Quantitative Models and Construction Methods - **Model Name**: AI Thematic Index Rotation Model **Model Construction Idea**: The model utilizes a full-spectrum price-volume fusion factor to score 133 thematic indices and constructs a weekly rebalancing strategy by equally allocating the top 10 thematic indices based on their scores [3][9][6] **Model Construction Process**: 1. **Thematic Index Pool**: Select thematic indices tracked by ETF funds classified by Wind, resulting in a pool of 133 thematic indices [9] 2. **Factor**: Full-spectrum price-volume fusion factor, which scores each thematic index based on the factor scores of its constituent stocks [9] 3. **Strategy Rules**: - On the last trading day of each week, select the top 10 thematic indices with the highest model scores - Allocate equally among the selected indices - Buy at the opening price of the first trading day of the following week - Weekly rebalancing with a transaction cost of 0.04% on both sides [9] **Model Evaluation**: The model demonstrates effective thematic index rotation and generates significant excess returns compared to the equal-weight benchmark [3][9] - **Model Name**: AI Concept Index Rotation Model **Model Construction Idea**: The model uses a full-spectrum price-volume fusion factor to score 72 concept indices and constructs a weekly rebalancing strategy by equally allocating the top 10 concept indices based on their scores [15][11][19] **Model Construction Process**: 1. **Concept Index Pool**: Select 72 popular concept indices from Wind [15] 2. **Factor**: Full-spectrum price-volume fusion factor, which scores each concept index based on the factor scores of its constituent stocks [15] 3. **Strategy Rules**: - On the last trading day of each week, select the top 10 concept indices with the highest model scores - Allocate equally among the selected indices - Buy at the opening price of the first trading day of the following week - Weekly rebalancing with a transaction cost of 0.04% on both sides [15] **Model Evaluation**: The model effectively identifies high-performing concept indices and generates consistent excess returns compared to the equal-weight benchmark [15][19] - **Model Name**: AI Industry Rotation Model **Model Construction Idea**: The model uses deep learning to extract information from full-spectrum price-volume data, scoring 32 primary industries and constructing a weekly rebalancing strategy by equally allocating the top 5 industries based on their scores [16][19][23] **Model Construction Process**: 1. **Industry Pool**: Includes 32 primary industries, with certain industries split into subcategories (e.g., food and beverage into food, beverages, and alcohol) [23] 2. **Factor**: Full-spectrum price-volume fusion factor, which scores each industry based on the factor scores of its constituent stocks [23] 3. **Strategy Rules**: - On the last trading day of each week, select the top 5 industries with the highest model scores - Allocate equally among the selected industries - Buy at the closing price of the first trading day of the following week - Weekly rebalancing without considering transaction costs [23] **Model Evaluation**: The model complements top-down strategies by leveraging AI's ability to extract patterns from multi-frequency price-volume data, achieving strong excess returns [16][23] - **Model Name**: AI CSI 1000 Enhanced Portfolio **Model Construction Idea**: The portfolio is constructed using the full-spectrum fusion factor to enhance the CSI 1000 index, aiming to achieve higher excess returns [27][29] **Model Construction Process**: 1. **Factor**: Full-spectrum fusion factor [29] 2. **Portfolio Construction Rules**: - Constituent stock weight must not be less than 80% - Individual stock weight deviation capped at 0.8% - Barra exposure limited to 0.3% - Weekly turnover rate controlled at 30% - Weekly rebalancing with a transaction cost of 0.4% on both sides [29] **Model Evaluation**: The portfolio demonstrates strong excess returns, high information ratio, and controlled tracking error [27][29] - **Model Name**: Text FADT_BERT Stock Selection Portfolio **Model Construction Idea**: The portfolio is based on the forecast_adjust_txt_bert factor, which is derived from upgraded text factors in earnings forecast adjustment scenarios, and selects the top 25 stocks for active quantitative enhancement [32] **Model Construction Process**: 1. **Factor**: Forecast_adjust_txt_bert factor, developed using text data related to earnings forecast adjustments [32] 2. **Portfolio Construction Rules**: - Select the top 25 stocks from the long side of the base stock pool - Active quantitative enhancement applied to the selected stocks [32] **Model Evaluation**: The portfolio achieves high annualized returns and excess returns relative to the CSI 500 index, with a strong Sharpe ratio [32] --- Model Backtesting Results - **AI Thematic Index Rotation Model** - Annualized return: 16.76% - Annualized excess return: 10.61% - Maximum drawdown of excess return: 20.79% - Excess Sharpe ratio: 0.82 - Year-to-date return: 24.22% [8] - **AI Concept Index Rotation Model** - Annualized return: 23.06% - Annualized excess return: 10.78% - Maximum drawdown of excess return: 19.48% - Excess Sharpe ratio: 0.91 - Year-to-date return: 25.27% - Year-to-date excess return: -0.98% [13] - **AI Industry Rotation Model** - Annualized return: 26.55% - Annualized excess return: 20.18% - Maximum drawdown of excess return: 12.43% - Excess Sharpe ratio: 1.96 - Year-to-date return: 23.70% - Year-to-date excess return: 1.52% [22] - **AI CSI 1000 Enhanced Portfolio** - Annualized return: 20.19% - Annualized excess return: 22.09% - Annualized tracking error: 6.07% - Maximum drawdown of excess return: 7.55% - Information ratio: 3.64 - Calmar ratio: 2.92 - Year-to-date excess return: 19.74% [27][30] - **Text FADT_BERT Stock Selection Portfolio** - Annualized return since inception: 39.96% - Annualized excess return since inception: 30.76% - Sharpe ratio: 1.39 - Year-to-date absolute return: 20.49% - Year-to-date excess return: -2.04% [32][37]

Quantitative Models and Construction Methods AI Theme Index Rotation Model - **Model Name**: AI Theme Index Rotation Model - **Model Construction Idea**: This model uses the all-frequency volume-price fusion factor to score 133 thematic indices and constructs a weekly rebalancing strategy by selecting the top 10 indices for equal-weight allocation[3][9] - **Model Construction Process**: 1. **Index Pool**: Select thematic indices tracked by thematic ETFs based on Wind's ETF classification, resulting in 133 indices[10] 2. **Factor**: Use the all-frequency volume-price fusion factor to score the constituent stocks of each thematic index[10] 3. **Strategy Rules**: - Select the top 10 indices with the highest scores on the last trading day of each week - Allocate equally among the selected indices - Execute trades at the opening price of the first trading day of the following week - Weekly rebalancing with a transaction cost of 0.04% (two-way)[10] - **Model Evaluation**: The model effectively captures thematic rotation opportunities and demonstrates strong performance relative to the benchmark[3][9] AI Concept Index Rotation Model - **Model Name**: AI Concept Index Rotation Model - **Model Construction Idea**: This model uses the all-frequency volume-price fusion factor to score 72 concept indices and constructs a weekly rebalancing strategy by selecting the top 10 indices for equal-weight allocation[11][15] - **Model Construction Process**: 1. **Index Pool**: Select 72 popular concept indices from Wind[15] 2. **Factor**: Use the all-frequency volume-price fusion factor to score the constituent stocks of each concept index[15] 3. **Strategy Rules**: - Select the top 10 indices with the highest scores on the last trading day of each week - Allocate equally among the selected indices - Execute trades at the opening price of the first trading day of the following week - Weekly rebalancing with a transaction cost of 0.04% (two-way)[15] - **Model Evaluation**: The model leverages AI-driven factor scoring to identify high-potential concept indices, achieving notable excess returns over the benchmark[11][15] AI Industry Rotation Model - **Model Name**: AI Industry Rotation Model - **Model Construction Idea**: This model applies the all-frequency volume-price fusion factor to score 32 first-level industries and constructs a weekly rebalancing strategy by selecting the top 5 industries for equal-weight allocation[4][22] - **Model Construction Process**: 1. **Industry Pool**: Includes 32 first-level industries, with certain industries split into subcategories (e.g., food and beverage, metals)[22] 2. **Factor**: Use the all-frequency volume-price fusion factor to score the constituent stocks of each industry[22] 3. **Strategy Rules**: - Select the top 5 industries with the highest scores on the last trading day of each week - Allocate equally among the selected industries - Execute trades at the closing price of the first trading day of the following week - Weekly rebalancing without transaction cost consideration[22] - **Model Evaluation**: The model complements top-down strategies by leveraging AI's ability to extract patterns from multi-frequency volume-price data[16][22] --- Model Backtesting Results AI Theme Index Rotation Model - **Annualized Return**: 16.78%[9] - **Annualized Excess Return**: 12.69%[9] - **Maximum Drawdown of Excess Return**: 16.56%[9] - **Excess Sharpe Ratio**: 1.00[9] - **YTD Return**: 10.75%[9] - **YTD Excess Return**: 3.90%[9] AI Concept Index Rotation Model - **Annualized Return**: 21.77%[13] - **Annualized Excess Return**: 11.36%[13] - **Maximum Drawdown of Excess Return**: 18.10%[13] - **Excess Sharpe Ratio**: 0.96[13] - **YTD Return**: 3.06%[13] - **YTD Excess Return**: -4.45%[13] AI Industry Rotation Model - **Annualized Return**: 25.70%[21] - **Annualized Excess Return**: 21.01%[21] - **Maximum Drawdown of Excess Return**: 12.43%[21] - **Excess Sharpe Ratio**: 2.02[21] - **YTD Return**: 5.14%[21] - **YTD Excess Return**: -1.70%[21] --- Quantitative Factors and Construction Methods All-Frequency Volume-Price Fusion Factor - **Factor Name**: All-Frequency Volume-Price Fusion Factor - **Factor Construction Idea**: This factor integrates high-frequency and low-frequency volume-price data using deep learning and multi-task learning to extract comprehensive stock selection signals[26] - **Factor Construction Process**: 1. Train 27 high-frequency factors using deep learning to obtain high-frequency deep learning factors[26] 2. Apply multi-task learning to low-frequency volume-price data for end-to-end extraction, resulting in low-frequency multi-task factors[26] 3. Combine the high-frequency and low-frequency factors into the all-frequency fusion factor[26] - **Factor Evaluation**: The factor demonstrates strong stock selection capabilities, with high RankIC and significant excess returns in backtesting[26][29] Forecast_Adjust_Text_BERT Factor - **Factor Name**: Forecast_Adjust_Text_BERT Factor - **Factor Construction Idea**: This factor upgrades text-based factors in earnings forecast adjustment scenarios using BERT to enhance stock selection performance[34] - **Factor Construction Process**: 1. Develop the forecast_adjust_txt_bert factor based on text data related to earnings forecast adjustments[34] 2. Construct a long-only portfolio using the top 25 stocks from the factor's high-score segment[34] - **Factor Evaluation**: The factor effectively captures alpha signals from textual data, achieving high returns and Sharpe ratios in backtesting[34] --- Factor Backtesting Results All-Frequency Volume-Price Fusion Factor - **5-Day RankIC Mean**: 0.114[29] - **Annualized Excess Return (Top Layer)**: 30.72%[29] - **YTD Excess Return (Top Layer)**: 3.80%[29] Forecast_Adjust_Text_BERT Factor - **Annualized Return**: 40.64%[41] - **Annualized Excess Return**: 32.25%[41] - **Annualized Volatility**: 28.77%[41] - **Maximum Drawdown**: 48.69%[41] - **Sharpe Ratio**: 1.41[41] - **Calmar Ratio**: 0.83[41]