Addressing Demand Forecasting Challenges in the U.S. Retail Supply Chains: An AI-Driven Machine Learning Evaluation of Historical, Operational, and Macroeconomic Information

Abstract

U.S. retail supply chains face persistent demand-forecasting challenges because product demand varies across items, stores, and time, complicating inventory and replenishment planning. Improving forecast accuracy while determining which available data sources provide meaningful predictive value is therefore important for data-driven retail supply chain management. This study addresses this challenge by applying machine learning to evaluate the incremental predictive value of historical demand, retail and operational information, and macroeconomic indicators for large-scale U.S. retail demand forecasting. Using the M5 dataset, three LightGBM specifications were evaluated under a common modeling framework across 25,940 item-store series and five rolling 28-day forecast windows. Model A incorporated lagged and rolling historical-demand features; Model B additionally incorporated calendar, SNAP, event, and selling-price information; and Model C further incorporated two-month-lagged consumer price index, unemployment, and retail sales indicators. Forecast performance was compared with a 28-day seasonal-naive benchmark using MAE, RMSE, and WAPE, supplemented by paired statistical and bootstrap robustness analyses. Mean WAPE decreased from 0.9006 for the seasonal-naive benchmark to 0.7022 for Model A, representing an improvement of approximately 22.03%. Model B further reduced mean WAPE to 0.6979, an additional improvement of approximately 0.61%, while Model C produced only a negligible further reduction to 0.6978, approximately 0.02%. Paired and series-average analyses showed a large and robust improvement for Model A, a small positive but heterogeneous average improvement for Model B, and negligible aggregate incremental forecasting value for Model C, whose bootstrap confidence interval for the mean MAE improvement included zero. These findings demonstrate that the history-based LightGBM specification generated the dominant forecasting improvement, while retail and operational information provided a modest incremental benefit and the examined lagged macroeconomic indicators contributed little additional aggregate predictive value. For U.S. retail supply chain forecasting, the findings demonstrate how machine-learning-based evaluation of progressively richer information sets can identify which data sources provide meaningful forecasting gains, supporting evidence-based model development for inventory and replenishment planning.