Where there is a goal / prediction, there are always uncertainties.
Credibility is tied to intended use. (the weak → strongest error and uncertainty control)
In case of limited amount of observations, the “uncertainty” can be classified as:
A deeper discussion about the classifications of “uncertainty”:
We cannot robustly quantify the behavior of the tails of the distribution from a small number of samples. This is because, by definition, our samples will, with high probability, not have any values from the tails of the distribution. Therefore, the best we can do is make a guess as to the tail behavior of the system. We need to acknowledge that we have made this assumption about the tail behavior of the distribution, and not make overly specific claims about the probability of a tail event is.
Changing the assumptions will not only mean changing predictions, but also changing uncertainty representation (e.g. iid)
