Monday, August 25, 2014

You don't need to understand how people make choices ...

... to get 80% of the way there.

Given what I've been doing with this blog, this is not the way to go for the future of macroeconomics:
The classical economic of choice is therefore far too simple as it does not capture what goes on in people’s brain when they make choices. “It is also much too static to capture the sensitivity and dynamics of the process,” [Daniel McFadden] said.

Maybe microeconomics might benefit from the study of human behavior, but macro seems to follow optimal information transfer. Optimal doesn't necessarily mean perfect, however. There is a microeconomic behavioral experiment at the beginning of the linked piece that shows a lot of information doesn't get through the market mechanism:
[McFadden] highlighted an experiment he carried out some time ago at his university where half of the students were given a chit saying they were entitled to a pencil and half did not. The two groups could trade as buyers and sellers.
While traditional economic theory said the market should clear with half the pencils sold at close to a median value. In fact less than a fifth were traded. “One answer is that people have agoraphobia – they don’t like markets and that influences resource allocation,” he said.
This brings up an interesting point about information transfer I've mentioned before (see the last paragraph). I've said ideal information transfer is the condition that the information transmitted by the demand is equal to the information received by the supply, I(S) = I(D). In real life, human rationality and behavior factors might put a limit on this so that I(S) = α I(D) for some α < 1. The thing is, α is completely unknown (at least right now). Maybe, according to the experiment mentioned, α = 2/5. This may seem like a problem for the theory, but in fact has no particular effect in any of the calculations and is essentially captured by the fitted value of "kappa" (the information transfer index). Another way to say this is that ideal information transfer might only refer to the ideal practically realizable information transfer.

To bring in an analogy with thermodynamics: there is a maximum efficiency of a Carnot cycle but this never reaches 100%. When we say we have ideal information transfer we are saying something analogous to saying we have a maximum efficiency Carnot cycle, taking that maximum efficiency to be an unknown parameter (it is fit to empirical data).

The most nihilistic way to put it is like this: if we only ever see (through a market mechanism) 2/5 of the total information available (at peak efficiency), what does it matter if that other information exists? The remaining 3/5 of the information is like an event outside of one's light cone. Saying I(S) = I(D) where I(D) is the accessible information is not mathematically different from saying I(S) = α I(D) for some α < 1.

1. I am not an economist or a physicist, and I’m not sure how I found your blog, but I’m interested in different ways of thinking about economics, so I’ve been reading some of your posts.

My professional background involved helping businesses and government organisations solve operational problems. I think of a business as a system which communicates with other systems (customers, suppliers, regulators, competitors) via flows of materials, energy, money and other types of information. At a macro level, this type of thinking perceives of the economy as a gigantic network and it is natural to think of information flowing through the network in “waves” which propagate trends. The waves can be specific e.g. buying the latest electronic gadget or sharing information on a news event, or generic e.g. optimism, pessimism, panic.

This type of thinking doesn’t appear to fit well with mainstream economics. It focuses on “what” happens; “when” it happens; “where” it happens; “how much” it happens; “who” is involved etc. Mainstream economics, on the other hand, appears to focus mainly on “why” something happens e.g. how do people make decisions. It feels to me that you are saying something similar from a very different starting point. However, I don’t understand the theory behind your views so I’d like to ask some questions to help me understand a little more.

First, you talk about I(S) and I(D) and say things like I(S) = α I(D) for some α < 1. This is very abstract. Can you give a concrete example of what you have in mind when you say this e.g. why is I(S) < I(D)? I can see that information can be lost but it is lost in all directions whereas your equation suggests that it is lost in one direction only. Can you explain please?

Second, I’d like to explore what you mean by information, and what you think you can and cannot forecast, using some example scenarios.

A) Imagine an agricultural economy where everyone consumes multiple foodstuffs. There are many different ways in which this could be organised. Three examples. 1) Everyone is entirely self-sufficient and produces every foodstuff. 2) Everyone specialises in foodstuff production and then trades using barter to achieve the same consumption outcome as in case 1. 3) Everyone specialises in foodstuff production and then trades using some form of money to achieve the same consumption outcome as in case 1. These three economies produce the same output. However, the trade involved, and what it is measured, are different in each case. How would your technique cope with these different scenarios? Does this type of difference matter to your technique?

B) Imagine the economy 100 years from now. Maybe all transactions are carried out using electronic money. At least theoretically, all of these transactions could be collated to create a gigantic database that would allow economists to slice and dice economic data in ways which are currently unimaginable. This is the equivalent of what individual businesses and households do today, at a micro level, to solve problems. It feels to me that this would be much more useful in uncovering the causes, or combinations of causes, of observed macro effects than any measurement/modelling technique using the highly aggregated data we have available today. What is your view?

C) Imagine the economy as it arrived at a major turning point such as the industrial revolution. Prior to this point there was one trend (little or no growth) which had been in place for centuries. After that point there was a completely different trend. I can’t see how any technique could forecast this. It is effectively a random event which changes the behaviour of the system. There are many smaller events such as this all the time e.g. the discovery of oil in a previously poor country in the Middle East. Would your forecasting technique be able to cope with this?

(cont’d below)

1. Thank you for your questions -- I will attempt to tackle them in chunks. I'll start with the paragraph that begins "First, you talk about I(S) and I(D) ... "

What I mean by I(D) (information coming from the demand) are all of the needs, theories and pieces of information that people come to the market with. Some examples are:

"Kevin thinks $5 is too high" "Kyoko thinks the price will rise to 1800 ¥ given the current trend of exchange rates before falling to 1200 ¥, so buys one today" "Steve buys$40 worth"

These vary in specificity and humans aren't necessarily writing these down and going out into the market with a list, but these are all instances of demand information.

Here's I(S) -- the information as received on the supply side of this product:

"A person doesn't buy the product"

This is the information that the supply receives and you can see that it is much less specific than the demand information, so in general I(S) < I(D). Complicated contracts can be drawn up that really specify demand over periods of years and things like derivatives can give some additional nuance beyond buying or selling ... these can help I(S) ≈ I(D). For example, a contract would probably help Kyoko buy more in the example above.

Now the supply side doesn't necessarily have to believe these demand theories, they just have to capture the information in them. In fact, many times the supply side has to think the opposite (a person sells a stock if he or she thinks it will go down, and buys one if he or she thinks it will go up -- these two sides have to find each other).

Since the model itself is set up as a communication channel (from information theory) going from demand to supply, the demand information sets a limit for how much information is received (you can't receive more information than is sent). So I(S) < I(D) ... however in looking at data, assuming I(S) ≈ I(D) seems to work pretty well. This is the information theory equivalent of the assumption of a "complete market". But! This set up allows I(S) < I(D) so doesn't preclude e.g. market failures.

The asymmetry (demand transmits information to the supply) has another function. In the first couple posts on this blog I derive supply and demand diagrams, and because demand is the source of information, the demand curve generally slopes downward. Since the supply receives the information, the supply curve slopes upward.

And one last function of the asymmetry: if I(S) < I(D) the market price is less than you'd get in an ideal market where I(S) = I(D). This makes intuitive sense if you think about it as if markets were perfect, more products would be cost effective to produce (you'd get a higher price for everything), but in imperfect markets people are hesitant to shell out money for goods so prices are generally lower (this is the result of the experiment described in the post above).

2. Regarding Scenario A, case 1 doesn't have a market, so output would depend mostly on external factors like the weather.

I address case 2 in a post from a few months back [1] (although there are only two goods: apples and bananas). What results is a system where there are multiple prices (there are N^2 - N of them where N products) -- the price of bananas in apples, the price of strawberries in apples, the price of bananas in strawberries, etc. Each of these prices are independently inefficient, so the overall market efficiency is less than if one good served as money (everyone uses the "apple price") -- or there was fiat currency. In the real world, case 2 would have to have a higher output than case 1 to feed everyone unless information transfer was ideal.

Case 3 is more efficient than case 2 since there are only N prices that are inefficient (the price of every good in terms of money), but then adds some complications that make macroeconomics interesting. I address this in [2] below.

[1] http://informationtransfereconomics.blogspot.com/2014/03/apples-bananas-and-information-transfer.html

[2] http://informationtransfereconomics.blogspot.com/2014/03/how-money-transfers-information.html

3. My response to B is sure it's possible, but it seems like a waste of time. I have a post on why I think that is [3]. Solving all of the individual dynamic equations for every molecule (micro) in an ideal gas will definitely lead you to the macro ideal gas law PV = nRT -- and even tell you what R is. But so would thermodynamics, which is much easier. Nature has found a way to reduce the information content of 6N-dimensional space (that's way over 10^23 dimensions) of individual atoms down to 3 dimensions (P, V and T).

In economics, the behavior of N economic agents (where N is hundreds of millions) seems to be captured in a few macro dimensions (NGDP, price level, unemployment rate, interest rates ...). If the indivdual micro behavior of the N economic agents matters at the macro level, then there must be hundreds of millions of macro variables that matter (the dimensions of the micro and macro spaces must be equal if there is no dimensional reduction). NGDP, price level and unemployment rate are just three of them. I have no idea what the remaining hundred million variables are -- and no one seems to even be suggesting they exist. Sure, people add in debt to GDP ratios, consumer debt or desired precautionary savings rates, but there are only hundreds of these indicators, not hundreds of millions -- so significant dimensional reduction must be occurring.

The loss of the information going from ~10^23 dimensions to a few dimensions is measured by entropy -- and that is analogous to the approach I'm taking here. I am essentially saying much of that micro information about economic agents and what they individually think is lost.

4. Regarding C, I'd agree those kind of events are probably totally unpredictable -- but industrial revolutions are rare events. So are monetary policy regime changes [4]. So are wars [5]. So, even, are recessions [6]. Most of the time an economy is just lumbering along. Mostly, the forecasts I make, I make ceteris paribus -- nothing major intervenes like a recession or a war or a major change in monetary policy. I have made some pretense to forecasting a recession recently [6], or at least, it's likelihood. I'm not sure that even if NGDP was above the trend line it is possible to know when the recession would strike (it's hard to predict exactly when a piece of metal will break even if you know the stress on it).

The information theory picture probably wouldn't work well with undiversified economies (the Middle Eastern oil example) or small economies. Also, the information theory picture concentrates on the trend, not the fluctuations [7].

PS One might be able to trace the industrial revolution to the creation of stable paper money by the Dutch (then imported into England at the founding of the Bank of England in 1699). This was not the first use of paper money, but is probably the beginning of the idea of a national monetary policy. In this sense, maybe we are seeing the efficiency of having only N prices vs N^2-N prices I mention in Scenario A above. This kind of history is really out of my league, though.

[4] http://informationtransfereconomics.blogspot.com/2014/08/monetary-regime-change.html
[5] http://informationtransfereconomics.blogspot.com/2013/09/exit-through-hyperinflation.html
[6] http://informationtransfereconomics.blogspot.com/2014/08/can-information-theory-predict.html
[7] http://informationtransfereconomics.blogspot.com/2014/01/what-is-and-isnt-explained-by.html

2. D) Some economies fail catastrophically e.g. Soviet communism, economies which face hyperinflation and revolutions. Would you forecasting technique be able to cope with this?

Note that all of my scenarios involve human input so I am sceptical of your suggestion that people can be removed from thinking about macroeconomics.

Third, in my systems-based thinking, it is not just information which flows but also money (a type of information, I guess), materials and energy. Materials and energy are subject to the laws of physics e.g. conservation. However, money and information are not. Money is conserved if, say, I buy a bicycle from you, but not if I borrow money from a bank. Information can be replicated/shared via education or the Internet. It can also be lost e.g. ancient civilisations may have had knowledge which was lost over time. Does this matter in your technique?

Fourth, have you investigated any non-mainstream brands of economics? From a systems-based perspective, I find Post-Keynesian economics much more approachable than mainstream economics (including mainstream Keynesian economics) and much more compatible with my own mental models.

Sorry for the number of questions. I’ll understand if you don’t have time to answer them all.

1. Regarding D, first I've made some attempts to look into hyperinflation:

http://informationtransfereconomics.blogspot.com/2014/01/rich-countries-poor-countries-japan-and.html

But like in C, human political events are outside the scope of the predictions of the models.

Setting up a market is kind of like setting up the internet. We know how internet traffic will flow for the most part based on queueing theory and information theory. Those theories tell you how the system works. They will not predict that Egypt will shut of Twitter during a revolution. The equations will not tell you that a big peak in video streaming traffic will happen in the evenings in each local time zone in the US (but they will tell you how the system will react).

The information theory picture of an economy seems to be able to tell you how a market functions, but doesn't tell you that the stock market closes on weekends, high school kids get summer jobs or the stock market will crash.

The weirdly simplifying piece that is different in the economics picture vs the internet analogy is that I am assuming as a starting point the internet is operating at full capacity -- there are no fluctuations in the evening and no one shuts down Twitter. What's weird about it is this simplifying assumption gets things right. It gets interest rates right on average and predicts inflation over 20 years -- assuming nothing on the scale of WWII happens (although it works for the interest rates right through WWII).

I'm not saying humans are always irrelevant -- they are just empirically irrelevant most of the time when there's no recession, war or monetary policy regime change. And those things don't happen very often.

I've been somewhat hyperbolic about removing humans from economics. But I have sort of a joke that gets at my point:

The human behavior independent prediction from classical economics of the "employment rate" (one minus the unenployment rate) is 100%. The actual result in the US since 1948 has been between 75% and 97% with an average of 94.2% -- that's less than 10% off from the human behavior independent result of 100%.

The neoclassical picture (as well as the information transfer picture) has a model with a natural rate ~95% -- and is only a couple of percent off on average.

All the complicated human behavior in some employment search model with 50 parameters is going after the difference between 95% and 94.2%.

Snark aside, until we know what the baseline is, we have no idea what we are modeling human behavior as a deviation from.

2. Regarding the point starting with "Third ...":

People often use the colloquial definition of "information" when they see the words "information theory". Information theory isn't about stuff you learn in school or even the content of a book. It is related to the replication and sharing, though. Information theory helps you figure out what kind of system you need to store data on a DVD or send a signal to Voyager 2 (error correction, required signal to noise ratios, bandwidth). The difference between common usage of the word information and information theory becomes stark when you see that a random string like "Ag@df#A6n" has more information in it than the string "the house" (the "the" is largely redundant, the space irrelevant, and you can probably guess what work "ho_se" is -- either horse or house; you can't guess a missing character in the first string, though).

In the information transfer picture, money most importantly defines the unit of information that is processed by the market. It turns out empirically, the best definition of money is "currency in circulation" ("M0"), which is a primary component of the monetary base (MB --base reserves have an impact on short term interest rates). When you get a loan from the bank, the transaction creates something measured by the monetary aggregate M1. These other aggregates (including e.g. M2, MZM) seem to be set more in response to aggregate demand. M0 appears to be able to unleash demand, but also create inflation.

I have more on information here:

http://informationtransfereconomics.blogspot.com/2014/06/money-unit-of-information-and-medium-of.html

3. Regarding your comment beginning "Fourth ..."

I am pretty much ignorant of what Post-Keynesian economics is. The descriptions out there range from going back to Keynes' original writing to what is called "modern monetary theory".

I have really only made some connections between the information transfer picture with the history of economics and mainstream topics, e.g. supply and demand, the quantity theory of money, the ISLM model, matching theory.

3. Thanks a lot for these extensive replies. I have scanned through them but will read them in more depth before replying further.

4. I noticed Noah Smith had a Bloomberg post up. He's comparing literature, science and economics:

"The other way that the economics culture differs from science or literature is in its purpose. Literature describes human behavior, while natural science ignores it. But economists want to understand and control human behavior, and that means the object of their study is as smart and free-willed as the economists themselves. Predicting the actions of humans is a lot harder than predicting the actions of particles, and requires you to ask different questions, such as “What would I do in this situation, if I were being smart?”"

http://www.bloombergview.com/articles/2014-08-27/the-economics-food-fight

1. It's funny because someone referenced that same CP Snow lecture a few days ago (I'm not sure where, Crooked Timber, maybe?).

I'm guessing Noah said "natural science" instead of "science" because psychology and anthropology study human behavior.

I still remain unconvinced that any behavior-based theory has ever described an economy even in general terms. If you put a bunch of individuals together and get some predictable behaviors out of it (e.g. Noah's example of an equilibrium arising in a trading game in class in one of his other posts), it makes me think human intelligence and free will has little to do with the outcome -- one thing it doesn't tell me is that it's important.

5. Thanks again for your replies to my previous comments. There are a couple of areas which I‘d like to explore further. Here is the first.

“This is the information that the supply receives and you can see that it is much less specific than the demand information, so in general I(S) < I(D)”.

I’m not following how are getting to this conclusion. Here is a diagram of a generic economic transaction which you should be able to access.

Potential customers and suppliers pick up information from the market. They then develop strategies for approaching an economic transaction and engage in a selection, negotiation and decision process which may or may not result in a successful transaction. Information on successful transactions is then added to the information available via the market.

Here are a few examples.

In any government procurement, there is one customer and many competing suppliers. The customer will probably set an upper price limit but the actual price is based mostly on competition between the suppliers. Each of the suppliers has less than full information on the customer’s strategy or the other suppliers’ strategies. The dynamic between the suppliers is at least as important as the dynamic between customer and suppliers.

In an auction, the roles are reversed. There is one supplier and many competing customers. The supplier will probably set a lower price limit (reserve price) but the actual price is based mostly on competition between the customers. Each of the customers has less than full information on the supplier’s strategy or the other customers’ strategies. The dynamic between the customers is more important as the dynamic between customers and supplier.

In the stock market, there are many customers (buyers) and many suppliers (sellers). Each customer and supplier enters the market based on their own insight into the market rather than based on the strategies of other customers or suppliers. The customer and supplier roles can be reversed from one transaction to the next. The most significant information issue is probably that large financial institutions have much more information than an ordinary investor both in terms of their ability to analyse the market and the speed at which they capture new information. This is true irrespective of whether the financial institutions are the customers or the suppliers in any transaction.

In day to day life there are similar imbalances, often in favour of the supplier. For example, a typical person engages in certain types of transaction infrequently (e.g. second-hand cars, major household improvements) so has little information on the market or what is a “good” price. Meanwhile the suppliers work in their specialist markets every day so have much more information at their disposal with which to negotiate a price. Customers can shop around or ask several suppliers to bid for the work. However, these customers are merely trying to minimize the innate information imbalance in this type of situation.

Economists do sometimes argue that “supply doesn’t create its own demand”. This is true in the sense that no supplier can guarantee customers just because the supplier brings a product to market. However, that is why marketing functions exist. Suppliers use advertising to inform customers of the advantages of their products. This suggests that there is an information imbalance in favour of the supplier which the supplier attempts to minimize via advertising and other marketing techniques.

Suppliers do also carry out market research on customer expectations. That is an example where there is an imbalance in favour of the customer. However, on its own, it’s not enough to justify a generic conclusion.

It’s not clear that I(S) < I(D) in most of these situations. How did you arrive at the conclusion that I(S) < I(D)? How much does it matter to your technique that this equation is true?

1. Sorry if your comment didn't appear initially -- it was put in the spam folder for some reason and I had to fish it out. And thanks for your great questions!

First off, I(S) < I(D) is not critical for the empirical results I present here -- in fact, I always assume I(S) = I(D) i.e. "ideal" information transfer. I've talked qualitatively about I(S) < I(D) in cases where the model seems to fail, but I've never put together any empirical results based those discussions.

I'm going to put a new post based on your question (David Glasner asked the same thing as well) and some discussion with Peter Fielitz (who sent me an email for how he considered your same question from reviewers). I'll address a couple of your points here first. I will say that this is a work in progress so the following discussion could be incorrect.

One thing to keep in mind is that suppliers can never sell more goods than people wish to buy (just using market exchanges) -- that is my intuition behind I(D) >= I(S).

I agree that there is lots of information that suppliers have -- market research, past sales, specific industry information, etc. However most of that information is received demand information -- it tells you about the market distribution of goods that have been and/or will be purchased.

How should we think of demand information? Simplifying a bit, demand information is a distribution of widgets over consumers, call it P. That distribution function is more valuable than gold -- if you knew it, you could sell everyone exactly how much they wanted at the maximum price they would buy. The market tries to uncover that distribution (it is trying to minimize the KL divergence) ... it's best guess is Q.

Q represents all the supplier information: market research, past sales data, specific market information.

But Q likely doesn't match P, so there is information loss.

Since Q and P are both probability distributions over a widget/price space, they both could have the same amount of information relative to a uniform distribution U over widgets and prices -- the KL divergences relative to U could be equal D(U||P) = D(U||Q). In that sense, P and Q represent the same "amount of information" in a very specific sense.

However, since you can't sell more widgets than any consumer is willing to buy, we have to measure Q relative to P and look at the information loss measured in the KL divergence D(P||Q).

Regarding your examples: in considering one consumer, the basic conditions for the information transfer model fail -- we must have many consumers otherwise the demand information is trivial (i.e. there is one consumer so there isn't a non-trivial distribution P of widgets and prices over consumers).

I agree suppliers can create demand by informing consumers of products -- but that isn't information sent via market mechanism (price signals don't tell you about new products).