量化策略系列之八：DFQ机器学习行业轮动模型

Quantitative Models and Factor Construction Quantitative Models and Construction Methods - Model Name: DFQ Machine Learning Industry Rotation Model - Construction Idea: Combines industry factors derived from machine learning-based stock selection factors (e.g., VAE, XGB, GP) to perform industry rotation[1][7][193] - Construction Process: 1. Train stock selection factors using machine learning models (e.g., VAE, XGB) without industry-neutralization[30][43] 2. Aggregate stock selection factors into industry factors by market-cap weighting[43] 3. Combine VAE, XGB, and GP industry factors equally to construct the DFQ Machine Learning Industry Rotation Model[165][193] - Evaluation: Demonstrates strong performance with a synergistic "1+1>2" effect, achieving high annualized excess returns and low drawdowns[165][194] - Model Name: DFQ Genetic Programming Industry Factor Mining System - Construction Idea: Utilizes genetic programming to mine industry factors with high adaptability and low correlation[51][193] - Construction Process: 1. Define fitness criteria as the minimum monthly excess return of top 5 industries across 20 paths[78] 2. Use rolling 10-year training windows and re-train annually[69][193] 3. Employ 145 features and 140 operators, including custom time-series and cross-sectional operators[74][76] 4. Optimize genetic programming with 7 key improvements, such as enhancing initial population quality and avoiding formula bloat[52][57][61] - Evaluation: Produces highly effective single factors with 60% showing positive out-of-sample monthly excess returns[194] Model Backtesting Results - DFQ Machine Learning Industry Rotation Model: - Annualized excess return: 18.42% - Maximum drawdown: -7.76% - Monthly win rate: 66.67% - Sharpe ratio of excess return: 1.77[165][167] - DFQ Genetic Programming Industry Factor Mining System (Dynamic XGB Weighted): - Annualized excess return: 11.10% - Maximum drawdown: -10.28% - Monthly win rate: 61.40% - Sharpe ratio of excess return: 1.16[128][194] Quantitative Factors and Construction Methods - Factor Name: VAE Industry Factor - Construction Idea: Derived from a Variational Autoencoder (VAE) model trained for stock selection[30][43] - Construction Process: 1. Train VAE model on stock data without industry-neutralization[30] 2. Aggregate stock-level factor values into industry-level factors using market-cap weighting[43] - Evaluation: Strong performance with annualized excess returns exceeding 10% for top 5 industries[43][165] - Factor Name: XGB Industry Factor - Construction Idea: Derived from an XGBoost model trained for stock selection[30][43] - Construction Process: 1. Train XGB model on stock data without industry-neutralization[30] 2. Aggregate stock-level factor values into industry-level factors using market-cap weighting[43] - Evaluation: Consistently high IC and RANKIC values, with annualized excess returns exceeding 10% for top 5 industries[43][165] - Factor Name: GP Industry Factor - Construction Idea: Mined using genetic programming to optimize fitness and reduce correlation[51][193] - Construction Process: 1. Define fitness as the minimum monthly excess return of top 5 industries across 20 paths[78] 2. Use rolling 10-year training windows and re-train annually[69][193] 3. Optimize genetic programming with 7 key improvements, such as enhancing initial population quality and avoiding formula bloat[52][57][61] - Evaluation: Demonstrates low correlation with other factors and strong out-of-sample performance when dynamically weighted[164][194] Factor Backtesting Results - VAE Industry Factor: - Annualized excess return: 10.01% - Maximum drawdown: -10.09% - Monthly win rate: 57.89% - Sharpe ratio of excess return: 1.03[165][166] - XGB Industry Factor: - Annualized excess return: 10.82% - Maximum drawdown: -10.61% - Monthly win rate: 59.65% - Sharpe ratio of excess return: 1.02[165][166] - GP Industry Factor (Dynamic XGB Weighted): - Annualized excess return: 11.10% - Maximum drawdown: -10.28% - Monthly win rate: 61.40% - Sharpe ratio of excess return: 1.16[128][194] Composite Factor Backtesting Results - VAE + XGB + GP (DFQ Machine Learning Industry Rotation Model): - Annualized excess return: 18.42% - Maximum drawdown: -7.76% - Monthly win rate: 66.67% - Sharpe ratio of excess return: 1.77[165][167]