Neural networks and times series forecasting: a theoretical approach

This paper presents some results of the authors´ work about forecasting by neural networks. The authors try to have an approach as close as possible to the classical statistical one. In this way, the authors show how a multilayer perceptron could process as a classical model (ARMA), and they present an algorithm to eliminate unnecessary connections of a network. Then, after a brief presentation of their data and forecasting methodology, the authors give a typology of daily electric consumption curves which was done by a self-organising Kohonen map