Thursday, February 14, 2019

JOLTS February 2019 (Dec 2018 data)

Sorry about the delay in adding the latest data to the forecast plots, but unfortunately something has happened in my upgrade to Mathematica 11.3 that's broken the single prediction errors on some of the JOLTS graphs. I've been trying to figure it out. Overall, it's another "status quo" data release with data basically in line with the models. The openings turned out fine, but the errors for the projection for the recession counterfactual broke — I left it off:

As a side note — if there's another major data revision in the same direction as the data release around the Fed's March meeting in 2018, the negative deviation could largely go away.

My guess for the source of the problem is that the treatment numerical precision was updated such that the silly coefficients in the single prediction errors started being treated properly in terms of precision instead of being cancelled algebraically. Silly coefficients? Yes — how's 7.306167312812677×101813 for you? I think the source is the years — 2016 is unnatural in many senses. Yes, I know, I should have subtracted out 2000 before putting them through the fitting algorithms. But it worked fine up until this week. I guess my hard drive has worked fine up until this week, too ...

In case you're interested, here's how some of the other graphs turned out:

The Mathematica 11.3 upgrade came along with a transition to Windows 10, so I can't just revert back to 11.2. I may have to move these to my old computer with Mathematica 10 point something.


Update 15 February 2019

Back to Mathematica 10.3, and we're up and running again (click to enlarge):

And here's the alternative Hires model based on this collection of dynamic equilibrium relationships:

The hires data still doesn't show a deviation. Based on this model which puts hires as a leading indicator, we should continue to see the unemployment rate fall through May of 2019 (5 months from December 2018, which is the data that was released this week).

Friday, February 1, 2019

Unemployment ticks up

I think this is the longest I've gone without blogging since I started my blog! It's due to a confluence of factors — writing my second book (now about 1/2 written, and I'm learning so much finally looking at all these dynamic equilibrium models together from the 30,000 foot level), and I've become ridiculously busy at my real job. It means I missed the 2nd anniversary of the dynamic information equilibrium model. The model came out as a series of posts, first for JOLTS data (10 Jan 2017), then for the unemployment rate (11 Jan 2017), and finally a forecast of the unemployment rate (18 Jan 2017). That last forecast was compared with a forecast from the FRBSF that went until the end of 2018 — it was flat at about 4.7% the entire period. As I said in that last post:
The interesting thing is that in either [dynamic equilibrium] model we shouldn't expect the flattening out over the two year period [Dec 2016 to Dec 2018] we see in the FRBSF model. We should expect either 1) a recession to start raising unemployment, or 2) a continuing decline (albeit at a slower rate). A constant unemployment rate won't happen, and in fact generally doesn't happen [1]. We might be able to test the various models here [2].
And we could!

The uptick in the latest data to 4% is being blamed on the US government shutdown in the business news, but it's really well within the normal fluctuations of the unemployment rate so any speculation as to its cause is pretty much unfounded without detailed micro data. As a side note, there are various reasons why people think the Fed paused its rate increases at it's meeting this past week — discussed thoroughly by Frances Coppola here. However, I think Tim Duy has the best theory of the Fed's reaction function: as long as inflation isn't a problem, the Fed doesn't hike rates if unemployment is flat. 

The most recent data shows a relatively flat unemployment rate for the past few months, much like in 2016. Basically, the FOMC has some model like the FRBSF model in its collective head of a constant equilibrium unemployment rate. And it's precisely that model which led the FRBSF forecast astray.

Wednesday, January 9, 2019

Wage growth accelerates (and will until there's a recession)

I haven't compared the wage growth forecast to the Atlanta Fed's wage growth tracker data in awhile (last time was here). The recent data remains consistent with the forecast accelerating growth (increase in the growth rate):

(The "Hatzius" marking is from this post comparing the highly paid Goldman Sach's chief economist Jan Hatzius with the dynamic information equilibrium models.)

This model comparing various labor market measures estimates that wage growth lags changes in JOLTS hires by 11 months, and changes in the unemployment rate by 6 months — that is to say a recession (defined as a spike in unemployment) precedes a drop in wage growth. An intuitive interpretation is that recessions "cut-off" wage growth.

What is also interesting is that wage growth appears to experience a negative shock when it reaches NGDP growth, and the latest data is beginning to rise above the NGDP model:

There is an intuitive explanation behind this effect: if average wages are growing faster than average growth, eventually it should start to degrade firm profitability. Over the next year we should get a good test of the usefulness of the dynamic information equilibrium framework and this hypothesis.

Tuesday, January 8, 2019

JOLTS day: January 2019

The JOLTS data for November 2018 came out this morning, but there isn't much change in the assessment of the possibility of a future recession (per here). The openings data is still showing a negative deviation:

One thing that is a bit more clear is that the log-linear slope (i.e. the "dynamic equilibrium") of the JOLTS openings data for the past 18 24 months is definitely lower:

That's a linear fit to the (log of the) openings data since 2011 up through 18 24 months ago alongside a quadratic fit to the full range.

The hires data still doesn't show a deviation. Based on this model which puts hires as a leading indicator, we should continue to see the unemployment rate fall through April of 2019 (5 months from November 2018):

The gray dashed line is a counterfactual recession of average size and width (steepness) that reaches the edge of the confidence band in April of 2019 — it gives a rough estimate of the soonest a typical recession will show up as a rise in the unemployment rate based on the linked model that combines several dynamic equilibrium models of different measures into a single system. However, as noted above, openings might well be the leading indicator in the next recession (hires led the 2008 recession and appears to have led the 2014 mini-boom, but that's a very limited number of shocks to work with). The hires measure could also be off by a couple months based on estimated error.

Part of the reason I put these models out there is to make predictions — right or wrong. Being wrong tells us something just as valuable as being right!

Monday, January 7, 2019

Employment situation: unemployment rate Dec 2018

I'm just now back from vacation, so I missed the December unemployment situation data. The (seasonally adjusted) unemployment rate ticked up in December to pretty much exactly where the dynamic information equilibrium model expected it two years ago back in January 2017:

As was discussed in early December, the DIEM was much more successful than the overshooting FOMC and FRBSF forecasts (click to enlarge any graph):

What's even more interesting is that the more recent forecasts from the FOMC and FRBSF (vintage September 2018 and June 2018, respectively) are now undershooting the data while a comparable vintage forecast from the DIEM (April 2018, to compare with the CBO forecast) is pretty accurate:

It's as if the FOMC and FRBSF forecasts "over-learned" from their tendency to be too high as unemployment continued downwards but now are too low as the log-linear decline (in the DIEM) flattens out.



Here is the DIEM forecast of the unemployment rate for African Americans getting it right over the same two year period:

There's also the "prime age" civilian labor force participation rate:

Tuesday, December 25, 2018

Markets continuing to fall

The US markets are closed today for the holiday, but the Nikkei dropped 5% (the Nikkei, however, is far more volatile than the US markets so this doesn't represent anything too out of the ordinary). I've updated the S&P 500 counterfactuals based on the latest data (keeping the "lol capitalizm iz doomd" fit that goes to zero because I love dark humor):

Funny enough (again, dark humor) the unconstrained fit now matches up with "median shock amplitude" fit (the two middle dashed blue curves are basically on top of each other). The truth is that fitting the leading edge of the data to a logistic function (and then taking the exponential) is not terribly stable, so expect many many revisions [1] to that counterfactual until we are about 1/2 the way through the shock.

Here are the Nikkei and the NASDAQ (the latter is pretty similar to the S&P 500). Click to enlarge:


Update 26 December 2018

And today it jumped back up a bit. New unconstrained counterfactual amplitude is now a bit smaller than the median shock. Probably a good place to link to this post about volatility regimes. Click to enlarge:


Update 7 January 2019

The estimates are back near the originals when we incorporate the recent gyrations:



[1] For an example, see the undershooting and overshooting on the estimate of the size of the Great Recession in the unemployment rate data (click to enlarge):

Saturday, December 22, 2018

Dynamic equilibrium model forecast performance: PCE inflation

This is another dynamic information equilibrium model forecast that has been ongoing for almost two years — core PCE inflation:

Part of the reason for this model is to lend credence to the idea that inflation is about the size of the labor force. We should expect 1.7% core PCE inflation (or 2.5% CPI inflation, all items) unless there are dramatic shifts in the size of the labor force like women entering the workforce in the 70s or men and women leaving the labor force in the aftermath of the Great Recession. The latter is the reason for that slight deviation in 2014 and the "lowflation" which has largely dissipated. Note that the relative sizes of the effects (surging or flagging inflation) agree with the relative size of the causes (labor force changes).

Thursday, December 20, 2018

This is fine.

Title reference.

We're out of the AR process volatility (estimated since 2010, blue region), but not yet out of the approximately 70-year 90% volatility (January 1, 1950-December 31, 2016, red region) of the S&P 500 estimated using the dynamic information equilibrium model (DIEM). In fact, the S&P 500 has been this low before (relative to the the DIEM) — in February 2016. That was towards the end of the period of decline in the months after the first Fed rate increase since the 2008 recession. The index subsequently recovered. Prior to that, we skimmed the lower edge of the 90% confidence interval in 2011 as Europe was going into a double dip recession. We can see both of these in the a longer view:

But also looking at the longer run, most of these dips are just that — dips associated with a recession. There are only two major collapses that warrant adding a shock to the model (dot-com bust and 2008). Of course, it's possible to model the longer negative shock in the 70s as a series of smaller shocks, but none are close to the scale of 2001 and 2008.

Given most of the history of the market, we should expect the current dip — should it extend significantly below the confidence limit — to be just that: a dip likely associated with a recession or at least (seemingly unnecessary) Fed rate hikes. However, given the recent history of the market, we might expect a larger collapse as part of what I've called the "asset bubble era". This period, roughly since the 90s, is period after the demographic shock of the 60s and 70s has faded, inflation has subsided, and e.g. labor force participation ceased rising rapidly.

Note that as part of the asset bubble era, the asset bubble doesn't have to be reflected as a bubble in the market itself. The dot-com bubble was, but the "housing bubble" was accompanied by basically equilibrium growth in the S&P 500 index. However as part of the asset bubble era it might be accompanied by a new shock. Of course, it's quite early so estimates of the shock parameters are going to be wildly uncertain. I included several counterfactuals (dashed blue) with different constraints in this graph:

Two of the paths constrain the amplitude of the shock to be the median (absolute value) amplitude of all the previous shocks. One of those leaves the shock duration and timing to be fit to data, the other just leaves the duration. The timing was constrained to be the value indicated by the timing estimated from yield curve inversion. The third (and smallest) fits all three parameters. This is probably showing the undershooting and overshooting that is a drawback of fitting logistic functions to partial shocks.

I also included my joke path ("lol capitalizm iz doomd") I showed on Twitter the other day where the S&P 500 collapses to zero. I'd say that is unlikely (it is, however, the result of doing the fit with just a fixed shock timing).

Have we entered a new era since the 1990s where recessions coincide with major collapses in the stock market? Or will we return to the era before the 1990s where markets fall during recessions, but rebound quickly? My hunch is the former and I'm not looking forward to the economy in 2019.


Update 21 December 2018

We've pierced the toast the 90% confidence interval (updated counterfactuals):

Friday, December 14, 2018

Information transfer economics: year in review 2018

We're coming up on the end of 2018. I'd first like to say thank you to everyone who has been reading and sharing the blog posts and the tweets. I've slowed down a bit — only 169 posts in 2018 up until today, well below the peak of 375 in 2015. And many of those posts were just checking the validity of the forecasts with each data release! In part, that's due to the fact that a project I was working on where I had to travel once a month to the middle of nowhere for one to two weeks ended in 2016. I don't have as many evenings sitting in a hotel room with nothing to do besides research and writing these days. Only one post from this year cracked the top 10 of all time on this blog — it was my critique of macro written from the perspective of having seen both the 10 years of econ criticism since the Great Recession alongside 10 years of defenses. I titled it Macro criticism, but not that kind in reference to the torrent of "lazy econ critiques" as well as Noah Smith's Bloomberg article (Econ Critics Are Stuck in the Past) written a few weeks before. I also think it's relevant that my critique of macro comes alongside empirically accurate forecasts of e.g. the unemployment rate compared to Fed macroeconomists (FRBSF, FOMC) — at least it's not lazy.

Dynamic equilibrium

This year kicked off with me posting the dynamic information equilibrium paper to SSRN. It also included the application to ensembles of information equilibrium relationships (markets) — i.e. "macro". While I had been working on it since the summer of 2017, a message from Fabio Ghironi at the UW econ department prompted me to finish it over the holiday break. That eventually lead to my participation in the "Outside the Box" economics workshop Fabio organized in October.


Another unexpected message came from one of my blogging heroes — Cosma Shalizi. Blogging is really about writing that wouldn't find a place in any traditional outlet. Although he's not blogging as much these days, he wrote one of the greatest examples of the form with a perfect title (hilarious enough on its own, although the reference now may be lost on the younger generations). A book review that ties together computational complexity and economics, I link to it any time I can find a way to work it in.

Cosma wrote me to let me know he was reading my book (!) and that he'd had the same interpretation of Gary Becker's paper written up in a draft blog post. I told him he should post it — it does a great job of addressing the supply side (which I sort of skipped over). He did, and I tweeted about it.

I should also add that I owe a debt of gratitude to another great blogger, economist David Glasner, for pointing the paper out to me a few years ago. 

Rethinking interest rates

This year prompted me to rethink the interest rate model that was actually one of the first decent models I put together with the information equilibrium framework. The large deviation that began with the 2016 election is well outside the normal range of errors — but such deviations have been signs of recessions (and model error is generally worse close to yield curve inversions). Recently, interest rates have been coming back down and a future recession could bring the model right back in alignment with data (i.e. we might have just underestimated the errors). As it is, it seems the dynamic equilibrium model where the interest rate is a price is accurate, but the supply and demand related to it are only approximated by the monetary base (minus reserves) and aggregate demand measured by NGDP.

A recession in late 2019?

I put my neck out and said that the negative deviation in the JOLTS openings data was the first signs of an upcoming recession (i.e. a slowing in businesses expanding). The shocks to the JOLTS data series appear to lead the rise unemployment rate by almost a year, and with the yield curve heading for inversion on roughly the same time scale I'm around 80% confident in this forecast. There's also wage growth rising to levels comparable with NGDP growth, which seems to be associated with the past few recessions. At the very least, I hope to learn something!

A new book

At the end of this year, I set myself a deadline for my second bookA Workers' History of the United States 1948-2020 — for the end of next year. Collecting the information together has actually led to some additional insights (e.g. the effect of unions on inequality and the (lack of) a housing bubble).

*  *  *

This has been another great year, and I'd like to thank everyone again for reading and sharing. Without you, this would likely just be a series of crackpot comments on other people's blogs (which is really not too far off the mark).

If you're interested, here's a link to the year in review for 2017. Also, click to enlarge any images above.

Wednesday, December 12, 2018

CPI forecast performance over the past two years

Unlike these forecasts, this one of US CPI inflation has been doing well (both year over year and continuously compounded, click to enlarge):

Here's the CPI level:

Here are the details about the re-estimate of the size of the post-2008 recession demographic shock (shown as the dashed line in the graphs).