A +316,000 NFP Print On Friday? The BLS Seasonal Fudge Factors Make It Very Likely

Tyler Durden's picture

While the ADP number today of -84k was not much of a surprise to consensus, everyone is focused on this Friday's much more important NFP release, which economists expect to post the first rise in 2 years at +10,000. Yet an analysis out of Stifel Nicolaus points out that due to various seasonal adjustments, an NFP print of up to +100,000 could be expected (which incidentally does not reflect anything favorable at all about the actual employment picture as it is due exclusively to seasonal fudge factors). In fact, Stifel argues, a print of +316,000 is theoretically possible (we await Goldman's whisper leak to provide additional color). In either case, should the NFP come at that level, we fully anticipate the market will react like a stung, rabid bull, as computers buy blindly on the headlines, with no regard for the underlying adjustments.

For those who care where such an aberration of a number could come from, here is one explation courtesy of Michael Widner at Stifel, Nicolaus.

  • Any rigid predictive model will occasionally be predictably wrong.
  • We believe we are coming up on such an event with employment data.
  • We believe the jobs reports later this week and for the next few weeks will paint a rosier picture of the employment market than is expected or warranted.
  • This is driven by the interaction of how the methodology works, how seasonal adjustment factors are calculated and applied, and the specific economic condition we are currently in.
  • Non-farm payroll appears the likely big beneficiary on Friday and we see a strong chance of a +100K or better print (vs. ADP data today at -84K).
  • Thursday's claims numbers are just a precursor to next week's as this mathematical peculiarity plays out but we believe there is a reasonable likelihood of seeing reported initial claims fall below 400K and continuing claims fall below 4M in the January 14 data.
  • The issue here is that the Department of Labor's fairly rigid adjustment models assume we had the usual surge in seasonal end-of-year hiring in 2009 and that the corresponding post-Christmas and post-New Year cutbacks will follow.
  • It doesn't account for the fact that average weekly hours worked before the holidays were at an all-time low (indicating excess labor capacity), that retail shopping expectations were muted, and in general seasonal hiring was fairly low. As a result we believe it overestimates impending seasonal firings and will subtract too high a number from actual results.
  • There are additional peculiarities in the models that we believe will have predictable effects. While the non-farm payroll data oddities are less predictable we believe there is a quantifiable positive bias which suggests we should see good numbers there this month and next.

And the most relevant bit for those computers among you who only read the first 2 letters of a headline before lifting any and all offers:

  • Our point is not to try to accurately predict what the numbers will be. There is far too much uncertainty to have any confidence in doing that.
  • Rather, our point is to demonstrate that the odds are skewed toward data looking better than expectations because the methodologies at this particular juncture have a bias.
  • This is independent of any actual jobs market improvement or decline.
  • The bad news is that the adjustment models are built by definition to be zero sum across the year. So the bias toward under-reporting now will come back as over-reporting as we move into the spring and the model expects seasonal hiring rather than firing.
  • We remain highly concerned about the employment situation and expect it to weigh on economic recovery much more than usual in this cycle.

Yet while January and February numbers will likely under-report, with the launch of the census hiring, those numbers will likely offset the adjustment shifts throughout the first half of the year, painting an abnormally and transiently rosy picture for a good six months. And no, this is not artificial consumer boosting. At least not in the purest sense, as that would be illegal, even for this government. And what is going on here is perfectly legal.

Some additional information from Stifel that indicates just how the DOL's adjustment model works in various situations:

In the simplest quantitative terms the Department of Labor's seasonal adjustment model will assume 33% of this week's initial claims are seasonal and will eliminate them from the as reported number. Next week it will jump to 45%.

This is based on historical patterns that are similar but quite a bit more sophisticated than simple regressions. But their chief limitation, and thus the basis of this commentary, is that they do not take into account the specifics of current labor markets or recent employment trends.

This will likely result in initial claims reaching their highest level since last February next week while the model reports it as the lowest level in 18 months.
 
Initial claims data is distorted in two ways by the adjustment model. The first we described in the introductory bullets. The summary is that it expects a large number of seasonal firings that we believe will come in light because the model doesn't know that the corresponding seasonal hiring was light. If we had to guess we'd say this adds a weekly bias of 15K – 20K currently. But don't look for that as a w/w change as the bias started to roll in during November and will continue into February. As a result the model is understating (over-
correcting) the data each week. This effect peaks next week.
 
The second issue is that the model works using multipliers rather than additions or subtractions. Specifically it expects a 15% w/w increase on an already inflated number. Put in terms of actual jobs the model will back out 223K expected seasonal job losses from headline numbers over the next two weeks. In a more normal environment (e.g., 2003 – 2006) this would about 160K.
 
Put differently, the model assumes that when firings are already running high the seasonal component will be amplified as well. Intuitively we expect the opposite. Firms have already cut to the bone and since the seasonal hiring was modest the corresponding firings will likely also be. We expect this to add a bias of roughly 60K over the next two weeks, with most of it coming next week rather than this week.
 
All in we believe there is a high likelihood that reported initial claims will fall below 400K next week (1/14) even if the jobs market doesn't actually improve. This week is a bit less clear and we expect the reported number to actually come in fairly flat w/w and expect last week's 434K number will likely be revised higher.
 
We are fond of the following chart that shows how these issues have impacted the reported data over the past six months. The bars show actual initial claims, the line shows the as-reported figures. The reported data has shown a steady decline since July while the actual initial claims have been on a steadily rising trend. In other words people have been losing jobs at a faster and faster pace since September, but the adjustment model expected job losses to rise even faster than this and thus turns the rising trend into a falling trend.

Source: Department of Labor
 
Mathematically if the model was built to assume seasonal hiring and firing was a percentage of total employment (e.g., 1% of total employment ebbs and flows seasonally) rather than a percentage of total firings (e.g., 33% of firings this week will be assumed seasonal) the chart would tell a much different story. But it bears repeating this is a zero sum game. What the model credits us with now it will demand of us later.

The same adjustment mechanism will also wildly distort the Continuing Claims picture:

Continuing claims data suffers from the same biases as initial claims data as it is built using the same models. When unemployment claims are running high it assumes a greater number of them are attributable to seasonality. And we happen to be heading into the weeks where the highest percentage of claims will be stripped out in the name of seasonality.
 
We spent 27 consecutive weeks this year (April through October 3) with the reported continuing claims numbers in the 6M+ range. Over the next 10 weeks the number gradually declined to just under 5M. In the next two weeks we believe the adjustment model can push the figure to 4.1M and possibly produce a below 4M print in January. (At this point we expect it to rise again in the spring.)
 
There are two factors at work here. The first is the fact that individuals can only remain in the data series for 26 weeks. After that they are eliminated, even if they are still jobless and collecting benefits (which can last up to 79 weeks now). As a result the number is generally pressured lower by the fact that people are getting kicked out the back end of the data faster than they're coming in the front end, even if they're still jobless.
 
The second factor is the seasonal adjustment that suffers the same biases we discussed for initial claims. Specifically 14% of continuing claims this week will be assumed to be seasonal (up from 2% last week) and 30% of next week's claims will be assumed seasonal. This will translate into 1.2M people being erased from the headline number next week, vs. only 109K last week.
 
Mathematically if the actual number of people in the continuing claims data rises by less than 120K over the next two weeks we'll see a reported number under 4M on January 14. This would represent a huge (apparent) improvement, as we barely touched below 5M last week for the first time since last February.
 
The following chart depicts the unadjusted and adjusted ongoing claims data.
The points we want to highlight are 1) the reported data (red line) has shown steady improvement since July, 2) the unadjusted sum of people collecting benefits (continuing claims plus those on extended benefits) has actually been steadily rising and is now at an all-time peak, and 3) the disparity between these trends is about to get much wider due to the adjustment methodology. We could conceivably have nearly 11M people collecting unemployment and see the data reported as a 3.something million figure.

The bottom line to us is that we expect the headline continuing claims numbers to drop sharply over the next two weeks driven purely by the methodology.
Whether the labor market actually improves or decays at all is going to be completely swamped by the adjustment factor in our view.
 
The big issue we see is that the sharp drop in reported continuing claims seems likely to give the impression that job creation is taking off, when in fact it will actually say nothing about job creation in our view.

But most disturbing for the topic at hand, NFP, here is why the Obama administration will likely be spinning "amazing" job data for the next several months:

For the initial and continuing claims data the seasonal adjustment factors are fairly rigid and published by the Department of Labor well in advance. The non-farm payroll adjustments used to be similar in that regard, but since 2003 that has changed. The Bureau of Labor Statistics no longer publishes adjustment factors in advance, and in fact now actually derives them on the fly as the data comes in. As a result it is now much more difficult to predict how seasonal adjustments may skew the non-farm payroll data. [Thank you Beijing ministry of data dissemination]
 
That said we do know that the seasonal adjustments need to be a zero sum game over the course of the year. And we do know the range of historical adjustment factors each month. As a result we can make predictions of the likely range they are going to exhibit in any given period.
 
The following chart depicts the historical seasonal adjustment factors by month for the non-farm payroll reports since 1996, when the current model was implemented. The vertical axis reflects the monthly adjustment factor, the horizontal reflects the month of the year. First we'll explain mechanically how this works and then talk about implications.

Mechanically the Bureau of Labor Statistics (BLS) surveys business each month and asks questions about their total payroll. If the BLS could talk to every employer the simple sum of all of their employees would be the unadjusted non- farm payroll figure. Of course they only make a statistical sample of businesses, biased toward large companies, but that's an issue for another day.We run with the assumption for now that their statistical sample is accurate and each month they have a correct tally of total employees in the unadjusted result.
 
To adjust for seasonality that unadjusted result is multiplied by the numbers depicted on the following chart. When jobs are seasonally weak the multiplier is greater than one, when they're seasonally strong the multiplier is less than one.
 

 
Here is the real crux of the issue, in our view, all the market ever really reacts to is the m/m change in seasonally adjusted non-farm payroll. And that number is very heavily influenced by the m/m change in the seasonal multiplier.
 
As we said earlier, we don't know where the December multiplier is going to come out. (Chances are the BLS is still deciding that.) But looking at the data plot of the past 14 years suggests it will likely be in a reasonably narrow range.
 
The November point (end of the heavy red line in the chart) was oddly at an all-time low. Effectively the BLS attributed a higher percentage of the current workforce to seasonality last month than any time in the past 14 years. Which means they erased a record high number of people out of the headline non-farm payroll number.
 
As we look to December data (reported this Friday) if this seasonal adjustment multiple returns to anything in the range of historical norms it should provide a huge lift to the reported m/m change. In quantitative terms a return to the 1996–2008 average would create a seasonal lift of 431K to the as-reported m/ m change. In comparison the 1996–2007 actual December m/m change (unadjusted) was 116K. Put differently, if this was an average December for job creation and the adjustment factor returns to a historical average we would see a non-farm payroll print of +316K on Friday.
 
We see this as an unlikely scenario as we don't think the BLS would make that large a change to the seasonal factor and they do have discretionary control.
 
We have no way of knowing what seasonal adjustment factor the BLS will use this month but we do know that eventually this imbalance will have to balance out. (Again, seasonal adjustments must work out to a zero sum across the year.) We suspect the BLS will smooth out that seasonal adjustment imbalance over the next several months, so we may get an upward bias of say 80K this month, 100K next month, then 120K, etc. Conspiracy theorists would suggest this might be an effective way to create the illusion of steady improvement that could potentially become self sustaining. (If we believe the economy could succumb to a placebo effect.) But it bears repeating, eventually these adjustments must all sum to zero.

We won't comment on the last paragraph, safe to say that there is nothing about the current administration that would provoke us to say they may have a data-manipulation agenda. Nothing.

And, once again, to those who trade merely off headline bullets instead of ever reading between the lines, here are Stifel's concluding remarks:

Every now and then circumstances come along that in our view lead to predictable distortions in the data that will create the appearance of a trend, whether or not that trend is real. We believe we are at one of those points. What is really happening and what appears to be happening with the economy can be two very different things for a certain period of time.
 
For those inclined to trade on such anomalies we reiterate that the probability bias seems to favor employment data coming in well above consensus expectations even if the true trend is mild worsening. We expect this to last through January for claims data and likely a few additional months for non-farm payroll.
 
There is most definitely a high degree of noise in actual results but in our view the methodologies force the probability distribution heavily in favor of better than consensus headlines.
 
As a final comment we remain fairly bearish on the economic recovery despite our expectations for data to look strong in the short term.

We will provide tomorrow's Jan Hatzius' NFP estimate the second it is released. We will not be surprised if Goldman once again leaves the consensus estimate pack and ends up with a highly favorable (and mostly imaginary) number, which will be within significant digits of Friday's NFP outcome.