The Wall Street Journal notes the Institute for New Economic Thinking was launched last year with $50 million from financier and theorist George Soros. The institute so far has approved funding for more than 27 projects, including efforts by aimed at developing new ways to model the economy. What are their ideas? Remember that Soros' Alchemy of Finance presented the big idea that prices are irrational, which he defined as biased 'in one direction or another' (his other idea, that markets can influence what they predict, playing off his role in the 1992 EMU crisis). He's sympathetic to researchers who would confirm he's not just rich, but a profound, original thinker. Here's the WSJ:
The problem, says Doyne Farmer, is that the models bear too little relation to reality. People aren't quite as rational as models assume, he says. Advocates of traditional economics acknowledge that not all decisions are driven by pure reason.
His proposal: Create a richly complex, computer-based simulation of the economy like those scientists use to model weather patterns, epidemics and traffic. Given enough computing power, such "agent-based" models can include millions of individual players, who don't have to be rational or agree with one another. Instead of equations that must be solved, the players have open-ended rules of behavior, such as, "If I've just turned 55 and I'm feeling blue, I'll buy a sports car."
Gee, macro models with lots of equations...hey, that was popular in the 1970s! Jan Tinbergen made the first one in the 1950s. They did not stop using them because they ran out of funding, but rather because they were such an obvious failure. With hundreds of equations based on mini-models of banking, savings, or industry behavior they predicted the past very well, but the future, not so much. Robert Lucas showed these models were internally inconsistent, in that the modelers assumed people would have, say, inflation expectations of 2% while the model implied 4%. Now, this is kind of tempting when applying models to real data, because data is historical, and with the benefit of hindsight, people are biased in one way or another. The problem is, at any one time we don't know which way. Chris Sims showed that much simpler vector auto regressions could predict just as well. By the late 80's these macro models were anachronisms, and the only purveyors are economists over 50, and global climate modelers.
Now, Doyne Farmer has been working on this problem for 20 years, and it's very naive to think all he needs is 10 million dollars and the resulting CPU to cleanly break out of the macroeconomic cul de sac. The idea now is to add new irrationalities, with a hat tip to Kahneman's behavioralist approach. The problem, again, is that irrationalities tend to explain everything, via too much anchoring or lack of base weighting, over-reaction and under-reaction.
The idea that adding new 'behavioralist' equations to the old macro modeling approach is naive because the old approach had hundreds of equations--assumptions about causal relations. Economists are very good at rationalizing behavior, so if some random assumption worked within this paradigm some economist would have accidentally found it the way the momentum effect was found in equity returns. Hundreds, if not thousands, of really bright people looked at this problem for decades, and then retired, and their interns eventually all decided to not follow them to obscurity.
Anything that works can be modeled as rational via clever theorizing; there are very few constraints on the equations in large-scale macro models. In that sense, irrationality with a complex model of this-causes-that is not 'new'. You are going to need something more specific.
I agree with your general sentiment about these agent based models - they have been around for 20 or more years, and have never lived up to their hype.
You really cannot get much from these models that you didn't put into the rules for choice.
Adding behavioral assumptions is not going to make any difference - it is just another utility function.
One should know at least roughly what drives the output (inner working of the model) and how sensitive the output is to input changes. Otherwise the model does not contribute anything to our understanding of the world (I don't even dare to mention forecasting).
May people forget that more often than not it's not feasible to come up with a full-fledged model. Recently we had this discussion about the effect of a x% rise in minimum capital requirements for banks on medium and long term GDP growth. Methinks making a list of possible positive and negative effects and thinking about it in an educated manner makes more sense then running a monster DSGE model that is partially calibrated to old data and to what the modeler thinks is a reasonable answer.
Can anyone think of a good way to short these models? :)
I agree, and this is a great blog - at least the parts I understand!
Looking at my own behavior - my investing and spending has changed completely in the last 12 months as awareness broke on the likely outcomes to the 2008 rescue. So I buy books and read different blogs, "Monetary Regimes and Inflation", "This time its different", "When money dies"...
Book sales and web-stats it seems to me could be a leading indicator of the behavior of this agent.
Every model is right (sometimes) and wrong (sometimes). Economic models are helpful when they expand our vision of what is possible. They make us stupid when we try to apply one model to every situation, and thus fail to consider some developments as possible.
If Farmer actually made as much money at Swiss Bank as is hyped, why is he schnorring research funding from Soros?
Post a Comment