Automatic ARMA identification using neural networks and the extended sample autocorrelation function: a reevaluation

Recently, several researchers have attempted to use neural network approaches in conjunction with the extended sample autocorrelation function (ESACF) to automatically identify ARMA models. The work to date appears promising, but generalizations are limited by the fact that the test and training sets for the neural networks were generated from random perturbations of prototype ESACF tables. This paper develops test and training sets by varying the parameters of actual ARMA processes. The results show that the ability of neural networks to accurately identify the order of an ARMA (p,q) model from its transformed ESACF is much lower than reported by previous researchers, and is especially low for time series with fewer than 100 observations.