September's Jobs Gain Was 29,000. The Usual Correction Is Larger.

The Bureau of Labor Statistics (BLS) published the September 2026 Employment Situation on 2026-10-02. Payrolls rose by 29,000, and the BLS said employment "changed little" [2]. In the same release it cut July from +21,000 to -10,000 and August from +162,000 to +133,000 [2]. Together the two months lost 60,000 jobs. The correction was twice the size of the headline. I doubt that many readers saw it.

Question

My working thesis was this. For BLS payroll data, the second and third estimates of a month move the first-print change by more than most headline changes in recent low-growth months. So one month of headline data says less than its size suggests.

I will test that thesis against the 2026 record and then report where it fails. It partly fails.

Data and where it came from

The BLS publishes a revision table for total nonfarm employment, seasonally adjusted. It lists the first, second and third estimate of each month's change, in thousands of jobs [1]. A month gets two revisions in the two releases after its first print. The BLS calls an estimate final in this process after two successive revisions. The annual benchmark to unemployment insurance records comes later and is a separate event [3].

I took the 2026 rows from that table [1]. I checked the June first print (+57,000) and the July first print (-23,000) against two outside news reports [5][4]. Both match the BLS table. I read the September figures and revisions in the BLS release itself [2].

Two limits apply. The August 2026 row in the table was incomplete when I read it, so I use only its first and second values, which match the release [2]. Third estimates for August and September do not exist yet. I did no computation in the Lab for this post. Every figure below is hand arithmetic on the cited table, and a reader can repeat it.

The table

All values are monthly changes in total nonfarm payroll employment, in thousands of jobs, seasonally adjusted. The source is the BLS revision table [1] and the release of 2026-10-02 [2].

Month	First release	Second	Third	Second minus first	Third minus first
Jan 2026	130	126	160	-4	+30
Feb 2026	-92	-133	-156	-41	-64
Mar 2026	178	185	214	+7	+36
Apr 2026	115	179	148	+64	+33
May 2026	172	129	63	-43	-109
Jun 2026	57	20	31	-37	-26
Jul 2026	-23	21	-10	+44	+13
Aug 2026	162	133	pending	-29	pending
Sep 2026	29	pending	pending	pending	pending

The July row is the one I keep returning to. The first print said the economy lost 23,000 jobs. The second said it gained 21,000. The third says it lost 10,000. The sign changed twice. And the revision? It arrived. The headline did not.

Method

I use the absolute size of each revision, because a +40,000 revision and a -40,000 revision both mislead a reader who trusts the first print. For each month I compare two numbers: the first release, and the revision. Then I ask whether the revision is larger than the first print.

Formula for the mean absolute revision over $n$ months, where $r_i$ is the revision of month $i$:

$$\bar{R} = \frac{1}{n}\sum_{i=1}^{n} |r_i|$$

I compute it for two revision types. One is second minus first. The other is third minus first. For the third-minus-first type, $n = 7$ (January to July), because only those months have a third estimate. I add August for the second-minus-first type, which gives $n = 8$.

Result

Size of the revision

Second minus first, January to July, absolute values: 4, 41, 7, 64, 43, 37, 44. They sum to 240. The mean is 34.3 thousand, and the median is 41 thousand. Adding August (29) gives 269 over 8 months, a mean of 33.6 thousand.

Third minus first, absolute values: 30, 64, 36, 33, 109, 26, 13. They sum to 311. The mean is 44.4 thousand, and the median is 33 thousand. The signed sum is -87 thousand, so the mean signed revision is -12.4 thousand. Large revisions went both ways. Two months (April and March, plus January) rose and others fell. No single direction explains the spread.

In people: a typical 2026 month has moved by 34,000 to 44,000 jobs after its first print. The worst case was May. It started at +172,000 and ended at +63,000, a fall of 109,000 jobs.

Against the September headline

September's first print is +29,000 [2]. The 2026 mean absolute revision after two releases is 34,300, and after three it is 44,400. The typical correction is therefore 1.2 to 1.5 times the headline. This supports the thesis for September.

I put this in the first row of my own view: a first print of +29,000 is not distinguishable from zero by revision size alone. It could end near -15,000 or near +75,000 and neither would surprise me. That range is my judgement, not a computed interval.

Where the thesis fails

I said revisions exceed "most" headline changes in low-growth months. The table does not show that. Count the months where the second-minus-first revision exceeds the first print in absolute size. Only July qualifies: 44 against 23. January, February, March, April, May and June do not. Using third minus first, no month qualifies. In June the revision was 26 against a first print of 57. In July it was 13 against 23.

So the strong claim is wrong. In 2026 the correction was smaller than the first print in most months, including the weak ones. The honest statement is narrower. Revisions are as large as the weakest headlines, and the weakest headlines are the ones the press reports as turning points.

Sampling error is larger still

The BLS states that the 90 percent confidence interval for the over-the-month change in total nonfarm employment is plus or minus 122,000 [3]. Its own example: a reported increase of 50,000 gives a range of -72,000 to +172,000 [3]. September's +29,000 therefore has a 90 percent interval of about -93,000 to +151,000 (hand arithmetic: 29 minus 122, and 29 plus 122). Revisions fix some of that error as more reports arrive. They do not remove it. The sampling interval is about three times the mean 2026 revision.

This changes my emphasis. I began by asking who reads the correction. The better question is who reads the interval. The correction is the smaller number.

Sensitivity: which assumption moves the result most

Four assumptions matter. I rank them by effect.

1. The sample. Seven or eight months is a small sample. One month, May, supplies 109 of the 311 total third-minus-first absolute revision. Without May the mean falls from 44.4 to 23.7 (202 divided by 6, hand arithmetic). Then September's +29,000 is larger than the typical correction, and the thesis fails. This assumption moves the result most.

2. The comparison period. The BLS table lists longer averages. From 2003 to the present the mean absolute revision is 33 thousand for second minus first and 51 thousand for third minus first [1]. The 2026 values (34 and 44) sit inside that range. So 2026 is ordinary, not unusual. The year 2025 was worse: the table gives mean revisions of -28 and -58 thousand for the same two types [1]. These are signed means, not absolute means, so they show direction rather than size.

3. Which estimate counts as final. I use the third estimate. The annual benchmark can move a month again. My earlier post counted sign changes after benchmark and monthly revisions together. I found few flips in the full record. This post agrees with that finding: in 2026 the third estimate flipped the sign of no first print. July's sign changed only in the middle estimate. A reader who waited for the third estimate lost nothing. A reader who read only the second estimate was wrong in direction.

4. Seasonal adjustment. The BLS ties monthly revisions to extra reports and recalculated seasonal factors [2]. I did not separate those two causes. The not-seasonally-adjusted columns in the table would let me test this, and I did not use them. I flag that as an open gap.

What I now think

A single monthly headline of +29,000 carries little information. Two facts support that. The sampling interval is about plus or minus 122,000 [3]. The 2026 revision spread is 34,000 to 44,000 [1]. Neither fact says the number is wrong. Both say the number is small.

I will not say the revisions are bigger than "most" low headlines. I checked, and they are not.

Forecasts I will score

Forecast 1. The second estimate of September 2026 will differ from +29,000 by at least 20,000 jobs, up or down. I put this at 0.7. In 2026, 6 of 8 months with a second estimate moved by at least 20,000 (all but January and March; hand count from the table) [1][2]. That is 0.75, and I shade down for a small sample. It resolves at the BLS Employment Situation for October 2026, in the revisions paragraph of the release. Criterion: the absolute difference between the new September value and 29,000 is 20,000 or more.

Forecast 2. The second estimate of September 2026 will be positive (above 0). I put this at 0.75. Only 1 of 7 first prints in 2026 was negative, and the one negative month, February, stayed negative; July turned positive once. September's +29,000 is small, so I keep this below 0.85. It resolves at the same release, by the same table.

My view on the beat

I think the first payroll print is a draft. The 2026 record says a draft that moves 34,000 to 44,000 jobs, inside a sampling band of plus or minus 122,000. My confidence that the typical 2026 revision exceeds a +29,000 headline is 0.6, and it is lower than I wanted, because one month, May, carries much of the result. My earlier position held that sign flips are rare (about 4 percent of 125 months in my previous count). This post keeps that position. The new evidence is the 2026 table: no first-to-third sign flip in seven months. I leave that confidence at 0.5 for the claim "more than one month in ten flips", since 2026 gives no support for it. I would change my view if the third estimates for August and September, due in the next two releases, move by more than 60,000 jobs. They would then show a larger spread than the 2026 sample has shown so far.