Intelligent agent modeling in economics

One request was this:

The role and future of intelligent agent modeling in economics.

Call me a stuck up sticky bit but I don’t see a bright future for this technique.  We already have, and have had, computable general (and partial) equilibrium models for decades.  In those models you try to estimate the parameters from empirical data.  The models rarely impress me but there are plenty of situations, such as estimating the effects of changes in the tax code, where we don’t have anything better.  And so those models survive, and will continue to survive.

What’s the important innovation behind intelligent agent modeling?  To introduce lots of arbitrary assumptions about behavior?  Greater realism?  Complexity?  Considerations of computability?  Learning?  We already have enough "existence theorems" as to what is possible in models, namely just about everything.  The CGE models already have the problem of oversensitivity to the initial assumptions; in part they work because we use our intuition to calibrate the parameters and to throw out implausible results.  We’re going to have to do the same with the intelligent agent models and the fact that those models "sound more real" is not actually a significant benefit.

What can be done will be done and so people will build intelligent models for at least the next twenty years.  But it’s hard for me to see them changing anyone’s mind about any major outstanding issue in economics.  What comes out will be a function of what goes in.  In contrast, regressions and simple models have in many cases changed people’s minds.

Sometimes a theory model tells you there are many many equilibria, as in much of game theory.
I believe that result is to be taken seriously and we should conclude
that many different things can happen in that situation.  I am
suspicious of trying to solve for the correct or most likely
equilibrium by introducing many more specific assumptions.

In my possibly overdogmatic view, economics is most useful when its models are relatively simple and intuitive.  We’ve run out of new models which are simple and intuitive.  So the theory game is over.  The standard, old data sets have been data mined to death.  We’re now on to the "can you build/create your own data set?" game.  That game can and will last for a long time; in some ways it will favor go-getter extroverts just as the theory game favored introverts.

I don’t yet see that there is a new game in town.  My preferred reform of economics involves more history and anthropology, I might add.

Addendum: Bob Murphy asks:

It may be apocryphal for all I know, but I once read that the editor
of the journal to whom Einstein sent his paper on special relativity
put it down and realized that physics would never be the same (or
something like that).

Is something like that even possible in economics?  What would it be like?  Say, the Lucas critique times 10?

The answer is no, in my view this is not possible, for reasons given above.

Comments

I agree that agent based models can offer the lure of 'realism' with little insight, but I still feel that more satisfactory explanations of business cycles and and bubbles will emerge from such models. I don't know much about it, but it looks to me like a more promising tool for capturing the dynamic effects of small deviations from rationality, information cascades and so forth. And interactions between banks, and banks and firms.

Well, I think you have provided some great insights and food for thought here, but that you are guilty of treading dangerously close to the "everything that ever will be invented has been invented" prediction, which so far has never yet been true. I agree that we need more history, more simple and common-sense and logic based theory. We need to toss the absurd and unrealistic aggregations and static theories and start fresh on the simple theories: which means theory is far from dead.

And on the agent side, I think you have good insights but have also missed the upside to the technique. You ask:

"What's the important innovation behind intelligent agent modeling? To introduce lots of arbitrary assumptions about behavior? Greater realism? Complexity? Considerations of computability? Learning? "

Complexity and learning are important parts, yes. But so is time, evolution, dynamic properties that matter in economics!

Think about the questions that have been so hard to answer with simple models (and may not be solved either by looking at empirical data): how does the time factor in the nature of competitive equilibrium change the assumption of market clearing, of one price, and so on? How do the many prices and products competing when out-of-equilibrium change the assumed outcome of the simple models in which one price is assumed?

If we are never *in* equilibrium, but also striving for it, does the GE model represent an approximation of the solution, or does the process matter when considering a policy (which may interfere with the equilibration process)?

The advantage of an agent model is that you have each agent separately represented, without aggregation. This means that you can make the same assumptions about behavior that you would normally make, but the agents can individually respond to policy changes over time, so that a return to equilibrium isn't *assumed* but either occurs or does not occur (or a new equilibrium occurs that looks entirely different) on the basis of all the factors involved.

You're stuck up sticky.

I don't follow the deduction:
"We don't have anything better than computable equilibrium models _therefore_ other approaches are unlikely to be useful."

You are currently engaged in a virtual debate with people who use simplified Keynesian models and who see challenges to such simplified to stupidity models as niggles. Surely that means that there is value in saying that the emperor is naked and that simplified models are deeply misleading for the most part.

And finally, there is no serious place for history in most equilibrium models, so you don't have a very coherent vision about the possible future developments in economics.

For what it's worth, I am not very enthusiastic about the computer sciency approach to economics -- but I think that more realistic modeling of agents can help in teaching us when coordination is and when it isn't a likely scenario. Equilibrium models can't do that at all.

You were lamenting the lack of understanding of such important issues in economics -- and yet here you are, digging your heals in the defense of the methods that are responsible for the lack of attention that such central issues received.

When simplified models are not very good at prediction and they're not realistic either, maybe we should try to focus on just one of those features (realism) and see what we can learn from that.

Tyler, I thought that your observation that "What comes out will be a function of what goes in. In contrast, regressions and simple models have in many cases changed people's minds." was useful, and I am surprised that the commentators have not engaged you directly on this point -preferring to sneer at your pretensions.

However, as far as I can see, nowhere in the foundations of economics is there an argument for delineating good theory/bad theory as a function of "what goes out is more than what goes in".

Do you have any papers that you could point me, and the other commentators to?

Dimitrios is right that Leigh Tesfatsion is the go-to gal on agent-based modeling,
not all of which involves "intelligent agents."

Luis is right that heterogeneous agent-based models are much better at modeling
speculative bubbles and crashes than CGE or pure theory models. Heck, there is a
whole category of pure theory models that "prove" that (rational) bubbles are
impossible, with a lot of economists actually believing that these theorems actually
said something useful about the economy, at least prior to October 19, 1987. I have
seen lots of agent-based models that can replicate things we have seen in the real
world financial markets that neither CGE nor pure theory have been able to do. I
shall not so modestly note a model I with Mauro Gallegati and Antonio Palestrini
have that shows the "period of financial distress" that Minsky and Kindleberger
described, and which Kindleberger argued occurred in 37 out of 47 historical cases.
So, agent-based models can replicate historical events others cannot.

The other obvious point made by some is that these models do not require equilibrium,
especially general equilibrium, which has probably never occurred anywhere ever, except
perhaps for one second by accident. Certainly lots of micro markets are in equilibrium
much of the time, especially if they are double auctions designed by Vernon Smith and
his gang, but general equilibrium is a mythical mirage.

I would agree, however, that simple models based on obvious and simple data are often
the most convincing. I know from experience that policymakers will not be convinced
of something unless they can see it in a bivariate graph that looks obvious. Econometrics
can show that such a relationship may not be meaningful, but nobody is convinced by a
very complex model or estimation that is not also obvious in this simple way (and, of
course, right now everybody is convinced that speculative bubbles and crashes do exist).

"Tyler, I thought that your observation that "What comes out will be a function of what goes in. In contrast, regressions and simple models have in many cases changed people's minds." was useful, and I am surprised that the commentators have not engaged you directly on this point -preferring to sneer at your pretensions."

I'll take this on. First of all, I do not sneer at Tyler, I think he has made some good observations. However, I think he is wrong about this.

Agent based modeling, unlike CGE and other aggregative computational techniques, has the ability to offer more than just a rehash of what went in.

A nice comment in a journal article I read once said that "mathematical economists write papers with 30 pages of equations, and at the end there emerge the same assumptions that were put in at the beginning."

That has long been a problem of economics - whether it is computational or not. But agent based models do allow for "emergence" --- you put in behavioral assumptions for agent actors, but the "equilibrium" for the whole emerges. Then you try out different policies, and discover how the system and equilibration changes. As always, it is completely dependent on the assumptions. There is no getting around that. But still, you are able to find out something that you could not with only scratch paper. It can tell you something that you, as an economist, could not predict beforehand.

This all sounds spot on to me. In particular, I agree that a deeper understanding of history would be more useful to most economists than a deeper understanding of chaos theory.

But I would say that there are some simple models (i.e., as simple as the linear regression) that might be useful to economists as an heuristic. The harmonic oscillator comes to mind. How else can you model a bubble and crash? I would say that this should absolutely become part of any economists' mental models after the crash of '08.

More specifically, economists should be working to make accounting rules and financial statements more user-friendly. Has there been any innovation in this area in the last 100 years? Financial statements right now report snapshots in time (balance sheet) and time-averaged measures (income and cash-flow statements). Sometimes you get turnover rates or the like in addition. But investors should have more information about the fluctuations in time (i.e., periodicity) of various inputs to and outputs from the firm. Why not give people a full time-series of various accounts? Wouldn't that make it much harder to engage in the off-balance sheet games that have been played over and over again by managers under ever increasing pressure to make the numbers?

Does anyone have any thoughts on the future of dynamic modeling in economics? Looking at the feedback nature of economic systems and trying to simulate the transient or dynamic behavior rather than the condition of equilibruium?

Barkley Rosser,

Agreed that harmonic oscillators are inapplicable, but it's not asymmetry that precludes them, it's simple EMH: convince me of an oscillator, and I will arb it into non-existance in 5 minutes. If it's real, it's gone. If it's spurious, well, I just added to the chance of a bubble or crash.

Modern economics proved to have been a miserable failure in many ways. Failure of modern monetary policy to achieve price stability and avoid crises. Failure of efficient markets hypothesis. Inability of equilibrium-based models to forecast effects of tax changes. Modern economics is mostly good to be taught in economics courses. It fails miserably in practice. Sorry to say the truth in a rough manner.

> What comes out will be a function of what goes in

This reminds me of what physics was like in the late 19th century. People believed that they could come up with a perfectly deterministic model of the universe based on a set of input parameters. This was all shaken up when things like the uncertainty principle and wave-particle duality were discovered and it was showed that physics could not be built off a deterministic model - hence, quantum mechanics was born.

If a discipline as predictable as physics cannot rely on deterministic models, what hope does any discipline based on human behaviour have?

Also, one major problem in any science is to assume that our models are perfect. As many of the commenters before me have pointed out, economic theory doesn't always apply in practice. Isn't this proof that there are problems with our models?

I think Tyler is right when he fears "intelligent agents" stuffed with code intelligible only to the modeller. I think the results of ABM should be driven by the interactions among the agents, not the agents themselves. Having ever more "realistic" agents is certainly not the right modelling strategy.

Tyler has missed to mention one area where ABM may be particularly useful: modelling emergence. This is curious because emergence is an old theme in Austrian economics: Menger thought it to be the most important issue in social science research, Hayek also talked about it a lot. But then, Austrians use words and don't employ mathematical or computable models.

The Cauchy distribution has no mean but it is symmetric about the
location parameter. You can balance it. What's your point?

My comment about the Cauchy distribution was in reply to another comment, now (apparently) erased. It was not a comment on the original post.

Red, my point is exactly what you miss: you cannot balance it.

pessimistic meta-induction

While I completely agree with Tyler that all models are based on assumptions, the important part in my opinion is the quality of the assumptions. Agent based modelling done right uses fairly simple assumptions about behaviour of people (they look for their personal advantage). Everything else follows from that.

Economic models on the other hand tend to use more complicated assumptions, which is a lot more tricky, since global economics is essentially a chaotic system, where minor influences can cause major disruptions, and catching those with assumptions is way more complicated.

So in my opinion you need simple models to tell you what might be feasable policy and complex agent based modelling to actually verify your assumptions.

sweeping statements here

Comments for this post are closed