February Employment Data: Watch Out!
Each month market participants wait with bated breath for the payroll employment report. No other piece of government data gets more attention.
This month, the significance is even greater. Many see employment changes as the true indicator of potential recession conditions. There will be plenty of commentary — all delivered in serious and authoritative tones.
None of the commentary will be accurate!!
To our continuing amazement, no one seems to understand the payroll employment report. There is one group that sees it as the official government number. It is the result against which everyone else’s estimates are measured. This happens because we really would like to know what the job change is each month. The government reports a number, and most take it seriously. Billions of dollars of market cap will swing on whether people think it shows a recession in progress.
Another group will maintain a consistent negative bias about the report, suggesting that job creation is inflated by the BLS. This is mostly because of the consistent and erroneous bashing of the Birth/Death adjustment.
No one seems to understand the process. The journalists in mainstream media publicize comments without doing any real education. Everyone who is invited to comment on TV or for the Wall Street Journal is happy to do so. There is a cottage industry in looking at the "internals" of the report. None of these analysts will mention sampling error or statistical significance.
A Dash of Insight on this Report
Let us state first that this report is generated with the highest level of skill and professionalism. We have interviewed the BLS researchers. Our own experts — highly skilled research professionals — have reviewed the methodology. It is excellent. There is no political bias. The revisions that occur come from delays in reporting companies, not from the BLS. There is no way to predict them. The much-criticized Birth/Death adjustment has improved the actual estimate in every period since it has been used. We know this by comparing the (eventual) state data, reported months later, with the BLS estimate.
So What is the Problem?
The BLS has been asked to do something that is nearly impossible. They estimate the number of jobs in the economy each month. That is about 140 million jobs. They then take the estimates for two consecutive months and subtract them to generate a difference. The process is a good survey of establishments with some excellent adjustments.
Even if they do a great job of estimating jobs in both months — a large number — the error in the difference between the months is quite large.
When they are all finished, all revisions, all employers reporting (which we do not see for two more months) they are subtracting one survey from another. The sampling error has a 90% confidence interval of +/- 100,000 jobs. (And by the way, the estimate from the household survey is +/- 400K jobs.)
Think about this. As our old stat prof explained it, suppose that God whispered in your ear and told you the TRUTH about the actual job change. One-third of the time, the official number would be 50K or more light and another third it would be 50K or more heavy. Meanwhile, the market draws inferences from much smaller deviations. You can know the truth, and still be voted wrong by the market.
The "internals" of the report have a similar proportionate sampling error.
Current Forecasts
ADP attempts to forecast the report with their own analysis of companies in their database. While their method has gotten a lot of criticism from some big misses, an analysis by Bespoke Investment Group shows that it is actually pretty good. ADP is looking for a gain of 2000 jobs.
Our own model, based upon the four-week moving average of initial claims, Michigan consumer sentiment, and the ISM report suggests a decline for February of 37,000 jobs. One must be careful in doing this analysis, using data from mid-month, since that is when the official survey is taken. Tomorrow’s jobless claims, for example, will not be reflected in Friday’s employment report. We are "solving" for a job change that is consistent with other concurrent economic data. No factor "causes" another. They are independent reads on the economy.
Economists have a consensus estimate of a gain of 20,000 jobs.
Any government result ranging from a job loss of 80,ooo to a gain of 60,00o does not prove any of these forecasts to be incorrect, in terms of statistical significance. And that is after we have all of the data — many months from now. The initial report has an even wider range, since the sample is not complete.
Resources
For more detail, interested readers might check out this article, or read more widely from our past analyses.
We also have a comprehensive article available upon request.
Readers can also see the effects at our sister site, The Payroll Employment Game. You can plug in your own guess at "truth" and see the range of possible BLS survey results.
Making Money from this Knowledge
Quite frankly, there is no good method for assuring an edge in trading. That is the point of highlighting the wide error band. Whatever happens will be given undue weight by the market. Having said this, we see significant downside risk in Friday’s report, mostly because our own forecast is so much lower than the consensus.
We all want so much to know what is happening with employment, that the monthly survey gets too much weight. It is a case of the data we have versus the data we need.
To the extent that I use the number, I try to use YOY change, because signal is greater than noise at that point. The only thing from a methodology standpoint that bothers me at all on the NFP number is the use of an ARIMA model in the birth-death adjustment (I haven’t read the methodology in two years, so if they have shifted, please forgive me). From my experience, ARIMA models are less reliable than structural models, particularly when forecasting.
At my interview on Bloomberg radio yesterday, I tried to make the point that people overanalyze period-to-period changes, but I don’t think I presented it that well. Perhaps that steps on their business model.
You illuminate a very pertinent point about the difficulty for the BLS to measure the NFP. When I was working on a mortgage trading desk, I build several models in attempts to forecast NFP, eventually finding a method producing very accurate forecasts.
My earlier models were based on other economic data reports, but weren’t highly effective.
I soon came to the conclusion that a dichotomy exists: 1) predicting the actual change in NFP 2) predicting the results of the NFP report.
I turned my attention the shortcomings of the BLS reporting (monthly reporting periods and frictional factors leading to lumpiness in additions) opposed to a purely economic variable based model. Building a GARCH based model, I theorized that there was an underlying true trend of job growth coupled with NFP report volatility.
The theory is that job growth is lumpy, when measured by monthly periods, yet there is a overall mean trend. Frictional factors lead to available jobs not being filled, thus some months not all openings will get filled, however that labor capacity is carried forward to future months where it is filled leading to above mean BLS NFP prints.
The model forecasted the mean trend based historical moving average with adjustments for GDP and productivity. The assumption is that the economy can create a specific number or new jobs, yet they are not added evenly month by month. Thus, some months when the additions are lower, that number is added to a cumulative deviation, which then will be added to future months forecasts to account for available jobs not being filled.
Example: F05- 262, M05 -est 215 act- 110, A05 est- 179 act 274. M05 est 180, act 78, I believe the trend was about 185, so if feb was high by 75 then, Mar should be low by 75, which it was. Of course, it much more complicated, since the volatility must be modeled, but I wanted to illustrate the reversion of the volatility, and that the consensus is usually close to the underlying mean trend, yet the reported figures, bounce around above and below.
It’s complicated by degree of sampling error of BLS, which later results in the revisions, which have to be modeled as well and also adversely affect the predictive power of the model. Another shortcoming with GARCH models is the breakdown stemming from large changes in the mean trend. That type of model would be inappropriate in current environment, however ’04-’06 it worked remarkably well. Just wanted to share my thoughts.
David and Turley —
I am working on several more articles on the BLS approach. You have some interesting ideas and I would like to consult further. I am intrigued.
Thanks,
Jeff