Labor Pains

Over on the main site I've written a (qualified) defense of the birth-death model the BLS uses to calculate the number of jobs gained and lost each month, and also argue that we (the media and the markets) pay too much attention to the job numbers in the first place.

One of the main arguments against the model is that it doesn't know how to handle a downturn, and for its part, the BLS admits to some of this (read page 6 here) but the Bureau also says:

a model-based approach is likely to have some difficulty producing reliable estimates at economic turning points or during periods in which there are sudden changes in trend. With the net birth/death model, this difficulty is significantly reduced, as imputation allows for such trends or turning points to be partially captured in the estimates. These imputed amounts are based on the active, reporting establishments which reflect changes in the economy. Therefore, contributions from the imputation for business births will also vary during economic changes. Furthermore, the imputation error from the net birth/death model is forecasted from the errors of the last five years. The error amounts from the most recent year are weighted more heavily, thus allowing recent changes to be more effectively captured.

In other words, since the model puts more weight on recent data, a downturn in jobs for the whole economy should get picked up, to an extent. And taking a look at the data between December and July, it seems that that's what happened.

The model makes predictions for the number of jobs gained or lost for ten major sectors. In December, the model's prediction and the reported non-seasonally adjusted number for each sector -- the b-d adjustments are applied to non-seasonally adjusted data -- were in agreement only 4 out of 10 times (by which I mean that both the prediction and the overall number for the sector were in the same direction and the predicted number was less than the overall number). In January, all 10 sectors were in agreement. In February it was back up to five disagreements, for the next three months it was down to two, June there was no disagreement, and July there were two. Not bad.

But things seem to have gone awry in August with eight disagreements. I've got a call into the BLS to see if they have any insight to offer here. For another excellent look at the b-d model, read James Hamilton here from last year.

UPDATE
I spoke to a very nice BLS economist who didn't really give me a good answer for the discrepancy, but did inform me that the b-d adjustments are set at every benchmark revision so the discrepancy may point to some old assumptions that need to get updated (and next month's job report will give us a look at the next round of benchmark revisions.)