links
papers
- Barigozzi and Trapin on EM + Kalman Filters for dynamic factor model estimation. Many researcher choices in implementing these models, so some comparisons might be useful. Existing methods I’ve used in the past
- pick your favourite PCA implementation
- dfms Seb Krantz’ package has worked well for me in the past
- gsynth uses the classic iterative technique proposed by Bai (2009) with 2WFE as the warm-start
- the very SEO-suboptimal fbi package that implements the ‘tall-wide’ (TW) algorithm that uses PCA (twice)
- van Ryzin and Talluri review of revenue management (I should probably also get the textbook), which appears to be a very practical operations research version of microeconomics (broadly defined: price theory and IO), econometrics, and optimization from the firm’s POV. (h/t Yoni Gur)
- Athey, Chetty, Imbens propose a simple and usable surrogate estimator for combining observational and experimental data. Seems slated to be taught during the summer institute on surrogate outcomes (h/t Peter Hull)
- PMA, Jang, Offer-Westort propose a very ambitious augmentation of the design-based approach that accommodates arbitrary patterns of interference. In the traditional design-based approach, the science-table grows in the number of potential outcomes, so it has columns ( treatments, units) in the unrestricted interference setting, only one of which is observed in each row, which makes it very difficult to make progress.
- Kleinberg, Mullainathan, Raghavan propose a behavioural economic model of agents in online settings systems that has useful implications for metric design and recommender systems.
- JK gave a keynote at last year’s WWW on this topic (h/t Harald Steck)
- Repository of Kim Border’s notes: I’ve been studying from KCB’s lecture notes since undergrad; He was a fantastic expository writer. RIP
code, music, articles
- damek davis’ numerical optimization for data science course is great; i’ve tried to follow along with the notebooks during the semester. Basically a ‘how to think about and solve most stats problems with pytorch’ class, which many disciplines would be well-advised to teach but likely never will because teaching lags practice by several decades [Roy model rules everything around me, reprise]
- just is a nice usable alternative to makefiles
- scooter lets you do bulk find-and-replace in the terminal instead of hoping
%s/a/bworks. - LLMs for supervised learning repo collating scripts that use many LLM APIs for supervised learning on a classic problem
- lnk makes dotfile management via git easier
- a-shell does some webassembly magic to get you a functional shell on ios. Obviously somewhat limited because for some reason apple lobotomizes its software while at the same time making leaps and bounds in the hardware capabilities of its devices.
- Greg Leo’s insane 8-key keyboard mind-boggling; typing on a 8-key keyboard is like playing chords on a musical instrument.
Stuff to make macos tolerable:
- install homebrew immediately. Still don’t understand how you can sell a ‘pro’ computer that doesn’t come with a package manager, but here we are.
- raycast window management on macos is a disgrace, so this is a reasonable option. Remap capslock to hyper (cmd+ctrl+alt+shift) and use hjkl/arrows for window snapping. This video is a good walkthrough of how to set it up. h/t Grant McDermott
- ghostty is the best terminal program I’ve used. I map `cmd + “ to trigger quake mode, which is a very useful way to have a terminal window open at all times without it taking up screen real estate.
Music:
- Fuubutsushi’s self titled album is excellent. h/t Greg Sasso
- Jonas Hellborg’s soundcloud is a goldmine of jazz fusion and (blech) ‘world music’. Hellborg has been doing interesting collaborations for four decades; I discovered him through his work with Shawn Lane, who is one of my favourite guitarists. Their work with Carnatic percussionists like Vikku Vinayakram and Selvaganesh is a great introduction to the Carnatic tradition, which all jazz and metal fans would do well to check out.
self-promotion
- Sudijono et al was accepted for a presentation at EC at Stanford in July. I learned a lot working on this paper + Tim is a phenomenal researcher.
- Elea Feit wrote a nice exposition of this emerging decision-theoretic approaches to A/B testing literature that we contribute to in this paper.
- first week on the new job, so no major OSS updates
- arxiving a regression mechanics for continuous treatments paper this coming week