Back propagation in time‐series forecasting

Abstract

One of the major constraints on the use of back propagation neural networks as a practical forecasting tool is the number of training patterns needed. We propose a methodology that reduces the data requirements. The general idea is to use the Box‐Jenkins model in an exploratory phase to identify the 'lag components' of the series, to determine a compact network structure with one input unit for each lag, and then apply the validation procedure. This process minimizes the size of the network and consequently the data required to train the network. The results obtained in eight studies show the potential of the new methodology as an alternative to the traditional time‐series models.