RSI 7 below 25 with ADX filter: rescued by regime fit

We tested an RSI mean-reversion strategy that enters long when RSI(7) drops below 25, filters out strong trends using ADX below 35, and exits on net profit targets. It failed our standard validation gates but was rescued and placed in observation status after we found it performed profitably in specific regime conditions.

The backtest numbers were ugly. Walk-forward validation showed an average loss of 0.338% per trade despite a 77% win rate, which tells you the wins were tiny and the occasional losses were large. The deflated Sharpe ratio, a measure that accounts for multiple testing and overfitting risk, came in at negative 1.98. The probability this edge is real rather than statistical noise: effectively zero under our multi-test correction. We soft-killed it in early May 2026 for insufficient out-of-sample trades and poor Sharpe performance.

Then we looked closer. When we segmented performance by regime, the strategy showed profitable behavior in ranging, low-volatility conditions on certain coins. That regime fit was strong enough that we restored it to observation status. It is not cleared for general selection or live trading, but we are continuing to collect data as our forge integration comes online.

Forward paper trading has logged 20 trades with an average profit of 0.058% per trade and a 60% win rate. That is a small sample and barely positive, but it is not bleeding, which matters when you are testing whether a regime hypothesis holds up in real market conditions. The strategy runs on BTC, ETH, XRP, and BNB using 5-minute bars, and only trades during the 12:00 to 22:00 UTC session window.

Be skeptical. This strategy exists in our library as a research artifact, not a deployable edge. The validation data said it should not work in general. We kept it alive because regime segmentation suggested otherwise, but until we have a larger forward sample and tighter regime detection, treat this as a curiosity.

If you want to test your own mean-reversion ideas with regime filters, head to /prove and run them through the same six-gate process. You will see exactly where they break.