Store sales evaluation and prediction using spatial panel data models of sales components

Abstract

This paper sets out a general framework for store sales evaluation and prediction. The sales of a retail chain with multiple stores are first decomposed into five components, and then each component is explained by store, competitor and consumer characteristics using random effects models for components observable at the store level and spatial error random effects models for components observable at the zip code level. We use spatial panel data over four years for estimation and a subsequent year for evaluating one-year-ahead predictions. Set against a benchmark model that explains total sales directly, the prediction error of our framework is reduced by 34% for existing stores during the sample period, by 5% for existing stores one year ahead and by 26% for new stores.