Predicting Bitcoin Returns by Machine Learning

Abstract

We construct a comprehensive set of 37 factors and employ a battery of 12 machine learning techniques to predict Bitcoin returns. We investigate the importance of factors and their influence on model outputs by using advanced SHAP method. Our findings underscore the superior predictabilities of tree models, particularly random forest. However, neural network models markedly underperform tree models. The miner reserves and market value to realized value ratio emerge as the most important factors for both tree and neural network models, exerting a significantly positive impact on model output. Our study provides valuable insights for investors, traders, and financial analysts.

Keywords: Bitcoin, Return prediction, machine learning