apoorva lal

Implementing Economic Models in pytorch The standard workflow for economic models is to specify a loss function (typically a negative likelihood or a quadratic form g’Wg derived from a moment condition g:= g(X;θ) = 0 to minimize), algebraic...

On the econometrics in torch point: I also have an earlier crash course
apoorvalal.github.io/lalgorithms/...

2025-08-04 links papers Liu et al is a practical guide to proximal causal inference using regressions [the semiparametric analogue is cursed].

apoorvalal.github.io/lalgorithms/... new linklog - the proximal pentagram, econometrics in torch, benn jordan wizardry

the bridge function h() is a solution to a conditional moment restriction [analogous to nonparametric IV] and is cursed; in general you need to let the basis (A, Z, X) get big [2nd screenshot] and then you have to handle non-uniqueness.

ref arxiv.org/abs/2208.08291

2025-07-14 self-promotion I’m presenting my ‘getting away with 2WFE’ paper at Polmeth on thursday afternoon ; ping me if you want to grab a beverage and/or go for a walk presented this pyensmallen poster at scip...

apoorvalal.github.io/lalgorithms/... boy do i have the rube goldberg machine for you

Even less of an excuse to not keep up tbh. If your fundamentals are strong, LLMs are a productivity multiplier. If you're barely hanging on, you get nothing out of them and go to dept meetings and bully juniors or RAs over slack.

attn: associate profs in academia and L6+s in tech

[getting through to people who once could do things and labour under the misapprehension that they still can is a unique kind of hell. if you've fallen off, just delegate and go quietly into the night; you're a glorified MBA now]

also i think computing the sandwich
M = np.einsum("ij,i,ik->jk", XX, epsi**2, XX)

is the only time i've used einstein summation correctly. what a diabolical piece of notation.

i like them even for linear IV because the SEs are one-step [HC1 on the instrumented variable is valid] while this isn't the case for 2SLS
github.com/apoorvalal/d...

Typical day on the farm

IKM had this panel data interpretation that i thought was promising apoorvalal.github.io/pdfhosting/p...
but bridge fn estimation is a minefield due to the standard CMR reasons.
I guess proximal joins the front-door-criterion at the "works in theory but nobody uses it in practice" table.

meant to ask you guys at the conference: what's the simplest credible proximal causal inference application you've seen ? Like Angrist/Krueger QoB for IV or Card/Krueger for DID? Feel like theory has far outstripped practice in that design, to its detriment.

Did assa abloy get Fredrik (Savje) as part of the deal

Looking forward to visiting the svinsundbrua building in eng quad

Universities should return to their original mission. Yale should stop making fiddly locks with poor haptic response.

News you can use

Or a man who frantically throws in @torch.compile / @jit

the first person to add \mid X] to the E[Y1 - Y0] parameter got their face bitten off by the 'functional parameters absolutely cursed' tiger

idk how i did this but the overplotting in the t-learner quadrant is pleasing. most literal "this is what neural-networks-go-brr gives you you dunderheads"

my basic teaching slides on this literature (likely already out of date)
apoorvalal.github.io/pdfhosting/p...

5 (V, Five!) CATE estimation+policy learning papers in a single volume. Could probably teach a good class with this issue alone.

ASOS Digital Experiments Dataset A novel dataset that can support the end-to-end design and running of Online Controlled Experiments (OCE) with adaptive stopping. Hosted on the Open Science Framework

osf.io/64jsb/ i think the ASOS one fits your use case? Their paper goes over some open experiments datasets with multiple metrics as of 2021
openreview.net/pdf?id=79shW...

said galactic cluster should be called sherman-morrison-woodbury

loosely related - the fact that a guy called 'Parmeter' works on nonparametrics is p funny

Boy do i have the regression library for Hari Seldon
py-econometrics.github.io/pyfixest/qui...

EAs will turn their ire on jensen's inequality after the shrimps have been saved/cooked/eulogized

Maximum Likelihood Estimates for Gaussian Mixtures Are Transcendental Gaussian mixture models are central to classical statistics, widely used in the information sciences, and have a rich mathematical structure. We examine their maximum likelihood estimates through the ...

The fact that GMM (Generalized Method of Moments) works very poorly for estimating GMMs (Gaussian Mixture Models) is morally wrong.

arxiv.org/abs/1508.06958

Pun Run - YouTube

For this slander, i must inflict on your Zaltzman's best
youtube.com/playlist?lis...

GitHub - daskol/typst-templates: A list of paper templates in the area of machine learning. A list of paper templates in the area of machine learning. - daskol/typst-templates

I like the jmlr one here
github.com/daskol/typst...

Annoying answer is typst defaults are so much better thought out that I am painfully unlearning the personal dialect of Tex I’d made for myself with a 1000 line boilerplate file.