This is a particularly fascinating area of inference, at least in my opinion.  While you are asking for an estimation of the probability of an unlikely event, I'll instead offer a paper that presents a heuristic for detection of model error.  It is explicitly probability free.

[Mathematical Definition, Mapping, and Detection of (Anti)Fragility][1]

Taleb has spent the majority of his professional life dealing with so-called "Black Swan Events" (a term he popularized with his book "The Black Swan").

To the point, finding an appropriate distribution of low frequency events is a particularly troublesome task that does not fit nicely into classical statistical inference.  Those events go to the heart of Hume's problem of induction, a topic that Taleb seems to have devoted his professional life to.
  [1]: https://arxiv.org/ftp/arxiv/papers/1208/1208.1189.pdf