Incorporating financial news for forecasting Bitcoin prices based on long short-term memory networks

Abstract

In this paper, we investigate how a deep learning machine learning model can be applied to improve Bitcoin price forecasting and trading by incorporating unstructured information from financial news. The two-stage model we propose that includes financial news significantly outperforms machine learning models without financial news. In the first stage, we leverage long short-term memory (LSTM) networks to extract structured information from financial news. In the second stage, we apply machine learning models with structured input from financial news to the prediction of Bitcoin prices. In addition to the superior performance relative to machine learning models without input from financial news, we find that the out-of-time rate of return attained with the proposed forecasting system is substantially higher than for a buy-and-hold strategy. Our study highlights how combining deep learning and financial news offers investors and traders support for the monetization of unstructured data in finance.

Keywords:

• Bitcoin price forecasting
• Sentiment analysis
• Deep learning
• Financial news
• Bitcoin trading

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed at https://doi.org/10.1080/14697688.2022.2130085.

Notes

1 Studies like Hirshleifer and Shumway (2003), Kamstra et al. (2003), Brown and Cliff (2004), Qiu and Welch (2004), Baker and Wurgler (2006), Baker and Wurgler (2007), Lemmon and Portniaguina (2006), Edmans et al. (2007), Tetlock (2007), Barber et al. (2009), Kaplanski and Levy (2010), Stambaugh et al. (2012), Huang et al. (2015), Sibley et al. (2016), and Bajo and Raimondo (2017) have investigated the effects of investor sentiment on different aspects of financial markets.