Forecasting Residential Displacement in Buffalo, NY

Introduction: Considerable effort has been put into developing ways to detect gentrification in its early stages and mitigate its negative effects. These systems are generally called Early Warning Systems (EWS). Most of the EWSs for gentrification developed in the U.S. have used traditional, widely-available data sources– in particular, yearly population estimates from the U.S. census at the tract or neighborhood level– to forecast gentrification. However, census tract-level analyses can potentially miss the piecemeal nature of gentrification. For example, it is possible that two homes being purchased above asking price on the same block may represent a critical early warning sign of gentrification. There is, consequentially, a need to explore the fine-grained spatiotemporal resolution of gentrification, and to do so with data that is available to city governments. We engage in a case study, using a new dataset and modern machine learning and statistical methods, that considers the extent to which it is possible to forecast one precursor to gentrification – residential displacement – in Buffalo, NY.

Method: We collected a dataset, somewhat unique in the academic literature but widely available to cities, which incorporates all purchases and tax information of homes in Buffalo, beginning in 1995 through June 2020. The dataset contains 134,312 transactions on all 51,425 homes in the city as of June, 2020. Using this dataset, we address two questions. First, can we actually predict displacement at finer-grained spatio-temporal areas than what census data provides? Second: to the extent we can forecast displacement at smaller scales, what are the factors that are most predictive, and do those factors vary over time or between neighborhoods? We address these questions with a series of predictive modeling experiments in which we vary spatial and temporal granularity in a variety of ways.

Results: We find that predictions at the single-home level on a short temporal scale are extremely difficult; models predict at chance levels whether or not a single home will sell in the next year. However, prediction accuracy rapidly increases as either spatial or temporal resolution increases. For example, we can predict whether or not more than three homes will sell in the same year on a single city block with an accuracy of around 70%, a relative increase of about 9% over baseline. We can also predict whether or not a single home will sell in the next five years with approximately the same accuracy. With respect to the second question, we find the top predictors of purchases for individual homes is information about nearby homes, another signal for the potential utility of studying displacement (and thus gentrification) at sub-census track levels.

Implication: Our preliminary result is that future residential displacement, and in turn a signal of gentrification, can be reliably predicted at spatial and temporal resolutions that are on a smaller scale than current EWSs consider, with data widely available to city governments. We argue potential intervention strategies that leverage these methods with policy may help decrease the effects of displacement on residents.