Bitcoin price prediction using sentiment analysis

Abstract

Aim: To explore the causal relationship between social media sentiment, related behavioral factors, and bitcoin price
performance, and to develop predictive models with higher accuracy in forecasting the bitcoin price by incorporating
sentiment analysis.
Methods: We used the Valence Aware Dictionary and sEntiment Reasoner module to perform a sentiment analysis
of 896,464 Twitter posts (tweets) published between November and December 2021, which we collected via web
scraping. We created several forecasting models using the average daily sentiment polarity, the average daily number
of tweets, and search interest for “bitcoin” on Google and Wikipedia as input variables. We predicted future bitcoin
prices using vector autoregression (VAR), Prophet, and long short-term memory (LSTM) artificial neural network models
and evaluated their predictive accuracy using the mean absolute percentage error (MAPE) as a performance measure.
Results: The results suggest a Granger causal relationship between social media sentiment and bitcoin prices. The
standard VAR model achieved a MAPE of 8%, while the LSTM model had a lower error rate of 5%. The Prophet model
had a MAPE of 11%.
Conclusion: Our results underline the highly speculative nature of bitcoin, especially in times of high prices. The inclusion
of behavioral variables in the development of bitcoin price prediction models significantly improved their prediction
accuracy, with the LSTM neural network model proving to be an extremely effective tool in this sense.