Originally published on July 9, 2019
By Robert Eisenbeis, Ph.D.
Earlier in the week, before July 4th, ADP released its jobs report for June and predicted that the economy had only created a modest increase of 102K jobs on the heels of May's 75K. The press picked up that number with headlines like "Job Creation Has Another Rough Month in June as Private Payrolls Rise by Just 102,000." Economists who responded to a Dow Jones survey on expected June jobs were nearly as pessimistic, forecasting only a 135K increase. Those forecasts on the eve of the July 4th holiday led markets to believe that a continued softening of the economy would provide the needed ammunition to justify a reduction in the FOMC's policy rate later in July, and as a result, markets closed at an all-time high on Wednesday.
Now we have the CES and Household Survey results for the month of June from the Bureau of Labor Statistics, showing that the economy actually created 224K jobs, more than double the number projected by ADP and above the average of 172K per month for the first six months of this year. Moreover, there are many positive elements to those job numbers. The unemployment rate held steady at 3.7%. There were healthy increases in business and professional services, healthcare, transportation and warehousing, manufacturing, and most notably, a 21K increase in construction jobs, despite what has been claimed to be a slowdown emerging in housing. Other major categories showed little change for the month. Finally, there was a gain in average hourly wages, which are now up 3.1% on a year-over-year basis.
All of this positive news disappointed markets, with the Dow down 130 or so points by midday Friday, July 5th, before ending above the day's opening. This response unfolded because the prospect of a rate cut, which had been priced in, now looked problematic and resulted in a reassessment.
What can we learn from this episode? First, it seems clear that it may be risky to put too much weight on a single data release, in this case on the ADP report. The size of the miss raises the question of how good of a predictor of the CES jobs report the ADP report is, keeping in mind that these two reports are compiled differently. The ADP report is based on data from approximately 400K businesses that are clients of ADP, and it is compiled by Moody's. The CES report is based on a monthly survey of a much smaller set of about 142K businesses, but covering over 680K job sites in the US.
If one runs a simple linear regression using the ADP release to predict what the CES jobs number will be, a number of interesting conclusions emerge. Below is a brief summary of that regression, spanning April 2002 through July 2019.
CES = 7.633 + .969 x ADP
Adjusted R-square = .754
The overall explanatory power of the regression within the sample is quite high and about on par with average results for time-series data like the employment series. That is, the equation explains about 75% of the variation in the CES survey. So, for every 100K jobs that were created according to the ADP survey, the prediction would be 90.7K jobs for the CES survey. However, even more interesting is what the ADP forecast errors look like from month to month. The absolute mean forecast error (that is, the average absolute size of the error regardless of whether it is positive or negative) is 78.9K.
The biggest prediction errors were in the immediate aftermath of the deepening of the financial crisis in 2009. In March of 2009, ADP predicted a decline of 417K jobs when the CES number showed a loss of 803K jobs. In November of 2009, ADP predicted a loss of 373.8K jobs, but instead the CES data showed a gain of 12K jobs.
The ADP survey consistently missed big during much of 2009 and 2010, and this record may explain why Moody's modified the methodology in November 2012. Looking at just the performance after the November 2012 methodology changes, a regression on the smaller sample size yields quite a different equation than a regression for the entire period since 2002. The following is the new equation:
CES = 68.57 + .695 x ADP
Adjusted R-square = .209
The overall power of the equation is substantially lower than that of the full-sample equation shown above explaining only 20% of the variation in the CES. However, given the relatively stable period after 2012, the absolute mean forecast error is now somewhat smaller, at 54.83. Looking at the over and under predictions since 2012, the biggest miss on the downside was for the month of July 2016, where ADP forecast 151.8K jobs being created, whereas the CES number came in at 336K. The biggest over-prediction was for the month of February 2019: ADP predicted that 219.9K jobs would be created when the CES reported 56K new jobs.
This brief dive into the ADP data again highlights the point made at the outset of the commentary, and that is, when it comes to economic forecasting, it is risky to bet on one given forecast or estimate. A colleague once put it this way: "When one lives by the crystal ball, one occasionally has to eat broken glass."
As for what the FOMC will do this month, our view has been and continues to be that the Committee will keep its powder dry and wait for more incoming data. The first look at second quarter GDP will now be extremely critical as will Chairman Powell's testimony later this week.
 The ADP data start in 2002 in the St Louis Fed's FRED database.
 As one last check, a regime-shift equation was estimated, allowing for a parallel shift in the estimated equation, separating the pre- and post-November 2012 period. The new equation is CES = 4.118 + .956 x ADP + 12.26 post-2012, and the adjusted R-square is now .75, about the same as that for the full sample. Interestingly, the equations would have all been off by about 100% for the July 2019 CES jobs number.
Editor's Note: The summary bullets for this article were chosen by Seeking Alpha editors.