机器学习因子选股月报（2026年3月）

Quantitative Models and Construction Methods 1. Model Name: GAN_GRU Model - Model Construction Idea: The GAN_GRU model combines Generative Adversarial Networks (GAN) for feature generation and Gated Recurrent Unit (GRU) for time-series feature encoding to create a stock selection factor[4][13][14] - Model Construction Process: - GAN Component: - The GAN consists of a generator (G) and a discriminator (D). The generator learns the real data distribution and generates realistic samples, while the discriminator distinguishes between real and generated data[23][24] - Generator loss function: $L_{G} = -\mathbb{E}_{z\sim P_{z}(z)}[\log(D(G(z)))]$ where $z$ is random noise, $G(z)$ is the generated data, and $D(G(z))$ is the discriminator's output probability for generated data being real[24][25] - Discriminator loss function: $L_{D} = -\mathbb{E}_{x\sim P_{data}(x)}[\log D(x)] - \mathbb{E}_{z\sim P_{z}(z)}[\log(1-D(G(z)))]$ where $x$ is real data, $D(x)$ is the discriminator's output probability for real data, and $D(G(z))$ is the discriminator's output probability for generated data[27][29] - GAN training alternates between updating the generator and discriminator parameters through backpropagation[30] - The generator uses an LSTM model to preserve the sequential nature of input features, while the discriminator employs a CNN model to process the 2D structure of the generated features[33][37] - GRU Component: - Two GRU layers (GRU(128, 128)) are used, followed by an MLP (256, 64, 64) to output predicted returns ($pRet$) as the stock selection factor[22] - Input features include 18 price-volume characteristics (e.g., closing price, turnover rate) sampled over the past 40 days to predict cumulative returns for the next 20 trading days[14][18] - Data preprocessing includes outlier removal, standardization, and cross-sectional normalization[18] - Training is conducted semi-annually with rolling updates, using Adam optimizer, a learning rate of $1e-4$, and IC as the loss function[18] - Model Evaluation: The GAN_GRU model effectively integrates GAN's feature generation capabilities with GRU's time-series encoding, making it suitable for capturing complex price-volume patterns in stock selection[4][13] --- Model Backtesting Results GAN_GRU Model - IC Mean: 0.1096*** (2019.02–2026.02)[41] - ICIR (Non-Annualized): 0.87[42] - Turnover Rate: 0.82X[42] - Recent IC: -0.0105*** (latest period)[41][42] - 1-Year IC Mean: 0.0517***[41][42] - Annualized Return: 38.13%[42] - Annualized Volatility: 23.18%[42] - IR: 1.64[42] - Maximum Drawdown: 27.29%[42] - Annualized Excess Return: 22.32%[41][42]