Volatility Prediction using Financial Disclosures Sentiments with Word Embedding-based IR Models

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1712–1721 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1157 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1712–1721 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1157 Volatility Prediction using Financial Disclosures Sentiments with Word Embedding-based IR Models Navid Rekabsaz1, Mihai Lupu1, Artem Baklanov2, Allan Hanbury1, Alexander D¨ur3, Linda Anderson1 1 3TU WIEN 2International Institute for Applied Systems Analysis (IIASA) 2N.N. Krasovskii Institute of Mathematics and Mechanics 1{family name}@ifs.tuwien.ac.at 2baklanov@iiasa.ac.at 3alexander.duer@tuwien.ac.at Abstract Volatility prediction—an essential concept in financial markets—has recently been addressed using sentiment analysis meth- ods. We investigate the sentiment of an- nual disclosures of companies in stock markets to forecast volatility. We specif- ically explore the use of recent Informa- tion Retrieval (IR) term weighting mod- els that are effectively extended by related terms using word embeddings. In paral- lel to textual information, factual market data have been widely used as the main- stream approach to forecast market risk. We therefore study different fusion meth- ods to combine text and market data re- sources. Our word embedding-based ap- proach significantly outperforms state-of- the-art methods. In addition, we investi- gate the characteristics of the reports of the companies in different financial sectors. 1 Introduction Financial volatility is an essential indicator of in- stability and risk of a company, sector or econ- omy. Volatility forecasting has gained consider- able attention during the last three decades. In addition to using historic stock prices, new meth- ods in this domain use sentiment analysis to ex- ploit various text resources, such as financial re- ports (Kogan et al., 2009; Wang et al., 2013; Tsai and Wang, 2014; Nopp and Hanbury, 2015), news (Kazemian et al., 2014; Ding et al., 2015), message boards (Nguyen and Shirai, 2015), and earning calls (Wang and Hua, 2014). An interesting resource of textual information are the companies’ annual disclosures, known as 10-K filing reports. They contain comprehensive information about the companies’ business as well as risk factors. Specifically, section Item 1A - Risk Factors of the reports contains information about the most significant risks for the company. These reports are however long, redundant, and written in a style that makes them complex to process. Dyer et al. (2016) notes that: “10-K reports are getting more redundant and complex [...] (it) re- quires a reader to have 21.6 years of formal ed- ucation to fully comprehend”. Dyer et al. also analyse the topics discussed in the reports and ob- serve a constant increase over the years in both the length of the documents as well as the number of topics. They claim that the increase in length is not the result of economic factors but is due to ver- boseness and redundancy in the reports. They sug- gest that only the risk factors topic appears to be useful and informative to investors. Their analysis motivates us to study the effectiveness …

Content truncated to 500 words