A wavelet hybrid method based on ARIMA and support vector regression forecasts

摘要

In this paper, we put forward a new hybrid methodology to generate forecasts of time series. Indeed, the proposed forecaster is a HWCF that integrates the following techniques: wavelet decomposition; ARIMA models; SVRs; wavelet combination of forecasts; and non-linear programming. Basically, the HWCF is able to capture, simultaneously, linear and non-linear auto-dependence structures exhibited by a time series, which are represented, at time t, by both the linear and non-linear combined forecasts: L_(c,t) and N_(c,t), respectively. After obtaining the combined forecasts L_(c,t) andNc,t, they are summed (i.e., L_(c,t) + N_(c,t) = y_(h,t)), producing the hybrid forecast y_(h,t), for each instant t. The numerical results show that HWCF achieved relevant accuracy gains in forecasting process of the annual time series of sunspot, when comparing with other ten competitive forecasters.