Quantitative Models and Factor Analysis Summary Quantitative Models and Construction - **Model Name**: Multi-modal Multi-scale Stock Price Prediction Model **Model Construction Idea**: The model integrates multi-modal (chart data and time-series data) and multi-scale (different frequency data) features to enhance stock price prediction accuracy. It employs four independent deep time-series models and convolutional models for feature extraction, using both regression and classification losses for end-to-end training[14][17][18]. **Model Construction Process**: 1. **Multi-modal Features**: Combines time-series price-volume data and standardized price-volume charts. Time-series models capture abstract numerical relationships, while convolutional models identify chart patterns[17]. 2. **Multi-scale Features**: Incorporates 1-minute high-frequency data, daily data, and weekly data. High-frequency data is factorized into 55 features, which are then input into time-series models[18]. 3. **Lightweight Design**: Reduces the parameter size of each sub-model to 1/4 of the initial version, minimizing overfitting and computational resource dependency[18]. 4. **Multi-head Output**: Outputs include absolute future returns and categorical predictions (up, flat, down), using mean squared error and cross-entropy as loss functions[19]. **Model Evaluation**: The model demonstrates significant improvements in prediction accuracy and excess returns compared to the initial version[14][17][19]. Model Backtesting Results - **RankIC Mean**: - All Market: 8.7% - CSI 300: 7.9% - CSI 500: 6.6% - CSI 800: 6.9% - CSI 1000: 8.2% - CNI 2000: 8.7% - ChiNext: 10.4%[21][116] - **RankIC Win Rate**: - All Market: 86.7% - CSI 300: 69.0% - CSI 500: 73.5% - CSI 800: 75.2% - CSI 1000: 84.8% - CNI 2000: 86.1% - ChiNext: 89.2%[21][116] - **Excess Annualized Returns**: - All Market: 12.97% - CSI 300: 9.17% - CSI 500: 5.30% - CSI 800: 8.38% - CSI 1000: 7.47% - CNI 2000: 7.47% - ChiNext: 11.52%[21][117] Quantitative Factors and Construction - **Factor Name**: Model-derived Factor **Factor Construction Idea**: Derived from the model's predictions, the factor captures both numerical relationships and chart patterns, leveraging multi-modal and multi-scale data[14][17][18]. **Factor Construction Process**: 1. Predictions from time-series models and convolutional models are combined. 2. Multi-frequency data (1-minute, daily, weekly) is processed to extract features. 3. Factor values are generated based on the model's outputs, including both regression and classification results[14][17][18]. **Factor Evaluation**: The factor shows low correlation with traditional Barra style factors, indicating its uniqueness[22][23]. Factor Backtesting Results - **Correlation with Barra Factors**: - Liquidity: -18% - Volatility: -16% - Size: -8%[22][23] - **RankIC Mean**: - All Market: 8.7% - CSI 300: 7.9% - CSI 500: 6.6% - CSI 800: 6.9% - CSI 1000: 8.2% - CNI 2000: 8.7% - ChiNext: 10.4%[21][116] - **RankIC Win Rate**: - All Market: 86.7% - CSI 300: 69.0% - CSI 500: 73.5% - CSI 800: 75.2% - CSI 1000: 84.8% - CNI 2000: 86.1% - ChiNext: 89.2%[21][116] - **Excess Annualized Returns**: - All Market: 12.97% - CSI 300: 9.17% - CSI 500: 5.30% - CSI 800: 8.38% - CSI 1000: 7.47% - CNI 2000: 7.47% - ChiNext: 11.52%[21][117]

广发证券首席金工分析师 安宁宁 anningning@gf.com.cn 广发证券资深金工分析师 陈原文 chenyuanwen@gf.com.cn 联系人：广发证券金工研究员 林涛 gflintao@gf.com.cn 广发金工安宁宁陈原文团队 摘要 神经常微分方程： 在机器学习国际顶会NeurIPS 2018上，Chen等人发表的论文《Neural Ordinary Differential Equations》获得了大会的最佳论文奖。简单来 说，一个常见的ResNet网络通常由多个形如h_{t+1}=f(h_t,_t)+h_t的残差结构所组成。在常规求解中，需计算出每一个残差结构中最能拟合训练数据的网 络参数。而该论文提出，假设当ResNet网络中的残差结构无限堆叠时，则每一个残差结构的参数都可以通过求解同一个常微分方程来获得。 液态神经网络： 基于上述工作，来自麻省理工学院的Ramin Hasani等人，创新性地以常微分方程的形式描述循环神经网络的隐藏状态变化，提出了一类被 称之为液态神经网络的模型，这些研究成果被发表在《Nature：Machine Intelligence》等国际顶级期刊上。此类模 ...

Core Insights - NVIDIA reported record revenue of $39.3 billion for Q4 FY2025, a 12% increase quarter-over-quarter and a 78% increase year-over-year, exceeding the expected $37.5 billion [1] - For the full fiscal year 2025, NVIDIA's revenue reached $130.5 billion, representing a 114% year-over-year growth [1] Financial Performance - Q4 FY2025 total revenue was $39,331 million, with a gross profit of $27,924 million, resulting in a gross margin of 73.4% [2] - Operating income for Q4 was $22,961 million, with an operating profit margin of 62.3% [2] - Net income for Q4 was $19,309 million, yielding a net profit margin of 55.6% [2] Data Center Segment - Data center revenue reached a record $35.6 billion in Q4, a 16% increase quarter-over-quarter and a 93% increase year-over-year, with FY2025 revenue at $115.2 billion [3] - The Blackwell product line saw Q4 sales exceed expectations at $11 billion, marking the fastest product ramp in company history [3] - Demand for AI infrastructure is driving significant growth, with large clusters starting at 100,000 GPUs [3] Gaming Segment - Q4 gaming revenue was $2.5 billion, a 22% decrease quarter-over-quarter and an 11% decrease year-over-year, attributed to supply constraints [3] - Full-year gaming revenue was $11.4 billion, reflecting a 9% year-over-year increase [4] Professional Visualization Segment - Q4 revenue was $511 million, a 5% increase quarter-over-quarter and a 10% increase year-over-year, with full-year revenue at $1.9 billion, up 21% [5] Automotive Segment - Q4 automotive revenue reached a record $570 million, a 27% increase quarter-over-quarter and a 103% increase year-over-year, with full-year revenue at $1.7 billion, up 5% [5] - Continued growth in autonomous vehicles is driving revenue, with partnerships announced for next-generation vehicles [5] Networking Segment - Q4 networking revenue saw a 3% decline, but a transition to larger NVLink and Spectrum X is expected to restore growth in the upcoming quarter [5] Gross Margin - Q4 GAAP gross margin was 73%, with non-GAAP gross margin at 73.5% [6]

Quantitative Models and Construction Methods 1. Model Name: Recurrent Neural Network (RNN) for Trend Strategy Enhancement - **Model Construction Idea**: Use RNN to predict whether a trend-following strategy will be profitable on a given trading day based on early morning market data. If the model predicts profitability, execute the trend-following strategy; otherwise, refrain from trading[3][22][61] - **Model Construction Process**: - Input: Early morning market data (e.g., opening price, closing price, high, low, volume, etc.) - Output: Binary classification (1 for profitable trend trading, 0 for non-profitable trend trading) - Model Structure: 13 input features → 200 LSTM units → 1 output node with Sigmoid activation function - Loss Function: Cross-entropy loss - Optimization: Gradient descent with backpropagation through time (BPTT) to minimize the loss function[46][61][66] - Training Data: Historical intraday data from 2010 to 2013, labeled based on profitability of trend strategies using metrics like Hurst index or K-line body ratio (R)[61][63] - Trading Rules: - Predict profitability at 33 minutes after market open - If predicted probability (p) > 120-day moving average of p (MA120(p)), execute trend-following trades; otherwise, no trades - Positions are closed at the end of the day or upon hitting stop-loss levels[63][66] - **Model Evaluation**: The RNN effectively filters out low-profitability signals, improving overall strategy performance. It demonstrates robustness to transaction costs and parameter stability[73][74][85] 2. Factor Name: Component Stock Consistency Indicator (R) - **Factor Construction Idea**: Measure the consistency of component stock movements to estimate market trend strength. Higher consistency indicates a stronger trend, suitable for trend-following strategies[14][21] - **Factor Construction Process**: - Formula: $$R = \frac{\lambda_1}{\sum_{i=1}^{300} \lambda_i} \times 100\%$$ where \( \lambda_1, \lambda_2, ..., \lambda_{300} \) are eigenvalues of the covariance matrix of component stock returns - Interpretation: \( R \) represents the variance contribution of the first principal component. Higher \( R \) values indicate stronger consistency among component stocks[14] - **Factor Evaluation**: The indicator effectively identifies market conditions suitable for trend-following strategies, as demonstrated by its ability to differentiate between high-consistency and low-consistency markets[14][21] --- Model Backtesting Results 1. RNN Model - **Annualized Return**: 18.47% (out-of-sample)[72] - **Cumulative Return**: 80.72% (out-of-sample)[72] - **Maximum Drawdown**: -8.63% (out-of-sample)[72] - **Win Rate**: 39.52% (out-of-sample)[72] - **Profit-Loss Ratio**: 2.27 (out-of-sample)[72] - **Average Return per Trade**: 0.17% (out-of-sample)[72] 2. Component Stock Consistency Indicator (R) - **High Consistency Example**: \( R = 86.4\), suitable for trend trading[16] - **Low Consistency Example**: \( R = 45.2\), unsuitable for trend trading[19] --- Factor Backtesting Results 1. Component Stock Consistency Indicator (R) - **High \( R \) Market**: Demonstrates strong trend-following profitability[16] - **Low \( R \) Market**: Demonstrates poor trend-following profitability[19]