Prophet and GPT-2 Algorithms for Demand Forecasting in Last-Mile Delivery: A Comparative Analysis and Optimization

Abstract

Demand forecasting is a central issue in last-mile logistics, where operational performance depends on a delicate balance between prediction accuracy and interpretability of results. This study offers a rigorous comparison between two radically different modeling approaches: Prophet, a transparent additive model well-suited to decomposable time series, and GPT-2, a modified transformer capable of capturing complex and nonlinear demand dynamics. Using real-world delivery data, we conduct a thorough evaluation of both models based on several criteria: forecast accuracy (MAE/RMSE), robustness to demand volatility, computational efficiency, and sustainability impact. To ensure a fair comparison, hyperparameter optimization is systematically conducted using Optuna, ensuring optimal configurations for each model. The results reveal that Prophet excels in stable demand environments, thanks to its ease of interpretation and regularity, while GPT-2 demonstrates a marked adaptability to unpredictable variations, at the cost of higher computational costs. The article concludes with practical recommendations for logistics stakeholders, proposing tailored selection criteria (accuracy, explainability, resource constraints) based on the specificities of operational environments. This work aims to bring theoretical advances in predictive modeling closer to the realities of last-mile logistics, providing a concrete decision-making framework for implementing forecasting solutions based on artificial intelligence.