Artificial neural networks are an important technique in nonlinear time series forecasting. However, training ofneural networks is a difficult task, because of the presence of many local optimal points and the irregularity ofthe error surface. In this context, it is very easy to obtain under-fitted or over-fitted forecasting models withoutforecasting power. Thus, researchers and practitioner need to have criteria for detecting this class of problems. Inthis paper, we demonstrate that the use of well known methodologies in linear time series forecasting, such as theBox-Jenkins methodology or exponential smoothing models, are valuable tools for detecting bad specified neuralnetwork models.