Neural and fuzzy models for economic forecasting. An econometric view and some practical experience

In this paper the new modelling techniques (neural networks, fuzzy systems, hybrid models) and the classical methods of econometrics and time series analysis are presented in an integrated view, as complementary rather than competitive tools. The role of neural, fuzzy and hybrid systems in the general framework of economic modelling and forecasting is analysed, with especial emphasis on their connections with mainstream tools of econometrics and time series modelling. The advantages of this integrated view emerge both at the theoretical and the empirical level. At a purely theoretical level, benefits derive from the availability of the methods of statistical inference, which enable the researcher to rigorously analyse neural learning processes. The method of sieves is especially useful to study the non-parametric estimation capabilities of neural networks, fuzzy systems and other flexible modelling tools. Some sieve estimation results for several classes of neural–fuzzy models are presented. From the practical viewpoint, the use of previous treatments of data and various statistical diagnoses seems, analogously to the case of standard statistical models, a good means to obtain parsimonious high-performance models. Some practical experiences with the modelling of Spanish economic time series are reported.