I will be at University of Tübingen on Tuesday Dec 2, and at NeuRIPS in San Diego Dec 3-8 spreading the good word of the Generalized Method of Moments in the age of AI. Econometrics lovers come say hi!

Global solutions with adaptive sparse grids are now implemented in Dynare.jl, thanks to @compsimon.bsky.social and team! This should substantially improve the ease of implementing fully nonlinear medium-scale macro models.

FWIW, there are data-based covariate selection procedures that allow for the possibility that some variables are mediators, not confounders. I even have one, though by now there are quite a few. All still require assumptions, and sometimes you get partial ID, but we can do better than ignoring it.

Is it true that I have completely lost my battle to make Double Machine Learning mean "cross-fitting Neyman orthogonal moments with an ML first stage" and not "a partially linear model fit by that procedure"? People should often be doing the former, useful thing, rarely the latter.

The new Rosalía album is astonishing. Run to your nearest source of music

Ethel Cain puts on a world-class live show.

Updated Fall time for free weekly open office hours for anyone in the world with Econometrics questions. Wednesdays 3:30-5:30PM Central European Time (US EDT 10:30/EST 9:30AM) or by appointment; sign up and drop by! Details and signup at: donskerclass.github.io/OfficeHours....

Updated Fall time for free weekly open office hours for anyone in the world with Econometrics questions. Wednesdays 3:30-5:30PM Central European Time (US EDT 10:30/EST 9:30AM) or by appointment; sign up and drop by! Details and signup at: donskerclass.github.io/OfficeHours....

NYC recommendations for a short trip tomorrow and Sunday? Looking for art, plays, LGBT nightlife, or econ/math/stats/CS seminars.

A nice Sunday: - Morning run - Shop for foundation at Sephora - Write short lit review on particle filters - Research meeting: find 2x2 matrix explanations for functional analytic ideas - Trans lesbian girl gang teaches me how to skateboard - Vegan döner - Pharmacy for bandages and ice packs

"A/S/L", also by @jeannethornton.bsky.social, in contrast, was written about and for downwardly-mobile 30-something women who used to be Boy Scouts and "pretended" to be girls in Final Fantasy IRC chatrooms, eg, precisely me. I cried continuously without stopping for the entire last half. A+

For a more reliable source to answer "what was the most foundational work in weak identification (and much more) over the past several decades?" I can think of no better place than the Cowles 2025 Conference on Econometrics Celebrating Don Andrews that I will attend next week.

New Substack post: the Nobel Prize goes to an economic historian and two economists of growth trying to answer the question, where do new technologies come from? someunpleasant.substack.com/p/growing-th...

Valid inference w/ LLM-simulated data: 1. Take subsample of texts, extract variables 2. Construct moments identifying param on those variables 3. Ask LLM to simulate variables on sample & remaining texts 4. Use same moments w/ simulated variables 5. Combine moments, estimate jointly w/ 2-step GMM

Everybody wants to join me at Universität Zürich now, huh. Herzlich willkommen in der Schweiz, Esther and Abhijit!

Dynamic investment under rational expectations: Yale preparing press materials in advance for Sveriges Riksbank season.

For today's meeting of the UZH ML+Macro Finance reading group, I read about Forward-Backward Stochastic Differential Equations. Among other things, these allow for a stochastic version of Pontryagin's maximum principle for optimal control along a path. Some notes I made explaining how they work:

I’ve been training my whole life for this moment

I was incredibly glad to catch #jasmine4t on tour last night! They put on an amazing live show, including new tracks from the extended album, out today!

Meta-discourse point: under DeGroot learning, signals from multiple sources are reinforced, regardless of whether such information is downstream of a common source (see e.g. stanford.edu/~arungc/CLX....) "Tagging" signals appears to reduce this bias (Möbius et al 2015) but is rarely done.

But the whole point of computer science and machine learning is to confuse operations and the implementations. 🤣

anybody have favorite resources for learning about decision making under uncertainty/reinforcement learning for someone with sheer curiosity and not a ton of time?

Does anyone have a favorite short intro to Reinforcement Learning? I'm putting together some instructional materials. Book chapters, articles, tutorials, etc. Preferably on the applied side over theoretical. #MLSky #Statsky

Not finding anything here: does anybody have references on properties of Kernel Mean Embeddings of Kernel Mean Embeddings (or some other numerically tractable nonparametric representation of random measures)? There's work on higher-order KMEs for stochastic processes, but this should be simpler.

PSA: "Summer Fun" by Jeanne Thornton is included free on Audible until 8/26. I didn't think from the synopsis, a book about a trans fan of a band that's a thinly veiled stand-in for the Beach Boys, that it would be for me, but I'm halfway through and now I think it might be the great American novel.

Etiquette question for German speakers: Am I supposed to go to the restroom labeled "alle Geschlechter" or the one labeled "FLINTA"? If it's "FINTA" instead, does it depend on how gay I'm feeling that day, or just how dykey my outfit is?

We are proud to have our very own @economeager.bsky.social as part of the first ever human MCMC chain traversing the space of all possible names and genders. We feel sincere congratulations are in order to Rafe, Rachel Leah and all other Bayesians involved.

Re BLS news, a few years ago I resigned from a CBO position before starting because I worried that regardless of what I did, the presence of someone like me would be used to impugn the credibility of a highly professional, nonpartisan agency. I can't know, but I can't say I made the wrong decision.

Professional Update: I am now a Senior Research Associate (postdoc) at the Department of Finance, Universität Zürich, sponsored by Felix Kübler and Yucheng Yang on a grant through the Swiss NSF! I'm looking forward to being part of the Computational Economics community in Switzerland and Europe!

On an invitation from a friend, I just played a game in a local women's softball league. My controversial opinion on 🏳️‍⚧️ women in sports: everyone on both teams was really nice and I had a really fun time! ⚾

"Useless estimator" story: I recently consulted with someone using Maximum Score for a real well-motivated empirical project! I was excited because I'd known it mostly as a theoretical curiosity, w/ the ugliest asymptotic distribution known to man that also breaks the bootstrap. 🧵

A case of regress in applied econometrics: Arellano Bond 91 on feedback in panel data was the most cited econ paper of all time. Recent focus on DiD, with or without heterogeneity, has given us methods which fail under feedback, which is omnipresent in dynamic models. A useful reassessment here:

Yes, I always tell people Grace Wahba figured out everything in the 70s and we’re all just rediscovering it constantly. And yet…

A shot in the dark: does anybody have a good reference on proving coercivity for ODE operators (eg A[f] =d/dt f - g(f(t)), <A[f],f> > c ||f||^2), for some suitable norm)? I feel like there should be a standard source for this...

A typical centerpiece of papers probing "monetary policy's effect on x" or "how does x affect MP's effects on y" is results from a LP or VAR with an off the shelf monetary surprise series. Some papers asking important questions about how to do and think about these exercises ⬇️ #EconSky

I just found a bug in my proof, where the corrected version ends up being enormously simpler and cuts out the need for an embarrassing amount of work done on results needed for the incorrect version.

Can't believe Society for Economic Dynamics overlapped with Twinks vs Dolls Olympics this year, smh.

New blog post: How chaotic is chaos? How some AI for Science / SciML papers are overstating accuracy claims. www.stochasticlifestyle.com/how-chaotic-... #julialang #sciml #chaos #ergodic #ai4science

Casey Corgi Childers September 28, 2010-May 23, 2025 He was a very good boy.

Harron Walker's new book is, unsurprisingly, very good. On women, journalism, and "women's journalism." www.goodreads.com/review/show/...

Congratulations to Grace Wahba for winning the 2025 International Prize in Statistics, in recognition of her ground-breaking work on smoothing splines, which has transformed modern data analysis and machine learning 👏👏👏 Read more: rss.org.uk/news-publica...

As the semester winds down I have a gap before student final papers come in in which to read general interest economics. Anyway, Suresh Naidu, Winston Chen, & @treballen.bsky.social have fascinating work in progress on a quantitative GE spatial model of US slavery. dl.dropbox.com/s/uvhbdrqr88...

Explaining the Kitagawa-Oaxaca-Blinder decomposition to a student: "Now, the goal here is descriptive. Labor economists are not interested in the causal effect of an experiment randomly assigning gender... Well, I'm interested in running that experiment, but most economists aren't."

Post: On Feral Library Card Catalogs, or, Aware of All Internet Traditions bactra.org/weblog/feral...

Went to look up textbook results after getting the nagging feeling that an ML paper was reinventing classical ideas, and found this gem: "Not reading to the end of Le Cam's papers became not uncommon in later years. His ideas have been regularly rediscovered." At least they're in good company.

I made it official! I am now legally Rachel Leah Childers.

Note that the feature of this generalization bound that allows it to improve quantitatively over VC/Rademacher bounds in spite of well-known lower bounds (cf "fundamental theorem of ML") is that is is *not* uniform: hypothesis h shows up in the 2nd summand. If *realized* h has high P, you are fine!

Big if true! If I follow, this means that many "optimal" online algorithms devised by theorists that use a reduction to some completely infeasible thing like exponential weights on a sequence of covering balls can replace it with a kind of GP Thompson sampling, and actually become usable.