Applying case-based reasoning to forecasting retail sales

Abstract This article describes the development and testing of a forecasting system for retailers who plan periodic promotions. When the number of variables describing the promotion is large relative to the historical database of past promotions, traditional forecasting approaches cannot be applied. In such cases, retailers must rely on the expertise of their buyers to subjectively estimate promotional unit sales. This research develops a Case-Based Reasoning system that allows all buyers to forecast promotional sales as accurately as the organization's expert buyer and, by making the subjective process explicit, also provides an avenue to improve forecast performance over time. The system (1) selects the historical analogs that are most similar to the planned promotion, (2) adjusts the sales of each analog to account for any differences between the analog and the planned promotion, and (3) combines the forecasts derived from the multiple analogs to arrive at a single sales projection. The performance of the system was tested and found to compare favorably to the performance of an expert buyer in a large national retail organization.