Measuring Gentrification: Using Data Science to Predict the Impacts of Gentrification

In this initial study, we chose to focus on the level of counties. We constructed a dataset of 630 US counties for which we had enough data to compute NGI values for each year between 2008 and 2021 and demonstrate their usefulness for real estate investors. We identified a statistical connection between the dynamics of NGI over time and the key performance metric of counties, measured by their residential property prices. We illustrated our findings on two use cases. While we chose residential property prices due to the data abundance, these findings may also be relevant for multifamily, which remains a dominant institutional asset class.

NICHE TRANSFORMATION

In recent years, multifamily investors have experienced a stronger bull market as compared to other asset classes.¹ However, as the rising tide has lifted (almost) all boats, a potentially receding tide necessitates a more selective investment strategy.

Just as asset allocation is a primary driver of portfolio returns in a mixed asset class portfolio (e.g., stocks, bonds, etc.),² selection decisions at the metropolitan statistical area (MSA) level or county level drive the bulk of returns within the real estate investment universe, with individual asset selection relegated to the second place in the return composition hierarchy. Our analysis indicates that—in the majority of cases—a residential property’s price can be predicted within a 20% margin solely based on its county, which validates the importance of county-level selection.

Real estate portfolio management can be enhanced by a proactive, “top-down” market selection approach highlighting locations with potential for above-market return profiles. “Bottom-up” individual investment-level strategies can be supplemented rather than replaced with these analyses. Knowledge of high-potential target counties within the investable universe can ultimately be an alpha enhancer in a competitive investment market.

But there are more than 3,000 counties in the US, so where shall we start?

Our goal is to go beyond trailing signals and momentum-driven strategies and to attempt to create forward-looking signals that may present above-average returns in future years. Specifically, can individual- or family-level wealth inflows predict residential real estate price patterns?

RETHINKING POPULATION DATA

Our primary model in this analysis can be characterized as the gentrification of a neighborhood, which is defined as the net wealth inflow relative to the wealth of the neighborhood. We propose an NGI of a neighborhood and use it to compare different neighborhoods.

The NGI is calculated as the difference between the estimated wealth excess of the in-migration and out-migration of the focal county. Counties with a high NGI have positive wealth inflows, meaning that when we estimate the wealth of in-migrants and out-migrants, there is a net increase. And vice versa for negative NGIs. We consider our NGI as a proxy for gentrification measure.

Obviously, NGI does not fully explain property price changes. In our dataset of 630 US counties, the NGI variation from 2010 to 2020 only weakly correlates with residential price change (r = 0.293), which has a marginally higher correlation compared to a simple demographic indicator such as a population increase estimate (r = 0.177). However, population changes are measured at a very low cadence (typically, once a decade) and cannot reliably be predictive. NGI, in contrast, can be measured quarterly. Moreover, we report a statistical connection between NGI and property price increase, which indicates a predictive power that NGI has on prices. While NGI is not yet fully studied, our first results look promising.

County-level housing supply can often be inelastic, which we assumed would increase prices as new residents bid up the existing real estate inventory. As expected, population growth is more correlated with residential real estate prices in high-barrier-to-entry markets that are characterized by low growth in new housing units.

However, as we analyzed the NGI of these counties based on the bifurcation between high- and low-barrier-to-entry counties, counterintuitively, we saw that the NGI had a stronger relationship with low-barrier-to-entry markets than high-barrier-to-entry markets (if we consider only low-barrier-to-entry counties, the correlation goes up from 0.293 to 0.456).

There are several potential explanations for this dynamic. Firstly, in supply-constrained markets, the utility of adding the wealth dimension to the net population growth may be reduced, because the primary effect will be on adding incremental demand, irrespective of wealth levels. Secondly, elastic supply markets may be lower-cost markets, which are more sensitive to incremental wealth inflows, because the economic effect of this net wealth addition might be more significant throughout the county. To illustrate these dynamics, we have analyzed two counties representing strictly divergent trends: Pinal County, AZ and Cape May County, NJ.

Our empirical study focuses on the statistical connection between NGI and single-family price dynamics, primarily because we have a large number of single-family transactions that ensure statistical signifi cance. We claim that our results are likely to be valid for the multifamily asset class as well. While there are two orders of magnitude fewer multifamily transactions than single-family transactions (at the per-county per-year level), multifamily prices and single family prices highly correlate (r = 0.9). Exhibit 1 illustrates of the similarity of median prices per square foot in the multifamily and single-family markets nationwide.

WHAT IS A NORMALIZED GENTRIFICATION INDEX?

Intuitively, the gentrification of a neighborhood depends on both inbound and outbound migration; namely, a neighborhood is gentrifying if higher-income people are moving into it and/or lower-income people are moving out. The notions of “richer” and “poorer,” for the sake of this article, are relative to the average wealth of the neighborhood’s residents.

Gentrification is cyclic: while richer people are moving in and becoming residents, the average level of wealth of the neighborhood residents rises until it becomes high enough so that people moving in are no longer richer than the average, and then the opposite process of “de-gentrification” starts until the average wealth goes down, for the gentrification to pick up again.

We measure gentrification by utilizing the notion of “wealth excess” of a household i with respect to the average wealth W_n of a household in neighborhood n. We define the wealth excess of household i in neighborhood n as WE_i,n = S(W_i–W_n), where W_i is an estimated wealth of i, and S is a smoothing function that levels down extreme differences between W_i and W_n. Smoothing is necessary to take care of edge cases and data issues as explained below. We highlight that WE_i,n might be negative if i is poorer than the average in n.

We use a propriety algorithm to estimate household wealth, which bases our estimation on home values, as homeownership is typically the primary source of wealth for a family. For the purposes of this analysis, if we could not obtain the value of a family’s home, we approximate it with an average home value in the same neighborhood.

As mentioned, the NGI of a neighborhood will be positive if many richer families are moving in, and many poorer families are moving out. It will be around zero if few families are moving in and out, or those moving families are of average wealth. The index will be negative if many poorer families are moving in, and many richer families are moving out. Note that if the smoothing function S was not applied, the NGI value might have been primarily influenced by very few wealth excess values WEi,n if they are truly outstanding. For example, if the wealth of a family happens to be estimated at three orders of magnitude compared to the average wealth of a neighborhood’s residents, then this family’s wealth excess would be the dominant factor in the neighborhood’s NGI, which is obviously an undesired outcome.

DATASET

NGI can be defined for “neighborhoods” at any level of granularity, such as MSAs, zip codes, and census tracts. In this work, we analyze NGIs at the coarse level of counties.

Out of 3,243 US counties and county equivalents, we select only the counties for which we have enough assessed property value data, transaction price data, and relocation data. We also make sure that assessed property values are not too low when averaged over all properties in a county (low average values per square foot might be due to a data artifact) and that property values and transaction prices would be in the same ballpark (very large difference between them would be caused by another data artifact). The selection process resulted in a dataset with 630 counties, covering the majority of the largest US counties that are also attractive destinations for investment capital. Out of the 398 largest US counties reported by the World Population Review,³ 334 counties are present in our dataset. Those that are not present are either located in non-disclosure states (so they lack transaction price data) or their data is only partially populated in public data sources. For example, assessed property value data of Cook County, IL (the second largest US county) is populated only from 2017. Exhibit 2 shows our dataset of 630 counties displayed on the map, where the color represents the size of the county in terms of the number of residential properties.

Our relocation dataset consists of 14.4 million relocation records spread over the 2000–21 time period. Each record contains the name of a person, their previous address, their new address, and the year of the relocation. Names of people are not used in this research. The relocation dataset was constructed from public data sources (property sale transactions and property tax records) using a proprietary record-matching algorithm. Each address is associated with the name of the county to which it belongs.

Our relocation dataset is only a subset of all US relocations. However, it does not show a selection bias: on the 630 counties, the Pearson correlation between the number of relocations into a county and the size of the county (measured in the number of residential properties) is high (r = 0.855). The correlation between the number of relocations out of a county and the size of the county is even higher (r = 0.94).