Reassessing Statistical and Hybrid Forecasting Models for Flexible Packaging Demand under Post-COVID Structural Changes

Abstract

Accurate demand forecasting is essential for production planning, inventory control, and supply chain management in the flexible packaging industry, particularly under structural changes following the COVID-19 pandemic. This study evaluates statistical, machine learning, and hybrid forecasting models for monthly demand prediction in the Indonesian flexible packaging market using data from January 2012 to December 2024. The evaluated models include ARIMA, SARIMA, SARIMAX with selected exogenous variables and a COVID-19 intervention term, standalone Multi-Layer Perceptron (MLP) and Support Vector Machine (SVM), and hybrid SARIMAX–MLP and SARIMAX–SVM models. To ensure a fair comparison, all models are assessed using RMSE, MAE, MAPE, and R² under a unified rolling-origin framework over the January 2023–December 2024 out-of-sample period, resulting in 24 one-step-ahead forecasts for each model. The empirical results show that ARIMA(1,1,3) achieves the lowest RMSE and highest R², indicating better performance in reducing larger forecast deviations, while SARIMA(0,1,1)(0,1,0$)_{12}$ achieves the lowest MAE and MAPE, indicating superior average absolute and percentage forecasting accuracy. In contrast, SARIMAX, standalone machine learning models, and hybrid models do not provide consistent improvements over simpler statistical benchmarks. The inclusion of exogenous variables and residual-based machine learning components does not improve forecasting performance, suggesting that the remaining external and nonlinear error structures are not sufficiently stable or learnable within the available monthly sample. The Harvey–Leybourne–Newbold corrected Diebold–Mariano test further indicates that the observed differences in forecast accuracy are not statistically significant at conventional levels. These findings show that increased model complexity does not necessarily lead to better forecasting performance. For practical industrial forecasting, simpler and more interpretable statistical models may be preferable when data are limited and short-term one-step-ahead forecasting accuracy is the main objective.