Forecast Accuracy Versus Waste Outcomes: Evaluating AI Based Demand Forecasting in Perishable Food Retail

Food retailers increasingly adopt artificial intelligence (AI) and advanced analytics to improve forecasting and inventory management in perishable categories, often evaluating models mainly through standard accuracy metrics such as Mean Absolute Percentage Error (MAPE). This article argues that such symmetric metrics are insufficient when the managerial objective includes food waste reduction. Using daily sales data from three grocery stores and ten perishable product families in the Corporación Favorita dataset, the study compares a seasonal naïve benchmark with a Random Forest forecasting model. The models are evaluated with traditional accuracy measures, a safety‑stock‑based waste proxy, and asymmetric error metrics that emphasise over‑forecasting. The results show that the seasonal naïve model achieves lower MAPE, yet the Random Forest model generates lower residual variability, lower safety stock, and a waste proxy approximately 20.7% below the benchmark, due to smaller over‑forecast magnitudes. Additional analyses indicate that promotions and holidays systematically increase over‑forecast errors for both models, but the Random Forest remains more robust in these high‑risk periods. The findings suggest that AI‑based forecasting can support waste reduction even when it does not improve conventional accuracy metrics, and that retailers should integrate waste‑sensitive indicators into the evaluation and selection of forecasting systems.