Cristiano Leite de Castro,  Antônio P. Braga,  Alessandro Vivas Andrade. 

On forecasting of time series, better results have been obtained from combination of different forecasting models than from the choice of that single specific model with the best individual performance. This paper describes the use of the Ensemble method for the attainment of a forecasting model with high capacity of generalization. In this method, the outputs of different forecasting models are combined linearly to produce a global output. The Ensemble method was applied for the combination of Multi-Layer Perceptron nets that supply distinct solutions for the forecasting of a nonstationary series. Each net was trained differently by means of the manipulation of training parameters.

http://www.lbd.dcc.ufmg.br/colecoes/enia/2005/055.pdf

Processando consulta.