Things are grim. But in more frivolous news...
@jamesbrandecon.bsky.social and I have been chipping away at `dbreg`, a 📦 for running big regression models on database backends. For the right kinds of problems, the speed-ups are near magical.
Website: grantmcdermott.com/dbreg/
#rstats
[1/2]
yeah your collaborator on dbreg is definitely a little too willing to vibe-code to get things finished quickly.... Been interesting to see when it bites. But my thinking is that the only way I can figure out how to manage the LLM well enough is to stretch the limits, break things, then fix/iterate
New blog post, on the impact of cheap verification on data science hiring, work, and compensation: jamesbrandecon.github.io/blog/posts_h...
You say MCMC but I've thought about this before and didn't immediately see the gains. GPU speeds up some bayesian stuff (variational inf) but MCMC is basically happening in serial so what's the benefit? Just that each draw requires matmul which is now gpu-fast?
+1 and because the (implicit) exclusion restriction isn't as front and center. Easier for most, I think, to wave hands and say generally "I hope unobservables aren't too bad" than to say "boy unobservables are really bad here but I found one special variable that fixes it"
Wrote a second blog post! This time it's about using LLMs as part of an ensemble imputation method, relying on their "knowledge" of the world to provide additional prediction signal. Seems to work well, especially for bigger models!
jamesbrandecon.github.io/blog/posts_h...
P.S. I also tried to make plotting easier. Here, for example, are estimates from a model with two product characteristics and correlated preferences. More to come, including docs to make it easier to dig through the "problem" objects to extract results.
As with FRACDemand (which is now registered!), the coolest thing here is that it's fast. less than a minute to define the problem, estimate it, and calculate price elasticities for 500 markets with 20 products each
Have a day off so I made some small updates to FKRBDemand.jl (new name!). Hopefully it's now easier to use the Fox-Kim-Ryan-Bajari method to estimate a random coefficient model with market-level data (2-step approach described here: www.jamesbrandecon.com/blog/0jxkvfr...)
github.com/jamesbrandec...
Congratulations, Jeff!
Nice, thanks for explaining! Agreed, it seems useful to help see how I'm using memory while building something (which I still wish I understood better)
Is the goal just that GC can catch when to run? Or is there another performance benefit I don't understand? Seems like a lot of additional code so I'm assuming there's more value
This point you've been on has stuck with me as an industry guy. How do firms succeed while being terrible at stats reasoning? Because some big ideas are robust to that bad logic and (like VCs?) you get a lot of shots and only need a few to pay off. That changes how to think of data work for the firm
Finally, we put this all in a Julia package! Trying to take seriously that these complex methods cost a lot of time and effort to implement, we tried to make it as easy as possible to use our methods for estimation, inference, and some basic counterfactuals 6/6
Third, we apply our method to real demand from 12 categories of products at a large retail food chain. Our estimates agree with our simulations — constraints matter, and imposing them generates more reasonable elasticities and counterfactuals (below) 5/
Second, we show in simulations that our method improves upon the previous state of the art for this model, both because we can fully enforce constraints and because doing so gives us more accurate estimates of price elasticities. 4/
This two-step approach not only helps us impose new constraints -- it also guarantees that our constraints are satisfied *everywhere*, which we show is often not true of the best previous approach 3/
First, we develop a new quasi-Bayes approach to solve this (hard!) problem. We show that imposing constraints on nonparametric demand curves is better suited to MCMC sampling approaches than frequentist optimization, and develop a two-step process for sampling in our setting 2/
New paper with @adam-n-smith.bsky.social
papers.ssrn.com/sol3/papers....
Our paper develops a new approach for estimating demand nonparametrically while imposing economic constraints and comes with a new package, NPDemand.jl! Some things we do in the paper 1/
The paper looks really nice, but Prop 2 (no ME) feels stronger than is useful in practice... do we hold ourselves to that standard for any other measurement tool? Or am I being unfair/missing something
Ditto, my brain hurts enough switching between the languages themselves. Adding UI as another thing to switch is a pain
Wow, as an increasingly frequent R user I wish I'd known this was possible months ago -- thanks for figuring this out in public
Been (finally) reading through it in detail recently -- great paper!
Agreed (and I get this argument from ML people a lot), but funny enough we'll never know the counterfactual there either! I could walk around in the dark and get where I'm going but it'd still be nice to have the lights on
My experience is that the former often dominates, and is the default even in settings where academics would lean toward the latter (when estimating the value in the world is statistically hard) but folks are surprisingly open to understanding the issues with "ML everything" when demonstrated
There's definitely some of "just do whatever works" in my experience but there are also a ton of people (often PhDs) trying to do things as right as possible under constraints, including experimentation and causal inference
Agreed, I wondered how reallocating regular up to premium is being counted. New planes with 50% premium seems high! Curious to see how premium margins decline (seems like they have to?) as lower WTP customers are forced into it just by prevalence
Awesome find. Silly question -- what does "incremental" here really mean? Net new seats mostly come from new planes/routes (I assume) and I don't see how 85% can be premium. Any idea?
My group at Microsoft is hiring again! This time looking for junior candidates specifically wanting to study advertising. Academic friends (econ or business school especially), please let me know if you have PhD students we should look at!
Listing here: jobs.careers.microsoft.com/global/en/jo...
Oh yeah? Well where else can you get a "taco" with a hot dog in it?