shameless self-promotion
-
My first foray into non-linklog posts on this new site: Implementing Economic Models in Pytorch is a tutorial post that walks through the use of torch, which is the most popular deep learning framework (and hence an excellent numerical programming environment that does linear algebra and optimization well), for estimating economic models that are defined as moment conditions.
- Related: Pytorch is coke, JAX is pepsi. I prefer JAX for its simplicity, but it is a losing battle to use it in production unless you work at deepmind. This repo has some snippets of standard econometrics methods in the JAX ecosystem.
-
I’ll be presenting some work at the Econometrics at Emory conference at the end of the month.
-
ollama_runner.py to run locally hosted LLMs (defaults to ‘big’ gemma3, which runs well on mac mini with the M4 chip). Also tried the much vaunted new deepcoder model, which claims to be competitive with O3 on maths/programming problems despite being miniscule in comparison. I asked it a basic econometrics question, and it proceeded to have a nervouse breakdown.
-
toy example of the robustness gains from solving m-estimation problems with SGD. The DGP is a simple Huber-style contamination setting where a small number of outliers throw off closed-form OLS and 2SLS. In contrast, the subsampling procedure used to update parameters using SGD seems to yield lower RMSE solutions.
-
llm-ocr-tex provides a blueprint for typesetting old pre-latex books with multi-modal LLMs (defaulted to gemini because it is good at this sort of thing, but happy to be corrected). The use-case is Pfanzagl (1982), which is a classic semiparametric statistics textbook that presages a lot of the modern ‘double-ml’ literature.
-
lemmatree is a useful little webapp I got gemini to write (and then fixed up a bit) that allows you to visualize the dependency structure of key mathematical results in an arxiv paper. Upload tex, send it to a good LLM, get mermaidjs output back.
links
-
Bai et al is a nice overview of modern methods for experimental design and analysis. Notable in its coverage of stratification, which is commonly taught but seldom implemented in large-scale A/B testing platforms due to a combination of engineering challenges and folk wisdom that regression adjustment is as good as stratification, which is not true.
-
Chen is a very nice short paper that establishes an exact equivalence between the ‘shocks-based’ view common in demand-modelling (and industrial organization more broadly) and the Neyman-Rubin potential outcomes framework. Optimistically, this paper helps clarify the connections between the ‘reduced-form’ and ‘structural’ views; pessimistically, people will continue to talk past each other.
-
Morris and Shin is a classic review paper on global games, which is very appropriate given the past week in the news. Equilibrium selection in games with multiple equilibria is a classic problem, and we are unfortunate enough to see real-world implications in the financial market, investments in the face of ever-changing trade policy, etc.
-
Good, short post on carcinisation in online entertainment ecosystems. In the marginal-user limit, everything tends to tiktok. Normative implications left as an exercise to the reader.
-
Feynman’s Cargo Cult Science commencement address has been coming to mind thanks to some discussions at work that I’m not at liberties to discuss.