Identifying Predictive Causal Factors from News Streams

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2338–2348, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2338 Identifying Predictive Causal Factors from News Streams Ananth Balashankar1, Sunandan Chakraborty2, Samuel Fraiberger1,3, and Lakshminarayanan Subramanian1 1Courant Institute of Mathematical Sciences, New York University 2School of Informatics and Computing, Indiana University-Indianapolis 3World Bank, Washington DC ananth@nyu.edu, sunchak@iu.edu, sfraiberger@worldbank.org, lakshmi@nyu.edu Abstract We propose a new framework to uncover the relationship between news events and real world phenomena. We present the Predictive Causal Graph (PCG) which allows to detect latent relationships between events mentioned in news streams. This graph is constructed by measuring how the occurrence of a word in the news influences the occurrence of an- other (set of) word(s) in the future. We show that PCG can be used to extract latent fea- tures from news streams, outperforming other graph-based methods in prediction error of 10 stock price time series for 12 months. We then extended PCG to be applicable for longer time windows by allowing time-varying fac- tors, leading to stock price prediction error rates between 1.5% and 5% for about 4 years. We then manually validated PCG, finding that 67% of the causation semantic frame argu- ments present in the news corpus were directly connected in the PCG, the remaining being connected through a semantically relevant in- termediate node. 1 Introduction Contextual embedding models (Devlin et al., 2018) have managed to produce effective repre- sentations of words, achieving state-of-the-art per- formance on a range of NLP tasks. In this pa- per, we consider a specific task of predicting vari- ations in stock prices based on word relationships extracted from news streams. Existing word em- bedding techniques are not suited to learn relation- ships between words appearing in different docu- ments and contexts (Le and Mikolov, 2014). Ex- isting work on stock price prediction using news have typically relied on extracting features from financial news (Falinouss, 2007; Hagenau et al., 2013), or sentiments expressed on Twitter (Mao et al., 2011; Rao and Srivastava, 2012; Bernardo et al., 2018), or by focusing on features present in a single document (Kalyani et al., 2016; Shynke- vich et al., 2015). However, relationships between events affecting stock prices can be quite com- plex, and their mentions can be spread across mul- tiple documents. For instance, market volatility is known to be triggered by recessions; this relation- ship may be reflected with a spike in the frequency of the word ”recession” followed by a spike in the frequency of the word ”volatility” a few weeks later. Existing methods are not well-equipped to deal with these cases. This paper aims to uncover latent relationships between words describing events in news streams, allowing us to unveil hidden links between events spread across time, and integrate them into a news-based predictive model for stock prices. We propose the Predictive Causal Graphs (PCG), a framework allowing us to detect latent relation- ships …

Content truncated to 500 words