Abstract
In this paper, a hybrid scheme for time series prediction is developed based on wavelet decomposition combined with Bayesian Least Squares Support Vector Machine regression. As a filtering step, using the Maximal Overlap Discrete Wavelet Transform, the original time series is mapped on a scale-by-scale basis yielding an outcome set of new time series with simpler temporal dynamic structures. Next, a scale-by-scale Bayesian Least Squares Support Vector Machine predictor is provided. Individual scale predictions are subsequently recombined to yield an overall forecast. The relevance of the suggested procedure is shown on the NINO3 SST anomaly index via a comparison with the existing methods for modeling and prediction.