Nice piece — the r=0.627 at the extremes is a clean result, and I like that you're being upfront that it's a 21-event snapshot rather than overclaiming.

One thing I'd gently push on, mostly as a framing question: the study establishes that the two sentiment measures agree with each other, but agreement between two estimators of the same latent thing is a bit of a foregone conclusion once both are well-built — if news sentiment and earnings sentiment are both proxying "how is this company doing," they sort of have to correlate at the tails. The more interesting (and harder) question is whether either one carries information that isn't already in price by the time you could act on it.

In my own poking around, the thing that's repeatedly humbled me is that sentiment level and forward returns decouple far more than sentiment level and each other — and that conditioning on an event (earnings, a catalyst, etc.) tends to absorb a lot of what the sentiment score looked like it was adding. So I'd be really curious to see this same extremes-agreement cut against a forward-return panel, even a crude one, rather than cross-validating the two scores against each other. My bet is the middle buckets staying undifferentiated (which you already flag) is the whole story.

Either way, appreciate you publishing the preliminary stuff instead of sitting on it until it's pretty. Following.