> With a valid measure I would expect a roughly even distribution over time between underestimates and overestimates

This is a valid hypothesis. It’s wrong, and I’ll explain why. (It’s a bad and invalid thing to conclude.)

If measurement errors were iid, you’d be correct. But they’re not. They’re well documented for not being so. Earlier survey results are biased by directional response bias inasmuch as the employers with the lease changes respond first. So the earliest releases tend to match whatever was going on before. Then the employers who had to do paperwork respond. And then, finally, someone gets around to calling the folks who never got back. Some of them aren’t around anymore.

So yeah, the directional tendency in revisions is well documented. And for a long time, the early releases were appreciated. But maybe American statistical and media literacy is such that only final releases should be released, which would mean we’d always be working with data 6 months to a year out of date.