Emergence (Arbie)
A stock-pick network bred rather than trained: whole trained networks crossed and mutated across generations.
- Status
- Concluded
- Started
- 2026-04
Written by Hollis from the project's own notes.
What it is
A distributed evolutionary system that breeds trained neural networks to pick equities. The network predicts which stocks will make a short swing up. Evolution then works on its learned weights, treating whole neurons as genetic traits, rather than on its architecture as most of the literature does. The first commit was on 2026-04-26 and the last run on 2026-05-31.
This card is the short version. The long one is chapter one.
How it works
The predictor was trained once, offline, on 5.7 million rows of daily equity data, behind a hard date wall and with its scaler fit on training dates only. From then on its weights changed only by breeding: fitness-weighted crossover plus small random mutation, in four isolated bloodlines crossed round-robin, selecting on pool averages rather than outliers. Fitness was simulated performance across several market eras, with difficult eras weighted double.
The predictor has real, modest skill: validation precision 40.58% against a 17.21% base rate, about a 2.4× lift.
Built with
- PyTorch, an LSTM predictor of about
5.8million parameters over a 20-day window of technical and market-context features. - PostgreSQL as the single source of truth: the population, its lineage and a task queue.
- A coordinator and worker daemons, so simulations ran on several of the lab's machines at once.
- A lineage engine with ageing, breeding, a sliding survival baseline, and "climate shifts" that reset every animal's fitness when the market era changed.
Where it stands
Concluded in August 2026 after a four-part review of how its backtests were built. The evolutionary machinery itself was confirmed working, and the review's lessons on measurement shaped everything after it. The lab's working rule, evolution is exactly as intelligent as the test you give it, comes from here, and so does Throng, whose worlds have no score to game. The public service built around this model family is described on the Stocko page.
Visit The full story on the research page ↗
← The whole portfolio