December Employment Report Preview
Each month I draw upon my team's original research to present a preview for the Payroll Employment Report. This is not a prediction of the announcement to be made Friday morning. It is better described as a forecast of the net change in payroll jobs once all of the data revisions are complete.
I'll start by explaining our approach and then preview some other estimates.
Each month we ask the question, "What change in payroll employment
would be consistent with other economic data from the same time period
(the middle of the prior month)?
This is not a forecast, per se, since we do not posit any
causal relationship among these variables. They are all concomitant
indicators of economic activity.
- We use the four-week moving average of initial unemployment claims,
culminating in the week of the employment survey. This is the best
direct indicator of new lob losses. This has improved in the last
three months to a loss of 468K. Note that the most recent
decline to 432K is not within the survey period for the monthly report.
- We look at the University of Michigan sentiment survey,
which we find to be more useful than the Conference Board's sentiment
index. Michigan uses a panel, where some families are carried over
from month to month. This is a good technique. Sentiment is
influenced by employment. When people have lost jobs, or are worried
about losing jobs, it shows up in sentiment. It is a good concurrent
indicator. The Michigan index is now at 72.5, rebounding from a dip over
the last two months.
- We us the ISM manufacturing index, which rebounded to 55.9 after a slight dip to 53.6 This is quite bullish for the overall economy.
Our long-term record has been pretty good, especially when compared
to the final revised data. This makes sense because our model was
derived from the final data. In recent months we have been too
bearish. The BLS benchmark revisions suggest that we have been much
better than first thought. I am working on a comparison with the final numbers.
This month, our estimate is for a net job loss of about 62,000.
It is always interesting to compare the job forecasts from different
sources. We follow several because of the interesting and widely
varying methods they use. A wise interpretation would be to consider
all of these disparate sources of information.
ADP has proprietary data because of its payroll management business. ADP sees losses of 84K. This estimate does not include government jobs.
All of these sources are valuable. The 90% confidence interval on
the BLS estimate, something that no mainstream media sources report, is
+/- 100K or so. And that is after revisions and benchmarking. It is a
survey — a good one — but it has an error band.
How to Interpret the BLS Report
There is an inherent problem with the BLS survey approach: How to deal with non-respondents. In most surveys the researchers treat non-respondents as similar to those who do answer. There is always the potential for self-selection by respondents.
In the payroll survey the problem is worse. We know that some of the non-respondents are no longer in business. The survey approach is an awkward method for determining the entire count of jobs. The BLS has done well to estimate job creation in various circumstances, but the methodology has broken down during 2009. While many observers point to the Birth/Death adjustment, this is a relatively minor element. Attention to the size of this adjustment actually distracts people from understanding the problem, as I described in this article.
Ironically, some other sources have been criticized for not matching the preliminary BLS reports. ADP has been off by about 85K a month according to Bloomberg. When we see the benchmark revisions in February, the BLS series will be adjusted lower, probably by nearly that amount. Perhaps Bloomberg will then run another story giving ADP a higher grade.
To summarize, no single methodology has a monopoly on truth. In about nine months we get the state data from actual unemployment premiums. That should be everyone's benchmark. Unfortunately, really accurate data often comes long after we need it.