[Update: Corrected an error in my analysis, though it oddly turns out to have very little effect on the relevant results. Corrections are in italics. I had fit the equation with AR(1) for the wrong time period, ending in 2005 instead of 2014. Note that all fits should now be 12/2000 to 6/2014, the period of available JOLTS data.]
Back in 2010, there was a jump in US job openings (from an extremely low level) that was not accompanied by a commensurate decline in the unemployment rate. Some saw this pattern as an indication of increased structural unemployment, with job openings becoming harder to fill from a given pool of unemployed. At the time, I argued that it was not so: job openings arise, and it takes time for them to reduce the unemployment rate; necessarily, there is a period when the unemployment rate remains higher than what would earlier have been associated with that number of job openings. Then in 2012, I changed my mind. A closer look at the data, including the additional two years that had passed, showed that, for a given number of job openings, the amount of hiring had declined. That shift in the “matching function” suggested a change in the underlying relationship between unemployment and job openings, not just a temporary dynamic effect associated with the time it takes to fill new openings.
Recently some research has come out of the Cleveland Fed (cited approvingly by Paul Krugman) purporting to show that I was right the first time. Specifically, Murat Tasci and Jessica Ice conclude that “there is no shift” in the Beveridge curve (the empirical relationship between job openings and unemployment). They show that, in the years since that initial jump in job openings, the unemployment rate has fallen faster, and vacancies (job openings) have risen more slowly, ostensibly leading them back to the relation they had before the apparent shift in 2010.
I must say, first of all, that I don’t quite see their charts even appearing to show what they claim. It’s true that, in vacancy-unemployment space, the point for 2014Q2 is very close to the point for 2008Q3; so, in a sense, any shift that was purported to have happened after 2008Q3 would now seem to have been an illusion. But when I look at their chart, it looks like the shift actually happened between 2008Q2 and 2008Q3, when the unemployment rate rose and the vacancy rate failed to fall. For the first two quarters of 2008, the not-yet-Great Recession looked much like the previous recession; then in 2008Q3 it appeared to shift to a new locus. That apparent shift has not been reversed.
Comparing recent experience to the previous business cycle, it’s clear that we’re seeing a very different pattern this time around, not just in the intensity of the recession but also in the relationship between vacancies and unemployment. Tasci and Ice have perhaps succeeded in demolishing the view that the large increase in vacancies in early 2010 represented a shift in the underlying relationship, but to my mind, that view has always been a straw man. In any case it’s not a view that I ever held: my research seemed to suggest changes in August 2006 and July 2010, the latter happening just after the alleged jump, not before or during.
Anyhow, in the light of the recent work, I decided to update my research, and, as my title suggests, my overall view hasn’t changed from 2012, though the details are a little different. In my 2012 post I presented a model of the Beveridge curve, and my updated results can be described in terms of that model, but for the sake of universality I’m going to present them in a more agnostic way.
Start with the conventional “matching function,” which gives new hires as a function of unemployment and vacancies. Using the JOLTS data (and using the absolute levels of hires, openings, and unemployment, as I did in my 2012 post), we can try to fit a matching function of the from lnH = a + b*lnV + c*lnU. When I do this, I invariably get a negative value for c (regardless of specification details such as the inclusion of terms for autocorrelated residuals). No plausible theory of the matching function gives a negative value for c. (Surely it’s easier, not harder, to find people to hire if there are more people looking for jobs.) So I re-fit, leaving out the U term. (To put this another way, I’m fitting the equation with the constraint that c is non-negative, and I find the constraint to be binding.)
So I fit the equation with c=0, and I get a=4.52 and b=0.48, which would imply that hires are approximately proportional to the square root of vacancies, the same result I got in 2012. Also as in 2012, I find that the residuals are autocorrelated (a Durbin-Watson statistic of 0.5, far from the ideal 2.0), presumably because the relationship has shifted over time. So again I fit with an AR(1) term,
and again I find that this gets but this time it is not sufficient to get rid of the Durbin-Watson problem. (The Durbin-Watson statistic goes to 2.9, still far from where it should be.) So I added an MA(1) term and an AR(2) term, and this finally seems to be enough to handle the serial correlation problem. This time, with the AR(1) term ARMA(2,1) terms, a=5.71 a=5.68 (although this isn’t very meaningful because the value of “a” is effectively shifted by the AR(1) term ARMA(2,1) terms), and b=0.34 b=0.33, which would imply that hires are approximately proportional to the cube root of vacancies (but in a way that shifts over time).
The interesting part, though, is what the residuals look like, so here they are:
[Note: This is the corrected chart. The picture that originally appeared here is still up on Twitter and looks pretty similar.]
A couple of things are pretty clear about this picture. First, there was a shift that took place between 2005 and 2008. (The shift seems to be gradual, but given the amount of noise, it’s plausible that the shift could have happened in certain particular months, or even in just one particular month. From the chart, the most decisive part of shift seems to happen between November 2007 and January 2008, which, probably not coincidentally, was also the turning point of the business cycle.) Second, the shift does not appear to have been reversed. (If you look closely, you might see another shift in 2010, which then seems to be reversed in 2013, but both the shift and the reversal could easily be noise, and in any case the original 2005-2008 shift has clearly not reversed.)
So here’s my conclusion: something really did happen to make the Beveridge curve shift, and it was a persistent change. Whether it was genuinely “permanent” we of course don’t know yet (since the current business cycle, one hopes, has a way to go, and the shift could be reversed later in the cycle), and whether it was truly a “structural” change is a question that is above my pay grade. But I’m going to go with my best guess based on the available data and say that it looks like there was an increase in structural unemployment associated with the 2008 recession (or with what preceded and/or followed it).
DISCLOSURE: Through my investment and management role in a Treasury directional pooled investment vehicle and through my role as Chief Economist at Atlantic Asset Management, which generally manages fixed income portfolios for its clients, I have direct or indirect interests in various fixed income instruments, which may be impacted by the issues discussed herein. The views expressed herein are entirely my own opinions and may not represent the views of Atlantic Asset Management. This article should not be construed as investment advice, and is not an offer to participate in any investment strategy or product