2025-10-20

links

papers

• Coulombe and Klieber on modern balancing-weights interpretation of Local Projection estimators widespread in macroeconometrics. The intepretation of OLS coefficients as linear combinations of the outcome $PY$ is well taught in intro econometrics, but the fact that this implies a counterfactual mean estimator $\omega Y_1$ with weights $\omega =X_1(X_0^{\prime }{X}_0{)}^{-1}{X}_0^{\prime }$ where $X_1$ is the target distribution matrix (say, the treated units’ covariates) and $X_0$ is the source distribution matrix (say, the control units’ covariates) is less well known (Robins et al 2007 and Kline 2011 show that this implies that OLS is ‘doubly-robust’ in a certain sense - for an implementation, see my cbpys package). Coulombe and Klieber show that this insight can be used to straightforwardly represent Impulse Response Functions (IRFs) as

$$\omega =(X_{\tau }^{\delta }(X^{\prime }X{)}^{-1}{X}^{\prime }-X_{\tau }^0(X^{\prime }X{)}^{-1}{X}^{\prime }=\underset{\text{Proximity Differential}}{\underbrace{X_{\tau }^{\delta }{X}^{\prime }-X_{\tau }^0{X}^{\prime }}}\underset{\text{Proximity Denominator}}{\underbrace{(X{X}^{\prime }{)}^{†}}}$$

where $X{X}^{\prime }$ is the Gram Matrix (more commonly seen in the the kernel-ridge regression lit), $†$ is the Moore-Penrose pseudoinverse, $X_{\tau }^{\delta }$ is the lagged covariate matrix with shock $\delta$ at horizon $\tau$, and $X_{\tau }^0$ is the lagged covariate matrix without shock (‘baseline’) at horizon $\tau$; in the simple case where the policy shock is binary, this is exactly the ATT at a given horizon. The first term is a proximity differential, capturing how the shock changes the proximity of the target distribution to the source distribution, while the second term is a proximity denominator. This lets one interpret IRFs as proximity-based imputation estimators in a specific function class, and opens the door to a variety of extensions including non-linear regression methods [they focus on random forests]. Basically, Local Projections are t-learners for proximity-based imputation of counterfactuals in time series settings; update your beliefs about their credibility accordingly.

• Cen et al is an ingenous experimental paper documenting strategic behaviour of users when faced with recommender systems. They build a small custom music player and randomise informing the users about the recsys. They find substantial strategisation by both explicit (likes) and implicit (dwell time) feedback. Heterogeneity by age is particularly interesting - younger (tiktok-brained) users are especially likely to use dwell-time strategically. Seeing as dwell-time is a common implicit feedback signal used in practice, this has worrying implications for both recsys and what songs look like in the wild - no more quiet classical intros if this finding generalises.

• Leudtke on autodiff for efficient semiparametric estimation. Uncharacteristically for a statistics paper, it comes with good, working code

code, music

• arc looks like a very efficient database for panel data; someone should use it to build a knn imputer for panel data.

• globalsearch looks like a nice general constrained nonlinear optimization library with reasonable python bindings.

• beads is a promising looking plugin to get around the epidemic of random partially-complete markdown files generated by coding agents. I’ve used it for an afternoon and it feels a bit too JIRA-like for my taste, but the memory enhancements appear to be solid.

• Spiro Dussias is a ridiculous guitarist. For most of my guitar-playing life, I’ve been able to at least pick up the main riff in guitar-based rock or jazz; post Tosin-Abasi, I’m completely lost but excited to hear genuinely new harmony and rhythm played on guitar [we don’t need yet more zeppelin ripoffs]. Time to retire the shred guitar and play the blues badly on a battered old tele.

• The new Lage documentary is very good.

self-promotion

• The agent is 1 month old and getting alarmingly effective at bargaining for feeds and walks.

• Difference in Differences with Event Data - scrappy WIP research note (written in Quarto, may eventually turn into an arxiv-able paper) on defining causal estimands for counting processes in DiD settings using inhomogeneous poisson processes, with the intention of displacing the current practice of arbitrary binning of event data into discrete time intervals. Comments welcome.

• torchonometrics now has a discrete choice submodule with several static choice models implemented in a hopefully flexible API to allow for easy estimation and counterfactual simulations. This notebook goes over a few examples, including the limitations of traditional MNL (IIA being too strong an assumption) and a matrix-completion approach to estimating mixed logit models that gets around this problem.

  • the library already has GPU-accelerated OLS, MLE, GMM, and GEL implementations; discrete choice is the next step towards building a general workhorse econometrics library in pytorch.
  • feel free to create issues or PRs if you have suggestions for features or find bugs.

• Causal inference in Tech - slide deck for a talk I gave at Jeong-Yoon Lee’s class at UChicago - probably won’t make much sense without my narration, but this slide has a decent summary of what I think are key directions for research in causal inference for tech companies. I was also proud of the Hippo → Taming the Hippo image.

• I’ve been messing around with deploying small personal webapps on bare metal on a VPS using hetzner and now have a household shopping and recipe app that is run via a CRUD fasthtml webpage hosted on a domain I bought on namecheap; this is a commitment device to document this.