A Motif-Based Graph Convolution Network for Stock Trend Prediction

Lei Zhou, Yuqi Zhang, Nancy Wang, Jian Yu, Guiling Wang, Xin Zheng

Abstract: The prediction of stock market trends remains a challenging task, garnering significant attention from economists and computer scientists. Recent studies have demonstrated that integrating stock prices with news data improves prediction accuracy. However, current approaches often overlook the complex inter-stock relationships embedded in news data. Deep learning, particularly Graph Convolutional Networks (GCNs), has shown promising results in stock trend prediction by leveraging a message-passing framework that allows nodes to aggregate information from their neighbors. In this paper, we propose a novel method, the Motif-based Graph Convolutional Network for Stock Prediction (MGSP), which addresses the over-smoothing problem by incorporating network motifs into the layer propagation. Our approach constructs a motif graph by correlating stocks with stock news, and employs scaled dot product attention from the transformer architecture to encode stock price and news features. The motif-based GCN is then applied to jointly optimize the embeddings of stock news and time series data, using a transformer encoder to estimate the probability of future price movements. Extensive experiments on U.S. stock data show that our method outperforms several state-of-the-art techniques.

External IDs:dblp:conf/iconip/ZhouZWYWZ24