Cluster-based hierarchical demand forecasting for perishable goods

Abstract

Demand forecasting is of particular importance for retailers in the context of supply chains of perishable goods and fresh food. Such goods are daily produced and delivered as they need to be provided as fresh as possible and quickly deteriorate. Demand underestimation and overestimation negatively affect the revenues of the retailer. Stock-outs have an undesired impact on consumers while unsold items need to be discarded at the end of the day. We propose a DSS that supports day-to-operations by providing hierarchical forecasts at different organizational levels based on most recent point-of-sales data. It identifies article clusters that are used to extend the hierarchy based on intra-day sales pattern. We apply multivariate ARIMA models to forecast the daily demand to support operational decisions. We evaluate the approach with point-of-sales data of an industrialized bakery chain and show that it is possible to increase the availability while limiting the loss at the same time. The cluster analysis reveals that substitutable items have similar intra-day sales pattern which makes it reasonable to forecast the demand at an aggregated level. The accuracy of top-down forecasts is comparable to direct forecasts which allows reducing the computational costs.