Monthly jobs-number revisions typically run ±30,000-60,000; the annual benchmark has recently skewed sharply negative

Summary

Every monthly jobs report gets revised twice more before it's final, and then again once a year in a larger "benchmark" revision. A single month's first revision (report 1 to report 2) has historically averaged about 32,600 jobs in either direction — sometimes up, sometimes down, with no strong long-run bias either way over 1963-2019. The bigger, once-a-year benchmark revision (which realigns the survey-based estimate to near-complete unemployment-insurance tax records) has averaged close to zero over 1980-2025 but recent years have broken that pattern: preliminary benchmark revisions nearly tripled in average size after 2020, and the last two annual benchmarks have both been unusually large and downward — a preliminary -911,000 jobs for the year through March 2025 (the second-largest downward benchmark revision on record, after the 2009 recession) and -598,000 for the year through March 2024. So: revisions are normally modest and roughly balanced between up and down, but the most recent vintage has leaned decisively toward the initial numbers being overstated.

How the revision process works

BLS's Current Employment Statistics (CES) survey samples about 121,000 businesses/agencies (roughly one-third of all nonfarm jobs) each month. Because responses trickle in over time, each month's payroll number gets revised twice as more surveys come back — response rates run about 55% at the first release, ~91% by the first revision, and above 93% by the second (final monthly) reading. Once a year, BLS also "benchmarks" the March estimate to near-universe counts of employment drawn from unemployment-insurance tax records (the same source as the Quarterly Census of Employment and Wages, QCEW), which corrects for sampling error and for the model BLS uses to estimate new-business "births" it can't yet see in survey data. That benchmark revision, announced each following February, ripples back through about 21 months of not-seasonally-adjusted data and, combined with re-estimated seasonal factors, affects the seasonally adjusted series over an even longer span.

Typical size and direction: month-to-month revisions

Using real-time vintages of the payroll series (Philadelphia Fed and St. Louis Fed archives), the San Francisco Fed calculates that the average absolute first revision to the monthly employment-gain figure was about 32,600 jobs over 1990-2019. The distribution is close to symmetric around zero — in the bin-by-bin breakdown of revisions since 1963, roughly 38% of revisions fall in clearly negative bins and about 42% in clearly positive bins, with the rest straddling zero — so on any given month a revision is about as likely to raise the initial estimate as to lower it.

The Cleveland Fed's more recent analysis (1980-2025, n=551 months) distinguishes two revision concepts: the benchmark revision (third monthly reading vs. the latest, benchmark-adjusted reading) averaged -4,510 jobs/month (median -2,000), while the cumulative revision (first release vs. latest estimate, combining monthly and benchmark effects) averaged +6,840 jobs/month (median +9,000) — i.e., over the long run the two effects have roughly offset, with a slight net upward drift once monthly revisions are included, but a mild downward drift in the annual benchmark component alone. Neither Cleveland Fed's structural-break tests (Bai-Perron, Chow) find a statistically significant regime change in this series over 1965-2025 or 1980-2025 — the process itself hasn't fundamentally changed even though outcomes have gotten more volatile lately (see below).

The annual benchmark revision: normally small, recently unusual

BLS and outside analysts treat -0.5% to +0.5% of the reported employment level as the "normal" range for the annual benchmark. Selected benchmark revisions to the March level: 2005 -158,000 (-0.1%); 2007 -293,000 (-0.2%); 2010 roughly -0.3%; 2024 (final) -598,000 (-0.4%, after a larger -818,000 preliminary estimate); and 2025 (preliminary, announced September 2025) -911,000 (about -0.6% of the level) — the second-largest downward benchmark revision on record behind the 2009 financial-crisis benchmark, and just outside BLS's historical "normal" band per the Cleveland Fed's own reading of the data (-0.54%). Reporting on the 2025 vintage also found the labor market showed outright monthly contraction in four separate months (January, June, August, October) once revised, cutting total 2025 job growth to roughly 181,000 for the year.

Separately, the average absolute size of the preliminary annual benchmark revision was about 0.12% of employment in 2016-2020 versus about 0.34% in 2021-2025 — roughly a threefold increase — a pattern researchers attribute mainly to declining CES survey response rates, which make the sample-based initial estimate less reliable relative to the administrative (UI tax) data used for the benchmark.

Which direction do revisions "usually" go?

• Long run (decades): roughly balanced/symmetric — first monthly revisions are about as likely to be positive as negative, and the Cleveland Fed's 1980-2025 sample shows a near-zero average with a mild net-positive cumulative effect once all revision stages are combined.
• Recent years (2023-2025): the pattern has leaned decisively downward — both of the last two annual benchmark revisions cut employment growth by hundreds of thousands of jobs, and 2025's monthly revisions turned four separate months of reported gains into outright job losses.
• Because past benchmark revisions are statistically correlated with future ones (a serial-correlation finding going back to Haltom, Mitchell & Tallman 2005 and reconfirmed by the Cleveland Fed through 2025), a run of downward revisions has some tendency to predict further downward revisions in the near term, though this is a modest statistical regularity, not a guarantee.