How do financial time series enhance the detection of news significance in market movements? A study using graph neural networks with heterogeneous representations

Forecasting trends in the financial market is a classic and challenging problem that attracts economists’ and computer scientists’ attention. This research area, characterized by its dynamic, chaotic, and nonlinear nature, is further complicated by the overarching influence of the efficient market hypothesis (EMH). The EMH posits that all available information, including historical prices and public news, is already reflected in current asset prices. It suggests that gaining consistent predictive advantages by leveraging such information is challenging. This paper evaluates different machine learning models to identify relevant news based on oscillations in a financial time series. Specifically, we explore the state-of-the-art in graph neural networks, which have the advantage of combining different representations of temporal series and textual data. As a result, we introduce three approaches to classify news as relevant or irrelevant and to model textual data and time series through graphs, taking into account the implications of the EMH. These approaches include text and time-series clusters with daily data, data occurring at perceptually important points in the time series, and data from moments when more than 70% of the news is classified as relevant. We find that similar to the challenge of using the news to enhance the prediction of financial series, the reverse is also true, highlighting the difficulty of identifying relevant news that potentially impacts commodity price fluctuations.