A Machine Learning Approach to Delineating Neighborhoods from Geocoded Appraisal Data

Identification of neighborhoods is an important, financially-driven topic in real estate. It is known that the real estate industry uses ZIP (postal) codes and Census tracts as a source of land demarcation to categorize properties with respect to their price. These demarcated boundaries are static and are inflexible to the shift in the real estate market and fail to represent its dynamics, such as in the case of an up-and-coming residential project. Delineated neighborhoods are also used in socioeconomic and demographic analyses where statistics are computed at a neighborhood level. Current practices of delineating neighborhoods have mostly ignored the information that can be extracted from property appraisals. This paper demonstrates the potential of using only the distance between subjects and their comparable properties, identified in an appraisal, to delineate neighborhoods that are composed of properties with similar prices and features. Using spatial filters, we first identify regions with the most appraisal activity, and through the application of a spatial clustering algorithm, generate neighborhoods composed of properties sharing similar characteristics. Through an application of bootstrapped linear regression, we find that delineating neighborhoods using geolocation of subjects and comparable properties explains more variation in a property’s features, such as valuation, square footage, and price per square foot, than ZIP codes or Census tracts. We also discuss the ability of the neighborhoods to grow and shrink over the years, due to shifts in each housing submarket.