1/ Interesting post by @Undercoverhist on large macro models of the 60ies echoing recent comments by @sjwrenlewis @VMRConstancio @ojblanchard1. Even though it is not well known in the academia, these so-called "semi-structural" old-fashioned models are still widely used in...
2/ ... policy institutions, and practitioners are happy with them (to be fair, they are much smaller than their counterparts from the 60ies). They are used for both medium term forecasting and scenario analysis. They have several advantages over other models for these tasks.
3/ First, they are build on detailed quarterly national accounts and so are fully consistent with them. As far as I know (correct me if I'm wrong), institutions using a DSGE for their macro forecast still needs an excel spreadsheet to properly map the forecast into national...
4/ ...accounts. With a semi-structural model, you do this task inside the model itself, which is a good idea. Second, instead of thinking in terms of underlying structural shocks, you think in terms of fit. With every update of national accounts, you invert the model and see...
5/...how that affects residuals, and whether you are still happy with your equations' fits. If an equation has consistently negative residuals in the past few years, you don't interpret it as a sequence of some structural shock that will fade away over the forecast horizon...
6/ ...because you imposed an AR(1) structure. You interpret it as a misspecified equation and you need to think hard and decide whether and for how long to extend those residuals in the future. How do you make that decision? Experts' jugdment. This is why...
7/ ... we still hire economists instead of simply relying on algorithms. (Of course, you can do the same with a DSGE, but then its structural interpretation and all the constraints it imposes on the model structure are kinda useless) Third, the lag structure of each equation...
8/ ...is unconstrained and just informed by the data. This gives you a good fit of short run dynamics. DSGE try to emulate this by piling up adjustment costs, habits, etc. The beauty of a hybrid approach is that you just fit the best equation you can without worrying about...
9/ ...where it comes from. (If you are shocked by this for theoretical reasons, remember you are aggregating decision rules over millions of individuals). Fourth, even though the model is quite flexible to match the data, it still has a strong theoretical content...
10/...compared for instance with VARs. Theory is what guides you in your choice of RHS variables for each equation and in setting several structural features of the model (long term behavior, where and how do you put a Phillips curve, should there be a role for expectations,...).
11/ But perhaps the main advantage of working with those models is that you *know* they are not perfect. You know some parts are clumsy, you know which equation has a bad fit, and which part of the model needs to be carefully looked at in simulations and forecasts.
12/ You never use it blindly without making a lot of adjustments. Again, this is jugdment, but that's what we expect from forecasters and policy analysts. The bottom line is: a macro forecast is a mix of theory, jugdment & time-series analysis with the discipline imposed...
13/ ...by national account data. Semi-structural models give you a flexible framework to do all this in one place and I think this is why they are still used by practitioners after all those years, and despite the criticism they faced in the academia.
