Consensus is for 215K on payrolls and for unemployment to fall a tenth to 6.6%.

Neil Irwin has a smart piece up on how, given the amount of noise in these data, most analysts read too much into the monthly job numbers (h/t: WC). I loved his set of hypothetical headlines–all different, and all consistent with a) a true growth rate of 150,000 jobs and b) the fact that there’s a 90% confidence interval of plus or minus 90,000 jobs around the published number.

To be clear, that means that if we hit the consensus tomorrow, there’s a 90% chance that the actual change in payrolls is between 125K (“A lousy month!”) and 305K (“Break out the champagne!”).

A few more thoughts about this:

–Irwin very correctly urges analysts and journalists to keep this variance in mind. I agree, but as I recently wrote on this very cite: *“…most everyone reads too much into these monthly jobs reports, given their relatively large confidence intervals and later revisions. (Truth be told, I usually say that right before I start reading stuff into it.) That’s why I always average a few months together to smooth out some of the noise.”*

–That latter point, about smoothing things out, leads me to posit a suggestion: maybe we should switch to quarterly jobs reports, like with GDP. It would save a bit on the reporting budget, though I’d keep the sample size the same. If I’m recalling correctly, that would reduce the confidence interval to about 50,000. (An assignment for the reader: with what frequency do other advanced economies publish employment statistics?–I seem to recall that at least some do so quarterly versus monthly.)

–For securities traders, however, what matters is less what’s actually true in a statistical sense then what other traders think is true. If we get 125K tomorrow, it’s not obvious that the smart trade is “based on the confidence interval, we cannot say this is a weak report, therefore I’ll stay put.” What matters, and I’m just talking about skittish daily movements here, is what the broader market decides about the number.

The initial readings tell us enough. And an initial reading of 215,000 means the final number is more likely closer to that then 150,000. Probability is not equal along a confidence interval.

Same with polls too, and the +/ 3 %, etc.. The stated number is the most likely along confidence intervals in polls.