Guided Attention Multimodal Multitask Financial Forecasting with Inter-Company Relationships and Global and Local News

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6313 - 6326 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Guided Attention Multimodal Multitask Financial Forecasting with Inter-Company Relationships and Global and Local News Gary Ang Singapore Management University gary.ang.2019@phdcs.smu.edu.sg Ee-Peng Lim Singapore Management University eplim@smu.edu.sg Abstract Most works on financial forecasting use infor- mation directly associated with individual com- panies (e.g., stock prices, news on the com- pany) to predict stock returns for trading. We refer to such company-specific information as local information. Stock returns may also be influenced by global information (e.g., news on the economy in general), and inter-company relationships. Capturing such diverse informa- tion is challenging due to the low signal-to- noise ratios, different time-scales, sparsity and distributions of global and local information from different modalities. In this paper, we pro- pose a model that captures both global and lo- cal multimodal information for investment and risk management-related forecasting tasks. Our proposed Guided Attention Multimodal Multi- task Network (GAME) model addresses these challenges by using novel attention modules to guide learning with global and local infor- mation from different modalities and dynamic inter-company relationship networks. Our ex- tensive experiments show that GAME outper- forms other state-of-the-art models in several forecasting tasks and important real-world ap- plication case studies. 1 Introduction Forecasting stock prices or returns is an important task in trading. Such forecasts can also be used in investment and risk management applications such as portfolio allocation and risk forecasting. Stock returns in financial markets are influenced by large volumes of textual information from diverse sources, e.g., news, blogs, social media. Such tex- tual information can be directly associated with a specific company (local), e.g, a company’s CEO stepping down; or relevant to multiple companies (global), e.g., disruptions in supply chains due to export curbs in key countries, airline industry bankruptcies. In this paper, articles with company tags are treated as local information. All articles are treated as global information as any article could be potentially relevant to a company. Direct and indirect relationships between com- panies also serve as channels through which the effects of information from both global and local textual and numerical information propagate and in- fluence stock returns, e.g., a disruption in company A could affect all its suppliers; a scandal involv- ing company A’s CEO may affect company B if the CEO is a member of company B’s board. We illustrate such diverse information and effects in Figure 1. Apart from low signal-to-noise ratios in finan- cial time-series due to market forces, there are other challenges in modeling such diverse information. Time scales of information from different modal- ities are of different granularity, e.g., numerical financial information may be available daily, while publication of financial text happens at irregular times. Companies’ local financial news are typi- cally sparse and long-tailed, e.g., a company may not be in the news for an extended period of time, but suddenly becomes the focus of many news …

Content truncated to 500 words