Quantitative Models and Construction Methods AI Full-Frequency Volume-Price Model - **Model Name**: AI Full-Frequency Volume-Price Model - **Model Construction Idea**: Utilize various frequencies of volume-price information in the market, including low-frequency data like daily, weekly, and monthly K-line data, as well as high-frequency data like minute lines, transaction-by-transaction, and order-by-order data[149] - **Model Construction Process**: Preprocess multi-frequency data and input it into the AI model for training, ultimately outputting stock selection signals "full-frequency fusion factor" which predicts the relative market excess return of individual stocks over the next 10 trading days[149] - **Model Evaluation**: The model has shown stable outperformance against the benchmark index CSI 1000 since the beginning of the backtest in 2017, with an annualized excess return rate of 21.86%, annualized tracking error of 6.05%, IR of 3.62, maximum drawdown of excess return of 7.55%, and Calmar ratio of excess return of 2.89[150][154] Master Factor - **Factor Name**: Master Factor - **Factor Construction Idea**: Based on the Master model with GRU and three-layer attention, introducing market characteristic variables through linear gating/attention gating modules to obtain the Master factor[156] - **Factor Construction Process**: The factor is derived from the Master model, which uses GRU to process time-series data and attention mechanisms to capture market characteristics. The factor's performance is evaluated based on IC, RankIC, and excess returns[156][157] - **Factor Evaluation**: The Master factor has shown sustained superior performance in out-of-sample tests, with an IC mean of 12.1%, RankIC mean of 14.6%, RankICIR of 1.37, and a RankIC win rate of 90.5%[161][162] Sentiment Triage Factor - **Factor Name**: Sentiment Triage Factor - **Factor Construction Idea**: Integrate the sentiment analysis capabilities of large language models (LLM) into AI volume-price models, dynamically selecting AI volume-price expert routing based on market sentiment[170] - **Factor Construction Process**: The factor combines alternative sentiment information with volume-price data, enhancing index enhancement portfolio performance. The factor is detailed in the report "LLMRouter-GRU: Sentiment Triage Empowering AI Volume-Price Factor"[170] - **Factor Evaluation**: The Sentiment Triage factor has shown promising results, with an annualized excess return rate of 8.47%, annualized tracking error of 4.59%, IR of 1.85, maximum drawdown of excess return of 4.84%, and Calmar ratio of excess return of 1.75[171][172] PortfolioNet2.0 Factor - **Factor Name**: PortfolioNet2.0 Factor - **Factor Construction Idea**: Introduce style models into the network, endowing the combination constraint items with differentiable capabilities, allowing AI volume-price factors to pursue high return elasticity and capture style returns in addition to Pure Alpha[174] - **Factor Construction Process**: The factor is derived from the PortfolioNet2.0 model, which integrates style models into the network to enhance the combination constraint items with differentiable capabilities, aiming to capture both Pure Alpha and style returns[174] - **Factor Evaluation**: The PortfolioNet2.0 factor has shown strong performance in the backtest, with an annualized excess return rate of 11.54%, annualized tracking error of 6.51%, IR of 1.77, maximum drawdown of excess return of 8.39%, and Calmar ratio of excess return of 1.38[175][176] LLM-FADT Text Strategy - **Factor Name**: LLM-FADT Text Strategy - **Factor Construction Idea**: Utilize large language models to enhance text-based stock selection by deeply analyzing analyst reports and extracting implicit information[178] - **Factor Construction Process**: The strategy involves posing multiple questions to the large language model regarding the analyst report's title and summary, requiring the model to provide insights on core information, potential risks, and future stock return guidance[178] - **Factor Evaluation**: The LLM-FADT strategy has shown stable performance, with an annualized return rate of 28.93%, annualized excess return rate of 26.43%, Sharpe ratio of 1.13, and IR of 2.08[179][180] Model Backtest Results AI Full-Frequency Volume-Price Model - **Annualized Return Rate**: 20.37% - **Annualized Volatility**: 23.31% - **Sharpe Ratio**: 0.87 - **Maximum Drawdown**: 33.08% - **Annualized Excess Return Rate**: 21.86% - **Annualized Tracking Error**: 6.05% - **Maximum Drawdown of Excess Return**: 7.55% - **IR**: 3.62 - **Calmar Ratio of Excess Return**: 2.89 - **Monthly Win Rate**: 79.25% - **Annualized Turnover Rate**: 32.57%[154] Master Factor - **IC Mean**: 12.1% - **RankIC Mean**: 14.6% - **RankICIR**: 1.37 - **RankIC Win Rate**: 90.5% - **Hedge Group Return**: 65.4% - **Long Group Excess Return**: 36.6% - **Short Group Excess Return**: -50.7% - **Long Group IR**: 5.35 - **Short Group IR**: -4.21[161] Sentiment Triage Factor - **Annualized Excess Return Rate**: 8.47% - **Annualized Tracking Error**: 4.59% - **IR**: 1.85 - **Maximum Drawdown of Excess Return**: 4.84% - **Calmar Ratio of Excess Return**: 1.75 - **Relative Benchmark Monthly Win Rate**: 75.00% - **Annualized Turnover Rate**: 10.51%[171] PortfolioNet2.0 Factor - **Annualized Excess Return Rate**: 11.54% - **Annualized Tracking Error**: 6.51% - **IR**: 1.77 - **Maximum Drawdown of Excess Return**: 8.39% - **Calmar Ratio of Excess Return**: 1.38 - **Relative Benchmark Monthly Win Rate**: 77.78% - **Annualized Turnover Rate**: 20.66%[175] LLM-FADT Text Strategy - **Annualized Return Rate**: 28.93% - **Annualized Volatility**: 25.63% - **Sharpe Ratio**: 1.13 - **Maximum Drawdown**: 36.70% - **Calmar Ratio**: 0.79 - **Annualized Excess Return Rate**: 26.43% - **Annualized Tracking Error**: 12.74% - **IR**: 2.08 - **Maximum Drawdown of Excess Return**: 22.89% - **Calmar Ratio of Excess Return**: 1.15 - **Monthly Win Rate**: 75.70% - **Annualized Turnover Rate**: 14.79%[180]