Vacancy Is Not One Number: Why CoStar, Yardi, and the Census Disagree

In the first quarter of 2026, the national vacancy rate for American rental housing was 4.8 percent. It was also 5.1 percent, 5.7 percent, 7.3 percent, and roughly 8.5 percent. Every one of those figures was published by a credible institution, for the same country, describing the same quarter. CBRE printed 4.8 (1). RealPage's occupancy of 94.9 implied 5.1 (2). Yardi Matrix's occupancy of 94.3 implied 5.7 (3). The U.S. Census Bureau's Housing Vacancy Survey said 7.3 (4). CoStar's total-inventory figure ran near 8.5, with its own analysts forecasting 8.8 by year-end (5).

That is a spread of roughly 370 basis points between the two most frequently cited numbers, and nearly 400 between the extremes. A lender who pulls one figure and a review appraiser who pulls another are not having a disagreement about the market. They are having a disagreement about definitions, and neither of them knows it.

This article is the autopsy. We reconstruct, provider by provider, why the numbers diverge, quantify how much of the spread each methodological choice explains, and close with the decision framework we apply in our own lender-facing work: which source governs which question. The short version, which we have argued before in our work on market saturation benchmarks, is that these are facts about methodologies, not facts about markets (6). What follows is the proof.

THE Q1 2026 TRIANGLE: five-source national vacancy comparison, same quarter

The word "vacancy" is doing five different jobs

Before any reconciliation is possible, one thing has to be understood: the five institutions above are not measuring the same quantity badly. They are measuring five different quantities well. The differences run along four independent axes, and each axis moves the headline number by an amount that can be estimated.

The first axis is the universe: what stock is counted before anything is counted vacant. The second is stabilization: whether properties still filling up after construction are in or out. The third is the concept itself: whether "vacant" means doors physically empty or revenue economically foregone. The fourth is production: whether the number comes from a probability sample with a published margin of error, or from a proprietary tracking operation with no disclosed error at all.

Take them in order.

Axis one: the denominator decides the answer before the survey begins

The United States has roughly 44.6 million occupied rental units (7). The single most consequential fact about that stock, for anyone comparing vacancy sources, comes from the Rental Housing Finance Survey: nearly 46 percent of American rental units sit in properties of one to four units (8). Almost half the rental market is a house, a duplex, or a small building with a landlord who owns three of them.

Now consider what each source actually tracks against that stock.

Yardi Matrix covers multifamily properties of 50 or more units — its own materials describe more than 63,000 such properties totaling more than 12 million units (9). Twelve million against 44.6 million is roughly 27 percent of the rental universe. The 50-unit floor removes the entire one-to-four-unit segment and everything from 5 to 49 units besides. Novogradac, in its teardown of Yardi's affordable-housing reporting, put the consequence plainly: smaller properties and many rural markets fall outside the lens (10).

CoStar tracks a broader universe — roughly 18 million multifamily units across a claimed 400 markets and 2,500 submarkets — which works out to about 40 percent of the rental stock (11). Still institutional, still concentrated in professionally managed buildings, still structurally blind to the scattered small-property stock, but wider than Yardi's aperture.

The Census Bureau tracks everything. Its Housing Vacancy Survey draws a probability sample of about 72,000 housing units per quarter from the Master Address File, with no property-size floor whatsoever (12). A single-family rental in rural Missouri and a 600-unit tower in Dallas carry proportionate weight.

Here is why this matters more than any other single fact in this article. For years, analysts — ourselves included — reasoned that the Census sits in the middle of the pack because it blends tighter-vacancy small-property stock into the loose institutional segment. That was an inference. The data now confirm it as a measurement. Census figures show single-family rental vacancy running at 5.7 percent in 2024, while the Census's own five-or-more-unit multifamily vacancy ran 8.2 percent in the fourth quarter of that year (13). The small stock the providers exclude is materially tighter than the large stock they track. Blending it in dilutes the headline downward. The Census is not "missing" vacancy that CoStar sees; it is counting occupied houses that CoStar does not.

So before a single unit is classified, before any survey week begins, the denominator choice alone has ordered the three main sources: narrow institutional universe with tight management reads lowest, all-stock blend reads middle, broad institutional universe carrying the supply wave reads highest.

THE UNIVERSE FUNNEL: 44.6M total rental units → CoStar ~18M → Yardi ~12M, with the 46% one-to-four-unit segment highlighted

Axis two: lease-up, or the wedge that moves with the cycle

The second axis is the one most likely to be pulled, unknowingly, inside the data product itself.

A newly delivered apartment building is mostly empty on the day it opens. That emptiness is a construction-timing artifact, not market distress — the Census Bureau's Survey of Market Absorption shows that only 49 percent of apartments completed in Q4 2025 had rented within three months, the sixth consecutive quarter below 50 percent, but that 90 percent had rented within twelve (14). The building fills. The only question is whether the data provider counts it while it fills.

A "stabilized" series excludes it. A "total" series includes it. And the two flagship providers made opposite choices.

CoStar's headline is a total-inventory figure. Its own reporting states the rule: a property is considered in lease-up if it has not yet reached at least 90 percent occupancy or has been open for less than 18 months, and CoStar publishes a separate stabilized series alongside the total (15). Yardi Matrix's headline is the opposite construction — its monthly national report is explicitly an occupancy rate in stabilized properties, with lease-up assets and "exception properties" excluded (16). The exact occupancy trigger and seasoning window behind Yardi's stabilization rule are not publicly disclosed, a point we return to below.

How big is the wedge? CBRE has quantified it. Had pre-stabilized properties been included in market statistics from 2015 to 2019, total vacancy would have run only about 60 basis points above the reported stabilized figure. As of the first quarter of 2025, the glut of pre-stabilized product from the construction boom would boost total vacancy by roughly 1.5 percentage points (17). The wedge is not a constant. It is a supply-cycle variable that ran 60 basis points in a calm decade and 150 in a delivery wave.

At the metro level, in the markets where the wave actually landed, the wedge dwarfs the national figure. In Austin, Yardi's stabilized occupancy of 92.3 percent in late 2025 implied vacancy near 7.7 percent, while CoStar's total-inventory vacancy printed 13.5 percent in the first quarter of 2026 — a spread of roughly 580 basis points in one metro, from one definitional choice (18). In supply-constrained New York, where there is almost no lease-up inventory to argue about, the equivalent spread collapses to roughly 180 basis points (19). The divergence between the sources scales with the size of the local construction pipeline. That is the signature of a methodological artifact, not a provider bias.

And here is the practitioner's trap. The number a subscriber sees first — the default in the interface, the one most likely to be copied into an appraisal or a credit memo — is the total figure in CoStar and the stabilized figure in Yardi. An analyst who pulls "the CoStar number" and "the Yardi number" without reading a single definition has compared a total to a stabilized, and the 280-basis-point gap between them is baked in before the analysis begins.

THE LEASE-UP WEDGE: stabilized vs. total vacancy, national and Austin vs. New York, 2015–2026

Axis three: doors empty versus dollars foregone

Everything so far concerns physical vacancy — units empty, a count of doors. Every one of the five sources in our opening triangle publishes a physical measure. Not one of them publishes the number a loan is actually repaid from.

Loans are repaid out of economic occupancy: the share of gross potential rent that actually arrives. Between the physical print and the economic reality sit four line items, none of which appear in any published vacancy rate. Concessions — and as of the first quarter of 2026, roughly a quarter of U.S. apartments were offering them, at an average value of 7.2 percent of rent, with the listings share in oversupplied metros like Denver, Charlotte, Dallas, and Austin running above 60 percent (20). Non-revenue units — models, employee units, doors down for renovation. Loss-to-lease, which in today's declining-rent Sun Belt markets has inverted into gain-to-lease, meaning in-place rents sit above market and revenue falls at renewal rather than rising. And bad debt, which the public REITs have worked back down toward one percent of revenue for institutional-quality stock, but which runs multiples of that in the class B and C assets that dominate agency and SBA collateral (21).

The profession's standards are unambiguous that the economic version is the required one. The Appraisal Institute defines effective gross income as potential gross income less a vacancy and collection loss deduction — collection loss, not merely vacancy (22). Fannie Mae's underwriting requires the combined deduction for vacancy, concessions, and bad debt to equal at least five percent of gross potential rent; Freddie Mac's floor is likewise generally not less than five percent; HUD's MAP Guide prescribes a seven percent minimum for market-rate assets even at full physical occupancy (23).

The failure mode is therefore precise, and we can put dollars on it. Take a 200-unit property in a high-concession Sun Belt market: 8 percent physical vacancy, concessions near 6 percent of gross potential rent, typical non-revenue and bad-debt loads. The analyst who applies the published 8 percent physical figure directly to gross potential rent overstates effective gross income by roughly 900 basis points of economic loss. Because operating expenses are fixed, the entire error drops to net operating income — an overstatement of roughly 21 percent — and a loan sized to a 1.25x debt service coverage ratio on the shortcut pro forma actually covers at about 1.04x (24). The credit file says the deal has cushion. The property says it is near breakeven. The difference is one column in a spreadsheet, populated from the wrong concept.

PHYSICAL-TO-ECONOMIC BRIDGE: waterfall from 8% physical to ~17% economic in a high-concession metro

Axis four: the number that admits its own error

The fourth axis is the least discussed and, for reviewers, the most important. The three source families are manufactured by different machines, and only one of the machines discloses its tolerance.

The Census Housing Vacancy Survey is a probability sample. Its national rental vacancy rate carries a published standard error — 0.15 percentage points in Q1 2026 — and a 90 percent confidence interval the Bureau states outright: 7.1 to 7.5 percent (25). More than that, the Census tells its users when a movement is noise. The Q1 2026 release states that the 7.3 percent rate was not statistically different from the 7.1 percent of a year earlier, nor from the 7.2 percent of the prior quarter (4). When the October 2025 federal funding lapse forced the Bureau to build a quarter from two months of data instead of three, it said so and widened the published margin accordingly (26).

The commercial providers are quasi-censuses of their tracked universes — research calls, listing feeds, revenue-management-system data, public records — plus modeling for the gaps. CoStar cites 1,600-plus researchers and 16 million monthly rent updates (27). Yardi describes telephone surveys conducted as a prospective renter (28). These are serious operations. But neither publishes a margin of error, a confidence interval, a coverage ratio, a verification rate, or a revision policy. We searched the methodology pages, the SEC filings, and the report footnotes; the disclosures do not exist. Yardi's terms of use state that its content is provided as is, without warranty of correctness, accuracy, or reliability (29). And because both providers revise history continuously as new properties enter the database, a figure cited "as of" a date in a feasibility study may be silently unreproducible a year later.

This produces an epistemic inversion worth stating bluntly. The reviewer instinct is to treat the provider figure as the precise one and the Census figure as the noisy one. The truth runs the other way. The Census figure is the only number in the set that quantifies its own uncertainty. A provider's 8.5 is not more precise than the Census's 7.3 plus-or-minus 0.25. It is merely silent about its precision. In our own reports, that silence is disclosed, not papered over.

There is one more consequence of the production difference, and it bites hardest exactly where much of our practice lives. Provider coverage thins to nothing in small and rural markets — a rural trade area frequently contains not a single 50-plus-unit property for Yardi to track. Meanwhile the Census's error widens as geography narrows: no ACS one-year data exists below 65,000 population, and in more than 70 small and midsize cities even the five-year rental vacancy estimates carry margins of error of three percentage points or more, meaning a published 7 could truly be 4 or 10 (30). There is a band — trade areas below roughly 20,000 people, which is to say nearly the entire eligible geography of USDA Business & Industry and Community Facilities lending — where both source families fail simultaneously. In that band, the defensible answer is not a database print at all. It is a primary survey of the competitive set to NCHMA standards, cross-checked against HUD's quarterly USPS address-vacancy data and, where subsidized stock dominates the local market, USDA's own Section 515 property-level occupancy files (31). We will treat that fallback stack fully in a companion piece; for present purposes the point is simpler. Anyone quoting a CoStar submarket vacancy rate for a town of 8,000 is quoting a number derived from a sample the provider will not size and does not flag.

THE RELIABILITY GRADIENT: source usability by geographic scale, national to rural trade area

Reconciling the triangle: where the 370 basis points actually come from

With all four axes on the table, the Q1 2026 spread can be walked, step by step, from the lowest print to the highest.

Start at the stabilized institutional floor. Yardi's implied 5.7, RealPage's 5.1, and CBRE's 4.8 are all the same species of number — physical vacancy in stabilized, professionally managed, larger-property stock — and they cluster accordingly, between roughly 4.8 and 5.9 percent.

Add the lease-up effect. CBRE's measured national wedge of roughly 150 basis points carries the stabilized figure toward 7.2 percent (17). This is the price of counting buildings while they fill.

Add the remaining universe and method effects. Moving from the stabilized 50-plus-unit universe to CoStar's broader total-inventory tracked universe adds roughly another 130 basis points, arriving at CoStar's realized 8.5 percent. We label part of this step honestly as residual: universe breadth, model imputation, and reference-period mismatch are entangled here, and no public data cleanly separates them. A forced full attribution would be tidier and less true.

Now locate the Census. Its 7.3 sits below CoStar's 8.5 not because it measures the institutional market more optimistically — the Census's own five-plus-unit vacancy of 8.2 percent in Q4 2024 essentially matched CoStar's total — but because the Census also counts the 5.7-percent-vacant single-family stock and the tighter small-property segment that the providers exclude (13). The all-stock blend dilutes downward. Middle position, fully explained.

Finally, extend past every published number to the one that belongs in a pro forma. Layering current concession prevalence and value onto the physical prints adds on the order of 180 basis points of gross-potential-rent loss from concessions alone, before loss-to-lease and bad debt, carrying true economic vacancy in soft markets past ten percent (20). No provider publishes that number. Every lender is repaid from it.

That is the whole spread, accounted for: a definitional floor near five, a lease-up wedge of one and a half points, a universe blend that explains the Census's middle seat, a total-inventory ceiling near eight and a half, and an economic reality above ten that none of the headlines print. Nothing in that walk required any provider to be wrong. It required only that the reader know what each one is counting.

One refinement matters for anyone applying this at the metro level. At the national scale, the gaps are statistically real: 120 to 160 basis points against a Census margin of error of 0.2 is genuine disagreement, definitionally driven. At the MSA scale, the Census margin widens to something on the order of two to four percentage points, and in a tight gateway market the entire provider-versus-Census "disagreement" can fall inside the Census confidence interval. In New York, it does: the sources do not statistically disagree at all (19). The lesson cuts both ways — a metro-level divergence should not be reported as a finding until it clears the published margin of error, and a metro-level agreement should not be celebrated as validation when the error band could swallow a three-point gap.

THE VACANCY STACK: full waterfall, stabilized floor → lease-up → total inventory → Census blend → economic vacancy

What a reviewer should trust, for which purpose

Everything above compresses into a working rule: match the source to the question, and label the basis on every figure that enters a credit file.

For the direction of the national market — is rental housing loosening or tightening — the Census Housing Vacancy Survey is the governing source. Full universe, consistent method across decades, disclosed error, and an explicit warning when a quarterly move is noise. Its weakness is that it says nothing about any particular submarket or asset class.

For positioning an institutional multifamily asset within its competitive set, the providers govern — on a stabilized basis. Yardi's stabilized occupancy, RealPage's same-store series, CoStar's stabilized toggle, or CBRE's stabilized print are the right comparables for a seasoned asset's pro forma vacancy, because they measure the population the subject will join. The total-inventory figure does not belong in a stabilized pro forma, and in a heavy-delivery metro the difference is not a rounding error; it is five points.

For lease-up and absorption risk on new construction, invert the rule. The stabilized figure now understates the truth, because the subject will spend its first eighteen months competing against exactly the pre-stabilized cohort the stabilized series excludes. Use the total figure, the four-and-five-star vacancy that proxies the new-delivery segment, and derive the absorption schedule from comparable lease-ups delivered within the prior 24 months, as the NCHMA content standards require — checked against the federal absorption benchmark from SOMA while it survives (14).

For any question touching the full rental stock — single-family rental exposure, small-property markets, affordability analysis, and any assignment where two-to-four-unit product is a real substitute — only the Census family covers the universe. The providers structurally cannot see nearly half the market.

For a single asset's stabilized revenue, no market print suffices at all. The governing number is economic vacancy built from the subject's own rent roll and the comp set's actual concessions, floored at the agency minimums, with the physical market figure serving only as one input to one component.

And for small and rural trade areas, the honest answer is that the secondary data fails, disclosed margins and all, and the study stands on primary research or it stands on nothing.

The discipline, stated once

The recurring error in credit files is not the use of any particular source. Every source in this article is professionally produced and fit for its designed purpose. The error is the unlabeled number — "market vacancy: 8.4%" — with no basis, no universe, no stabilization flag, no date convention, and no acknowledgment that a different, equally credible institution printed 5.7 for the same quarter. Our own standard, applied across every feasibility study we deliver, is that a vacancy figure never appears without its passport: source, universe, stabilized or total, physical or economic, reference period, and, where one exists, the margin of error. That is a single sentence of disclosure per figure. It is the difference between an analysis a review appraiser can audit and a number that happens to be in a database.

Vacancy is not one number. It never was. The institutions publishing it are not in conflict; they are answering different questions with different instruments, and the spread between them is not noise to be averaged away but information about universe, cycle position, and concept. Read that way, the disagreement between CoStar, Yardi, and the Census stops being an inconvenience and becomes what it always actually was: the most efficient diagnostic available for what kind of market you are actually looking at — and a standing test of whether the analyst quoting the number understands what it measures.

July 11, 2026, by Michal Mohelsky, J.D. Principal of MMCG Invest, LLC, feasibility study company serving feasibility studies for USDA projects. 

Reach out to discuss how our methodology supports your lending or investement decision.

MMCG Invest, LLC prepares lender-grade feasibility studies and market analyses for SBA 7(a), SBA 504, USDA, and conventional loan programs nationwide. The methodological standards described in this article are applied in every engagement.

Request a Feasibility Study → Book a Meeting: https://calendar.app.google/EJzWEz3GCqLY2jU86

Michal Mohelsky, J.D. | Principal | mmcginvest.com 

Contact: michal@mmcginvest.com

Phone: (628) 225-1125

Disclaimer: This report is provided for informational purposes only and does not constitute investment advice. Data presented herein is derived from proprietary MMCG databases and third-party sources believed to be reliable; however, MMCG Invest makes no representation as to the accuracy or completeness of such information. Figures from third-party industry databases have been independently verified and, where appropriate, adjusted to reflect MMCG's proprietary analytical methodology. Past performance is not indicative of future results.

Sources

(1) CBRE Research, U.S. Multifamily Figures, Q1 2026 (stabilized vacancy 4.8%; net absorption 78,100 units). (2) RealPage Market Analytics, Q1 2026 Data Update (occupancy 94.9%; concessions 25.5% of units at 7.2% average). (3) Yardi Matrix, Multifamily National Report, February 2026 (occupancy in stabilized properties, 94.3%, one-month reporting lag). (4) U.S. Census Bureau, Quarterly Residential Vacancies and Homeownership, First Quarter 2026, released April 28, 2026 (rental vacancy 7.3%; not statistically different from Q1 2025 and Q4 2025). (5) Apartments.com / CoStar Group, U.S. Multifamily Vacancy Forecast, May 7, 2026 (Grant Montgomery, National Director of Multifamily Analytics; realized total-inventory vacancy ~8.5% Q1 2026; forecast 8.8% year-end 2026, 8.4% year-end 2027). (6) MMCG Invest, "Market Saturation Benchmarks: The Hidden Denominators That Decide Whether a Feasibility Study Is Bankable," May 2026. (7) U.S. Census Bureau, American Community Survey, 2019–2023 five-year estimates (renter-occupied units, 44.6 million). (8) HUD / U.S. Census Bureau, Rental Housing Finance Survey, 2021 (share of rental units in one-to-four-unit properties). (9) Yardi Matrix product documentation (63,000+ properties of 50+ units; 12M+ units). (10) Novogradac, "Yardi Matrix Report: Breaking Down What It Reveals About Affordable Housing's Competitiveness," February 2026. (11) CoStar Group product documentation and market analytics materials (tracked multifamily universe; markets and submarkets). (12) U.S. Census Bureau, CPS/HVS Methodology and Source and Accuracy of Estimates (sample design, Master Address File, ~72,000 units per quarter). (13) U.S. Census Bureau, Housing Vacancy Survey detail by units in structure (5+ unit rental vacancy 8.2%, Q4 2024) and single-family rental vacancy (5.7%, 2024). (14) U.S. Census Bureau / HUD, Survey of Market Absorption of New Multifamily Units, Q4 2025 completions (49% three-month absorption; 90% twelve-month); HUD Request for Information, Federal Register, September 9, 2025. (15) CoStar News reporting on lease-up definition (below 90% occupancy or open less than 18 months) and stabilized-versus-overall vacancy series. (16) Yardi Matrix Multifamily National Report, Definitions (occupancy in stabilized properties; exception-property exclusions). (17) CBRE Research / CBRE Econometric Advisors, "Pre-Stabilized Properties Limit Overall Multifamily Rent Growth," June 2025 (60 bps 2015–2019 wedge; 1.5 percentage points Q1 2025). (18) Yardi Matrix Austin market reporting, December 2025; Matthews Real Estate (citing CoStar), Austin Multifamily Market Report, Q1 2026. (19) CBRE U.S. Multifamily Figures Q1 2026 (New York vacancy 2.9%); U.S. Census Bureau HVS annual MSA table (New York metro rental vacancy, 2024). (20) RealPage Market Analytics concession tracking, Q1–Q2 2026; Zillow Rental Report, April 2026 (listings-share concessions by metro); Colliers via Multifamily Dive, Q1 2026. (21) Equity Residential Form 10-Q disclosures (residential bad debt ~1.0% of rental income, 2025); Essex Property Trust and Camden Property Trust supplemental disclosures; MAA investor materials (competitor concessions ~1.25 months). (22) Appraisal Institute, The Dictionary of Real Estate Appraisal and The Appraisal of Real Estate (effective gross income; vacancy and collection loss). (23) Fannie Mae Multifamily Selling and Servicing Guide (minimum 5% of GPR combined vacancy, concessions, bad debt); Freddie Mac Multifamily Seller/Servicer Guide §19.2; HUD MAP Guide Ch. 7 (7% market-rate minimum). (24) MMCG Invest worked underwriting example; assumptions and full arithmetic available on request. (25) U.S. Census Bureau, HVS Q1 2026 Source and Accuracy (standard error 0.150; 90% MOE ±0.247). (26) U.S. Census Bureau, HVS Q4 2025 release, February 3, 2026 (October 2025 collection suspension; widened margin of error). (27) CoStar Group, Market Analytics product materials and SEC Form 10-K business description (research operation; monthly rent updates). (28) Yardi Matrix, "How We Report Rental Market Conditions" (telephone survey as a prospective renter; monthly update cadence). (29) Yardi terms of use (content provided "AS IS"; no warranty of correctness, accuracy, or reliability). (30) NYU Furman Center Housing Solutions Lab / Local Housing Solutions, "Interpreting Rental Vacancy Rates for Small and Midsize Cities" (Roanoke, VA example; 70+ cities with 5-year MOE ≥3 percentage points); U.S. Census Bureau ACS publication thresholds. (31) National Council of Housing Market Analysts, Model Content Standards (September 2025); HUD Aggregated USPS Administrative Data on Address Vacancies; USDA Rural Development Section 515 Multi-Family Housing property datasets.