机器学习因子选股月报（2025年7月）-20250630

Quantitative Factor and Model Analysis Quantitative Models and Construction 1. Model Name: GAN_GRU Model Model Construction Idea: The GAN_GRU model combines Generative Adversarial Networks (GAN) for generating realistic price-volume sequential features and Gated Recurrent Units (GRU) for encoding these sequential features into predictive signals for stock selection [2][9]. Model Construction Process: - GRU Component: - Input features include 18 price-volume features such as closing price, opening price, turnover, and turnover rate [10][13]. - Training data consists of the past 400 trading days' features, sampled every 5 trading days, forming a 40x18 feature matrix to predict the cumulative return over the next 20 trading days [14]. - Data preprocessing includes outlier removal and standardization at both time-series and cross-sectional levels [14]. - The GRU network consists of two layers (GRU(128, 128)) followed by an MLP (256, 64, 64), with the final output being the predicted return (pRet) [18]. - GAN Component: - The generator (G) uses an LSTM model to preserve the sequential nature of the input features, while the discriminator (D) employs a CNN to process the two-dimensional price-volume feature "images" [29][32]. - The generator's loss function is: $L_{G} = -\mathbb{E}_{z\sim P_{z}(z)}[\log(D(G(z)))]$ where $z$ represents random noise, $G(z)$ is the generated data, and $D(G(z))$ is the discriminator's output probability [20][21]. - The discriminator's loss function is: $L_{D} = -\mathbb{E}_{x\sim P_{data}(x)}[\log D(x)] - \mathbb{E}_{z\sim P_{z}(z)}[\log(1-D(G(z)))]$ where $x$ is real data, $D(x)$ is the discriminator's output for real data, and $D(G(z))$ is the output for generated data [23][25]. - Training alternates between updating the discriminator and generator parameters until convergence [26]. Model Evaluation: The GAN_GRU model effectively captures both sequential and cross-sectional price-volume features, leveraging the strengths of GANs and GRUs for stock selection [2][9][29]. --- Quantitative Factors and Construction 1. Factor Name: GAN_GRU Factor Factor Construction Idea: The GAN_GRU factor is derived from the GAN_GRU model's output, representing the encoded price-volume sequential features as a stock selection signal [2][9]. Factor Construction Process: - The factor is derived from the predicted return (pRet) output of the GAN_GRU model [18]. - The factor undergoes industry and market capitalization neutralization, followed by standardization [18]. Factor Evaluation: The GAN_GRU factor demonstrates strong predictive power across various industries, with consistent performance in both IC and excess returns [36][40]. --- Model Backtest Results 1. GAN_GRU Model: - IC Mean: 11.54% - ICIR: 0.89 - Turnover Rate: 0.83 - Recent IC: 8.34% - 1-Year IC Mean: 11.09% - Annualized Return: 37.71% - Annualized Volatility: 24.95% - IR: 1.56 - Max Drawdown: 27.29% - Annualized Excess Return: 24.95% [36][37]. --- Factor Backtest Results 1. GAN_GRU Factor: - IC Mean: 11.54% - ICIR: 0.89 - Turnover Rate: 0.83 - Recent IC: 8.34% - 1-Year IC Mean: 11.09% - Annualized Return: 37.71% - Annualized Volatility: 24.95% - IR: 1.56 - Max Drawdown: 27.29% - Annualized Excess Return: 24.95% [36][37].