Nonlinear time series analysis by neural networks: an example

This paper presents an application of feedforward neural networks to nonlinear time-series analysis of a univariate system. The appropriate network size is determined by studying statistical criteria for the fits on the training set and on an independent test set. By a proper choice of network size, the problems of model over-parametrization can be avoided. The reason for the worse performance of the over-sized networks is analyzed and discussed.