Learning Target-Specific Representations of Financial News Documents For Cumulative Abnormal Return Prediction

Proceedings of the 27th International Conference on Computational Linguistics, pages 2823–2833 Santa Fe, New Mexico, USA, August 20-26, 2018. 2823 Learning Target-Specific Representations of Financial News Documents For Cumulative Abnormal Return Prediction Junwen Duan†∗, Yue Zhang‡, Xiao Ding†, Ching-Yun Chang‡, Ting Liu† †Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, Harbin, China {jwduan,xding,tliu}@ir.hit.edu.cn ‡Singapore University of Technology and Design yue zhang@sutd.edu.sg,chang.frannie@gmail.com Abstract Texts from the Internet serve as important data sources for financial market modeling. Early statistical approaches rely on manually defined features to capture lexical, sentiment and event information, which suffers from feature sparsity. Recent work has considered learning dense rep- resentations for news titles and abstracts. Compared to news titles, full documents can contain more potentially helpful information, but also noise compared to events and sentences, which has been less investigated in previous work. To fill this gap, we propose a novel target-specific abstract-guided news document representation model. The model uses a target-sensitive repre- sentation of the news abstract to weigh sentences in the news content, so as to select and combine the most informative sentences for market modeling. Results show that document represen- tations can give better performance for estimating cumulative abnormal returns of companies when compared to titles and abstracts. Our model is especially effective when it used to combine information from multiple document sources compared to the sentence-level baselines. 1 Introduction Texts from the Internet was shown to be statistically correlated with stock market trends (Antweiler and Frank, 2004). Natural language processing (NLP) techniques have been applied to extract information from company filings (Lee et al., 2014), financial news articles (Xie et al., 2013) and social media texts (Bollen et al., 2011) in order to gain understandings of financial markets. In particular, traditional meth- ods exploited statistical signals, such as lexical and syntactic features (Wang and Hua, 2014; Schumaker and Chen, 2009), sentiment (Bollen et al., 2011) and event structures (Ding et al., 2014; Ding et al., 2015; Ding et al., 2016) from these text sources, which suffers from feature sparsity problems. With the recent trends in deep learning for NLP, neural networks have also been leveraged to learn dense representations for text elements, which can address the sparsity of discrete features in statistical models. Such representations can implicitly represent sentiment, event and factual information, which can be extremely challenging for sparse indicator features. In particular, Ding et al. (2015) show that deep learning representations of event structures yield better accuracies for stock market prediction com- pared to discrete event features. Chang et al. (2016), Duan et al. (2018) use neural networks to directly learn representations of news abstracts, showing that it is effective for predicting the cumulative abnormal returns of public companies. One limitation of Ding et al. (2015) and Chang et al. (2016), however, is that these methods only model news titles and abstract texts, which are typically single sentences. Ding et al. (2015) show that a model that uses only news content gives inferior results compared to one …

Content truncated to 500 words