证券研究报告：金融工程报告 研究所 分析师:肖承志 SAC 登记编号:S1340524090001 Email:xiaochengzhi@cnpsec.com 研究助理:李子凯 SAC 登记编号:S1340124100014 Email:lizikai@cnpsec.com 近期研究报告 《OpenAI 发布 GPT-5，Claude Opus 4.1 上线——AI 动态汇总 20250811》 - 2025.08.12 《融资余额新高，创新药光通信调整， 指数预期仍将震荡上行挑战前高—— 行业轮动周报 20250810》 - 2025.8.11 《ETF 资金偏谨慎流入消费红利防守， 银行提前调整使指数回调空间可控— — 行 业 轮 动 周 报 20250803 》 - 2025.08.04 《ETF 资金持续净流出医药，雅下水电 站成短线情绪突破口——行业轮动周 报 20250727》 – 2025.07.28 《ETF 资金净流入红利流出高位医药， 指数与大金融回调有明显托底——行 业轮动周报 20250720》 – 2025.07.21 《大金融表现居前助指数突破，GRU 行 业轮动调入非银行金融—— ...

- Model Name: Diffusion Index Model; Model Construction Idea: The model is based on the principle of price momentum; Model Construction Process: The model tracks the weekly and monthly changes in the diffusion index of various industries, ranking them accordingly. The formula used is $ \text{Diffusion Index} = \frac{\text{Number of Upward Trends}}{\text{Total Number of Trends}} $; Model Evaluation: The model has shown varying performance over the years, with significant returns in some periods and notable drawdowns in others[27][28][31] - Model Name: GRU Factor Model; Model Construction Idea: The model utilizes GRU deep learning networks to analyze minute-level volume and price data; Model Construction Process: The model ranks industries based on GRU factors, which are derived from deep learning algorithms processing historical trading data. The formula used is $ \text{GRU Factor} = \text{GRU Network Output} $; Model Evaluation: The model performs well in short cycles but has mixed results in longer cycles[33][34][36] - Diffusion Index Model, Average Weekly Return: 2.06%, Excess Return: -0.00%, August Excess Return: -0.45%, Year-to-Date Excess Return: -0.41%[31] - GRU Factor Model, Average Weekly Return: 2.71%, Excess Return: 0.65%, August Excess Return: 0.32%, Year-to-Date Excess Return: -4.35%[36] - Factor Name: GRU Industry Factor; Factor Construction Idea: The factor is derived from GRU deep learning networks analyzing minute-level trading data; Factor Construction Process: The factor ranks industries based on GRU network outputs, which are calculated from historical volume and price data. The formula used is $ \text{GRU Factor} = \text{GRU Network Output} $; Factor Evaluation: The factor has shown significant changes in rankings, indicating its sensitivity to market conditions[6][14][34] - GRU Industry Factor, Steel: 2.82, Building Materials: 1.72, Transportation: 1.3, Oil & Petrochemicals: 0.27, Construction: -0.46, Comprehensive: -1.87[6][14][34]

- Model Name: Diffusion Index Model; Construction Idea: The model is based on the principle of price momentum, capturing industry trends through diffusion indices; Construction Process: The model tracks the weekly and monthly changes in the diffusion indices of various industries, ranking them accordingly. The formula for the diffusion index is not explicitly provided; Evaluation: The model has shown varying performance over the years, with significant drawdowns during market reversals[24][25][28] - Model Name: GRU Factor Model; Construction Idea: The model utilizes GRU (Gated Recurrent Unit) deep learning networks to process minute-level volume and price data, aiming to capture trading information; Construction Process: The model ranks industries based on GRU factors, which are derived from the deep learning network's analysis of trading data. The specific formula for GRU factors is not provided; Evaluation: The model has performed well in short cycles but has shown general performance in longer cycles[31][32][35] - Diffusion Index Model, Average Weekly Return: 0.89%, Excess Return Since July: -3.47%, Excess Return YTD: -0.45%[28] - GRU Factor Model, Average Weekly Return: 4.27%, Excess Return Since July: 1.34%, Excess Return YTD: -4.25%[35] - Factor Name: Diffusion Index; Construction Idea: The factor is based on the momentum of industry prices, capturing upward trends; Construction Process: The factor is calculated by observing the weekly and monthly changes in the diffusion indices of various industries. The specific formula is not provided; Evaluation: The factor has shown varying performance, with significant drawdowns during market reversals[24][25][28] - Factor Name: GRU Factor; Construction Idea: The factor is derived from GRU deep learning networks, capturing trading information from minute-level volume and price data; Construction Process: The factor is calculated by ranking industries based on the GRU network's analysis of trading data. The specific formula is not provided; Evaluation: The factor has performed well in short cycles but has shown general performance in longer cycles[31][32][35] - Diffusion Index Factor, Top Industries: Comprehensive Finance (1.0), Steel (1.0), Non-Bank Finance (0.999), Comprehensive (0.998), Non-Ferrous Metals (0.997), Home Appliances (0.995)[25] - GRU Factor, Top Industries: Banking (3.3), Real Estate (0.58), Oil & Petrochemicals (-1.26), Textile & Apparel (-1.73), Light Manufacturing (-2.49), Electric Power & Utilities (-2.83)[32]

- Model Name: Diffusion Index Model; Model Construction Idea: The model is based on the principle of price momentum; Model Construction Process: The model tracks the weekly changes in the diffusion index of various industries, ranking them based on their diffusion index values. The formula used is $ \text{Diffusion Index} = \frac{\text{Number of Stocks with Positive Momentum}}{\text{Total Number of Stocks}} $; Model Evaluation: The model captures industry trends effectively but may face challenges during market reversals[5][27][28] - Model Name: GRU Factor Model; Model Construction Idea: The model utilizes GRU (Gated Recurrent Unit) deep learning networks to analyze minute-level price and volume data; Model Construction Process: The model ranks industries based on their GRU factor values, which are derived from the GRU network's analysis of trading information. The formula used is $ \text{GRU Factor} = \text{GRU Network Output} $; Model Evaluation: The model performs well in short cycles but may struggle in long cycles or extreme market conditions[6][13][33] - Diffusion Index Model, IR value 2.05%, weekly average return 0.24%, monthly excess return -1.00%, annual excess return 2.05%[25][30] - GRU Factor Model, IR value -4.52%, weekly average return 1.32%, monthly excess return 0.77%, annual excess return -4.52%[32][37] - Factor Name: GRU Industry Factor; Factor Construction Idea: The factor is derived from GRU deep learning networks analyzing minute-level trading data; Factor Construction Process: The factor values are calculated based on the GRU network's output, ranking industries accordingly. The formula used is $ \text{GRU Factor} = \text{GRU Network Output} $; Factor Evaluation: The factor captures short-term trading information effectively but may face challenges in long-term or extreme market conditions[6][13][33] - GRU Industry Factor, IR value -4.52%, weekly average return 1.32%, monthly excess return 0.77%, annual excess return -4.52%[32][37]

- The diffusion index model tracks industry rotation and has achieved an excess return of 0.37% since 2025[26][27][31] - The diffusion index ranks industries weekly based on momentum, with top industries including non-bank finance (1.0), comprehensive finance (1.0), and media (0.976)[4][28][30] - The diffusion index suggests monthly industry allocation, recommending sectors such as non-bank finance, banking, and media for June 2025[27][31] - GRU factor model focuses on industry rotation based on transaction data, achieving an excess return of -4.76% in 2025[33][36][34] - GRU factor ranks industries weekly, with top industries including textile & apparel (3.7), construction (3.34), and real estate (3.28)[5][13][34] - GRU factor suggests weekly industry allocation, recommending sectors such as real estate, transportation, and coal for the current week[36][34][33]

Quantitative Models and Construction Methods 1. Model Name: Diffusion Index Model - **Model Construction Idea**: The model is based on the principle of price momentum, aiming to capture upward trends in industry performance[27][28] - **Model Construction Process**: The diffusion index is calculated for each industry, ranking them based on their momentum. Industries with higher diffusion index values are considered to have stronger upward trends. The model selects industries with the highest diffusion index values for allocation. - Formula: Not explicitly provided in the report - **Model Evaluation**: The model has shown mixed performance over the years. It performed well in 2021 and 2022 but faced significant drawdowns in 2023 and 2024 due to market reversals and failure to adjust to cyclical changes[27] 2. Model Name: GRU Factor Model - **Model Construction Idea**: This model leverages GRU (Gated Recurrent Unit) deep learning networks to process high-frequency price and volume data, aiming to identify industry trends and generate excess returns[34][39] - **Model Construction Process**: The GRU network is trained on historical minute-level price and volume data to predict industry rankings. The model then allocates to industries with the highest GRU factor scores. - Formula: Not explicitly provided in the report - **Model Evaluation**: The model has shown strong adaptability in short-term cycles but struggles in long-term trends and extreme market conditions. It has faced challenges in capturing excess returns in 2025 due to concentrated market themes[34][39] --- Model Backtesting Results 1. Diffusion Index Model - **2025 YTD Excess Return**: 0.37%[26][31] - **June 2025 Excess Return**: 1.99%[31] - **Weekly Average Return (June 2025)**: -0.65%[31] - **Weekly Excess Return (June 2025)**: 0.79%[31] 2. GRU Factor Model - **2025 YTD Excess Return**: -3.83%[34][37] - **June 2025 Excess Return**: 0.25%[37] - **Weekly Average Return (June 2025)**: -1.18%[37] - **Weekly Excess Return (June 2025)**: 0.25%[37] --- Quantitative Factors and Construction Methods 1. Factor Name: Diffusion Index - **Factor Construction Idea**: Measures the momentum of industries by ranking them based on their upward trends[28] - **Factor Construction Process**: The diffusion index is calculated for each industry weekly. Industries are ranked based on their index values, with higher values indicating stronger momentum. - Example Values (as of June 20, 2025): - Top Industries: Comprehensive Finance (1.0), Non-Bank Finance (0.973), Banking (0.97)[28] - Bottom Industries: Coal (0.174), Food & Beverage (0.313), Oil & Gas (0.387)[28] - **Factor Evaluation**: The factor effectively captures upward trends but may underperform during market reversals[27][28] 2. Factor Name: GRU Factor - **Factor Construction Idea**: Utilizes GRU deep learning to analyze high-frequency trading data and rank industries based on predicted performance[34][39] - **Factor Construction Process**: The GRU network processes minute-level price and volume data to generate factor scores for each industry. Industries are ranked based on these scores. - Example Values (as of June 20, 2025): - Top Industries: Coal (3.48), Non-Bank Finance (3.15), Utilities (2.65)[35] - Bottom Industries: Communication (-17.95), Media (-15.45), Defense (-11.87)[35] - **Factor Evaluation**: The factor is effective in short-term trend identification but struggles with long-term stability and extreme market conditions[34][39] --- Factor Backtesting Results 1. Diffusion Index Factor - **Top Weekly Changes (June 20, 2025)**: - Oil & Gas: +0.09 - Textiles: +0.044 - Metals: +0.036[30] - **Bottom Weekly Changes (June 20, 2025)**: - Agriculture: -0.229 - Defense: -0.086 - Building Materials: -0.078[30] 2. GRU Factor - **Top Weekly Changes (June 20, 2025)**: - Non-Bank Finance: Significant increase - Consumer Services: Significant increase - Comprehensive: Significant increase[35] - **Bottom Weekly Changes (June 20, 2025)**: - Communication: Significant decrease - Electronics: Significant decrease - New Energy Equipment: Significant decrease[35]

- Model Name: Diffusion Index Model; Model Construction Idea: The model is based on the observation of industry diffusion indices; Detailed Construction Process: The model tracks the weekly changes in diffusion indices for various industries, ranking them based on their performance. The formula used is $ \text{Diffusion Index} = \frac{\text{Number of Advancing Stocks}}{\text{Total Number of Stocks}} $; Model Evaluation: The model has shown varying performance over the years, with significant returns in some periods and notable drawdowns in others[6][14][27] - Model Name: GRU Factor Model; Model Construction Idea: The model utilizes GRU (Gated Recurrent Unit) deep learning networks to analyze minute-level price and volume data; Detailed Construction Process: The model ranks industries based on GRU factor scores, which are derived from the GRU network's analysis of trading data. The formula used is $ \text{GRU Factor Score} = \text{GRU Network Output} $; Model Evaluation: The model has achieved substantial excess returns by capturing trading information, though it has faced challenges in certain market conditions[7][14][34] Model Backtest Results - Diffusion Index Model, Average Weekly Return: 0.72%, Excess Return: 0.11%, Year-to-Date Excess Return: -2.26%[32] - GRU Factor Model, Average Weekly Return: 1.07%, Excess Return: 0.44%, Year-to-Date Excess Return: -3.71%[37] Factor Construction and Evaluation - Factor Name: GRU Industry Factor; Factor Construction Idea: The factor is constructed using GRU deep learning networks to analyze minute-level trading data; Detailed Construction Process: The factor scores are calculated based on the GRU network's output, which evaluates the trading data to rank industries. The formula used is $ \text{GRU Factor Score} = \text{GRU Network Output} $; Factor Evaluation: The factor has shown significant improvements in certain industries, indicating its effectiveness in capturing trading information[7][14][35] Factor Backtest Results - GRU Industry Factor, Top Industries: Automotive (2.84), Steel (1.85), Media (1.48), Power Equipment and New Energy (1.35), Communication (0.88), Coal (0.66)[7][14][35]

Quantitative Models and Construction 1. Model Name: Diffusion Index Model - **Model Construction Idea**: The model is based on the principle of price momentum, aiming to capture upward trends in industry performance[28][38] - **Model Construction Process**: The model calculates the diffusion index for each industry, ranking them based on their relative performance. Industries with higher diffusion indices are recommended for allocation. The model tracks weekly and monthly changes in the diffusion index to adjust allocations dynamically[5][14][29] - **Model Evaluation**: The model has shown strong performance in capturing momentum trends during upward markets but may underperform during market reversals[28][38] 2. Model Name: GRU Factor Model - **Model Construction Idea**: This model leverages GRU (Gated Recurrent Unit) deep learning networks to process high-frequency volume and price data, aiming to identify industry rotation opportunities[39] - **Model Construction Process**: The GRU network is trained on historical minute-level data to predict industry factor rankings. The model dynamically adjusts allocations based on the predicted rankings, focusing on industries with higher GRU factor scores[6][34][39] - **Model Evaluation**: The model performs well in short-term scenarios due to its adaptability but may face challenges in long-term or extreme market conditions[39] --- Backtesting Results of Models 1. Diffusion Index Model - **2025 YTD Excess Return**: -2.75%[27][32] - **April 2025 Excess Return**: -0.68%[32] - **Weekly Portfolio Return**: -0.18%[32] 2. GRU Factor Model - **2025 YTD Excess Return**: -3.54%[34][37] - **April 2025 Excess Return**: 0.68%[37] - **Weekly Portfolio Return**: -0.78%[37] --- Quantitative Factors and Construction 1. Factor Name: Diffusion Index - **Factor Construction Idea**: Measures the breadth of industry performance to identify upward trends[5][14] - **Factor Construction Process**: The diffusion index is calculated as the proportion of stocks in an industry with positive momentum. Weekly and monthly changes in the index are tracked to adjust rankings dynamically[5][14][29] - **Factor Evaluation**: Effective in capturing momentum trends but sensitive to market reversals[28][38] 2. Factor Name: GRU Industry Factor - **Factor Construction Idea**: Utilizes GRU deep learning to analyze high-frequency trading data and predict industry rankings[39] - **Factor Construction Process**: The GRU network processes minute-level volume and price data to generate factor scores for industries. Industries with higher scores are prioritized for allocation[6][34][39] - **Factor Evaluation**: Strong adaptability in short-term scenarios but limited in long-term or extreme market conditions[39] --- Backtesting Results of Factors 1. Diffusion Index Factor - **Top 6 Industries (as of April 30, 2025)**: Banking (0.988), Non-Banking Financials (0.94), Comprehensive Financials (0.928), Computers (0.884), Retail (0.88), Automobiles (0.872)[5][14][29] - **Weekly Change Leaders**: Steel (0.17), Comprehensive (0.095), Automobiles (0.065)[5][31] 2. GRU Industry Factor - **Top 6 Industries (as of April 30, 2025)**: Real Estate (4.62), Textiles & Apparel (4.14), Comprehensive Financials (2.89), Transportation (1.71), Light Manufacturing (1.7), Construction (1.41)[6][35] - **Weekly Change Leaders**: Pharmaceuticals, Real Estate, Comprehensive Financials[6][35]

证券研究报告：金融工程报告 发布时间：2025-04-28 研究所 分析师:肖承志 SAC 登记编号:S1340524090001 Email:xiaochengzhi@cnpsec.com 研究助理:李子凯 SAC 登记编号:S1340124100014 Email:lizikai@cnpsec.com 近期研究报告 《OpenAI 发布 GPT-4.1，智谱发布 GLM- 4-32B-0414 系 列 — — AI 动态汇总 20250421》 - 2025.04.23 《国家队交易特征显著，短期指数仍交 易补缺预期，TMT 类题材仍需等待—— 行业轮动周报 20250427》- 2025.04.21 《小市值强势，动量风格占优——中邮 因子周报 20250420》 – 2025.04.21 《基本面与量价共振，如遇回调即是买 点——微盘股指数周报 20250420》 – 2025.04.21 《Meta LIama 4 开源，OpenAI 启动先 锋计划——AI 动态汇总 20250414》 - 2025.04.15 《融资盘被动爆仓导致大幅净流出， GRU 模型仍未配置成长——行业轮动周 报 2025 ...