The following analysis sets out inconsistencies in the payroll employment data as published by the Bureau of Labor Statistics (BLS), versus the actual data as adjusted and recalculated each month by the BLS using its “concurrent” seasonal-adjustment process.  With this product, we make available the SGS calculations of the actual data produced internally by the BLS, using the published BLS data methodologies and programs. 

These numbers are provided in a format whereby data users may explore them not only at the aggregate nonfarm payroll level, but also at numerous industry levels.

No attempt is made here to address any issues in terms of possible seasonal-factor distortions, sampling problems or bias-factor and birth-death modeling issues. The data here, again, are unpublished BLS calculations, not an SGS-Alternate Payroll Employment Measure.


  • Unpublished Seasonally Adjusted Time Series for the Payroll Survey. The headline number is published by the BLS, but the public cannot see how seasonally adjusted jobs have been moved between past months.  Why is this important and how do we obtain the real picture?
  • Trend Extrapolations Showing Seasonal Residuals Available.  How a by-product of the seasonal adjustment process gives us insights into the next month’s Payroll Survey’s seasonal adjustments.
  • Data Sample AvailableWe make available a current sample of data as an Excel workbook.


Figure 1


1.     The Problem with the Published Payrolls Data

The Bureau of Labor Statistics (BLS) produces the much-watched Establishment Survey,1 popularly referred to as the “Payroll Survey.” Its headline, monthly change in total employment is much anticipated and analyzed by economists and investors.

Although the BLS also publishes revisions to the prior two months’ numbers, it is not widely understood that the concurrent seasonal adjustment (SA) process used each month revises every past month.2

As we understand it, the BLS does not publish these revisions because of the effort and confusion involved in publishing a fully revised history each month.3 Prior to the introduction of the current methodology in 2003, only the prior two months were revised and published.  Perhaps, the effort involved in re-working the established publication process was deemed not worthwhile when the new methodology was introduced.

The BLS also might argue that public should be concerned only by the trend of the most recent three months, and that the changes made to prior months are not significant. 

But, there are reasons for wanting to see the real picture:

  1. The monthly change as recently as two months before the current release is incorrectly stated by the BLS.  For example, for the March 2012 release, the published monthly change for January is the difference between the new January number and the old December number. What we really want is the difference between the new January and the new December numbers (283,000 versus the reported 275,000).
  2. Part of the current month’s headline number contains a re-adjustment to the seasonal factors. One would like to see how much of this may have been re-allocated from past, un-reported months.  For example, a few months ago the published headline number may have benefitted from a certain seasonal adjustment.  Some of the benefit now may have been revised away and now shows up in today’s headline number. Without knowing this the public has an incorrect picture of the true pattern of growth, free of any possible “double-counting” of seasonal effects.
  3. If data users want to use the BLS time-series in any statistical model, they need to have an internally consistent series showing the latest view of what the seasonally adjusted numbers were in the past, not a series which is the co-mingling of data from separate monthly revisions.

As an example, the chart below shows the published and actual SA monthly changes for the total private payroll numbers.


Figure 2

Dark red shows the unpublished picture, the light blue the numbers published by the BLS. These red, “Actual” numbers are the product of the BLS’s X-12-ARIMA seasonal adjustment model.  They are produced each month by the BLS but not published.  We reproduce them each month using the same software, input data and configuration files used by the BLS, with the differential plotted and discussed in the regular, monthly SGS Newsletter commentary on the Payroll Survey.

The actual numbers provide a self-consistent time-series. The number for, say, October 2011 was produced as a result of the same seasonal adjustment process that created the number for March 2012, and indeed all of the red monthly numbers. This process made use of all known data up to and including the March 2012 release, and was the process that produced that March headline number. All of the seasonality has been spread self-consistently throughout the year.

The blue numbers however are a combination of the SA processes made at different times in the past.  For example, the published monthly change for November 2011 is derived from the job-levels for October and November 2011.  But the values used for those two levels were those created by two different seasonal adjustment data sets, the former was part of a pre-benchmark process carried out using data up to and including October, while the November number was produced when the January 2012 release was produced.  To some extent, apples are being compared with oranges.  

One can see that there is a moving around of 20,000 job swings here and there, as the seasonal adjustment factors have changed.  Compare the October through December numbers in Fig. 2 above – the seasonal adjustment process has changed its mind markedly about the distribution of seasonal effects -  and, this is for the high-level aggregate total.  At a specific industry level, those swings could be much more extreme.


2. Trend Extrapolations

A by-product of the X-12-ARIMA process is a trend “forecast” for each industry series. These give future seasonal factors, and estimates for future job levels based on the extrapolation of the fitted seasonally adjusted trend. These industry series can be aggregated to give “forecasts” for higher level sectors and totals.

It must be emphasized that these forecasts are purely statistical extrapolations of the existing, seasonally adjusted “trend” which has been fitted to the data.  They cannot predict what underlying, non-trend changes may take place out in the real world. In effect they are saying that if the “trend” in each industry continues, this is what will be reported by the seasonal adjustment process next month.

These forecasts are useful, in that:

  1. They show the actual, calculated trend. This is not as simple as saying that next month will be the same as this month, or some simple average of past months.
  2. When an industry reports a number the next month that is markedly off-trend, data users can be alerted to the fact that something of significance may be happening in that industry.
  3. The X-12-ARIMA model cannot fully remove all seasonality. The forecast can alert users ahead of the next month’s report, as to what residual seasonal effects may contribute to the final headline number.  These are artifacts of the whole process, not real, but are nevertheless part of the reported headline number and, as such, of interest in predicting what might be published.

A rather egregious example of this last point occurred prior to the December 2011 payroll report.  Here is the picture as of the November 2011 release, having been run through our X-12-ARIMA process.


Figure 3

The trend extrapolation for total private sector jobs in December was a gain of 191,000 jobs.  In drilling down to where this was coming from, the first stop was one sub-sector which stood out.


Figure 4

A rather large forecast jump in December employment!

On further examination, we saw that the X-12-ARIMA model was having a problem with an underlying component series, “Couriers and Messengers” (not charted separately here).  This had highly seasonal spikes around December which were not being completely removed in the seasonal adjustment process.  The forecast was telling us that the model was likely to add a very large number of jobs in December purely as part of the SA process, regardless of what the underlying, not-seasonally adjusted numbers might come in at.

When the BLS December numbers did come out, this indeed was what happened:


Figure 5

Fifty thousand jobs reported added to the sector in December after seasonal adjustment, with total private payroll growth coming very close to the trend estimate at a 212,000 gain.

This odd-looking gain was commented on at the time in the BLS release,4 but was not noted as a seasonal adjustment problem, and raised some eyebrows after the event. The modeling of this series has since been improved by an adjustment to the X-12-ARIMA parameters made by the BLS as part of the 2012 Benchmark revision, and this jump has now disappeared from the currently published series.

The modeling of this series has been improved, but there are still some residual seasonalities in the system, witness the current forecast for total Private payrolls shown here going out for the next three months:


Figure 6

If there were no residual effects, one would expect the next three months to be at the same level. 


3.  Data Samples

We have packaged up a sample data-set as an Excel Workbook with menus and charts for browsing the data.    

The data-set covers:

  • Data from the March 2012 BLS Payroll Survey (published April 6th.)  (Please contact us to request next release data-set.)
  • Our cacluations of the unpublished, BLS seasonally-adjusted data series
  • Trend extrapolations

The sample data set contains:

  • Data back to January 2010
  • Next three months of trend extrapolations
  • 131 industry and aggregate series
  • 221kB Excel Workbook   Download 
    (Windows Excel 2007/2010 - Sorry, not Mac Excel compatible, due to our use of ActiveX controls.)

A full data-set is available to SGS Newsletter subscribers. It contains:

  • Data back to January 2006
  • Next 12 months of trend extrapolations
  • 201 industry and aggregate series (including some whose seasonally adjusted numbers are never published by the BLS)
  • 620kB Excel Workbook   Download   
    (Windows Excel 2007/2010 - Sorry, not Mac Excel compatible, due to our use of ActiveX controls.)



1.  Current Employment Statistics

2.  “Because CES revises only two months of estimates each month, the fourth month back from the current first preliminary estimate is adjusted using a different set of seasonal factors than the third month back. For example, with the release of October first preliminary data, factors are revised for September and August, but not July."    Technical information: Revisions to CES data for late sample reports, annual benchmarking, and other factors

 3. "Revisions to data for previous months also may produce gains in accuracy, especially when the original data are themselves regularly revised on a monthly basis. Numerous revisions during the year, however, should be avoided, because they tend to confuse data users and substantially increase publication costs."   Employment and Earnings, January 2004, Tiller, Richard B. and Evans, Thomas D., "Revision of Seasonally adjusted Labor Force Series in 2004," , Page 2.  [Note: This quote is taken from a discussion of seasonal adjustment of the related Current Population Survey, but is equally applicable to the Payroll (Current Establishment) Survey.]

4. "Employment in transportation and warehousing rose sharply in December (+50,000). Almost all of the gain occurred in the couriers and messengers industry (+42,000); seasonal hiring was particularly strong in December." Bureau of Labor Statistics

5. Free Exchange, The Economist. January 6, 2012  3rd Paragraph