"All societal problems are people problems, and all people problems are fundamentally real estate problems." — Probably Peter Thiel
Allocations are set by metro, research is published by metro, and the rent growth assumption in the Excel model is a metro number with a submarket adjustment bolted on by hand. The unit of ownership is a building on one side of a county line, but if it were on the other side the price might collapse by 20%.
These gaps are, however, measurable to a far more granular level. But first, back to the problem: Across the 47 metros where a public rent index covers at least three counties, the best and worst are 9.3 points apart in five-year rent growth in the median market, 15.1 points at the ninetieth percentile; and counties are coarser than any submarket, so that is a floor. The demand side is worse: the median metro spans 17 people per thousand residents between its highest and lowest county migration rate. Put the two together and the metro assumption stops being a forecast of the asset. In rent, the range inside a typical metro is about as wide as the range across the middle half of all metros; in migration, the median metro holds more dispersion than separates the tenth-percentile metro from the ninetieth. In plain terms: picking the right metro was never most of the decision. Which county the building sits in moves the outcome about as much as which metro it sits in, and the street it sits on moves it further still. The county is simply the finest level at which most public data can show that.
Show every county as a table
| County | Population | Asking rent, 5-year growth | Asking rent, 1-year growth | Home value, 5-year growth | Net domestic migration per 1,000 | Units permitted per 1,000 | Renters who can afford the local rent | Rent as a share of median renter income | Income of arrivals vs leavers | Migration balance |
|---|---|---|---|---|---|---|---|---|---|---|
| Dallas County, TX | 2,656,028 | +16.8% | +0.6% | +14.7% | -17.8 | +4.8 | 41.6% | 36.2% | 0.92× | −0.108 |
| Tarrant County, TX | 2,230,708 | +17.7% | +0.6% | +16.4% | -1.4 | +8.4 | 39.1% | 37.8% | 0.96× | +0.010 |
| Collin County, TX | 1,254,658 | +13.3% | -1.4% | +18.2% | +15.5 | +15.0 | 51.6% | 29.0% | 1.04× | +0.113 |
| Denton County, TX | 1,045,120 | — | -1.0% | +16.4% | +14.7 | +10.1 | 45.6% | 33.0% | 1.17× | +0.095 |
| Ellis County, TX | 232,387 | +20.9% | +0.5% | +17.0% | +31.7 | +12.8 | 34.3% | 41.9% | 1.14× | +0.226 |
| Johnson County, TX | 210,547 | +22.9% | +0.8% | +17.7% | +30.1 | +10.4 | 40.0% | 37.7% | 1.21× | +0.183 |
| Kaufman County, TX | 197,829 | +16.0% | +1.9% | +7.1% | +46.3 | +8.8 | 33.9% | 44.3% | 0.92× | +0.218 |
| Parker County, TX | 179,707 | +25.0% | +2.3% | +20.4% | +30.0 | +2.4 | 39.6% | 38.2% | 1.24× | +0.202 |
| Rockwall County, TX | 137,044 | +15.3% | -0.3% | +16.0% | +33.1 | +13.2 | 46.5% | 32.3% | 1.22× | +0.173 |
| Hunt County, TX | 118,729 | — | +2.2% | +13.9% | +36.1 | +10.8 | 38.8% | 38.2% | 1.16× | +0.196 |
| Wise County, TX | 81,275 | — | -1.6% | +16.1% | +37.6 | +5.7 | 37.1% | 39.7% | 1.14× | +0.242 |
A headcount is the wrong unit for a rent forecast
Migration is published as a headcount, but a household's contribution to rent is bounded by its income. Watch fifteen years of it move county by county and the metro's own core is usually the county losing people while its ring gains.
Zoom into our own example and the county is where the view stops being two-sided. Below it, public data can still see who arrives in each census tract, but never who leaves.
The IRS tracks tax returns that change county from one year to the next, and publishes for every county pair how many moved and what they earned. That turns migration from a headcount into a ledger: the income arriving can be set against the income leaving.
The ceiling on rent is an income distribution, and it is published
What arrivals will pay is set by the income distribution of the households who would live there, and for renter households it is published per county, by bracket, every year. Fit a log-normal to the ACS brackets (median R² 0.991 across 2,759 counties), shift it by twelve months and the affordability convention, and the share who clear a rent is a demand curve.
In 64% of counties the median renter's ability to pay at 30% of income is below the local asking rent. Sitting tenants are past the convention too: the median county spends 30% of median renter income on rent and 50% are over thirty. In the median metro that share spreads 10.4 points between its strongest and weakest county.
Show every county as a table
| County | Renter households | Median renter income | Can pay at 30% | Asking rent | Median gross rent | Clear the ask | Realised burden |
|---|---|---|---|---|---|---|---|
| Dallas County, TX | 478,237 | $54,501 | $1,363 | $1,646 | $1,469 | 41.6% | 32.3% |
| Tarrant County, TX | 312,653 | $51,994 | $1,300 | $1,639 | $1,447 | 39.1% | 33.4% |
| Collin County, TX | 143,751 | $71,839 | $1,796 | $1,736 | $1,792 | 51.6% | 29.9% |
| Denton County, TX | 120,115 | $62,111 | $1,553 | $1,706 | $1,642 | 45.6% | 31.7% |
| Ellis County, TX | 16,352 | $52,275 | $1,307 | $1,827 | $1,449 | 34.3% | 33.3% |
| Johnson County, TX | 16,203 | $50,120 | $1,253 | $1,573 | $1,343 | 40.0% | 32.1% |
| Kaufman County, TX | 10,641 | $50,221 | $1,256 | $1,856 | $1,408 | 33.9% | 33.7% |
| Parker County, TX | 9,719 | $50,842 | $1,271 | $1,618 | $1,440 | 39.6% | 34.0% |
| Rockwall County, TX | 6,890 | $72,968 | $1,824 | $1,962 | $1,899 | 46.5% | 31.2% |
| Hunt County, TX | 10,969 | $48,712 | $1,218 | $1,551 | $1,184 | 38.8% | 29.2% |
| Wise County, TX | 4,292 | $50,316 | $1,258 | $1,666 | $1,239 | 37.2% | 29.5% |
The slope matters more than the level: at the rent a county is asking, the curve says what another hundred dollars costs in pool depth. What it cannot do is date anything. Affordability headroom carries no information about subsequent rent growth, which is why our own model moved from what they can afford to what fair market rent is, a question about lease-up timing rather than income.
Which signals carry information, and which only feel like they do
If the spread inside a metro were noise, the answer would be to diversify across it and stop paying for granular data. So does anything observable in advance rank the counties inside a metro? Signal and outcome are both measured as departures from the county's own metro that year, which discards the Sun-Belt-beats-Rust-Belt gap on purpose. A year counts only where the sign holds and clears its own noise band. Bold clears the pooled band too, and holds in at least seven of every ten base years the signal covers.
The sharpest signal is not a demand signal. The multifamily share of permitted units (what a county lets through, not how much) runs at −0.27, worth 2.7 points between the top and bottom fifth, and its sign is the opposite of the instinct: where the pipeline concentrates, rent then lags.
Both migration measures work and the IRS view works better: +0.21 for the balance ratio against +0.16 for the Census headcount rate, income ratio +0.16 behind them.
Three things that feel like signals are not: churn, at −0.04; affordability headroom; and trailing twelve-month rent growth, the headline in every quarterly report, at +0.04 against ±0.02, a sign that survives in only 4 of its 7 base years and turns to −0.06 once you shorten the horizon to a year.
2,046 county-years. Outside the noise band, and the same sign beyond each year's own band in 7 of 8 base years measured separately. 25 points sit outside the axes and are drawn at the edge.
Census data says whether demand is arriving. MSCI says what it costs.
None of that is price. The public layer tells you which side of a metro is gaining households, whether they out-earn the ones leaving, and how deep the pool is at a given rent. It cannot tell you what any of it trades at, or whether the spread has been bid away.
The public half: exhaustive, forward, free. Every county, every year. It describes the demand still arriving after the last comp has aged out. Its strength is that there is no survivorship: a county with no transactions still reports its population. Its limits: no price, and the floor is the county, because the census asks where you lived a year ago, so a tract reports who arrived and never who left.
The licensed half: priced, granular, backward. MSCI Real Capital Analytics: consideration actually paid, with cap rate, buyer and capital group, on its own submarket tags. The one widely held record that prices a geography smaller than a metro from closed transactions. Its strength is what someone paid, not an index of what things are worth. Its limits: it is backward-looking, a submarket-quarter cell can be a handful of trades, and only aggregates of it can be published, which is what the box below holds.
Four uses that survive the evidence:
Sourcing. Screen for the disagreement: public signals positive, priced record unmoved. The output is a ranked list to spend a week on.
Underwriting. The metro assumption imports a 9.3-point range of realised outcomes as a point estimate. A submarket curve is a real improvement; presenting it without a distribution is not, because a correlation of 0.3 moves the centre of an IRR distribution a little and widens the honest confidence interval a lot.
Portfolio management. Permit composition is the one signal with a stable sign, visible years ahead because an authorisation takes one to three years to deliver: the cheapest deterioration alarm available where you already own.
Pricing. Use the curve's slope, not its ceiling: what another hundred dollars costs in pool depth goes into the lease-up plan.
Firms operate at metro level not because metros are the right unit, but because the county layer means a dozen sources on incompatible clocks, a boundary change every decade, and a licensed record that shares no keys with any of them. That work is not clever. It is just hard to keep correct, and it does not decay.
The largest allocator in the world has stopped buying the metro
Norway's sovereign wealth fund holds about $75 billion of real estate, at the bottom of its own target range, after a decade its new global head of real estate, Alex Knapp, describes as "not been good enough". The strategy that replaces it is written in the vocabulary of this article. NBIM's own plan states that investment decisions "will primarily be driven by asset and sector fundamentals", not by the cities it used to concentrate in.
“We're going to hold it because it's got return potential.”
The test Knapp applies to every standing asset is a single question, would we buy it at today's pricing, and the answer is a sale when it is no. He expects "more cycling for sure", with performance driven by "smart new investments and smart divestments", into a market he reads as having "more net sellers than net buyers". Notably, he also says the fund itself will work "at a strategy level, rather than an asset level", through partners and platforms. The asset-level judgment does not disappear in that model. It moves to the operating partner, who now has to earn the mandate on it.
And so we find ourselves in a stock-picker's market. The phrase is ours, not Knapp's, but it is what a return-potential test applied asset by asset amounts to. When the allocator's edge is which assets it holds and sells rather than which metros it is in, the evidence that matters is the evidence in this article, at the resolution of this article: which side of the county line, who is arriving and what they earn, how deep the renter pool is at the asking rent, and whether the permit pipeline is about to arrive on top of you. None of it is visible from the metro number every allocation memo still quotes.
That is the layer Starboard keeps correct. Every county series in the figures above, on its own clock, joined to the priced record and scored at the asset, so that the question Knapp asks of a building he already owns can be asked of one you do not yet. In a market where the biggest buyer says the metro is no longer the unit, being able to answer it is the whole edge. If you would rather it were kept correct for you, see it read a rent roll.
Sources: Zillow ZORI · Zillow ZHVI · Census Bureau population estimates · Census Bureau Building Permits Survey · IRS Statistics of Income · ACS 5-year (B25118, B25064). Rent and value through June 2026; migration to July 2024; permits 2024; IRS filing pairs to 22-23; ACS 2023 five-year. Built 4 August 2026. Alex Knapp's remarks are from PERE and IPE Real Assets, both December 2025; the strategy language is from NBIM's Strategy 28. MSCI figures are aggregates of licensed MSCI Real Capital Analytics data: US closed sales, cached extract through 2026-07-24; price changes are medians of what traded in the year to June 2021 against the year to June 2026 in cells of at least 20 sales, not a constant-quality index, so a shift in the mix of what sold moves them without any repricing. The county map is Census Bureau county population estimates, Vintage 2019 and Vintage 2025 components of change, net domestic migration per year ending July 1 with the April to July 2020 stub omitted; 3,143 of the 3,144 counties and county-equivalents in the fifty states and DC (Kalawao, HI is unreported), Connecticut as its nine planning regions from 2020, on Census cartographic boundaries. The tract map is ACS 5-year 2019 to 2023 table B07204 for the eleven counties of the Dallas metro, on the 2023 cartographic tract boundaries; inflow is movers from a different county or from abroad over population aged one and over. The county map uses the Vintage 2025 estimates, which revise earlier years; Figure 1 uses the Vintage 2024 release, so the same county and year can differ slightly between the two (Dallas County in 2024 reads −18.2 per 1,000 on the map and −17.8 in Figure 1 with Dallas selected, which is the metro Figure 1 opens on unless your own is one the panel covers). Header photograph: downtown Dallas across the Trinity River floodplain, by Max Fray on Unsplash.