RSI-7 mean reversion killed by its own validation data

We killed this RSI mean-reversion strategy before it could do more damage. The gene entered long when 7-period RSI dropped below 25, filtered trades to low-ADX regimes (below 35) to avoid strong trends, and ran only during 12:00 to 22:00 UTC sessions. It exited on net profit targets. The idea was sound: buy oversold dips in choppy conditions. The execution failed every gate that mattered.

Walk-forward validation across BTC, ETH, XRP, and BNB on 5-minute bars told the story. Average P&L per trade came in at -0.136%, meaning the strategy bled value on every trade after fees. Win rate clocked 60.2%, which sounds promising until you realize the wins were smaller than the losses. The deflated Sharpe ratio (a measure that adjusts for the multiple testing we run in the lab) landed at -0.57, deep in negative territory. The probability this edge was real, not a fluke: 0%. We ran 590 out-of-sample trades in validation and the signal was clear: this doesn't work.

The forward test briefly confused things. Fifteen paper trades returned 0.270% average P&L with a 46.7% win rate. That's the danger of small samples. Fifteen trades tell you almost nothing, especially when they contradict 590 validation trades. We soft-killed the strategy on audit when the strict gates caught what the forward test hadn't yet revealed: the backtest validation itself said insufficient edge.

The regime filter likely hurt more than helped. Restricting to low-ADX environments shrinks the trade pool, and RSI mean reversion needs volume and volatility to pay off after fees. The session filter (12:00 to 22:00 UTC) may have clipped the highest-liquidity hours depending on the coin. And a 7-period RSI is fast, which means more whipsaws and more fee drag.

This is a reversible soft-kill. If we fix the gate logic or the strategy proves itself under new conditions, it could return. For now, it's a textbook case of why we trust the validation sample over a handful of forward trades.

Want to test your own mean-reversion idea with better filters? Head to /prove and run it through the same gates. Or browse strategies that survived the gauntlet at /survivors.