Most multi‑agent trading demos target US markets and assume Western data feeds and trading rules. This fork focuses on the often‑overlooked problem of adapting an LLM‑based multi‑agent research pipeline to the realities of the Chinese A‑share market—data heterogeneity, policy sensitivity, short‑term hot money behavior, and T+1 / daily limit constraints.
What Sets It Apart
- A‑share data engineering: native connectors to mootdx (TCP), Tencent, Eastmoney, Sina and Tonghuashun to fetch K‑lines,龙虎榜, 解禁, F10 and on‑chain style indicators without paid APIs. This means the pipeline can run offline in China‑facing environments where Yahoo/AlphaVantage are unusable.
- Seven analyst roles tuned to A‑share dynamics: market, social, news, fundamentals plus three A‑specialists (policy analyst, hot‑money/龙虎榜 tracker, lockup/解禁 watcher). Each analyst produces tool‑augmented reports that feed bull/bear debates and a portfolio manager.
- Operational realism: enforces A‑share trading constraints (T+1 settlement, daily涨跌停, minimum lot size, ST rules) and provides both a Streamlit web UI and CLI for one‑click analyses or batch runs.
- LLM‑agnostic orchestration: dual‑LLM design (quick_think for analyst tooling; deep_think for global managers), configurable to multiple providers (MiniMax, DeepSeek, Qwen, OpenAI, Anthropic, Google, etc.), but requires API keys and incurs many model calls per run.
Who It's For and Tradeoffs
Great fit if you want to prototype LLM‑based multi‑agent investment research specifically for Chinese A‑shares, need free direct data connectors, and want an out‑of‑the‑box Streamlit UI plus a CLI for batch reports. Look elsewhere if you need a production trading system (this is for research/demonstration and explicitly disclaims investment advice), want a zero‑cost LLM usage (the pipeline needs many paid API calls for reliable tool usage), or require institutional‑grade order execution and compliance plumbing out of the box.
This repository is a deep fork of TauricResearch/TradingAgents, released under Apache‑2.0, and emphasizes practical engineering choices for running multi‑agent LLM workflows on China market data rather than proposing new model architectures.
