Intelligent techniques for forecasting multiple time series in real-world systems

Abstract

Purpose – The purpose of this paper is to describe a real-world system developed for a large food distribution company which requires forecasting demand for thousands of products across multiple warehouses. The number of different time series that the system must model and predict is on the order of 105. The study details the system's forecasting algorithm which efficiently handles several difficult requirements including the prediction of multiple time series, the need for a continuously self-updating model, and the desire to automatically identify and analyze various time series characteristics such as seasonal spikes and unprecedented events. Design/methodology/approach – The forecasting algorithm makes use of a hybrid model consisting of both statistical and heuristic techniques to fulfill these requirements and to satisfy a variety of business constraints/rules related to over- and under-stocking. Findings – The robustness of the system has been proven by its heavy and sustained use since being adopted in November 2009 by a company that serves 91 percent of the combined populations of Australia and New Zealand. Originality/value – This paper provides a case study of a real-world system that employs a novel hybrid model to forecast multiple time series in a non-static environment. The value of the model lies in its ability to accurately capture and forecast a very large and constantly changing portfolio of time series efficiently and without human intervention.