Nof1's "proto scaling law for financial intelligence"
Subject
Nof1 describes itself as an AI research lab building models for financial markets. It previously ran Alpha Arena, in which several large models traded crypto with real money.
The claims (nof1.ai homepage, fetched 2026-10-12)
- "We've identified a proto scaling law for financial intelligence, with signs of skill-transfer to other domains."
- Its models, once released, will "perform well-calibrated research … make forecasts".
- The same page says current LLMs are "shockingly bad" at markets and show minimal improvement between releases.
Three inputs
| Input | Published? |
|---|---|
| Model and training-data cutoff | No (models not yet released) |
| Test window | No |
| Attempts and evaluation protocol | No; "general market intelligence" is described as an internal measure |
Finding
Cannot assess: the claim is public, the evidence is not. A scaling law needs at least model sizes, an evaluation metric, and an out-of-sample window. None is on the page. The former Alpha Arena pages (/alpha-arena, /leaderboard) returned 404 on 2026-10-12, so historical results could not be rechecked.
Worth noting
Nof1's view of current models (poor at markets, little progress between releases, non-stationary environment) matches the premise of this audit series. Treating markets as a benchmark that cannot be memorised is the closed-book idea in its purest form: only forward testing is closed-book. When the models and evaluations are released, the three inputs can be requested.
Recheck: on model release.