A new approach for combining neural networks during training for time series modeling

A great deal of data in business, economics, engineering and the nature science occur in the form of time series. Linear models have been frequently used for time series modeling. Recently, neural networks (NN's) have been used as nonlinear models. Many researchers have performed a combination of several NN's in order to attain a better accuracy model. In this paper, we propose a new approach for combining NN's. We develop a number of combination models during the training of the NN's and then select the best performer among them on a validation set as the eventual combination model. Experimental results show the significant improvements of the proposed method compared to the apparent best NN, and combination of the trained NN's. Copyright (C) 2007 Praise Worthy Prize S.r.l - All rights reserved.