Nonlinear prediction of time series using radial wavelet networks

A nonlinear prediction method based on a multiresolution approximation with nonorthogonal wavelets is presented. The method relies on a progressive learning algorithm which retains the essential features of the author´s data series in a coarse-to-fine approach. It builds up a network of radial wavelet units which grows as the new features at the finer scales are learned. It can also be seen as a cascade of whitening filters in the phase space domain. Experimental results for a nonlinear time series from the real world are presented