Abstract

Demand analysis is one of the cornerstones of any supply chain management system, and most of the critical operational decisions in the supply chain rely on accurate demand predictions. Although a large body of academic literature proposes various forecasting methods, there are still important challenges when using them in practice. The common problem is that firms need to decide about thousands of products, and the demand patterns could be very different between them. In this setting, frequently, there is no single forecasting method that works well for all products. While some autoregressive models might work well in some cases, the demand for other products might require an ad-hoc identification of trend and seasonality components. In this chapter, we present a methodology based on meta-learning that automatically analyzes several features of the demand to identify the most suitable method to forecast the demand for each product. We apply the methodology to a large retailer in Latin America and show how the methodology can be successfully applied to thousands of products. Our analysis indicates that this approach significantly improves the firm's previous practices, leading to important efficiency gains in the supply chain.