The seasonally adjusted number came in at a level everyone described as steady. The unadjusted number for that month was the weakest for that month in eleven years. Both are published. One of them was on television.
Seasonal adjustment is not a fudge and I have spent enough time with the methodology to be irritated by people who imply it is. Retail employment rises in November everywhere, every year, for reasons that have nothing to do with the economy, and a series that reports that rise as news is a series that reports the calendar. Removing it is the only way to see anything at all.
What is worth understanding is how the adjustment factors are produced, because that is where the interesting failure lives.
The factors are estimated from the recent history of the series itself. The procedure looks at how much this month has typically differed from the trend over the last several years, and removes that much. It is re-estimated regularly, so the factors move as the history moves.
Now consider a genuine structural change — something that alters what a given month looks like, permanently. A hiring pattern shifts because an industry reorganises. A holiday retail season stretches earlier and flattens. Something real happens to when work occurs.
For the first year, that shows up as a deviation from the seasonal pattern, which is exactly right: the adjusted series moves and the analyst sees something. By the second and third year, the change is in the estimation window. The procedure is now learning that this is what that month looks like, because for three years it has been. The factor absorbs it. And the adjusted series goes quiet, not because the change reversed but because the machine has decided it is seasonal.
A structural change and a new seasonal pattern are, to a procedure that infers seasonality from repetition, the same object. There is no version of the method that separates them, and the statisticians who build these series say so plainly in their own documentation, which almost nobody reads.
This is not a hypothetical and I want to be precise about the size of it, since precision is the thing I keep demanding of other people. It does not fabricate movements and it does not hide a recession. What it does is delay recognition of a persistent shift by roughly two to three years, and shift it into the revision history, where the annual re-estimation quietly rewrites the recent past and nobody reports the rewrite because a revision to a factor is not a story.
The defence, and I hold it myself most days, is that the alternative is worse. An analyst eyeballing unadjusted series against their own sense of what November should look like will find whatever they came in expecting, and I have watched that done, and it is not better. The procedure is honest, documented, and consistent, and consistency is what makes comparison possible at all.
So my practice is small and unsatisfying, which by now readers of mine will recognise as where I generally end up. I look at the adjusted series and I look at the unadjusted one against the same month in prior years, and when they disagree about direction I write down the date. Four times in ten years the unadjusted comparison was seeing something first.
The other six times it was noise, and I could not tell which was which at the time, and I still cannot.
