Sequential network construction for time series prediction

This paper introduces an application of the sequential network construction (SNC) method to select the size of several popular neural network predictor architectures for various benchmark training sets. The specific architectures considered are a FIR network and the partially recurrent Elman network and its extension, with context units also added for the output layer. We consider an enhancement of a FIR network in which only those weights having relevant time delays are utilized. Bias-variance trade-off in relation to the prediction risk estimation by means of nonlinear cross-validation (NCV) is discussed. The presented approach is applied to the Wolfer sunspot number data and a Mackey-Glass chaotic time series. Results show that the best predictions for the Wolfer data are computed using a FIR neural network while for Mackey-Glass data an Elman network yields superior results.