Ext. Dagobah — Swamp
Jason: It seems monetary policy and inflation are completely uncorrelated. It seems reasonable to believe monetary policy doesn't actually affect inflation.
[Milton Friedman appears as a force ghost dressed as a Jedi from behind some foliage.]
Milton: I see you haven't heard my thermostat argument! Imagine a car ...
Jason: Actually, I have but ...
Milton: [Undaunted] ... driving on a hilly road trying to keep the same speed. If the driver was really good, the speed on the speedometer would be constant and you'd see the gas pedal go down and up in perfect correlation with the hills. But what you wouldn't see is speed changing — it'd be uncorrelated with the hills and the gas pedal.
Jason: Wait, is this a speedometer or a thermostat?
Milton: Quiet, you! I'm not finished ... Now if the driver wasn't very good, you might think you could tease out the relationship by looking at the gas pedal and the hills. But no! All you'd see in the gas pedal data when compared to the hills are the driver's random errors. No information about the relationship between the gas pedal and speed is available.
Jason: Ok, but how did we figure out looking at the gas pedal was important?
Milton: This isn't about whether we know about the gas pedal. We could be ignorant of the gas pedal — the point is that the model could exist!
Jason: So assume a complex model relationship between gas and speed when there appears to be no correlation?
Jason: Sounds kind of like the opposite of Occam's razor to me. I think I'll stick with Occam.
Milton: Wait, I mean no! Anyone can see monetary policy affects inflation.
Milton: Look at hyperinflation!
Jason: Ok, but can we extrapolate from 100% inflation down to 2% inflation? That's equivalent to extrapolating processes that happen on a time scale of a year to a time scale of 50 years ...
Milton: Gah! Physicists!
Jason: In fact, data seems to show a definite change in behavior around 10% inflation, which is remarkably close to the time scale between recessions ... [trails off, staring up at the sky]
Milton: Look, you. We have lots of evidence that monetary policy affects inflation.
Jason: Awesome! Why didn't you just show me that evidence instead of basically telling me that Occam's razor isn't always right? I mean, Occam's razor is a heuristic, not a theorem ... of course it's not always right. So are the models built using this evidence pretty good at forecasting, then?
Milton: Well, not exactly ...
Jason: Hmm. Can I see your evidence monetary policy affects inflation?
Milton: Here you go! All the evidence that monetary policy affects inflation!
Jason: Thanks, wow! Why didn't you just show me this in the first place?
Milton: I wanted to teach you about the thermostat!
Jason: But the reason we don't go with Occam's razor in this case is that we have all this evidence you just showed me ... it has nothing to do with thermostats or speedometers ... that's just question begging ... assuming we already have all this evidence ...
Milton: You're welcome!
[The force ghost suddenly vanishes.]
But here's what really happened ...
Jason: Hmm. Can I see your evidence monetary policy affects inflation?
Milton: You see, you won't be able to tease it out of the data. Imagine the Fed is a thermostat keeping a constant temperature ... the turning on and off of the heater is going to be completely uncorrelated with the temperature inside the house.
Jason: That's the same argument as the speedometer. Are you just trying to get out of showing me evidence because you don't have any?
Milton: You see, what I said is true ... from a certain point of view.
Jason: Certain point of view!??
[The force ghost suddenly vanishes.]
Great! I love your parable / story thingies. It's been a while since you've posted one of these, hasn't it?ReplyDelete
In my own opinion, this particular form can come across as really pretentious if not done with a significant humor component, so I tend to use it only in particular cases.Delete
… dialogues… that's the word I was searching for.ReplyDelete
Jason, I hope you and your family are safe and well in the current crisis.ReplyDelete
I haven’t been reading your blog much in the last year or so as we ran out of road for meaningful discussion a while ago. However, as I have been stuck at home recently, I read this post, and a few others. This post is a great illustration of the difference in our perspectives on the world, so I thought I would write a detailed comment.
When we try to understand the behaviour of natural system like gravity, we observe the behaviour of the system from a single perspective, in the manner of physicists, and try to infer how it works, starting with no prior knowledge. People like Newton and Einstein are rightly applauded for their genius.
However, when we try to understand the behaviour of any man-made system, we can observe the system from multiple perspectives. For example, a design perspective or a control perspective or a historical perspective (based on how the design & control systems have changed over time). We know a lot more about man-made systems than we do about natural systems because we design and operate man-made systems. We can also check the usefulness of any one perspective by comparing what we can see from that perspective to what we can see from other perspectives.
How does this relate to Milton Friedman’s thermostat?
Jason: “Is this a speedometer or a thermostat”?
The principle applies to the control of any man-made system from cars to heating systems to computer networks to businesses to economies. I’ll focus on cars as they are more complex than heating systems, from a control perspective, but simple and tangible enough so everyone can understand my arguments.
Jason: “How did we figure out looking at the gas pedal was important”?
We used a combination of our specialist knowledge of the design of cars and our general knowledge of the operation and control of cars. Friedman’s challenge is asking whether a mathematical model of a car’s normal behaviour can figure out what we already know from other perspectives. Friedman knows how drivers control cars. So does everyone else reading this comment.
Jason: “So assume a complex model relationship between gas and speed when there appears to be no correlation”?
You have misunderstood the point. No assumptions are required. See previous point. There IS a complex correlation between gas pedal and speed. We know that. The question is whether a mathematical model of the car’s normal behaviour can figure this out for itself. You are claiming that such models represent the ONE TRUE WAY of understanding how the macroeconomy works. Friedman is asking why we should believe you if equivalent models can’t even figure out how cars work?
Jason: “Sounds kind of like the opposite of Occam's razor to me. I think I'll stick with Occam”
You have misunderstood the point again. Occam’s razor is wrong here. That is a large part of the point that Friedman is making. I thought science was about bending over backwards to challenge your own assumptions. You are arguing that humanity doesn’t know how cars work even though we design, build, drive and repair them.
Jason: “Can I see your evidence monetary policy affects inflation”?ReplyDelete
Friedman’s challenge is not specific to either monetary policy or inflation.
There is a general principle in human systems thar I call the paradox of risk. The paradox is that if you identify a risk to the normal operation of a system BUT you also devise and implement a mechanism for managing that risk AND that mechanism is successful in mitigating the risk THEN the behaviour of the system will be identical to the situation where the risk did not exist. That is why Occam’s razor is inappropriate in such situations. The only way to prove that the risk really exists is for the adverse impact of the risk to happen e.g. for the system to fail.
Here is an example. Assume man-made climate change is real. Assume it will have major adverse effects if we do nothing. Assume that someone invents a clever but very expensive technology to mitigate the risks. Assume that the resultant political wars over funding can be won. Assume that the technology is implemented. Assume that the technology successfully mitigates the risks of climate change.
The result of this process would be that there were no major adverse impacts from climate change. People who don’t believe in climate change would then say that the whole thing was a hoax and that we had wasted billions for no good purpose. And they would justify their view using high level stats of climate data and Occam’s razor – similar to your arguments here.
The same thing happened with the Year 2000 date problem in IT. Something similar seems to be happening with inoculations for diseases like measles.
Occam’s razor is bullshit in such cases.
Regarding the specifics of monetary policy, similar points pertained during Friedman’s lifetime i.e. you can’t disprove the effectiveness of a risk mitigation strategy until it fails or until it stops being applied. The whole edifice of monetary policy fell apart in 2008 precisely because the risks of economic instability were not mitigated by monetary policy. Friedman was wrong in his view on monetary policy. However, he was dead by 2008, so it is unfair on him to make a case against his policy arguments based on evidence that came to light after his death. However, that doesn’t invalidate the very smart argument he made about controlling man-made systems using the thermostat analogy.
The same argument pertains with fiscal policy. In 2008, Paul Krugman argued that we needed huge fiscal stimulus – much more than was eventually implemented. Assume he was correct. If such huge stimulus had been implemented, and that stimulus had mitigated the adverse consequences of 2008 further, many people would have argued that Krugman was wrong in his policy prescription because the effects of 2008 were very mild, so the huge stimulus was pointless. The only way to prove that the huge stimulus was required would have been to abstain from any stimulus and let the economy tank. However, that would have been completely irresponsible.
Jason: “So are the models built using this evidence pretty good at forecasting, then?”ReplyDelete
You constantly overrate the importance of your type of prediction. Let’s pursue the car analogy further.
When a car breaks down, a mathematical model of the car’s normal behaviour would identify only an unspecified shock. It has no useful value for fixing the problem. This is analogous to most economic models. It would be a humble mechanic who would identify that the battery was flat and needed to be replaced. Mathematicians whose elegant behavioural models thought of the car as a single point in space would be astonished as their models didn’t acknowledge the existence of a battery or even an engine. This is analogous to economists (and you) asking why anyone would want to know how individual components of the economy work when your mathematical models have no role for such components
When we manage human systems, we do not do so by predicting the timing of breakdowns or crashes. We manage them based on risk assessments. For example, we get our cars serviced to reduce the risk of a later breakdown. If you used a mathematical model to analyse the car’s behaviour before and after the service, you would see no difference. Using this model and the type of thinking you are employing here you would then conclude that the service had no effect. However, all that would prove is that you hadn’t understood why people get their cars serviced.
Your models would see any anomaly in the car’s behaviour as an unexplained shock to which you would attach a generic name (like a recession in economics). You would then ask what is the cause of this phenomenon? However, there are many possible causes, and combinations of causes, of car breakdowns and crashes. Your model would have no insight into any of them.
We manage the risks of breakdowns and crashes in cars in many ways/ We take out insurance policies; we wear seat belts; we set and police speed limits; we drive as little as possible when the weather is bad; we make people take tests before they are allowed to drive on busy roads; we replace used components before they wear out; we monitor the components on a real-time basis to detect faults as quickly as possible etc. We don’t write mathematical models to predict the timing of the next car crash or breakdown.
We also manage rare adverse events in many natural systems based primarily on risk assessments. For example, earthquakes, volcanic eruptions, avalanches, medical epidemics. We have talked before about analogies between some of these and economic recessions. We don’t try to predict the timing of such rare events or take credit for predictions of normal behaviour in the absence of these rare events. If you disagree, what is your prediction for the adverse economic impact of the coronavirus, and what is the basis for that prediction given that we don’t even know the full medical impact or the lengths of any lockdowns?
Based on what you have written in this post, you haven’t yet understood any of this. As I have said before, any useful form of economics for policymaking would have to be more like medicine (or car maintenance) than physics. Medicine focuses primarily on risk management based on a detailed understanding of how the various components of the body work and how they interact with each other. However, even then, best practice policymaking is called evidence-based policymaking rather than scientific policymaking as policy ALWAYS involves value judgements and the subjective weighting of facts. It is not science.
JM Keynes: “If economists could manage to get themselves thought of as humble, competent people on a level with dentists, that would be splendid”
Note dentists – not physicists. Problem solvers – not mathematicians.
And yes, my family is doing fine — we're on lockdown here in Seattle and have been for weeks (initially voluntarily, and then as ordered by Inslee). I hope you and your family are doing well, too!
I'll take your points on one at a time — the first is your extension of the car analogy.
First, it's fine from a simple system illustration, but it lacks quite a bit of nuance which makes makes it not fit for purpose.
1. A car was designed. No one designed the economy.
2. A car operates on forces (mechanical, electrical, etc) we understand to various degrees. The economy does not.
In fact, if we were to think of the economy using an analogy of something we humans built, the best one I can think of is deep learning algorithms. We have no idea how they work, but we can kind of get them to do things that we want (but often there are unintended consequences).
I think there is actually a deeper connection to markets than just an analogy here — wrote about it here:
"When a car breaks down, a mathematical model of the car’s normal behaviour would identify only an unspecified shock."
is not true for cars (there are actually lots of indicators and all kinds of codes that specify exactly what's broken in a modern car). I work on satellite systems and this kind of telemetry is very important because you can't actually bring a satellite back to earth to fix it — you fix it on orbit.
It is true for deep learning and in economics. The news tends to point to this or that event as the "cause" of a recession, but sometimes that's just pareidolia. Even the current economic crisis that seems brought on by COVID-19 may have been brewing for months and the outbreak was just the sound of the snapping twig that triggered the avalanche.
"You constantly overrate the importance of your type of prediction."
It's not that I think forecasting time series is the only & best test of a model, it's that current macro models like DSGE underperform AR processes — the linear extrapolation of noisy time series. And when there exist models (like mine!) that outperform those AR processes, it's a reason to consider them.
I am fully aware of other ways to validate models ("retrodiction" and describing past data with fewer parameters), but I will not back down from theory or more generally ideas needing to be grounded in empirical reality. Fantasy claims about how people reckon the economy works that are at odds with data are a definite pet peeve.
I'll get to more later!
Jason: “I'll take your points on one at a time — the first is your extension of the car analogy. First, it's fine from a simple system illustration, but it lacks quite a bit of nuance which makes it not fit for purpose”ReplyDelete
The post you wrote was about Friedman, cars and heating systems. You can’t decide that your post is fit for purpose (particularly when you have missed the point of what Friedman was saying even about simple man-made systems) but then decide that my response is not fit for purpose for continuing with the same analogy.
Jason: ”A car was designed. No one designed the economy”
That’s not true. All human systems are designed. For example, the US constitution was designed and many of the behaviours we observe flow directly from the components and rules in that design. The UK follows a different design, so we observe different behaviours. Democratic systems are also designed. We observe different behaviours with different designs there too. No-one says that we don’t understand the US constitution or democracy unless we can predict the results of elections.
All human systems are made up of components and rules and their interactions. Components are made up of sub-components and their interactions. All economic policymaking is about making changes to the design and rules of the economic system by influencing the design and behaviour of components and component groups e.g. government departments, banks, businesses, households.
I have never understood why you don’t understand this. It feels like you must be doing it deliberately. Particle physicists have spent the last hundred years or so breaking natural phenomena into ever-smaller components and sub-components on the assumption that an understanding of these components will provide us with a greater understanding of the natural world. Why do they do that if system components have no relevance?
Jason: “A car operates on forces (mechanical, electrical, etc) we understand to various degrees. The economy does not”
You are confusing two separate concepts: operation and behaviour. A car operates the way it was designed to operate based on its components and their interaction, which is why a mechanic knows that he can fix the car by replacing a flat battery. The battery can be considered as a black box by the mechanic. He doesn’t need to understand the chemical processes involved in making the battery work. He just needs to understand how the battery interacts with other components.
The car’s overall behaviour, such as its acceleration, is based on forces. However, the car behaves differently under different scenarios. Generic models that focus on normal behaviour have little useful to say about rare/catastrophic scenarios.
If you don’t understand the difference between system operation and behaviour, or the fact that different behavioural models are required for different scenarios, you haven’t understood what I am saying.
Jason: “I work on satellite systems and this kind of telemetry is very important”
Yes, timely measurement is vital to the effective control of any system. That’s true of cars too – I mentioned that in my earlier reply. Many of the problems in economics arise from infrequent and inaccurate measurement. No-one in their right mind would imagine that it is a good idea to manage the economy based on very high summary GDP figures estimated (badly) only once every three months. If macroeconomics were ever to develop into anything remotely scientific or useful, it would have to start by designing more accurate, more timely and more granular measurement systems much more akin to how we manage satellites and cars ... and businesses and banks and pretty much everything else.
I note that one of the biggest problems in the current coronavirus epidemic is our inability to test everyone for the virus. This has major implications for individuals i.e. do I have the virus or not, and for society i.e. how quickly is it spreading/evolving and how do we control it.
I imagine that someone will design an app to allow us to self-test and share the results centrally before the next such epidemic. In the meantime, any predictions are hampered by the absence of such measurements. That is not anyone’s fault, but it limits our ability to manage the epidemic effectively at both individual and societal levels.ReplyDelete
Note also that many of the prediction models involved in managing the epidemic use untested “best guess” assumptions as their basis, so they are only ever as good as those assumptions. That is the nature of making decisions under uncertainty irrespective of who is making the decisions or the subject of the decisions. You would say that using such assumptions is not “science” but in the real world we should never let the perfect be the enemy of the good.
You will get no argument from me on the need for timely and accurate system measurement. The issue I am raising is not about measurement. It is about massive system failures, whether their timing can be predicted and how we should manage the related risks.
Jason: “It's not that I think forecasting time series is the only & best test of a model, it's that current macro models like DSGE underperform AR processes — the linear extrapolation of noisy time series. And when there exist models (like mine!) that outperform those AR processes, it's a reason to consider them”
Yes, I agree. You have done a fantastic job of showing that you can make better predictions than academic economists with simpler models. You should be applauded for that.
However, that is not relevant to what I am saying. The current crisis is likely to show that you (and other modellers) failed to predict upcoming changes in GDP and unemployment by a VERY long way. One of the points I am making is that this failure is inevitable as such crises are unpredictable. The epidemic probably started when a single animal infected a single human in a Chinese market. History tells us that such things can occur, with very approximate frequencies, and that they can cause havoc in human societies. Similarly, market crashes, runs on currencies etc. That is why JM Keynes said:
“Economists set themselves too easy, too useless a task if in tempestuous seasons they can only tell us that when the storm is long past the ocean is flat again”
Keynes might as well have been talking directly to you. His argument is about whether your type of prediction has any value it only ever works when the ocean is flat. I recall that the error bands in your retrospective predictions for the 2008 crisis were so wide you might as well have said “I have no idea what will happen next”. The issues under discussion are probability, risk and uncertainty, and their importance for taking policy decisions. You are not even having the correct discussion.
Jason: “I will not back down from theory or more generally ideas needing to be grounded in empirical reality. Fantasy claims about how people reckon the economy works that are at odds with data are a definite pet peeve”
Yes, I agree. That is not the point I am making. You need to take that up with mainstream academic economists. Mainstream economists do not make accurate observations even at the level of a single market. Instead, they base their thinking on absurd assumptions - like perfect markets – which could not possibly exist. And they don’t understand basic accounting – which is like a mathematician failing to understand basic arithmetic.
Jason: “Even the current economic crisis that seems brought on by COVID-19 may have been brewing for months and the outbreak was just the sound of the snapping twig that triggered the avalanche”
No. The crisis was brought about by the virus and our management of the consequences of the virus. The crisis is global and applies to countries who economies were previously in very different states. The crisis is more akin to a tsunami overwhelming a prediction of the normal oceanic tides. Your recent predictions are irrelevant as they did not see the tsunami coming.
Jason: “A car operates on forces (mechanical, electrical, etc) we understand to various degrees. The economy does not”ReplyDelete
I meant to say earlier that trade between any two components of the economy is a similar concept to a measure of the interactions (forces) between two bodies. If you start with the global GDP-related economy, the number of sales and purchases will be the same i.e. income = expenditure. When you divide the global economy in any way, you are left with trade imbalances. For example, if you split the global economy into the US versus the rest of the world, the imbalances are US imports and exports.
You can keep going with further sub-divisions. You can split the US component into internal supply and internal demand. Internal supply is the business sector.
When Keynes and his contemporaries created the subject of macroeconomics, they did so to study the interactions between the supply (business and imports) and demand (non-business and exports) components, and to what extent those interactions could be influenced by government activity, so they further separated the internal demand component into households and governments.
It’s the study of these components and their interactions that give rise to the accounting identities quoted in economic textbooks.
Of course, you can separate the economy into different sets of components than that original set. There is nothing unique about that subdivision. From a UK perspective, we could sub-divide the rest of the world into the EU and the residual rest of the world. We could sub-divide households into rich households and poor households.
The most appropriate sub-divisions will depend on the questions you ask.
One of the long cons of mainstream economics is the view that the macroeconomy should be conceived ONLY as the interactions of millions of undifferentiated agents. It is difficult to see how you would show any impact of government activity if you don’t isolate the government into a separate component from the rest of the economy. In my view, that is the ultimate victory of Milton Friedman and other libertarians/right-wingers over left-leaning economists who have been duped into a conceptual model which makes justifying government activity difficult to achieve.
I said that your models were libertarian a couple of years ago. You thought I meant that your politics were libertarian. I meant that your conception of the economy is libertarian as your models don't conceive of a government. You have stated in the past that you can see little evidence that government policy makes much difference to the economy. However, the government makes up 30-50% of GDP-related spending in most countries so that is not a credible view IMHO.
"I meant that your conception of the economy is libertarian as your models don't conceive of a government."Delete
This is an incorrect interpretation.
The models do often say that typical government policies don't have the effects that people think (e.g. moderate changes in UI do not cause the unemployment rate to fall faster or slower — it falls at the same rate regardless of policies in the post-war era). But I think that's just a function of economic theory being wrong than any details of information equilibrium.
But the big takeaway from the empirical success of the models is that equilibrium economics is *boring* and that most of the action (non-equilibrium economics) is in sociology (government policies, institutions, sexism, racism).
The big takeaway is that you're not going to go from human behavior to an equilibrium price — you're more likely go from studying human behavior to understanding fear or racism to put in place policies that mitigate or dissuade them that will reduce the impact of recessions or make housing more affordable (I write much more about this in my book). The equilibrium price bit is boring and is pretty basic in terms of math.
Wrote about this years ago:
(Which you appear to have read, and even commented on.)