The simple version

The Census Bureau reported on Tuesday that sales of new single-family houses ran at a seasonally adjusted annual rate of 607,000 in July, which it published as 10.5 percent below June. That percentage led the coverage.

A percentage change is a comparison, and comparisons need two numbers. The July figure is new. The June figure it is measured against is not the June figure that was published in June, because the same release revised June upward by 50,000. Against the number that was actually reported at the time, the drop is roughly a third the size.

There is a second caveat printed in the same sentence as the first. Census attaches a margin of error of plus or minus 14.0 points to that 10.5 percent and marks it with an asterisk, and the asterisk is defined in the release: the confidence interval includes zero, so there is insufficient statistical evidence to conclude the change is different from zero.

The numbers

• Sales of new single-family houses ran at a seasonally adjusted annual rate of 607,000 in July 2026 (U.S. Census Bureau and HUD, Monthly New Residential Sales, July 2026, release CB26-128, August 25, 2026)
• Census published that as 10.5 percent (plus or minus 14.0 percent) below the June 2026 rate of 678,000, marked with an asterisk meaning the 90 percent confidence interval includes zero (Census and HUD, CB26-128)
• The June rate as originally published on July 24, 2026 was 628,000 (Census and HUD, Monthly New Residential Sales, June 2026, release CB26-121)
• June was therefore revised up by 50,000 units between the two releases, an increase of about 8.0 percent (computed from the Census figures above)
• Measured against the 628,000 that June originally reported, July's 607,000 is about 3.3 percent lower rather than 10.5 percent (computed from the Census figures above)
• The gap between the two framings is about 7.1 percentage points, and all of it comes from the revised base rather than from anything in the July data (computed)
• Census states that on average the preliminary seasonally adjusted estimate of total sales is revised about 5.0 percent (Census and HUD, explanatory notes, CB26-128)
• Census defines a sale as a deposit taken or a sales agreement signed, which can occur before a building permit is issued (Census and HUD, explanatory notes, CB26-128)
• July sales were also reported 6.3 percent (plus or minus 19.6 percent) below the July 2025 rate of 648,000, a figure carrying the same asterisk (Census and HUD, CB26-128)

A percentage has two numbers in it

Every month-over-month change is a fraction: this month over last month. Attention goes almost entirely to the numerator, because that is the new information. The denominator is treated as settled history.

It usually is not. Statistical agencies publish an early estimate and revise it as more complete data arrives, and the revision appears inside the next month's release rather than as its own announcement. So the same document that delivers a new number also quietly changes the number it is being compared against.

When the prior month is revised upward, the current month's decline looks larger, because it is falling from a higher starting point. When the prior month is revised down, a decline shrinks or can become an increase. Nothing about the current month changed in either case.

That is what happened here. June was first reported at 628,000 and now stands at 678,000. July's 607,000 sits 3.3 percent below the first figure and 10.5 percent below the second. The July number did not move. The comparison did.

Why this series in particular

New home sales are unusually prone to this, and the Census Bureau says so plainly in its own materials. The sample is small relative to the total housing market, the series is volatile month to month, and Census puts the average revision to the preliminary seasonally adjusted estimate of total sales at about 5.0 percent.

The same release carries a caveat wider than the revision. The 10.5 percent decline is published with a margin of error of plus or minus 14.0 points, and the asterisk beside it is defined in the release itself: the 90 percent confidence interval includes zero, so there is insufficient statistical evidence to conclude that the actual change is different from zero.

The two caveats stack rather than cancel. The recomputed 3.3 percent decline sits inside that same band, which means the smaller figure is no more statistically solid than the larger one. What the release supports is narrower than either headline: Census measured 607,000 in July, and it cannot say with confidence how that compares to June.

None of this is a criticism of the Census Bureau, which publishes its revision practice, its methodology, and its confidence intervals openly, and prints the interval in the same sentence as the estimate. The agency is doing exactly what it documents. The gap opens between what the release says and what a headline built from it implies.

The Real Cost lens on a number that moved twice

The practical translation is about how much weight a single monthly figure can carry, which is less than it appears.

• A reported month-over-month change is measured against a base that may itself be revised again next month
• So the same July figure could be described with a different percentage in September, without July changing at all
• Census puts the average revision to this series at about 5.0 percent, which is large next to most of the monthly changes it reports
• When the published margin of error includes zero, the direction of the change is not established, let alone its size
• For anything that matters, the trend across several months carries information a single comparison cannot
• This is why revisions are published: an early estimate is a first attempt rather than a provisional lie, and the correction is the system working

For a household watching housing conditions, the useful read is direction over several months rather than the size of any one month's percentage. The percentage is the least stable number in the release.

What this means

When a monthly economic release produces a dramatic percentage, two questions resolve most of it. Was the prior month revised, and by how much. Does the margin of error printed beside the change include zero. Statistical agencies publish both answers in the release, usually within the first few paragraphs.

The habit transfers to every monthly indicator. Retail sales, durable goods, payrolls, and housing all publish early estimates and revise them. A percentage change is a claim about two numbers, and only one of them is new.

What this is NOT

This is not a claim that the Census Bureau published anything improperly; revisions are its documented, standard practice, and the agency prints its methodology, its average revision size, and its confidence intervals openly. This is not a claim that news outlets reported dishonestly; they reported the figure as released. This is not a claim that new home sales did not fall, only that the release's own margin of error does not let Census conclude that they did. This is not a prediction of housing activity, home prices, or mortgage rates. This is not advice about buying, selling, or timing any housing or financial decision, and it is not advice about any security or fund. Figures are as published on the dates stated and are subject to further revision. This is not investment or financial advice of any kind.

Sources

• U.S. Census Bureau and U.S. Department of Housing and Urban Development, Monthly New Residential Sales, July 2026 (CB26-128), released August 25, 2026: https://www.census.gov/construction/nrs/pdf/newressales_202607.pdf
• U.S. Census Bureau and U.S. Department of Housing and Urban Development, Monthly New Residential Sales, June 2026 (CB26-121), released July 24, 2026: https://www.census.gov/construction/nrs/pdf/newressales_202606.pdf
• U.S. Census Bureau, New Residential Sales: https://www.census.gov/construction/nrs/index.html
• U.S. Census Bureau, Survey of Construction methodology, which documents how the new residential sales estimates are produced: https://www.census.gov/construction/soc/methodology.html