When LLMs Play Devil’s Advocate With Chinese Stock Data

What it does

This Python project is a concept-validation system for Chinese A-share investing that routes market data through a panel of specialized agents. A bull researcher and a bear researcher ingest fundamentals, news, and technical data—sourced largely via AkShare—then present opposing cases in a structured debate. An LLM acts as an independent third-party referee, and the system synthesizes their arguments into a final decision signal.

The interesting bit

The core mechanic is adversarial validation disguised as an investment committee. Rather than trusting a single prompt, the system makes opposing agents argue and uses an LLM to evaluate the confidence differential, theoretically catching biases that a lone model might miss.

Key highlights

• Multi-agent pipeline with dedicated bull, bear, and macro-analyst roles
• “Debate room” module where an LLM judges arguments and feeds a hybrid confidence score into the final decision
• SQLite cache layer (akshare_cache.db) with configurable TTL to reduce AkShare API failures and news-fetch retries
• Incremental news caching and deduplication across Sina, NetEase, and East Money feeds
• Dual-licensed (MIT & GPLv3+NonCommercial) with both CLI and FastAPI backend interfaces

Caveats

• The backtesting module is explicitly flagged as still in testing
• No benchmark results or performance metrics appear in the README
• The authors repeatedly disclaim any investment advice; this is framed strictly as an educational research tool

Verdict

A solid sandbox if you are experimenting with multi-agent reasoning or adversarial LLM architectures. If you are looking for a trading bot to deploy with real capital, the README itself tells you to look elsewhere.