Monday, March 8, 2010

Should Economics adopt methods from Physics?

In their quest for universal laws, physicists use data mining methods on large datasets and uncover regularities that beg for a theory. Should economic adopt similar methods?

Austin Gerig thinks so, and he bases his entire argument on the distribution of daily returns of the stock market. It is true that Wall Street is full of Physics PhDs who do data mining, looking to beat arbitrage and the efficient market hypothesis. But that is not Economics. There is much more to Economics than studying the daily returns on the stock market, even if your family or neighbors think this is what you do as an economist. In particular, Economics is about studying how agents' behavior changes as the environment changes, something purely statistical methods will never uncover. And do not get me started on theory-less data mining. We already have too much of that in Economics, so do not let physicists do it, too.

4 comments:

Min said...

I agree wholeheartedly with your main point. Economics is not physics.

However, there is one thing that economics might profit from by taking from physics, and that is operational definitions. OC, some economic definitions are operational. But the lack of operational definitions, coupled with the ideological aspects of public economic debates, makes it hard to tell what is meaningful and what is hot air. :(

James Reade said...

There's actually far too little proper data mining in economics, and most empirical work suffers as a result. Economists, tied to their economic theories, start with a very small econometric model which suffers from many biases and other problems, meaning it doesn't support what they want. So they either abandon the model or tweak things until they look right (choose the right set of weak instruments, for example). That's bad data mining, and economics is full of that.

But decent, useful data mining, there isn't enough of. This data mining starts with economic theory to motivate a problem of interest, considers all the theoretical explanations and begins with a general econometric model that satisfies the statistical assumptions it's based on. From there a smaller model is selected if possible, removing irrelevant variables along the way.

A search for "general-to-specific" or "David Hendry" in Google will give a great volume of information on useful data mining for economics...

Austin said...

Thank you for pointing me to your blog and for linking to my paper. Your comments are well taken: economics is certainly more than just analyzing stock data, and data mining is not the best way to formulate a theory. This isn't quite what I was trying to say, however. My point was simply this: physicists often search for regularities in the natural world and formulate simple theories to explain these regularities. I think it would benefit economists to use this methodology when researching economic phenomena. This isn't simply data mining and it's different from what quants do on Wall Street (My guess is, if you're using data mining techniques to find a regularity, then whatever you find isn't going to be that interesting to a theoretician). I agree that this methodology is not applicable to all of economics, but it is useful in researching more than just financial markets, even if the specific examples I gave were limited to this.

jenizaro said...

Austin: I think economists already do what you suggest. In particular the empirical regularities you point out in your paper have been extensively studied by economists in the so called "Efficient Markets Hypothesis" literature.