Estimation of Heterogeneous and Nonstationary Retail Demand with Aggregate Data

Expected demand is an important measure for a retailer to decide whether to open a new store at a candidate location. Retail chains need to understand the store choice and temporal preferences of different customer segments to estimate future customer visits to such a new store. As mixed-use zoning becomes more common, retailers are faced with increasingly heterogeneous and nonstationary demand, making the demand estimation problem more complex. We develop a method to make such estimation while relying only on aggregate data. Given the geographic and demographic information of the trading areas of existing stores, we propose a new parametric segmentation-temporality (ST) model that learns both store choice and temporal preferences of different customer segments from just historical aggregate data. We investigate the data requirements for the local identifiability of our ST model and compare the ST model with three misspecified models, including models that consider segmentation only, temporality only, and neither segmentation nor temporality. We develop an iterative two-step algorithm to obtain segment-wise parameter estimates of the ST model. Based on the geographic and demographic information in Beijing, we illustrate the feasibility of using limited customer visit data and publicly available information of the trading areas to obtain reliable estimates of the expected customer visits to candidate stores. The numerical experiments show that it is necessary to consider both customer segmentation and demand temporality to obtain reliable estimates of the expected customer visits. We find that assuming stationarity causes the estimates of the expected aggregate customer visits on different days to be biased, but does not affect the estimates of the expected daily customer visits over a cycle of each segment. We also show that when ignoring customer segmentation, even the estimates of the expected daily aggregate customer visits can be highly biased