Time Series Modeling and Forecasting With Feedforward Neural Networks: A Comparative Study Using the Resex Data

It is considered in this paper the modeling and forecasting, through feedforward neural networks, of a time series previouslystudied in the literature (Brubacher, 1974; Martin, 1980; Stahlbut, 1985; Allende, 1989) called RESEX series, which presentsseasonality and outliers. Some elements of the network architecture such as the input variables definition are suggestedby a previous time series analysis and other elements such as the number of intermediate layers and corresponding knots are defined through numerical methods. Not only traditional backpropagation based algorithms are used but also a robust learning algorithm is considered in order to deal more properly with the outliers. All models and methods (traditional and NN based) are then compared considering different predictive performance measures and theresults are discussed.