⚠️ 免责声明 / DISCLAIMER本项目仅为个人学习与技术探索的玩具项目（Toy Project）。 所有分析结果、预测信号、评分及任何输出均不构成任何形式的投资建议或交易决策依据。 作者不对任何因参考本项目内容而产生的投资损失承担责任。股市有风险，投资须谨慎。
This is a personal hobby / toy project for learning and experimentation only. All analysis results, prediction signals, scores, and any output do NOT constitute investment advice or trading recommendations of any kind. The author assumes no responsibility for any financial loss resulting from using this project. Investing involves risk. Always do your own research.
一个基于大语言模型（LLM）驱动的 A 股市场智能分析平台。用 AI 探索 A 股市场的信号分析、新闻情报、量化回测与自主决策循环——纯粹出于对技术的好奇心。
An LLM-powered intelligent analysis platform for the A-share (Chinese stock) market. Built out of curiosity to explore how AI models can be applied to signal analysis, news intelligence, quantitative backtesting, and autonomous agent loops.
| 模块 | 说明 | Module | Description |
|---|---|---|---|
| 🧠 自主 Agent 循环 | OODA 循环：信号聚合 → 贝叶斯预筛 → 多空辩论 → 风险闸门 → Kelly 仓位（仅模拟） | Autonomous Agent Loop | OODA cycle: signal aggregation → Bayesian prescreen → bull/bear debate → risk gates → Kelly sizing (simulation only) |
| 📡 市场情报管线 | 5 层信源 + 7 维评分 + 因果影响链 + NetworkX 知识图谱 | Market Intelligence | 5-layer sources + 7-dimension scoring + causal impact chains + NetworkX knowledge graph |
| 🎯 智能选股 | 多风格量化筛选 + LLM 复核 + T+1 隔夜风险 + 胜率追踪 | Smart Stock Picks | Multi-style screener + LLM review + T+1 overnight risk + win-rate tracking |
| 📊 市场状态识别 | HMM 三态（牛 / 熊 / 震荡）+ 情绪周期闸门 | Regime Detection | 3-state HMM (bull / bear / consolidation) + sentiment-cycle gating |
| 🤖 多模型 AI | Claude / Gemini / OpenAI / DeepSeek 路由 + 共识投票 | Multi-LLM | Claude / Gemini / OpenAI / DeepSeek routing + consensus voting |
| 🌊 事件总线 | Redis Streams 事件驱动微 OODA（行情 / 新闻 / 情绪 / 信号） | Event Bus | Redis-Streams event-driven micro-OODA (market / news / sentiment / signal) |
| 🛡️ 风险引擎 | 熔断器 + Kelly 仓位 + VaR + A股约束（T+1 / 涨跌停 / 100 股） | Risk Engine | circuit breaker + Kelly sizing + VaR + A-share constraints (T+1 / price limits / 100-share lots) |
| 🌐 全球情报 + 新闻 | 全球指数 / 大宗 / 汇率关联 + AI 新闻聚合 | Global Intel + News | global indices / commodities / FX correlation + AI news aggregation |
| 📈 量化回测 | backtrader 策略回测，可选 Qlib 自定义 alpha 因子 | Backtesting | backtrader strategy backtesting, optional Qlib custom alpha factors |
| 📱 Discord | 交易信号 / 情报自动推送到 Discord | Discord | auto-push trade signals / intel to Discord |
| 🖥️ Web UI | FastAPI + React 19：ControlTower / Portfolio / Recommendations / Review | Web UI | FastAPI + React 19: ControlTower / Portfolio / Recommendations / Review |
| ⚙️ 自动化调度 | Celery 45+ 定时任务 + 常驻 agent 守护进程 | Automation | Celery beat (45+ tasks) + always-on agent daemon |
v2 架构以「自主 Agent OODA 循环」为中心，由三路信号来源驱动，下游接风险闸门与（模拟）执行：
数据层 / Data (src/data/)
AKShare · EastMoney push2 · QMT · 多源健康路由回退
│ 行情 / OHLCV / 资金流 / 交易日历
▼
┌──────────────────────── 信号来源 / Signal Sources ────────────────────────┐
│ 市场情报 Intelligence 量化 Quant 智能选股 Recommendation │
│ src/intelligence(_hub)/ src/quant/ src/recommendation/ │
│ 5层信源·7维评分·因果链·知识图谱 HMM状态·Alpha·信号库 多风格筛选·LLM复核·T+1风险 │
└─────────────────────────────────────┬──────────────────────────────────────┘
│ signals
▼
自主 Agent 循环 / Agent OODA Loop (src/agent_loop/)
SignalAggregator → DecisionPipeline（贝叶斯预筛 · 多空辩论 · 风险闸门 · Kelly 仓位）
InvestmentDirector（7 团队）· 情绪周期闸门 · ThesisTracker · OutcomeTracker → 置信度校准
│ TradeProposal （⚠️ 仅模拟 / simulation only）
▼
风险 & 执行 / Risk & Execution 事件总线 / Event Bus
src/risk/ 熔断·Kelly·VaR src/event_bus/ Redis Streams
src/trading/ 执行闸门·KillSwitch·A股约束 7 streams → 微 OODA（行情/新闻/情绪/信号）
横切 / Cross-cutting:
src/llm/ 多模型网关（Claude/Gemini/OpenAI/DeepSeek + 共识投票）
src/web/ FastAPI · frontend/ React SPA · src/discord_bot/ · openclaw/ Celery + 常驻 daemon
| 层 / Layer | 技术 / Technologies |
|---|---|
| 数据 / Data | AKShare, adata, EastMoney push2 (curl_cffi), QMT/XtQuant (optional), pandas, numpy, pyarrow, yfinance |
| 分析 / Analysis | ta (technical indicators), plotly, matplotlib |
| AI 预测 / AI | Anthropic Claude, Google Gemini, OpenAI, DeepSeek, Claude Code bridge (fallback) |
| 量化 & Agent / Quant | hmmlearn (HMM 市场状态), networkx (知识图谱), scikit-learn, Qlib custom alpha factors (optional) |
| 策略回测 / Backtest | backtrader, Qlib (optional) |
| 后端 / Backend | FastAPI, uvicorn, Redis (cache + Streams 事件总线), Celery + Beat |
| 前端 / Frontend | React 19, TypeScript, Vite, shadcn/ui, Tailwind CSS 4, React Query |
| 通知 / Notification | Discord (bot + webhook) |
| 基础设施 / Infra | Docker Compose, nginx, searxng (self-host search, optional) |
pip install -r requirements.txt
make demo # 用样例数据跑 v2 回测 / runs the v2 backtest on bundled sample dataSee docs/how-it-works.md for what it does. For the full
stack (web UI, agent loop, automation), continue below.
- Docker & Docker Compose
- 至少一个 LLM API Key（Anthropic / Google / OpenAI）
- （可选）Discord Bot Token，用于推送通知
git clone https://github.com/Jcstack/ashare-ai-analyst.git
cd ashare-ai-analyst
cp .env.example .env
# 编辑 .env，填入你的 API Key
# Edit .env and fill in your API keys# 编辑 config/stocks.yaml，添加你关注的 A 股代码
# Edit config/stocks.yaml to add your A-share stock codesmake up # Docker 构建 + 启动所有服务 / Build + start all services访问 / Visit: http://localhost
make logs # 查看服务日志 / View logs# Python 后端
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# 代码检查 / Lint
.venv/bin/ruff check src/ tests/
.venv/bin/ruff format --check src/ tests/
# 单元测试 / Unit tests (external deps mocked; this is what CI runs)
.venv/bin/pytest tests/unit -q
# 前端 / Frontend
cd frontend
npm install
npx tsc --noEmit # 类型检查 / Type check
npm run build
npm run dev # 开发模式 / Dev mode| 文件 / File | 用途 / Purpose |
|---|---|
config/stocks.yaml |
股票自选池 / Stock watchlist |
config/llm.yaml |
LLM 模型路由（调用方→模型映射）/ LLM model routing (caller→model) |
config/agent.yaml · config/trading_loop.yaml |
Agent / OODA 循环参数 / Agent & OODA-loop params |
config/recommendation.yaml |
智能选股风格/过滤/权重 / Screener styles, filters, weights |
config/intelligence_hub.yaml · config/event_bus.yaml |
情报信源 / 事件总线 / Intel sources & event bus |
config/risk.yaml · config/trading_constraints.yaml · config/broker.yaml |
风险 / A股约束 / 券商 / Risk, A-share constraints, broker |
config/openclaw.yaml |
Celery beat 定时任务 / Celery beat schedule |
.env |
API Keys 与密钥（不提交！）/ API keys (never commit!) |
.
├── src/
│ ├── data/ # 多源行情采集 / Multi-source market data (AKShare·EastMoney·QMT)
│ ├── intelligence/ # 市场情报：因果链·辩论·知识图谱 / Causal chains, debate, knowledge graph
│ ├── intelligence_hub/ # 信源聚合·7维评分·存储 / Source aggregation, scoring, store
│ ├── agent_loop/ # 自主 OODA 决策循环 (+ daemon/) / Autonomous OODA loop
│ ├── quant/ # HMM 状态·Alpha·信号库 / HMM regime, alpha, signal library
│ ├── recommendation/ # 智能选股引擎 / Smart stock screener + LLM review
│ ├── trading/ # 执行闸门·A股约束·KillSwitch / Execution gates, constraints
│ ├── risk/ # 熔断·Kelly 仓位·VaR / Circuit breaker, sizing, VaR
│ ├── event_bus/ # Redis Streams 事件总线 / Event bus
│ ├── prediction/ # LLM 分析与预测 / LLM analysis & prediction
│ ├── analysis/ # 技术指标·情绪 / Technical indicators, sentiment
│ ├── llm/ # 多模型网关 / Multi-LLM gateway + router
│ ├── web/ # FastAPI 后端 / FastAPI backend
│ ├── discord_bot/ # Discord 机器人 / Discord bot
│ └── strategy/ backtest/ market_intelligence/ audit/ …
├── frontend/ # React 19 SPA (ControlTower / Portfolio / Recommendations / Review …)
├── openclaw/ # Celery 任务 + 常驻 daemon / Celery tasks (45+) + always-on daemon
├── config/ # YAML 配置文件 / YAML config files
├── mcp_server/ # MCP 数据桥 / MCP data bridge
├── research/ # 研究工作站 / Research workstation
├── tests/ # 测试 / Tests
├── docs/ # 文档 / Documentation
├── docker-compose.yaml
└── .env.example
docs/— documentation indexdocs/how-it-works.md— how the agent works + what the backtest taught us (newcomer overview)docs/guides/development-guide.md— architecture, tech stack, data flow (start here)docs/guides/runbook.md— local setup & rundocs/backtest-v2-results.md— honest out-of-sample backtest of the v2 stack (spoiler: no demonstrated alpha vs buy-and-hold)docs/testing/— test strategy & casesdocs/research-workstation-README.md— research workstation usage
This repo is built to be picked up efficiently with Claude Code (or any AI coding agent). It ships project context and ready-made commands:
CLAUDE.md— project memory loaded into context automatically: stack, architecture, setup, commands, conventions, and gotchas. Read this first..claude/commands/— project slash commands:/verify,/lint,/testrun the exact checks CI gates on..claude/settings.json— shared, secret-free settings (a safe read/test command allowlist so you get fewer permission prompts). Personal overrides go in.claude/settings.local.json(git-ignored).research/CLAUDE.md— a separate analyst-persona project root:cd research && claude.
# from the repo root
claude # start a session; CLAUDE.md context loads automatically
> /verify # run lint + unit tests + frontend build, the way CI doesLLM configuration (Claude via the Claude Code bridge or the Anthropic API, with a Gemini
fallback) lives in config/llm.yaml. Model IDs follow Anthropic's
current catalog (claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5).
LICENSE— MIT licenseCONTRIBUTING.md— contribution and verification expectationsCODE_OF_CONDUCT.md— community standardsSECURITY.md— private vulnerability reporting guidanceCHANGELOG.md— release notes.github/ISSUE_TEMPLATE/— structured bug / feature intake.github/CODEOWNERS— maintainer review ownership
MIT License — 自由使用，但请阅读上方免责声明。/ Free to use, but please read the disclaimer above.
🧪 Toy Project · 玩具项目 · For Learning Only · 仅供学习