Research: what we learned trying to make crypto trading pay

Six weeks of building, running and auditing trading systems with an AI assistant, and a backtest of 31,407 indicator setups. Written up with every loss included.

What the data says

• Which timeframe works for crypto trading? 31,407 backtests after costs

  The shorter the bar, the less money survived the fee and the tick.

• Win rate vs. expectancy: why the indicators that win most often lose money

  The rules that won most often lost money. The arithmetic behind it.

• 34 indicator rules on hourly crypto charts: all of them lost

  Textbook rules, unchanged, on 1-hour bars of 22 coins: 0 of 34 made money.

• The one-minute backtest trap: how a fee-only backtest invents profit

  Fee-only 1-minute results in the tens of thousands of percent, and why they were never real.

• The hidden cost of a price tick on Upbit

  At 128 won, one step is 0.78% of the price: fifteen times the fee.

• The best backtest we found, and why we did not trust it

  +63% a year after costs in replay. Six reasons not to believe it yet.

• Our backtest said 75% a year. Survivorship bias said 41%.

  What changes when you pick coins the way you could have picked them at the time.

• How long does it take to prove a trading strategy works?

  About 34 months of live trades for a typical trend rule. Two weeks prove nothing.

• Trend following wins 36% of the time. That is the point.

  Nine years, 1,037 trades: most trades lose, a few trends pay for everything.

What happened when we tried

• Six weeks of letting AI build a crypto trading system

  −44% in the first week, ₩25.2 million of paper losses across 918 exits, and a rule-perfect server that still lost.

• One bug caused 86% of our trading losses

  157 exits held a median 12 seconds, and a wrong conclusion we withdrew a day later.

• Can a language model pick crypto trades? We tested 6,536 candidates

  The trades the model liked most did worst. Plus three counting mistakes.

• Two weeks of a rule-perfect trading server: −3.93%

  Five trades, five losses, zero errors, and what the audit found.

• Do AI trading bots make money? What the public record shows

  Verified profits sit with market makers. No retail bot publishes audited results.

• Strategy graveyard: crypto trading strategies that failed

  Every approach we tested, how we tested it, and why it failed.

Trading research is paused until better AI models are available. If it restarts, results will be published here first, good or bad.