Forecasting Bitcoin returns with long short-term memory networks and wavelet decomposition: A comparison of several market determinants

Abstract

Investigating Bitcoin price forecasting has attracted academic attention recently. However, despite some studies on potential economic determinants of Bitcoin price, a consensus on the best predictors is not reached yet. This paper investigates different predictors from various markets including Gold, Oil, S&P500, VIX, USDI, Ether and Ripple as well as Bitcoin historical price in predicting one-step-ahead Bitcoin returns. We propose a two-stage forecasting that comprises discrete wavelet transform as the decomposition method and a deep long short-term memory network as the forecasting algorithm. Beside analyzing forecasting for both univariate and multivariate regression, we design a simulated trading system to put the forecasts into practice and analyze the economic profitability of the predictors. In addition, we shed light on the black box method by implementing sensitivity analysis. To investigate the predictors’ efficacy through time and consider the effects of early 2018 price spike, the dataset is split into two periods: (1) prior to and including the spike and (2) after the spike. According to the experiments, it is hard to choose one predictor over the other in the first period. However, in the second period, Gold and Oil show the highest statistical accuracy, while S&P500 is the most profit-making predictor.