Williams %R + MFI mean-reversion killed by gate audit

We killed this strategy after a gate audit found it should never have graduated. The entry looked for double-oversold conditions using Williams %R below -85 and a 14-period Money Flow Index (MFI, a volume-weighted RSI) below 30, active only when ADX stayed below 30 to avoid trending markets. The idea was to catch mean-reversion bounces in quiet, range-bound periods.

The backtest was brutal. Walk-forward validation showed -0.227% average P&L per trade with a deflated Sharpe ratio of -1.90 across 1,364 out-of-sample trades on BTC, ETH, XRP, and BNB on 5-minute bars. The win rate barely broke even at 50.7%, but winners were smaller than losers after fees. The multi-test corrected probability of being a fluke was 100%, meaning the system concluded this signal had no edge whatsoever.

The strategy somehow slipped through to forward test, where it logged 19 trades and posted 0.013% average P&L with a 42.1% win rate. The forward numbers were noisy but not encouraging enough to override the validation red flags. When we ran a strict-gate audit in May 2026, the original validation metrics failed the threshold and we soft-killed the gene. It will be re-evaluated when we deploy the corrected gate logic, but the underlying signal looks too weak to rescue.

The failure teaches two lessons. First, filtering for low ADX does not guarantee clean mean-reversion setups. Choppy, directionless markets can grind sideways without bouncing, and double-oversold conditions can stay oversold. Second, tight exit rules on 5-minute bars expose you to noise and slippage. The net-profit exit mode with a 1.5 percent target and 1.5 percent stop sounds symmetrical, but in practice fees and spread eat the edge before the target hits.

If you are working on mean-reversion ideas, check out the full survivors registry to see which oscillator combinations held up under fees, or head to /prove to test your own momentum or regime-switching logic before committing capital.