A Bayesian approach to combined neural networks forecasting.
Maurits D. Out, Walter A. Kosters · 2000
. Suitable neural networks may act as experts for time series predictions. The naive prediction is in a Bayesian manner used as prior to steer the weighted combination of these experts. The paper was presented at ESANN'2000, The 8th European Symposium on Articial Neural Networks, in Brugge (Belgium), April 26-28, 2000. 1. Introduction Predicting the near future using information from the past is a challenging task. In this paper we try to do so by combining neural networks and techniques from Bayesian statistics (cf. [1] and [6]) in order to generate a prediction for an observable y t+1 , given the time series y 1 ; y 2 ; : : : ; y t . We start from the following observations. First it is hard to beat naive predictions|y t in the setting above. Secondly, it is a complicated matter to combine expert forecasts into better ones. We feel that a joined eort might generate better results, especially if neural networks support the process. For this purpose we use so-called FIR networ...