Graph Structured Semantic Representation and Learning for Financial News.

This study links stock prices of publicly traded companies with online financial news to predict direction of stock price change. Previous work shows this to be an extremely challenging problem. We develop a very high-dimensional representation for news about companies that encodes lexical, syntactic and frame semantic information in graphs. Use of a graph kernel to efficiently compare subgraphs for machine learning provides a uniform feature engineering framework that integrates semantic frames in document representation. Evaluated on a news web archive against two benchmarks, only our approach beats the majority class baseline, and with statistically significant results.