Predicting Neighborhood Change Using Publicly Available Data and Machine Learning

Abstract

Gentrifying and declining neighborhoods affect nearly all major cities in the United States. Although gentrification tends to increase the economic value of a neighborhood, its subsequent impact on demographics and affordability often draws criticism from the local population. Additionally, decline in living conditions create socioeconomic stress across affected neighborhoods. Current rules-based approaches for identifying these neighborhoods are backwards-looking and do not allow local policymakers to get ahead of neighborhood decline and gentrification to mitigate their effects. This paper proposes a methodology for identifying neighborhood change in near real time by using machine learning. We define four neighborhood change types – gentrifying, declining, inclusively growing, and unchanging – and leverage publicly available US Census American Community Service (ACS) data, Zillow home value and rent data, and US Department of Housing and Urban Development (HUD) Housing Choice Voucher (HCV) data to predict which of these categories a census tract is likely to be a part of in the coming year. We train individual models across eight different metropolitan core based statistical areas (CBSA). The average performance of our models was 74% accuracy and 74% precision, an improvement over the rules-based baseline of 61% accuracy and 61% precision. These results suggest a promising application of the data to enable community intervention to produce more inclusive urban development strategies. The approach outlined in this paper can also be expanded to other metropolitan areas, to determine whether these results hold for those areas. The methodology enables supplementing data from the study with more Metropolitan area focused data, when available, to improve the precision of the models. This can provide reliable results for the policymakers looking to mitigate adverse consequences associated with neighborhood changes.