The best backtest we found, and why we did not trust it
Out of 31,407 setups, a 4-hour Bollinger Band breakout with a 2% stop-loss made +63% a year after fees and slippage, passed every check we had, and held up over nine years of replay. We still did not put money on it.
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 rule
On 4-hour bars: buy when the close rises above the upper Bollinger Band (30 bars, 2.5 standard deviations); sell when it closes below the middle line or falls 2% below the entry price. Adding a filter that only buys above the 200-bar average gave the single best result in the whole study, +64.3% a year with a worst fall of −11.3%, but the plain version is easier to trust because it has fewer moving parts.
| 4-hour BB(30, 2.5σ) breakout + 2% stop | |
|---|---|
| Return a year, fee only | +71.1% |
| Return a year, fee and one tick | +63.3% |
| Worst fall, with slippage | −16.9% |
| Trades, win rate | 675, 35% |
| First period / second period | +67.7% / +58.2% a year |
| By year, fee only | 2024 +113%, 2025 +33%, 2026 to October +56% |
| Coins that made money, fee only | 12 of 12, from +20% to +156% a year |
| Nine years, 2017 to 2026 | +98.3% a year fee only, +86.0% with slippage; worst fall −30% to −33%; one losing year, 2022 (−15.8%) |
Why it looked real
- It made almost the same in the second period, which played no part in choosing it, as in the first.
- Its neighbours did nearly as well: 20 or 50 bars, 2 or 3 standard deviations and stops from 1% to 3% all made between +43% and +69% a year with the fee alone. A lucky point would sit alone.
- It made money on every one of the twelve coins, and in a falling second period in which holding the same coins returned −13.7% a year.
- Breakout rules have decades of academic support, in crypto too. The Bollinger page has every setting.
Why we did not trust it
- We chose the coins in 2026. All twelve were large and had been listed for at least two years when we picked them. Coins that collapsed are missing, and our own survivorship test cut a similar rule from +75% to +41% a year.
- It is the best of 31,407. The top of a long list is partly luck by construction. The two-period test narrows that but cannot remove it.
- Breakouts are where slippage is worst. We charged one tick per trade. Breakouts happen when prices jump and order books thin out, which is exactly when fills are worse than one tick.
- The model is simpler than an account. Equal money in twelve coins, each traded on its own, is not how a real account with limited slots and a changing coin list behaves.
- One year did much of the work. 2024 alone returned +113%.
- Our best live-tested rule had looked as good. The rule we ran on a server made +88% a year in the same nine-year replay and lost 3.93% in its first two live weeks. The server test.
How we would test it
Paper-trade it live for months, not weeks; compare every live decision with the replay bar by bar; measure real fills on breakout bars; and expect a fraction of the backtest. We paused trading research before doing that, so this remains a backtest. You can sort every Bollinger setup yourself in the explorer.