Negative results rarely get written up. These are mine — experiments killed with evidence, kept here so I don't quietly re-run them. Each one names the test that killed it.
Followed profitable-looking leaders with paper money for four months: 3,122 trades, −$224, a 28.3% win rate, negative every single month. A leader-quality filter didn't restore the edge. Retired. The dry run's main value was making those losses fake.
Tested every fixed 2-hour window across the sample. Best gross edge: about one basis point. Fees: six times that. Dead on arrival.
Gross carry ran between −3% and +2% annualised in 2026. There's nothing to harvest. Shelved unless the funding regime changes materially.
Full-sample t-stats looked healthy. Split by year, the edge died in 2022 and never came back. Lesson kept: never trust a full-sample statistic without a year split.
The ratio is a random walk with a 241-day mean-reversion half-life. There is no spread to trade.
Real structure exists — a BTC-to-altcoin lead-lag with a t-statistic of 7.7 — but it sits roughly 15× below taker fees. Real and unreachable: the worst kind of result.
Buy every outcome when the set prices under $1.00. Four months of scanning found zero qualifying candidates. The market is tighter than the strategy assumed.
Team form, referee histories, fantasy-football sentiment: public information the bookmakers price before you do, behind a 2–8% vig. Killed at the research stage without building the pipeline.
What survived is on the now page.