A Comparative Study of Forecasting Techniques for Reducing Food Waste in Retail Operations

Abstract

Food waste represents a critical challenge in retail supply chains, with approximately 30-40% of the food supply being discarded annually, resulting in substantial economic and environmental consequences. This study presents a systematic evaluation of demand forecasting techniques designed to mitigate food waste in retail operations. We compare statistical methods including ARIMA and SARIMA, machine learning approaches such as Random Forest, XGBoost, and LightGBM, alongside deep learning architectures including LSTM and Bidirectional LSTM networks. The evaluation framework encompasses multiple dimensions: forecasting accuracy across different product categories, computational efficiency, scalability considerations, and potential waste reduction impacts. Results demonstrate that machine learning techniques achieve superior performance in capturing complex demand patterns, with XGBoost and LightGBM delivering optimal accuracy-complexity tradeoffs. Deep learning models exhibit particular strength in handling long-term dependencies and seasonal variations. The findings provide actionable guidance for retail practitioners seeking to implement data-driven forecasting systems for waste reduction initiatives.