An applied study on recursive estimation methods, neural networks and forecasting

Abstract We compare three modelling approaches to univariate time series forecasting, based on recursive estimation and supervised learning methods. The models considered range from relatively simple time-varying parameter damped trend models to non-linear models based on radial basis function ‘networks’ or multi-layer perceptrons. The estimation methods considered are the Kalman filter procedure, the Recursive Least Squares algorithm and variants, and the Levenberg-Marquardt algorithm, which we try to describe under a common framework. As our main goals, we discuss some of the main identification and estimation issues associated with those approaches, and illustrate their application through the study of selected data from the Lisbon stock exchange index.