Averaging several RSI lookbacks into an ensemble feels like a robustness win — no single parameter to overfit, so surely the result generalizes. This note runs the ensemble through combinatorially-symmetric cross-validation (CSCV) and finds the opposite: each added lookback is another degree of freedom, and the Probability of Backtest Overfitting climbs fast.
Ensembles are not free regularization
The intuition that averaging reduces variance is correct only when the components are chosen without reference to the test data. When you select which lookbacks to include because the blend backtests well, you’ve simply hidden the parameter search inside the ensemble construction. The PBO reflects that hidden search.
The honest version
An RSI ensemble can still be worth trading — but the lookbacks have to be fixed a priori from an economic rationale, not tuned. Report the DSR against the full trial count, ensemble members included, and let it stand or fall on that.