Independent audit
Can an AI search find a winning Bitcoin strategy?
We chose this strategy because it was posted publicly, and tested it on past data. Nobody asked or paid for this audit, and the accounts named are not Autrady customers.
Short answer
FailsNo. We ran the same kind of search on Bitcoin: 152 common indicator strategies, 2020 to 2026, after fees and funding. The best one made +697% in its backtest, less than simply holding Bitcoin (+1,083%). On shuffled prices where no pattern can exist, the same search still found a best strategy worth +457%.
- strategies beat holding Bitcoin
- 0 of 152
- median "best strategy" on patternless prices
- +457%
- for the 2020–2023 winner in 2024–2026, while Bitcoin rose 84%
- −55%
The claim
“found 151 strategies and the best ran 6,724% in the backtest”
The post describes an AI agent with a backtesting tool that searched trading strategies on Bitcoin. It reports 849 trades and a 53% drawdown for the best one. The same thread describes a 22-hour live run on a $10,000 crypto futures account: 16 trades, up to 11 positions open at once, ending on 8 straight winners.
The thread publishes no rules, test period or leverage for the backtest, and no profit, fees or drawdown for the live run. So the strategy itself cannot be tested. We tested the process instead: let a search pick the best of about 150 standard strategies by backtest return, then check whether the winner is real.
What we tested
Everything below was fixed in the script before any result was looked at.
- Market
- Bitcoin perpetual futures (BTCUSDT on Bybit), hourly bars from 25 March 2020 to 29 October 2026, and 4-hour bars built from them.
- Strategies
- 76 common indicator rules: moving-average crosses, price above a moving average, momentum, MACD, Donchian and Keltner channels, Bollinger breakout and reversion, RSI and stochastic. Each long-only and long/short, on 1-hour and 4-hour bars: 152 strategies (the thread: 151).
- Trading
- Signal at the bar's close, position from the next bar. No leverage.
- Costs
- 0.065% per side (0.055% taker fee plus 0.01% slippage), plus the funding actually paid or received every 8 hours.
- Pass bar
- Walk-forward, 2022 to 2026: each year, trade the strategy with the best Sharpe ratio on all earlier data. Pass if its alpha against Bitcoin has a t-statistic of at least 2.33 (Newey-West) and is positive in both halves of the period.
Five tests
- The thread's process: the best of 152 by full-period backtest return, against holding Bitcoin.
- Pick the best on 2020–2023, then trade it in 2024–2026.
- The walk-forward test above, which decides the verdict.
- A noise control: the same search on 20 shuffled copies of the real hourly bars. Every move is real, but their order is random, so no rule can have an edge.
- The deflated Sharpe ratio of the best strategy, which allows for having tried 152.
Results
| Test | Result | Holding Bitcoin |
|---|---|---|
| Best of 152, full backtest | +697%, 239 trades, max DD −45%28-day momentum on 4-hour bars, long only | +1,083%, max DD −77% |
| Strategies that beat holding | 0 of 152 | |
| Best of 2020–2023, traded 2024–2026 | −55%, rank 108 of 152It made +1,293% in 2020–2023 | +84% |
| Walk-forward 2022–2026 the verdict test | +33%, alpha t +0.21 Fails | +69% |
| Walk-forward with no costs at all | +118%, alpha t +1.03 Still fails | +69% |
| Same search on shuffled prices 20 runs, no pattern exists | Best strategy: median +457%, highest +688% | +1,083% |
| Deflated Sharpe of the best | Probability 0.03 that it beats luck A pass needs 0.95 |
Why a search on Bitcoin "finds" big winners
Bitcoin rose almost twelvefold over the period. Any rule that is long most of the time collects that rise. That is why the search on shuffled prices, where no rule can have an edge, still reports a best strategy of +457%. A big backtest number on Bitcoin is mostly Bitcoin.
The real best (+697%) sits at the top of that noise range. At most it is a faint four-week trend, which published research on Bitcoin also reports. It still loses to holding Bitcoin and fails the walk-forward test.
Picking the past winner is the trap. The best strategy of 2020–2023 made +1,293%, mostly by going short in the 2022 crash. In 2024–2026 it lost 55% while Bitcoin rose 84%.
What about +6,724%?
We could not get near it. Allowing leverage from 1x to 5x, the best rule with a drawdown under 53% is still the unleveraged +697%. Holding Bitcoin at 2x made +1,496%, with a 97% drawdown on the way. A figure like +6,724% needs a longer history, heavy tuning of parameters, or a backtest error: look-ahead, where the code trades on a price before it exists, is common in AI-written backtests. Without the rules we cannot say which.
The 22-hour live run: eight straight winners
Positions opened at the same time in crypto tend to win or lose together. To measure how much, we replayed 2,769 baskets with no skill at all, from December 2024 to October 2026: every 6 hours, open Bitcoin and 10 large altcoins at once, all long or all short, each with an equal stop and target, closed after 22 hours at the latest.
- Outcomes of positions in the same basket were strongly correlated (0.43 to 0.50). Eleven positions behaved like about two independent bets.
- Eight or more winners out of 11 happened in 34% to 40% of the no-skill baskets.
- Read as independent coin flips, 8 wins in a row would look rare (0.4%). That reading is the illusion.
- Sixteen trades over one day cannot show skill either way. One bad morning at 10x leverage, the level the thread's video shows, erases a lucky day.
Verdict
Fails. A search over about 150 standard rules on Bitcoin finds whatever rode Bitcoin's rise. It does the same on data with no patterns at all, loses to holding Bitcoin, and fails walk-forward. We could not reproduce +6,724%, and the live run amounts to about two correlated bets on one day.
What would change the verdict
- The exact rules, test period, leverage and costs behind +6,724%. We would run them as written and publish the result, pass or fail.
- A strategy from such a search that beats holding Bitcoin after costs in a walk-forward test, with alpha t of at least 2.33 and positive alpha in both halves.
- Months of live results with every trade, fee and drawdown shown, not one day.
About this audit
This is Autrady research, not yet in the self-serve product: the self-serve audit tests signals on US stocks and ETFs, and we ran this crypto check ourselves. Price data and funding come from the exchange's public history. For the script and the full table of 152 strategies, email [email protected].
Have the signals behind a strategy? Audit them before you trade them.
Upload one row per signal, the ticker and the date it went out, and get a verdict after costs, judged against the market. 14 days free, no card, then €19 a month. Self-serve audits cover US stocks and ETFs.
Autrady tests historical evidence. This is research, not investment advice, and past results do not predict future returns. If you posted one of the strategies above and think we misread a rule, email [email protected] with the exact rules: we will run them as written and publish the result, pass or fail.