Market Regime Classification vs. Optimization Overfitting - page 2

[Frederic Jean Andre Jouillerot]

Two checks before trusting a regime layer. First, it's still optimization one level up: states, lookback and thresholds need their own out-of-sample test. Second, filters inflate metrics by shrinking the sample. If the filter lifts PF from 1.3 to 1.7 but removes 60% of trades, compare it with randomly skipping the same 60% a few thousand times. If it doesn't clearly beat that distribution, it's selecting good trades in hindsight, not detecting regimes. And a complex classifier should beat a plain ADX or ATR-percentile filter to justify its extra parameters.