A neural-FIR predictor: Minimum size estimation based on nonlinearity analysis of input sequence

Abstract

In this paper, a hybrid model of multy-layer neural network combined with a finite-impulse-response filter is proposed for a nonlinear time series prediction. We introduce an important analysis of the input sequence to determine the effective minimum combination of the input samples and hidden neurons. Through computer simulations, using both sunspot and computer generated time series, the proposed analysis has shown its effectiveness and the proposed predictor has demonstrated its superiority. It is of a faster convergence and smaller residual error than the conventional nonlinear predictor.