Stock Price Movement Prediction based on Relation Type guided Graph Convolutional Network

Stock Price Movement Prediction (SPMP) task tends to predict the future price fluctuation of some stocks, which is challenging because of the volatile nature of financial markets. Recent researches widely introduce Graph Convolutional Network (GCN) and achieve some competitive performances for SPMP. However, their works usually only extract the signal representations of individual stock in the phase of feature extraction whereas they ignore the real-time interactions among companies along the signal channels. In addition, when the financial relations among different entities (i.e., companies and executives) participate in the aggregation of GCN during feature encoding, the recent advances fail to consider the dynamics of the relation type representations and the iterative update between entities and relations. To address this problem, we propose a novel method consisting of an External Attention (EA) module and a Relation Type guided Graph Convolutional Network (RTGCN) for SPMP. In detail, to achieve the real-time interactions among companies along the signal channels, we introduce an external attention mechanism to share the company memory of the price and sentiment signals. Moreover, the proposed RTGCN considers the dynamics of the relation type embeddings and achieves iterative update between entity nodes and relation edges in the graph. The experiments on two open-source datasets (i.e., CSI100E and CSI300E) demonstrate that the proposed SPMP method can achieve state-of-the-art performances and can assuredly bring benefits for the performance improvement of SPMP.