Predicting Bitcoin Prices via Machine Learning and Time Series Models

Abstract In this study, we predict Bitcoin price trends using the back propagation neural network (BPNN), autoregressive integrated moving average (ARIMA), and generalized autoregressive conditional heteroscedasticity (GARCH) models. Based on principal component analysis (PCA), we extract two new input components for BPNN from Bitcoin’s three-day closing prices, MA5, MA20, daily trading volume, Ether price, and Ripple price. The training set covers the period between September 1, 2015 and March 31, 2020, and the forecasting set covers the period between April 1, 2020 and June 30, 2020. Empirical results reveal (1) the predictive ability of BPNN over that of the ARIMA models; (2) BPNN with two hidden layers is able to predict price trends more precisely than that with only one hidden layer; (3) in terms of time series models, the ARIMA-GARCH family of models demonstrates better predictive performance than ARIMA models; and (4) among the ARIMAGARCH family of models, the ARIMA-EGARCH model is proven to produce the best predictive results on price, and the ARIMA-GARCH model predicts more accurately than the ARIMA-GJR-GARCH model. Specifically, our findings provide a reference on Bitcoin for market participants. JEL classification numbers: C32, C45, C53, G17. Keywords: Bitcoin, Back propagation neural network, Autoregressive integrated moving average, Generalized autoregressive conditional heteroscedasticity, Principal component analysis.