Less is More: Bitcoin Volatility Forecast Using Feature Selection and Deep Learning Models

Abstract

Utilizing a large set of variables that include transaction information, public attention, blockchain information, macroeconomic variables and technical indicators, we compare different deep learning models with baseline methods, such as statistical and machine learning models, on Bitcoin volatility forecast. We find that feature selection approach strongly affects model performance. The results show that a simple Long Short-Term Memory (LSTM) model outperforms other models when using individual feature selection method.