RSI Mean Reversion: The Indicator That Fails Two Different Ways

The Window 1 paradox (2015)

Window 1 shows a 70% OOS win rate — the highest in the table — alongside a near-zero Sharpe (+0.06) and negative net expectancy (−1.29 pips/trade). This is not a contradiction. It is a signature of mean reversion fighting a trend. In a trending environment, the strategy catches many small pullback bounces (wins) while occasionally entering into continuation moves that produce outsized losses. The 70% win rate is real. The 30% of losing trades are sufficiently large to overwhelm it.

Window 4 (2021) — same disease, worse outcome

Window 4 produced a −0.64 Sharpe despite a 60% OOS win rate. The same dynamic: wins are frequent but small; losses are infrequent but large. The EUR/USD downtrend of 2021 (approximately 1.23 to 1.13 across the year) created a persistent environment where oversold entries into a falling market repeatedly produced small bounces followed by trend resumptions.

Window 2 (2017) — why the aggregate Sharpe looks good

Window 2's +2.38 Sharpe with 75% OOS win rate and +10.45 pips/trade is the outlier that pulls the aggregate upward. EUR/USD in 2017 oscillated in a range with trend that produced clean mean-reversion setups. Remove Window 2 and the remaining five windows average a Sharpe of roughly +0.43 — below the survival threshold. The aggregate Sharpe of +0.81 is, in a meaningful sense, Window 2's Sharpe diluted by the others.

The alternating pattern below threshold=36

Between thresholds 20 and 34, net-expectancy and drift kills alternate. Threshold=24 drifts; threshold=26 loses money. Threshold=28 drifts; threshold=30 loses money. This is not noise — it reflects a genuine interaction between how frequently the signal fires and which specific market environments it captures. The odd-threshold variants happen to capture more losing trade distributions in the historical data.

Threshold=34: the clearest edge case

Threshold=34 generated 210 OOS trades and a Sharpe of 0.60 — the second-highest in the family — but posted a net expectancy of −0.41 pips/trade and was killed at Stage 1. With 210 trades and a −0.41 pip average, total losses across the sweep were modest in absolute terms, but the kill filter is per-trade expectancy: any negative average fails, regardless of aggregate performance metrics.

Above threshold=34: drift dominates

Once signals fire more than 200 times across the sweep, net expectancy is no longer the bottleneck. Thresholds 36, 38, and 40 all produce positive average expectancy. They are killed by drift: high Sharpe variance across windows that cannot be predicted in advance. The threshold dial moves the strategy from 'losing money steadily' to 'making money inconsistently' — but not to 'making money stably.'

Finding 1: RSI's failure windows differ from MACD's

RSI underperforms in trending environments (Windows 1 and 4 in the best variant). MACD underperforms in regime-change environments — windows where the character of the market shifts between training and test periods. These are not identical. A combination signal that uses MACD as a trend filter and RSI as an entry trigger would at minimum face different structural failure conditions than either signal alone. Whether it survives those conditions is an empirical question. This is not an assumption — it is a hypothesis for Phase 3.

Finding 2: Threshold 38 is the only RSI configuration with any residual interest

Four of six OOS windows produced positive Sharpe for HYP_0010. The two negative windows have a clear regime interpretation. If a directional filter could suppress RSI signals in trending environments — specifically by requiring a trend-neutral or mean-reverting regime condition before entry — the 2015 and 2021 window losses might be reduced. This is not validated here. It is a testable proposition for Phase 3. No other RSI threshold configuration merits Phase 3 inclusion: the alternating net-expectancy kills at lower thresholds have no obvious fix.