Revisions to monthly payroll estimates are supposed to be roughly mean-zero. Over the last fifteen months, twelve have been downward, several materially so.
The likeliest explanation is unexciting: the birth-death model, which imputes jobs created by firms too new to be in the sample, is calibrated on a business formation and failure rate that no longer matches reality.
A model that is wrong in one direction for a year is not a model with noise. It is a model with a bias.
This matters beyond statistical hygiene. The first print moves markets and shapes policy commentary. The revision, three months later, moves nothing, because by then it has been superseded twice.
The fix is methodological and slow. The consequence in the meantime is that the highest-attention economic release in the world is systematically optimistic at the moment of maximum attention.
