Examine Bitcoin price predictability with machine learning algorithms

Abstract

As the most-traded digital asset, Bitcoin receives a tremendous increase in investment interests. The primary purpose of this paper is to examine and compare two commonly applied machine learning algorithms on their capability and feasibility in Bitcoin price prediction. The regression models of Extreme Gradient Boosting and Long Short-Term Memory are selected as the investigated objects in this paper. The experiments are evaluated by considering the extra impacts of sample dimensions and time-series frequency, along with the trade-offs between prediction accuracy (measured by residual error) and computational efficiency (measured by computing time). Long Short-Term Memory, as a more convoluted deep learning method, achieves better accuracy when a daily dataset with limited input features is used. However, its predictability has considerably decreased and thus become less efficient than XGBoost, when more peripheral information and higher frequency data points (trading price every 15 minutes) are available. Besides algorithmic complexity, Long Short-Term Memory also takes much longer computing time than Extreme Gradient Boosting, making it a less applicable model to use when dealing with large sample sizes.