Dynamic Stochastic General Equilibrium (DSGE) models are designed to do policy analysis. They are based on microfoundations and calibrated or estimated to provide quantitative answers to policy scenarios and in particular study environments for which there is no historical precedent. These model are not designed for forecasting, traditional statistical tools, where ad-hoc models with minimal theory are fitted to the data, are optimized for this. But one may still ask whether DSGE models are any good for forecasting.

Rochelle Edge, Michael Kiley and Jean-Philippe Laforte take the the DSGE model used by the Federal Reserve Board to check on its forecasting performance and are surprised by its success. If fact, the specific DSGE model they use, dubbed "Edo", outperforms both the time-series model of the Fed and the staff predictions. Now, "Edo" is not a simple and small real business cycle model, it is a heavy beast with lots of details. Adding complexity allows a model to provide richer results, but contrarily to statistical models, it may lower the performance of a DSGE model. Indeed, adding new features may undo well-performing components of the model, as everything is interlinked ("general equilibrium"). It is all the more remarkable that this large model works so well.

Interestingly, this exercise is performed with real-time data, i.e., data that was available at the time the forecast would have been made, neglecting subsequent revisions. This is important from a policy point of view, as real-time forecasts are those that really matter for policy decisions.

To all naysayers, DSGE models are good for forecasting, and better than traditional models. The fact that they may not have predicted the current crisis does not mean that they need to be rejected en bloc. Did traditional, statistical models do any better? Of course not, as they are notoriously bad at predicting turning points. Also, the fact that DSGE models (among others) failed this time does not mean that suddenly all lessons learned from these models are not valid anymore and that Keynesian policies are suddenly effective again.

## Tuesday, July 28, 2009

Subscribe to:
Post Comments (Atom)

## 4 comments:

How do you respond to this?

http://www.economics.ox.ac.uk/students/james.reade/J_James_Reades_Website/Blog/Entries/2009/7/28_Another_Defence_of_DSGE_Models.html

James Reade seem to believe that RBC=DSGE. the research nowadays, as far as I can read it, does not assume perfect competition, complete markets and representative agents. In fact, it seems you will have a lot of difficulties getting anything published without heterogeneity and frictions of some sort.

It is sad to see a literature being criticized on grounds that are twenty years old.

"Adding complexity allows a model to provide richer results, but contrarily to statistical models, it may lower the performance of a DSGE model."

adding complexity to a statistical model also diminishes its forecasting performance; that's one of the reasons for preferring parsimonious models.

EL's critique of statistical models is weird. He doesn't seem to realize that all of modern econometrics (since Haavelmo) is statistical modelling built on top of economic theory.

Fascinating. I should say that the link above is broken but I'm sure you can find it by subbing the bit before "/J_James_Reades_Website..." with "www.james.reade.co.uk".

I didn't realise anybody actually read my blog! I should also add that I don't think DSGE=RBC, I simply think the emphasis of RBC remains in DSGE - a fundamental disregard for what economic data actually say.

The funny thing, of course, is the kind of work I do is criticised on grounds that were refuted about 20 years ago too - the Lucas critique and its implications for econometric modelling.

The other funny thing is that this blog is trying to point out that what happened to Keynesian models 20 years ago when they failed to predict stagflation should not happen to DSGE models today because of their predictive failure.

I agree. As economists we should not be throwing out models en masse because they go wrong.

The difficulty remains that large gap between economic data and DSGE models. Econometric methods and economic models are both vitally important but hopefully we'll one day soon get past the usual DSGE paper which sets up a straw man in an "econometric model", which the DSGE model beats (surprise surprise).

I also hope one day to actually generate some research of my own on this topic...

Post a Comment