I have a derivation somewhere. I think it starts with probability of claim and assumes a normal distribution. Then you have to choose your confidence interval.
But also, that only comes up with the standard for the severity, no? Since it depends solely on the # of claims, and not considering the exposure at all.
My two cents:
If it’s non-traditional, you’ll just have make your own credibility standard.
And how would you do that?
Well, you’ll have to model some large population, see if its size allows you to predict future claims. If it doesn’t, you’ll have to model a larger size.
If there is no finite size of population such that its claims are predictable (in a range you’ll accept), then too bad.
If my notes are right, the formula for credibility is {Phi^(-1)[1 - (1-0.95)/2]sigma/mu}^2
where Phi is the normal/gaussian cdf, 0.95 the confidence interval for the derivation, sigma is the standard deviation of the estimate, and mu is the mean.
if you are estimating counts then you can set sigma = sqrt(mu).
remember that this aims to limit fluctuations and gives no account of how reliable is the other measurement.
however, especially for claim counts, it has basically a single parameter, so it’s easier to estimate with less data.