深度学习因子月报：Meta因子5月实现超额收益3.9%-20250611

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 a dynamic market state factor[19][21]. - Construction Process: - Matrix decomposition is applied to fund-stock investment networks to derive intrinsic attributes of funds and stocks. - Static intrinsic attributes are updated semi-annually using fund reports and transformed into dynamic attributes by calculating their similarity to the market's current style preferences. - These dynamic attributes are combined with LSTM outputs and fed into an MLP model to enhance factor performance[19][21]. - Evaluation: The factor effectively captures dynamic market preferences and improves model performance[19][21]. 2. Factor Name: Meta_RiskControl - Construction Idea: Integrate factor exposure control into deep learning models to mitigate risks during rapid style shifts, leveraging meta-incremental learning for market adaptability[25][28]. - Construction Process: - Multiply model outputs by corresponding stock factor exposures and include this in the loss function. - Add penalties for style deviation and momentum to the IC-based loss function. - Use an ALSTM model with style inputs as the base model and apply a meta-incremental learning framework for periodic updates[25][28]. - Evaluation: The factor reduces style deviation and volatility, effectively controlling model drawdowns[25][28]. 3. Factor Name: Meta_Master - Construction Idea: Incorporate market state information into the model, leveraging deep risk models and online meta-incremental learning to adapt to dynamic market conditions[35][37]. - Construction Process: - Use deep risk models to calculate new market states and construct 120 new features representing market preferences. - Replace the loss function with weighted MSE to improve long-side prediction accuracy. - Apply online meta-incremental learning for periodic model updates, enabling quick adaptation to recent market trends[35][37]. - Evaluation: The factor demonstrates significant improvements in long-side prediction accuracy and market adaptability[35][37]. 4. Factor Name: Deep Learning Convertible Bond Factor - Construction Idea: Address the declining excess returns of traditional convertible bond strategies by using GRU neural networks to model the complex nonlinear pricing logic of convertible bonds[50][52]. - Construction Process: - Introduce convertible bond-specific time-series factors into the GRU model. - Combine cross-sectional attributes of convertible bonds with GRU outputs to predict future returns[50][52]. - Evaluation: The factor significantly enhances model performance compared to traditional strategies[50][52]. Factor Backtesting Results 1. DL_EM_Dynamic Factor - RankIC: 12.1% (May 2025)[9][12] - Excess Return: 0.6% (May 2025), 10.4% YTD[9][23] - Annualized Return: 29.7% (since 2019)[23] - Annualized Excess Return: 23.4% (since 2019)[23] - IR: 2.03[23] - Max Drawdown: -10.1%[23] 2. Meta_RiskControl Factor - RankIC: 12.8% (May 2025)[9][14] - Excess Return: -0.7% (HS300), 0.8% (CSI500), 0.5% (CSI1000) in May 2025; 3.0%, 4.8%, and 8.3% YTD respectively[9][30][34] - Annualized Return: 20.1% (HS300), 26.1% (CSI500), 34.1% (CSI1000) since 2019[30][32][34] - Annualized Excess Return: 15.0% (HS300), 19.2% (CSI500), 27.0% (CSI1000) since 2019[30][32][34] - IR: 1.58 (HS300), 1.97 (CSI500), 2.36 (CSI1000)[30][32][34] - Max Drawdown: -5.8% (HS300), -9.3% (CSI500), -10.2% (CSI1000)[30][32][34] 3. Meta_Master Factor - RankIC: 14.7% (May 2025)[9][17] - Excess Return: -0.5% (HS300), 0.5% (CSI500), 0.4% (CSI1000) in May 2025; 4.2%, 3.3%, and 5.0% YTD respectively[38][44][47] - Annualized Return: 22.0% (HS300), 23.8% (CSI500), 30.7% (CSI1000) since 2019[38][44][47] - Annualized Excess Return: 17.5% (HS300), 18.2% (CSI500), 25.2% (CSI1000) since 2019[38][44][47] - IR: 2.09 (HS300), 1.9 (CSI500), 2.33 (CSI1000)[38][44][47] - Max Drawdown: -7.2% (HS300), -5.8% (CSI500), -8.8% (CSI1000)[38][44][47] 4. Deep Learning Convertible Bond Factor - Absolute Return: 1.7% (偏股型), 2.6% (平衡型), 1.7% (偏债型) in May 2025[52][55] - Excess Return: 0.1% (偏股型), 1.0% (平衡型), 0.2% (偏债型) in May 2025[52][55] - Annualized Return: 13.2% (偏股型), 11.8% (平衡型), 12.7% (偏债型) since 2021[52][55] - Annualized Excess Return: 5.8% (偏股型), 4.0% (平衡型), 4.4% (偏债型) since 2021[52][55]