> There is a popular derivative heuristic that the economy is "good" when ADP > BLS.

Surely that doesn't apply when both metrics are bad? I don't know. I'm not familiar with job reporting statistics. So info / perspective would be appreciated.

Also, when you say "over fits" I'm used to that in a machine learning context where it means data outside the fit is worse. Did you mean that ADP is more accurate with private sector and BLS is more accurate with public sector. Or is it more a matter of ADP comes in higher with private sector and BLS comes in higher with public sector. Usually, "overfit" is a bad thing in a machine learning context, but I wouldn't expect ADP to do worse in private and BLS to do worse in public. But the opposite. I would appreciate it if you could clarify the terminology from your perspective.

Yes, ADP is more accurate with private sector jobs and BLS is more accurate with government jobs, because their data sources are heavily biased toward having accurate data in those subsets respectively. They both generally assume that the data they don’t have looks like the data they do have despite the selection bias and disjointedness of their data. BLS extrapolates private sector employment from government employment and census data in ways that no one thinks is representative but that is what they work with. ADP has no government employment data in their model but has a relatively exhaustive sample of private sector employment.

People that know account for these biases. This is why the ADP > BLS heuristic exists. If private sector jobs growth is less than government job growth, reflecting the biases in their respective models, that is interpreted as either the government sandbagging the numbers with make-work jobs or private sector employment being so poor that the monotonic government employment eclipses private sector job growth. Either way, it is a bad sign.