AI-driven quantitative factor mining for China's A-share market.
Let AI generate, validate, score, backtest and store quantitative factors — you only set the research direction.
FactorLab 是一套全自动 A 股量化因子挖掘系统。核心理念：
人类设定研究方向 → AI 生成因子代码 → AST 安全沙箱执行 → 8 维量化评分 → 冲击成本回测 → 4 维稳健性检验 → 自动入库
你只需告诉它"我要找反转类因子"，系统便自动生成 Python 代码，在 1,482 只 A 股、15 年历史 上跑回测，全部通过则自动存入数据库。跑完 50 轮后自动生成一份带图表的 HTML 研究报告。
| 特性 | 说明 |
|---|---|
| AI 自动生成因子 | 接入 DeepSeek API，根据研究方向自动生成 Python 因子代码 |
| AST 安全沙箱 | 纯语法树级别代码审查，4 类危险模式拦截，零误报 |
| 8 维评分关口 | IC、IR、覆盖率、相关性、换手率、方向性、稳定性、衰减比 — 全部达标才入库 |
| 真实冲击成本 | 逐日逐股计算冲击成本，非固定扣费，大盘股 < 0.05%，小盘股 > 0.5% |
| 4 维稳健性检验 | 单调性、样本外稳定性、IC 衰减、分年验证 |
| 多样性门控 | 新因子与库存因子 Spearman 相关 > 0.70 自动拒绝，防止同质化 |
| 全自动批量 | 一键 50 轮，双层熔断 + 断点续跑，单轮成本 ≈ ¥0.10 |
| HTML 报告 | 单文件 base64 内嵌图表，双击打开，无需网络 |
| 自然语言检索 | "有哪些量价背离的因子？"→ 自动分词匹配 |
# 1. 克隆仓库
git clone https://github.com/YippeeXu/FactorLab.git
cd FactorLab
# 2. 创建虚拟环境（推荐）
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# 3. 安装依赖
pip install -r requirements.txt
# 4. 配置 API Key
cp .env.example .env
# 编辑 .env，填入你的 DEEPSEEK_API_KEY
# 5. 下载数据（可选，首次约 10 分钟）
python scripts/download_data.py --fetch
# 6. 运行演示（不消耗 API 额度）
python src/main.py --batch 5 --dry-run
# 7. 实战挖因子
python src/main.py --batch 50FactorLab/
├── src/ # 核心源码（29 个 Python 模块）
│ ├── main.py # CLI 入口
│ ├── engine.py # AI 因子生成引擎
│ ├── checker.py # AST 安全审查
│ ├── sandbox.py # 代码执行沙箱
│ ├── score.py # 8 维评分系统
│ ├── backtest.py # 冲击成本回测
│ ├── database.py # SQLite 数据库
│ └── ...
├── tests/ # 测试套件
├── docs/ # 文档
│ ├── QUICKSTART.md # 快速上手
│ ├── ARCHITECTURE.md # 系统架构
│ ├── API.md # 接口文档
│ └── FAQ.md # 常见问题
├── examples/ # 示例报告
├── scripts/ # 工具脚本
│ ├── download_data.py # 数据下载
│ └── schema.sql # 数据库结构
├── data/ # A 股数据（.gitignore 忽略）
└── requirements.txt
A 股历史行情数据来源于 AKShare（基于东方财富/同花顺公开数据）。首次运行需下载约 169MB（1,482 只 × 15 年日线 OHLCV）。数据仅供研究学习，不构成投资建议。
- Python 3.9+
- pandas / numpy（数据处理）
- AKShare（A 股数据）
- DeepSeek API（因子生成）
- SQLite + WAL（因子数据库）
- matplotlib（图表生成）
| 层级 | 机制 |
|---|---|
| 代码安全 | AST 语法树审查，4 类危险模式拦截，白名单验证 |
| 执行隔离 | 沙箱超时 30s，异常不崩溃主程序 |
| 权力分割 | 人类锁死核心规程，AI 只能写建议 |
| 熔断机制 | 单轮 3 次失败 + 连续 10 轮无入库自动暂停 |
| 防注入 | 100% 参数化 SQL |
欢迎提交 Issue 和 Pull Request！
- Fork 本仓库并创建分支：
git checkout -b feature/my-feature - 确保代码通过测试：
pytest tests/ -v - 遵循现有代码风格（PEP 8）
- 提交 PR，描述清楚改动内容和动机
详细规范请参见 docs/ARCHITECTURE.md。
MIT License — 2026 YippeeXu
FactorLab is a fully automated quantitative factor mining system for China's A-share market. The core idea:
Human sets direction → AI generates factor code → AST sandbox execution → 8-dimension scoring → Impact cost backtest → 4-dimension robustness check → Auto storage
Just tell it "I want reversal factors", and the system generates Python code, runs backtests on 1,482 stocks over 15 years, and stores passing factors automatically. After 50 rounds, it generates an HTML research report with embedded charts.
| Feature | Description |
|---|---|
| AI Factor Generation | DeepSeek API integration, generates Python factor code from research directions |
| AST Security Sandbox | Pure syntax-tree level code review, 4-category danger interception, zero false positives |
| 8-Dimension Scoring | IC, IR, coverage, correlation, turnover, directionality, stability, decay ratio — all must pass |
| Real Impact Cost | Per-day per-stock impact cost calculation, not fixed fee deduction |
| 4-Dimension Robustness | Monotonicity, out-of-sample stability, IC decay, yearly validation |
| Diversity Gate | Auto-reject new factors with Spearman correlation > 0.70 to existing ones |
| Full Auto Batch | One-click 50 rounds, dual fuse + resume support, ~¥0.10 per round |
| HTML Report | Single-file base64 embedded charts, open offline |
| NL Query | "Show me divergence factors" → auto token matching |
# 1. Clone
git clone https://github.com/YippeeXu/FactorLab.git
cd FactorLab
# 2. Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# 3. Install
pip install -r requirements.txt
# 4. Configure API Key
cp .env.example .env
# Edit .env with your DEEPSEEK_API_KEY
# 5. Download data (first run, ~10 min)
python scripts/download_data.py --fetch
# 6. Demo (no API cost)
python src/main.py --batch 5 --dry-run
# 7. Mine factors
python src/main.py --batch 50FactorLab/
├── src/ # Core source (29 Python modules)
│ ├── main.py # CLI entry
│ ├── engine.py # AI factor generation engine
│ ├── checker.py # AST security review
│ ├── sandbox.py # Code execution sandbox
│ ├── score.py # 8-dimension scoring
│ ├── backtest.py # Impact cost backtest
│ ├── database.py # SQLite database
│ └── ...
├── tests/ # Test suite
├── docs/ # Documentation
│ ├── QUICKSTART.md # Quick start guide
│ ├── ARCHITECTURE.md # System architecture
│ ├── API.md # API docs
│ └── FAQ.md # FAQ
├── examples/ # Sample reports
├── scripts/ # Utility scripts
│ ├── download_data.py # Data download
│ └── schema.sql # Database schema
├── data/ # A-share data (.gitignore ignored)
└── requirements.txt
Historical A-share data from AKShare (based on East Money/Flush public data). First download ~169MB (1,482 stocks × 15 years daily OHLCV). For research only, not investment advice.
- Python 3.9+
- pandas / numpy (data processing)
- AKShare (A-share data)
- DeepSeek API (factor generation)
- SQLite + WAL (factor database)
- matplotlib (chart generation)
| Layer | Mechanism |
|---|---|
| Code Safety | AST syntax tree review, 4-category danger interception, whitelist validation |
| Execution Isolation | Sandbox timeout 30s, exceptions don't crash main program |
| Power Separation | Humans lock core protocols, AI can only write suggestions |
| Fuse Mechanism | Single round 3 failures + 10 consecutive rounds without storage auto-pause |
| Anti-Injection | 100% parameterized SQL |
Issues and Pull Requests are welcome!
- Fork this repo and create a branch:
git checkout -b feature/my-feature - Ensure tests pass:
pytest tests/ -v - Follow existing code style (PEP 8)
- Submit PR with clear description of changes and motivation
See docs/ARCHITECTURE.md for detailed guidelines.
MIT License — 2026 YippeeXu