The Short Version
AIOKA runs a council of AI agents that votes on Bitcoin around the clock. Each agent reads a different slice of the market: on-chain data, macro, liquidity, technicals, risk. A judge then turns their votes into one verdict, from STRONG_BUY to AVOID.
In October 2026 we asked the question every trading system should be asked: do the verdicts actually predict anything?
We took every council verdict from 1 April to 5 October 2026, 6,504 of them, and compared each one with what Bitcoin's price did over the next 4, 24 and 72 hours. The answer was no. A bullish verdict was followed by the same returns as any random hour.
This article shows how we tested it, what we found, why it happened, and what we changed. We are publishing it because a signal service that only shows its wins is not worth trusting, and that includes ours.
How We Tested It
We used three tests, all on data from our own production database:
Forward returns. For every hour in which the council ruled, we measured Bitcoin's return over the next 4, 24 and 72 hours, using 1-minute prices. We then compared those returns with the return after *any* hour in the same period. If bullish verdicts are worth something, the hours after them should do better than average.
Confidence. Each verdict comes with a confidence score. We checked whether higher confidence meant better outcomes.
A trading test. We simulated actually trading on the verdicts with AIOKA's real exit rules (take-profit levels, trailing stop, cooldown after each trade), one position at a time. Then we ran the same rules 300 times with random entry times. A real signal should beat most of the random runs.
To keep ourselves honest, we put 95% confidence intervals on the averages. Hourly signals overlap (a 72-hour window starting at 10:00 shares most of its hours with one starting at 11:00), so we resampled whole days rather than single hours. We also subtracted trading fees, because a signal that only works before fees does not work.
Finding 1: The Council Was Bullish Almost All the Time
Out of 3,251 hours in which the council ruled, it was bullish in 2,925. That is 90% of the time.
A signal that says "buy" nine hours out of ten cannot tell good hours from bad ones. The numbers confirm it:
Any hour: Bitcoin rose +0.14% on average over the next 24 hours.
After a bullish verdict: +0.13% over the next 24 hours.
Difference: -0.01%, with a 95% confidence interval from -0.26% to +0.27%. In other words, zero.
The strongest verdict did no better. After a STRONG_BUY, Bitcoin fell 1.07% on average over the next 72 hours, while any hour was followed by a 0.45% gain. With only 119 STRONG_BUY hours, that gap could still be chance, but it certainly is not an edge.
Finding 2: Confidence Barely Mattered
If the council knew when it was right, high-confidence verdicts should do better. The link between confidence and the next day's return was very weak (a rank correlation of +0.077, where 0 means no relationship).
The confidence threshold that our trader used to enter trades was 0.68. In the trading test, entering on bullish verdicts above 0.68 landed at the 48th percentile of the 300 random runs. That is exactly what random entries would do.
One slice did stand out. Bullish verdicts with confidence of 0.75 or higher produced 26 trades averaging +0.50% each before fees, better than 99% of the random runs. That looks good until you subtract fees:
Before fees: +0.50% per trade
With low (maker) fees, 0.25% per side: 0.00%
With standard (taker) fees, 0.40% per side: -0.30%
Twenty-six trades is also a small sample. We tested seven different entry signals, and with seven tries, one or two can look good by luck alone.
Finding 3: A One-Line Rule Did Just as Well
For comparison, we tested the simplest rule we could think of: buy when the 1-hour RSI drops below 30 (a standard "oversold" reading). No AI, no on-chain data, no judge.
It produced 31 trades averaging +0.45% before fees, better than 98% of the random runs. That is about the same as the council's best slice. After fees it was also roughly zero.
And over the whole period, simply buying Bitcoin on 1 April and holding it returned +25.7% ($68,263 to $85,786 on 5 October). None of the signals came close to that.
Why It Happened
When we looked at how each agent voted, two causes stood out.
One agent never changed its mind. Our on-chain agent, Chain Oracle, was described to the judge as the most important member of the council. It voted bullish in 97% of its votes. Its main input, the MVRV Z-Score, is a valuation metric that moves over months, not hours. For most of the period it read "undervalued", so the agent said "bullish" every hour. A slow metric makes a stubborn vote.
The judge turned ties into "buy". Across all agents, only about 42% of individual votes were bullish. Yet 90% of the final verdicts were bullish. When the bullish and bearish votes were evenly split, the judge (itself an AI model) still ruled bullish 73% of the time. It also had no way to say "wait": a neutral verdict appeared only when the council was openly divided.
Neither of these is exotic. They are the kind of design flaws any multi-agent system can have, and you only find them if you test the output against the market.
What We Changed
Paper trading gate back to 0.75. We had been testing a lower confidence gate (0.68) in paper trading. The study showed it was no better than random, so we reverted it on 5 October.
Fixed missing inputs. One agent had been missing half of its data and voted HOLD every single time. Its data sources were repaired.
Chain Oracle now reads changes, not levels. Instead of "MVRV is low", it looks at how on-chain metrics moved over the last 7 days.
A rule-based verdict, tested side by side. We built a second verdict that is computed from the agents' weighted votes with a fixed formula, including a real "neutral" zone, and let the AI write only the explanation. It runs next to the original verdict without trading, and we will compare the two against the market before either one is trusted.
No profit claims. Our published track record is practice money, shown before fees, and says so. Since 8 October a small real-money test (about $18 per order from a $50 account) checks that orders execute correctly. It is not a profit strategy.
The Limits of This Study
We want to be as clear about the weaknesses of this test as about its results:
One market period. April to October 2026 was mostly a rising market. A council that is always bullish looks better in a rising market than it would in a falling one.
Small trade counts. The trading test produced 20 to 83 trades per signal. That is enough to rule out a large edge, not enough to prove a small one.
Many tests. We tested seven entry signals. When you test several things, some look good by chance, so the two "better than random" results are moderate evidence at best.
Our own system. These results are about AIOKA's council as it was built from April to October 2026. They do not prove that AI councils can never work.
What This Means If You Use Trading Signals
Whoever you get trading signals from, an AI service, a Telegram group or a newsletter, you can ask for the same test:
How often does it say "buy"? If the answer is "most of the time", it cannot be telling you much.
Compared with what? A signal's returns mean nothing on their own. Compare them with random entry times over the same period, and with simply holding.
After fees? Round-trip trading fees of 0.5% to 0.8% wipe out most small edges. Ask for results after fees.
Is the full record public? Every signal, including the losers, not just the screenshots of the winners.
If a provider cannot answer these, treat their signals as unproven. That is how we now treat our own.
You can follow the council's current vote every hour on aioka.io/live, and every practice trade, wins and losses, is on the track record.
*This article is for informational purposes only and does not constitute financial advice. Past performance does not guarantee future results. Always do your own research before making any investment decisions.*