A transformer trained on stock returns beats every factor we know.

This new paper introduces StockGPT, a GPT-style model trained directly on 70M U.S. stock returns (no text, no fundamentals, just price). The result?

• 119% annual return, Sharpe ratio of 6.5 (daily L/S portfolio),

• Still earns 69% after trading costs,

• Outperforms momentum, reversals, and all 11 major stock factors,

• Orthogonal to known signals.

A decoder-only transformer, trained once on data up to 2000, works 23 years later. No retraining needed.

Note: The paper doesn’t address short availability, which is a real challenge in small caps. However, even the long-only leg shows strong performance, making this paper a must-read.