Once Ground Is Broken, Housing Units Leave a Rate and Join a Count

A monthly housing report arrives as one headline, and the figure that travels is usually the one carrying the largest percentage. In the July 2026 New Residential Construction release from the U.S. Census Bureau and the Department of Housing and Urban Development, published August 18, that figure was a 12.4 percent monthly decline in privately-owned housing starts, to a seasonally adjusted annual rate of 1,239,000 from a revised June estimate of 1,415,000.

Nothing in that number describes how many places anyone can move into. A start and a finished unit are separated by the length of a construction project, and the same release measures that separation in two tables that are not rates at all. They are counts of units standing at the end of the month. Once the five tables are sorted into rates and counts, the headline stops being a supply story and becomes something narrower and more useful.

What the Five Tables Measure, and Which Two Are Not Rates

The release carries five numbered tables, and they do not share a unit of measurement. Three are seasonally adjusted annual rates, which answer the question of how fast something is happening if the month repeated for a year. Two are levels: counts of units sitting in a particular state at the end of the period.

• Table 1, permits. An annual rate. July 2026: 1,443,000, which the release puts 5.0 percent above the revised June rate of 1,374,000 and 3.1 percent above the July 2025 rate of 1,400,000. Single-family authorizations ran at 894,000 and units in buildings with five units or more at 490,000.
• Table 2, authorized but not started. A level. The heading reads New Privately-Owned Housing Units Authorized, but Not Started, at End of Period. July 2026: 279,000 units, of which 149,000 were single-family and 127,000 were in buildings with five units or more.
• Table 3, starts. An annual rate. July 2026: 1,239,000 total, 808,000 single-family, 421,000 in buildings with five units or more.
• Table 4, under construction. A level. The heading reads New Privately-Owned Housing Units Under Construction at End of Period. July 2026: 1,262,000 units, of which 579,000 were single-family and 666,000 were in buildings with five units or more.
• Table 5, completions. An annual rate. July 2026: 1,212,000 total, 878,000 single-family, 329,000 in buildings with five units or more.

The distinction is not a pedantic one, because it has an arithmetic consequence. A level measured in units divided by an annual rate measured in units per year yields a number whose unit is years. That single operation is the only bridge in the release between the count of unfinished buildings and the pace at which they are being finished, and it is available in every month without any modelling.

The Interval the Release Prints Beside Each Change

Table 1 stands apart from the other four, and the release says so in its own explanatory notes. It states that "The statistics in Table 1 in this release are based on a non-probability sample and not subject to sampling error." That is why the permit figures appear with no tolerance attached, while the starts and completions figures carry one.

The release also explains how to read that tolerance: "Whenever a statement such as '2.5 percent (±3.2 percent) above' appears in the text, this indicates the range (-0.7 to +5.7 percent) in which the actual percentage change is likely to have occurred." And it defines the asterisk it prints beside certain changes: "* The 90 percent confidence interval includes zero. In such cases, there is insufficient statistical evidence to conclude that the actual change is different from zero."

Applied to July, that footnote sorts the month's six headline comparisons into two piles. The 12.4 percent monthly decline in total starts carries an interval of ±9.5 percent, so the range runs from 21.9 percent down to 2.9 percent down and does not include zero. The same is true of the year-over-year comparisons: total starts 13.5 percent below July 2025 with an interval of ±11.0 percent, and total completions 16.8 percent below July 2025 with an interval of ±13.7 percent. Those three changes clear their own tolerance.

The other three do not, and the release marks each with an asterisk. Single-family starts are printed 9.9 percent below the revised June figure with an interval of ±10.4 percent. Total completions are printed 9.1 percent below revised June with an interval of ±10.2 percent. Single-family completions are printed 5.8 percent below revised June with an interval of ±8.7 percent. By the release's own rule, none of those three monthly moves can be distinguished from no change at all. This is the same reading problem that shows up whenever a statistical agency publishes a point estimate and its own uncertainty in the same line, as with the payroll benchmark discussed in the September note on how benchmark revisions land on one month.

The Stock That Barely Moved While the Flows Swung

Table 4 is where the supply question actually lives, and through the first seven months of 2026 it hardly moved. The count of units under construction at the end of the month ran 1,270,000 in January, 1,270,000 in February, 1,262,000 in March, 1,268,000 in April, 1,252,000 in May, 1,264,000 in June and 1,262,000 in July. From January to July the level fell by 8,000 units, about 0.6 percent of the series.

Set that beside the flow. Between June and July the starts rate moved from 1,415,000 to 1,239,000, a swing of 176,000 units per year, while the total count of units under construction moved 2,000 units. The two numbers are not comparable in any direct sense, which is precisely the point: the monthly percentage that gets quoted belongs to a rate, and the thing that determines how many units finish this quarter is a count that the same percentage barely touches.

The composition moved a little more than the total. Single-family units under construction ran between 578,000 and 591,000 across the seven months, a spread of 13,000 or about 2 percent. Units in buildings with five units or more ran between 647,000 and 674,000, a spread of 27,000 or about 4 percent. In July the larger-building segment stood at 666,000 units, above the single-family count of 579,000, which is the reverse of the starts table, where single-family units outnumbered larger-building units 808,000 to 421,000 at annual rates. The completion rate for larger buildings ran 386,000 a year in June against 329,000 in July, and because the release prints no confidence interval for that pair, the distance between those two rates is arithmetic rather than a measured decline.

How Long the Pipe Is

Dividing the level by the annual rate gives a clearance time. At July's completion rate, the 1,262,000 units under construction represent about 12.5 months of output. The two segments sit far apart. Single-family: 579,000 units standing against a completion rate of 878,000 a year, or about 7.9 months. Buildings with five units or more: 666,000 units standing against a completion rate of 329,000 a year, or about 24.3 months.

Those ratios can be checked against a measured duration. Census publishes the average number of months from start to completion for buildings started in permit-issuing places. In 2022, the most recent year in that table, the average across all buildings was 8.3 months. Single-family houses built for sale averaged 7.6 months, contractor-built houses 10.1 months and owner-built houses 12.1 months. Buildings with 2 to 4 units averaged 17.0 months, 5 to 9 units 15.7 months, 10 to 19 units 15.5 months and 20 units or more 18.1 months. The pattern is stable in earlier years: the all-buildings average was 6.7 months in 2018, 7.0 in 2019, 6.8 in 2020 and 7.2 in 2021.

The single-family comparison lines up well. An implied 7.9 months sits between the 7.6-month built-for-sale average and the 8.3-month all-buildings average. The larger-building comparison does not: an implied 24.3 months runs well above the 18.1 months measured for buildings of 20 units or more. A ratio of a current stock to a current completion rate is not a build-time estimate, and the two will only agree when the stock accumulated at roughly the pace it is now being drawn down. The release does not decompose the gap, and two readings remain open: build times longer than the 2022 averages, or a stock that accumulated faster than it is now being drawn down. Nothing in the release separates them.

The Backlog Sitting in Front of the Pipe

Table 2 counts units that have cleared permitting but have not broken ground. In July 2026 that backlog stood at 279,000 units, up from a revised 268,000 in June, with single-family units at 149,000 against 144,000 and units in buildings with five units or more at 127,000 against 121,000. Measured against July's starts rate of 1,239,000 a year, the backlog equals about 2.7 months of starts.

This is where a permit number and a start number come apart. Permits ran at 1,443,000 in July against starts of 1,239,000, and the release attaches no confidence interval to the permit figure because, by its own note, Table 1 is not subject to sampling error. A rising not-started backlog alongside a falling starts rate is consistent with authorizations accumulating faster than ground is being broken, which is an observation about sequencing rather than about demand. Whether a permit becomes a start is decided later and under different conditions, the same kind of gap that separates two published estimates of one quantity in the August piece on the GDP and GDI discrepancy.

Where This Doesn't Apply

The lag argument is weakest exactly where single-family building dominates. At a measured 7.6 months for built-for-sale houses, a change in single-family starts reaches completions inside three quarters. In a market or a month where the single-family share is high, treating starts as a distant signal understates how quickly it arrives.

The clearance ratios are arithmetic, not forecasts. Dividing 666,000 by 329,000 assumes July's completion rate holds for two years, which no part of the release claims. The ratio is a way of stating the current relationship between a stock and a flow, and it changes every month that either one moves.

The accounting identity does not close on seasonally adjusted data. If the count under construction rose by starts and fell by completions and nothing else, June to July should have added about 2,250 units, since the gap between the two annual rates, 1,239,000 less 1,212,000, works out to roughly 2,250 units in a single month. The count instead fell by 2,000, from 1,264,000 to 1,262,000. The series are seasonally adjusted independently, so a residual of a few thousand units is a property of the adjustment rather than a finding about buildings.

The duration table is dated. Its most recent year is 2022, and build times since then are not in it. Any comparison between a 2026 stock-to-flow ratio and a 2022 average duration carries that four-year distance, and the 2022 figures were themselves elevated relative to 2018 through 2021.

Finally, the three monthly changes flagged with an asterisk stay flagged. Nothing in the pipeline arithmetic converts a change that fails its own confidence interval into a measured decline, and none of the reasoning here treats single-family starts, total completions or single-family completions as having fallen between June and July.

What to Watch Next Week

Census has scheduled the August 2026 New Residential Construction release for September 17, 2026 at 8:30 a.m., with revised building permits for August following on September 24. Four things in that release are checkable against this month's reading.

• The revision to July's 1,239,000 starts rate. June was revised before July's headline was computed. The size and direction of July's revision determines whether the 12.4 percent decline survives in the form it was published.
• Whether the total starts change clears its interval again. A second consecutive monthly decline that excludes zero is a different signal from one that does not, and the release does the arithmetic in its own text.
• Whether units under construction leave the 2026 range. The total has held between 1,252,000 and 1,270,000 all year. A move outside that band would be the first month in which the stock, rather than the rate, carried the news.
• Whether the not-started backlog keeps building. It rose from 268,000 to 279,000 between June and July. A third and fourth month of increase would put more weight on the gap between authorization and groundbreaking.

The larger-building completion rate is worth a separate line, for the reason given earlier: the release prints no confidence interval for it, so a second month of movement in that series still would not arrive with a test attached.

Concrete Framework

A repeatable procedure for reading this release, or any release that mixes rates with counts.

1. Sort the tables before reading any percentage. Write down which series are seasonally adjusted annual rates and which are end-of-period levels. In this release the split is Tables 1, 3 and 5 against Tables 2 and 4.
2. Read the interval before the point estimate. If the release prints an asterisk, record the change as not distinguishable from zero and do not carry it forward into any later sentence.
3. Keep the comparison windows apart. A change against the prior month and a change against the same month a year earlier are different measurements with different intervals. In July 2026 they landed on opposite sides of the asterisk rule for total completions.
4. Divide the level by the annual rate. Units divided by units per year gives years. Multiply by twelve for months. Do this separately for each structure type, because the segment answers differ by a factor of three.
5. Compare the implied clearance time with a measured duration. Use the published start-to-completion averages, note the year they cover, and treat any gap as a question rather than as a result.
6. Reconcile the stock change against starts less completions. Expect a residual on seasonally adjusted series and treat a small one as an adjustment artifact.
7. Log the prior month's revision every time. The revised base is what the new percentage is computed against, and a revision moves the headline without anything happening on any site.

Read this way, the July release says something narrower than the headline. The rate at which ground is being broken fell by an amount that clears its own margin of error. The count of unfinished units barely moved, and at the current pace of completions it stands for about 12.5 months of supply, roughly 7.9 on the single-family side and roughly 24.3 on the larger-building side. Only the second of those sentences is about how many places will exist to live in. The length-of-time series and the Survey of Construction methodology set out how the sample is drawn and how the durations are compiled, and a similar gap between what a headline index is named and what it measures runs through the September note on the trade services line in producer prices.