Predicting Bitcoin price discontinuities from realized metrics and Twitter sentiment via machine learning

We consider a comprehensive framework to predict the probability of Bitcoin price discontinuities (jumps). To this end, we employ six classifiers namely, XGBoost, AdaBoost, Random Forest, Logistic Regression, Support Vector Machines (SVM), and Artificial Neural Networks (ANN). We integrate realized measures based on high-frequency price data, along with unique sentiment indices derived from Twitter data as inputs into the machine learning models. Sentiment indices are constructed using VADER and TextBlob analysis and each is combined with realized variance, realized skewness, realized kurtosis, and realized semi-variances. Our results reveal that XGBoost shows strong predictive performance with lower missclassification cost, while Random Forest and ANNs are less efficient. Additionally, TextBlob-based sentiment exhibits improved results relative to the models with VADER. These findings support the effectiveness of the proposed jump detection framework for predicting Bitcoin price discontinuities. • Predicts Bitcoin price jumps using realized volatility metrics. • Incorporates Twitter sentiment into forecasting models. • Applies machine learning to detect price discontinuities. • Improves jump prediction accuracy over benchmark models. • Combines market microstructure and social media signals.