Eliseu Celestino Schopf,  Paulo Martins Engel. 

This paper describes a new forecast technique over highly nonlinear noisy time series. The neural statistical method uses a multi-layer perceptron and the Extended Kalman Filter. The method combines the modeling of the noise of the Kalman Filter with the nonparametric modeling and treatment of high nonlinearities of the neural network. The filter is adjusted with internal data of the net, filtering the data so that the net can have better results. The method reaches satisfactory performance in experiments with complex time series with noise addition. The method presents a satisfactory filtering of series values. The feedback with filtering values increments the method's performance.

http://www.lbd.dcc.ufmg.br:8080/colecoes/waamd/2007/005.pdf

Processando consulta.