基于 LLM 的因子挖掘与 ETF 轮动策略
端到端量化平台：LLM Agent 因子发现（GLM-4.5）、多层信号评估、滚动 XGBoost/线性组合、实盘 ETF 轮动交易。
A complete quantitative trading platform integrating LLM-driven factor discovery, signal evaluation, combination optimization, and live ETF rotation trading.
LLM Agent Factor Mining
Designed a multi-model Agent system (GLM-4.5, Gemini) that simulates senior quant researchers. The system constructs context-aware prompts with market state analysis, knowledge base of historical quality factors (IC > 0.03, IC IR > 0.5, Sharpe > 1.5), and innovation direction identification. LLM outputs structured JSON factor expressions parsed through safe_eval sandbox execution with AST-level security checks (future function detection, code injection prevention, complexity limits).
Signal Evaluation Pipeline
Built a multi-tier evaluation framework: Rank IC analysis, 5-layer backtesting, and comprehensive scoring (IC mean/std/IR, Sharpe, max drawdown, Calmar, turnover, coverage). Implemented strict screening: IC > 0.02, IC IR > 0.1, Sharpe > 1.0, coverage > 80%, inter-factor correlation < 0.7. OOS validation uses IC/IR decay rate filtering (50% threshold) with direction consistency checks across in-sample (pre-2020) and out-of-sample (post-2021) periods.
Signal Combination
Supports equal-weight, ICIR-weighted, rolling linear regression, rolling XGBoost, and rolling LightGBM ensemble methods. Rolling models retrain periodically to adapt to market regime changes. HRP (Hierarchical Risk Parity) portfolio optimization with Leidoit-Wolf shrinkage for covariance estimation.
ETF Rotation Strategy
Daily rebalancing across A-share on-exchange ETFs with flexible price execution (open, close, VWAP, open-VWAP 1-5min, cross-day). Supports top-N selection, inverse volatility weighting, and multiple benchmarks (HS300, CSI500, CSI1000). Complete live trading system with automated data pipeline, scheduler, and WeChat notifications.
System Architecture
Modular design with operators registry, Parquet storage, Joblib parallel evaluation, and automated data update pipeline.
核心成果
- LLM Agent factor mining (GLM-4.5 / Gemini)
- Multi-tier evaluation: IC → layer backtest → OOS validation
- Rolling XGBoost/linear/ICIR combination methods
- Live trading with scheduler & notifications