One thing that might play into it is that actual talent distributions are leptokurtic, ie they have a fat tail, while our internal model is more of a normal dis

One thing that might play into it is that actual talent distributions are leptokurtic, ie they have a fat tail, while our internal model is more of a normal distribution, or maybe even some kind of truncated normal based on what we have previously seen. So these outliers that occur naturally in the leptokurtic have a much lower perceived likelihood. PS: the bit about talent distributions being leptokurtic is pure conjecture, no idea if it is true. Seems like something Nassim Taleb would say.