Forecast-driven discount optimization in FMCG distribution: an ensemble learning review

Abstract

Trade discounts are among the fastest-acting profit levers available to a distributor of fast-moving consumer goods (FMCG). A distributor typically receives a limited discount budget from the manufacturer and must decide how to spread it across the matrix of stock-keeping units (SKUs) and customers so that company profit grows without sacrificing sales volume or the customer base. The dominant analytical architecture for such decisions is the two-stage predict-then-optimize pipeline, in which a demand model estimated from historical data feeds a constrained profit-maximization program. This paper reviews the international and regional literature on both stages of that pipeline. On the forecasting side, we examine the empirical evidence – led by the M5 competition and a series of recent journal studies – that tree-based ensemble methods (XGBoost, LightGBM, CatBoost, random forests, and stacked combinations of these) now define the state of the art for SKU-level retail demand prediction. On the optimization side, we analyze field-validated and model-based systems for price, markdown, and promotion optimization, from the Rue La La deployment to nonlinear integer programming for supermarket promotions and prescriptive price optimization via binary quadratic programming. The review documents a consistent 3-10% profit or revenue uplift band reported in peer-reviewed applications and identifies a clear gap: no published study combines ensemble demand forecasting with constrained discount allocation at the SKU-customer level, and none does so for Central Asian FMCG distribution.