FisherQuant · stock-report

An autonomous AI agent system for A-share quantitative research — from daily market judgment to candidate scoring, strategy tracking, and retrospective optimization, all in one self-updating public dashboard.

这不是一份静态选股报告，而是一套每日自动运行的 AI 量化研究操作系统。
This is not a static stock report — it is a daily self-running AI quantitative research operating system.

What It Does

Most quant dashboards answer one question. FisherQuant answers five, in sequence, every single trading day:

① Can we trade today?      →  Market regime classification + risk scoring
② Which strategy to run?   →  Multi-strategy activation & consensus scoring
③ Which stocks to watch?   →  AI-scored candidate cards with entry/exit conditions
④ Did we get it right?     →  Automated retrospective vs. actual price movement
⑤ How to improve?          →  Factor evolution tracking + strategy parameter tuning

All five layers are publicly visible, fully traceable to run_id / trade_date / generated_at, and rebuilt automatically on each trading day.

Live Dashboard

Layer	Page	Description
🏠 System Console	`index.html`	Run status, system verdict, overall health
📊 Market	`market-overview.html`	Morning brief, midday analysis, industry heatmap
🎯 Strategy	`strategy-vs-market.html`	Active / watch / degraded strategy states
📋 Candidates	`decision-candidates.html`	Per-stock AI decision cards
🔬 Research Lab	`research-lab.html`	Alpha191 factor research, T1 model analysis
🔁 Review	`recommendation-review.html`	Post-hoc validation against real market data

Screenshots

System Console	Market Overview

Candidate Decisions	Strategy vs Market

Architecture

┌─────────────────────────────────────────────────────────┐
│                 FisherQuant Agent Pipeline               │
│                                                         │
│  Market Data ──▶ Market Agent ──▶ market_state.json     │
│                       │                                 │
│                       ▼                                 │
│             Strategy Orchestrator ──▶ strategy_state    │
│              (PreBreakout v4.1 / T1 / O2C)              │
│                       │                                 │
│                       ▼                                 │
│          AI Factor Research (Alpha191)                  │
│                       │                                 │
│                       ▼                                 │
│         Candidate Scoring ──▶ candidate_state.json      │
│                       │                                 │
│                       ▼                                 │
│       system_verdict.json  ◀──  Review & Validation     │
│                       │                                 │
│                       ▼                                 │
│        data/latest/*.json  ──▶  GitHub Pages            │
└─────────────────────────────────────────────────────────┘

The entire pipeline is orchestrated by a Python agent backend and published to this repository via GitHub Actions every trading day. The public site reads exclusively from data/latest/*.json — a strict versioned data contract ensuring every displayed result is reproducible and auditable.

Data Contract

All public pages consume a single source of truth: data/latest/.

File	Purpose
`run_manifest.json`	Run identity: `run_id`, `trade_date`, validation gates
`market_state.json`	Market regime, risk score, morning & midday context
`strategy_state.json`	Strategy activation states and consensus scores
`candidate_state.json`	Per-stock decision cards with AI reasoning
`review_state.json`	Retrospective summary vs. actual outcomes
`research_state.json`	Factor research and alpha signal analysis
`system_verdict.json`	Final system-level go/no-go verdict with gate checks

Every artifact is tagged with trade_date, run_id, and generated_at — enabling full audit trails across runs.

Key Design Decisions

Single data contract: pages never parse raw analytics directly; all display logic reads from data/latest/ only
Layered responsibility: market / strategy / candidate / review are strictly separated — no layer skips upstream validation
Hard gates: system_verdict.json enforces freshness, date consistency, and pipeline completion before anything is published
Zero client-side secrets: the public site is fully static; all AI inference and data processing happens in the private agent backend

Tech Stack

Layer	Technology
Agent Backend	Python (multi-model: T1, O2C, Alpha191 pipeline)
Frontend	Vanilla JS + Chart.js, semantic HTML5
Styling	Custom CSS design system (`assets/styles/`)
Deployment	GitHub Actions → GitHub Pages
Data Format	JSON (versioned contracts) + CSV (factor outputs)

Roadmap

✅ Stable

Multi-layer pipeline: market → strategy → candidate → review
Unified data/latest/*.json data contract with full traceability
Daily auto-deploy via GitHub Actions
AI-generated market context, Alpha191 factor research, multi-strategy scoring
System verdict with hard gate validation

🚀 Codex-Driven Evolution

Phase 1 — Architecture Migration (#1)
Transition agent memory structures and multi-model routing into a native OpenAI Codex ecosystem for deeper contextual reasoning across runs.
Phase 2 — Automated Maintenance (#2)
Integrate Codex into GitHub Actions for autonomous PR review, incremental refactoring, and AI-driven issue triage.
Phase 3 — Developer Ecosystem (#3)
Self-generating API documentation and contributor tooling powered by Codex code comprehension.

Contributing

See CONTRIBUTING.md for guidelines. Contributions welcome in:

Frontend: improving the dashboard UI/UX (*.html, assets/)
Data schema: proposing extensions to the data/latest/ contract
Tooling: generate_github_pages.py, deploy.sh improvements

Please open an issue before submitting a PR for non-trivial changes.

Security

See SECURITY.md for the vulnerability reporting policy.

License

MIT — data outputs and frontend code are freely reusable. The private agent backend and proprietary strategy logic are not included in this repository.