Our backtest said 75% a year. Survivorship bias said 41%.
A trend rule on eight large coins replayed at +75% a year from 2018 to 2026. Then we asked which eight coins we would actually have picked in 2018.
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
A trend ensemble: on daily and four-hour charts, three trend states (RSI regime, 20/10-bar channel breakout, price above its 100-bar average) vote; buy when both timeframes agree, sell when either turns off, and stay out when Bitcoin is below its 100-day average. Costs were included.
The mistake
We had tested it on BTC, ETH, XRP, SOL, DOGE, ADA, XLM and TRX, the large coins of 2026. In 2018 nobody knew which coins would still be large eight years later. Picking today's winners quietly hands the backtest information from the future.
The fix
We collected full history for 289 Upbit markets and, on each date, let the rule trade only the eight coins with the highest trading volume at that time.
The point-in-time version is the honest one, and it is still optimistic: coins that were delisted are missing from the exchange's data. Holding that same changing top-8, rebalanced monthly, lost 58% in total (−9.8% a year, with a −91.5% drawdown), which shows how much of the rule's result came from getting out of falling coins.
Two more checks we ran
- Look-ahead: delaying every trade by one bar still gave +68% a year on the fixed coins, while deliberately leaking the next bar's price produced +291% a year. The gap is how a hidden leak would look.
- Even after both checks, our live two-week test lost money (server test).
The lesson
Choose every input the way you could have chosen it at the time: coins, parameters, even the decision to use a filter. Then expect less than the backtest.