The Impact of Ensemble Stacking with Machine Learning and Statistical Models on Improving Sales Forecasting Accuracy

Abstract

Sales forecasting in retail remains a challenge, as changing customer behavior and seasonal patterns undermine predictive accuracy. For example, during promotional periods, managers tend to respond to changes in the weather or abrupt swings in preferences, not anticipating. For this purpose, we propose an ensemble stacking framework for machine learning and statistical models to improve forecasts in dynamic demand. We used a meta-learner of RidgeCV, trained on past retail sales data, to ensemble base learners including Random Forest, Extra Trees, XGBoost, and Multiple Linear Regression. Tree-based models captured non-linear interactions, but linear regression preserved interpretability of trends. The meta-learner learned to assign adaptive weights based on recent error patterns in time windows. We assessed performance using RMSE and MAE with particular focus on high volatility periods (e.g., holiday spikes and post-promotion troughs). The stacking model achieved an RMSE of 0.7689 and an MAE of 0.3066, outperforming individual base models consistently. Stacked predictions also showed less week-to-week variance and faster recovery after sudden shifts in demand. These results show that stacking can capture heterogeneous demand dynamics better than single models and, hence, provide a more robust structure. In practice, it allows for more confident safety-stock decisions and timely replenishment adjustments. Retailers can run a number of models in parallel and continuously rebalance them with new data as it comes in. Future work may include online updates of meta-learners and external covariates such as local events or competitor pricing.