The Field With No Error Bar Missed by Eighty Thousand. The Field With a 400,000-Unit Error Bar Beat by Fifty-Nine — and the Beat Is Not Significant.

公開: 更新: 2026/11/18 23:17 UTC
X Facebook LinkedIn

The October new residential construction report landed at 13:30 UTC on 18 November, and it settled a three-way consensus dispute this desk published at the pre-US slot of 18 November by beating every figure in it. Housing starts printed 1,309,000 at a seasonally adjusted annual rate against published consensus values of 1.23, 1.245, 1.25 and 1.27 million — a beat of between 39,000 and 79,000 depending on whose number you were holding. Then the issuer told you the beat might not be a beat at all. The Census Bureau prints the monthly change as 9.9 per cent plus or minus 16.8 per cent, with an asterisk whose footnote says the 90 per cent confidence interval includes zero. Building permits, in the same document, printed 1,295,000 against three consensus figures that agree with each other to within 5,000 — and missed all three by about 82,000, with no margin of error published at all, because the issuer says that series is not subject to sampling error. One release, two fields, and the field a system can measure its own uncertainty on is the one that moved in the direction nobody forecast.

Four consensus values, all beaten, and the issuer will not call it significant

Here is the scoreboard, with every consensus figure named to the publisher that carried it. Three of the four came from pages this desk read before the release; the fourth is a Reuters poll of economists quoted in a Reuters report.

  • 1.23 million — FinanceCalendar, read by this desk at the pre-US slot. Beaten by 79,000, or 6.42 per cent.
  • 1.245 million — a Reuters poll of economists, as quoted in a Reuters report published 16:34 UTC. Beaten by 64,000, or 5.14 per cent.
  • 1.25 million — Trading Economics’ consensus column, read after the print; also the figure a single analytics post carried before it. Beaten by 59,000, or 4.72 per cent.
  • 1.27 million — Trading Economics’ consensus column as this desk read it at the pre-US slot, before the release. Beaten by 39,000, or 3.07 per cent.

Read those last two lines again, because they are the same field at the same publisher. At 11:4x UTC on 18 November this desk recorded Trading Economics’ consensus for October housing starts as 1.27 million, with its own separate model forecast at 1.20. After the release the consensus column on that page reads 1.25 million and the model forecast still reads 1.20. The model field is byte-stable and the consensus field is not.

We are not going to tell you which direction that edit went or why, because we cannot read the page’s history and we will not invent one. What we can tell you is the operational consequence, and it is the whole point of this section: the surprise you compute depends on when you read the consensus. Score this release against that vendor before the print and the beat is 39,000. Score it afterwards and the beat is 59,000. A system that stores the consensus at release time and a system that scrapes it the following morning are scoring two different events, and neither of them is wrong about its own arithmetic.

And then the issuer undercuts all four. The 9.9 per cent monthly rise carries an asterisk, and the footnote behind that asterisk says the 90 per cent confidence interval includes zero. On the Census Bureau’s own convention that means the change is not statistically significant. The year-over-year comparison, 2.8 per cent plus or minus 16.5 per cent against the October 2025 rate of 1,273,000, carries the same asterisk. So does the single-family line: 864,000, up 4.0 per cent plus or minus 9.0 per cent from a revised 831,000. Three published increases, three asterisks, none of them significant on the publisher’s own test.

The error bar is ten times the entire consensus spread

Do the arithmetic on that margin of error in levels rather than percentages, because percentages hide the scale. The revised September base is 1,191,000. Plus or minus 16.8 per cent of that base is plus or minus 200,090 units. The full 90 per cent interval on the change runs from minus 6.9 per cent to 26.7 per cent, which in levels is a band from about 1,108,800 to about 1,509,000 — 400,180 units wide.

The four published consensus values span 1.23 to 1.27 million. That is a spread of 40,000 units. The ratio of the issuer’s confidence interval to the entire disagreement among forecasters is 10.004.

This desk carries a standing note that vendor dispersion is not an uncertainty estimate and that the ratio between the two is about ten. That note was built from consensus spread against realised surprise on other releases, which is an indirect measurement. This is the direct one, and it is the cleanest instance we have had: the issuing agency publishes its own sampling uncertainty on the same line as the number, and it is ten times wider than the forecasters’ disagreement, to three decimal places. Against the realised surprise rather than the spread, the interval is 6.25 times the 64,000 beat.

What a system should do with that is narrower than it sounds. It does not mean the print is uninformative — the single-family and multifamily components move for real reasons, and 433,000 multifamily starts against a prior 345,000 is a 25.5 per cent jump that is not noise in anybody’s telling. It means that a release blackout sized by consensus dispersion is sized by the wrong number, and that a rule which fires on “starts beat consensus” is firing on a quantity the issuer itself declines to call a change. If you want a filter on this release, build it on the components and on permits, which is where the signal that survives its own error bar actually is.

The prior was revised away inside the same release

Every one of those four consensus figures was constructed against a September housing starts level of 1,221,000. We know that because all three publishers this desk read before the print agreed the prior was 1.221 million, and one of them lists it as the actual for the September release. The October document replaces it. September starts are marked revised at 1,191,000 — 30,000 lower, a downward revision of 2.457 per cent — and the 9.9 per cent rise the headline reports is computed off the revised base.

Measure the same print against the base the forecasters were actually using and you get a different number: 1,309,000 over 1,221,000 is 7.207 per cent, not 9.9 per cent. The 2.7 percentage points between those two figures are not a data error and not a forecasting error. They are a base change that arrived in the same document as the print being scored against it.

Permits did not do this. The September permits level this desk held before the release was 1,398,000, and the revised September permits figure in the October document is 1,398,000. Same release, same table, two series: one prior revised by 30,000 and one prior untouched. There is no way to know in advance which of the two you are holding, which is why this desk’s standing rule is to check the prior for revision in the release itself rather than carrying one forward — and the useful refinement is that a release can revise one of its own series and not the other, so checking the field you care about is not the same as checking the release.

The document also tells you how often this should be expected. Preliminary seasonally adjusted permits, starts and completions are revised on average by 3.8 per cent or less, and the total quantity response rate behind the headline table is 76.6 per cent. A 2.457 per cent revision is therefore well inside normal, which is the uncomfortable part: nothing unusual happened, and it still moved the denominator under every forecast in the market.

Permits: three publishers inside five thousand, a miss of eighty-two, and no error bar at all

At the pre-US slot this desk published that it could not find a building permits consensus at two of three publishers and that the single figure it had, 1.375 million, was single-sourced. That absence is now closed, and it closed in the direction that makes the earlier complaint look mild rather than wrong.

  • 1.375 million — a single analytics post, read before the release. Missed by 80,000.
  • 1.377 million — attributed in a Reuters report only to “economists”, with no named poll. Missed by 82,000.
  • 1.38 million — a calendar aggregator, unattributed, and tagged by that page as a low-impact event. Missed by 85,000.

The permits consensus spread is 5,000 units — one eighth of the starts spread. Forecasters agreed almost exactly on permits and disagreed about starts, and they were wrong about permits and right about the direction on starts. The permits miss, 82,000 against the middle figure, is 1.281 times the 64,000 starts beat. The field everybody agreed on produced the larger error.

And this is the field with no margin of error. The October document publishes no plus-or-minus figure and no significance asterisk for permits, and its explanatory notes say the permit estimates in the headline table come from a non-probability sample and are therefore not subject to sampling error. That is a statement about sampling, not a claim of precision, and the distinction matters enormously to anyone building a filter: the absence of an error bar is not a tighter error bar. The series that reports its uncertainty looks noisy and the series that reports none looks clean, and the one that looks clean is where the 82,000 miss is.

The levels themselves: permits 1,295,000, down 7.368 per cent from the revised September figure and 8.674 per cent below the October 2025 rate of 1,418,000. Single-family permits 772,000, down 8.747 per cent from a revised 846,000 — the sharper of the two components. Multifamily permits 475,000, down 5.2 per cent. Completions, which nobody forecasts, printed 1,272,000, up 6.533 per cent from a revised 1,194,000, and that line carries an asterisk too. Two publishers reached for a “lowest since” framing on the permits number and reached for different series to do it — one said single-family permits were the lowest since January 2023 and the other said the overall series was the lowest since 2020. We print the levels and leave the superlatives where we found them, because they are two different objects being described with one shape of sentence.

The next release is 17 December 2026. The issuer states the date and states no time and no time zone, so the 13:30 UTC hour in anybody’s calendar row came from a vendor. Worth noting for anyone with a long schedule: the document says that from the January 2027 release, due 18 February 2027, the Survey of Construction completes a sample redesign, and estimates for non-permit-issuing areas are discontinued. Those areas are under 1 per cent of the estimates, so the level effect is small — but the series definition changes on a dated boundary, and a model trained across it will not be told.

What this does not tell you

  • We cannot show you the consensus edit. We hold our own note that a vendor’s consensus column read 1.27 before the print and we read 1.25 after it. We have no page history, no timestamp on that field, and no way to distinguish a correction from a late survey update from a rendering difference. The finding is that the field is not stable, not that anyone changed it improperly.
  • Two of the five pages we went to for consensus values would not open. Two wire-service pages returned client errors, and one of them has the expected starts figure in its own URL slug. We did not publish anything from a slug. There may be consensus values we did not see, and if there are, the 40,000 spread is a floor rather than the full range.
  • The 1.23 and 1.375 figures are from this desk’s own pre-release notes, read at the pre-US slot and not re-read after the print. We are scoring against what we recorded, which is the honest way to score a forecast, but it means two of the seven consensus values in this article rest on one reading each.
  • The “economists” behind the 1.377 permits figure are not named. One page attributes its starts consensus to a Reuters poll and its permits consensus to “economists” with no poll named, in the same article. We do not know whether those are the same economists.
  • We do not know the original September permits figure — only that the revised one equals what we held. If September permits were revised and happened to land back on 1,398,000, we could not tell from this document.
  • The margin of error is a sampling interval, not a total error estimate. It says nothing about seasonal adjustment error, which on a series this volatile is a live question, and nothing about the response-rate gap.
  • Nothing here is a view on housing. We have not formed one, and a release whose headline change the issuer declines to call significant is a poor place to start.

Related

Primary source: the United States Census Bureau and Department of Housing and Urban Development joint release New Residential Construction, October 2026 (Release CB26-187), read in full, 18 November 2026. Every level, margin of error, significance asterisk, revision, response rate and schedule note in this article is from that document.

Consensus and secondary readings, all 18 November 2026: a Reuters report carried at GV Wire (starts consensus, permits consensus, component changes) · Trading Economics (post-print consensus column, regional detail, the revised prior) · TipRanks (permits forecast). The 1.23 million starts consensus and the 1.375 million permits consensus are from this desk’s own pre-release notes of 18 November.

Pages we tried and could not open, so that no figure here is implied to rest on them: an InvestingLive report of the starts print and a citybiz report of the same, both returning client errors. No figure in this article comes from either, and in particular we did not take the consensus value that appears in one of those URLs.

All times UTC. Commentary and interpretation are our own. Nothing here is investment advice, a recommendation or an offer; you are responsible for your own risk.


Calendar Desk
Calendar Desk