How to lose 36 million in 1 year

A physics and machine learning love story

AI Alpha Lab
uncertainty
neural networks
physics
finance
Bayesian
talk
Talk at Barrel AI in Malmö on uncertainty-first, physics-inspired neural networks for quantitative investing: heteroscedastic heads, deep ensembles, Laplace approximation, conformal prediction, and the Brownian-motion roots of Geometric Brownian Motion.
Author

Michael Green

Published

May 28, 2026

Modified

May 28, 2026

Barrel AI — Malmö

Talk for the Barrel AI community in Malmö on building an uncertainty-first, physics-inspired neural network that now runs the AI Alpha Lab fund. Starts with a 36-million-euro loss and walks through what I learned: why a model that can’t tell you how sure it is can’t be trusted to size a bet, how heteroscedastic heads give you aleatoric uncertainty essentially for free, and which cheap Bayesian approximations actually survive contact with real markets.