What predicts city rent growth?

What we measured

One row per city, top-250 by population. Target = the 5-year compound annual growth rate of the Zillow ZORI rent index from 2021-04 → 2026-04, computed in this script directly from the monthly ZORI city series. Cities missing either endpoint were dropped, not imputed (no fabrication of the target). 6 of 250 cities dropped → n = 244.

Features: ZHVI 5y CAGR (asset-price control, mirror of BAH's H1); median household income, median age, population, lat/lon (city-level from ACS / Census); bachelor_plus_pct, unemployment_pct, IRS net migration (joined from the nearest ZCTA centroid in zip_master.json). Cities' centroids were matched to the closest ZCTA from the 2023 Census Gazetteer by haversine.

Read: a one-standard-deviation increase in ZHVI 5y CAGR is associated with a +0.71 percentage-point increase in city ZORI 5y CAGR — the dominant signal. The lon and abs_lat coefficients say that, after controlling for ZHVI, eastern and northern cities grew rents faster than western and southern ones (about +0.42 and +0.31 pp/yr per sigma respectively). The IRS-migration coefficient is negative and significant — the opposite of the pre-registered direction.

The city regression beats the BAH regression on every fit-quality metric. R² test is positive (0.28) rather than negative (−0.27). City-level rent CAGR is a more tractable prediction problem than MHA-level BAH CAGR — larger n, the target is the rent index itself rather than an administrative pass-through, and the geographic gradient that died in the BAH study survives here at p<0.001.

IRS net migration is a statistically significant negative predictor of rent CAGR at p < 0.001 with the wrong pre-registered sign. We expected “more people moving in → faster rent growth”; we got “more people moving in → slower rent growth.”

The likely mechanism is a COVID base-effect inside the 2021→2026 window. Destination cities for the 2020–22 migration wave (Sun Belt, Florida, Tennessee) had already spiked their rents by 2021. The 2021→2026 CAGR captures their subsequent normalization — high level, low subsequent CAGR. Origin cities (NYC, SF) bottomed out in 2021 and recovered, giving them higher CAGR off a low base. This is exactly what a naive “follow the migration map” thesis would miss.

Operational implication: WindMayor should NOT underwrite a long-rent thesis in high-inflow destination metros purely on net-migration data. The migration signal in this window is contrarian.

• Supported: ZHVI growth and ZORI growth comove at r = +0.57 at the city cross-section. If you believe a city's asset prices will keep growing, you have substantial basis to expect its rents to grow alongside.
• Out-of-sample fit is positive (R² test = +0.28) — the city regression actually generalizes, unlike the BAH study where R² test went negative.
• Geographic gradient is real after controls — eastern (lon ↑) and northern (abs_lat ↑) cities outperformed at p < 0.001 within ZHVI peers. We don't fully know why — possibly Sun Belt rent growth was already fully priced into ZHVI, so the residual tilts the other way.
• Migration story is inverted — high-inflow destinations are low-CAGR cities in this 5y window. Base effect from COVID-era spike then normalization.
• Pricing-power thesis weakly rejected — high-income cities grew rents slower bivariately (r=−0.29). Income level may be a contrarian indicator over the 2021–2026 window.
• Operationally: ZHVI + geographic-gradient + base-effect awareness is the actionable trio. Education, age, unemployment, and population add no signal here.

• n=244 is small. Nine features ÷ 244 obs ≈ 27:1 is OK for naive OLS but partial-correlation power is modest. Out-of-sample R² of +0.28 with 9 features looks plausible but not exceptional.
• Spatial autocorrelation not corrected. Cities near each other (Dallas/Plano, SF/Oakland) are not independent. Reported p-values are suggestive, not confirmatory.
• Nearest-ZCTA proxy for bachelor_plus, unemployment, and IRS migration is approximate. Large cities span many ZCTAs with different ACS profiles; picking the closest ZCTA centroid is a known weakness. A polygon-weighted ACS join might recover signal H2 / H4 missed here.
• IRS migration is 2022-only — a single-year snapshot inside a 5y CAGR target. The window mismatch may explain the wrong-sign result; a 5y cumulative net-flow feature would be cleaner.
• Window-specific artifact risk. 2021→2026 is the COVID-recovery window. Many results here may be base-effect signatures of pandemic dislocation. A 2010–2019 or 2015–2020 window might give very different answers.
• Top-250 by population is a selection bias. Findings generalize to mid-large U.S. cities, not to rural areas or sub-100k towns.
• Discarded features: raw population (replaced with log_population); rent level, yield, DSCR, price-to-rent (mechanically related to the target); state dummies (50 features for 244 rows is unsafe).
• Gradient boosting was the planned supplementary nonlinear model. sklearn is not installed in the working Python environment, so the GBM result is null rather than hand-rolled.
• Causation: we cannot claim it. Observational cross-section. ZHVI → ZORI, common drivers behind both, or ZORI → ZHVI would all produce the partial correlations we see.

The regression is one Python script. It reads the cities panel, the cached Zillow ZORI/ZHVI city CSVs, the Census ZCTA gazetteer, and zip_master.json — computes per-city features in parallel via ThreadPoolExecutor, fits OLS, and writes the JSON result. No external network calls; data is from cache.

[run] loading cities, ZORI, ZHVI, ZCTA centroids, zip_master...
[run] 250 cities loaded
[zipmaster] loaded 31,646 ZCTA records
[zillow] ZORI: 4,430 city series
[zcta] loaded 33,791 ZCTA centroids
[zillow] ZHVI all-homes: 21,356 city series
[panel] computing features for 250 cities (parallel)...
[panel] 250 rows built
[panel] dropped 6 rows missing zori_5y_cagr target -> 244 cities for fit
[panel] 244 rows after final dropna
[fit] OLS on n_train=195 features=9

=== OLS coefficient summary (standardized) ===
  const                          beta=+0.0508  t=+67.20  p=0.000
  zhvi_5y_cagr                   beta=+0.0071  t= +8.57  p=0.000 ***
  bachelor_plus_pct              beta=+0.0002  t= +0.29  p=0.772
  irs_net_migration_2022         beta=-0.0030  t= -3.79  p=0.000 ***
  unemployment_pct_acs           beta=+0.0003  t= +0.36  p=0.717
  median_income                  beta=-0.0017  t= -1.67  p=0.097 *
  log_population                 beta=-0.0010  t= -1.24  p=0.215
  abs_lat                        beta=+0.0031  t= +3.85  p=0.000 ***
  lon                            beta=+0.0042  t= +4.45  p=0.000 ***
  median_age                     beta=+0.0003  t= +0.42  p=0.678

  n        = 244
  n_train  = 195, n_test = 49
  R^2 train = 0.4819
  R^2 test  = 0.2762
  MAPE test = 0.1676