Boosting Store Sales Through Ensemble Learning-Informed Promotional Decisions

Many real-world analytics problems, such as forecasting sales of fashion products, involve uncertain and heterogeneous demand, requiring prescriptive analytics to incorporate multiple covariates and address the inherent challenge of model uncertainty. Traditional predict-thenoptimize (PTO) approaches typically rely on a single predictive model, overlooking model uncertainty. To address this, we propose an ensemble learning framework that integrates Mallows-type model averaging into the PTO paradigm, leveraging diverse candidate models with varying covariates to enhance forecast accuracy and decision robustness. Theoretically, we prove that the weighted forecasts achieve asymptotic optimality under mild conditions and establish finite-sample risk bounds, ensuring stable performance even in limited-data settings. We empirically evaluate the proposed framework using weekly store-level sales data from an internationally recognized footwear brand in China. The forecasting exercise demonstrates that our approach consistently achieves the lowest prediction risk, improving forecast accuracy by 4.72% to 7.41% compared to the best-performing alternatives without weighted forecast features. In the subsequent decision optimization exercise, we identify gift, combo, and discount promotions as key decision variables and show that our framework delivers the highest predicted sales responses on average, outperforming alternative forecasting methods and existing data-driven decision frameworks.