Cryptocurrencies have become an important dual entity (asset and currency) of the global financial system. Investing in cryptocurrencies as well as using them as a payment method is exponentially expanding. Prediction of cryptocurrency prices is a challenging task mainly because they represent a relatively new phenomenon, exhibiting high volatility. To better understand cryptocurrency trends and to improve price predictability, we propose to use sentiments of cryptocurrency-related news and micro-blogs. Using bitcoin-related tweets and Bitcoin prices between 2016 and 2020, we propose a transfer learning methodology [1] for Bitcoin price prediction based on sentiment analysis of finance micro-blogs. We use deep learning-based natural language processing (NLP) transformers like RoBERTa [2], which currently outperforms most of the methods used in NLP related tasks. We fine-tune the pre-trained RoBERTa model by using a labeled dataset of general financial tweets. Next, we leverage the fine-tuned model to evaluate the sentiment of the bitcoin-related tweets, which we use as an input dataset. As output, we use the “softmax” function in order to obtain probabilities for the positive or negative sentiment of the observed sample of tweets. Afterward, we compose two temporal numeric streams for the positiveness and negativeness of the news related to crypto assets. We use both streams as input for recurrent and convolutional
networks used for Bitcoin price prediction. We align temporally the sentiment and price vectors of Bitcoin with related tweets’ time of publishing. We evaluate different model architectures in order to identify the model that best fits the empirical data. The evaluation of the models is performed by using F1 and Matthews Correlation Coefficient (MCC) scores. The initial results show that the RoBERTa model shows accuracy of F1=0.908 and MCC=0.765 for the Bitcoin sentiment classification task. Additionally, we find that the Pearson correlation between the volume of published tweets and the price of the Bitcoin is 0.68. The future direction of this research is to evaluate networks of tweets related to different cryptocurrencies to study their comovements and establish communities of cryptoassets.
References:
[1] Sebastian Ruder, Matthew E Peters, Swabha Swayamdipta, and Thomas Wolf. Transfer learning in natural language processing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Tutorials, pages 15–18, 2019.
[2] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.