Lalgorithms

2025-05-09

4 min read

papers

Demand estimation bumper crop:

  • Chambers et al : an alternative to logit demand that parsimoniously parametrizes demand as a combination of softmax (the logit piece) and a salience function. Randomness arises in this framework due to variation in salience rather than T1EV utility shocks, which strikes me as a much more intuitive in a world where recommender systems and advertising are huge parts of the economy and afaict nobody has made a viable business out of handing out gumbel shocks at safeway. Since it is in the AER and written by theorists, the article is a bit of a slog and I haven’t quite grokked the identification conditions yet, but this promises to be a very useful framework.
  • Ursu, Seiler, Honka propose a version of the classic Weitzman sequential search model that can be taken to data. Pairs well with the Chambers et al paper.
  • Ma proposes a simple model of recommender systems with heterogeneous users that enter and exit. Recsys is motivated by the result that optimizing recommendation quality aligns with user retention, but this breaks down in the presence of unobserved user heterogeneity (confounding / ‘frailty’). In this framework, A/B tests that measure short-term recommendation quality and retention may not map to long term effects due to (1) steady state convergence being slow, and (2) the mix of effects on the intensive margin (utility for users who remain in the system), and the extensive margin (users who churn). This needs an economics rewrite.
  • Structural Workshop Papers from a 2011 issue of marketing science are useful didactic resources for structural models with marketing applications [which is basically standard microeconomics afaict; convinced that marketing departments are a tricksy way to hire more IO people into business schools]. Seems to have been a one-off, which is a pity.

Other papers:

  • Policy Learning with Competing Agents: policy learning in a capacity-coonstrained environment with myopic competing agents converges to a nice quantity that admits to a policy gradient estimator
  • Dynamax for state-space modelling: state space in jax.
  • Bernardi and Farne is a thorough study of the Low-rank + sparsity setting in covariance matrix estimation (L + S where L is low rank and S is sparse), which has applications to panel data, time series modelling, recsys, etc.
  • Faridini proposes an interesting method to characterize statistical power under counterfactual sample size choices (e.g. if sample sizes were doubled) and finds that most economics experiments are quite well powered
  • Michal Kolesar’s econometrics notes are extremely thorough and lucid, and recently got an update with DML material
    • github handles pdfs terribly; these can be rendered in readable form with the following bookmarklet I wrote a while ago to replace github with nbviewer.
javascript: (function () {  var currentURL = window.location.href;  var newURL = currentURL.replace(    "https://github.com/",    "https://nbviewer.org/github/",  );  window.location.href = newURL;})();

create a bookmark (on chrome: Bookmark manager > Add bookmark) and paste this whole chunk of code into the url field, and put it in your bookmarks bar. Open one of Kolesar’s pdfs, then click on this bookmarklet, and voila, a full pdf that doesn’t have garbage UI elements around it. You can use a similar trick to open ipynb files in colab [exercise left to the reader], or redirect urls in your intranet.

code, music, articles

self-promotion

  • Econometrics at Emory was excellent; do not take redeyes to conferences that you’re slated to present at.
  • Today is my last day at Netflix. I’ll be starting at the Central Economics and Science team at AWS in a couple of weeks; excited for a new set of challenges to work on with some very smart people.

Graph View