Ben Wellington – Complex Feature Engineering at Two Sigma (S7E32)
Summary
Corey Hofstein interviews Ben Wellington of 2Sigma on the critical role of feature generation in quantitative investing. They discuss how large language models are democratizing data creation, allowing researchers to rapidly prototype novel signals like CEO micro-expressions. The conversation highlights the tension between AI’s efficiency and the need for orthogonal, diverse alpha, warning that over-automation risks homogenizing strategies.
Topics discussed
Intro, Return Stacking Symposium, and podcast overview
Ben's background: NLP, 2 Sigma, and feature generation
Defining features: From price changes to news sentiment
Pod shop vs. shared platform models at 2 Sigma
Where alpha lives: Data, features, and forecasting
Human priors, creativity, and avoiding noise in features
Collinearity, orthogonality, and signal decay
Hiring for ingenuity and lateral thinking skills
LLMs as data generators and research accelerators
Democratization of tools and the risk of convergence
Timing technology adoption and infrastructure choices
Model control, hallucinations, and preventing overfitting
Data upgrades, diminishing returns, and knowing when to ship
Scaling idiosyncratic insights and the changing role of coders
AI as an amplifier and Ben's obsession with city data
Listen ad-free on Castria