Stock Embeddings Acquired from News Articles and Price History, and an Application to Portfolio Optimization

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3353–3363 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3353 Stock Embeddings Acquired from News Articles and Price History, and an Application to Portfolio Optimization Xin Du Department of Advanced Interdisciplinary Studies, Graduate School of Engineering, The University of Tokyo duxin@cl.rcast.u-tokyo.ac.jp Kumiko Tanaka-Ishii Research Center of Advanced Science and Technology, The University of Tokyo kumiko@cl.rcast.u-tokyo.ac.jp Abstract Previous works that integrated news articles to better process stock prices used a variety of neural networks to predict price movements. The textual and price information were both encoded in the neural network, and it is there- fore difficult to apply this approach in situa- tions other than the original framework of the notoriously hard problem of price prediction. In contrast, this paper presents a method to en- code the influence of news articles through a vector representation of stocks called a stock embedding. The stock embedding is acquired with a deep learning framework using both news articles and price history. Because the embedding takes the operational form of a vec- tor, it is applicable to other financial problems besides price prediction. As one example ap- plication, we show the results of portfolio op- timization using Reuters & Bloomberg head- lines, producing a capital gain 2.8 times larger than that obtained with a baseline method us- ing only stock price data. This suggests that the proposed stock embedding can leverage textual financial semantics to solve financial prediction problems. 1 Introduction News articles influence the dynamics of financial markets. For example, after the release of breaking news, the share prices of related stocks are often observed to move. This suggests the possibility of using natural language processing (NLP) to aid traders by analyzing this influence between news article texts and prices. Recent studies (Ding et al., 2015; Hu et al., 2018; Chen et al., 2019; Yang et al., 2018) have indeed reported that news articles can be leveraged to im- prove the accuracy of predicting stock price move- ments. These previous works have used deep learn- ing techniques. They train neural networks with article texts and financial market prices, attempting to improve price prediction. In these approaches, the overall mutual effect between texts and prices is distributed over the neural network, which makes it difficult to extract this effect and apply it to tasks other than price prediction. Therefore, we take a new approach by explicitly describing this mutual effect in terms of a vector. A stock is represented by a vector so that its inner product with an embedding of a text produces a larger value when the text is more related to the stock. In the rest of the paper, we call this vector a stock embedding. The names of stocks, such as “AAPL” (the ticker symbol for Apple Inc.), typically appear in a fi- nancial news article text. Because these names form part of the text, usual NLP techniques can be applied to acquire an embedding of …

Content truncated to 500 words