Macro criticism, but not that kind

With all the tired and plain wrong critiques of economics out there that are easily shot down by even the most critical student of economics, I thought I'd try my hand at writing at one that might pass muster. I did write a book, but it was more aimed at taking a new direction; this will be a more specific critique.

First, let me avoid the common mistake of using the word "economics" but then exclusively talking about macroeconomics: my critique is being leveled at macroeconomics (macro). This is not to say I don't also have criticisms of microeconomics or growth theory, but rather let me just focus on macro because that is what most people are interested in. I'm pretty sure the comeback "Auction theory is successful!" isn't really going to cut it with the Post Crash Economics Society or in general anyone whose life was turned upside-down by the Great Recession.

Second, let me avoid the common mistake of saying macroeconomists don't think about X. They do. There's a good chance they've thought about X much more than you have. Instead, let me focus on how macroeconomists think about thinking about X — the context, the spoken and unspoken narratives, the institutional knowledge.

And finally, let me avoid the common mistake of decrying the use of math in economics (this time in general). Mathematics is an extraordinarily useful tool. I know — I'm a physicist. I don't think economists have "physics envy", but the charge does carry a nugget of truth that I'll get to later.

Many critics claim that macroeconomists failed because they were unable to predict the global financial crisis and global recession. Some critics of "mainstream" macro echo that claim and further claim that unrepresented schools of economic thought did in fact predict the crisis. Regardless of the truth of those claims, the real issue is that it is not currently known with any empirical certainty if financial crises or recessions are predictable (or whether the former cause the latter). There are decent thought experiments as to why financial crises or asset bubbles should be impossible to predict. This is frequently misinterpreted as an argument that bubbles can't exist. However I have to admit as a scientist that if you can only identify a bubble after it has popped it is at least plausible that the concept might not be useful.

But that real issue — that we don't know if major macroeconomic events are predictable — is further complicated by the fact that macro doesn't even know if macro time series of measurements like GDP or the unemployment rate are predictable outside of recessions. The best performing forecasting models tend to be things like vector autoregressions (VARs), but these models essentially 1) choose some set of macro observables, 2) measure their drifts, periodicities, variances and covariances, and 3) project into the future based on that knowledge. While this is a perfectly scientific undertaking, the understanding it delivers is little more than saying the time series are correlated randomness of a certain variety. The relative success of these kinds of models compared to models based on actual macroeconomic theory derived from thinking about people making decisions in some way or another should be a source of deep embarrassment for macro theory and theory-based models. It's as if the proverbial monkeys were able to type up Hamlet while the theorists were replacing the ribbon. However we still have economists like Olivier Blanchard and Lawrence Christiano [pdf] touting DSGE models as still useful for running policy experiments, or claiming they aren't designed for forecasting.

This failure points to the failure of what Noah Smith called big unchallenged assumptions. Theory-heavy models have lots of these, from the Euler equation (relating agents' view of the future to present consumption) to the Phillips curve (relating inflation and unemployment, i.e. the real economy). VARs have fewer assumptions — some assumptions still go into the choice of which macro time series to use. It is my impression that these unchallenged assumptions about the kinds of ingredients to use in macro models are the reason why including more of them leads to worse forecasting even when the economy is not in a recession. This is also why I think the foray into machine learning (championed by, for example, Susan Athey) might be extremely helpful for macroeconomics. Imagine if every machine learning model put zero weight on the interest rate!

This brings us to those meta-narratives. A lot of those theory-heavy macro models are generally based on the idea that the central bank's monetary policy and the government's fiscal policy are the drivers of GDP and inflation, and the sources of recoveries from recessions — or even their cause. The theoretical variables are interest rates, the price level, unemployment, and output (alongside their future values as expected by agents with ideal or bounded rationality). Economists like Paul Romer claim that the so-called Volcker disinflation of the 1980s represented "a clean test" of the importance of monetary policy, and that questioning whether the Fed under Volcker caused the 1980s recessions should be seen as a "yellow caution flag". Now Paul Romer isn't the only economist, but similar sentiments are expressed in graduate texts like David Romer's Advanced Macroeconomics. There are even people not named Romer that have written papers [pdf] about it. There is actually insufficient evidence to confirm or reject this narrative of events, and there are signs it might be the result of post hoc ergo propter hoc reasoning. For example, the yield curve inverts — often a good indicator of an upcoming recession in the US — and the stock market tanks in 1978, prior to Volcker's nomination to the Fed or the implementation of any change in monetary policy. Also note that a fall in conceptions [pdf] appears to precede economic decline measured by the unemployment rate by several quarters. Turnaround in Job Openings and Labor Turnover Survey data also tend to come before more traditional measures of economic decline. Neither of these were measured as indicators by macroeconomists at the time. There could have been many signs that the 1980s recession was already underway that went unreported because they are difficult to measure, or, as in the case of conceptions, not discovered until later. Note that the existence of possible leading indicators aren't the same thing as predicting recessions because macroeconomists don't know if they even have in hand the indicator with the longest lead nor know whether that indicator can be predicted.

Up another level of metaphysics, macroeconomics does not fully understand what a recession is aside from a general slowdown in economic activity. While heuristics such as two consecutive quarters of negative GDP growth are sometimes used, the NBER assembles a group of economists after a candidate for a recession appears to look at a large number of indicators and declare when that recession, if it is one, started and ended. Sometimes their results are not universally accepted — e.g. the early 2000s recession is given a fairly low probability using this metric. Now there is nothing wrong with this. Astronomers have revised their definition of what a planet is as recently as 2006. Physics has no idea what dark energy is. However, this lack of understanding does not seem to give macroeconomists pause when making assumptions about what a recession is or its cause in specific situations such as Del Negro et al adding financial frictions to explain the 2008 recession [pdf] that are independent of whether the model can describe other recessions. As Dani Rodrik says in his book Economics Rules, one model with a particular set of assumptions is applied to one specific situation while another with another set of assumptions is applied to another situation — usually the models and assumptions are selected post hoc. He says that's the "art" of economic theory. 

It's not the assumptions' realism or lack thereof that's the issue. The issue is that these models are the mathematical analog of Kipling's "just-so" stories. The leopard got its spots in this particular way, and that doesn't help you understand how the cheetah got its spots. As Feynman said in his famous Cargo Cult Science commencement address at CalTech in 1974:

When you have put a lot of ideas together to make an elaborate theory, you want to make sure, when explaining what it fits, that those things it fits are not just the things that gave you the idea for the theory; but that the finished theory makes something else come out right, in addition.

The model that is used to make the 2008 recession come out right doesn't make something else come out right, in addition. All too often, that's what the defense "macroeconomists actually do study X" really means: there is a just-so model that has been used to explain X. 

It's not just macro that has this problem; it has been brought up in evolutionary biology for example. It doesn't seem that there is a lot of internal criticism in macro of this kind of model-building, and in fact some economists actively seek out these kinds of stories. I also believe that it is a lack of an immune response of internal criticism to this kind of model-building lets a lot of "schools of thought" (usually just different sets of story elements) proliferate. Before I get the "there are no 'schools of thought'" rejoinder from a macroeconomist, let me just say that the aforementioned graduate macro text says:

Where the major macroeconomic schools of thought differ is in their hypotheses concerning [recessionary] shocks and propagation mechanisms.

While Romer says hypotheses, these tend to be more sets of assumptions about what a recession is.

The close relative of the just-so story is the escape from the self-imposed straitjacket; I can't do any better than the blog Mean Squared Errors in describing this:

Consider the macroeconomist. She constructs a rigorously micro-founded model, grounded purely in representative agents solving intertemporal dynamic optimization problems in a context of strict rational expectations. Then, in a dazzling display of mathematical sophistication, theoretical acuity, and showmanship (some things never change), she derives results and policy implications that are exactly what the IS-LM model has been telling us all along. Crowd — such as it is — goes wild. 

And let's be clear: not even the most enthusiastic players of the macroeconomics game imagine that representative agents or rational expectations are, in any sense, empirical realities. They are conventions, "rules of the game." That is, they are arbitrary difficulties we impose on ourselves in order to demonstrate our superior cleverness in being able to escape them. 

They are, in a word, Houdini's straightjacket [sic].

The meta-narratives that require these particular mathematical modeling elements end up making even the simplest macro models incredibly complex. The model of Del Negro et al contains an entire DSGE model, but the idea that a financial crisis can cause a recession doesn't really need that complex of a model — unless you're proving something else. It's only the fact that a DSGE model with Euler equations, Phillips curves and various microfoundation assumptions is the starting point for adding financial frictions that requires the complexity. And then, for all that complexity, the model doesn't actually do all that well (e.g. the shape of post-recession inflation is completely wrong):

This lackluster model has over 20 parameters (left-most column; the financial frictions correspond to the parameters that aren't in the other versions):

This is where that nugget of truth about physics envy comes in. Proving the existence of a just-so story that keeps the meta-narrative faith requires a level of complexity that far exceeds the accuracy of the model. When George Box said "all models are wrong", he was advising against building exactly this kind of Rube Goldberg device. The physics envy charge is leveled at exactly this kind of unnecessary complexity that doesn't result in better, more empirically accurate models. The charge is best understood as an attempt to answer the question: Why would macroeconomists do this? It can't be because they just enjoy algebra. I don't know; maybe they're just happy to write out lots of LaTeX symbols like in physics papers ... physics envy? Well, that's the best answer I've heard because otherwise it doesn't make sense!

One of the problems with the now standardized critique of unrealistic assumptions, failing to predict the crisis, or failing to add whatever "better" ingredients the author of the critique either personally researches or simply likes more is that it's so easily batted down because it's a caricature of macroeconomic theory from the 1990s or even the 1890s (in a similar way that many critiques of string theory are based on the state of string theory in the 1990s). More often than not the "better" ingredients (evolutionary biology, nonlinear dynamics, more accurate accounting) are simply another set of assumptions chosen to fit a narrative and build just-so stories — but a narrative and just-so stories the critic prefers. They aren't any more empirically accurate than the macro they're criticizing (and often haven't been used to construct models to compare with data at all — which is hilarious when coupled with the standard critique that macro isn't empirical).

Macro has no natural immunity to just-so stories because it doesn't have a robust internal criticism of them; it has to stick to debunking the caricature. This made me think that the "standardized critique" may well have adapted to macro like a virus adapts to a cell. When a macroeconomist sees the standard elements of the critique, the immediate response is to attack those: Macro does study X! Macro is empirical! The rest of economics is fine! Auction theory! These smack-downs increase the profile of the critique, and allow the critic's just-so story to invade the minds of many more readers.

There you have it: my critique of macro that avoids many of the pitfalls of the "standard critique". Macroeconomic theory simply isn't good enough to have any big unchallenged assumptions. They should all be challenged. Challenge the meta-narratives. Does monetary policy even matter? Is inflation always and everywhere a demographic phenomenon? Do people's decisions have any effect at all? Shut down just-so stories. Ask what else the model makes come out right, in addition. It's fine to use math and unrealistic assumptions to question these narratives, just make sure to use data.

*  *  *

Update 12 May 2018

It didn't fit in the narrative above, but one other criticism I had (that I talked about here in my post Lazy econ critique critiques) where some set of unrealistic assumptions or just-so story is used to explain macro/aggregated data but then the results are turned around to draw conclusions about the agents obeying those unrealistic assumptions. Unrealistic microfoundations are fine if they lead to empirically accurate theories, but you cannot then turn around and use those empirically accurate theories to draw conclusions based on those "microfoundations". A representative agent in a DSGE model may yield something reasonable for macro data (GDP, inflation), but you cannot turn around and say that individual rational behavior yields the result and policy impacting that behavior at the individual level would cause things to change. The representative agent (just as an example here of some kind of assumption) may give you a way to describe the macro data well, but it's an effective agent. You can't cross levels from an effective agent at the macro scale to actual agents at the micro scale and assume your unrealistic assumptions don't wildly impact the micro-level results.

Aside from the example given in the link about infectious disease, the neo-Fisher debate where Woodford applied bounded rationality/finite belief revision crosses scales from properties of micro agents to macro effects. I talked about it here (where I also found that the way macroeconomists treat limits is problematic). When the finite time to revise beliefs is taken as an actual property of actual humans that is proposed to have a the macro effect in the model, Woodford mistakes his effective agents for real ones.

Gender differences in unemployment

The St. Louis Fed has an article about the difference between the unemployment rate for women and for men, producing the data in this graph:

If we look at the data alone, it looks like this measure is positive until it drops to zero/negative  after the 1980s recessions. However, women's labor force participation was accelerating during this time (effectively adding more women looking for work) — which had other effects such as potentially creating the Philips curve. If we subtract an estimate of this effect (I admit I just eyeballed this fit using a dynamic equilibrium shock which is approximately Gaussian in shape), we essentially get a flat curve with dips for recessions:

It's imperfect (especially the 1950s, which may have a bit of e.g. post-war labor force re-entry), but this representation of the data helps mitigate an "optical illusion" — the sudden drop in the 80s now just looks like a recession dip super-imposed on the declining dynamic equilibrium shock.

...

PS The other differences noted in the article are likely due to the fact that the dynamic equilibrium is logarithmic — falling from a higher unemployment rate falls faster than falling from a lower unemployment rate. The figure about gender differences in time shows the difference between the era of women entering the workforce and  e.g. the 2008 recession where women's labor measures become correlated with men's (click for larger image):

Validating my CPI inflation forecast

I forgot that CPI data (all items) was going to come out today when I wrote my post from yesterday (inflation oscillations as "gravity waves" due to labor force changes), but I'm glad I didn't wait because the update is pretty much what the forecast said (and has been saying for the past year). The original forecast overshot the size of the post-Recession shock by a small amount (original forecast is the solid line, updated shock size is the dashed line), but it was well within the model error. Here are the continuously compounded and year over year CPI inflation forecasts as well as the CPI level forecast (where that shock over-estimate makes the most difference):

Labor force participation and gravity waves

The latest in a long line of arguments about how monetary policy controls the economy comes to us in the form of a report by William Gavin [pdf] published by the St. Louis Fed about how monetary policy and the Fed is responsible for a bunch of macroeconomic observables (of course). The report produces this graph of CPI:

I guess that's one way to look at it. However, the dynamic information equilibrium model of CPI, that I  have been tracking the forecasts of for some time, provides an alternative macroeconomic history. Let's look at the overall model:

The dynamic equilibrium is about 2.5% CPI inflation, with three major shocks — a post-war slowdown, the demographic shift of women entering the workforce (see Twitter thread here, or download pdf here), and the post-Great Recession drop in the labor force. This model is fairly accurate, averaging only a few percent error over the 70 years. But the difference between the model and the data is also interesting:

The periods of large error are associated with transitions of people into the labor force (as well as a strong Phillips curve per the link above). Aside from those periods, the error is only about ~1% (in CPI level, i.e. a CPI of 100.0 ± 1.0). From this, we can re-make the graph at the top of this post:

So was it the Fed's actions, or are the trends and fluctuations in CPI primarily due to the labor force? It's true that any view of CPI data is going to be model dependent (of course, the dynamic equilibrium model is empirically accurate and has done a good job of forecasting). But what I found to be fascinating is that those fluctuations in CPI highlighted above in green and red look just like chirps. In fact, the data looks remarkably like the gravitational wave chirps measured with LIGO:

These chirps are formed by inspiraling black holes or neutron stars — as their orbits around each other shrink (due to shedding energy via gravity waves), the period of the orbit gets shorter and shorter resulting in higher and higher frequencies until the orbiting pair shed all of their orbital energy and collapse into a black hole.

What if the mechanism of people entering the workforce creates these oscillations in CPI, which cause more people to fluctuate in and out of the workforce with the accompanying recessions and recoveries? It's an hypothesis, but the chirp wave forms (red and green above) fit the data pretty well.

Thinking like a physicist: MaxEnt edition

Nick Rowe is really great at putting together explanations of economic models in a very intuitive form, and I would have loved to have had him as a professor. But sometimes "thinking like an economist" and "thinking like a physicist" clash so badly it makes my eye twitch involuntarily. Nick put together a simple model where he wanted to show that it was possible that people cluster in cities even though they'd prefer not to. He considered a scenario where there was a possible city in the east and a possible city in the west. He considered it to be a two-point domain; I imagined it as two islands, and I guess as long as there were more than half the people on one it was "a city".

Nick then introduces utility and proceeds to show that you can get any result you want. This made me think: how would I approach the idea of the formation of cities in the simplest model possible?

Well, first off I would look at a continuous domain rather than a discrete point set (because the idea of cities is that people form areas of higher density, not just choose one place over another). In math, we write this as [0,1]. Think of it as modeling Canada along the Trans-Canada Highway from Montreal (1) to Vancouver (0) — corresponding to Nick's cities in the east and west, respectively.

If I assume I know nothing about humans (agents), my best guess is that they'd be uniformly distributed on [0,1]. That's the maximum entropy (least informative prior) distribution on a finite domain given no constraints. But what if we observe cities? That macro observation provides a constraint on our micro (agent) model. If x is the position between zero and one, then the macro observation results in a constraint on the expected values of ⟨x⟩ and and ⟨1−x⟩. The maximum entropy distribution on [0,1] with those constraints is called the beta distribution:

Depending on the parameters (i.e. macro observables), you can get any results you want just like in the economic version. The uniform distribution on [0,1] is also a special case of  the beta distribution for α = β = 1. Note that you could treat the problem like how Nick does where instead of x being position along the Trans-Canada Highway, it is the population of one of the two cities on the original discrete point set (with the other population being 1 − x). His possible solutions (most people live in one or the other city, or evenly split) would look like this:

Since these are maximum entropy distributions, we actually are assuming we know nothing else besides what we observe (i.e. the cities). We don't know if agents like this arrangement or hate it. But we do know that it is entirely possible that it can arise from random, inconsistent (as Gary Becker put it, "irrational") choices with only slightly different weights on the state space (i.e. where or which city to live in).

What else is maximum entropy (MaxEnt) reasoning good for? Part of the reason for writing this post was the connection between Nick Rowe's post and email discussion I was having. Is the income (or wealth) distribution what it is simply because income is (roughly) bounded at the bottom, but unbounded at the top? Well, that's part of it. The maximum entropy distribution on [0, ∞) [1] with an observed average ⟨x⟩ (i.e. average income, 1/λ) is an exponential distribution:

However, this distribution has fixed inequality with a Gini coefficient of 0.5. If income was exponentially distributed, then inequality wouldn't change. That's not what we see in the real world.

Now Vilfredo Pareto made some of the first observations about income distributions, and there's a maximum entropy distribution named after him that can have different Gini coefficients. It's not on [0, ∞), but rather [x₀, ∞) with a minimum value x₀:

The constraint in this case is on the expected value ⟨log x⟩ (i.e. the average log income), and the Gini coefficient is 1/(2 α - 1) with alpha of about 1.7 matching the US Gini data. But while this is a decent description of the ("fat") tail (i.e. the highest incomes), it doesn't do well with the lower income. Is there a maximum entropy distribution that does a bit better at all scales? This actually took a bit of effort to show, but the piece of knowledge that we'd want to add (the constraints) is really about how quickly agents acquire money: the time to observe α events that come at a rate β described by a Poisson process (which generates the above exponential distribution). This is called a gamma distribution, but it's not really what we're looking for. If the time to acquire a lot of money is extremely short, you have high income; if it is long, you have low income. So the expected time T to get M money yields an income of M/T (i.e. a salary in units of dollars per year or wage of dollars per hour). If T is gamma distributed, then 1/T is inverse gamma distributed. It looks like this:

We have our Pareto tail, but with a more sensible distribution at the lower end. We're also back to [0,∞) instead of having an explicit minimum value x₀.

Now I didn't just pull this distribution out of the air; it was computed from a model that included agents exchanging, "betting" on a market, as well as taxes that redistribute wealth by Bouchaud and Mezard almost 20 years ago. A more recent paper (also discussed in that email) by Berman et al changes up the taxes such that income is redistributed upward (think economic rent, or just plutocracy) and gets different results [2].

Anyway, I hope you can see the sort of things the maximum entropy approach are good for (and not good for: income distributions don't really fall out simply in the approach except for an effective Pareto distribution at the high end). And I hope it elucidates the way at least this physicist looks at problems e.g. assuming only things you know or observe, thinking in terms of scales.

...

Footnotes:

[1] Unlike economists, mathematicians know that infinity isn't an actual number and so it doesn't make sense to include it in a domain (set). The square brackets mean "included", so that [0,1] includes 1, but [0,1) doesn't. The latter domain has every number as close as possible to 1 as you can imagine, but just not 1.

[2] I do not like the way this paper is framed at all. You cannot claim to be surprised by non-ergodicity when you set up your model to remove its ergodicity. We added a piece to the model that keeps people in one part of the state space, and are now surprised that people stay in one part of the state space. Your assumptions are invalidating the ergodic hypothesis.

No really; they said they were surprised:

"Our findings invalidate the ergodic hypothesis. The fitted reallocation rate is not robustly positive for any dataset we analyze. Indeed, for one dataset we find it to be consistently negative for the last thirty years or so. We cannot overstate our surprise at this finding."

I'm guessing Ole Peters wrote that line as he has a knack for putting something in a model and not realizing he put it in the model. I'm also putting "ergodicity" and "non-ergodicity" on my list of red flags that someone is about to embark on some dubious economic theory.

JOLTS data

Update: Ha! Apparently economists see today as extraordinary, but I predicted this back on 10 January 2017.

There's now 1.0 unemployed worker per job vacancy, the lowest rate on record. #JOLTStwitter going nuts. (Series starts in 2001.) pic.twitter.com/q2tokzgIqe

— Justin Wolfers (@JustinWolfers) May 8, 2018

Update: Here is the most recent data (black) overlaid on the original forecast from here (which also shows the original code):

Original post:

I was busy running around doing work stuff today, so I won't belabor the fact that the latest JOLTS data doesn't really change things from the previous assessment.

Recessions and special snowflakes

The 1991, 2001, and 2008 recessions.

My issues with Dirk Bezemer's academic credibility seem to have nudged the Post Keynesian hornets' nest such that they seemed to assume I was attacking Post Keynesianism in general, which I gladly took on because one thing that I really do dislike is cult-like adherence to ideology regardless of what that ideology is. But first let me clarify a few things.

• My own work here would probably best qualify as Marxist econophysics if we're putting labels on things in the sense that Marx today would probably be a neoclassical economist but think of the results as sucky and doomed to end in revolution. I'm uncertain about the inevitability revolution, so in my own work I suggest a lot of neoclassical economics is often a fine description of reality empirically but the results are sucky and unavoidable.
• I really don't have any problems with many thinkers associated with Post-Keynesianism (Robinson, Minksy, Godley & Lavoie), but rather the fans that try to turn a bunch of disparate ideas into a "school" and make claims about these beatified economists they likely would not themselves. Various Post Keynesian supporters have told me that Post Keynesianism is well-defined and then proceeded to give me a novel set of commandments not listed by the previous person to do so. The definition I'm going by is Marc Lavoie's [1] because it is the only one that seems to be a stable kernel.
• Monetarism at least has hyperinflation as an empirical success; Post Keynesian empirical work is limited at best (and the right data might not even be available). As I suggested to Jo Michell on Twitter, it feels like a band that has logos and merch designs before they've played any gigs.

If you want to argue with these things, please do note that I am most convinced by theoretical curves passing through data points or their point-estimate equivalents.

Now that is out of the way, one of the other things I noticed in the trolling, I mean, discussion is that people have a lot of different stories about how different recessions happen. These stories are told with the kind of conviction that looks awkward in the context of several people telling different stories starring the same heroes and the same villains (or with the heroes and villains interchanged). The film Rashomon comes to mind. On the other hand Post Keynesian economics explains every recession, but every recession is a special snowflake with entirely different causes.

Now this isn't just a Post-Keynesian phenomenon, but is in fact extremely common from random men mansplaining the early 90s recession to economists giving what they think are unassailable descriptions of the Volcker Fed causing the 1980s recessions. The diversity of post hoc ergo propter hoc arguments identifying the cause of recessions is more likely a result of the fact that most of the time series take a turn for the bad in a recession weighted by politics [pdf].

This suggests an interesting hypothesis. Maybe the 90s recession was caused by the preceding saving and loan crisis, the 2000s recession by the dot-com bust, and the 2008 recession by the housing crisis ... but what if instead these recessions are caused by some other factor and the recession process simply undermines everything (with the news reporting on the largest collection of things that were undermined)? Like an avalanche taking everything with it, the oncoming recession undermines every source of growth if they are a bubble or not. Sure, this is just basic common sense in causal analysis: you observe X and Y seem to cause Z, but what if W causes X and Y? But I think it might be even more helpful in this case.

Now I'm not saying housing bubbles are a good thing as long as there's no recession, and it seems very likely the size of the US housing market boom contributed to the magnitude of the 2008 recession. Economic bubbles and even sustainable booms likely add snow to the eventual avalanche. But the idea that recessions pop bubbles (instead of popping bubbles causing recessions) helps us understand a few things:

• There is as yet no consensus in economic theory as to what a recession is, and even the disparate theories individually do not describe recessions with a great deal of empirical accuracy
• Several so-called housing bubbles (including many places in the US and in Norway) seem to have picked right back up and continued in the aftermath of the Great Recession
• Australia seems to be having a continuing housing boom/bubble, but no recession meant no crisis
• The theories of collapsing over-investment often are independent of what the over-investment is being invested in (e.g. over-investment in infrastructure, tulip bulbs, dot-com stocks, places to live, industry stocks). 

That last point is one of the more curious aspects. You would imagine over-investment in housing (employing millions of people and producing real assets) would be different than over-investment in random internet start-ups (employing only thousands of people and producing intellectual property). However, the two boom-bust cycles only differed by roughly a factor of 2 in scale. You could replace "housing" with "credit card debt", "corporate bonds", or "student loans" and the bubble analysis would be mostly unchanged. In fact, there are many stories ongoing today (the most common being a stock/asset bubble) that would likely be seen in the aftermath of a future recession as evidence that over-investment caused it [2]. The "dot com" or "[Dow Jones] industrial" adjectives in the market crashes are just adjectives -- not critical components of the analysis. Now this could be evidence that financial crises are universal processes, but another possible interpretation is that there's a universal process behind them — a recession cutting the booms off. An analogy:

Even though a drunk at a bar is being cut-off by last call, last call wasn't caused by how drunk he is. His drinking binge was ended by a separate process. 

I am not in any way saying this is conclusive evidence [3], but rather serves best as a palliative for confirmation bias and post hoc ergo propter hoc reasoning. Maybe each recession is a special, unique snowflake: the result of a process that starts when air with water in it reaches a certain temperature.

...

Footnotes:

[1] Marc Lavoie (pdf, H/T Jo Michell):

Essential Post-Keynesian Features (Lavoie 2006)

• The principle of effective demand

(demand-led economies)

– Both in the short and in the long run

• The importance and irreversibility of time

– Historical time

– Dynamics, the traverse

– Path dependence, multiple equilibria

– Tracking financial stocks

[2] And whoever wrote those analyses would be lauded as the ones that predicted the crisis.

[3] For one: what the @#&* is this underlying process? I am sympathetic to it being a more social process than economic one, but I'm really not convinced by anything at this point.

Robots versus shipping containers

There is an ongoing debate as to whether automation or trade is responsible for the decline in manufacturing jobs in the US (or as I put it, robots versus shipping containers); a new article in Quartz makes the case that it's the latter. Leaving aside the question of whether it is good or bad for a country to lose manufacturing jobs (and another country gain them), I decided to try and look at this with the dynamic information equilibrium model to answer the question: robots, or shipping containers?

I used the ratio of manufacturing jobs to all jobs (MANEMP over PAYEMS on FRED), and tried to see if a single shock can explain the data. The result is plausible (I ignored the WWII build-up and decline):

Click to enlarge. This model has one shock centered on 1989.6 with a width of about 40 years (1969 to 2009). However, the empirical accuracy can be improved by adding a second shock:

These shocks are centered on 1981.6 and 2004.1 with widths of 40 and 9 years respectively. We have overlapping eras from 1962 to 2002 and 1995 to 2014 that would could tentatively label the robot shock and the shipping container shock. It is actually surprising that we can resolve two shocks (they could plausibly merge together into a single shock like the model above). I am not sure of the source, but an article here says that GM introduced its first industrial robot prototypes in 1961. Here is an Atlantic article about industrial robots and their origins that sources the 1961 date to here. There's some more information here with an early appearance on the Tonight Show in 1966:

The shipping container shock appears in the data on the maximum size of shipping container vessels, which starts to take off after 1995:

It's not that the size of the vessels was driving shipping, but more likely the other way around. In the aftermath of the end of the cold war, globalization began as barriers to trade came down (and the free market ideology behind it flourished).

The relative size of the two shocks is about 3 to 1 (i.e. it was mostly robots), but the modern politics behind a move toward increasing trade barriers and tariffs is more likely due to the more recent shock. The robots had already taken over the US in the early 2000s. The shipping container shock came to manufacturing that was more difficult to automate. However, that shock is also mostly over in the US.

So the story behind manufacturing jobs in the US looks like it was first decimated by robots, but then finished off by shipping containers.

Down, down, down: the unemployment rate

The unemployment rate showed its first fall in a few months, to 3.9% from 4.1%. Overall, the forecasts (model described here) from the beginning of 2017 (!) are doing fine almost one and a half years out. At least my forecasts. The historical forecasts from the FOMC and FRBSF look more and more like they're just playing catch-up with reality (and the Minneapolis Fed remains biased high). Eyeballing it, the lifetime of accuracy from the FRBSF forecast is about four to six months.

Three sigma deviation in the 10-year rate

So I'm continuing to track the 10-year interest rate forecast from nearly 3 years ago. While the forecast did well before the 2016 election, today we're above a 3-sigma deviation from the estimated model error (the 99.9% percentile, or 1 in 1000). Of course with nearly 800 data points, we might expect to see at least *one* 3-sigma event. A similar deviation happened in the early 80s (Sep 1981 to Jun 1982) making this the second period of such a deviation.

This is extremely interesting because that time period represents exactly the time period where the Fed raised the discount rate to its maximum level, which (according to the standard narrative) kicked off the the second dip of the double-dip recession. However, like in the 80s, there appear to be signs of an upcoming recession in other data that might be a leading indicator.

I will admit it is speculative, but given the timing/timeline of the previous 3-sigma event it may become clear the US is in recession in the next 6 months (NBER won't officially declare it until a few quarters later).

Now you might wonder how raising interest rates to only about 2% could trigger a recession today in the same way raising interest rates to 14% did in the 80s. I admit I don't have a good answer to this except to say increasing labor force participation in the 80s probably provided a sufficient tailwind that Fed had to do do much more.

In any case, this makes for an excellent test of the model. Interest rates should come back down in the near term (about 6 months). A possible mechanism to bring them down is recession. The longer they stay at the 99.9% of their range or further, the more likely the model can be rejected.

Here's a zoomed in and non-log scale version of the graph at the top of the page (the green band was the forecast of the green line while the gray bands represent 50% and 90% confidence limits on the model error from the observed path):