Wavelets in multi-step-ahead forecasting

This paper investigates the possibility of obtaining long-into-the-future reliableforecasts of observed nonlinear cyclical phenomena. Unsmoothed monthly sunspotnumbers that are characteristically cyclical with nonlinear dynamics as well as theirwavelet-transformed and wavelet-denoised series are forecasted through October 2008.The objective is to determine whether modelling wavelet-conversions of a series providesreasonable forecasts. Two computational techniques – neural networks and geneticprogramming – are used to model the dynamics of the series. Statistical comparison oftheir ex post forecasts is then used to identify the data set and computational technique touse under the circumstances. Copyright © 2005 IFAC