Gene Callahan makes a good point. Economic models, he says, are things you should use, not believe. I'd like to amplify this.

For me, economics is primarily a practical discipline. It is not about pompous armchair handwaving, but a practical guide to better decision-making in the the real world: policy-making and institutional design are subsets of this. However, the real world is a complex place. And the solution to complexity is often to satisfice - to pick decision-rules, or models, that are good enough. As someone once said, what matters is what works.

Take, for example, the efficient market hypothesis. Strictly speaking, this is not true: for me, the strongest evidence against it isn't windy theorizing about the Grossman-Stiglitz paradox but the empirical fact that momentum stocks beat the market. But it is useful. Sure, the investor who acts as if it were true would miss out on momentum profits (and on the out-performance of defensive stocks), but he'd avoid countless big mistakes such as paying high fees to mediocre fund managers; trading too much; overpaying for new flotations or "growth" stocks; and so on.

The EMH might not be true, but it's good enough for practical purposes*.

In fact, in a complex world, there is a positive danger in seeking the truth. Gene gives the example of LTCM whose trading rules were true until they suddenly weren't. There are other examples. Goldman Sach's David Viniar's famous bleat that "We were seeing things that were 25-standard deviation moves, several days in a row" expressed the fact that his risk management models were true for a particular dataset but false when that dataset changed.

This highlights an important and overlooked trade-off - between optimization and resilience. A model that gives us a true answer in some states of the world can give us fatally wrong ones when those states change: this is what happened to banks' risk management models in 2007-08. A model that was good enough - such as "shit happens" - might have served banks better** than ones that were "true" in the sense of perfectly fitting a small data sample.

My point here should be a trivial one. We should ask two questions of any model. One is: is it good enough? The other is: what's the worst that can happen if we trust it? As some guy once said, it's better to be roughly right than precisely wrong.

* especially if accompanied by the "sell in May, buy on Halloween" rule.

** I mean better in the sense of aiding their survival and banks' shareholders. They might not have been better in the sense of maximizing bankers' pay - which is of course the only purpose banks have.