Sentiment-Based Bitcoin Volatility Prediction: A Comparative Study of Linear and Non-Linear Models

Abstract

. The paper explores the forecasting ability of social media aggregate sentiment measures, with historic variance, of Bitcoin return volatility. The researchers use a data set consisting of more than 16 million 2014-2019 tweets and the results of linear Ordinary Least Squares (OLS) benchmark predictions are compared with those of nonlinear ensemble methods (Random Forest and XGBoost). As a result of empirical findings, it is clear that although the XGBoost method has the lowest Mean Absolute Error, as it strongly represents the median trend, the linear OLS model has the lowest Mean Squared Error as a result of avoiding colossal over-predictions. On the other hand, random forest model shows a high level of overfitting. Those results indicate that sentiment and structural volatility clustering are valuable predictors, although the selection and regularization of the model are essential to identifying the nature of operations within cryptocurrencies markets without being noise.