This study was designed: (a) to investigate a simple neural-network solution to forecasting the special class of time series corresponding to a wide range of ARMA(p,q) structures; (b) to study the significance of matching the input window size with the nature of time series. The study adopted a simulation approach in conjunction with an experimental design. It is discovered that a simple two-layered network, with proper input window size, is able to consistently outperform the multi-layer feedforward network and that the two-layered network is comparable to the Box–Jenkins modelling approach for a majority of the ARMA(p,q) time series studied and better than the Box–Jenkins modelling approach when the ARMA structure gets more complex and generates more variability. The results affirm that it is unnecessary to use multi-layer feedforward networks for this special class of linear time series and that the two-layered network can be a useful forecasting alternative to the widely popular Box–Jenkins model.
