Thursday, November 5, 2009

The Establishment Survey: Musings on Seasonal Adjustments and the Birth/Death Factor

This piece is a long time coming for me, as I wanted to examine exactly what is going into the very important (to the markets at least) Establishment Survey data that is released the first Friday of every month with the Employment Situation Summary by the BLS. As most of those reading this know, there are two integral factors that are used to adjust the data gleaned from the monthly Establishment Survey before it is released; the birth/death adjustment and a seasonal adjustment (applied after the birth/death adjustment). These adjustments greatly influence the actual reported employment data and are yet poorly understood and almost never analyzed by the financial media. First I will examine the seasonal adjustments. Then I will look at the much maligned by the birth/death adjustments. And finally I will draw some conclusions on what we are really seeing with the data.

Every month the BLS adjusts the data gained from the Establishment Survey by a "seasonal factor" in an effort to reduce the noise in the survey from well known regular employment changes (ie Christmas hiring/firing, summer jobs, etc). These adjustments are important because they help us glean a more accurate and timely picture of what is really happening with employment across the nation. Since the non-adjusted data is also published each month, it is relatively easy to get a rough (and these are rough) estimate of the seasonal adjustments for each month. To further reduce "noise", I averaged the adjustments over the 2004-2008 (2009 where possible) period.

January - 1.62
February - 1.15
March - .63
April - .03
May - (.49)
June - (.79)
July - .16
August - .12
September - (.26)
October - (.72)
November - (.80)
December - (.56)

This shows us that seasonally we "add" in a lot of jobs in January and February, as this adjustment is applied to a current base of around 130 million (ie so every .1% adds or subtracts 130,000 jobs) and subtract a fair number of jobs going into the end of the year (which is why it will be difficult to see gains in jobs prior to January/February). The adjustments have been very consistent over the time period, with the biggest fluctuation coming in August, which went from a .2 add in 2004 to a .04 add in 2009 (but decreased gradually). the largest recent change was from a 1.65 add in January in 2008 to a 1.53 add in 2009 (a difference of about 150,000 jobs). But in most cases, the seasonal adjustments are pretty consistent (as one would expect them to be, as seasonally little if anything has changed over the last several years).

On the the maligned birth/death adjustment. This adjustment made each month to the Establishment Survey (before the seasonal adjustment) is essentially a mark to model approach to estimate new business formation following the loss of jobs or closing of businesses across the country. Since this adjustment is applied prior to the seasonal adjustment and since it is supposed to be reflective of business formation or destruction, one would expect this number to vary greatly both on a monthly basis and also on a yearly basis depending on underlying economic conditions. What I found though is the exact opposite of my expectations (again taking the average of the 2004-2008 period):

January - (269,000), but remember the January number also includes the benchmark revision for the previous year and is thus a poor indicator
February - 116,000
March - 147,000
April - 276,000
May - 206,000
June - 174,000
July - (38,000)
August - 123,000
September - 36,000
October - 65,000
November - 36,000
December - 65,000

Now, if you are wondering why I took averages, it is because the monthly numbers were very consistent over the time period, which implies that the birth/death factor is more of an additional seasonal adjustment than it is a model of business creation/destruction. With the exception of July, most months had a variance of less than 10% each year, which is quite odd considering that this "model" was supposed to estimate creation/destruction from survey respondents/non-respondents and one would expect the adjustments to have much more variance, especially during a deep recession like we experienced this year (which also included very tight credit conditions that should have greatly curtailed new business formation), but what we have actually seen this year are birth/death adjustments that mimic the previous averages. This leads me to believe that the model either a) is broken and cannot take into account adverse economic conditions or b) that the model is essentially a second seasonal adjustment, in which case it is probably more appropriate to eliminate the birth/death adjustment and simply change the seasonal factor (in most cases this would be by about .1%).

To conclude, the data above should enable us to more accurately examine the monthly Establishment Survey data by keeping close tabs on any changes in the seasonal adjustments or birth/death adjustments that fall outside their averages. This is important as we begin to emerge from this recession, as I would be very concerned if we see the seasonal adjustments for October-December move up (ie become less negative), as this would "create" more jobs that should be factored out under normal circumstances. i would also question any birth/death adjustments that fall outside their typical average as well (and since the seasonal factors are all relatively small you can essentially count the entire birth/death adjustment into the seasonally adjusted employment number, since it will only change by a statistically insignificant couple thousand jobs at most). Finally, keep in mind that the margin of error on the Establishment Survey is around 100,000 jobs and thus any number inside of that (plus or minus) tells us very little.