TradingAgents-astock is a fork of TradingAgents adapted to China's A-share market. The upstream framework targets US equities, with data from Yahoo Finance and Alpha Vantage and analysts that assume US market rules. This fork replaces the data layer with free, key-less Chinese sources, adds three A-share-specific analyst roles, and applies A-share trading constraints (T+1 settlement, daily price limits, minimum lot sizes, trading sessions, and ST stocks). Reports are written in Chinese while the internal debate stays in English, and the alpha benchmark is the CSI 300 rather than SPY. The project positions itself as an engineering implementation of the TradingAgents paper for research and teaching rather than an investment product: it outputs a direction, a rating, and reasoning, with no entry-price, stop-loss, position-size, or target-price outputs.
The pipeline runs seven analysts, each with a tool loop: market, social sentiment, news, and fundamentals, plus the added policy analyst, hot-money tracker (dragon-tiger lists and large-order flows), and lock-up monitor (share unlocks, major-holder reductions, and pledges). Their reports feed a bull-versus-bear researcher debate, a Research Manager that produces an investment plan, a Trader that applies the A-share constraints, a three-way aggressive, conservative, and neutral risk debate, and a Portfolio Manager that issues the final rating. Two model tiers are used: quickthinkllm for analysts, researchers, trader, and risk debaters, and deepthinkllm for the two managers. role_llms assigns a different provider or model to any single role, for example giving the bull and bear debaters different vendors.
Data sources need no API key: mootdx (TCP) for OHLCV bars, financial snapshots, and F10 text; Tencent Finance for real-time PE, PB, market cap, and turnover; East Money for dragon-tiger lists, unlocks, sector quotes, and stock information; Sina for historical bars and financial statements; 10jqka for EPS consensus; CLS for global news; and Baidu for concept sectors and fund flows. East Money requests go through a serial throttle with jitter and a keep-alive session to avoid temporary IP bans during batch runs.
LLM providers include MiniMax (default), DeepSeek, Zhipu GLM, Qwen, OpenAI, Anthropic, Google Gemini, xAI, OpenRouter, Ollama, Kimi through the Anthropic-compatible endpoint, and any OpenAI-compatible gateway via a custom base URL. An optional [agentsdk] extra routes some or all nodes through the Claude Agent SDK and a logged-in claude CLI so they consume a personal Claude Pro or Max subscription instead of API billing, falling back to a paid provider when the quota is hit. Each analysis makes roughly 30 to 50 LLM calls.
Features
- Seven analysts: four upstream roles adapted to A-shares plus policy, hot-money, and lock-up analysts
- Debate chain: bull and bear research debate, Research Manager synthesis, Trader plan under A-share constraints, three-way risk debate, and Portfolio Manager decision
- Free data layer: mootdx, Tencent, East Money, Sina, 10jqka, CLS, and Baidu, with East Money rate limiting
- Per-role models: role_llms sets provider and model per role; identical provider and model pairs share one client instance
- Streamlit Web UI: choose provider and model in the sidebar, enter a six-digit code and date, follow a 12-stage progress view, and export Markdown or PDF reports
- CLI and Python API: the tradingagents interactive CLI and TradingAgentsGraph.propagate(code, date); examples/run_cases.py batches several tickers
- Decision performance stats: tradingagents performance reads the memory log and reports direction_accuracy on an alpha basis against the CSI 300, marking samples under 20 as noise
- Claude subscription mode: a claudeagentsdk provider override with opus and sonnet aliases and quota fallback
- Checkpointing: optional SQLite resume of interrupted runs
