Forecasting Digital Asset Return: An Application of Machine Learning Model

Abstract

In this study, we aim to identify the machine learning model that can overcome the limitations of traditional statistical modelling techniques in forecasting Bitcoin prices. Also, we outline the necessary conditions that make the model suitable. We draw on a multivariate large data set of Bitcoin prices and its market microstructure variables and apply three machine learning models, namely double deep Q-learning, XGBoost and ARFIMA-GARCH. The findings show that the double deep Q-learning model outperforms the others in terms of returns and Sortino ratio and is capable of one-step-ahead sign forecast of the returns even on synthetic data. These critical insights in forecasting literature will support practitioners and regulators to identify an economically viable cryptocurrency forecasting return model.