Being the connection weights set the rprop algorithm

66 Cortez, Rocha & Neves

Figure 8. The best fitness value (BIC) obtained in each generation by the evolutionary

vector machines, nonlinear techniques that present theoretical advantages (e.g., ab-sence of local minima) over neural networks. For instance, Cao and Tay (2003) have applied support vector machines to financial series, obtaining competitive results when compared with back-propagation networks.

Another potential non-conventional approach to forecasting relies in the use of evolu-tionary computation, which is expected to increase in importance, motivated by advan-tages such as explicit model representation and adaptive global search. Currently, there are two promising approaches: first, the optimization of traditional forecasting methods, such as the ARMA model (Equation 4), with evolutionary algorithms based on real value representations (Cortez, Rocha, & Neves, 2004); second, the use of genetic program-ming, by building a numerical expression made of a set of time lags and operators, such as the ones taken from the alphabet {‘+’,‘-’,‘*’ and ‘/’} (Kaboudan, 2003).

Comparative experiments, among conventional (e.g., Holt-Winters & Box-Jenkins) and connectionist approaches, with several real and artificial series from different domains, were held. These have shown that the Holt-Winters method, although very simple,

Copyright © 2006, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of Idea Group Inc. is prohibited.