Forecasting nonstationary time series based on Hilbert-Huang transform and machine learning

Abstract

We propose a modification of the adaptive approach to time series forecasting. On the first stage, the original signal is decomposed with respect to a special empirical adaptive orthogonal basis, and the Hilbert's integral transform is applied. On the second stage, the resulting orthogonal functions and their instantaneous amplitudes are used as input variables for the machine learning unit that employs a hybrid genetic algorithm to train an artificial neural network and a regressive model based on support vector machines. The efficiency of the proposed approach is demonstrated on real data coming from Nord Pool Spot and Australian National Energy Market.