Raw Material Demand Forecasting Using Hybrid Prophet-XGBoost to Reduce Food Waste

Abstract

Inaccurate raw material inventory management in the commercial food and beverage sector frequently causes overstock, escalating costs and exacerbating the global food waste crisis. To address this inefficiency, this research develops a highly accurate predictive framework estimating daily demand for perishable raw materials. Utilizing historical point-of-sale transaction data from Matra Coffee between July and December 2025, the study proposes a serial hybrid machine learning architecture integrating Facebook Prophet and Extreme Gradient Boosting algorithms. Prophet is deployed as a base learner to model long-term trends and robust seasonal periodicities, while XGBoost operates as a residual meta-learner capturing short-term non-linear fluctuations and calendar anomalies. Empirical evaluation demonstrates exceptional predictive fidelity, yielding Mean Absolute Percentage Error values consistently below twenty percent across all primary perishable commodities, peaking at maximum accuracy for vegetables. The architecture is operationalized via a web-based graphical user interface featuring purchase simulation and financial waste calculation modules. Simulations indicate implementing this data-driven procurement strategy can reduce raw material overstock by twenty percent, mitigating food waste and bolstering supply chain sustainability.