Quantitative Models and Construction 1. Model Name: barra1d - **Model Construction Idea**: This model is part of the GRU factor family and is designed to capture short-term market dynamics through daily data inputs[4][6][8] - **Model Construction Process**: The barra1d model uses daily market data to calculate factor exposures and returns. It applies industry-neutralization and standardization processes to ensure comparability across stocks. The model is rebalanced monthly, selecting the top 10% of stocks with the highest factor scores for long positions and the bottom 10% for short positions, with equal weighting[17][28][29] - **Model Evaluation**: The barra1d model demonstrated strong performance in multiple stock pools, showing resilience in volatile market conditions[4][6][8] 2. Model Name: barra5d - **Model Construction Idea**: This model extends the barra1d framework to a five-day horizon, aiming to capture slightly longer-term market trends[4][6][8] - **Model Construction Process**: Similar to barra1d, the barra5d model uses five-day aggregated data for factor calculation. It follows the same industry-neutralization, standardization, and rebalancing processes as barra1d[17][28][29] - **Model Evaluation**: The barra5d model experienced significant drawdowns in recent periods, indicating sensitivity to market reversals[4][6][8] 3. Model Name: open1d - **Model Construction Idea**: This model focuses on open price data to identify short-term trading opportunities[4][6][8] - **Model Construction Process**: The open1d model calculates factor exposures based on daily opening prices. It applies the same industry-neutralization and rebalancing methodology as other GRU models[17][28][29] - **Model Evaluation**: The open1d model showed moderate performance, with some drawdowns in recent periods[4][6][8] 4. Model Name: close1d - **Model Construction Idea**: This model emphasizes closing price data to capture end-of-day market sentiment[4][6][8] - **Model Construction Process**: The close1d model uses daily closing prices for factor calculation. It follows the same construction and rebalancing methodology as other GRU models[17][28][29] - **Model Evaluation**: The close1d model demonstrated stable performance, with positive returns in certain stock pools[4][6][8] --- Model Backtesting Results 1. barra1d Model - Weekly Excess Return: +0.57%[29][30] - Monthly Excess Return: +0.75%[29][30] - Year-to-Date Excess Return: +4.38%[29][30] 2. barra5d Model - Weekly Excess Return: -2.17%[29][30] - Monthly Excess Return: -3.76%[29][30] - Year-to-Date Excess Return: +4.13%[29][30] 3. open1d Model - Weekly Excess Return: -0.97%[29][30] - Monthly Excess Return: -2.85%[29][30] - Year-to-Date Excess Return: +4.20%[29][30] 4. close1d Model - Weekly Excess Return: -1.68%[29][30] - Monthly Excess Return: -4.50%[29][30] - Year-to-Date Excess Return: +1.90%[29][30] --- Quantitative Factors and Construction 1. Factor Name: Beta - **Factor Construction Idea**: Measures historical market sensitivity of a stock[15] - **Factor Construction Process**: Calculated as the regression coefficient of a stock's returns against market returns over a specified period[15] 2. Factor Name: Size - **Factor Construction Idea**: Captures the size effect, where smaller firms tend to outperform larger ones[15] - **Factor Construction Process**: Defined as the natural logarithm of total market capitalization[15] 3. Factor Name: Momentum - **Factor Construction Idea**: Identifies stocks with strong recent performance[15] - **Factor Construction Process**: Combines historical excess return mean, volatility, and cumulative deviation into a weighted formula: $ Momentum = 0.74 * \text{Volatility} + 0.16 * \text{Cumulative Deviation} + 0.10 * \text{Residual Volatility} $[15] 4. Factor Name: Volatility - **Factor Construction Idea**: Measures the risk or variability in stock returns[15] - **Factor Construction Process**: Weighted combination of historical residual volatility and other measures[15] 5. Factor Name: Valuation - **Factor Construction Idea**: Captures the value effect, where undervalued stocks tend to outperform[15] - **Factor Construction Process**: Defined as the inverse of the price-to-book ratio[15] 6. Factor Name: Liquidity - **Factor Construction Idea**: Measures the ease of trading a stock[15] - **Factor Construction Process**: Weighted combination of turnover rates over monthly, quarterly, and yearly horizons: $ Liquidity = 0.35 * \text{Monthly Turnover} + 0.35 * \text{Quarterly Turnover} + 0.30 * \text{Yearly Turnover} $[15] 7. Factor Name: Profitability - **Factor Construction Idea**: Identifies stocks with strong earnings performance[15] - **Factor Construction Process**: Weighted combination of various profitability metrics, including analyst forecasts and financial ratios[15] 8. Factor Name: Growth - **Factor Construction Idea**: Captures the growth potential of a stock[15] - **Factor Construction Process**: Weighted combination of earnings and revenue growth rates[15] --- Factor Backtesting Results 1. Beta Factor - Weekly Return: +0.14%[21] - Monthly Return: +1.65%[21] - Year-to-Date Return: +5.29%[21] 2. Size Factor - Weekly Return: +0.36%[21] - Monthly Return: +1.00%[21] - Year-to-Date Return: +6.37%[21] 3. Momentum Factor - Weekly Return: +2.21%[24] - Monthly Return: +8.80%[24] - Year-to-Date Return: +23.30%[24] 4. Volatility Factor - Weekly Return: +2.82%[24] - Monthly Return: +12.29%[24] - Year-to-Date Return: +25.25%[24] 5. Valuation Factor - Weekly Return: +1.47%[21] - Monthly Return: +2.30%[21] - Year-to-Date Return: -2.26%[21] 6. Liquidity Factor - Weekly Return: +1.80%[21] - Monthly Return: +5.91%[21] - Year-to-Date Return: +19.70%[21] 7. Profitability Factor - Weekly Return: +4.57%[21] - Monthly Return: +7.53%[21] - Year-to-Date Return: +27.56%[21] 8. Growth Factor - Weekly Return: +2.76%[24] - Monthly Return: +6.51%[24] - Year-to-Date Return: +14.51%[24]

二 〇 二 五 年 度 2025 年 08 月 28 日 国债期货系列报告：多通道深度学习模型 在国债期货因子择时上的应用 | 虞堪 | 投资咨询从业资格号：Z0002804 | yukan@gtht.com | | --- | --- | --- | | 宋子钰（联系人） | 从业资格证号：F03136034 | songziyu@gtht.com | 报告导读： 本报告直面传统量化因子在当前债市震荡行情中普遍性能下降的核心痛点，创新性地提出了融合日频与分钟 频数据的深度学习双通道模型(LSTM 和 GRU)。实证研究表明，该模型能有效捕捉不同时间尺度的市场信息，显著 提升策略在样本外(尤其是市场下行期)的预测准确性与稳定性，为重构债市量化择时体系提供了具有强泛化能力 的新思路。 在模型的使用上，主要使用了 RNN、LSTM、GRU 三种适用于时序预测的深度学习模型，重点阐述了 LSTM/GRU 通过门控机制解决长程依赖问题的优势。在传统深度学习模型设计的基础上，突破单一频率输入的局限，设计使 用了双通道模型架构。研究发现，深度学习方法在国债期货日频择时上可能会存在较大的过拟合风险，但是在加 入分钟频信息后 ...

编译丨王聪 编辑丨王多鱼 排版丨水成文 神经精神疾病 具有复杂的病理机制、显著的临床异质性以及漫长的临床前期，这给早期诊断和精准干预策略的制定带来了挑战。 随着大规模多模态神经影像数据集的发展以及人工智能 （AI） 算法的进步，将多模态成像与 AI 技术相结合已成为早期发现神经精神疾病以及为其量身定制个性 化治疗方案的关键途径。 2025 年 5 月 20 日，北京师范大学 舒妮 教授、南京航空航天大学 黄伟杰 副研究员在 Cell 子刊 Cell Reports Medicine 上发表了题为 ： AI-powered integration of multimodal imaging in precision medicine for neuropsychiatric disorders 的综述论文。 在这篇综述中， 作者概述了多模态神经影像技术、人工智能方法以及多模态数据融合策略，强调了基于神经影像数据的多模态人工智能在神经精神疾病精准医疗 中的应用，并探讨了其在临床应用中的挑战、新兴解决方案以及未来的发展方向。 神经精神疾病，例如阿尔茨海默病、自闭症、抑郁症等，如同一幅复杂的拼图：症状多样、病因隐蔽， ...