Decomposition-Based Dynamic Inductive Graph Embedding Learning Method to Forecast Stock Trends

The stock market is a profit-oriented, chaotic, and nonlinear market game platform. Because price changes are directly related to investors’ returns, the accurate prediction of the short-term trend of stock prices has garnered significant attention. Recently, the application of a graph neural network (GNN) has become a research hotspot because of its ability to mine the momentum spillover effect between asset prices to achieve more effective forecasting. However, the graph data used in previous studies are difficult to obtain and process, and there is much room for improvement at the feature processing level. In this study, we construct a dynamic inductive predictive graph neural network (DIP-GNN) model, which introduces a sequence decomposition algorithm to separate the wave modes of the data, which can considerably simplify the learning difficulty of the model, thereby improving the overall performance. In addition, on the basis of the information captured by a recurrent neural network in the time dimension, the model uses Pearson’s correlation coefficient, Manhattan distance, and dynamic time warping to dynamically evaluate the short-term correlation between daily stock prices, and then the identified dynamic relation network to extract valuable cross-sectional information. By combining the cross-sectional information with the temporal information and applying it to the graph representation learning task, stock dynamics can be revealed more effectively. Multiple comparative experiments show that DIP-GNN exhibits better predictive performance than the benchmark models and has robust and much superior profitability over traditional strategies in several markets.