A hybrid ensemble framework for cost-optimized demand forecasting in fresh produce retail

The high perishability, demand volatility, and asymmetric inventory costs make forecasting retail demand for fresh produce challenging. This paper proposes a hybrid two-stage ensemble method of cost-optimized demand forecasting, evaluated on 46,599 daily transaction records from a Chinese supermarket. Phase 1 cross-validates eight models in a rolling-window cross-validation. Phase 2 combines Bayesian hyperparameter optimization, an asymmetric loss function that encodes the 20:1 cost ratio into the training gradient, and ensemble weight optimization to minimize business costs. In this single-retailer setting, the hybrid ensemble achieves a 46.3% reduction in per-window inventory cost relative to the Phase 1 XGBoost benchmark. The most important finding is that the accuracy-optimized stacking ensemble incurs 48.5% higher inventory costs than the hybrid, even with higher demand R 2 , indicating that the accuracy-optimal combination is not cost-optimal when the cost structure is asymmetric. The results confirm the predict-and-optimize design philosophy in a real-world perishable retail environment.