A plain-English guide to payroll revisions: the first, second and third estimates, the 122,000 confidence interval, and why a plan fixes its line on the first print.

Written Friday November 6, 2026, accompanying the Friday November 6 analysis note, which graded a pre-registered plan against the October payrolls first print of +165,000 and recorded that August was revised from +133,000 to +186,000 and September from +29,000 to +60,000 in the same release. Every Employment Situation carries three payroll numbers that can move: the new month, and the two months before it. This guide explains what the first, second and third estimates are and what gets folded in between them, how large the revisions usually are on the BLS's own record, what the 122,000 confidence interval around the monthly change means for a line drawn at +100,000, how the annual benchmark differs from the monthly revisions, and why a plan on this site fixes its thresholds on the first print rather than on a number that will not exist for two months.

Three estimates of one month

The payroll figure comes from the Current Employment Statistics survey, a sample of establishments that report their payroll counts for the pay period including the 12th of the month. Not every establishment reports in time for the first release. In the BLS's words, "estimates for the most recent 2 months are based on incomplete returns; for this reason, these estimates are labeled preliminary," and "the prior 2 months are routinely revised to incorporate additional sample reports" along with "recalculated seasonal adjustment factors." A month's estimate is published three times: a first estimate on the first Friday after the reference month, a second a month later and a third a month after that, and "it is only after two successive revisions to a monthly estimate, when nearly all sample reports have been received, that the estimate is considered final" (BLS, Employment Situation technical note, last modified November 6, 2026).

Two things change between estimates, and they are different in kind. The first is more data: late reports arrive and replace the imputation that stood in for them. The second is the seasonal adjustment itself: "new seasonal factors are used each month to adjust the three most recent monthly estimates" (same technical note), so a month's seasonally adjusted change can move even if its unadjusted count did not. The BLS publishes both adjusted and unadjusted revisions, and the table below is the adjusted series, which is the one the market trades.

How large the revisions usually are

The BLS keeps a running table of first, second and third estimates for every month since 1979 (BLS, Nonfarm payroll employment: revisions between over-the-month estimates, last modified November 6, 2026). For 2026 so far, in thousands, seasonally adjusted:

Month	First	Second	Third	Third minus first
January	130	126	160	+30
February	-92	-133	-156	-64
March	178	185	214	+36
April	115	179	148	+33
May	172	129	63	-109
June	57	20	31	-26
July	-23	21	-10	+13
August	162	133	186	+24
September	29	60	(December 4)
October	165	(December 4)	(January)

Source: the BLS revisions table above; these monthly figures do not reflect later benchmark or seasonal-factor changes. The same page gives the average absolute revision since the 2003 sample redesign as 33,000 from first to second estimate, 34,000 from second to third and 51,000 from first to third. By that yardstick Friday's revisions were ordinary: September's +31,000 (29 to 60) is about the average second-estimate revision, and August's net +24,000 (162 to 133 to 186) is below the average first-to-third. May is the year's cautionary row: a first print of +172,000 that ended at +63,000. The sign of the revision is not predictable from the first print; this year four of the eight finished months revised up and four down.

The 122,000 confidence interval

The revisions are one kind of error, the gap between an early estimate and a later one from the same sample. Sampling error is another, the gap between any sample estimate and the count the full population of establishments would give. The technical note quantifies the second: "there is about a 90-percent chance, or level of confidence, that an estimate based on a sample will differ by no more than 1.6 standard errors from the true population value because of sampling error," and "the confidence interval for the monthly change in total nonfarm employment from the establishment survey is on the order of plus or minus 122,000." The BLS's own illustration: a reported change of +50,000 has a 90 percent interval of minus 72,000 to +172,000, and "since this range includes values of less than zero, we could not say with confidence that nonfarm employment had, in fact, increased that month."

Apply that to Friday. A first print of +165,000 carries a 90 percent interval of roughly +43,000 to +287,000. The plan's firm line was +100,000, which is inside the interval, and its soft line was zero, which is outside it. So the print was firm by the plan's table, and the plan can say with the BLS's confidence that payrolls rose, but it cannot say that the true change exceeded 100,000. That is not a defect in the plan. The plan's row is a classification of the published number, chosen before the release so the grade cannot be argued afterwards; it is not a claim about the population. What the interval forbids is a sentence like "payrolls grew strongly in October," stated as a fact about the economy from a single first print. The analysis note does not write that sentence.

The annual benchmark is a different animal

The monthly revisions improve a sample estimate with more of the same sample. Once a year the BLS does something else: "the sample-based estimates from the establishment survey are adjusted once a year (on a lagged basis) to universe counts of payroll employment obtained from administrative records of the unemployment insurance program." The gap between the March sample estimate and the March universe count "is known as a benchmark revision, and serves as a rough proxy for total survey error." On the BLS's record "the absolute average benchmark revision for total nonfarm employment over the prior 10 years is 0.2 percent," with a range of minus 0.4 to plus 0.3 percent (technical note). On a payroll level of about 159.3 million, 0.2 percent is roughly 320,000 jobs, an arithmetic illustration and not a forecast of the next benchmark. The benchmark also re-anchors the birth-death model that estimates jobs at establishments too new to be in the sample; the NFP anatomy guide covers that model and the February timing of the benchmark's incorporation, and the data revisions guide covers how revisions work across series.

The distinction matters for reading a release. A monthly revision of +31,000 to September is the sample filling in. A benchmark revision of several hundred thousand, spread across a year, is a statement that the sample was drifting from the population. The first kind happens every month and should surprise no one; the second is the one that rewrites a year's narrative, and it arrives on a known date rather than with the monthly print.

Why the plan fixes its line on the first print

1. The first print is the number the market trades at 8:30. The second and third estimates arrive one and two months later, after the grade has been published. A line on the "final" number could not be graded on the day, and a grade that waits two months invites regrading.
2. The lines are fixed levels, not distances from the prior month. The Friday plan's rows were at or below zero and at or above +100,000, written as levels, so the +31,000 revision to September could not move them. The data revisions guide has the general rule; the Nov 6 plan is the live case.
3. The band's width is set from first prints for the same reason. The plan set its width from the spread of the last three first prints (minus 23,000, plus 162,000, plus 29,000, a span of 185,000). The same three months on their latest estimates read minus 10,000, plus 186,000 and plus 60,000, a span of 196,000. Using revised values would mean the plan's own width changes every month as old months are revised, which is the moving-goalpost problem in a different coat.
4. The revisions are recorded, not graded. The +84,000 to August and September is in the note's data list with its source. It is not folded into the classification, and no line is drawn on it. If a future plan wants to grade revisions it will say so in advance and state a line.

What the revisions do not tell you

• Where October will settle. Upward revisions to August and September say nothing about the direction of October's own revisions in December and January; the 2026 table shows no pattern a reader could have used in advance.
• Whether the economy is stronger than the first prints said. The three-month average of the latest estimates (186,000, 60,000, 165,000) is about 137,000; on the first prints it was about 119,000. Both are inside each other's sampling error.
• Anything about the agency's motives. Revisions in both directions are a property of a sample that fills in over time, documented by the BLS itself.

A checklist for reading the revisions line

• Quote the first print as the first print, and say which estimate every other number is.
• Compare the revision with the BLS averages (33,000, 34,000 and 51,000); a revision near them is routine.
• Put the 122,000 interval around the headline before writing a sentence about the economy.
• Separate the monthly revision from the benchmark; only the second is a statement about survey drift.
• Check whether your thresholds were levels fixed before the release; if they were distances from a prior that just moved, say so.

Where this shows up in practice

• November 6 (Friday): October payrolls +165,000, firm, with August and September revised up a combined 84,000 (the live example)
• October 2 (Friday): September payrolls first printed +29,000 (now +60,000 on the second estimate)
• August 7: July payrolls first printed minus 23,000 (now minus 10,000 on the third estimate, after a second estimate of plus 21,000)

Related reading

• The paired analysis: Friday November 6
• The NFP report
• Data revisions
• ADP vs nonfarm payrolls
• The no-call band

Nothing on this site is investment advice or a recommendation to trade. Most retail traders lose money.