Assessing integration orders for SARIMA modeling A hypothesis testing approach with information criterion hyperparameter selection, case of predicting gas consumption in central Tunisia

Forecasting natural gas demand is critical for enhancing energy efficiency, optimizing infrastructure planning, and supporting Tunisia's transition toward sustainable energy. As the fastest-growing fossil fuel and one of the cleanest among non-renewable energy sources, natural gas plays a pivotal role in balancing energy security and environmental objectives. However, accurate demand forecasting remains challenging. While SARIMA models are valuable for this, a critical limitation in existing studies is the insufficient attention to integration orders determination (d, D) in SARIMA models-a step often overlooked in conventional time series analysis and automated tools. This study focuses on Tunisia's central region, leveraging hourly gas consumption data from ten stations between January 2015 and August 2023. The data is downsampled into weekly, monthly, quarterly and yearly frequencies to develop mid- and long-term forecasts. We rigorously compare two methodologies for identifying integration orders: (1) traditional Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests, and (2) a hybrid approach combining OCSB and Canova-Hansen tests with ADF/KPSS validation. Each method is evaluated through SARIMA models optimized using AutoARIMA and assessed via performance metrics, including MAPE, sMAPE, MASE, RMSSE, and a novel Overall Weighted Average (OWA) score. Our findings demonstrate that expert-driven selection of integration orders significantly outperforms automated methods, particularly for low-frequency data, while intercept-no-trend SARIMA configurations yield more robust differencing strategies. For instance, our best-performing model achieves a MAPE of 12 %, underscoring the practical value of methodological precision. The hybrid methodology, despite some data structure sensitivities, showed computational efficiency advantages. These insights not only refine gas demand forecasting but also offer actionable guidance for energy policymakers and operators in Tunisia and similar regions, enabling more efficient resource allocation and infrastructure development in the context of evolving energy landscapes.