I was wondering if anyone had any thoughts on this paper and blog post from Daniel J. McDonald and Cosma Rohilla Shalizi. Their thesis is that the estimation of DSGE models is plagued with issues. Namely, they find that when estimating the Smets-Wouters model on simulated data, it struggles to estimate the true parameters even as the length of the data grows large. My impression from doing this in the past was that estimating DSGE models on simulated data typically performs well in this regard, so I was surprised to see their results. They also find that swapping out series with each other, e.g. relabelling wages as the interest rate, leads to a better fit of the model to the data. Notably, they do everything in R rather than using Dynare.
I was just curious to see what people made of this.
I did not read it completely but I intend to do it. It does seem interesting – contrarian viewpoints and evidence is always intriguing, isn’t it? Thank you for bringing it to our attention. I am curious too – to learn what others make of it.
I think the results of the paper are fairly known. Like not all parameters can be or should be estimated. In the paper, it says, “Here we can see that while many parameters are well-estimated with decreasing variability as the training sets grow, others converge to the wrong values or fail to converge at all.” So many parameters seem ok. Quite a few are not ok, it seems.
The following statement appears in one estimated DSGE paper by Stéphane et al. (2007), “Some parameters are fixed prior to estimation. This concerns generally parameters driving the steady state values of the state variables for which the econometric model, including detrended data, is quasi uninformative. The discount factor is calibrated to 0.99, which implies annual steady-state real interest rates of 4%. The depreciation rate is set equal to 0.0025 per quarter. Markups are set to 1.3 in the goods market and 1.5 in the labor market. The steady state is consistent with labor income share in total output of 70%. Shares of consumption and investment in total output are respectively 0.65 and 0.18.”
So it seems it is well known that some few parameters are better not to be estimated. Maybe there are more of them besides the ones mentioned above. Hope it helps.
I absolutely don’t trust that paper. People like SGU (2012) in their Econometrica paper have done similar Monte Carlo exercises and they worked fine. I should also point out