Systematic framework for ARIMA and SARIMA model selection using multiple error metrics: An application to export forecasting

This study proposes a systematic and practical framework for selecting autoregressive integrated moving average (ARIMA) and seasonal autoregressive integrated moving average (SARIMA) forecasting models based on multiple error evaluation metrics, including root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE). In practical forecasting applications, different error metrics often produce conflicting model selection outcomes, resulting in uncertainty in decision-making. Accordingly, this study introduces a percentage-based consistency evaluation approach to identify the most reliable forecasting model across multiple datasets. This study proposes a systematic framework for ARIMA and SARIMA model selection that integrates multiple forecasting error metrics to address conflicting model rankings and improve the consistency of model-selection decisions. This framework is applied to Thailand’s export value data, consisting of 69 monthly observations across 12 datasets derived from major industrial product groups (HS-84, HS-85, and HS-87) and key trading partners, including China, Japan, and South Korea. Results indicate that ARIMA(1,2,1) achieves the highest level of consistency across error metrics, with 83.33% under RMSE and 75% under MAPE and MAE. In terms of seasonal models, SARIMA (2,2,2)(2,0,0)₁₂ demonstrates the best overall performance. The findings confirm that the choice of error metrics can significantly influence forecasting outcomes. The proposed consistency-based framework reduces model selection conflicts and enhances the reliability of decision-making. This approach provides practical guidance for forecasting applications in economic planning, industrial systems, and export management.