Robust time series prediction by optimal algorithms theory

This paper presents a new approach to time series forecasting which can be considered alternative to classical statistical techniques when dealing with a limited number of observations or whenever statistical hypotheses reveal to be inadequate. The method leads to efficient forecasting techniques based on recent results of the theory of optimal algorithms. One of the most attractive features of the approach consists in avoiding the usual two step procedure of model fitting and predicticn with the fitted model. Two real classical and widely studied time series have been analysed and the obtained results compare favourably with those of advanced statistical techniques, especially with regard to multistep ahead predictions.