Quantitative Models and Construction Methods Model Name: Excess Return Attribution Model - **Model Construction Idea**: Decompose the excess return of a fund's portfolio relative to the benchmark into three dimensions: style, industry, and stock selection[12] - **Model Construction Process**: The excess return of the portfolio is decomposed as follows: $ \text{Portfolio Excess Return} = \text{Style Return} + \text{Industry Return} + \text{Stock Selection Return} $ This decomposition allows for the identification of the primary sources of excess return, highlighting that active equity funds tend to lose from style, remain neutral in industry, and gain from stock selection[12][14] - **Model Evaluation**: The model effectively identifies that stock selection is the primary driver of alpha, while style and industry contributions are less significant or negative[14] --- Model Name: Core-Satellite Strategy (Scheme ①) - **Model Construction Idea**: Allocate a portion (W%) of the portfolio to replicate the benchmark index (core) and the remaining (1-W%) to active management (satellite)[19] - **Model Construction Process**: 1. Allocate W% of the portfolio to replicate the benchmark index (e.g., CSI 300) 2. Allocate the remaining (1-W%) to active stock selection based on the fund manager's views 3. Optimize the portfolio to minimize tracking error and performance deviation[19][21] - Example: For W=50%, the optimized portfolio reduced daily absolute deviation from 0.80% (simulated portfolio) to 0.40%[21] - **Model Evaluation**: This strategy effectively controls tracking error and performance deviation without reducing excess returns. It is particularly effective for large sample sizes and can be adjusted based on specific performance evaluation requirements[23][24] --- Model Name: Industry Neutralization Strategy (Scheme ②) - **Model Construction Idea**: Ensure the portfolio's industry allocation matches the benchmark (e.g., CSI 300) while focusing on stock selection to outperform industry indices[40] - **Model Construction Process**: 1. Adjust the portfolio's stock weights to achieve industry neutrality relative to the benchmark 2. Replace uncovered industries in the simulated portfolio with industry indices 3. Optimize the portfolio to minimize tracking error and performance deviation[40][43] - Example: For a specific fund, the optimized portfolio reduced daily absolute deviation from 1.03% (simulated portfolio) to 0.24%[43] - **Model Evaluation**: This strategy effectively controls tracking error and performance deviation while maintaining excess return potential. It is particularly suitable for portfolios with broad industry coverage[46][49] --- Model Name: Style Neutralization Strategy (Scheme ③) - **Model Construction Idea**: Minimize style deviation relative to the benchmark by optimizing stock weights without changing the stock selection[53] - **Model Construction Process**: 1. Use a weight optimizer to adjust stock weights in the portfolio 2. Minimize style exposure deviation relative to the benchmark (e.g., CSI 300) 3. Optimize the portfolio to reduce tracking error and performance deviation[53][54] - Example: For a specific fund, the optimized portfolio reduced daily absolute deviation from 0.47% (simulated portfolio) to 0.27%[54] - **Model Evaluation**: This strategy is simple, cost-effective, and achieves significant improvements in tracking error and performance deviation. It is particularly effective for large sample sizes[55][58] --- Model Name: Barbell Strategy (Scheme ④) - **Model Construction Idea**: Combine extreme growth and extreme value strategies to reduce tracking error and smooth portfolio volatility[61] - **Model Construction Process**: 1. Allocate 50% of the portfolio to a growth strategy (e.g., Wind Growth Fund Index) 2. Allocate the remaining 50% to a value strategy (e.g., Dividend Low Volatility Index) 3. Optimize the portfolio to balance risk and return[64][65] - Example: The combined portfolio achieved an annualized excess return of 3.20%, with a tracking error of 8.35% and a maximum drawdown of 44.52%[64][65] - **Model Evaluation**: This strategy is effective for managers with extreme style biases, significantly reducing tracking error and portfolio volatility while improving the holding experience[66][67] --- Backtesting Results of Models Core-Satellite Strategy (Scheme ①) - Annualized Tracking Error: 7.56% (W=50%)[30] - Maximum Deviation: 1.58% (W=50%)[30] - Average Deviation: 0.37% (W=50%)[30] - Annualized Excess Return: 1.74% (W=50%)[30] - Maximum Excess Drawdown: 5.78% (W=50%)[30] - IR: 0.1651 (W=50%)[30] Industry Neutralization Strategy (Scheme ②) - Annualized Tracking Error: 10.00%[51] - Maximum Deviation: 2.50%[51] - Average Deviation: 0.60%[51] - Annualized Excess Return: 2.00%[51] - Maximum Excess Drawdown: 6.00%[51] - IR: 0.2000[51] Style Neutralization Strategy (Scheme ③) - Annualized Tracking Error: 6.00%[60] - Maximum Deviation: 1.50%[60] - Average Deviation: 0.40%[60] - Annualized Excess Return: 3.00%[60] - Maximum Excess Drawdown: 4.00%[60] - IR: 0.5000[60] Barbell Strategy (Scheme ④) - Annualized Tracking Error: 8.16% (W=50%)[67] - Maximum Deviation: 2.42% (W=50%)[67] - Average Deviation: 0.39% (W=50%)[67] - Annualized Excess Return: 8.51% (W=50%)[67] - Maximum Excess Drawdown: 20.62% (W=50%)[67] - IR: 1.0420 (W=50%)[67]

- The DFQ-Diversify model effectively addresses the out-of-distribution generalization problem by introducing a self-supervised domain recognition and adversarial training mechanism, achieving explicit decoupling of label prediction and domain recognition tasks[2][3][10] - The model's training process includes three core modules: update_d, set_dlabel, and update, which work together through adversarial training to complete domain recognition and label prediction tasks, achieving explicit decoupling of the two[3][22][23] - The update_d module is responsible for domain recognition, using a GRU-based feature extractor, a domain bottleneck layer, a domain classifier, and a label adversarial discriminator to enhance domain representation accuracy and robustness[23][24][25] - The set_dlabel module updates the domain labels of samples through inference and clustering optimization, ensuring that the domain labels reflect the actual distribution of features in the feature space[28][29] - The update module focuses on label prediction, using a shared GRU feature extractor, a label bottleneck layer, a label classifier, and a domain adversarial discriminator to enhance label prediction accuracy and robustness[30][31][32] - The model employs a self-supervised dynamic domain partitioning mechanism, which helps the model autonomously identify potential domain information, enhancing its flexibility and generalization adaptability[34][36] - The DFQ-Diversify model constructs a three-level adversarial training mechanism, including inter-module task adversarial updates, intra-module dual loss adversarial balance, and gradient reversal layer mechanism, to achieve feature decoupling and robust transfer learning[42][43][47] - Compared to the Factorvae-pro model, the DFQ-Diversify model introduces self-supervised learning to dynamically identify potential domains, enhancing flexibility and generalization ability[50][53] - The DFQ-Diversify model shows superior performance in multiple stock pools, especially in large-cap stocks, with significant excess returns in the CSI All Share Index, CSI 300, and CSI 500 stock pools[5][6][107] - The model's backtesting results indicate that it achieved an IC of 12.22%, rankIC of 14.58%, and an annualized excess return of 32.52% in the CSI All Share Index stock pool from 2020 to 2025[5][107] - In the CSI 300 and CSI 500 enhanced strategies, the model achieved an IR of 1.89 and 1.67, and annualized excess returns of 11.27% and 12.19%, respectively[6][172][180]