FLAG: Stock Movement Prediction via Fusing Logic and Semantic Graphs of Financial News

Angela Lil, Jiduan Liu, Yuyong Lil, Fan Meng

Abstract: Financial news is widely used in stock movement prediction, and the critical point is to extract valuable infor-mation from news. Some previous works utilize graph-based approaches to represent news in a topological structure and construct relationships between nodes by semantic similarity. However, most of these models ignore depicting contextual information and logical structures in the news. In this work, we propose FLAG (Network with Fusion of Logic and semAntic Graphs), a novel model that captures both the logical structure and the latent semantic connection in the news for stock move-ment prediction. In FLAG, we construct two graphs for each news item, a logical graph based on continuation and coordination relations and a semantic graph based on semantic encoding. Then, we fuse the two graphs to represent news. Moreover, we combine news representations with global market information and historical prices for final prediction. FLAG achieves the state-of-the-art accuracy on the dataset we collected.