MSMixer-GCN: A Hybrid Deep Learning Model for Stock Price Prediction Using Multi-Scale Convolution, MLP-Mixer and Graph Convolution Networks

Stock price prediction is a challenging problem in the field of finance that receives significant attention. Researchers have developed combinations of neural network models, such as RNNs, GNNs, and Transformers, which show great potential for stock forecasting. However, these models often struggle with the complex nonlinear dynamics of the stock market and neglect the intertwined spatial relationships and temporal features among stocks. This oversight can adversely affect prediction accuracy. Therefore, we propose a stock price prediction model named MSMixer-GCN in this paper, which is based on a multi-scale convolutional MLP Mixer and a spatio-temporal graph convolutional network. Firstly, we adopt a multi-scale convolution combined with the MLP-Mixer approach to construct the MS-MLP Mixer module, which captures trends in stock indicators and temporal dimensions, thereby enhancing the utilization of historical time series information. Next, we develop a dynamic spatio-temporal graph convolutional network that allows spatial structural information to interact with features across different time scales, significantly improving the accuracy of stock price predictions. Finally, through extensive experiments on three datasets from the United States stock market, our proposed MSMixer-GCN model demonstrates outstanding performance, surpassing state-of-the-art models in terms of IC ranking metrics and the Sharpe Ratio, achieving average relative performance gains of 6.25% and 8.67%, respectively. This confirms the superior performance and generalization capability of the model.