Six weeks of letting AI build a crypto trading system
From 2 September to 13 October 2026 a one-person company in Seoul asked an AI assistant to build a crypto trading system, test it, run it, and audit it. Every trade was on paper. This is what happened, in order.
How this was made: the code, the research runs and the first drafts of this text were produced by an AI assistant (Claude) working under the operator's direction; the operator made every decision and checked the numbers. All trading was on paper: real prices, no real money. Nothing here is investment advice.
The short version
- Minute-scale trading lost every way we tried it. Across 42 paper accounts and 918 exits, the loss was ₩25,161,360 (about US$18,900), with a 32% win rate.
- The only rules that survived nine years of replay were slow trend-following rules on large, liquid coins. Even those shrank a lot once we removed hindsight from the backtest.
- The best of them ran for two weeks on a dedicated server, followed its rules exactly, and lost 3.93%. Holding Bitcoin over the same two weeks lost 0.25%.
- The operator's conclusion: the problem is the method, not the market. Trading research is paused until better AI models arrive.
2–5 September: a radar for order books
The first goal was a free website that shows price gaps between exchanges, plus a trading signal. The AI built a data engine that read order books and trades from Upbit, then from twelve exchanges, and a dashboard that scored every coin as "buy, wait or sell" each second, with spoken briefings in English and Korean.
6–11 September: the first automated trades, −44.2%
The first auto-trader bought one coin with all its cash at market price whenever the dashboard said buy, and sold on warning signals. The paper result after the first audit was −44.2%. Only 10% of trades won and the median holding time was 12 seconds; 157 of 158 exits were "emergency" exits.
The cause was arithmetic, not bad luck: a market-order round trip on small Korean-won coins cost a median 0.46%, while the signal's edge was +0.02% to +0.08% per trade.
12–17 September: patches that did not fix the problem
We tried limit orders, longer holding, a lead signal from Binance to Korean prices, and parallel "shadow" accounts on every exchange. Chasing coins that were already surging lost 62% in a 63-hour replay. A five-day study of 1.91 million rows showed the signals guessed direction 58–63% of the time, but the average move they caught was only +0.04% to +0.09%. A good hit rate cannot pay for costs ten times larger than the move.
A paper-ledger audit found one more thing: 86.5% of all losses came from a single code path that sold on unconfirmed alarms. Fixing it removed a bug, not the underlying problem.
18–25 September: looking everywhere else
We tested language-model trading decisions, leaderboard copying, pump strategies, dip buying and fee tiers. We tested 34 classic indicator rules on hourly charts: 34 out of 34 lost. A survey of published results found verified profits only where firms have a structural edge (speed, market making, scale). We found no verified track record for retail bots, copy trading or AI bots.
26 September: counting the damage and slowing down
Across the whole project so far: 42 paper ledgers, 918 exits, ₩25,161,360 lost, 32% wins, average win half the size of the average loss. All of it came from rules that held positions for minutes to hours. We switched to daily and four-hour trend rules, and a replay on eight large coins showed +75% a year. Then we noticed the eight coins had been chosen in 2026. Choosing the top eight by trading volume at each point in time cut the result to +41% to +46% a year. Details.
28 September: start over, test everything
We threw away earlier assumptions and tested 44 strategy families, about 9,500 settings, across seven timeframes, choosing settings only from past data each year (walk-forward). Only trend following on liquid coins at four hours or one day survived. The pick: buy when the four-hour close breaks the 55-bar high, sell when it falls two average ranges below its peak, stay out when Bitcoin trades below its 100-bar average, and cut losses at 5%. Nine-year replay: +92% a year. After adjusting for the traps we knew about, we expected +28% to +46% a year with drawdowns of 27–33%.
We also worked out how long live trading would take to prove a rule like this: about 34 months. The calculation.
29 September – 13 October: a dedicated server
A small server in Seoul ran the rule around the clock with a standalone program, Telegram alerts and an external heartbeat. Paper money: ₩10,000,000. Over two weeks it processed all 87 four-hour bars on time, logged zero errors, and matched an independent replay trade for trade. It also sold five times and lost five times: −₩392,842, −3.93%. From 7 October it had been sitting in cash because Bitcoin fell below its average, which avoided a further drop. The full ledger and audit.
A final study across 26,154 indicator setups confirmed the pattern: the shorter the timeframe, the less money survived costs. Fees eat the edge.
What we would tell anyone building a bot
- Compare your edge per trade with your cost per round trip before anything else.
- A hit rate means nothing without the size of the move.
- Pick backtest coins the way you could have picked them at the time.
- Look at losses by exit reason; one bug can look like a bad strategy.
- Weeks of live results cannot prove a strategy. Use them only to check the system follows its rules.
- A system that follows its rules perfectly can still lose money.
Everything we tested is listed in the strategy graveyard.