[yourqiwi]

> I am going to guess you have never written down an economic model, much less presented one at an economics seminar. If your model can be "tuned to forecast whatever", you are going to be annihilated. You will never be able to get away with presenting a model like that. There is no place for it in the field.

A poorly specified model will get ripped apart when presented in an academic context. I've seen it happen in seminars. Questions about the standard errors (robust? clustered?), omitted variable bias, whether things have been detrended, fixed effects, quality of the instrumental variable, etc. When econometrics is done properly and by the book, it can be a useful tool for analyzing observational data.

But these tools get regularly applied outside of academia and the results are presented to people who can't critically evaluate them. Nobody knows that you ran a million other regressions in STATA, and kept the significant one. Or that your result isn't robust when you excluded/included such-and-such variables. When research is motivated in a particular direction (as it often is in politics and consulting), then you can often pull the levers and turn the dials until you get the result you want.

[huitzitziltzin]

> Nobody knows that you ran a million other regressions in STATA, and kept the significant one.

Totally fair criticism and I agree with it. Not at all specific to econometrics though!

> Or that your result isn't robust when you excluded/included such-and-such variables.

100%! Also definitely not specific to econometrics.

> When research is motivated in a particular direction (as it often is in politics and consulting) ...

The CBO is non-partisan. I have more faith in them than that (unless you have inside information?). The whole tone of the specific report linked is very much "here is what we know and don't know and here is our best estimate, which has limits." That is research as it should be at its best. Litigation consulting is different.

> ... then you can often pull the levers and turn the dials until you get the result you want.

Here I don't agree. There are limits. It's not that as simple as that. I can't turn an estimated -5 into positive 500 just because I want to.

[yourqiwi]

> I can't turn an estimated -5 into positive 500 just because I want to.

I didn't have quite that much precision in mind. More like "flip the coefficient's sign" or "make the coefficient bigger" or "make it statistically significant", which are all often doable with enough experimentation (as much as I hate to admit it).