Enhancing Outbound Pallet Forecasting in Warehousing: A Comparative Analysis of VARX, ARIMAX, and Process-Level MLP Models

This study compares forecasting models for outbound pallet demand in warehouse operations: VARX-Based Chain Ratio (VCR), ARIMAX, and Multi-Layer Perceptron (MLP). Two MLP variants are evaluated: MLP1 integrates multi-stage operational data (receiving, replenishment, picking), while MLP2 uses only historical goods issue process (GIP) data. MLP1 significantly outperforms VCR and ARIMAX, achieving the highest R² (above 0.88) and lowest MAE and RMSE. In contrast, MLP2 performs notably worse, highlighting the value of upstream process-level features. While traditional models offer simplicity and transparency, they struggle to capture the nonlinear dynamics of pallet flows. As warehouse operations increasingly prioritize automation and data-driven responsiveness, accurate forecasting becomes critical. This study provides empirical and methodological contributions by validating the benefits of detailed, process-aware features, offering insights to enhance planning and resource allocation in complex logistics environments.