ROC-12 reversal with ADX filter killed at audit

We killed this strategy before it ever took a live trade. The idea was simple: buy when 12-period rate of change drops below -2%, targeting mean reversion in ranging markets (ADX below 35 to filter out trending regimes). Exit when net profit hits 1.25% or stop at 1.5% loss. Despite a 73.6% win rate in both training and validation, the strategy posted a -0.004% average P&L per trade and a deflated Sharpe ratio of -0.02. That means the wins were too small to overcome the losses and fees, and the statistical adjustment for multiple testing confirmed this was noise, not edge.

The high win rate is exactly the trap this kind of strategy sets. Small frequent wins feel good in a spreadsheet, but the math doesn't lie. With 208 out-of-sample trades across BTC, ETH, XRP, and BNB on 5-minute bars, we had enough data to see the pattern: the strategy was giving back its gains through a combination of whipsaw and fee drag. The 12-hour UTC session window (12:00 to 22:00) captured active trading hours, and the ADX filter correctly screened for low-trend environments, but the entry threshold was too aggressive. ROC below -2% on a 5-minute bar is a minor dip, not a meaningful pullback, so we were catching falling knives more often than true reversals.

We soft-killed this gene at audit on May 2, 2026, under strict gates that demand positive out-of-sample Sharpe and non-trivial edge. The kill is reversible if we refactor the gate logic, but right now the evidence says this configuration doesn't work. The deflated Sharpe's 100% fluke probability after multi-test correction means we have no reason to believe this edge would persist.

The lesson: win rate alone is a vanity metric. A strategy that wins 73% of the time but loses money is just an expensive coin flip. If you're building mean-reversion systems, check our /learn section on sizing exits relative to entries, or head to /prove to test your own idea with realistic fee assumptions baked in from day one.