Enhancing forecast accuracy in the paper scrap sector: a comparative study of ensemble and machine learning models
摘要
Despite the extensive use of econometric and machine learning models in trade forecasting, empirical evidence on hybrid ensemble methods in the context of paper scrap supply chains remains limited, particularly at a cross-country level. To address this gap, we employ a combination of statistical and machine learning techniques, including Holt-Winters method, ARIMA, Long Short-Term Memory (LSTM) networks, XGBoost, and a dynamic ensemble methodology. The proposed approach contributes by empirically validating a dynamically weighted ensemble framework within an underexplored recycling-oriented trade domain, rather than by introducing a novel forecasting algorithm. The global paper industry is undergoing substantial transformation, driven by evolving consumer demands—particularly those associated with e-commerce and environmental sustainability. This study examines the dynamics of paper scrap supply chains, emphasizing key trends in recycling, international trade, and the growing influence of e-commerce packaging on the sector. Over the past decade, global trade in paper products has declined at an average annual rate of 0.4%, largely due to reduced demand for printed materials. In contrast, the expansion of e-commerce has significantly increased the demand for packaging materials, reshaping paper supply chain structures. We evaluate the out-of-sample forecasting performance of ARIMA, LSTM, and a four-component Ensemble (Holt–Winters, ARIMA, LSTM, and XGBoost) across ten countries using MSE-based accuracy metrics (RMSE, MAE, and MAPE). Relative to ARIMA, the Ensemble reduces average RMSE by 6.6%, MAE by 8.3%, and MAPE by 6.4%. Relative to LSTM, the improvements are substantially larger (RMSE 31.1%, MAE 31.9%, MAPE 26.9%), reflecting that LSTM, despite hyperparameter tuning, exhibits higher variance on several series. Furthermore, the analysis identifies critical data points required to monitor the evolution of circular supply chains within the paper industry, underscoring the importance of robust data collection and reporting frameworks. As the industry adapts to shifting market conditions, the integration of sustainable practices into paper scrap supply chains will be vital for ensuring resilience and long-term growth. In conclusion, the interplay between e-commerce, recycling initiatives, and international trade is reshaping the paper scrap supply chain landscape, presenting both challenges and opportunities for stakeholders across the industry. This study offers empirical insights into the behavior of adaptive ensemble weighting mechanisms across heterogeneous market regimes, with implications for policymakers and industry stakeholders engaged in circular supply chains.