Economist Writing Every Day

Are you tired of hearing about revisions to jobs data? Well, there was another hot one released by BLS yesterday. Known as the “preliminary estimate of the Current Employment Statistics (CES) national benchmark revision to total nonfarm employment,” this change isn’t yet incorporated into the official jobs data. But it will, possibly slightly modified, be included with the January 2026 jobs release, altering jobs data back to April 2024. It is part of the normal annual process of reconciling the monthly, survey-based jobs data with the near-universe data from unemployment insurance records. Normally, this is a quiet affair, especially the preliminary estimate which is just giving a heads up to researchers about what will be coming in a few months.

I wrote about these preliminary figures last year, when the initial estimate was a negative revision 818,000 jobs. When revised and actually incorporated into the data, it was a somewhat smaller 598,000 jobs, which I then used in a post just last month to show that BLS hasn’t been getting worse at estimating jobs. If anything, they have been getting better. Yesterday’s report showed that the revision could be negative again, this time 911,000 jobs. That’s a little bigger than last year, but maybe it will end up being smaller in the final number. So, no big deal again?

Maybe not. The 911,000 jobs revision would actually be much larger than last year’s revisions because it’s coming on top of a slower growing labor force already. The initial report for March 2024 showed 2.9 million jobs added in the past year, so the 818,000 revision was a much smaller share than this most recent data, since the March 2025 initial report showed just 1.9 million jobs added in the prior year. And the March 2025 jobs numbers have already been revised down by over 100,000 jobs since the initial report, meaning that potentially half or more of the initially reported job gains would be lost due to the revision, as opposed to about 20 percent last year.

Is losing half of the job gains large? Yes. In fact, almost unprecedented:

(note: I am trying out a new chart template. Let me know what you think!)

The preliminary revision for March 2025 certainly stands out. It would be roughly equal in size to the 2020 revision, which we know was a highly unusual and hard to measure time. Before that, you have to go back to 2002-03 to find revisions this large, but crucially, the 2002-03 and 2020 revisions were of a different nature. These were negative job growth numbers that got even more negative after revision. They were either during recessions or immediately after, when the labor market was still losing jobs year-over-year. The same goes for 1990-92 revisions: during and after a recession, and making bad jobs numbers even worse. None of the large negative revisions in the last 40 years were robust job growth years, that subsequently had half of their jobs taken away by the revision.

(Note: I truncated the chart for 1992, because it was a bit of an odd one — the initial report was just -16,000 jobs lost in the prior year, but it ended up being -201,000. That’s an 1,100% error!! But this is because the denominator was so small, so it’s not really a good comparison with this approach).

I think all of this means that the 2025 number really is a big deal, after several years of the BLS pessimists crying wolf about the numbers being way off. They now appear to actually have been way off!

Importantly, I don’t attribute this to incompetence or malfeasance on the part of BLS. Survey response rates are dropping, which is a problem, but no doubt BLS would like to have better resources to improve their methods (and they always are updating and improving their models). Sometimes a survey, even a survey of thousands of employers, can be wrong, especially if the non-responders or those not in the survey pool are different from the responders. This is especially challenging around recessions, when many firms are closing shop and possibly not reporting those losses. BLS has to model those, with its birth-death model, and it may not be especially good during a recession.

But we weren’t in a recession in 2024, which makes the preliminary, large revision number for March 2025 (covering the prior 12 months) an even bigger puzzle.

Of course, one other approach is to use the total size of the labor force as the denominator (rather than job growth) in the calculation of the revision error. For example, Ernie Tedeschi does this in a nice chart he made. But notice that even with this approach, the errors have been getting larger over the past 2-3 years, and April 2024-25 looks especially bad, especially outside of recessions!