Forecasting with Universal Approximators and a Learning Algorithm

Abstract

This paper applies three universal approximators for forecasting. They are the Artificial Neural Networks, the Kolmogorov-Gabor polynomials, as well as the Elliptic Basis Function Networks. We are particularly interested in the relative performance and stability of these. Even though forecast combination has a long history in econometrics focus has not been on proving loss bounds for the combination rules applied. We apply the Weighted Average Algorithm (WAA) of Kivinen & Warmuth (1999) for which such loss bounds exist. Specifically, one can bound the worst case performance of the WAA compared to the performance of the best single model in the set of models combined from. The use of universal approximators along with a combination scheme for which explicit loss bounds exist should give a solid theoretical foundation to the way the forecasts are performed. The practical performance will be investigated by considering various monthly postwar macroeconomic data sets for the G7 as well as the Scandinavian countries.