A neural network approach to forecasting model selection

The literature has shown that no one model provides the most accurate forecasts. The focus has instead shifted to identifying the characteristics of the time series in order to provide guidelines for choosing the most appropriate extrapolation model. In this paper we test the feasibility of employing the neural network structure for model selection. To accomplish this objective, a set of time series characteristics, representing the domain knowledge, is established. A back propagation neural network is then constructed with eleven input nodes representing six time series characteristics. The output nodes of the neural network represent nine time series forecasting methods grouped into three categories. The results indicate that the neural network approach can assist the practitioner in the selection of the appropriate forecast model.