The U.S. jobs report is one of the most closely watched economic indicators in the world. Yet the number that moves markets each month is not necessarily the number that ultimately enters the historical record. The Bureau of Labor Statistics (BLS) releases its first payroll estimate roughly three weeks after the reference month and subsequently revises it twice as additional information becomes available. In most years revisions are typically made both in the positive and negative direction. However, 2025 was an unusual year, with every final revision being lower than the initial estimate. This is the first time that all 12 months had initially overestimated the number of jobs created going back to at least 1979.
With every month eventually revised downward, broader questions have been raised about how much confidence investors and policymakers should place in economic data before the underlying activity is fully observed. It is also worth questioning how the initial numbers are covered by the media, who typically report the number as a fact versus an estimate, but that is another piece for a different time. We wanted to look at what goes into these numbers and how much faith should people place in the initial values.
The 2025 experience exposed the difficulty of measuring a changing labour market in real time. Survey responses can be incomplete, newly created businesses may not yet be visible to the statistical system, and models based on historical relationships can struggle when economic conditions shift. BLS has responded by modifying its birth-death methodology and so far, the early 2026 revision pattern is more balanced than in 2025, but it is still too early to draw a firm conclusion.
The most useful question may therefore not be whether BLS “got 2025 wrong.” It may be whether the changes introduced in 2026 can make the first number more reliable when it matters most: before the final data are available.
BLS was obviously very aware of the pattern and did a deep dive into the why this was happening. One of the big reasons is definitely more in the wonky category (persistent errors in birth-death forecasts) the other big reason is definitely less economic speak but arguably more troubling: response bias and non-response bias.
There is less the BLS can do about the latter, but beginning with the January 2026 estimates, BLS modified the ARIMA component (ARIMA is an acronym for Autoregressive, Integrated, and Moving Average) of its birth-death model to incorporate current survey information into the forecast. Since then the revisions have followed a more familiar pattern: four months were ultimately adjusted down, and three month adjusted up.
The importance of quality government data cannot be overstated. It is the foundation of how decisions are being made that effect 100% of Americans, and if decisions are being made on low-quality data, the outputs of these decisions are likely to be as good as the inputs. Given this level of importance, we thought it would be helpful to put together a quick rundown of the numbers, and how the changes made by the BLS seem to have things back on the right track, at least for the things they have in their control.
2025: The First Estimate Kept Getting Smaller
Revisions are a typical part of the monthly employment report. The first estimate is followed by additional rounds as more employer responses become available, allowing the BLS to update its assessment of employment growth. What made 2025 unusual was not simply that the numbers changed, but that they repeatedly changed in the same direction.
Every month with an initial estimate (the October estimate was skipped due to the government shutdown) was eventually revised downward. The average first-to-third revision was approximately 58,000 jobs lower than the initial estimate. Historically, revisions have moved in both directions, but the 2025 pattern consistently presented a stronger picture of employment growth initially than subsequent data supported.
These are revisions to the estimated monthly change in employment, not to the employment level. October has no normal initial estimate because of the lapse in appropriations.
May and June: The Big Change
| Initial | Third | Revision | |
| May 2025 | +139K | +19K | -120K |
| June 2025 | +147K | -13K | -160K |
May’s initial estimate of a139,000 job increase and was later reduced to 19,000. June was initially reported at 147,000 before eventually being revised to a loss of 13,000 jobs. At the time of the July report, however, the figure attracting attention was 258,000, the combined May–June downward revision before the subsequent third revisions (the 280,000 drop was the full change from initial to third revision). Reuters described the 258,000 adjustments as the largest combined two-month downward revision, outside the pandemic period, since at least 1979. It is worth noting that the former Commissioner Erika McEntarfer was fired after the July 2025 jobs report was released, with President Trump claiming, with no evidence, that she was rigging reports.
The Bigger Correction Was Hidden Beneath the Monthly Numbers
The monthly revisions were only part of the story. Each year, BLS benchmarks its survey-based employment estimates against the Quarterly Census of Employment and Wages (QCEW), which is derived primarily from state unemployment-insurance records and covers nearly all employers. The 2025 benchmark revised the March employment level downward by approximately 898,000 jobs, or 0.6%, on a seasonally adjusted basis.
Why the Benchmark Matters
The size of the correction was also unusual. Over the previous decade, annual benchmark revisions had averaged approximately 0.2% of total nonfarm employment in absolute terms.
| MONTHLY CHANGE | EMPLOYMENT LEVEL |
| Revised as additional survey responses arrive. | Revised through the annual benchmark against more comprehensive administrative data. |
Source: Bureau of Labor Statistics, 2025 CES National Benchmark Article and benchmark documentation.
Why Did the Number Move?
There are three main reasons for the strong movement of benchmark numbers:
Response Error
Starting in early 2025, BLS noted a decline in response rates overall. Once numbers could be compared to administrative records (actual payrolls) those companies more willing to take the survey were typically overestimating the number of jobs they had, or were going to create.
Non-response error
While some of this is discussed above, there was also an important finding that non-responding businesses ultimately had a lower level of employment than businesses that did respond. While this also likely was affected by the factors in the response error, there is also likely an effect of companies that began to struggle in early 2025 having other priorities besides filling out the BLS survey.
Businesses outside of the frame
Back to more jargon category, but the short version is that it takes a while for new businesses to provide data to the BLS. Given the large number of new companies that have been formed since the pandemic, this factor needed to be adjusted.
The implication of these factors is important: the discrepancy was not necessarily the result of a statistical formula simply producing the wrong answer. Some of the underlying information entering the system was itself imperfect. BLS has emphasised that its analysis is preliminary and not comprehensive, so these findings should be viewed as important contributors rather than a complete explanation of the benchmark revision.
When the Model Has to Guess
Above is a simplified version of how the estimates need to be configured. The BLS is using all data available and one of those factors is how many births and deaths there have been, which is an input in the calculation of the overall number of jobs in the United States. Since 2020 (aka the pandemic) the models have not performed as well, and the overall estimate for the birth-death forecast had to be dropped by 117,000 jobs.
This overestimate has a ripple effect and BLS has reported that birth-death forecasts have experienced persistent and relatively large errors. Over the 12 months from April 2024 through March 2025, the model forecast 1.102 million net private-sector jobs associated with births and deaths, compared with 985,000 in the actual benchmark data. This 117,000 job gap was one contributor to the broader 861,000 job benchmark revision described below, not the whole story, since response error and non-response error (discussed next) also played a role. Corrections are nothing new, but the size of this revision is well above average and represents the second year in a row with a large number of jobs being removed.
| −861K or -0.5% 2025 benchmark revision | 0.2% 10-year average absolute revision |
This −861K figure is the not-seasonally-adjusted benchmark measure. It is a different basis from the −898K seasonally adjusted March employment-level revision used elsewhere in this report — the two should not be combined.
2026: A New Method
Beginning with the January 2026 employment estimates, BLS modified the ARIMA component of its birth-death model so that current information from the monthly sample could be incorporated into the forecast. The aim is to make the estimate more responsive to current economic conditions rather than relying as heavily on historical relationships.
| BEFORE JANUARY 2026 Historical information → ARIMA-based birth-death forecast | FROM JANUARY 2026 Historical information + current CES sample information → Modified ARIMA-based birth-death forecast |
The approach itself was not entirely new. BLS had already used similar adjustments during parts of the post-benchmark periods following the 2024 and 2025 benchmarks. Beginning in January 2026, however, the modified approach became part of the regular first preliminary estimate.
| EARLY 2026 4 upward revisions 3 downward revisions January–July 2026 months with subsequent estimates. Too early to determine whether the change has materially improved reliability. |
Conclusion
So far, the changes BLS introduced in January 2026 appear to be helping: the early revisions are smaller and more evenly split between upward and downward than the one-directional pattern that defined 2025. That is a meaningful improvement, but it only addresses part of the problem. The birth-death model’s forecasting error was one contributor to the 2025 revisions, response error and non-response error, which come down to how honestly and how completely businesses answer the survey in the first place, are harder to fix with a formula, and BLS itself has said its analysis of those causes is still preliminary. It is too early to say whether 2025 was a one-time event or the start of a longer trend, and too early to say 2026’s numbers will hold up once they go through their own benchmark revision.
For readers, the practical takeaway is less about any single number and more about how to read it. The first jobs report each month is a provisional estimate, not a final verdict. It will be revised twice more before the year is out, and checked again the following year against a far more complete data set in the annual benchmark. A single month’s surprise is normal; what made 2025 unusual was the same direction repeating for twelve months straight. That is the kind of pattern worth watching for going forward, more than any one headline figure.
References
1. U.S. Bureau of Labor Statistics. CES Revision Tables: Nonfarm Payroll Employment — Revisions between Over-the-Month Estimates, 1979–Present.
2. U.S. Bureau of Labor Statistics. 2025 CES National Benchmark Article and CES Benchmark Revision Documentation.
3. U.S. Bureau of Labor Statistics. Preliminary Analysis of the 2025 CES Benchmark Revision: Response and Nonresponse Error.
4. U.S. Bureau of Labor Statistics. CES Birth-Death Model Documentation and FAQ.
5. U.S. Bureau of Labor Statistics. Upcoming Changes to the Establishment Survey Birth-Death Model, effective January 2026.
6. Federal Reserve Board. Christopher J. Waller, speeches on labour-market data and monetary policy, 2025–2026.
7. ADP Research. ADP National Employment Report and Methodology.
8. Reuters. Coverage of the 2025 BLS employment revisions, BLS Commissioner Erika McEntarfer’s dismissal, and research on combining private and official employment data.
