Tuesday, December 01, 2009
CBO is now saying that the stimulus is working as planned. Menzie Chinn provides an outline; Derek Thompson calls it "the handiwork of dirty Keynesians." If by dirty Keynesians you mean people who use large-scale macroeconometric models, then I would say you're right. But that's what everyone uses in this business.
You would think we'd dump them; there's a nice bit of analysis from April by Eric Falkenstein explains why they fell out of favor: Complexity in modeling made them pretty much useless in forecasting. I appreciate Chinn's defense of the ceteris paribus assumption: "without the stimulus, output would have been x% lower and employment yyy thousand workers fewer", etc., without having lived in that world. That's quite true. But you need somehow to decompose the error from this nasty graph in to what was because of modeling error, what was an external shock (something that we used to teach in forecasting with Theil's U statistic.)
So why aren't they dumped? The answer comes from having all those equations: It's quite easy to mine a large macro model for a story of why your forecast went wrong than it is to use one of those three-equation VARs that Chris Sims showed were so much better for not having heroic assumptions. I can better talk my way out of a bad forecast, the more moving parts I have to wave my hands in front of. Try to explain a black box, as Scott Beaulier did with his students, and you get confusion and a pink slip as a forecaster.
And this is, by the way, why I think most of the estimates offered for long term budget impacts of health reforms and cap-and-trade bills need to be taken with a very large serving of skepticism; if you rejoice when the model goes your way, you have to accept when it doesn't, and you have to accept the large errors their models can make. They come from models most graduate programs do not teach our graduate students any more -- those students who have to work with them later learn them after they get their doctorates, if they are economists. You don't need even to be one to run one: Fairmodel, discussed earlier, is a 30-equation plus 100-identities model that you can try at home.
(BTW, in our program, all applied economics masters students get a graduate0level course in forecasting and do know a thing or two about these models. Alumni report it helps in the job interview, FWIW.)