Popularity prediction and optimal location: Application to restaurants

Abstract

We consider the location problem for retail service facilities, consumer-facing storefronts that provide a service and compete with other retailers to some degree or the other. Location is one of the most important strategic decisions for a retail firm. It is a risky and often an irrevocable decision, in the sense that it involves a large investment, is very difficult to rectify, and affects profits and operations for many years in the future. This problem is especially challenging for the following reasons: (i) Location models require estimates of how demand will expand and shift when we locate a new facility, but the firm, since it has not yet started operations, has no historical demand data to calibrate the models; (ii) Future entry as well as exits of competitors affect the firm’s revenues and profitability, but predicting such future strategic developments is rather complicated. In this paper, we consider forward-looking competitive entry and exit decisions using a simple equilibrium framework, solvable by integer programming and estimable from public data. To capture the taste of local demographics, we build a model based on online reviews of the incumbent establishments where facilities have latent characteristics and customers have preference for these latent characteristics. This serves as an input to predict customer demand which drives our optimal location solution and gives firms an easy and tractable toolkit for their decision-making. We apply the model to a service industry, specifically the restaurant industry, to illustrate how it can be made operational. Our estimation results show that customers differ significantly in their willingness to travel and rating sensitivities across restaurant types. Apart from a tractable toolkit to help their decision process, we show, via counterfactuals, that optimized location decision-making can increase chances of survival by up to 37.5%. Managerial insight into the nature of competitive location dispersion is also provided.