A comparative study of AR order selection methods

Autoregressive (AR) modelling has assumed an important role in many application areas of signal processing. When fitting an AR model to an observed data sequence the selection of the model order is of great importance. Many criteria for selecting the AR model order have been proposed. However, these are scattered throughout the literature and little analysis of their relative performance is available. The principle aim of this paper is to compare the performance of some of the more recent methods of AR model order selection so that a broad perspective of the subject may be obtained. We compare the order selection methods using both narrowband and wideband data, with and without additive noise. The results of the study allow some basic recommendations to be made about the best types of method to be used in each case. It is observed that methods using the observed data perform best in the case of narrowband signal with no noise. Methods which employ the autocorrelation (second-order cumulant) sequence of the data show promise when used with wideband or noisy signals.