The association between financial news sentiment and cryptocurrency price movements

Abstract

Investors, market traders and financial institutions are primarily interested in making financial decisions that result in the highest monetary gains. The majority use several methods and indicators, to attempt at predicting the trend or price of the specific stocks markets they are investing in. One popular approach involves utilising financial sentiment analysis to gain insights on future market trends. Financial Sentiment Analysis is a particularly challenging task because of the domain specific financial terminology used within financial text, and the lack of manually labelled financial data. In literature, we find two primary approaches to accomplish this task, mainly lexicon based and deep learning based. Lexicon based approaches use sentiment lexicons to annotate sentiment labels in text, and do not require any pre-training. On the other hand, deep learning models are trained on manually labelled financial text to learn how to label sentiment. This leaves a question as to which approach performs best when it comes to financial sentiment analysis. Apart from the technique itself, it is still unclear from the literature as to which source of online data will contribute most, to provide a concise indication of sentiment. Typically, the two online sources of data utilised are social media platforms such as Twitter, Reddit or Telegram, or news platforms such as Reuters and the New York Times. This research investigates the relationship between the sentiment of financial news with cryptocurrency price movements. Both lexicon based and deep learning methods are evaluated, whilst also experimenting with various types of data sources. This is done to determine which combination of techniques and data will have the most significant effect on the prices. We introduced an enhanced deep learning model, called PBLREC, which is based on the existing FinBERT model. The task of this PBLREC model is to label the sentiment of financial text into three classes. These being positive, negative and neutral. The performance of this model was evaluated against several configurations that utilised different data sources as well as other techniques from published research such as the original FinBERT model, VADER lexicon and an extended version of the VADER lexicon we called VADERMCD, that contains financial terminology. In our experiments, we show the superiority of our model, achieving an average F1-score of 92.28%. Financial News was then classified using both the PBLREC model, and the standard VADER lexicon separately. Price direction prediction utilising both the sentiment annotated text and price data was then conducted, using a Logistic Regression model for cryptocurrencies and traditional stocks. We show that text annotated by the PBLREC model provided the best indication of sentiment, for the task of price direction prediction, by achieving an overall accuracy of 56%.