Abstract

An Application of Neural Networks to Spatial Retail Data
Track: Retail
Authors: Serene Ong, Xin Zhao

Publicly available retail data provide a detailed portrait of county and city-level retail activity, but have two major drawbacks: they have a time-delay, and due to the choice of geography by the provider, large and small retail areas are represented with the same level of detail. We propose an approximation procedure to overcome these weaknesses with a three-step procedure: Employing OLS regression analysis and Multi-Layer Perceptron Neural Networks, we estimate determinants of retail sales at county level with data available at county- and zip-code levels. We then use these estimates to project retail sales at zip-code level. Finally, we determine the dependence of each retail area on concurrently available data to estimate current levels of retail sales. We validate the procedure by using geographically higher aggregate data and find that relatively simple Multi-Layer Perceptron Neural Network Estimators outperform linear regression models due to non-linear components of spatial dependencies.