Core Insights - The article presents a composite enhancement strategy for the CSI 500 index, which has achieved an annualized excess return of 16.6% relative to the benchmark from January 2014 to February 2026, with a year-to-date annualized excess return of 9.6% in 2023 [1]. Group 1: Characteristics of the CSI 500 Index - The CSI 500 index has evolved with an upward shift in market capitalization percentiles, moving towards a mid-to-large-cap characteristic; the weight of tail stocks has decreased, leading to increased concentration [2]. - Since 2022, the volatility of factor returns has increased, prompting improvements in the enhancement strategy for the CSI 500 index [2]. Group 2: Factor Weighting and Dynamic Adjustment - In the context of increased factor return volatility, using ICIR weighting, which considers volatility, has proven to be more stable compared to the traditional IC mean weighting method. The annualized excess return of the CSI 500 enhancement strategy using ICIR weighting is 5.21%, significantly higher than the 1.43% from IC mean weighting [2]. - Dynamic adjustment of factor exposure can enhance alpha sources, especially in a declining overall factor return environment. The annualized excess return of the CSI 500 enhancement strategy, based on PLS model dynamic adjustments, has risen to 7.02% since 2023 [2]. Group 3: Satellite Strategy Construction - A GARP50 strategy has been constructed within the CSI 500 index constituents, yielding an annualized excess return of 14.2% relative to the benchmark since 2014, with an annualized tracking error of 6.0% and an information ratio of 2.20 [2]. - Allocating 30% of the weight to the GARP50 strategy has resulted in an increase in the composite enhancement strategy's annualized excess return to 9.6% relative to the benchmark since 2023, with a tracking error of 4.38% [2].

Core Insights - The report emphasizes the development of a deep learning factor selection system for stock prediction and portfolio optimization, shifting from traditional logic-driven methods to data-driven approaches [7][10]. - The system integrates diverse data sources, including daily and minute market data, to enhance the performance of alpha signals [7][10]. - The report outlines the construction of multiple models that utilize different network architectures to extract unique alpha signals, demonstrating low correlation among them [8][54]. Data and Network - The input data consists of three categories: daily market data, minute market data, and manually crafted features, with neural networks independently extracting alpha features from each dataset [11]. - The report describes the use of Long Short-Term Memory (LSTM) networks combined with self-attention mechanisms to capture long-term dependencies in time series data [19]. - A Graph Attention Network (GAT) is employed to model the complex relationships between stocks, providing a global analysis perspective [20]. Alpha Models - The report presents various alpha models, including simple equal-weight, tree model weighting, and network weighting, with a focus on combining multiple signals to enhance robustness [3][3.1][3.2]. - The average Information Coefficient (IC) for the combined factors since 2019 is reported as 11.3% for 5-day IC and 12.4% for 10-day IC, indicating strong predictive power [31][32]. Risk Models - The report highlights the use of neural networks to identify high-dimensional non-linear risk patterns directly from raw price and volume data, enhancing risk control in portfolio construction [9]. Index Enhancement Strategies - The report details the performance of enhanced index strategies based on deep learning alpha signals, with annualized returns reported as follows: - CSI 300 enhanced portfolio: 18.2% annualized return, 14.2% excess return over the index [3][5.1]. - CSI 500 enhanced portfolio: 22.4% annualized return, 17.2% excess return over the index [3][5.2]. - CSI 1000 enhanced portfolio: 29.8% annualized return, 24.5% excess return over the index [3][5.3].