Comparison of automated procedures for ARMA model identification.

Area of Science:

• Statistics
• Econometrics
• Time Series Analysis

Background:

• Accurate time series model identification is crucial for reliable forecasting.
• Automated procedures aim to simplify the complex process of Autoregressive Moving Average (ARMA) model selection.
• Existing software like SAS offers multiple automated tools for ARMA model identification.

Purpose of the Study:

• To evaluate and compare the performance of three automated ARMA model identification procedures: MINIC, SCAN, and Extended Sequential Correlation Function (ESACF).
• To assess the effectiveness of these methods across various model structures, parameter values, and sample sizes for stationary, nonseasonal time series.

Main Methods:

• Monte Carlo simulation experiments were employed to generate time series data with diverse characteristics.
• The performance of MINIC, SCAN, and ESACF was assessed based on correct identification rates and selection of parsimonious representations.
• Analyses considered different dependency levels (low, medium, high) and sample sizes.

Main Results:

• On average, all three procedures correctly identified or selected equivalent ARMA structures in at least 60% of trials.
• MINIC excelled for purely autoregressive models, SCAN for mixed structures, and ESACF for moving-average processes.
• Model identification accuracy decreased with lower dependency processes, and larger sample sizes influenced SCAN and ESACF towards higher-order structures.

Conclusions:

• The evaluated automated ARMA identification procedures (MINIC, SCAN, ESACF) demonstrate reasonable effectiveness for stationary, nonseasonal time series.
• Method performance varies by model structure, with SCAN and ESACF showing sensitivity to sample size and dependency levels.
• These findings offer practical guidance for selecting appropriate automated tools in statistical software for time series analysis.