Building Easier AI
One Month Is Not a Trend: Read the Margin Before You Act
On 24 September, Census reported new home sales up 6.4% in August, to a seasonally adjusted 684,000 a year. Beside the 6.4% it printed ±19.5% and an asterisk. The footnote: "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" (Census release).
A month earlier, July's first print of 607,000 was "the slowest since January of this year," per NAHB on 25 August. July now stands at 643,000. A builder who cut a crew on that headline cut it on a number since revised away.
The release's explanatory notes add an instruction: "It takes 4 months to establish a trend for new houses sold." The agency that produces the number is telling you how little one month proves.
None of this is peculiar to houses. A fabricator's orders, a service company's calls and a remodeler's signed contracts are exact counts, yet one signature landing on the 1st instead of the 31st moves them. Before you hire, cut a price, buy a lot or order ahead on one month, there is a check that takes an afternoon — and some numbers pass it.
Census prints the margin beside the change
Reading a Census margin takes one subtraction and one addition. For August's +6.4% (±19.5%), the 90 percent interval on the change in new home sales runs from a fall of 13.1% to a rise of 25.9% (our arithmetic). The year-on-year move, −2.0% (±15.7%), fails the same test, and so does the median price: $393,700, down 5.8% on the year against a margin of ±8.2% (Census release).
The survey is also revised. Census says the preliminary estimate of total sales "is revised about 6.7 percent" on average (Census release). July's move from 607,000 to 643,000 was +36,000, or 5.9% (our arithmetic): smaller than the average revision. An ordinary revision still moved the number an owner would have acted on by 36,000 homes a year.
Starts and permits come with their own instructions. The construction release of 17 September says "It may take three months to establish an underlying trend for building permit authorizations, six months for total starts, and six months for total completions" (Census, New Residential Construction).
This newsletter printed a starts change as news a week ago. Issue 06 ran single-family starts at "▲ 7.6% M/M" in its numbers strip and led Bright Spots with the rebound, +7.6% on the month and +5.2% on the year. Census's margins on those two changes were ±14.0% and ±10.9%; neither is distinguishable from zero. The caveat underneath carried the changes that did clear: single-family starts down 4.7% year to date (±3.2%) and single-family completions down 22.9% on the year (±10.1%) (Census, New Residential Construction). The signal and the noise sat in one section, and the noise went on top.
A revision is a different matter from misreading noise, and it is no one's mistake. Feature 07 printed July's 9.6 months' supply as the highest since January, which is what the first print showed, per NAHB on 25 August. Census has since revised July to 9.0, and February (9.3) and May (9.2) now sit higher (Census release). The sentence was right the day it ran. The number under it moved, which is what a first print does.
What one month costs when you spend on it
Four decisions ride on a monthly number in a business that builds things: the crew, the lot, the price and the order. Each commits money in a day. The Census releases behind those numbers need three to six months to establish a trend, by their own account.
The crew is the expensive one to get wrong. Cut two people in late August on July's 607,000 and the saving lasts exactly as long as the slowdown does; a month later July read 643,000, and if the work comes back you are hiring again. Price that on your own books: the weeks a job runs short-handed, the ad, the first month a new hire works at part speed, the job you turn down in between. Then price the other side, carrying the same two people for the four months Census says a trend takes. That is the most waiting can cost, and it assumes they produce nothing for four months.
The other three work the same way. A lot you pass on in a slow-looking month belongs to someone else by the time the revision lands. A price cut on a finished spec, once it closes, becomes the next appraisal's comparable. An order pulled forward on a rebound is inventory you finance whether or not the rebound was real.
The jobs report a field-service owner reads is revised too, and once a year BLS benchmarks it to "comprehensive counts of employment" drawn mostly from unemployment-insurance tax records (BLS release). BLS's preliminary benchmark, published 28 August, revised construction employment for March 2026 up 62,000 (+0.8%): the monthly survey had been undercounting the trade. Manufacturing, the fabricator's row in the same table, went the other way, down 67,000 (−0.5%): 67,000 jobs in the monthly estimate that the benchmark says weren't there (BLS, preliminary benchmark table). Both figures are preliminary; BLS issues the final benchmark in February 2027 (BLS release).
Some numbers clear the bar
Census's two September releases also carry changes that pass its own test, and waving them away is the opposite mistake. In the construction release, total housing completions in August were down 27.1% on the year, against a margin of ±8.9%, and single-family completions fell 10.4% in a single month, at ±9.3%: a monthly move that clears. In the sales release, new home sales in the West were down 26.8% on the year (±22.9%), and the average new-home price fell 8.8% on the year (±7.7%). A builder selling new homes in the West is not reading noise.
Revisions are not always large, either. The same BLS benchmark that moved construction and manufacturing moved total employment by −79,000, or −0.1% (BLS), against a ten-year absolute average of 0.2% (BLS release).
And a price is not an estimate. The rate a lender quotes your buyer is what they will pay. Freddie Mac's 30-year figure is an average of the rates on thousands of loan applications submitted through its Loan Product Advisor, and Freddie publishes no margin of error beside it. It averaged 7.03% in the week to 24 September (PMMS), up from 5.98% in the week to 26 February and 6.65% in the week to 20 August, per Freddie Mac's weekly history. That is 105 basis points since February and 38 in five weeks (our arithmetic). Anyone who waited for the rate to establish a trend paid for the wait. Four months is the right patience for a survey estimate and the wrong patience for a price.
Put your own margin on your own count
Pick the one monthly count that moves money in your business: signed contracts, leads, service calls, orders booked, units shipped. It has its own wobble, and measuring it takes a spreadsheet and an afternoon.
Pull the thirteen months before this one, which gives you twelve month-to-month changes. Write each as a percentage and drop the sign. Then ask one question of this month's move: how many of those twelve were as big?
Small counts swing hard. At ten jobs a month, three signatures either way is 30%. A move matched several times in the twelve is the business's ordinary wobble. A move bigger than anything in the twelve has cleared your bar, the way completions cleared Census's, and it earns a plan. Before that plan touches payroll, our rule of thumb (not a statistic) is three months on the same side of your twelve-month average. Three months is the shortest window in either Census release above.
Run the test on the good months too. Staffing up for a rebound is the same bet in the other direction, and August's national +6.4% is the example.
Keep prices out of it. A supplier's quote, a lender's rate sheet and a sub's new day rate carry no margin, because they are what you will pay. Re-price the week they move.
Run it once on an illustrative remodeler, with round numbers invented for the arithmetic. They signed 10 jobs in July and 7 in August, a 30% drop. In the twelve changes before that, the count had moved 30% or more four times, and the year closed within one job of where it began. A move that size, four times in a flat year, is an ordinary month for that business: the two people they were about to let go stay on Monday's schedule.