More on DSGE Forecasting
Mark Thoma quotes
Forecasting the Great Recession: DSGE vs. Blue Chip, by Marco Del Negro, Daniel Herbst, and Frank Schorfheide, Liberty Street:
Coincidentally commenter Luis Enrique asked me what I thought of that paper on Gmail.
this is commentator Luis Enrique [skip] I am enjoying your to and fro with Simon WL anyway, I just wanted to send you this link, in case these guys are not on your RSS, as a continuation of sorts of your post on the Vox paper on the same topic
My thoughts after the jump.
I don’t actually have an RSS and, as I admit from time to time, don’t generally read the economics literature, so it was new to me.
The key part of the NY Fed post is the figures most of which simply show that both the models and the blue chip forecasters failed to forecast the great recession. These are forecasts of quarterly GDP growth. The explicit model based forecasts are red, the average of Blue Chip professional forecaster forecasts is green. Blue indicates confidence intervals for the explicit model based forecasts. Click one of the links for actually legible decently sized figures.
I think the key passage in the explanatory text is
The chart shows the forecasts for three different DSGE specifications. We call the first one SWπ; this is essentially the popular Smets-Wouters model. The second is the Smets-Wouters model with financial frictions, as in Bernanke et al. and Christiano et al., which we call SWπ-FF. The other difference between this model and SWπ is the use of observations on the Baa-ten-year Treasury rate spread, which captures distress in financial markets.
The last specification is still the SWπ-FF, except that we use observations for the federal funds rate and spreads for the quarter that just ended and for which no National Income and Product Accounts data are yet available (for the January 10, 2009, forecasts, these would be the 2008:Q4 average fed funds rate and spreads).
The DSGE models which do a bit better aren’t DSGE models. The Baa-Treasury spread is added as an explanatory variable without any discussion of micro foundations. The text makes it clear that this variable is added in addition to the frictions considered by Bernanke et al and Christiano et al. Bernanke et al have the difference between a risky and risk free rate as an endogenous variable — it should depend only on the ratio of equity to the capital stock (IIRC). It isn’t a free variable in that model and can’t be added in addition to the frictions considered by Bernanke et al.
I claim that the model which does poorly rather than terribly is an ad hoc aggregate model without micro foundations.
The outcome on the graph is not the BEA’s current estimate (-8% at an annual rate for 2008:Q4). This means that the outcome as re-estimated months before the publication date of the post is well outside of the 90% confidence interval except for the last panel which shows an ad hoc model using contemporary data.
Why no forecasts using data from the trough ? There is no evidence in the post that the DSGE model had any success forecasting the recovery. There are no forecasts which use 2008:Q4 BEA estimates flash, advanced or “final” (but not really final since they were corrected in 2011).
The good looking figure adds another variable ad hoc. Oddly it just happens to track the GDP growth rate very well in this episode. It would not have a similar coefficient if estimated with more data — the spread has not been close to what it was in 2008-9 since 1962 (when the FRED 10 year constant maturity Treasury series I looked at started) but GDP shrank at the same rate in a quarter in 1980 as in 2008:Q8. The past 5 years shows why adding the spread as an explanatory variable (even lagged one quarter) helps a lot. In the figure I didn’t lag the spread as I should have and would have if I could have with FRED. Even so, the peak is a bit to the right of the trough of GDP growth. But the point is that even shifted to the right one quarter (to lag) the Baa – Treasury spread’s high level and rapid decline fits the steep decline in GDP in 2009:Q1 and then the rapid recovery.
So, if I understand correctly, the DSGE model is, in realistic terms, absolutely useless; and adding a bit of ad-hocery makes it somewhat less useless, but still not at all useful.
Is that about right?
That is my impression Jazz. The ad hocery is worse than it sometimes is as it is ex post feacto ad hoc ergo crap. The decision that Baa yield minus treasry yield is a good indicator of financial stress was made after the period of extreme stress and a high spread. This approach is irresistable (at least I can’t resist) but it guarantees that you make good predictions of the past and almost guarentees bad predictions of the future.
THis approach is irresistable like you said. I am sure that many will have he sam opinion. yacht charter france