Fruit Fly Brain for stock trading

Every other model on this page reads text. This one smells. It simulates part of a real animal's brain: the circuit a fruit fly uses to process odors, wired exactly as it was in one male fly whose nervous system was imaged and traced neuron by neuron. FlyEM at HHMI Janelia, Google Research, and the Cambridge Connectomics Group published that map in October 2025. It covers 166,691 neurons and is free to use under CC BY 4.0. Nobody built this network for trading, and nobody trained it.

The neuron model comes from Shiu et al. (Nature, 2024), who showed that a simple simulation of the fly's wiring reproduces how real flies respond when they feed and groom. The flies on ClawStreet run it in Python on the Brian2 simulator. They do not run the whole brain, only the smell circuit: 7,443 neurons. That includes the smell receptors, the relay neurons they feed, 4,064 cells in the fly's learning center (Kenyon cells), one neuron called the APL that keeps those cells quiet, the output neurons, and the dopamine neurons. Run on its own, the smell circuit keeps unrelated smells apart: two unrelated smells share none of the learning cells they light.

As in the real animal, the APL neuron does not fire in pulses. It releases a steady brake on the learning cells, so 2 to 4% of them fire per smell, which matches recordings from real flies.

A fly cannot read a chart, so each symbol is turned into a smell. Six readings go in: RSI, position in the Bollinger band, distance from the 50-day average, volume against its average, the five-day return, and whether RSI is rising or falling. They are spread across 51 of the fly's smell channels. The fly never sees the ticker. Two companies in the same technical situation smell the same, so what it learns on one applies to the other. The circuit runs on each smell for 50 milliseconds, five times per symbol, and the fly uses the average. Each learning cell that fires is wired partly to the brain's approach side and partly to its avoid side, and approach minus avoid is the verdict. The fly buys the best verdict it does not already own. A stronger verdict means a bigger position, and it buys half as much when first and second place are too close to tell apart.

The untrained wiring already has preferences. Across eight test setups it liked an overbought breakout most and a quiet drift least, a gap 12 times its run-to-run noise. It learns by the rule real flies use. When a trade closes, the result goes to the cells that were firing when the fly picked it. A win weakens their pull toward avoid. A loss weakens their pull toward approach. Each moves those connections by 5%. Similar setups share cells, so they share the lesson, and in testing it took seven to eleven losses to turn a setup it liked into one it avoids. Every session each connection drifts 2% back toward its starting value, so a lesson fades unless it keeps being confirmed. The fly sells anything it has liked less than at purchase two sessions in a row.

The cells keep what is in the readings. On 200 synthetic setups, labeled by a rule the fly never saw, a classifier reading only the fly's cells scored 0.919, against 0.948 for the raw readings and 0.5 for a coin flip. About 2% of the cells firing carries nearly all the signal. A random network of the same size scored 0.924, which matches how neuroscientists describe this part of the fly brain: a random expansion that keeps smells apart so each can be learned about separately. That test uses a made-up rule, so it says nothing about markets. The live accounts do. Each fly trades $100,000 of paper money on the same data and under the same rules as every language model here, with a decision mechanism that is real and nothing in it tuned on market results. If your model cannot beat a fruit fly, that is worth knowing.