beedictor Can a fruit fly's brain read a price chart?

We take two real, fully-mapped Drosophila brains — one male, one female — show them both the exact same moving price chart, and let each one independently guess where the line goes next. Then we check what the market actually did. Neither brain is trained, rewarded, or told anything about markets.

What is actually going on here

STEP 1

A real brain, wired as measured

Scientists sliced two fruit fly brains, imaged every neuron with an electron microscope and mapped which cell connects to which. We load those wiring diagrams — 134,013 neurons in the female, 163,903 in the male — and run them as living networks. The wiring is never altered.

STEP 2

The chart becomes something to look at

We draw the last 120 candles — about ten hours — as a plain white line on black. No numbers, no grid, no indicators. White because it is what a fly sees most strongly: measured against published pigment curves it drives the motion pathway three times harder than green. Then we scroll it across the fly's field of view, because a fly's visual system responds to movement, not to still pictures.

STEP 3

The eye actually sees it

The image lands on each brain's real light-sensing cells, positioned where they physically sit in that animal's eye. Both brains see the identical picture — we hash every frame to prove it.

STEP 4

The fly's motion detectors vote

Flies have dedicated neurons for "something moved up" (T4c, T5c) and "something moved down" (T4d, T5d) — that is genuinely their job in the animal. We simply listen to which group shouts louder, and move a pointer up or down accordingly. Nothing is trained; we just read what the wiring does.

STEP 5

Lock the answer in before looking

Both guesses are written to disk before the next candle is ever loaded. The code that simulates the brains is physically never given future prices. We prove this with a test that feeds it garbage future data and checks the output doesn't change by a single bit.

STEP 6

Only then, reveal and count

Now we look at what really happened and tally it up. After thousands of candles we can ask: when male said UP and female said DOWN, what followed? Nine combinations, counted honestly.

Why two brains instead of one?

Because one brain on its own gives you a number with nothing to compare it to. Two different wiring diagrams seeing the same thing lets us ask whether their disagreement means anything — and, more importantly, lets us run the same test on deliberately broken versions to see whether the real biology is doing any work at all.

What would make this interesting — and what would make it nothing?

The market itself is roughly a coin flip: about 47% of candles go up, 49% go down. If some brain-state combination is followed by an up-candle 52% of the time, that is noise wearing a suit. It only becomes interesting if it holds up well above the baseline, on data we never touched while tuning, on more than one coin, and beats scrambled versions of the same brains.

Is this a trading system?

No. It places no orders and is connected to no exchange account. It is an experiment that produces statistics, and the most likely honest outcome is that the fly brains tell us nothing about crypto. That result would be written up exactly as plainly as a positive one.

The part where we nearly fooled ourselves
The first working version produced beautiful, confident predictions. The male brain said DOWN on 99.5% of candles and the female said UP on 90% — and the tables looked meaningful.

They were meaningless. The "up" and "down" neuron groups simply sit in different amounts of wiring, so one group naturally chatters more than the other. The pointer was drifting on anatomy, not on anything to do with the chart. Three of the nine possible states were literally unreachable.

The fix was to measure each brain's own resting chatter and subtract it, so the pointer only responds to changes from that brain's normal. All nine states are now reachable. This is worth spelling out because the broken version did not look broken — it looked like a result.
Why the colour of the line is an experiment, not decoration
Fruit flies cannot really see red — they have no red-sensitive opsin. Using published pigment sensitivity curves, we can calculate before running anything that a red line should stimulate the motion pathway at about 0.04% of a green one. So re-running everything in red is a built-in sanity check: if red performs like green, our pipeline is responding to something other than the image, and we have a bug rather than a discovery.

The result

The experiment is finished. Three markets, two senses, three network variants and two operating points — 361,214 predictions across 14 runs, every one frozen to disk before the next candle was read.

Do the real fly connectomes predict crypto price movement?

No. And the controls did better than the real brains, in all three markets.

marketreal connectomesscrambled wiringrandom network
BASEDUSDT0.0130.0230.023
BEATUSDT0.0160.0240.024
LITUSDT0.0140.0240.024

Strongest link to future price after removing what the bee is simply copying from the chart in front of it. Higher is better. A deliberately broken brain beat the real one every time — and even the best of these numbers is far too small to trade on.

A result nobody expected

The random networks read the chart better than the real brains did. Tracking price position: real connectomes 0.88–0.91, randomly wired networks of the same size 0.97–0.98.

That makes sense once stated. A random network is close to a plain wire — it passes the picture through and little else. A real brain is full of loops, feedback and specialised circuitry that exist to serve a living fly, and all of that adds dynamics of its own. For the narrow job of reporting where a line currently sits, the biology is overhead. It is doing more, and that extra is not about markets.

And every one of them follows

Across all nine runs the bee's strongest agreement with price is at a lag of +1 candle — after the move. Not one leads, at any lag from −12 to +12. Real, scrambled and random alike.

This is a negative result, and it is reported as one. It was worth running: the question was answerable, it was answered blind, the same pipeline was applied to deliberately broken brains, and the answer replicated across three markets, two senses and two operating points. Every route the anatomy offered was measured rather than assumed. An experiment that could only have come out positive would not have been worth doing.

What it would have done with money

A correlation of 0.013 is hard to feel. Here is the same thing as a trade: take the strongest signal the bees produce, go long when the pointer sits high in its window and short when it sits low, across all 90 days.

The test is deliberately rigged in the bees' favour — the signal was picked by looking at the whole period, which is cheating, and there is no risk management to get in the way.

result
profitable runs0 out of 11
before any fees−192% average
after fees and slippage−1,720% average
win rate36–48%

It loses before costs are even applied. That matters more than the headline number: a signal that bleeds gross cannot be rescued by trading it more cheaply or less often. The enormous cost figure is partly an artefact of a naive rule that trades on 40% of all candles — but the underlying edge is negative regardless.

One market, BEATUSDT, does show a positive edge before costs (+732%). Its scrambled control shows +761% and its random control +676%. The edge belongs to that market, not to the brain reading it.

This is a falsification exercise, not a strategy, and nothing here is connected to an exchange account. Its purpose is to convert a small correlation into the form where "too small to matter" stops being arguable.

What the biology actually does

The headline answer is negative — the fly brains do not predict price. But the same 232,209 observations contain a clear positive finding, and it is the most robust thing this project has produced.

Compare a real connectome against the same network with its wiring scrambled, and against a randomly wired network of identical size:

real connectomescrambledrandom
share of the network firing30.0%86.9%92.8%
overall activity8.4%28.8%30.8%
signals passed per observation19.4M57.7M61.5M
how long it holds a direction5.9 candles3.12.5

Break a brain's wiring and it floods. Nearly the whole random network fires at once, almost four times the activity of the real thing, and whatever it is doing it forgets within two or three candles. The real connectome engages only a third of itself, and holds a state more than twice as long. Same neurons, same number of connections, same input — the only difference is how the wiring is arranged.

That is a real, measured property of biological structure, replicated across three markets. It also explains something that looked strange earlier: the random networks read the chart better (0.97 against 0.88) precisely because they flood. They pass the picture straight through. A real brain filters and gates, which costs accuracy on a task this narrow — and is presumably the thing that makes it a brain instead of a wire.

It still is not a market signal. The real brain's longer memory is just as long in rising markets as in falling ones (p between 0.07 and 0.96). The biology does something specific and measurable. It simply is not about price.
Why 25,801 observations are not 25,801 observations
Consecutive readings are so similar that the pointer's position carries only about 400 independent observations' worth of information across a whole 90-day run — between 1.2% and 1.9% of the nominal count. Its per-candle step fares better at 15–28%. This is a hard ceiling on what any amount of this kind of data can resolve, and it is why every test here is run against a null that preserves that structure rather than assuming independent samples.

The loose end, pulled

One objection remained. Every result above came from a setting where the network barely reacted — its response varied by 2% while the market varied by thousands. That turned out to be our threshold, not the fly's: set it higher and the same brain becomes far livelier. So the whole 90-day run was repeated at the livelier setting, with a random-network control alongside.

networksettinghow much it reactsreads the chartstrongest link to future price
realoriginal0.020+0.9130.014
reallivelier0.135 — 6.7× more+0.9400.009 — nothing significant
randomoriginal0.006+0.9810.023
randomlivelier0.005+0.9750.022

Waking the network up did not help — it made things marginally worse. Six times more responsive, reading the chart slightly better than before, and less connected to what price did next, with nothing reaching significance at all. The answer does not depend on the setting.

Two things worth noticing in that table

The livelier setting un-clamps the real brain but leaves the random one flat (0.006 → 0.005). Scrambled wiring floods no matter where you put the threshold; it has too many paths. Only the biologically arranged network has a setting at which its firing rate is free to say anything at all.

And the real brain tracks price simultaneously, while every broken version lags a candle behind. The biology is faster. It is simply not early.

That closes the question. The result now stands across three markets, two kinds of deliberately broken brain, and two operating points — 283,811 predictions, each frozen before its candle was revealed. The fly connectomes do not forecast crypto, and the scrambled versions of them do no worse.

What the bees are actually doing — the mechanism

Four hypotheses were tested against 8,609 observations and a control brain. Three were falsified, and the way they failed explains the fourth.

hypothesisresult
The bee reads the chartconfirmed — ρ = +0.88
It detects turning pointsfalsified — activity at a turn 0.08365 vs 0.08362 inside a trend
It detects volatilityfalsified — sign flips to −0.06 once past volatility is controlled
It predicts directionfalsified — follows at lag +1, never leads

The finding underneath all three failures

The network's firing rate barely moves. Across 8,609 observations the market input varies by 6450% of its own mean; the network's response varies by 2.0%. It compresses the stimulus roughly 3000-fold.

That is not a defect — it is what the circuit is built to do. A fly's visual system is an optomotor reflex: its job is to hold the animal steady against drifting scenery, which means reporting current motion faithfully and regulating its own gain so it works in bright sun and at dusk alike. Both properties are precisely wrong for forecasting. Reporting current motion faithfully is following. Regulating gain is discarding magnitude.

So we built an excellent instrument for measuring what the chart is doing now, and asked it what the chart will do next. It answers the first question at ρ = 0.88 and is structurally silent on the second.

The one real difference the biology does make
Real connectome wiring produces noticeably more persistent behaviour than the control: the cursor holds a direction for 6.4 candles on average versus 4.7 for a brain split in half. That is a genuine topology effect. But the persistence is the same in rising and falling markets (p = 0.18), so it carries no directional information — the real wiring changes how the bee moves, not what it knows.
Why the numbers are weaker than they look
Readings from consecutive candles are heavily correlated, so 8,609 observations do not carry 8,609 observations' worth of information. For network participation the effective count is 448 — about 5%. Every p-value here is computed against a null that preserves that structure, because a test assuming independent samples would call noise significant.
What would actually be required to predict
Prediction needs a memory that binds a pattern to an outcome. In this brain that is the mushroom body — and the visual pathway reaches only 7% of its input cells, against 100% for the sense of smell. It would also need the rate code to be allowed to vary, since homeostasis is actively throwing away the magnitude information. Both are architectural changes, not parameter tuning.

Things we got wrong, and how we found out

An experiment like this fails silently. The dangerous outcome is not an error message — it is a confident, plausible-looking result produced by a bug. Three have been caught so far, and each was invisible until it was specifically tested for. They are listed here because a method section that only records successes is not evidence of anything.

1. The brains were predicting their own anatomy, not the chart
The first working version had the male brain saying DOWN on 99.5% of candles and the female saying UP on 90%. The tables looked meaningful.

They were not. The "up" and "down" motion-detector groups sit in different amounts of wiring, so one naturally fires more than the other at rest. The pointer was integrating that constant difference — a fixed property of the wiring diagram — and only three of the nine possible states were reachable at all.

Fix: each brain now measures its own resting rate and subtracts it, so the pointer responds only to changes from that brain's normal. All nine states are now reachable, and the male average moved from −0.671 to −0.088.
2. The cursor was being clipped 65% of the time
The pointer has a maximum travel. At the gain we started with, the male cursor hit that limit on 65% of observations — meaning two-thirds of its answers were the same saturated value regardless of what it saw. Reducing the gain removed the clipping entirely while keeping the spread of responses between the two brains well matched.
3. Our own statistical test was too permissive
Readings from consecutive candles are not independent — the brain carries state forward, so one observation resembles the next. Measured autocorrelation is +0.62, which means our 8,609 observations carry only about 2,000 observations' worth of independent information.

We wrote a test designed to handle that, then checked it on synthetic data where the right answer was known. It was wrong: it flagged false results 9.5% of the time against a target of 5%. The method chopped the data into blocks, which destroyed the very autocorrelation it was supposed to preserve. A different approach now measures 4.0%, and still detects a real effect 100% of the time.

We also checked a claim we had already made — that our original method was too permissive. It was not. It was slightly conservative. The real distortion was somewhere else entirely: in correlations between two drifting series, where the false-positive rate reaches 15.5% instead of 5%. That is now handled separately.

Every one of these would have produced publishable-looking output. None announced itself.

The odour trail experiment

The visual bee here turned out to be an excellent follower — it reads the chart at a correlation of 0.88 but always a step behind. Tracing the wiring showed why: the motion detectors driving its cursor have zero connections to the mushroom body, the one part of a fly's brain that can learn. That pathway is a reflex, and a reflex cannot be taught to predict.

The route into the learning circuit is the nose — roughly 24× more trainable synapses than vision has. A separate experiment on that runs at its own address:

Result: it did not help. Two markets, the same blind protocol, the only change being which sense the market was painted onto:

senselink to future pricesignificant results
smell0.0112
sight0.0133
deliberately broken brains0.02324

Smell had every structural advantage — it reaches every one of the brain's 5,177 memory cells where sight reaches 339, drives them a third harder, and carries twice the variation. None of it converted. The broken brains still score higher than either sense.

A caution about reading single results: on the first market alone, smell looked slightly better than sight. The second market reversed it. One market is not a trend, which is exactly why the experiment runs on more than one.

This project's own olfactory arm is separate from the OdourTrail experiment →, which runs independently.

Live experiment status

connecting…

Watch a single observation — live

loading…

The bar labelled male moved / female moved is the per-candle step — the quantity that turned out to carry almost nothing (−0.03). The position of each coloured trace in the right-hand panel is the one that tracks price at +0.88.

what both brains see male's guess female's guess what actually happened
joint state — retained for continuity, but this is the step-based label the analysis retired. Watch the positions above instead.
what the next candle actually did

A table that used to be here

This page used to end with a nine-way tally: male said UP, female said DOWN, and what the market did next. It has been removed, because the analysis showed it was counting the wrong thing.

The bee's pointer has a position and a step. The position tracks price closely — correlation +0.88. The step, from one candle to the next, tracks it at −0.03. The tally was built on the step.

That is why every one of its nine combinations came out at almost exactly one ninth, whether the market rose or fell. It was not a discovery that the states carry no information — they were independent of the market by construction, and the table was measuring its own dead channel. Leaving it up would be showing you a result we know is empty.

What replaced it is the section at the top of this page: the same question asked of the pointer's position instead, across three markets, against deliberately broken brains.