Combining FIR filters and artificial neural networks to model stochastic processes

Three FIRNN structures (neutral networks involving FIR filters) are proposed to predict time series. Based on the structure of a feed-forward neutral network with one hidden layer, the structures use algorithms similar to the back-propagation algorithm. Performance comparisons are performed with some other methods of prediction, such as the autocorrelation method, the covariance method, the Durbin algorithm and the MYWE and LSMYWE methods. Some tests of the quality of prediction are performed. From these comparisons, it was concluded that these new structures give a small error whatever the type of data at the expense of increased computation time.