Modeling food-related business closure in select New York City communities using multi-scale and spatial features

Abstract

This paper introduces an extensible framework to predict small-business closures to inform urban planners, lenders, and business owners as to factors to improve business resilience. This paper couples machine learning with two point of interest (POI) datasets and infrastructure data and uses New York State’s COVID-19 PAUSE as a stressor for investigating small-business resiliency. The study included 2537 food-related, non-chain, retail businesses across select New York City zip codes, of which 17.7% closed permanently. Macro-, meso-, and micro-levels of features included the neighborhood profile, street dynamics, and venue-specific, location-related characteristics. A Gaussian Mixture Neural Network model achieved 74.1% precision, 92.5% recall, and an 82.3% F1-score without use of financial data. High-end restaurants located further than average from public transit were most at risk for closure, while non-restaurant, food businesses in commercially diverse areas having higher-than-average social media ratings were least at risk. This paper introduces a model for timely prediction of pandemic-induced, food-related, small-business closures without reliance on private or protected financial data, and provides insights into urban design to promote small, food business survivability.