Can an AI model predict perfectly and still have a terrible world model?
What would that even mean?
Our new ICML paper formalizes these questions
One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵
Yeah I feel economists (esp structural ones) in Amazon share the same pattern. Good at defining a MVP solution then transfer it to others to horn details.
Economists have contributed some of the core ideas to many fields. We played a big role in ad auction development. Many of the core ideas of finance (Black-Scholes, Modigliani-Miller, factor models, CAPM, Glosten/Milgrom/Kyle, etc) were created by, well, finance/econ academics
My friend found a ring while snorkeling in Rábida, Galapagos. It was lost by a man about a yr ago & has significant sentimental value. If you or sb you know lost a ring there, please contact me with the inscription inside the ring. We want to reunite it with its owner. #lostring
Hot take: The predoc system exists because of modern applied micro.
Research via ‘labs’ has increased the amount of delegatable busywork. In the hard sciences this work is done by PhD students. But because econ PhDs are on a pure mentorship model, this leads to predocs.
Two things I really echo based on my industry experience:1) interpreting the data is often more important than fancy statical methods; 2) the magnitude of causal effects is often not stressed enough. Always ask: what does the number mean, why it matters? https://t.co/vvw79aFMGz
We describe our re-analysis here (comments welcome!): https://t.co/gDpuALIGUg.
The Journal of Population Economics is taking a look at these issues.
We thank @RiccardoCiacci for facilitating this re-analysis by sharing parts of his do-file with us!
11/11
Maybe the solution is to treat package references differently. Say put them in the appendix etc. It should be put somewhere to give credit to the package creators.
@just_cameron I think it's one of the ways people reduce the number of references, especially since some journals have restrictions. You can easily exhaust half of the references just by citing all the packages you used in the script. I'm not saying this is the right thing to do btw..
One question: people in industry often appreciate the first order principle (me too!). IMO econ applies first order principle quite well, but then people decry it for oversimplification (see all those investment books). Why?
I was a bit shocked to see Penn State ranking so high up there. During my 6 years there, rarely did anyone bring up anything about EJMR (except for JM rumors during JM season). One possibility is that a few fervent users could contribute disproportionately many posts…
Useful for forming a prior about Econ PhDs from these schools?
I always knew about EJMR during my PhD but I was already so depressed and anxious about my future that I didn’t think I needed an excuse for more. After the job market with multiple job offers in the bag, I went on EJMR one day bc I was curious how everyone else was doing. I took a cursory look. I thought “oh man this place is a cesspool of losers, mostly single lonely sad men stuck in interminable PhD programs prob sitting in some moldy dimly lit basement of some old building, angrily typing away snidely snarky anonymous comments. The posts were full of arrogance, bitterness and quiet despair. What a truly sad existence!” I clicked close and never went back. I went through a very dark period during my PhD, which wasn’t unique as I came to learn. It was hard and I was often very lost. I was very glad that I didn’t go on EJMR though and learned instead on my friends inside and outside the program. I learned a great deal from my PhD experience and I’d like to think that it made me a more resilient person.
https://t.co/Nj9MDhd25k
Great observation. Many structural empirical work now carries the torch of old-school price theory. I would also add that the most interesting reduced form paper also uses price theory to inspire research questions, though more implicitly.
To some extent. True, no one teaches Micro like that (see Mascolell et al. (great) textbook, which reads like a real analysis book with Axioms and Proofs and hemicontinuous correspondences and various Fixed point theorems). But a lot of the "structural" stuff Levitt decries is in part mostly applied Chicago Price Theory.
Consider two excellent examples of this year's job market:
4/n
I really feel upset that Price Theory was a lost skill now. I was lured into the economics world purely because of it, and I always feel that’s the economists’ comparative advantage. Such a pity
Levitt on the legacy of the Chicago School of Economics, including how "Every macro department feels a lot like heavily influenced by Chicago", "Chicago price theory really has lost" and a phone call with Friedman in the 1990s with Friedman being very upset about the decline of price theory:
"Go back to the '70s, there's a real gap between how Chicago thought about macro and how Chicago thought about micro compared to the rest of the world. And interestingly, the Chicago view essentially won in macro and our students placed well and influenced... Every macro department feels a lot like heavily influenced by Chicago. And I think for better or worse that has been a success story for the Chicago way of thinking. I think just really the opposite for Chicago micro; we have not had very many students who've gone out and been influential, maybe Ed Glaeser being a clear counter-example to that. And I think in the marketplace for ideas, I gotta say that the Chicago price theory really has lost. And it hasn't caught people's imagination. And I remember I was on a call with Milton Friedman as long after he left. He left Chicago in 1978, but this must have been 20-something years later where he was upset that Chicago price theory was not doing well, that it wasn't being appreciated. And I remember Casey Mulligan saying, "Hey, Milton, I thought you believed in markets. Let's just face it, price theory is losing in the market for ideas." And Milton Friedman got so upset about that. He believed in markets until it applied to Chicago price theory where he thought that it was the right way, so markets shouldn't have any bearing on it. But I think that's just the truth. That the people who you think of as being the logical heirs to Chicago price theory, the two that come to mind really are Ed Glaeser and Jesse Shapiro, they're not at Chicago. And with Kevin [Muprhy] retiring, there isn't, when Kevin Murphy retiring, there just isn't anybody around now really, other than Casey Mulligan, who could really keep the torch going. And the movement in terms of textbooks and what people are taught is just so away from what I think of Chicago price theory, which is not as mathematical, it's more about how you take the simple tools the very old tools you know tools that go back to people like Marshall and how you use them. It's really the skill that I see in Chicago price theory (one that I don't have) is how do you look at a problem and understand it to the lens of the right tool. And it's not complicated. It's usually very simple. Once you can see it, it's usually not much more than intermediate micro is what gets applied. And it's more artistic. And I really feel like our field has moved towards technicality. Harder proofs. More mathematical. And I don't see any going back. I think it is essentially lost to posterity at this point.”
@upanizza@joachim_voth My theory is that it’s a time mismatch. People do predoc because the Econ market was good in 2010s with many international universities hiring + strong private sector demand . Things are changing, but behavior adjustment is slow
@huihan_zhang By 3) I mean student treat themselves as “consumers” and therefore have the right to complain and bossy around. But they don’t have the same way to the real boss.
Happy Holidays #EconTwitter. I just published my Structural Estimation lectures (with text, code, and exercises) from my time @uchicago, @BeckerFriedman, and @UChi_CompSocSci as section of my new online @jupyterbook@ExecutableBooks entitled, "Computational Methods for Economists using Python".
https://t.co/bWHSBs8GtQ
The key chapters are:
- Maximum likelihood estimation: https://t.co/ga8nOz0Jvm
- Generalized method of moments estimation: https://t.co/Y2CWoVWvvf
- Simulated method of moments estimation: https://t.co/9MCc5hGEJ0
@comp_simon@MarlonAzinovic@john_stachurski@QuantEcon