Monthly employment figures are reported as facts and are actually early estimates. Later revisions frequently change the direction of the story the first release told.
The first release is built on partial returns
The published figure comes from a survey of employers, and not every employer responds before the publication deadline, which is set for timeliness rather than completeness.
Statistical agencies estimate the missing responses using established methods, then replace those estimates with actual returns in subsequent months.
Each revision therefore rests on more real data than the one before, which is why later figures are more reliable and receive less attention.
Two surveys measure different things
Employment is typically measured both by surveying employers about positions and by surveying households about whether people are working.
The two produce different numbers because they count differently: someone holding two jobs appears twice in one and once in the other, and the self-employed are treated differently again.
Divergence between them is informative rather than contradictory, and analysts generally read them together rather than treating either as the definitive count.
Seasonal adjustment carries assumptions
Hiring follows a strong annual pattern around agriculture, retail seasons and school calendars, so raw figures are adjusted to make months comparable.
The adjustment factors are derived from previous years, which means an unusual year is measured against a pattern that no longer describes it well.
After a large disruption, seasonal factors take several years to settle, and revisions during that period tend to be larger than usual.
Annual benchmarking is the largest correction
Periodically the survey estimates are reconciled against administrative records covering nearly all employers, such as unemployment insurance filings.
That benchmark can shift the level of employment across an entire year, and it is the point at which sustained estimation errors become visible.
Errors of this kind arise mainly from the model estimating how many businesses opened and closed, which performs poorly around turning points in the economic cycle.
How to read a release sensibly
A single month's change is small relative to the survey's margin of error, so a movement within that range is not evidence of a turn.
Three-month averages and the revisions to prior months usually carry more information than the headline figure that generates the coverage.
Reading the revision alongside the new number is the simplest correction available, and it is published in the same release for anyone who looks.