Must-Read: Ray Fair (2010): Convergence in Macroeconomics: Hoisted from Ray Fair’s Archives
Must-Read: Bluntly, your macroeconomic model–whatever it is–needs to mimic a Simsian VAR in-sample. If it does not, it has no claim on our attention: it is imposing assumptions that are neither the true structure nor even useful epicycle-like forecasting hypotheses. And if it cannot fit the data we have, it has no ability to claim to provide useful policy multipliers.
The Lucas critique remains true: a model can mimic a VAR and still not be useful for the purpose of providing policy multipliers. But the anti-Lucas critique–that a model that does not mimic a VAR has no claim to our attention for any purpose whatsoever–is much truer:
Convergence in Macroeconomics: Hoisted from Ray Fair’s Archives: “There have been a number of recent papers arguing…(2010):
…that there has been considerable convergence in macro research and to the good. Blanchard (2009, p. 2)… Woodford (2009, pp. 267, 269)… Chari et al. (2009, p. 242) state: “Viewed from a distance, modern macroeconomists… are all alike.”… Galí and Gertler (2007, p. 26)… state: “Overall, the progress has been remarkable. A decade ago it would have been unimaginable that a tightly structured macroeconometric model would have much hope of capturing real-world data, let alone of being of any use in the monetary policy process.”… There has been convergence… [to what] I will call ‘macro 2’, [which] dominates… refereed journals….
My non-macro friends often ask why macroeconomists cannot just compare models in terms of how well they fit the data and choose the model that fits best?… It is not, however, common…. The only case I am away of is in Fair (2007, Table 1), where a DSGE model in Del Negro et al. (2007) is compared to the US model in Fair (2004)…. The four-quarter-ahead RMSE for real GDP for the DSGE model is 2.62%, which compares to 1.33% for the US model in which autoregressive equations are specified for the exogenous variables…. The eight-quarter-ahead RMSE for the DSGE model is 6.05%, which compares to 1.84% for the US model. The DSGE model is thus not accurate. This is, of course, only one example, and in future work more comparisons like this should be done…
I am with Ray fair here: Whenever somebody shows up with a DSGE model that they attempt to use for any purpose, my first question is: how does this fail to mimic a VAR? My second question is: how much do the factors in the model that caused it to fail to mimic a var–the factors that we know are wrong–corrupt your answers to the question of interest right now? My third question is: what validation can you present that this is in fact a useful linear approximation to the emergent properties generated by the true microfoundations–which true microfoundations your model definitely lacks?
More often than not, presenters give little evidence of having thought about any of these three questions before…