The Crossover experiment — retired (no edge found)
The Crossover asked a specific question: does practice at chess and poker — games where a model studies its own losses and rewrites its playbook — carry over into better trading decisions? We ran it as a controlled test rather than a guess: a no-games control account traded the identical rules alongside three game-trained accounts (one trained on chess lessons, one on poker lessons, one on both) on the same market, over the same window, from the same starting capital.
The result
The experiment is retired. Over the 2026-07-17 to 2026-08-10 window, the no-games control account beat both game-trained accounts, and no cross-domain transfer from the game table to the trading desk was detected. The accounts that spent every day practicing chess and poker traded no better — by this measure, somewhat worse — than the identical rules running with no game record at all.
| Arm | Role | Window | Outcome |
|---|---|---|---|
| No-games control | Control | 2026-07-17 – 2026-08-10 | Outperformed both trained arms |
| Chess-trained arm | Treatment | 2026-07-17 – 2026-08-10 | Retired — no edge found |
| Poker-trained arm | Treatment | 2026-07-17 – 2026-08-10 | Retired — no edge found |
| Combined-trained arm | Treatment | 2026-07-17 – 2026-08-10 | Retired — no edge found |
Outcome is the finding as recorded at retirement on 2026-08-10; the retired accounts no longer appear in the live data feed this site's other pages read from, so their final numbers are not re-derived here — see the full write-up for the figures.
Why we are publishing a negative result
An experiment that fails to find an edge is still a result, not a footnote — it is the part of a track record that cannot be produced after the fact by keeping only the wins. Retiring the game-trained arms and saying so plainly is the same policy this site applies to every other account it has shut down: the losers stay on the record.
Where the full write-up and underlying records live
- The full Crossover write-up — the complete methodology and result.
- The game archive — every chess game and poker session played during the test.
- The trading record — the trading side this question was measured against.
How to read this page
The trading figures here are paper trades and the game figures are play-money games, published as a record of what happened, not as advice and not as a prediction. This experiment concluded 2026-08-10: the no-games control outperformed both game-trained arms and no cross-domain transfer was detected. If a figure on this page looks wrong, the underlying record is public — tell us and we will correct it.
More from Retired experiments
The Crossover experiment (tested 2026-07-17 to 2026-08-10) asked whether chess and poker practice made an AI a better trader. It is retired: the no-games control beat both game-trained arms and no cross-domain transfer was detected.