Find Comparable Neighborhoods to Any Census Tract

Comparable-tract analysis answers a question that standard market searches cannot: which census tracts, regardless of geography, are statistically most similar to a given subject tract on the dimensions that drive housing market behavior? The Verus-AI comparables engine constructs an 8-feature vector for each tract in the curated launch-metro tracts, then computes cosine similarity of weighted, robust-scaled feature vectors across that population. The eight features are: Verus score, historical CAGR percentage, appreciation percentage, log median income, education ratio, log current value, population year-over-year percentage, and income year-over-year percentage.

This choice matters: it reduces the leverage that extreme outliers would otherwise exert on the distance calculation, so a tract with an unusually high absolute home value does not automatically cluster with other high-value tracts unless the rest of its feature profile also aligns. The raw cosine output, which ranges from -1 to 1, is then rescaled onto a 0-to-1 interval for readability.

Critically, the forecast is excluded from the feature vector.

Coverage is constrained to the curated launch-metro tracts.

Five subject tracts are examined here, spanning Atlanta, Austin, and their respective neighbor sets across Houston, Nashville, Dallas-Fort Worth, Raleigh, and Charlotte.

Buckhead (Atlanta) produces the third-tightest cluster among the 5 worked examples on this page, with similarity scores ranging from 0.9879 to 0.9947 across its five neighbors. All five neighbors are drawn from Houston or Raleigh, which is notable given that Buckhead is an Atlanta tract: the algorithm finds no Atlanta neighbors for this subject, suggesting its combination of Verus score, income profile, and appreciation history is more common in Houston's western suburbs than in the Atlanta metro itself. The nearest neighbor, Houston - Tract 4403, scores 0.9947 and carries a historical CAGR of 5.75% and a median home value of $702,800.

Bouldin Creek (Austin) returns neighbors from Nashville, Dallas-Fort Worth, and Atlanta, with similarity scores from 0.9894 to 0.9932. The nearest neighbor, Nashville - Tract 7800, scores 0.9932 and posts a historical CAGR of 7.71%. The income range across Bouldin Creek's neighbors, from $126,883 to $181,061, is wide relative to the tight similarity band, reinforcing that the model is weighting the proportional structure of the feature vector rather than absolute income levels.

Inman Park (Atlanta) presents a different character. Its five neighbors span Atlanta, Dallas-Fort Worth, and Houston, with similarity scores from 0.9887 to 0.9942.

Decatur (Atlanta) produces the highest single similarity score among the 5 worked examples on this page: Avondale Estates scores 0.9972, a margin that stands out against the other nearest-neighbor pairs. The five Decatur neighbors span Atlanta, Charlotte, Raleigh, and Dallas-Fort Worth, with similarity scores ranging from 0.9755 to 0.9972, the widest spread among the five subject tracts examined here. That spread of 0.0217 between the top and bottom neighbor is more than twice the spread observed for Buckhead (0.0068) and Bouldin Creek (0.0038), suggesting that while Decatur has a very close statistical twin in Avondale Estates, the broader neighborhood of similar tracts disperses more quickly. All five Decatur neighbors carry historical CAGRs between 7.29% and 7.64%, a notably compressed range that contrasts with the wider value dispersion across the set.

Midtown (Atlanta) is one subject tract among the 5 worked examples on this page whose neighbor set includes a tract flagged with a feature divergence. Dallas-Fort Worth - Tract 1653 scores 0.9834, the lowest similarity score within Midtown's neighbor set, and differs from Midtown on population momentum. Houston - Tract 0202 scores 0.9873 and omits Verus score from its shared-feature list, another partial alignment worth flagging. The Midtown neighbor set spans home values from $456,000 to $611,100 and historical CAGRs from 4.70% to 6.56%, a moderate spread consistent with a mid-tier urban tract that has analogs across multiple Sun Belt metros.

The Buckhead example illustrates this: two of its five neighbors score above 0.9924, meaning the model found multiple tracts in the curated launch-metro tracts that align tightly on the full 8-feature profile.

What the score does not capture is equally important. Physical characteristics, walkability, transit access, school district quality, crime patterns, zoning constraints, and proximity to employment centers, are absent from the feature vector. Two tracts can be statistically near-identical on income, appreciation, and education share while differing substantially on neighborhood character. The honesty note embedded in the coverage block states this directly: neighbors are statistically similar on the measured features, not identical.

The exclusion of forecast from the feature vector also means that similarity does not imply shared forward trajectory. The model applies a forecast cap of 37.01% to projected appreciation across the framework, which constrains the upper bound of projected appreciation.

The geographic dispersion of neighbors is itself an interpretive signal. When a subject tract's five nearest neighbors are drawn from two or more metros, as is the case for Buckhead, Bouldin Creek, Inman Park, Decatur, and Midtown, it indicates that the subject's statistical profile is not locally unique; it has close analogs in other Sun Belt markets. Conversely, when neighbors cluster within the same metro (Inman Park draws three of five from Atlanta), it suggests the subject's profile is more locally common. Neither outcome is inherently favorable or unfavorable; the distribution simply describes where statistically comparable conditions exist within the curated launch-metro tracts.

The comparables engine operates on an 8-feature vector drawn entirely from measured and historical data. The features are: Verus score, historical CAGR percentage, appreciation percentage, log median income, education ratio, log current value, population year-over-year percentage, and income year-over-year percentage. Logarithmic transformation is applied to median income and current value before scaling, which compresses the right tail of those distributions and reduces the influence of very high-value tracts on the similarity calculation.

Robust scaling is applied after log transformation.

After robust scaling, cosine similarity is computed between the subject tract's weighted feature vector and every other tract's vector in the curated launch-metro tracts. Cosine similarity measures the angle between two vectors in feature space, not their absolute distance, which is why proportional profile alignment matters more than absolute level. The raw cosine output, which ranges from -1 to 1, is rescaled onto a 0-to-1 interval.

The forecast is excluded from the feature vector by design. Including a forward-looking estimate in a similarity calculation would conflate the model's predictions with the measured inputs, potentially creating circular outputs. The forecast cap of 37.01% applies to the separate forward-projection module and has no bearing on similarity scores. Coverage is limited to the curated launch-metro tracts; the model does not score or match tracts outside that population.