Not only market signals but also disturbances of related companies influence the stock volatility of a company. Currently, most approaches that utilize inter-stock correlations rely on predetermined graphs, such as industry graphs, shareholder graphs, or event graphs, necessitating supplementary financial data. Consequently, this approach often leads to limited generalization ability of the model and incomplete representation of inter-stock relationships. Thus, this study introduces VGC-GAN, a multi-graph convolutional adversarial framework, for predicting stock prices. Initially, the proposed model generates multiple correlation graphs by analyzing historical stock data, providing a comprehensive depiction of inter-stock correlations from diverse perspectives. Subsequently, a Generative Adversarial Network (GAN) framework, augmented with Mean Square Error (MSE) loss, is constructed to enhance the predictive performance of the model. The framework combines Multi-Graph Convolutional Network (Multi-GCN) and Gated Recurrent Unit (GRU) as the generator. It undergoes supervised and adversarial training with Convolutional Neural Network (CNN), facilitating the in-depth exploration of hidden correlations between stocks and time dependence of stocks. To mitigate the impact of noise, VGC-GAN uses subsequences after Variational Mode Decomposition (VMD) with optimized parameters as input to the generator. The proposed model is evaluated on several real datasets, and the experimental results confirm its effectiveness in stock price prediction.
Read full abstract