Comparison of automated procedures for ARMA model identification

Abstract

This article evaluates the performance of three automated proceduresfor ARMA modelidentification commonly available in current versions of SAS for Windows: MINIC, SCAN, and ESACF. Monte Carlo experiments with different model structures, parameter values, and sample sizes were used to compare the methods. On average, the procedures either correctly identified the simulated structures or selected parsimonious nearly equivalent mathematical representations in at least 60% of the trials conducted. For autoregressive models, MINIC achieved the best results. SCAN was superior to the other two procedures for mixed structures. For moving-average processes, ESACF obtained the most correct selections. For all three methods, model identification was less accurate for low dependency than for medium or high dependency processes. The effect of sample size was more pronounced for MINIC than for SCAN and ESACE SCAN and ESACF tended to select higher-order mixed structures in larger samples. These findings are confined to stationary nonseasonal time series.