A novel decomposition-forecasting approach towards Bitcoin price prediction: Hybrid of Theta-SVR

Abstract

The Theta method has attracted academic attention lately due to its simplicity and superior performance. This paper proposes a new hybrid forecasting approach based on combining the Theta decomposition method and support vector regression (SVR) for forecasting highly volatile and noisy Bitcoin price time series. Using Theta decomposition with coefficients ranging from 0 to 2 with 0.1 steps, we extracted 20 Theta lines from the original time series. Each of these 20 lines is used for a univariate regression. Then the results of each forecasts aggregated to construct the final predicted values. Moreover, we used the Theta lines to construct a predictor space for multivariate regression using SVR. However, due to poor performance of the multivariate regression and to further enhance its performance, we eliminated inefficient Theta lines from the predictor space. Enhanced MASE by 10.45% and 5.68% in comparison to the Theta-SES (classic Theta) and SVR, the results indicate the superiority of the proposed hybrid Theta-SVR.