Performance of Autoregressive Order Selection Criteria: A Simulation Study

Proper selection of autoregressive order plays a crucial role in econometrics modeling cycles and testing procedures. This paper compares the performance of various autoregressive order selection criteria in selecting the true order. This simulation study shows that Schwarz information criterion (SIC) , final prediction error (FPE) , Hannan-Qiunn criterion (HQC) and Bayesian information criterion (BIC) have considerable high performance in selecting the true autoregressive order, even if the sample size is small, whereas Akaike’s information criterion (AIC) over-estimated the true order with a probability of more than two-thirds. Further, this simulation study also shows that the probability of these criteria (except AIC) in correctly estimating the true order approaches one as sample size grows. Generally, these findings show that the most commonly used AIC might yield misleading policy conclusions due to its unsatisfactory performance. We note here that out of a class of commonly used criteria, BIC performs the best for a small sample size of 25 observations