My new paper asks, "theorizing after looking at the data is bad, right?"

Actually, post-hoc theorizing might be OPTIMAL in the modern era of huge datasets and mature theories. 🧵 There's a reason for the sacred cow. Post-hoc theorizing leads to overfitted ideas. As a result,
measured quality > actual quality,
for ideas based on post-hoc theory

Not a video, but here's a thread on how to install VScode / Copilot + Latex. (On Windows). Actually, I've jumped from VScode / Copilot to Cursor AI, so let's do that. (Cursor is a VScode fork, so the same steps should work for VScode) 🧵

https://twitter.com/AlexZevelev/status/1775905842602352945

First download and install the Cursor AI IDE. Go to and click Download. (If you already use VScode, Cursor will port over all your settings, easy peasy).cursor.com

🚨My paper "Missing values handling for machine learning portfolios" (w/ Jack McCoy) is accepted @J_Fin_Economics!🚨 No one likes imputing missing data. But in ML you are often forced to impute. We provide guidance for imputation based on three facts about return predictors🧵 Fact 1: Missingness occurs in blocks organized by time. If a stock is missing book/market this month, it's 80% likely to be missing book/market in every previous month. So there's not much time-series info you can use to impute.

That banger 🎸of a paper, Fama and French (1993), doesn't look like much of anything using data post-1993. The CAPM works about as well as their 3-factor model----using FF's original test portfolios! Bananas. H/T *Dino Palazzo* for discovering this. (1/4) FF's size and value factors are not beautiful nor theoretically well-founded. And they fail to explain most anomalies. But they are supposed to provide an improved empirical description of, well, **size and value sorted portfolios**!! I guess they did, until 1993.

New working paper!! The paper is called "High-Throughput Asset Pricing" (w/ @ChukwumaCDim). But in my heart, it will always be called "How I learned to stop worrying and love data mining." ❤️🧵 How can you stop worrying and ❤️ data mining? You can if you do a good job!---if you mine data *rigorously*. We demonstrate by mining 73,108 long-short strategies based on accounting, past returns, and ticker symbol data using empirical Bayes shrinkage.

@TomZ_Econ and I just finished a review of “Publication Bias in Asset Pricing Research.” (arxiv.org/pdf/2209.13623…). Rather than list methods, @JohnHCochrane's reviews inspired us to focus on stylized facts. Here are 4 facts from meta-studies of hundreds of market anomalies 🧵 Fact 1: almost all anomalies can be replicated.

One fav papers, of the few that changed my views, is Amromin and Sharpe 2005. They show that households have exactly the -opposite- view on expected returns and recessions that I was taught in my Ph.D. This paper is also the saddest example of the Matthew effect I know of. 🧵 2/5 I was taught that expected returns are high in recessions. In recessions, risk is high, so returns are high. My JMP (and first pub) is a quantitative GE model of this fundamental idea.

Open source asset pricing is massively updated!! The code is so so much more user friendly now. Anyone w/ WRDS + Stata should be able to successfully replicate basically the entire cross-sectional predictability lit. Please spread the good word!! (1/5)

github.com/OpenSourceAP/C… Wanna see how we get a t-stat of 15 for DivSeason? Just go to the DivSeason.do file in Signals/Code/Predictors/. This file can be run independently of all the other Predictor/*.do files. Easy-peasy.

github.com/OpenSourceAP/C…

(2/5)