Quantitative Models and Construction Methods Model Name: GRU(10,2)+NN(10) - **Model Construction Idea**: This model combines Gated Recurrent Units (GRU) with a neural network (NN) to capture temporal dependencies in high-frequency data[4] - **Model Construction Process**: The model uses a GRU with 10 units and 2 layers, followed by a neural network with 10 units. The GRU processes sequential data, and the NN captures non-linear relationships[4] - **Model Evaluation**: The model shows strong performance in capturing temporal patterns and generating significant returns[4] Model Name: GRU(50,2)+NN(10) - **Model Construction Idea**: Similar to the GRU(10,2)+NN(10) model but with more GRU units to capture more complex temporal dependencies[4] - **Model Construction Process**: The model uses a GRU with 50 units and 2 layers, followed by a neural network with 10 units. This setup allows for deeper temporal feature extraction[4] - **Model Evaluation**: The model is effective in capturing complex temporal patterns and generating significant returns[4] Model Name: Multi-Granularity Model (5-day label) - **Model Construction Idea**: This model uses multiple granularities of data to improve prediction accuracy[4] - **Model Construction Process**: The model labels data with a 5-day horizon and uses a combination of features from different time scales to enhance prediction[4] - **Model Evaluation**: The model shows strong performance in capturing multi-scale patterns and generating significant returns[4] Model Name: Multi-Granularity Model (10-day label) - **Model Construction Idea**: Similar to the 5-day label model but with a 10-day horizon to capture longer-term dependencies[4] - **Model Construction Process**: The model labels data with a 10-day horizon and combines features from different time scales to enhance prediction[4] - **Model Evaluation**: The model is effective in capturing longer-term patterns and generating significant returns[4] Model Backtesting Results - **GRU(10,2)+NN(10)**: - Multi-Period Return: -1.32% (last week), -0.71% (November), 44.83% (2025)[4] - Excess Return: -0.77% (last week), -1.01% (November), 7.21% (2025)[4] - **GRU(50,2)+NN(10)**: - Multi-Period Return: -1.5% (last week), -1.23% (November), 44.56% (2025)[4] - Excess Return: -0.83% (last week), -0.92% (November), 7.9% (2025)[4] - **Multi-Granularity Model (5-day label)**: - Multi-Period Return: 0.75% (last week), 2.56% (November), 63.15% (2025)[4] - Excess Return: 1.07% (last week), 2.36% (November), 24.44% (2025)[4] - **Multi-Granularity Model (10-day label)**: - Multi-Period Return: 0.91% (last week), 2.55% (November), 57.7% (2025)[4] - Excess Return: 0.98% (last week), 2.27% (November), 24.14% (2025)[4] Quantitative Factors and Construction Methods Factor Name: Intraday Skewness Factor - **Factor Construction Idea**: This factor captures the skewness of intraday returns to identify asymmetric return distributions[4] - **Factor Construction Process**: The factor is calculated using the skewness of intraday returns over a specified period[4] - **Factor Evaluation**: The factor is effective in identifying stocks with asymmetric return distributions[4] Factor Name: Downside Volatility Proportion Factor - **Factor Construction Idea**: This factor measures the proportion of downside volatility to capture risk characteristics[4] - **Factor Construction Process**: The factor is calculated as the proportion of downside volatility relative to total volatility over a specified period[4] - **Factor Evaluation**: The factor is effective in identifying stocks with higher downside risk[4] Factor Name: Post-Open Buy Intention Proportion Factor - **Factor Construction Idea**: This factor measures the proportion of buy intentions after market open to capture investor sentiment[4] - **Factor Construction Process**: The factor is calculated as the proportion of buy orders relative to total orders after market open[4] - **Factor Evaluation**: The factor is effective in capturing investor sentiment and predicting stock movements[4] Factor Name: Post-Open Buy Intensity Factor - **Factor Construction Idea**: This factor measures the intensity of buy intentions after market open to capture investor sentiment strength[4] - **Factor Construction Process**: The factor is calculated as the intensity of buy orders relative to total orders after market open[4] - **Factor Evaluation**: The factor is effective in capturing the strength of investor sentiment and predicting stock movements[4] Factor Backtesting Results - **Intraday Skewness Factor**: - Multi-Period Return: -0.26% (last week), 0.49% (November), 22.76% (2025)[4] - Excess Return: 0.42% (last week), 1.46% (November), 6.14% (2025)[4] - **Downside Volatility Proportion Factor**: - Multi-Period Return: 0.38% (last week), 1.35% (November), 20.32% (2025)[4] - Excess Return: 0.41% (last week), 1.08% (November), 3.54% (2025)[4] - **Post-Open Buy Intention Proportion Factor**: - Multi-Period Return: 0.28% (last week), -0.01% (November), 19.33% (2025)[4] - Excess Return: 0.47% (last week), 0.28% (November), 8.78% (2025)[4] - **Post-Open Buy Intensity Factor**: - Multi-Period Return: 0.27% (last week), 0.57% (November), 26.36% (2025)[4] - Excess Return: -0.22% (last week), -0.55% (November), 10.06% (2025)[4]

Quantitative Models and Construction - **Model Name**: DecompGRU **Model Construction Idea**: The model improves information interaction between time-series and cross-sectional data by introducing two simple de-mean modules on the GRU baseline model[18] **Model Construction Process**: 1. The DecompGRU model architecture is based on GRU as the baseline 2. Two de-mean modules are added to enhance the interaction between time-series and cross-sectional data 3. The model is trained using IC and weighted MSE loss functions[18] **Model Evaluation**: The model demonstrates improved interaction between time-series and cross-sectional data, enhancing prediction accuracy[18] Model Backtesting Results - **DecompGRU TOP200 Portfolio**: - Cumulative absolute return: 41.11% - Excess return relative to WIND All A equal-weight index: 13.98% - Maximum drawdown: 10.08% - Weekly win rate: 64.52% - Monthly win rate: 100% - October absolute return: 1.78%, excess return: -0.07%[11] - **ETF Rotation Portfolio**: - Cumulative absolute return: 19.06% - Excess return relative to benchmark: -2.00% - Maximum drawdown: 7.82% - Weekly win rate: 62.50% - Monthly win rate: 57.14% - October absolute return: -2.04%, excess return: -1.18%[14][15] Quantitative Factors and Construction - **Factor Name**: Sentiment Heat Factor **Factor Construction Idea**: The factor aggregates stock-level sentiment heat metrics (e.g., browsing, self-selection, and clicks) to represent broader market sentiment[19] **Factor Construction Process**: 1. Individual stock sentiment heat is calculated as the sum of browsing, self-selection, and click counts 2. The sentiment heat is normalized by dividing by the total market sentiment on the same day and multiplying by 10,000 3. Aggregated sentiment heat is used as a proxy for market sentiment at the index, industry, and concept levels[19] **Factor Evaluation**: The factor effectively captures market sentiment and its impact on pricing errors[19] Factor Backtesting Results - **Broad-based Index Sentiment Heat Rotation Strategy**: - Annualized return since 2017: 8.74% - Maximum drawdown: 23.5% - 2025 portfolio return: 38.5% - Benchmark return: 32.9%[28] - **Concept Sentiment Heat BOTTOM Portfolio**: - Annualized return: 15.71% - Maximum drawdown: 28.89% - 2025 portfolio return: 42.1%[41][44]

- The GAN_GRU factor is based on the GAN_GRU model, which utilizes a Generative Adversarial Network (GAN) for processing volume-price time series features and then uses a GRU model for time series feature encoding to derive the stock selection factor[4][13][14] - The GAN_GRU model includes two GRU layers (GRU(128, 128)) followed by an MLP (256, 64, 64), with the final output prediction return (pRet) used as the stock selection factor[22] - The GAN model consists of a generator and a discriminator. The generator aims to generate data that appears real, while the discriminator aims to distinguish between real and generated data. The generator's loss function is $L_{G} = -\mathbb{E}_{z\sim P_{z}(z)}[\log(D(G(z)))]$[23][24][25] - The discriminator's loss function is $L_{D} = -\mathbb{E}_{x\sim P_{data}(x)}[\log D(x)] - \mathbb{E}_{z\sim P_{z}(z)}[\log(1-D(G(z)))]$[27][28][29] - The GAN_GRU model's training process involves alternating training of the generator and discriminator until convergence[30] - The GAN_GRU factor's performance from January 2019 to September 2025 shows an IC mean of 0.1136, an annualized excess return of 22.58%, and a recent IC of 0.1053 as of September 28, 2025[41][42] - The GAN_GRU factor's IC mean for the past year is 0.0982, with the highest IC values in the coal, building materials, social services, non-bank finance, and food & beverage industries[42][44] - The top-performing long portfolios in September 2025, based on the GAN_GRU factor, include sectors like building materials, steel, social services, coal, and non-bank finance, with excess returns of 5.78%, 5.13%, 1.91%, 1.55%, and 1.21%, respectively[45] - Over the past year, the top-performing long portfolios based on the GAN_GRU factor include home appliances, building materials, food & beverage, utilities, and textiles & apparel, with average monthly excess returns of 5.04%, 4.96%, 3.92%, 3.53%, and 3.10%, respectively[46] - The top stocks in each industry based on the GAN_GRU factor as of September 28, 2025, include companies like Baolaite, Yutaiwei-U, Cangge Mining, Tuowei Information, Hengtong Co., Angang Co., and others[49][50]

Quantitative Models and Construction Methods 1. Model Name: Weekly Rebalancing AI-Enhanced CSI 500 Wide Constraint Portfolio - **Model Construction Idea**: This model aims to enhance the CSI 500 index performance by leveraging AI-based factors while applying wide constraints on portfolio construction [72][73] - **Model Construction Process**: - The model uses deep learning factors (e.g., multi-granularity model with 10-day labels) as the basis for stock selection [72] - Constraints include: - Stock weight: 1% - Industry weight: 1% - Market cap weight: 0.3 - Turnover rate constraint: 0.3 - The optimization objective is to maximize expected returns, represented by the formula: $$ max \sum \mu_{i}w_{i} $$ where \( w_{i} \) is the weight of stock \( i \) in the portfolio, and \( \mu_{i} \) is the expected excess return of stock \( i \) [73][74] - **Model Evaluation**: The model demonstrates moderate performance under wide constraints, with cumulative excess returns shown over time [75][77] 2. Model Name: Weekly Rebalancing AI-Enhanced CSI 500 Strict Constraint Portfolio - **Model Construction Idea**: Similar to the wide constraint model but applies stricter constraints to control risk and enhance robustness [72][73] - **Model Construction Process**: - Constraints include: - Stock weight: 1% - Industry weight: 1% - Market cap weight: 0.1 - Additional constraints: - Market cap squared: 0.1 - ROE: 0.3 - SUE: 0.3 - Volatility: 0.3 - Component stock constraint: 0.8 - Optimization objective remains the same as the wide constraint model [73][74] - **Model Evaluation**: The stricter constraints result in a more stable performance, with cumulative excess returns displayed over time [76][80] 3. Model Name: Weekly Rebalancing AI-Enhanced CSI 1000 Wide Constraint Portfolio - **Model Construction Idea**: This model targets the CSI 1000 index, applying wide constraints while leveraging AI-based factors for enhanced returns [72][73] - **Model Construction Process**: - Constraints are similar to the CSI 500 wide constraint model, with a focus on smaller-cap stocks [73] - **Model Evaluation**: The model shows significant cumulative excess returns, particularly in recent years [79][86] 4. Model Name: Weekly Rebalancing AI-Enhanced CSI 1000 Strict Constraint Portfolio - **Model Construction Idea**: Similar to the wide constraint model but applies stricter constraints to manage risk and improve consistency [72][73] - **Model Construction Process**: - Constraints are similar to the CSI 500 strict constraint model, tailored for the CSI 1000 index [73] - **Model Evaluation**: The model demonstrates strong performance under strict constraints, with cumulative excess returns highlighted [85][87] --- Model Backtesting Results 1. Weekly Rebalancing AI-Enhanced CSI 500 Wide Constraint Portfolio - **Weekly Excess Return**: -1.36% (last week), -3.85% (September), 0.94% (YTD 2025) [13][78] - **Weekly Win Rate**: 23/39 weeks [13] 2. Weekly Rebalancing AI-Enhanced CSI 500 Strict Constraint Portfolio - **Weekly Excess Return**: -1.35% (last week), -1.33% (September), 3.70% (YTD 2025) [13][81] - **Weekly Win Rate**: 24/39 weeks [13] 3. Weekly Rebalancing AI-Enhanced CSI 1000 Wide Constraint Portfolio - **Weekly Excess Return**: 0.40% (last week), 0.42% (September), 9.15% (YTD 2025) [13][83] - **Weekly Win Rate**: 26/39 weeks [13] 4. Weekly Rebalancing AI-Enhanced CSI 1000 Strict Constraint Portfolio - **Weekly Excess Return**: -0.19% (last week), 0.67% (September), 14.01% (YTD 2025) [13][90] - **Weekly Win Rate**: 25/39 weeks [13] --- Quantitative Factors and Construction Methods 1. Factor Name: Intraday Skewness Factor - **Factor Construction Idea**: Captures the skewness of intraday stock returns to identify potential outperformers [6][8] - **Factor Construction Process**: Referenced in the report "Stock Selection Factor Series Research (19)" [13] - **Factor Evaluation**: Demonstrates strong performance with IC values of 0.027 (historical) and 0.042 (2025) [9][10] 2. Factor Name: Downside Volatility Proportion Factor - **Factor Construction Idea**: Measures the proportion of downside volatility in realized volatility to assess risk-adjusted returns [6][8] - **Factor Construction Process**: Referenced in the report "Stock Selection Factor Series Research (25)" [18][20] - **Factor Evaluation**: Shows moderate performance with IC values of 0.025 (historical) and 0.036 (2025) [9][10] 3. Factor Name: Post-Open Buying Intensity Factor - **Factor Construction Idea**: Quantifies the intensity of buying activity after market open to identify short-term momentum [6][8] - **Factor Construction Process**: Referenced in the report "Stock Selection Factor Series Research (64)" [22][26] - **Factor Evaluation**: Displays stable performance with IC values of 0.035 (historical) and 0.030 (2025) [9][10] 4. Factor Name: Deep Learning Factor (Improved GRU(50,2)+NN(10)) - **Factor Construction Idea**: Utilizes a gated recurrent unit (GRU) and neural network (NN) architecture to predict stock returns [6][8] - **Factor Construction Process**: Combines GRU with NN to capture temporal dependencies in high-frequency data [61][62] - **Factor Evaluation**: Strong performance with IC values of 0.066 (historical) and 0.050 (2025) [12][61] --- Factor Backtesting Results 1. Intraday Skewness Factor - **IC**: 0.027 (historical), 0.042 (2025) [9][10] - **Multi-Long-Short Return**: 3.82% (September), 16.22% (YTD 2025) [9][10] 2. Downside Volatility Proportion Factor - **IC**: 0.025 (historical), 0.036 (2025) [9][10] - **Multi-Long-Short Return**: 2.86% (September), 13.58% (YTD 2025) [9][10] 3. Post-Open Buying Intensity Factor - **IC**: 0.035 (historical), 0.030 (2025) [9][10] - **Multi-Long-Short Return**: 0.65% (September), 11.29% (YTD 2025) [9][10] 4. Deep Learning Factor (Improved GRU(50,2)+NN(10)) - **IC**: 0.066 (historical), 0.050 (2025) [12][61] - **Multi-Long-Short Return**: 2.13% (September), 7.40% (YTD 2025) [12][61]

Quantitative Factors and Models Summary Quantitative Factors and Construction Methods 1. **Factor Name**: DL_EM_Dynamic - **Construction Idea**: Extract intrinsic stock attributes from public fund holdings using matrix decomposition, and combine these attributes with LSTM-generated factor representations to create dynamic market state factors[19][21]. - **Construction Process**: - Matrix decomposition is applied to fund-stock investment networks to derive intrinsic matrices for funds and stocks[19]. - Static intrinsic attributes are updated semi-annually using fund reports and transformed into dynamic attributes by calculating their similarity to current market preferences[19]. - These dynamic attributes are combined with LSTM outputs and fed into an MLP model to enhance performance[19]. - The factor is used to construct a CSI 1000 enhanced index portfolio with constraints on tracking error (5%), industry exposure (±0.02), style exposure (±0.5), and individual stock weight (3%). Weekly rebalancing is applied, and transaction costs are set at 0.2% for both sides[21]. 2. **Factor Name**: Meta_RiskControl - **Construction Idea**: Incorporate factor exposure control into deep learning models to mitigate drawdowns during rapid style factor changes[26]. - **Construction Process**: - Multiply model outputs by corresponding stock factor exposures and include this in the loss function[26]. - Add penalties for style deviations and style momentum to the IC-based loss function[26]. - Use an ALSTM model with style inputs as the base model and integrate it with a meta-incremental learning framework for dynamic market adaptation[26]. - Construct enhanced portfolios for CSI 300, CSI 500, and CSI 1000 indices with constraints on market cap deviation (±0.5), industry deviation (±0.02), and individual stock weight (5x benchmark weight). Weekly rebalancing and 0.2% transaction costs are applied[29]. 3. **Factor Name**: Meta_Master - **Construction Idea**: Leverage market-guided stock transformer models (MASTER) and deep risk models to capture market states and improve factor performance[36]. - **Construction Process**: - Incorporate market state vectors derived from recent price-volume data of CSI 300, CSI 500, and CSI 1000 indices into the MASTER model[36]. - Construct 120 new features representing market states based on the styles of recently best-performing stocks[36]. - Replace the loss function with weighted MSE to enhance long-side prediction accuracy and use online meta-incremental learning for periodic model updates[36]. - Construct enhanced portfolios for CSI 300, CSI 500, and CSI 1000 indices with constraints on market cap deviation (±0.5), industry deviation (±0.02), and individual stock weight (5x benchmark weight). Weekly rebalancing and 0.2% transaction costs are applied[38]. 4. **Factor Name**: Deep Learning Convertible Bond Factor - **Construction Idea**: Address the diminishing excess returns of traditional convertible bond strategies by using GRU deep neural networks to model the complex nonlinear pricing logic of convertible bonds[50]. - **Construction Process**: - Introduce convertible bond-specific time-series factors into the GRU model[50]. - Combine cross-sectional bond attributes with GRU outputs to predict future returns, significantly improving model performance[50]. --- Factor Backtesting Results 1. **DL_EM_Dynamic Factor** - **RankIC**: 11.3% (July 2025, CSI 1000)[7][10] - **Excess Return**: 1.3% (July 2025, CSI 1000); 11% YTD (2025)[7][10] - **Annualized Return**: 29.7% (since 2019)[23] - **Annualized Excess Return**: 23.4% (since 2019)[23] - **IR**: 2.03 (since 2019)[23] - **Max Drawdown**: -10.1% (since 2019)[23] 2. **Meta_RiskControl Factor** - **RankIC**: 15.5% (July 2025, All A-shares)[7][13] - **Excess Return**: - CSI 300: 1.9% (July 2025); 6.4% YTD (2025)[31] - CSI 500: 1.4% (July 2025); 4.4% YTD (2025)[33] - CSI 1000: 1.3% (July 2025); 9.3% YTD (2025)[35] - **Annualized Return**: - CSI 300: 20.1% (since 2019)[31] - CSI 500: 26.1% (since 2019)[33] - CSI 1000: 34.1% (since 2019)[35] - **Annualized Excess Return**: - CSI 300: 15.0% (since 2019)[31] - CSI 500: 19.2% (since 2019)[33] - CSI 1000: 27.0% (since 2019)[35] - **IR**: - CSI 300: 1.58 (since 2019)[31] - CSI 500: 1.97 (since 2019)[33] - CSI 1000: 2.36 (since 2019)[35] - **Max Drawdown**: - CSI 300: -5.8% (since 2019)[31] - CSI 500: -9.3% (since 2019)[33] - CSI 1000: -10.2% (since 2019)[35] 3. **Meta_Master Factor** - **RankIC**: 18.9% (July 2025, All A-shares)[7][16] - **Excess Return**: - CSI 300: 2.0% (July 2025); 7.9% YTD (2025)[39] - CSI 500: 1.6% (July 2025); 5.5% YTD (2025)[45] - CSI 1000: 1.4% (July 2025); 8.1% YTD (2025)[47] - **Annualized Return**: - CSI 300: 22.0% (since 2019)[39] - CSI 500: 23.8% (since 2019)[45] - CSI 1000: 30.7% (since 2019)[47] - **Annualized Excess Return**: - CSI 300: 17.5% (since 2019)[39] - CSI 500: 18.2% (since 2019)[45] - CSI 1000: 25.2% (since 2019)[47] - **IR**: - CSI 300: 2.09 (since 2019)[39] - CSI 500: 1.9 (since 2019)[45] - CSI 1000: 2.33 (since 2019)[47] - **Max Drawdown**: - CSI 300: -7.2% (since 2019)[39] - CSI 500: -5.8% (since 2019)[45] - CSI 1000: -8.8% (since 2019)[47] 4. **Deep Learning Convertible Bond Factor** - **Absolute Return**: - July 2025: 5.8% (equity-biased), 3.8% (balanced), 3.3% (debt-biased)[52] - Annualized (since 2021): 13.2% (equity-biased), 11.8% (balanced), 12.7% (debt-biased)[52] - **Excess Return**: - July 2025: 1.5% (equity-biased), -0.4% (balanced), -0.9% (debt-biased)[55] - Annualized (since 2021): 5.8% (equity-biased), 4.0% (balanced), 4.4% (debt-biased)[55]

Quantitative Models and Construction Methods - **Model Name**: DecompGRU **Model Construction Idea**: The model improves the GRU baseline by introducing two simple de-mean modules to enhance the interaction between temporal and cross-sectional information[14] **Model Construction Process**: 1. The DecompGRU model architecture is based on GRU with added de-mean modules for trend decomposition[14] 2. Two versions of the model are trained using different loss functions: IC and weighted MSE[14] 3. The IC-based model and MSE-based model are used to score stocks, and the top 200 stocks are selected for portfolio construction[8][14] **Evaluation**: The model effectively captures temporal and cross-sectional interactions, leading to improved stock selection performance[14] Model Backtesting Results - **DecompGRU Model**: - Cumulative absolute return: 24.54% - Excess return relative to WIND All A equal-weight index: 9.80% - Maximum drawdown: 10.08% - Weekly win rate: 72.22% - Monthly win rate: 100%[10] - **ETF Rotation Portfolio (Based on DecompGRU Scores)**: - Cumulative absolute return: 12.97% - Excess return relative to WIND ETF index: 8.65% - Maximum drawdown: 6.16% - Weekly win rate: 68.42% - Monthly win rate: 75%[12] Quantitative Factors and Construction Methods - **Factor Name**: Total Heat Indicator **Factor Construction Idea**: The indicator aggregates stock-level attention metrics (views, favorites, clicks) to represent market sentiment at broader levels (indices, industries, concepts)[17][18] **Factor Construction Process**: 1. Calculate the sum of views, favorites, and clicks for each stock[18] 2. Normalize the sum as a percentage of the total market activity on the same day[18] 3. Multiply the normalized value by 10,000 to derive the final indicator, with a range of [0, 10,000][18] **Evaluation**: The factor serves as a proxy for sentiment-driven mispricing, particularly effective at the stock level[18] Factor Backtesting Results - **Broad Index Heat Rotation Strategy**: - Annualized return since 2017: 8.74% - Maximum drawdown: 23.5% - 2025 YTD return: 18.8% - Benchmark return: 17.1%[27] - **Concept Heat BOTTOM Portfolio**: - Annualized return: 15.71% - Maximum drawdown: 28.89% - 2025 YTD return: 27%[44] Additional Observations - **Broad Index Heat Changes**: - Largest increase: CSI 500 (+10.21%) - Largest decrease: CSI 2000 (-6.02%)[27][29] - **Industry Heat Changes**: - Top 5 positive changes: Building Materials (+83.5%), Building Decoration, Social Services, Steel, Food & Beverage - Top 5 negative changes: Light Manufacturing (-32.5%), Textile & Apparel, Automotive, Real Estate, Utilities[38] - **Concept Heat Changes**: - Top 5 concepts: Dairy (+233.5%), Football (+194.9%), NMN (+115), Short Drama Games (+113.6%), Rent-Sale Rights (+109.6)[39][47][48]

Quantitative Models and Construction Methods 1. Model Name: GRU(50,2)+NN(10) Factor - **Model Construction Idea**: This factor leverages a deep learning architecture combining Gated Recurrent Units (GRU) and Neural Networks (NN) to capture high-frequency trading patterns and predict stock returns[4][55] - **Model Construction Process**: - The GRU(50,2) component processes sequential high-frequency data with 50 units and 2 layers - The NN(10) component is a fully connected neural network with 10 neurons in the output layer - The model is trained on historical high-frequency data to predict stock returns, optimizing for multi-class classification or regression tasks[4][55] - **Model Evaluation**: Demonstrates robust performance in capturing high-frequency trading signals and generating stable returns[4][55] 2. Model Name: Multi-Granularity Model (5-Day Label) - **Model Construction Idea**: This model uses multi-granularity data to predict stock returns over a 5-day horizon, leveraging bidirectional AGRU (Attention-based GRU) for feature extraction[57][60] - **Model Construction Process**: - Input data is segmented into multiple granularities (e.g., daily, intraday) - Bidirectional AGRU is applied to extract temporal features from the data - A 5-day label is used as the prediction target, and the model is trained to optimize for this horizon[57][60] - **Model Evaluation**: Effective in capturing medium-term trading patterns and generating consistent returns[57][60] 3. Model Name: Multi-Granularity Model (10-Day Label) - **Model Construction Idea**: Similar to the 5-day label model, this version extends the prediction horizon to 10 days, using bidirectional AGRU for feature extraction[60][65] - **Model Construction Process**: - Multi-granularity data is processed with bidirectional AGRU - A 10-day label is used as the prediction target, and the model is trained to optimize for this extended horizon[60][65] - **Model Evaluation**: Provides a longer-term perspective on trading patterns, with slightly lower returns compared to the 5-day model but still effective[60][65] --- Model Backtesting Results GRU(50,2)+NN(10) Factor - **IC**: Historical: 0.066, 2025: 0.039[4][55] - **e^(-RankMAE)**: Historical: 0.336, 2025: 0.334[4][55] - **Long-Short Return**: Weekly: 0.70%, June: 3.58%, 2025 YTD: 19.78%[4][55] - **Long-Only Excess Return**: Weekly: -0.30%, June: 0.92%, 2025 YTD: -1.06%[4][55] Multi-Granularity Model (5-Day Label) - **IC**: Historical: 0.081, 2025: 0.070[57][60] - **e^(-RankMAE)**: Historical: 0.344, 2025: 0.343[57][60] - **Long-Short Return**: Weekly: 1.56%, June: 5.97%, 2025 YTD: 35.45%[57][60] - **Long-Only Excess Return**: Weekly: 0.40%, June: 2.16%, 2025 YTD: 11.87%[57][60] Multi-Granularity Model (10-Day Label) - **IC**: Historical: 0.074, 2025: 0.065[60][65] - **e^(-RankMAE)**: Historical: 0.342, 2025: 0.343[60][65] - **Long-Short Return**: Weekly: 1.66%, June: 5.76%, 2025 YTD: 33.44%[60][65] - **Long-Only Excess Return**: Weekly: 0.71%, June: 2.06%, 2025 YTD: 11.11%[60][65] --- Quantitative Factors and Construction Methods 1. Factor Name: Intraday Skewness Factor - **Factor Construction Idea**: Measures the skewness of intraday returns to capture asymmetry in price movements[4][10] - **Factor Construction Process**: - Calculate intraday returns for each stock - Compute the skewness of these returns using the formula: $ Skewness = \frac{E[(X - \mu)^3]}{\sigma^3} $ where $X$ is the return, $\mu$ is the mean, and $\sigma$ is the standard deviation[4][10] - **Factor Evaluation**: Effective in identifying stocks with asymmetric return distributions, though performance varies across periods[4][10] 2. Factor Name: Downside Volatility Ratio - **Factor Construction Idea**: Focuses on the proportion of downside volatility relative to total volatility to capture risk-averse behavior[4][14] - **Factor Construction Process**: - Calculate downside volatility as the standard deviation of negative returns - Compute the ratio of downside volatility to total volatility[4][14] - **Factor Evaluation**: Useful for identifying stocks with higher downside risk, though returns are sensitive to market conditions[4][14] 3. Factor Name: Opening Buy Intensity - **Factor Construction Idea**: Measures the intensity of buy orders during the opening period to capture early trading sentiment[4][17] - **Factor Construction Process**: - Aggregate buy orders in the first 30 minutes of trading - Normalize by total trading volume during the same period[4][17] - **Factor Evaluation**: Captures short-term sentiment effectively, though performance is volatile[4][17] --- Factor Backtesting Results Intraday Skewness Factor - **IC**: Historical: 0.027, 2025: 0.047[4][10] - **e^(-RankMAE)**: Historical: 0.324, 2025: 0.330[4][10] - **Long-Short Return**: Weekly: -0.51%, June: 1.48%, 2025 YTD: 14.73%[4][10] - **Long-Only Excess Return**: Weekly: -0.03%, June: 0.18%, 2025 YTD: 2.59%[4][10] Downside Volatility Ratio - **IC**: Historical: 0.025, 2025: 0.046[4][14] - **e^(-RankMAE)**: Historical: 0.324, 2025: 0.328[4][14] - **Long-Short Return**: Weekly: -0.04%, June: 1.86%, 2025 YTD: 12.84%[4][14] - **Long-Only Excess Return**: Weekly: 0.09%, June: 0.50%, 2025 YTD: 1.07%[4][14] Opening Buy Intensity - **IC**: Historical: 0.031, 2025: 0.028[4][17] - **e^(-RankMAE)**: Historical: 0.322, 2025: 0.322[4][17] - **Long-Short Return**: Weekly: 0.77%, June: 1.85%, 2025 YTD: 11.44%[4][17] - **Long-Only Excess Return**: Weekly: 0.04%, June: 0.61%, 2025 YTD: 5.91%[4][17]

Quantitative Models and Construction Methods Quantitative Factors and Their Construction 1. **Factor Name**: Intraday Skewness Factor **Construction Idea**: Captures the skewness of intraday stock returns to identify potential return asymmetry[3][6] **Construction Process**: Referenced in the report "Stock Selection Factor Series Research (19) - High-Frequency Factors on Stock Return Distribution Characteristics"[11] **Evaluation**: Demonstrates mixed performance with positive returns in some periods but underperformance in others[3][6] 2. **Factor Name**: Downside Volatility Proportion Factor **Construction Idea**: Measures the proportion of downside volatility in intraday price movements to assess risk[3][6] **Construction Process**: Referenced in the report "Stock Selection Factor Series Research (25) - High-Frequency Factors on Realized Volatility Decomposition"[16] **Evaluation**: Shows consistent positive returns in certain periods but limited robustness in others[3][6] 3. **Factor Name**: Post-Open Buy Intention Proportion Factor **Construction Idea**: Quantifies the proportion of buy orders after market open to gauge investor sentiment[3][6] **Construction Process**: Referenced in the report "Stock Selection Factor Series Research (64) - Low-Frequency Applications of High-Frequency Data Using Intuitive Logic and Machine Learning"[20] **Evaluation**: Exhibits moderate performance with occasional strong returns[3][6] 4. **Factor Name**: Post-Open Buy Intention Intensity Factor **Construction Idea**: Measures the intensity of buy orders after market open to reflect market momentum[3][6] **Construction Process**: Referenced in the report "Stock Selection Factor Series Research (64) - Low-Frequency Applications of High-Frequency Data Using Intuitive Logic and Machine Learning"[24] **Evaluation**: Performance is inconsistent, with periods of underperformance[3][6] 5. **Factor Name**: Post-Open Large Order Net Buy Proportion Factor **Construction Idea**: Tracks the proportion of large net buy orders after market open to identify institutional activity[3][6] **Construction Process**: Derived from high-frequency trading data[30] **Evaluation**: Generally positive performance with strong returns in specific periods[3][6] 6. **Factor Name**: Post-Open Large Order Net Buy Intensity Factor **Construction Idea**: Measures the intensity of large net buy orders after market open to capture market trends[3][6] **Construction Process**: Derived from high-frequency trading data[35] **Evaluation**: Mixed results with moderate returns in some periods[3][6] 7. **Factor Name**: Improved Reversal Factor **Construction Idea**: Enhances traditional reversal factors by incorporating high-frequency data[3][6] **Construction Process**: Derived from intraday price reversals[40] **Evaluation**: Limited performance improvement over traditional reversal factors[3][6] 8. **Factor Name**: Tail-End Trading Proportion Factor **Construction Idea**: Measures the proportion of trading activity near market close to capture end-of-day effects[3][6] **Construction Process**: Derived from high-frequency trading data[45] **Evaluation**: Underperformance in most periods[3][6] 9. **Factor Name**: Average Single Transaction Outflow Proportion Factor **Construction Idea**: Tracks the proportion of outflows in single transactions to assess liquidity[3][6] **Construction Process**: Derived from high-frequency trading data[50] **Evaluation**: Limited effectiveness in predicting returns[3][6] 10. **Factor Name**: Large Order Push-Up Factor **Construction Idea**: Measures the impact of large orders on price increases to identify market movers[3][6] **Construction Process**: Derived from high-frequency trading data[55] **Evaluation**: Moderate performance with occasional strong returns[3][6] 11. **Factor Name**: Deep Learning High-Frequency Factor (Improved GRU(50,2)+NN(10)) **Construction Idea**: Combines GRU and neural networks to capture complex patterns in high-frequency data[3][6] **Construction Process**: Utilizes GRU(50,2) and NN(10) architectures for feature extraction and prediction[59] **Evaluation**: Strong performance in certain periods but underperformance in others[3][6] 12. **Factor Name**: Deep Learning High-Frequency Factor (Residual Attention LSTM(48,2)+NN(10)) **Construction Idea**: Incorporates residual attention mechanisms with LSTM and neural networks for enhanced prediction[3][6] **Construction Process**: Utilizes LSTM(48,2) and NN(10) architectures with residual attention layers[61] **Evaluation**: Consistently strong performance across multiple periods[3][6] 13. **Factor Name**: Deep Learning Factor (Multi-Granularity Model - 5-Day Label) **Construction Idea**: Uses multi-granularity modeling with 5-day labels for short-term predictions[3][6] **Construction Process**: Trained using bidirectional AGRU[64] **Evaluation**: Strong performance with high returns in most periods[3][6] 14. **Factor Name**: Deep Learning Factor (Multi-Granularity Model - 10-Day Label) **Construction Idea**: Uses multi-granularity modeling with 10-day labels for medium-term predictions[3][6] **Construction Process**: Trained using bidirectional AGRU[65] **Evaluation**: Consistently strong performance across multiple periods[3][6] AI-Enhanced Portfolio Construction 1. **Portfolio Name**: CSI 500 AI Enhanced Wide Constraint Portfolio **Construction Idea**: Maximizes expected returns under wide constraints using deep learning factors[69][70] **Construction Process**: - Weekly rebalancing - Constraints on individual stocks, industries, market cap, and other factors - Objective function: $$ max\sum\mu_{i}w_{i} $$ where \( w_i \) is the weight of stock \( i \) and \( \mu_i \) is its expected excess return[71] **Evaluation**: Strong cumulative excess returns since 2017[72] 2. **Portfolio Name**: CSI 500 AI Enhanced Strict Constraint Portfolio **Construction Idea**: Similar to the wide constraint portfolio but with stricter constraints[69][70] **Construction Process**: Same as above with stricter constraints on market cap, ROE, SUE, and volatility[71] **Evaluation**: Moderate cumulative excess returns since 2017[73] 3. **Portfolio Name**: CSI 1000 AI Enhanced Wide Constraint Portfolio **Construction Idea**: Maximizes expected returns under wide constraints using deep learning factors for smaller-cap stocks[69][70] **Construction Process**: Same as CSI 500 portfolios but applied to CSI 1000 index[71] **Evaluation**: Strong cumulative excess returns since 2017[76] 4. **Portfolio Name**: CSI 1000 AI Enhanced Strict Constraint Portfolio **Construction Idea**: Similar to the wide constraint portfolio but with stricter constraints for smaller-cap stocks[69][70] **Construction Process**: Same as above with stricter constraints on market cap, ROE, SUE, and volatility[71] **Evaluation**: Strong cumulative excess returns since 2017[79] Backtest Results for Factors 1. **Intraday Skewness Factor**: IC (2025): 0.057, Multi-Period Returns: 14.35% (2025)[3][6] 2. **Downside Volatility Proportion Factor**: IC (2025): 0.055, Multi-Period Returns: 11.77% (2025)[3][6] 3. **Post-Open Buy Intention Proportion Factor**: IC (2025): 0.033, Multi-Period Returns: 10.32% (2025)[3][6] 4. **Post-Open Buy Intention Intensity Factor**: IC (2025): 0.026, Multi-Period Returns: 11.19% (2025)[3][6] 5. **Post-Open Large Order Net Buy Proportion Factor**: IC (2025): 0.039, Multi-Period Returns: 12.32% (2025)[3][6] 6. **Post-Open Large Order Net Buy Intensity Factor**: IC (2025): 0.028, Multi-Period Returns: 6.78% (2025)[3][6] 7. **Improved Reversal Factor**: IC (2025): 0.003, Multi-Period Returns: 9.34% (2025)[3][6] 8. **Tail-End Trading Proportion Factor**: IC (2025): 0.022, Multi-Period Returns: 5.43% (2025)[3][6] 9. **Average Single Transaction Outflow Proportion Factor**: IC (2025): 0.012, Multi-Period Returns: 0.82% (2025)[3][6] 10. **Large Order Push-Up Factor