How the “Fly Brain” model works
We took the complete wiring map of the fruit-fly brain (FlyWire, version 783) and taught this brain to look at a crypto chart and answer one question: will the price keep rising?
66.5 %% correct among the 10 % most confident
66.5 % — the fly brain with an eye (mean of 5 runs: 66.3 · 65.6 · 68.6 · 66.2 · 65.9). That is above the linear model (52.9 %), gradient boosting (55.2 %) and neural nets (55.0 % from scratch, 64.8 % pre-trained), and above the threshold of 65.8 % set in advance.
1. What the model is
- 138,639 neurons and 2.7 million connections — the real fly-brain wiring, unchanged.
- Only each neuron’s “volume” and threshold and a small read-out at the output are learned. The wiring stays the fly’s.
- A signal passes through the brain in 6 steps; at each step neurons sum their inputs and “fire”.
2. How the fly “sees” the market
1Candle chart of the last hours
2A 512-point picture
3Eye photoreceptors (L1–L3, R7, R8)
4Optic lobes and central brain · 6 steps
596 mushroom-body outputs + 93 visual neuron groups
6Score: “top floor” or further growth
- The right eye looks at the coin, the left eye at bitcoin. The “sense of smell” receives a few market indicators.
- Much like a real fly deciding where to fly from the picture in front of its eyes.
3. The exam
- The moment: a coin’s price is at its 24-hour high. The question: is this the “top floor” (over the next 24 h the price will not rise more than another 2 %) or will the rise continue?
- The measure: the share of correct “top floors” among the 10 % of events the model is most confident about. Random choice gives 43.6 %.
- Training: pre-training on 137,051 events of public candles (before April 2025) → fine-tuning on 2,708 events → exam on 19,392 events, 79 coins, April 2025 – October 2026. Rules and the winning threshold were written down in advance.
4. Result: the fly brain outperformed conventional models
% correct among the 10 % most confident
- The fly’s wiring matters: the same brain with shuffled connections gives 59.5 %; the real wiring is better in all 5 runs out of 5.
- The path: without an eye the brain only caught up with a linear model (51–53 %). The breakthrough came from vision (the chart picture) and pre-training on a large public corpus.
5. Trading: what history and demo-account tests showed
The exam tests “vision”: whether the brain recognises the moment. Trading is a separate test. Trades are counted with exchange fees, the rules are written before the calculation, and the verdict is counted on the last stretch the model never saw.
| Test | Where and how many | The fly | For comparison | Verdict |
|---|---|---|---|---|
| Short near the 24-hour high; exit at a 2 % take or after 24 h | Binance, 79 coins, May – October 2026, 592 trades | −0.21 % per trade | all such moments −0.45 %; shuffled wiring −0.30 % | 🔴 the fly cuts the loss but does not turn it positive |
| Same entries + position management by the copy-trading leaders’ “formula”: averaging in, exit from the average price, emergency stop | same 592 trades | −0.37 % per trade, 2.6× deeper drawdown | without the formula −0.21 % | 🔴 the formula makes it worse |
| Which copy-trading leader to copy now: the fly picks a third of entries per day | Binance, 72 leaders, 29.09–17.10.2026, 4,612 entries (copied 4 min late) | +0.055 % per entry | copy all −0.013 %; boosting −0.021 %; shuffled −0.003 %; “best by past” −0.110 % | 🟨 better than every comparison, but within chance; re-checked weekly |
| Demo account | Bybit Demo, 18 coins, since 11.10.2026 | running | — | verdict once enough trades accumulate |
- Summary as of 18.10.2026: the fly brain recognises the moment on a chart better than ordinary models, but there is no money edge after fees yet.
- The model trades no real money — only history and a demo account. Every test is pre-registered; we publish negative results just like positive ones.
Honestly about the limits
- An ensemble of 5 pre-trained convolutional nets comes close to the brain (66.2 %) — the margin is small.
- The final word belongs to the check on new data: it runs from 15 October to December 2026, with the models frozen.
- This is research on historical data. Not trading signals and not investment advice; no real money involved. Market data — Binance.
Updated 18.10.2026 · FlyWire 783