Can an AI build a trading bot that beats holding crypto?
Short answer: We had an AI design a trend-following bot for BTC and ETH and tune it on 2018 to 2022, where it made about 58% a year against 4% to 10% for holding. On 2023 to 2026, data it had never seen, it made about 21% a year while holding BTC made 54% and holding ETH 24%. It stayed profitable with smaller drops, but it did not beat holding.
“I built a trading bot with an AI and it hit big” is a popular post format. The screenshots usually show one trade or one backtest. We wanted the version that matters: a bot designed and tuned by an AI, then judged on data it never touched.
The bot
- Designed by an AI (Claude) before any test was run: 4-hour candles, long only, no leverage.
- Buy when the 50-period average is above the 100-period average, the price is above the slow average, and the price breaks the highest high of the last 20 candles.
- Sell when the price closes below the fast average, or falls 4 times the average true range below its best close since entry.
- Costs: 0.10% fee plus 0.05% slippage on every buy and every sell.
Why rules, and not the AI deciding each trade: a language model has read about past prices during training, so a backtest of its own trade-by-trade decisions would quietly use the future. Rules it designs can be tested honestly on held-out years; an AI deciding live can only be judged going forward.
How we tested
- The rule was written and timestamped first.
- The AI tried 24 settings on 2018 to 2022 only and kept the one with the best risk-adjusted return.
- That one setting was then run once on January 2023 to September 2026.
- Pass rule: on both BTC and ETH, a higher yearly return than holding, and a better risk-adjusted return (Sharpe ratio) than a simple daily 200-day trend filter.
Results
| Tuning years 2018-2022 | Unseen years 2023-2026 | |||
|---|---|---|---|---|
| Per year | Worst drop | Per year | Worst drop | |
| BTC bot | 58.8% | 31% | 20.8% | 27% |
| Hold BTC | 4.2% | 81% | 53.9% | 53% |
| BTC daily 200-day filter | 24.1% | 55% | 32.9% | 36% |
| ETH bot | 57.9% | 36% | 20.7% | 33% |
| Hold ETH | 10.3% | 94% | 23.8% | 68% |
| ETH daily 200-day filter | 33.9% | 73% | 23.9% | 38% |
On the years it was tuned on, the bot looked spectacular. On the years it had never seen, it roughly doubled the money on each coin, with about half the worst drop of holding, while in the market only about 28% of the time. It still trailed holding on both coins, and on BTC its risk-adjusted return (0.92) was no better than the one-line daily filter (0.93). On ETH it beat the filter on that measure (0.78 against 0.70), but the rule needs both.
The gap between 58% and 21% is the lesson. Most of the tuned result came from choosing settings that fit the past. A bot screenshot rarely tells you which part you are looking at.
Live paper trading
The same bot now runs on live 4-hour candles for six weeks without real money, so its live behaviour can be compared with the test. We will add the result here when it ends.
Limits
- One strategy family, 24 settings, two coins. A different design could do better or worse; the point is the method, which any bot claim should pass.
- Fills are assumed at the next candle’s open.
- 2023 to 2026 was mostly a rising market, which favours holding. A bot that sits out most of the time gives up the biggest up moves.
Verdict
FAIL. The AI-designed bot made money on unseen data and cut the worst drops, but it did not beat holding either coin, and on BTC it did no better than a one-line trend filter. Not financial advice.