Best of #econtwitter: The Lucas Critique vs. machine learning, special supplement
The tweet that started it all…
^much discussion ensued. Another claim — an econometric claim rather than a “macro” claim:
Do note these econometric claims are, I think?, different from the Lucas critique. See especially Nathaniel Bechhofer’s comments below. Here’s a useful, simple concrete example of the Lucas critique.
The discussions below will ping-pong a bit between the two…
Some good back-and-forths:
Back on the econometrics side:
And a long post from Jason Abaluck here
Part of the issue — maybe all of the issue?? — alas is confused and/or confusing terminology and/or different people using terms differently:
^can continue downthread for several more interesting back-and-forths. Eventually…:
[…much tweeting ensued…]
[…much tweeting ensued…]
^me neither, but I liked reading the discussion — here’s where this line of conversation ended up — thank you to all involved tweeters for your service (genuinely!)
A different take
^longer thread on this claim here, “You can’t scale your way to data from counterfactuals that haven’t happened”.
Am I wrong to say that: this is both the heart of the matter and clearly wrong? If “carbon-based learning” can get around the Lucas Critique, then so can “machine learning”, at least in principle. Or as @EconTraina himself puts it “optimistically”:
The second volley
An extraordinarily constructive suggestion:
^“why think when you can code”, as an old boss used to put it to me. Well:
^alas the exercise in the post was not the right test:
^…continued discussion… also. There is an update here but unless I am mistaken it does not address the issue
Meta: “I definitely wouldn’t be doing any of this pre-tenure”
On a meta note: all of this discussion is beautiful
On a substantially more depressing note:
^🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲🥲
If you ever wonder why your humble correspondent is anonymous…
A related paper
Another analysis
This is the right test of the Lucas critique if I’m not mistaken:
Fin
Excellent curation. The terminlogy confusion point is crucial, when different camps use Lucas Critique differently the debate fractures. What I found valuable is seeing how econometrics vs macro framings diverge on whether counterfactual generalization is fundamentally impossible or just needs better inductive biases.