AN APPLIED STUDY ON RECURSIVE ESTIMATION METHOD, NEURAL NETWORKS AND FORECASTING

We compare three modelling approaches to univariate time series foreca sting, based on recursive estimation and supervised learning methods. The models considered range from relatively simple time-varying parame ter damped trend models to non-linear models based on radial basis fun ction 'networks' or on multi-layer perceptrons. The estimation methods considered are the Kalman filter procedure, the Recursive Least: Squa res algorithm and variants, and the Levenberg-Marquardt algorithm, whi ch we try to describe under a common framework. As our main goals, ive . discuss some of the main identification and estimation issues associ ated with those approaches, and illustrate their application through t he study of selected data front tile Lisbon stock exchange index. (C) 1997 Published by Elsevier B.V.