Isla Délice, a leading French halal food manufacturer, faces significant demand-planning challenges across a portfolio of 424 active SKUs, with diverse demand patterns ranging from high-volume, stable products to low-volume, intermittent items. The current forecasting approach lacks the granularity and statistical rigor needed to minimize forecast error, leading to suboptimal inventory levels, service failures, and excess safety-stock costs. This capstone project develops a segmentation-driven demand forecasting framework. SKUs are first clustered by demand behavior using K-means on Average Demand Interval (ADI) and squared Coefficient of Variation (CV²). For each cluster, the best-performing forecasting model is selected from a candidate set that includes exponential smoothing methods, AutoSARIMA, the Syntetos-Boylan Approximation (SBA), Prophet, and LightGBM. Results from cluster-level time-series modeling indicate strong separability into four demand pattern types (Smooth, Intermittent, Erratic, Lumpy), each with a distinct best-performing forecasting model. The framework produces transparent, cluster-based recommendations that planners can interpret and apply alongside their existing ABC-XYZ view, with cluster-winning models leading to meaningful WMAPE reductions in the Intermittent and Smooth clusters (which together cover roughly two-thirds of active SKUs) relative to a uniform forecasting strategy.
