Intermittent demand forecasting for long tail SKUs

Abstract

This paper provides systems and methods for the demand planner to improve forecasting intermittent long-tail demand by leveraging cluster-based processing. The proposed framework has been established on the demand data for a global power-generation business with reliable forecast number accuracy. The exploratory analysis encompasses both demand profiling and product classification stages, and the forecasting system identifies a cluster from historical demand data. Clustering aims to partition n products into k clusters, in which each product belongs to the cluster with the nearest product attribute. Demand of products within each cluster are aggregated, and the Unobserved Components time series Model (UCM) has been used to forecast at cluster level. Cluster-level forecasts are then disaggregated into child products based on the ratio of recent consumption.