Roxx
Evolving traders instead of training predictors, graded on what the market actually did.
- Status
- Concluded
- Started
- 2026-08
Written by Hollis from the project's own notes.
What it is
Roxx followed directly from Emergence. Instead of scoring animals with a fitness function written by hand, it scored them on the market itself. Profit is what actually happened. It can't be rescaled, and it doesn't care what anyone meant to reward. That turns a design question (what should we reward?) into a fidelity question (does the simulator honestly report what the market did?), which can be checked.
It also had a measured reason to use evolution rather than gradient descent. A market offers no label for the correct trade on a given day, and profit through discrete fills isn't differentiable. Evolution needs only a ranking, and a simulator provides one.
How it works
A herd of networks traded simulated portfolios across a set of rotating market windows, with further windows held back to test out of sample. Each generation the weakest were culled and the survivors bred with mutation. The final configuration ran 48 animals over 26 rotating windows and six held-out windows. That run lasted about 60 hours on the lab's GPU.
Across the notebook entries the selection method itself was the main subject of study. The project moved from a round-based cull to a cull ranked by an Elo rating, and produced the first clean measurement of whether the herd's picks carried out-of-sample signal: +0.113, positive in every block of the final run. The numerics stayed clean for all 726 generations of that run, and a pre-registered check on trading activity held at 47 to 49 trades throughout.
Built with
- PyTorch networks of about
791,000parameters, benchmarked at the start against a throughput gate. - A portfolio simulator ported from the lab's earlier financial-data design, with transaction costs and a null baseline to beat.
- Rotating and held-out windows, so no fixed test could be memorised.
- Checkpointed runs with a structured log of every generation, and a test suite for the simulator and the cull.
- An append-only lab report, one entry per session, numbers verified from the logs.
Where it stands
Concluded on Jason's decision on 2026-08-12, when the lab set aside the goal of a profitable algorithm to focus entirely on evolving intelligence. His closing note gives the scientific reason:
Without an antagonist to intelligently remove the weaker members of the group, ideal candidates and unworthy ones contributed nearly equal shares to the gene pool.
A market doesn't adapt to the animals trading in it, so it can't supply that antagonist. The work moved the same day to Throng, where the other animals do.
Visit Roxx on the research page ↗
← The whole portfolio