Pii: S0925-2312(01)00648-8

This paper deals with nonlinear time series prediction. The proposed method combines the wavelet decomposition (as a *ltering step) and neural networks to provide an acceptable prediction value. Basically, the wavelet decomposition uses a pair of *lters to decompose iteratively the original time series. It results in a hierarchy of new time series that are easier to model and predict. These *lters must satisfy some constraints such as causality, information lossless, etc. We prove here that our method reduces the empirical risk comparatively to the classical ones. As an illustration, the results obtained on both sunspot and MacKey–Glass time series are shown. c © 2002 Elsevier Science B.V. All rights reserved.