Pump it: Twitter Sentiment Analysis for Cryptocurrency Price Prediction

Predicting the prices of the cryptocurrencies can be done by only using historical data related to the price, but adding other sources of information can be beneficial. In this work, we propose to analyse the market sentiment and add that information to the models. This sentiment was analyzed across 567 thousand tweets about 12 coins to get a daily grasp of the sentiment, polarity and subjectivity of the market. The tokens were separated into classes: established, emerging and "meme" tokens. We trained various algorithms, such as OLS, LOGIT, LSTM and NHITS. Two periods were analysed: one corresponding to a bear market and one to a bull market. Due to the high intra-day volatility of cryptocurrencies, LSTM that takes longer periods into consideration did not seem to perform better than the ones without "memory", like OLS and LOGIT. NHITS was the best performing model accuracy wise, but lacked in returns, which we associated with our over-simplistic trading strategy. The information extracted from social media proved to be helpful across the range of models and coins. We successfully showed that "meme" tokens do not represent a viable investing strategy in our study. The forecasting error does not increase significantly from a bear market to a bull market, even though the market changes drastically.