MERGE: Multi-view Relationship Graph Network for Event-Driven Stock Movement Prediction

Stock movement prediction has long been an attractive task in financial data mining, with banks and investment institutions attracted by its wide range of applications and potentially high value. In contrast to the conventional time series prediction tasks, the intrinsic characteristics of stocks render the incorporation of additional information a crucial factor in the prediction of stock movements. Inter-stock relationships and financial texts emerge as the most popular auxiliary information in this task. However, the acquisition of reliable inter-stock relationships is often difficult, while financial texts frequently contain substantial noise, which further complicates the task. In this work, we propose MERGE, a novel graph-based framework for the stock movement prediction task that efficiently exploits information from multiple sources and takes into account the interplay between them. MERGE involves a Multi-View Relationship Graph Network module that constructs multiple dynamic graphs by mining relational information in prices to model the various types of stock interactions in the market from different perspectives. In addition, to sufficiently consider the impact of external information on stock behavior, the Dualistic Event Encoder module extracts the most valuable parts from financial texts to capture the event-driven factors of stock volatility. Furthermore, extensive experiments on three real-world datasets also validate the effectiveness of our proposed MERGE framework compared with state-of-the-art baseline methods.