xuangu

XuanGu is an intelligent stock selection research system for A-shares, Hong Kong stocks, and US stocks, targeting a future horizon of approximately two weeks (10 trading days). It integrates candidate recall, market data validation, factor scoring, deep research, event evidence, risk gates, and historical validation into a single auditable pipeline, aiming to identify stocks worth prioritized research based on post-fee returns over the next 10 trading days. The project explicitly states that the homepage is an explainable rule-based research ranking, not a validated model capable of predicting price increases, and does not constitute investment advice. In terms of core capabilities, the system provides dynamic candidate pools across three markets: A-share dynamic recall targets 300 stocks, while Hong Kong and US markets re-read public market cross-sections on each snapshot, with recall targets of 200 and 300 stocks respectively, no longer relying on static repository lists. All valid A-share market data candidates first complete basic scoring and technical scoring; up to 300 stocks that pass tradability filtering proceed to deep research using Legacy, V2, Dual-Low, Chan/CZSC, Serenity, UZI, and the jury model. The scoring mechanism includes an independent "Two-Week Return Opportunity Research Ranking" composed of six evidence contributions: multi-period momentum, same-market relative strength, momentum acceleration, trend volume-price structure, industry resonance, and deduplicated event catalysts. It then deducts overextension penalties based on past 5/10-day gains and deviation from MA20, distinguishing three entry states: pending open review, cautious review, and observation-only pullback. The system emphasizes that these scores are deterministic rule scores, not probabilities of price increase, and do not represent return commitments. Production output uses a two-layer structure: production_decision V4 runs two rule qualification channels—event catalyst and quality trend—separately after shared safety gates, producing QUALIFIED_PICK or NO_QUALIFIED_PICK; global_decision handles independent strictly calibrated probabilities. Rule qualification scores are not equivalent to price increase probabilities, and the Shadow model is not authorized simply because the rule track has candidates. For automation, the system uses GitHub Actions native schedule and a 30-minute watchdog. After configuring a repository-specific token, Cloudflare Cron provides independent primary triggering and re-dispatch; concurrency locks and the online schedule_gate handle deduplication, with bounded retries for timeouts and transient failures. When snapshots expire, the system switches to HISTORICAL_RESEARCH_ONLY, and currently executable candidates are forced to 0. The page provides six tabs: Today's Answer, Candidate Pool, Event Evidence, Historical Validation, Model Logic, and Data Health. All numbers come from published snapshots; the browser only displays server-side published and consistency-verified results. Historical validation uses six categories of physically isolated evidence, including opportunity ranking subsequent performance, rule qualification results, formal executable track, Shadow research track, and complete observation track, with raw snapshots kept immutable. The project also includes a 10-trading-day shadow probability model and net excess return ranking V2, both in research or collection stages, with calibrated=false and production_eligible=false fixed, and they will not independently obtain buy permissions. The README records release acceptance records for 2026-09-10 and 2026-09-11, indicating that this version can generate, publish, and display verifiable research results, but does not prove maximization of future returns, nor does it constitute a continuous on-time SLA.