The ADP numbers are missing the increase in BLS employment

ADP reported today that private non-farm employment increased by 55,000 in May. ADP is widely thought of as a good predictor of the BLS figure which comes out tomorrow. But lately the ADP numbers have been way off, as the graph below shows. Last month, for example, ADP showed an increase of 32,000 - two days later the BLS came out with an increase of 231,000. Since January ADP has understated the BLS numbers by about half a million jobs. The difference between ADP and BLS numbers appears to be larger and more persistent now than it was in 2009; I wonder why. At any rate, the fact that ADP is predicting sluggish jobs growth is not convincing evidence that the BLS number will not be large.

And yet...

ADP reports private employment down 23,000 in March. But: that doesn't count Census workers, of which the government probably hired around 100,000 this month; ADP has been 70,000 too low on average since October; ADP numbers weren't affected by the blizzard in February while BLS's were artificially suppressed. So it's not hard to see BLS reporting +200,000 even in light of ADP's numbers. But I won't be satisfied with anything less than 250,000.

February employment

The ADP National Employment Report says the economy lost 20,000 private sector jobs in February. The ADP report seems to be a good proxy for private employment as it appears in the BLS report. For reasons given in the link above, the ADP numbers for February will not be affected by the blizzards in the way that the BLS data are. The data suggest to me that total employment growth in February was near zero (-20,000 plus whatever government employment growth there was). The BLS report should show a large negative number as a result of the blizzards. We're behind schedule here, people!

The construction of the ADP figures, described here, is interesting (in an incredibly techno-geeky kind of way). One reason the ADP numbers fit the BLS numbers so well is that they are benchmarked to the BLS numbers through a regression technique that I don't understand.

The BLS data (as currently reported) for growth of employment by industry is regressed on: (a) the matched-sample growth rates by industry based on the ADP data; (b) a weighted average of the historical average growth rates of employment in each cell based on QCEW data; (c) a weighted average of the historical average

growth rates of employment in each cell based on the ADP data; (d) lagged values of BLS estimates of growth of employment by industry; (e) initial unemployment claims filed during the week immediately before the week that includes the 12th of the month.

The regressions are estimated concurrently. The coefficient on term (b) above is restricted to unity. The coefficient on (c) is restricted to the negative of the coefficient on term (a). This method allows different trends of employment by size of payroll within industries, while assuming that the other industry-wide relationships implied by the regressions hold for all size classes within an industry. Inclusion of lagged employment in the regressions5 controls for shifting differences between the BLS sample and the ADP customer base, while inclusion of initial unemployment claims controls for differences in the definitions of employment used by BLS and ADP.

A level of employment is established in each cell by cumulating the predicted value of the matched sample growth rate in each cell forward and backwards from the most recently benchmarked March estimate of employment in that cell. Such referencing effectively weights the growth rates of the ADP data in each cell by

the observed distribution of employment by industry and size classification.

These levels are then summed to the aggregates by select industry and size of payroll that are shown in the summary table of the monthly report.

In addition, every year the ADP are revised when the BLS does its benchmark revisions. So the phenomenal fit of the ADP data (correlation of differences is 0.95) is by construction. I wonder how well the ADP figures predict the BLS figures before revisions, i.e. how closely is the -20,000 number reported for February correlated with the BLS number we get on Friday? Apparently Macroeconomic Advisors has studied this question, but I don't know what they found.

Bob is telling me that the seasonal adjustment factors that the BLS uses are probably skewed this winter because the weather has been much more severe than it has been in the past few years. (Basically, the BLS's seasonal adjustment procedure uses the last 5 or 6 years of data to net out seasonal effects. So if we had a mild Januaries in recent years, the January effect will be small. Then when we get socked by a bad January as happened this year, and employment declines as a result, the seasonal adjustment process only partially offsets that, interpreting the plunge as a bad economy rather than bad weather.) ADP uses the same seasonal adjustment procedure, so whatever ails the BLS data on that dimension will ail ADP data as well.

Playing around with employment data

Some back of the envelope calculations. The ADP employment report says private sector employment rose by 43,000 from September to October. The ADP numbers have consistently underestimated the BLS numbers for private sector employment all year. But it's more confusing than that: the ADP revises the previous month's numbers at the same time they report new estimates, so the change is from a revised estimate, not their initial estimate. That confuses the exercise of using the data to predict the BLS numbers, which also are computed relative to a revised estimate of the previous month's employment. So what if instead we try to predict the level of BLS employment from the level of ADP employment, then use that information to infer the change in BLS employment?

Ok, the first thing I find is that the ADP numbers on the level of employment are falling further and further below the BLS numbers.



Let's say that in October the trend continues, and ADP is 0.99 of the BLS private sector number. ADP private sector employment is reported at 107,056, so that implies a BLS number of 108,137. The BLS reports that in September private employment was 107,970 - this implies an increase in private sector employment of 167,000, before taking account of revisions.

But government employment has been falling. I've lost track of what's going on with the Census (I think all the temporary Census workers have been let go by now, but I don't know for sure), but the average monthly change in total government employment from February (before Census hiring) and September was -32,000. Therefore we get a total increase in payroll employment of +135,000.

Hmm, meanwhile the Institute for Supply Management's index of employment shows an acceleration of hiring in both manufacturing and services in October. I run a regression from January 2000 to September 2010 of the change in payroll employment on a constant and the ISM employment index for services and the index for manufacturing. I then forecast into October and I get.... drumroll please.... +133,000.

Ok, so two complete separate methods give me almost exactly the same result. I'm going to stop there before a third method gives me something completely different (specifically, I'll ignore this morning's initial claims numbers, which were not good), and place my bet:

Payroll employment for October up 134,000.

I'm a little confused about the December employment situation

My little statistical model of payroll employment, based on initial claims and ISM manufacturing and non-manufacturing employment indexes, says that payroll employment will grow by about 130,000 in December. The ADP employment survey, on the other hand, says private sector employment grew by 297,000. ADP has consistently underestimated private sector employment by about 50,000 this year, government employment has consistently fallen by a few tens of thousands per month. So I'm going to go out on a limb and say that payroll employment in December rose by somewhere between 100,000 and 300,000.

In judging the strength of the employment numbers, keep this in mind. At the pre-recession peak in December 2007 (hey, happy three-year anniversary!) payroll employment was 137.951 million. In November 2010 it was 130.539 million, so we're down 7.4 million jobs from the peak. However, population growth over the last three years means we needed an extra 4.68 million or so just to keep pace, so the actual shortfall in jobs now is about 12 million. We need 130,000 or so jobs per month to keep pace with population growth going forward; anything in excess of 130,000 jobs lets us eat into the 12 million job shortfall. So if employment growth is 300,000 per month (which is way above current forecasts), it will take us just over 70 months - that's almost 6 years - to get back to normal. We're in a very deep hole.