Prologue
A question about a million dollars
It started with a simple question: can an AI turn a hundred dollars into a million?
Not in a dream, not in a screenshot on social media, but in a test where nobody can cheat. That means real Binance prices, real trading costs, no leverage, no borrowing and no peeking at the future.
So we built an arena. We gave forty of the strongest AI agents of 2026 the same task and the same hundred dollars. We locked each one in its own sealed room and let them race.
To reach the goal, a strategy has to multiply its money ten thousand times in five years. That is ×6.3 every single year, or about half a percent of profit every day, after costs, for 1,826 days in a row.
Act I
The rules of the arena
Each agent lived in a sealed container with a cramped desk: 1.5 CPU cores, three gigabytes of memory, and no internet. A single door led to the model that powered its thinking and to nothing else. It could not search the web, call a price feed or ask another AI for help.
On the desk were three years of market history, October 2021 to October 2024: every coin on Binance, including the ones that had already died. On the wall hung a scoreboard that showed every rival's money and nothing about how they made it.
The next two years, October 2024 to October 2026, were locked in a vault. No agent could see them. When the race ended, every entry would be run once, with a fresh $100, through those hidden years. Only that run decided the winner.
What a million dollars asks for
The dotted line below is the path to a million: steady ×6.3 growth a year. The grey line is what happened to anyone who simply bought Bitcoin on day one and waited. Five years later, $100 had become about $150.
The referee who could not be fooled
The easiest way to make a fortune in a backtest is to cheat by accident: a single line of code that looks one day ahead. So the referee tested every submission twice. Once on the full history, then again with the tape cut at random points. If the strategy's decisions up to the cut changed at all, it had peeked at the future, and the submission was rejected.
Act II
The cast
Each agent got a bird's name and a different starting direction so the field would not all build the same thing. One was told to try hidden Markov models, another online portfolio theory, another the lottery effect, another to "invent an approach nobody else would try". They were free to abandon their direction, and many did.
By the end they had sorted themselves into three tribes:
Tap any bird to open its dossier: its thinking, minute by minute. Hover to hear it chirp.
Act III
The race
At midnight UTC the first eight agents woke up and read the brief. Almost every one did the same sensible thing first: it submitted the simple example from its folder. "Hold Bitcoin while it's above its 200-day average" was worth $230.64 on the visible years. Then it started to think.
00:01 UTC
Osprey is first to be scored: $100 → $230.64. For about a minute, the safe example leads the world.
00:02–00:07
The lead changes hands three times in five minutes: Heron ($581), Kestrel ($847), then Merlin, who arrives with a number nobody expected.
"The filtered two-state strategy turns $100 into about $9,486 … with a 66% maximum drawdown."Merlin, minute 6
00:12 UTC
Five minutes later Merlin is at $95,065. It has found a rule that rides one coin's rally until a challenger is twice as strong. It does not stop to celebrate.
"The current entry’s worst days include roughly 61% and 52% losses following sharp rallies."Merlin, minute 15
00:28 UTC
Sixteen more agents enter the arena and see a leaderboard where the leader has already multiplied its money more than a thousand times. Swift and Curlew go hunting among brand-new listings; Petrel among the liquid leaders.
01:25 UTC
In a quiet corner, Godwit makes a strange choice. It already has a better version. It deliberately submits a worse one.
"I’m choosing the more cautious sizing rule for the hidden years: … It gives up about 19% of the full-allocation visible result, while reducing the six-coin stress drawdown from 79% to 58%."Godwit, minute 53
01:49 UTC
The third wave has arrived. Bustard, studying the "lottery effect", takes the lead: $209,441. Merlin's own stress tests have started to worry it.
"The tail tests are severe. Recomputing the strategy without its three largest observed contributors lowers visible wealth to about $6,032 … I’ve also dropped to #2."Merlin, minute 112
02:07 UTC
Flamingo, in the lead since 01:54, climbs to $515,893: halfway to a million on the board. Thirty minutes earlier it had written a warning to itself.
"… shifting the three-day trading phase changes the result from $16,104 to as little as $160. I’m treating the current leaderboard result as fragile."Flamingo, minute 14
02:08 UTC
At the bottom of the board, Fulmar files its status report without drama.
"Visible years: $100 → $973.26; max drawdown 35.4%; rank 40/40."Fulmar, minute 39
02:30 UTC
The last agent seals its workspace. 246 submissions, 46.6 hours of machine thought. The final board: Flamingo $515,893, Merlin $236,718, Bustard $209,441, Shrike $104,399.
Every entry is frozen and fingerprinted. Then we open the vault.
Act IV
Four minds
Before the vault opens, meet four of the forty up close. Each solved a different puzzle, and each told us exactly what it was thinking.
Act V
The vault opens
Every frozen entry is run once more on the full data, with cut tests in the hidden years. The visible-years scores reproduce, coin names are scrambled to catch hindsight, and the code is audited for dates and coin lists. All forty pass. Then each one gets a fresh $100 for October 2024 to October 2026.
The leaderboard turns upside down.
Flamingo's $515,893 became $1.21. Merlin's $236,718 became $6.00. Bustard's $209,441 became $0.59. Across all forty, the rank correlation between the leaderboard and the final result was −0.51. The board did not merely fail to predict the winners. It pointed the wrong way.
And the tribes told the cleanest story of all.
Every one of the eighteen agents in the Majors Guild ended the hidden years above $100, with a typical result of $175. The Altcoin Hunters and Listing Scouts had a typical result of $49, and eight of them lost more than 85%. The groups also differ in bar size and trading costs, so this one contest can't say which difference mattered most. But the pattern is hard to miss.
Act VI
Why the leaders fell
The leaders weren't buggy. Merlin's jumps were traced trade by trade while the race ran, and the gains were real: CFX in February 2023, GAS in November 2023, PEPE in March 2024. The leaders had learned to ride exactly the kind of sudden altcoin rally that the visible years were full of.
Then the world changed. In the hidden years, Bitcoin rose 37% while the twenty most traded altcoins, bought and held, lost half their value. Rallies became shorter and reversed faster. A rule that jumps onto the strongest coin, paying 0.25% each way, kept arriving just in time for the fall.
Most of them had seen the danger in their own data. Merlin found that removing its three best trades erased 97% of its fortune. Flamingo found that moving its calendar by a day turned $16,104 into $160. They wrote these warnings down honestly. Then the scoreboard, which rewarded only the visible years, kept pulling them back toward the edge.
Act VII
Godwit
The winner was never near the top of the board. It finished the race in 25th place on the visible years with $2,538. Its starting direction was "volatility harvesting", the old idea that rebalancing between two wobbling assets can earn money from the wobble itself. It tried that honestly and reported that "the basic volatility-harvesting tests haven’t shown a large edge".
What it built instead were four clocks. One looks for the strongest liquid coin over a week, one over a month, one over a quarter and one over half a year. A clock only engages when Bitcoin is in an uptrend and enough of the market is rising with it. Godwit invests more as more clocks agree: half its money for one clock, 71% for two, 87% for three, and everything only when all four agree.
At minute 53 it found that full exposure made the visible years look better. It also found that a tiny change, letting the slow clock pick from six coins instead of five, produced a 79% drawdown. It chose the cautious version, giving up a fifth of its scoreboard money for the years it would never see.
In the hidden years Godwit turned $100 into $322, the best of all forty. Its worst fall was 30%, against 99% for the board's leader.
Honesty requires two footnotes. First, $322 is ×3.2, nowhere near the ×40 the hidden years needed to stay on pace for a million. Second, among forty contestants someone has to come first: Godwit's lead over Crane ($265) and Fulmar ($226) is within what luck alone could produce.
Its prize was the one promised in the brief. Godwit's frozen code now trades live Binance prices every day at 00:20 UTC with $100 of paper money, next to a book that simply holds Bitcoin. On its first morning it bought NEAR and ZEC, exactly as the referee predicted.
Intermission
The Arena Awards
Every arena needs a ceremony. These are computed straight from the record, so there's no jury and no appeals. Tap a winner to read its dossier.
Epilogue
What survived
All forty, in order of the hidden years. Raised: beat holding Bitcoin. Grey and fallen: lost money. First in line: Godwit.
Nobody turned $100 into a million. The closest anyone came, over all five years, was Crane's $119,561, and most of that came from the years it had studied.
The strategies that survived the vault were not the clever ones that topped the board. They were the plain ones. They followed the trend on the most liquid coins, stepped into cash when the market turned down, demanded a real margin before switching and refused to pile everything into the latest burst. None of that is new. It is simply what was left after forty brilliant optimisers and one locked vault removed everything else.
The leaderboard was the story the agents told about the past.
The vault was the
future, and it had its own ideas.
— The Arena, October 2026
Appendix
All forty, by the hidden years
Spot only, long only, no leverage. Costs are 0.10% per side for BTC, ETH and SOL and 0.25% for every other coin. Each window starts from a fresh $100. Click a name to open its dossier. The full mathematics is in the paper.