Nobody asks this out loud on the first Friday of the month, and everything else on the page rests on it. If the labor market has beaten the forecast this consistently for two years, what is the forecast actually a forecast of?
March payrolls came in at 303,000 against a consensus expectation of about 200,000, with unemployment at 3.8 percent and participation ticking up. It is the latest in a run of these, and the run is now long enough that the misses are the pattern rather than the exception.
Before the question can be answered, the number has to be understood, and almost nobody outside the business knows how it is made. So: a primer.
It is not one number. It is two surveys that disagree.
The monthly employment report is built from two entirely separate instruments, run by the same agency, asking different people different questions.
The establishment survey goes to employers. It samples business and government payroll records and asks how many people were on the payroll in the reference period. That produces the headline jobs figure, the 303,000.
The household survey goes to households. It asks people about their own situation: whether they worked, whether they looked for work, whether they wanted work. That produces the unemployment rate, the 3.8 percent.
The two are not the same measurement expressed differently. A person with two jobs counts twice in the establishment survey and once in the household survey. A self-employed person appears in the household survey and not the establishment one. The surveys have diverged for months at a time, and when they do, commentary tends to quote whichever supports the argument being made.
Both numbers are estimates with error bars nobody prints
This is the part that would change how the report is read if it were stated in every story about it.
These are surveys, not counts. The payroll figure carries a confidence interval around it that is wide enough to swallow most of the monthly variation people write about, and the household figure carries its own. A move of thirty or forty thousand between months is inside the noise.
The figures are also revised, twice, as more complete payroll records arrive. Those revisions are frequently substantial and they are reported thinly or not at all, because the revision arrives on the day a fresh headline number is published and the fresh number takes the coverage.
So the sequence for any given month is: a widely reported estimate, then two quiet corrections, and the corrected figure is what eventually enters the historical record that everybody later analyzes.
Where the forecasts come from
The consensus expectation reported as the benchmark is a survey of economists, and it is worth being clear about what those economists are doing.
They are not forecasting the labor market. They are forecasting the survey result, which means modeling both the underlying economy and the behavior of a statistical instrument, including its seasonal adjustment.
Seasonal adjustment deserves its own sentence here because it does a great deal of the work. Hiring is enormously seasonal, and the published figure is the raw count adjusted by a factor estimated from prior years. In a period when the seasonal pattern itself has shifted, which it plainly has since 2020, the adjustment is estimated from a pattern that no longer quite holds, and a systematic miss is exactly what that produces.
The answer to the question
What is being forecast is a model of the economy that was built on the relationship between the labor market and everything else, and that relationship changed.
The standard framework said that reducing inflation of this magnitude required demand to fall, that falling demand shows up as job losses, and that unemployment therefore had to rise substantially. That is not an unreasonable theory. It is close to the central result of decades of macroeconomics, and it is why a recession was forecast repeatedly and confidently.
What appears to have happened instead is that a large part of the adjustment came through the supply side: labor force participation recovering, immigration adding workers, and the extraordinary distortions of 2020 and 2021 unwinding. An economy that can add workers can add jobs without adding wage pressure, which breaks the link the forecast depended on.
So the misses are not sloppiness. They are a model being wrong in a consistent direction, which is the most informative kind of wrong, and the profession has been slow to say so because saying so means retiring the framework rather than adjusting a parameter.
Why the report exists in this form at all
A word about the institution, because trust in it is doing quiet work throughout.
The figures are produced by the Bureau of Labor Statistics on a published schedule, with the methodology documented, the confidence intervals stated, the revisions disclosed, and the release time fixed so that nobody sees it early. That combination is why markets move on it.
It is also why the revisions matter more than they appear to. An agency that quietly corrected its own numbers would have a credibility problem; an agency that publishes the corrections alongside the fresh estimate has handed the reader everything and is not responsible for the reader taking only the headline.
The failure in this ritual is therefore not statistical. The statistics are the most carefully documented part of it. The failure is in the layer above, where an estimate with a wide error band is compared against a forecast of an estimate, and the difference is reported as news about the economy.
How to read the next one
Three habits, all of which cost nothing.
Read the revisions before the headline. A month that adds 300,000 while revising the two prior months down by 80,000 between them is a different month from the one the headline describes, and the revision is in the same release.
Read the household survey alongside the payroll figure rather than instead of it. Sustained divergence between the two is a real signal about the composition of employment, and it is invisible if you only ever look at one.
And treat the consensus miss as information about the forecasters rather than about the economy. A single surprise is noise. Two years of surprises in one direction is a statement about the model, and the model is the thing that will eventually be revised.
The practical consequence is smaller than it sounds and more useful than it looks. A single month’s figure is not a signal worth reacting to; the three month average is. And the forecast that figure gets compared against has been, for two years running, the least reliable element in the entire ritual.
