It's actually both! The US NWS basically figures out both of these things and multiplies them together to get a final probability. The "chance of rain" comes from evaluating the output of multiple weather forecast models, along with meteorologist expertise and pattern recognition.

I was just playing with this problem yesterday evening actually. Perturbation approaches are the best once hw float precision is exhausted. It feels so much like a gpu problem but it just doesn't execute well on shader cores once you gotta deal with arbitrary precision

this place was awful back when invites were handed out on a clout basis