[[AutoGluon]] kept telling me volume-based features were top-ranked. But when I dropped them, model performance improved.
Turns out they were overfitting to recent action, basically cheating by capturing too much recent context that won’t generalise.
Now I use SHAP over rolling holdout sets, and look for consistency in top-ranked features across market regimes. If a feature is only important during rallies, it’s not robust.
[[AutoGluon]] [[ML]] [[Serendipity]]