Dimensioning of the error of neural network models applied to the forecast of time series

Artificial neural networks are an important technique in nonlinear time series forecasting. However, training of neural networks is a difficult task, because of the presence of many local optimal points and the irregularity of the error surface. In this context, it is very easy to obtain under-fitted or over-fitted forecasting models without forecasting power. Thus, researchers and practitioner need to have criteria for detecting this class of problems. In this paper, we demonstrate that the use of well known methodologies in linear time series forecasting, such as the Box-Jenkins methodology or exponential smoothing models, are valuable tools for detecting bad specified neural network models.