Diagnosing and Tuning a Statistical Forecasting System: A Case Study

Commercial forecasting systems are commonly used in manufacturing businesses to generate sales forecasts for thousands of products. These systems typically feature a number of built-in options to select and estimate the statistical forecasting model, from a pre-specified collection of models, and are targeted to provide a low cost and efficient method of forecasting. However, it is often observed that after prolonged use (e.g. two or more years), these systems suffer from serious performance degradation in terms of large forecast errors for a significant number of products. An immediate and tangible consequence of inaccurate forecasts is increase in inventory levels, rendering the entire supply chain less efficient. It is therefore important to diagnose and tune the statistical forecasting systems as part of their regular maintenance and operation.

This chapter describes a case study in which simple but useful tools were devised for the forecast practitioner to (i) diagnose a statistical forecasting system systematically and identify products that require forecast performance improvement; and (ii) tune the parameters of a statistical forecasting system to improve its overall forecast performance. The proposed tools were developed for a large manufacturer of consumer and industrial products in the USA.