Quantitative Models and Construction Methods 1. Model Name: Multi-Granularity Model (5-Day Label) - **Model Construction Idea**: This model leverages deep learning techniques to capture multi-granularity features of stock data over a 5-day horizon[66] - **Model Construction Process**: The factor is trained using a bidirectional AGRU (Attention Gated Recurrent Unit) model, which processes sequential data to extract temporal dependencies and patterns[66] - **Model Evaluation**: The model demonstrates stable performance across different time periods, indicating its robustness in capturing market dynamics[66] 2. Model Name: Multi-Granularity Model (10-Day Label) - **Model Construction Idea**: Similar to the 5-day label model, this model extends the horizon to 10 days to capture longer-term patterns in stock data[70] - **Model Construction Process**: The factor is also trained using a bidirectional AGRU model, with adjustments to accommodate the extended time horizon[70] - **Model Evaluation**: The model shows consistent performance, with slightly different characteristics compared to the 5-day label model, making it suitable for longer-term strategies[70] 3. Model Name: AI-Enhanced Index Strategies - **Model Construction Idea**: Combines multiple deep learning factors (e.g., 5-day and 10-day multi-granularity models) to construct AI-enhanced index strategies with risk constraints[72] - **Model Construction Process**: - The combined factor is a weighted sum: `0.5 * Multi-Granularity Model (5-Day Label) + 0.5 * Multi-Granularity Model (10-Day Label)`[72] - Optimization objective: Maximize expected returns, represented by the function: $$ max \sum \mu_{i} w_{i} $$ where \( w_{i} \) is the weight of stock \( i \), and \( \mu_{i} \) is the expected excess return of stock \( i \)[75] - Risk control constraints include limits on individual stock weights, industry weights, market capitalization, and turnover rates[73][75] - Backtesting assumes next-day average price execution and deducts a 0.3% transaction cost[76] - **Model Evaluation**: The model effectively balances return maximization and risk control, with different configurations (e.g., wide vs. strict constraints) tailored to specific index benchmarks[72][73] --- Model Backtesting Results 1. Multi-Granularity Model (5-Day Label) - **IC**: Historical: 0.079; 2026: 0.040[14] - **e^(-RankMAE)**: Historical: 0.343; 2026: 0.334[14] - **Long-Short Return**: March: 1.68%; 2026 YTD: 9.31%[14] - **Long-Only Excess Return**: March: 1.21%; 2026 YTD: 4.95%[14] - **Monthly Win Rate**: 9/10[14] 2. Multi-Granularity Model (10-Day Label) - **IC**: Historical: 0.072; 2026: 0.040[14] - **e^(-RankMAE)**: Historical: 0.342; 2026: 0.336[14] - **Long-Short Return**: March: 2.35%; 2026 YTD: 8.19%[14] - **Long-Only Excess Return**: March: 1.48%; 2026 YTD: 4.72%[14] - **Monthly Win Rate**: 8/10[14] 3. AI-Enhanced Index Strategies - **AI Air Quality Index Strategy**: - **Weekly Rebalancing**: Excess Return: -0.12% (last week), 0.65% (March), 4.17% (2026 YTD); Absolute Return: -5.47% (last week), -7.86% (March), 6.70% (2026 YTD)[15][81] - **Daily Rebalancing**: Excess Return: -0.78% (last week), -0.08% (March), 4.41% (2026 YTD); Absolute Return: -6.12% (last week), -8.59% (March), 6.94% (2026 YTD)[15][81] - **CSI 500 AI Enhanced (Wide Constraint)**: - **Weekly Rebalancing**: Excess Return: 1.43% (last week), 5.62% (March), 2.71% (2026 YTD); Absolute Return: -4.40% (last week), -4.76% (March), 6.66% (2026 YTD)[15][83] - **Daily Rebalancing**: Excess Return: 0.60% (last week), 1.79% (March), -2.71% (2026 YTD); Absolute Return: -5.23% (last week), -8.58% (March), 1.24% (2026 YTD)[15][83] - **CSI 500 AI Enhanced (Strict Constraint)**: - **Weekly Rebalancing**: Excess Return: 0.35% (last week), 3.51% (March), 2.73% (2026 YTD); Absolute Return: -5.47% (last week), -6.87% (March), 6.68% (2026 YTD)[15][89] - **Daily Rebalancing**: Excess Return: 0.31% (last week), 2.10% (March), 1.42% (2026 YTD); Absolute Return: -5.52% (last week), -8.27% (March), 5.37% (2026 YTD)[15][89] - **CSI 1000 AI Enhanced (Wide Constraint)**: - **Weekly Rebalancing**: Excess Return: 0.79% (last week), 3.52% (March), 4.19% (2026 YTD); Absolute Return: -4.46% (last week), -5.56% (March), 6.67% (2026 YTD)[15][91] - **Daily Rebalancing**: Excess Return: -0.20% (last week), 1.81% (March), 1.92% (2026 YTD); Absolute Return: -5.44% (last week), -7.27% (March), 4.40% (2026 YTD)[15][91] - **CSI 1000 AI Enhanced (Strict Constraint)**: - **Weekly Rebalancing**: Excess Return: 0.57% (last week), 2.55% (March), 3.67% (2026 YTD); Absolute Return: -4.68% (last week), -6.53% (March), 6.15% (2026 YTD)[15][97] - **Daily Rebalancing**: Excess Return: 0.75% (last week), 1.87% (March), 3.72% (2026 YTD); Absolute Return: -4.49% (last week), -7.21% (March), 6.20% (2026 YTD)[15][97]

- The report highlights the performance of high-frequency factors, deep learning factors, and AI-enhanced portfolios, summarizing their historical and 2026 metrics such as IC, RankMAE, multi-long-short returns, and monthly win rates[9][10][11] - High-frequency factors include intraday skewness, downside volatility proportion, post-open buying intention proportion, post-open buying intention intensity, post-open large-order net buying proportion, post-open large-order net buying intensity, improved reversal factor, end-of-day transaction proportion, average single-order outflow proportion, and large-order-driven price increase factor[7][9][10] - Deep learning factors include improved GRU(50,2)+NN(10), residual attention LSTM(48,2)+NN(10), multi-granularity model with 5-day labels, and multi-granularity model with 10-day labels[7][9][10] - AI-enhanced portfolios are constructed based on deep learning factors, combining multi-granularity models (5-day labels and 10-day labels) with constraints such as stock weight, industry weight, market cap, PB, ROE, SUE, volatility, turnover rate, and constituent stock limits. The optimization goal is to maximize expected returns using the formula $$max\sum\mu_{i}w_{i}$$, where \(w_i\) is the stock weight and \(\mu_i\) is the expected excess return[73][74][76] - The report provides detailed backtesting results for high-frequency factors, deep learning factors, and AI-enhanced portfolios, including weekly, monthly, and annual returns, as well as win rates across different timeframes[9][10][11]