Friday, November 05, 2010

A Slight Illness - The Doctor Jests, A King Today - Tomorrow He Is Dead

Has the global financial crisis discredited Dynamic Stochastic General Equilibrium (DSGE) models? It seems to me that that may be so, but I wonder if new criticisms of DSGE models have been put forth. It seems to me the fundamental flaws of such models have been unaddressed for decades. Cosma Shalizi describes these models in a critical way. Joseph Stiglitz is not impressed with them either:
"It is hard for non-economists to understand how peculiar the predominant macroeconomic models were. Many assumed demand had to equal supply - and that meant there could be no unemployment. (Right now a lot of people are just enjoying an extra dose of leisure; why they are unhappy is a matter for psychiatry, not economics.) Many used "representative agent models" - all individuals were assumed to be identical, and this meant there could be no meaningful financial markets (who would be lending money to whom?). Information asymmetries, the cornerstone of modern economics, also had no place: they could arise only if individuals suffered from acute schizophrenia, an assumption incompatible with another of the favoured assumptions, full rationality." -- Joseph Stiglitz

As I understand it, Smets and Wouters (2007) is an example of a DSGE model widely approved of by mainstream economists. Sbordone et al. (2010) is a recent presentation of an introductory DSGE. One can see that the output of these models is a set of stochastic processes meant to model certain time series available in empirical data. Nominal interest rates, (real) income, the inflation rate, the volume of one-period government bonds, employment, and nominal wages are all examples of such time series. The input into such models is another set of stochastic processes. These inputs are given names that suggest they are random terms in functions characterizing either government entities - e.g., monetary policy shock - or agents in microeconomic models. Examples of the latter kind of names are a household discount rate shock, productivity shock, markup shock, and firm discount rate shock. Stochastic processes are specified by parameters of certain probability distributions. As one can see from the names of these inputs, the agents are supposed to be optimizing, including across time. A story, expressed in mathematics and supposedly of microeconomic equilibrium, connects the inputs to the outputs in the model. That is, the DSGE models are supposed to have microfoundations.

But they do not have microfoundations. I look for a number of mistakes in such models:
  • Are inputs into production function measured in numeraire units? (The numeraire is often taken to be a basket of consumer goods.) Joan Robinson (1953-54) explains why measuring the quantity of capital in production functions in numeraire units is an error. Notice this is not solely a question of the aggregation of capital. A model can have a continuum of capital goods, yet still exhibit this mistake.
  • Are representative agents used? Kirman (1992) explains why the use of representative agents is unfounded.
  • Is money modeled? Frank Hahn (1965) explains why money does not matter in General Equilibrium models, even though it does seem to matter for actually existing capitalist economies. Mainstream economists have a couple of strategies for introducing money in an ad hoc way into DSGE models. But I am not convinced the typical modeler has ever managed to address Hahn's point.
  • Is the possibility of multiple equilibria taken seriously? Is it demonstrated that non-equilibrium dynamic processes converge to the modeled equilibrium? Richard Goodwin (1990) illustrates what a macroeconomics looks like that, in contrast to typical DSGE models, takes dynamics seriously. Kirman (1989) shows that ignoring muliple equilibria and stability issues was demonstrated to be unfounded by the Sonnenschein-Mantel-Debreu results. Shiller (1978) long ago raised the issues of multiple equilibria and convergence. Shiller was critiquing the tradition out of which DSGE models evolved.
I haven't read much in the literature of DSGE models. It seems to me, however, that these issues are routinely ignored by many mainstream economists. Perhaps a wider range of macroeconomic models should be considered by serious researchers.

To read:
  • Wynne Godley and Marc Lavoie (2007) Monetary Economics: An Integrated Approach to Credit, Money, Income, Production and Wealth, Palgrave MacMillan.


BruceMcF said...

A broader range of models? Heck, I might go for an equally narrow range of models, if the range consisted of not-known-invalid models as opposed to known-invalid models.

Tomboktu said...

The URL for Sbordone et al isn't working. I got it here:

Robert Vienneau said...

Thanks for the comments. I've updated the post with the link suggested by Tomboktu. I might not mind so many models known to be invalid if the invalidity was discovered just last year, say, and their retention was just a matter of inertia that one had some reason to believe would be overcome in maybe half a decade.

Cosma Shalizi said...

I wouldn't mind if the systematic errors were persistent, well-known and ineradicable, so long as they were small. (I'm fine with mechanical engineers using classical mechanics and not quantum field theory.) But I don't think the DSGE-mongers have even this on their side.

JW Mason said...

There's another problem I've been wondering about, which is how one can square agents optimizing across time with unanticipated shocks. Are rational expectations and comparative statics even logically compatible?

Robert Vienneau said...

You might find of interest David Hendy and Grayham Mizon's paper On the Mathematical Basis of Intertemporal Optimization. It argues that the typical rational expectations approach is inappropriate when time series exhibit unanticipated breaks.

Anonymous said...

Great blog with lot of thought provoking entries!

Actually,all problems you mentioned are present in the DSGE framework.

However,there is another interesting "feature" of these models.When estimating a DSGE (or RBC,CGE) model,a lot of sophisticated econometric inference methods are used.But,in general,the parameters you get by these econometric methods are patently absurd.

So,the model gets "calibrated",which means that you search for parameters that 1.are consistent with the equilibrium concept 2.meaningful in the sense that the system won't explode or households don't consume twice their wages in the steady state etc. 3.and you do this in a systemic manner (all equations should be meaningful,simultaneously - a feature that the econometric inference does NOT provide in most of the times)

In short,in real life,on top the econometric inference,a lot of rule-of-thumb and (albeit intelligent) guesswork are embodied in a DSGE model.

Godley and Lavoie in their book also mention this "calibration" issue.(In the foreword,if I'm not mistaken.)

So I'd compare a real-life DSGE to a very high order polynom: you can fit it to every function (curve) in the world by calibrating its coefficients - but its performance out-of-the sample will be questionble the least.


Robert Vienneau said...

I agree that "calibration" is a problem with DSGE models.