// the find
simonlin1212/TradingAgents-astock
A股多Agent投研框架 — 适配A股数据源(龙虎榜/游资/解禁等)，7位分析师基于A股规则的辩论决策，基于TradingAgents深度改造，适配大A。A-share multi-agent investment research framework — 7 AI analysts, bull/bear debate, risk assessment。
A fork of TauricResearch's TradingAgents that rewires the multi-agent debate pipeline for China A-shares. It pulls data from mootdx, Eastmoney, Sina and Tonghuashun instead of Yahoo Finance, and adds policy, Dragon-Tiger-list hot-money and lockup-expiry analysts. It suits researchers who want to study how LLM debate behaves on Chinese market data, and the README is explicit that it is not a source of trade ideas.
The A-share rules are built into the pipeline rather than bolted on: T+1, daily limit up and down, minimum lot sizes and ST handling are all modeled, and the benchmark is CSI 300, which matters because much of the upside on a bullish call in A-shares is market beta. The performance command is the most honest piece: it scores direction on alpha, leaves Hold out of the count, drops unparseable records instead of counting them as zero, and checks whether the five rating tiers are monotonic in mean alpha. Eastmoney traffic goes through one throttled `_em_get()` with a reused session and jitter, which is more care than most scrapers give the endpoints they depend on. The per-role LLM config fails loudly on misspelled role names, shares one client across identical provider and model pairs, and logs truncated outputs instead of returning half a report.
Most A-share data comes from unofficial HTTP endpoints on Eastmoney, Tonghuashun, Baidu and CLS with no contract, so a renamed field or changed response will break a stage without warning. The throttle reduces ban risk but does nothing about schema drift. Report quality depends on the model's tool-calling, and the README admits lightweight models sometimes return empty analyst reports, which are skipped, so a run can finish with fewer than the seven analysts it advertises. A full analysis is 30 to 50 LLM calls per ticker and the README gives no cost or latency estimate, which matters when the batch script loops over a watchlist. The output is a direction and rating with no price levels, by design, and the performance statistics are the only check on it; those use overlapping holding windows, ignore transaction costs and have no position sizing, so a good score there is weak evidence.