Abstract: A hybrid deep learning stock index futures price prediction model based on VAE-ATTGRU is proposed, using variational autoencoder (VAE) and the recurrent neural network (RNN), to address the difficulty of predicting high-volatility, non-stationary, non-linear, and high signal-to-noise ratio characteristics in the stock index futures market. Firstly, the technical indicators of the stock index futures are learned using the VAE, and the latent factors learned by VAE are fused with the original data to achieve data augmentation and obtain a richer factor representation. Secondly, RNN is used to predict the stock index futures prices. It is found that the gated recurrent unit (GRU) combined with the attention mechanism (ATTGRU) can fully learn from the stock index futures data enhanced by VAE, capture key feature information, and reassign weights. The VAE-ATTGRU model is evaluated on datasets such as the CSI300 stock index futures, the CSI 500 index futures, and the SSE 50 index futures using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination R2. The experimental results demonstrate that the VAE-ATTGRU model outperforms other models in terms of prediction accuracy.

Key words: stock index futures prediction, variational autoencoder (VAE), data augmentation, attention mechanism, gated recurrent unit (GRU)

摘要： 针对股指期货市场高波动、非平稳、非线性和高信噪比等特性造成的预测难度大的问题，利用变分自编码器（VAE）和循环神经网络（RNN）提出一种基于VAE-ATTGRU的混合深度学习股指期货价格预测模型。利用变分自编码器对股指期货技术指标进行学习，将VAE学习到的潜在因子与原始数据融合实现数据增强，得到更丰富的因子表示；使用循环神经网络对股指期货价格进行预测，发现结合了注意力机制的门控循环单元（ATTGRU）可以对VAE增强后的股指期货数据进行充分学习，对关键特征信息进行捕捉并重新赋予权重。在沪深300股指期货、中证500股指期货和上证50股指期货数据上进行实验，通过均方根误差（RMSE）、平均绝对误差（MAE）、平均绝对百分比误差（MAPE）和决定系数R2对VAE-ATTGRU模型进行评估，发现其在预测精度上优于其他模型。

关键词: 股指期货预测, 变分自编码器（VAE）, 数据增强, 注意力机制, 门控循环单元（GRU）