Each month the market reacts and over-reacts to the payroll employment report.  The problem is that the BLS has been asked to measure the monthly change in employment.  Why?  Because that is what we want to know.

The problem?  It is an impossible job.  The BLS attempts to solve the problem in a very professional fashion, by estimating the total employment in each of two months and subtracting one from the other.  They use excellent statistical methods in their process, but even a small error in the total number of jobs leads to an apparently large error in the change.

At some point we will have something approaching an actual count using the state unemployment data.  That is not available for six months or so and by then, no one cares any more.

The Fresh Look

A wise stat prof once explained problems like this by telling us to imagine that God whispered into our ear and told us the Correct Answer!  A statistical process is only an estimate of that answer.  The methodology, if it includes a survey, involves both sampling and non-sampling error.

The BLS attempts to learn the correct answer and announces their result on the first Friday of each month.  (See our Payroll Employment Game and play yourself to get some idea of the range of results even when one knows the Correct Answer.)

ADP has plenty of data about private firms and uses this actual data to measure monthly changes.

The NewArc Investments staff has developed a model that asks what change in jobs corresponds to other economic indicators.  Please note that the job change is not a “true” dependent variable.  Each indicator in the analysis is a measure of some other cause — change in the economy.

The Twin Danger for Modelers

There are many problems in analyzing payroll employment, but two stand out:

1. The BLS, even though it is only doing an estimate of the Correct Answer, is deemed to have the official result.  Put another way, anyone making a prediction about non-farm payrolls is not trying to predict the truth, but rather to predict the BLS estimate of the truth.  This estimate has a confidence interval of +/- 100,000 from sampling error alone.
2. Most of us base models on the final, official data — after both the monthly and the benchmark revisions.  This makes sense because we are attempting to represent what is actually happening, something that is not known until later.  It also makes sense because there is no predictable pattern to BLS revisions.  They come from a delayed response from some companies and eventual benchmarking.  We have seen no good method of predicting revisions, despite frequent claims by some pundits.  Claims that the BLS does revisions to improve the outlook for political reasons are just silly.

The Modelers May be Right

At “A Dash” our methods are based upon the final revisions.  We use various economic indicators and ask what job change would be consistent with things like Michigan sentiment, the ISM report, and the four-week moving average of weekly jobless claims.  Our result for this month suggests a job loss of about 200,000, a very disturbing result.

There is a problem.  One of our indicators, the initial claims series, may be distorted.  We described this today on RealMoney as follows:

Why not accept the result at face value? There was a big spike in initial claims a few weeks ago. It was attributed to the emergency extension of benefits and state efforts to communicate with those who might be eligible. The result was that many people who had been eligible for some time all showed up for first-time claims. I would expect this effect to last only for a few weeks, but claims have remained high. My guess is that the unusual circumstances caused the spike, but there has also been a build-up in claims. As a result, the four-week average is a bit higher than it otherwise would be.

To conclude, this many-faceted report is even more difficult to preview than usual. Interested readers can get more background on the methods I use in my article from last year on little known facts about the payroll employment report.

This situation causes us to have even less confidence than usual in a prediction.

ADP also works from final BLS data.  In a typically excellent article, David Altig analyzes the ADP estimate of a job loss of 33,000.  Readers should see his entire article for a complete analysis and charts, but here is his conclusion on ADP:

Our Conclusion

The employment data is showing, so far, job losses consistent with slightly positive economic growth from the payroll jobs perspective.  If one uses the employment rate or hours worked (noted by Menzie Chinn), the data are at mild recession levels — so far.

And finally, none of this has much to do with the much-maligned birth/death adjustment, which has consistently improved BLS forecasts when checked against actual data.  We expect some downward revisions from the upcoming benchmarking, but not nearly as great as those forecast by the leading critics of the BLS.