Abstract
This project introduces a segment-based forecast strategy to enhance tactical demand planning for a leading water bottling company. SKUs were segmented based on key demand pattern characteristics — average demand, coefficient of variation, and intermittency — resulting in four distinct clusters that reflect varying levels of demand predictability. A variety of forecasting models, including ARIMA, Exponential Smoothing, XGBoost, TiDE, Croston, and N-BEATS, were evaluated within each cluster using performance metrics such as MAPE, APE, MAE, and RMSE. The analysis also incorporates exogenous variables, such as natural events and holiday periods, to evaluate their impact on forecast accuracy. Testing results showed that implementing a Best-Fit Model strategy improved forecast accuracy by 1.56 percentage points in Absolute Percentage Error (APE). The results indicate that statistical time-series and machine learning models perform well in more stable and high-volume SKU clusters. In contrast, highly intermittent and low-volume SKUs remain challenging to forecast; thus, a more collaborative planning approach such as CPFR is recommended. This segmentation-driven approach improves forecast accuracy, enhances planning efficiency, and provides a scalable and interpretable framework for demand forecasting across the product portfolio.
