New Regularization Improves Retail Demand Forecast Stability Without Losing Accuracy

Retail demand forecasts are crucial for various operational planning cycles, including replenishment, capacity, labor, and transportation. However, traditional point-error objectives in forecasting models often fail to constrain abrupt, undesirable movements between consecutive forecasts, leading to operational inefficiencies. While post-hoc smoothing can address this, it only acts after model fitting and can degrade point accuracy. This research investigates whether a training-time penalty on within-series movement can improve forecast-path stability without materially affecting point accuracy. The proposed penalty is integrated into a temporal-structured pipeline that combines recent-demand embeddings with various features like calendar, price, hierarchy, item, and store data. The model was evaluated on selected M5 demand series at scales of 1000, 3000, and 4000 series. Results show that this stability-aware hybrid model significantly improves the Forecast Stability Score over XGBoost by 6.91%, 6.66%, and 7.68% across the different scales. Crucially, the Root Mean Squared Error (RMSE) changes remained minimal, within 0.72% across three random seeds, demonstrating that stability gains were achieved without a significant accuracy trade-off. In comparison, post-hoc exponential smoothing achieved lower raw movement but incurred a larger RMSE cost. These findings advocate for extending forecast evaluation beyond mere point-error minimization to include an accuracy-stability trade-off perspective, particularly for operational retail forecasting.