simonlin1212/TradingAgents-astock

TradingAgents‑Astock – A China‑stock‑focused multi‑LLM research framework

What it is

• A fork of the open‑source TradingAgents project (the paper TradingAgents: Multi‑Agents LLM Financial Trading Framework). It implements the same multi‑agent pipeline but rewrites the data layer, agent roles and trading rules for China’s A‑share market.
• Licensed under Apache 2.0, installable with a single pip install -e . and runs locally (no Docker required, though a Docker image is provided).

Key ideas

Dimension	Original TradingAgents (US‑stock)	TradingAgents‑Astock (China)
Data sources	Yahoo Finance, Alpha Vantage	Free Chinese data via mootdx (TCP), East‑money, Sina, Tonghuashun, Tencent Finance, etc. – no API keys needed
Analyst roles	4 analysts (market, sentiment, news, fundamentals)	7 analysts – the original 4 plus Policy, Hot‑Money (track “龙虎榜”), and Lockup (share‑unlock monitoring)
Trading constraints	US rules (T+0, no price limits)	A‑share rules (T+1 settlement, daily涨跌停 limits, minimum lot size, trading windows)
Output language	English	Chinese reports (internal debate stays English for better reasoning)
Benchmark	SPY	CSI‑300 (沪深300)

Architecture (high‑level)

1. 7 Analyst agents generate individual research notes using tool‑calls to the data layer.
2. Bull vs. Bear researchers debate the notes (configurable number of rounds).
3. Research Manager (deep‑think LLM) synthesises a trading plan.
4. Trader converts the plan into an A‑share‑compliant order (T+1, price‑limit checks, lot‑size rounding).
5. Risk debaters (Aggressive / Neutral / Conservative) discuss risk.
6. Portfolio Manager (deep‑think LLM) issues a final rating and rationale.

Two LLM tiers are used:

• quick_think_llm for the fast‑turnover agents (analysts, researchers, trader, risk debaters).
• deep_think_llm for the two managers that need the whole context.

How to get started

git clone https://github.com/simonlin1212/tradingagents-astock.git
cd tradingagents-astock
pip install -e .
# optional: install a specific LLM provider, e.g.
# pip install --no-deps "langchain-google-genai>=4.0.0"

Create a .env file with the API key of the LLM you want (MiniMax, DeepSeek, ZhiPu, Qwen, OpenAI, Anthropic, etc.).

MINIMAX_API_KEY=sk-…   # or DEEPSEEK_API_KEY, OPENAI_API_KEY, …

Run either the CLI or the Streamlit UI:

tradingagents               # interactive CLI
tradingagents-web           # starts the Streamlit UI
# or
streamlit run web/app.py

Supply a 6‑digit A‑share ticker and a target date; the system will fetch data, run the full 12‑step pipeline, and output a markdown (or PDF) report containing:

• individual analyst notes
• the bull/bear debate
• risk‑debate outcomes
• a final Buy / Hold / Sell rating with a short justification.

Performance‑tracking feature

• After each run the decision and the subsequent market outcome (relative to CSI‑300) are logged.
• tradingagents performance aggregates these logs and reports direction‑accuracy (how often the rating’s direction beat the benchmark), up_rate, outperform_rate, and a monotonicity test of the rating tiers.
• No extra LLM calls are needed for this statistics; it simply reads the stored logs.

Configuration highlights

• llm_provider, quick_think_llm, deep_think_llm choose models.
• max_debate_rounds and max_risk_discuss_rounds control how many back‑and‑forth exchanges occur.
• role_llms lets you assign different providers/models to specific roles (e.g., give the Bull and Bear debaters different LLMs to avoid echo‑chamber effects).
• backend_url enables any OpenAI‑compatible gateway (useful for domestic proxies or self‑hosted inference).
• output_language is currently limited to Chinese (reports) while internal reasoning stays English.

Typical use‑cases

• Research/teaching – explore how multi‑agent LLM debates behave on Chinese market data.
• Prototype development – swap in your own data vendor or custom analyst logic while keeping the graph infrastructure.
• Metric collection – run the framework on historical dates, gather direction‑accuracy statistics, and experiment with prompt or model changes.

What it is not

• Not an investment advisory service; it emits no actionable price targets, position sizing, or execution instructions.
• Not a back‑testing engine – the performance view only records the outcome of each single‑day decision without transaction‑cost modeling.
• Not a turnkey “stock‑picker” app – you must provide your own LLM API keys and run the code on your own hardware.

---

Bottom line: TradingAgents‑Astock is a fully‑open, pip‑installable research framework that adapts the multi‑LLM TradingAgents architecture to the specifics of China’s A‑share market, offering free data connectors, seven specialized analyst agents, and a complete debate‑to‑rating pipeline suitable for academic study or experimental trading‑research projects.