenAI-compatible gateway
- Claude subscription mode: optional routing of nodes through the Claude Agent SDK and a signed-in Claude CLI instead of per-token API billing
- Decision performance: reports direction accuracy, alpha against the CSI 300, and whether ratings separate outcomes, from logged decisions
- Web UI: a Streamlit interface with live stage progress, full reports, history, and Markdown or PDF export
