How Bad Data Warped Everything We Thought We Knew About the Jobs Recovery

You know something is really boring when economists say it is.

That's what I thought to myself when the economists at the Brookings Institution's Panel on Economic Activity said only the "serious" ones would stick around for the last paper on seasonal adjustmentzzzzzzz...

... but a funny thing happened on the way to catching up on sleep. It turns out seasonal adjustments are really interesting! They explain why, ever since Lehmangeddon, the economy has looked like it's speeding up in the winter and slowing down in the summer.

In other words, everything you've read about "Recovery Winter" the past few winters has just been a statistical artifact of naïve seasonal adjustments. Oops.

Okay, but what are seasonal adjustments, and how do they work? Well, you know the jobs number we obsess over every month? It's cooked, in a way -- but not how Jack Welch thinks. For example, the economy didn't really add 169,000 jobs in August. It added 378,000 jobs. But that 378,000 number doesn't tell us too much. See, the economy pretty predictably adds more jobs during some months more than others. Things like warmer weather (which helps construction), summer break, and holiday shopping create these annual up-and-downs. So to give us an idea of how good or bad each month actually is, the Bureau of Labor Statistics adjusts for how many jobs we would expect at that time of year. This doesn't change how many jobs we think have gotten created over the course of the year; it changes how many jobs we think have gotten created each month of the year. You can see how that smooths out the data in the chart below from Johns Hopkins professor Jonathan Wright's Brookings paper. It compares the adjusted (blue) and unadjusted (red) numbers for total employment going back to 1990.

But there's a problem. The BLS only looks at the past 3 years to figure out what a "typical" September (or October or November, etc.) looks like. So, if there's, say, a once-in-three-generations financial crisis in the fall, it could throw off the seasonal adjustments for quite awhile. Which is, of course, exactly what happened. The BLS's model didn't know about Lehman. It only knew about the calendar. So it saw all the layoffs in late 2008 and early 2009, and interpreted them the only way it knew how: as seasonality, not a shadow banking run.

And that messed things up for years. Because the BLS's model thought the job losses from the financial crisis were just from winter, it thought those kind of job losses would happen every winter. And, like any good seasonal model, it tried to smooth them out. So it added jobs it shouldn't have to future winters to make up for what it expected would be big seasonal job losses. And it subtracted jobs it shouldn't have from the summer to do so. You can see Wright's estimate of just how much this changed the monthly jobs in the chart below, which I've annotated with when the Fed stopped and started its unconventional policies. Notice a pattern?

The Fed has stepped on the gas when seasonal adjustments have made the recovery look weaker than it actually was. And the Fed has stepped off the gas when seasonal adjustments have made the recovery look stronger than it actually was. Now, this is certainly suggestive, but it's not dispositive. As Wright points out, Fed economists are aware of Lehman's seasonal distortions: it's why they changed their seasonal adjustments for calculating industrial production. But there is still a question how aware the policymakers on the Federal Open Market Committee are of this. Indeed, St. Louis Fed president James Bullard said one reason they didn't taper their bond purchases in September was weak data -- and that "sometimes the jobs report can change the whole contour of how the [FOMC] look at the data." (Though, to be fair, House Republicans threatening to blow up the world economy again was probably a bigger reason for the no-taper). In other words, bad data might be influencing the Fed's bad stop-start policy.

Just how bad are the data? Well, keep in mind that the jobs report's margin of error is supposed to be about 90,000. But these post-crisis seasonal errors have almost doubled it to about 170,000. That's right: the jobs report's real margin of error has been about as big as the average jobs report itself the past few years. Now, the one bit of good news here is this effect has already faded away for the most part. Remember, the BLS only looks back at the past 3 years of data when it comes up with its seasonal adjustments -- so the Lehman panic has fallen out of the sample.

Here are two words we should retire: Recovery Winter. It was never a thing. The economy wasn't actually accelerating when the days got shorter, nor was it decelerating when the days got longer. It was mostly growing at the same, kind-of-miserable pace. Of course, we journalists (myself included) scrambled to explain what turned out to be a spurious trend: it was the pentup demand for housing or cars or ... something that had the economy looking up every winter. Eventually some Wall Street firms, and journalists like Cardiff Garcia of FT Alphaville, began to suspect something was screwy with the seasonals. But in the meantime, everyone else showed off our infinite capacity for rationalization. There's always a story you can tell, and we certainly told them. After all, stories are more interesting than disclaimers about margins of error and seasonal adjustments.

Now, seasonal adjustments might not sound sexy, but there's nothing sexier than getting the jobs numbers right. They matter for the Fed. They matter for markets. And they matter for our own understanding of the economy.

The BLS can, and should, do better.