Model Identification of Integrated ARMA Processes.

Area of Science:

• Time Series Analysis
• Statistical Modeling
• Econometrics

Background:

• Autoregressive Integrated Moving-Average (ARIMA) models are crucial for time series forecasting.
• Automated model selection methods aim to simplify ARIMA identification.
• Existing methods like SCAN and ESACF are available in statistical software (e.g., SAS).

Purpose of the Study:

• To evaluate the performance of Smallest Canonical Correlation Method (SCAN) and Extended Sample Autocorrelation Function (ESACF) for identifying integrated ARIMA models.
• To compare the effectiveness of SCAN and ESACF on both untransformed and differenced time series data.
• To develop an improved strategy for ARIMA model selection by integrating various techniques.

Main Methods:

• The study applied SCAN and ESACF to identify integrated ARIMA models.
• Performance was assessed by comparing automated identifications with known model structures.
• A novel model selection strategy was proposed, combining different identification approaches.

Main Results:

• SCAN achieved 79% correct identifications, while ESACF reached 80% accuracy.
• Both methods demonstrated disappointing accuracy for certain ARIMA models and parameterizations.
• Inconsistent model identifications were observed for both automated tools and human experts.

Conclusions:

• Automated ARIMA model selection tools (SCAN, ESACF) exhibit limitations in accuracy and consistency.
• Human expertise alone does not guarantee reliable ARIMA model identification.
• An integrated strategy combining multiple identification techniques is recommended for robust ARIMA model selection.