- The report highlights the strong performance of high-frequency factors, with notable multi-granularity factor returns and differentiated excess returns in AI-enhanced portfolios [6][10][11] - High-frequency factors such as intraday skewness, downside volatility proportion, and opening buy intention proportion recorded significant multi-long-short returns and excess returns across weekly, monthly, and YTD periods [6][10][11] - Multi-granularity models (5-day and 10-day labels) achieved robust multi-long-short returns, with the 5-day label model yielding 0.91% (weekly), 2.6% (monthly), and 10.22% (YTD), while the 10-day label model delivered 0.71% (weekly), 3.06% (monthly), and 8.9% (YTD) [6][10][11] - AI-enhanced portfolios, including the Air Index Increment and CSI 500/1000 AI Enhanced portfolios, demonstrated varying excess and absolute returns under both wide and strict constraints, with weekly and daily rebalancing strategies [6][15][73] - The optimization objective for AI-enhanced portfolios is to maximize expected returns, represented by the function: $$max\sum\mu_{i}w_{i}$$ where \(w_i\) is the weight of stock \(i\) in the portfolio, and \(\mu_i\) is the expected excess return of stock \(i\) [75][76]

Quantitative Models and Construction Methods - **Model Name**: Multi-granularity model (5-day label) **Construction Idea**: The model leverages deep learning techniques to capture multi-granularity features in stock data over a 5-day horizon[70][66] **Construction Process**: Based on bidirectional AGRU (Attention Gated Recurrent Unit) training, the model combines historical data and predictive features to optimize stock selection[68][66] **Evaluation**: Demonstrates consistent positive returns across multiple testing periods[70][66] - **Model Name**: Multi-granularity model (10-day label) **Construction Idea**: Similar to the 5-day label model, but extends the horizon to 10 days for broader feature extraction[70][68] **Construction Process**: Utilizes bidirectional AGRU training, integrating longer-term predictive signals for enhanced stock selection[68][70] **Evaluation**: Provides robust performance with slightly lower returns compared to the 5-day label model[70][68] - **Model Name**: AI-enhanced index optimization models **Construction Idea**: Combines deep learning factors (multi-granularity models) with risk control constraints to optimize portfolio returns[70][71] **Construction Process**: 1. **Objective Function**: $$ max\sum\mu_{i}w_{i} $$, where \( w_i \) represents stock weights and \( \mu_i \) represents expected excess returns[74][75] 2. **Constraints**: Includes stock-level, industry-level, market capitalization, PB, ROE, and turnover rate constraints[71][74] **Evaluation**: Effective in balancing risk and return, with varying performance across different constraint levels[70][71] Model Backtesting Results - **Multi-granularity model (5-day label)**: - IC: Historical 0.079, 2026 0.046 - RankMAE: Historical 0.343, 2026 0.337 - Multi-long-short returns: 8.05% (2026), 0.43% (March)[10][13][70] - **Multi-granularity model (10-day label)**: - IC: Historical 0.073, 2026 0.044 - RankMAE: Historical 0.342, 2026 0.340 - Multi-long-short returns: 6.25% (2026), 0.42% (March)[10][13][70] - **AI-enhanced index optimization models**: - **Air Quality Index Model**: - Weekly excess returns: 4.28% (2026), 0.55% (March) - Daily excess returns: 5.51% (2026), 0.61% (March)[79][76][70] - **CSI 500 Wide Constraint Model**: - Weekly excess returns: -0.93% (2026), 2.61% (March) - Daily excess returns: -4.25% (2026), 0.73% (March)[82][81][70] - **CSI 500 Strict Constraint Model**: - Weekly excess returns: 1.02% (2026), 2.09% (March) - Daily excess returns: 0.42% (2026), 1.29% (March)[84][83][70] - **CSI 1000 Wide Constraint Model**: - Weekly excess returns: 2.98% (2026), 2.43% (March) - Daily excess returns: 2.54% (2026), 2.36% (March)[88][87][70] - **CSI 1000 Strict Constraint Model**: - Weekly excess returns: 2.34% (2026), 1.33% (March) - Daily excess returns: 3.10% (2026), 1.25% (March)[93][91][70] Quantitative Factors and Construction Methods - **Factor Name**: Intraday skewness factor **Construction Idea**: Captures the asymmetry in intraday stock return distributions[15][17] **Construction Process**: Referenced from the report "Stock Selection Factor Series Research (19)"[15][17] **Evaluation**: Strong performance in capturing short-term market movements[15][17] - **Factor Name**: Downside volatility proportion factor **Construction Idea**: Measures the proportion of realized downside volatility in stock returns[20][22] **Construction Process**: Referenced from the report "Stock Selection Factor Series Research (25)"[20][22] **Evaluation**: Effective in identifying risk-averse stocks[20][22] - **Factor Name**: Opening buy intention proportion factor **Construction Idea**: Quantifies the proportion of buy orders during the opening period[23][27] **Construction Process**: Referenced from the report "Stock Selection Factor Series Research (64)"[23][27] **Evaluation**: Demonstrates strong predictive power for short-term returns[23][27] Factor Backtesting Results - **Intraday skewness factor**: - IC: Historical 0.026, 2026 0.036 - RankMAE: Historical 0.324, 2026 0.334 - Multi-long-short returns: 3.69% (2026), -0.62% (March)[10][11][15] - **Downside volatility proportion factor**: - IC: Historical 0.025, 2026 0.043 - RankMAE: Historical 0.324, 2026 0.333 - Multi-long-short returns: 5.65% (2026), -0.17% (March)[10][11][20] - **Opening buy intention proportion factor**: - IC: Historical 0.031, 2026 0.043 - RankMAE: Historical 0.322, 2026 0.326 - Multi-long-short returns: 4.27% (2026), 1.30% (March)[10][11][23]

Investment Rating - The report indicates a positive performance of large order factors and stable performance of multi-granularity factors, with AI-enhanced portfolios showing strong results [2][3]. Core Insights - The report highlights the differentiation in high-frequency factors, with large order factors performing well. The AI-enhanced portfolio continues to show strong performance [2][3]. - The multi-granularity factors have maintained stable performance, with significant returns reported for both 5-day and 10-day label models [6][9]. - The AI-enhanced portfolios have demonstrated substantial excess and absolute returns across various combinations, indicating effective investment strategies [14][73]. Summary by Sections High-Frequency Factors and Deep Learning Factors - High-frequency skewness factor reported a multi-directional return of -1.11% for the last week and 2.89% for February 2026 [6][10]. - The downward volatility proportion factor showed a return of -0.61% for the last week and 4.14% for February 2026 [6][10]. - The opening buy intention proportion factor had a return of -0.04% for the last week and 3.82% for February 2026 [6][10]. - The opening large order net buy proportion factor reported a return of 0.34% for the last week and 2.89% for February 2026 [6][10]. - The AI-enhanced portfolios, including the weekly rebalancing AI air value increase portfolio, reported excess returns of 3.63% and absolute returns of 3.29% for the last week [14][73]. Multi-Granularity Factors - The multi-granularity model (5-day label) reported a multi-directional return of 0.65% for the last week and 6.15% for February 2026 [6][10]. - The multi-granularity model (10-day label) reported a multi-directional return of 0.53% for the last week and 4.60% for February 2026 [6][10]. AI Enhanced Portfolios - The weekly rebalancing of the CSI 500 AI enhanced wide constraint portfolio reported an excess return of 2.25% and an absolute return of -0.42% for the last week [14][73]. - The daily rebalancing of the CSI 1000 AI enhanced strict constraint portfolio reported an excess return of 1.65% and an absolute return of -0.82% for the last week [14][73].

- The high-frequency skewness factor had long-short returns of 0.67% last week, -1.18% in December, and 22.39% year-to-date 2025[5] - The intraday downside volatility factor had long-short returns of 0.87% last week, -1.33% in December, and 19.08% year-to-date 2025[5] - The post-open buying intention proportion factor had long-short returns of 0.66% last week, 0.61% in December, and 21.12% year-to-date 2025[5] - The post-open buying intention intensity factor had long-short returns of 0.46% last week, 0.94% in December, and 28.09% year-to-date 2025[5] - The post-open large order net buying proportion factor had long-short returns of -0.21% last week, 0.17% in December, and 22.11% year-to-date 2025[5] - The post-open large order net buying intensity factor had long-short returns of -0.25% last week, 0.38% in December, and 12.5% year-to-date 2025[5] - The intraday return factor had long-short returns of 0.35% last week, 0.91% in December, and 22.33% year-to-date 2025[5] - The end-of-day trading proportion factor had long-short returns of -0.94% last week, 1.04% in December, and 16.73% year-to-date 2025[5] - The average single transaction outflow proportion factor had long-short returns of -1.15% last week, -2.15% in December, and -8.11% year-to-date 2025[5] - The large order push-up factor had long-short returns of 0.41% last week, -0.93% in December, and 7.19% year-to-date 2025[5] - The GRU(10,2)+NN(10) factor had long-short returns of 1.13% last week, -0.47% in December, and 47.04% year-to-date 2025, with long-only excess returns of -0.2% last week, -0.26% in December, and 7.1% year-to-date 2025[5] - The GRU(50,2)+NN(10) factor had long-short returns of 1.66% last week, 0.19% in December, and 47.39% year-to-date 2025, with long-only excess returns of 0.15% last week, 0.06% in December, and 8.92% year-to-date 2025[5] - The multi-granularity model (5-day label) factor had long-short returns of 2.46% last week, 1.12% in December, and 68.13% year-to-date 2025, with long-only excess returns of 0.74% last week, -0.18% in December, and 24.48% year-to-date 2025[5] - The multi-granularity model (10-day label) factor had long-short returns of 2.26% last week, 1.11% in December, and 62.71% year-to-date 2025, with long-only excess returns of 0.76% last week, -0.5% in December, and 24.3% year-to-date 2025[5] - The weekly rebalanced CSI 500 AI-enhanced wide constraint portfolio had excess returns of 0.41% last week, -2.64% in December, and 5.46% year-to-date 2025[5] - The weekly rebalanced CSI 500 AI-enhanced strict constraint portfolio had excess returns of 0.92% last week, -1.62% in December, and 9.23% year-to-date 2025[5] - The weekly rebalanced CSI 1000 AI-enhanced wide constraint portfolio had excess returns of 1.55% last week, -2.69% in December, and 15.39% year-to-date 2025[5] - The weekly rebalanced CSI 1000 AI-enhanced strict constraint portfolio had excess returns of 1.48% last week, -1.45% in December, and 19.02% year-to-date 2025[5]

Core Viewpoint - The article discusses the integration of return factors into an orthogonal layer within deep learning models to enhance stock selection effectiveness while maintaining low correlation with existing return factors [1][2]. Group 1: Deep Learning Model Enhancements - By incorporating return factors into the orthogonal layer, deep learning factors can maintain good stock selection performance while ensuring low correlation with these return factors [1]. - The deep learning model's black-box nature makes it challenging to manually adjust factor weights during significant market style shifts; thus, the orthogonal layer allows for easier manual adjustments without compromising stock selection effectiveness [1]. Group 2: Performance Metrics - After adding return factors to the orthogonal layer, deep learning factors still exhibit strong stock selection capabilities, achieving an Information Coefficient (IC) above 0.02 and an IC Information Ratio (IR) exceeding 6 [2]. - The combination of deep learning factors with manually constructed return factors leads to significant improvements in overall market long positions compared to using deep learning factors alone, although the enhancement varies across different index-enhanced portfolios [2]. Group 3: Correlation and Performance - The correlation between deep learning factors and multi-granularity factors remains low after integrating return factors into the orthogonal layer, with high-frequency data inputs showing a correlation of no more than 0.01 [2]. - Utilizing deep learning factors alongside multi-granularity factors can significantly enhance the performance of overall market long positions, although the deep learning factors show limited predictive capability for mid to large-cap stock returns, resulting in less noticeable improvements for index-enhanced portfolios [2].