Crypto Volatility Forecasting: Mounting a HAR, Sentiment, and Machine Learning Horserace

Abstract The relationship between investor sentiment and cryptocurrency market volatility remains an area of growing interest in empirical finance. In this study, we present an innovative forecasting approach by utilizing a unique dataset of AI-generated sentiment from a comprehensive database of crypto market news. In a horserace fashion, we first evaluate the Heterogeneous Autoregressive (HAR) model and then compare its forecasting performance to five advanced machine learning (ML) methods. ML performs reasonably well and improves the accuracy of the benchmark HAR model. Interestingly, including sentiment does not improve the forecasting accuracy of the HAR model. However, our findings highlight that investor sentiment seems to influence crypto market volatility in a nonlinear fashion that can (only) be captured by ML methods. In other words, LightGBM, XGBoost, and LSTM models show enhanced predictive accuracy when sentiment data is incorporated, improving no-sentiment forecasts in 54.17% of the cases studied. Overall, our results emphasize the significant potential of integrating machine learning and sentiment analysis as a promising avenue for improved forecasting, offering potential benefits for risk management strategies and provide valuable insights for researchers and practitioners.