A data-driven and context-aware approach for demand forecasting in the beverage industry

Accurate demand forecasting is essential for logistics and supply chain management as it enables efficient inventory planning, reduces operational costs, and ensures high service levels across the network. However, in practice, diverse demand patterns of items make this task challenging, and a one-size-fits-all forecasting approach is inadequate. This paper proposes a data-driven and context-aware forecasting framework and tests it by using both endogenous data from a large private-label beverage manufacturer and exogenous features (such as holidays and temperature). Our method begins by classifying SKUs based on demand volume, volatility, and intermittency, and then refining the derived clusters by taking volume distribution into account. Totally, we obtain four distinct clusters, which are (i) stable and high volume, (ii) stable with low volume, (iii) erratic and intermittent, and (iv) lumpy. To explore the appropriate forecasting models for different demand patterns, we employ statistical models (exponential smoothing, ARIMA, and Croston), machine learning models (XGBoost), deep learning models (TiDE and N-BEATS), and even qualitative approaches such as collaborative planning, forecasting, and replenishment (CPFR). Our experimental results suggest which forecasting models are recommended for each demand pattern, and insightful implications are provided for the managers.