The POT model can be used to study tail behaviour, events exceeding a certain threshold.
Given a sequence of random variables X1, X2. . .XK and a threshold level m, only
Xk:Xk]m are considered. We can think of the sequence as all accidents during T years,
where X0
j denotes the death toll in accident j and that we are only interested in accidents
claiming at least m lives, discarding X0
j if X0
j Bm and putting XkðjÞ ¼ X0
j if X0
j m, where
kðjÞ ¼ jfi; 1
i
j;X0
i mgj is the catastrophe number that accident j corresponds to.
The POT model assumes that the number Km of Xk is Poisson distributed and that the
exceedances Xkm are independent and identically Pareto distributed.
To justify the use of the POT model, the threshold parameter m must be large enough
so that the exceedances are in the tail of the distribution. What constitutes large enough
cannot be known a priori, one must look at data and use for example quantile-quantile
plots (QQ-plots) to decide a level of m that is consistent with the model.
In the case were we study the distribution of lost lives in deadly accidents, it is known
that the far majority of such events are single accidents, that is, claiming one life. Using
the (perhaps outdated) formula (1) gives at hand that 97% of victims were in accidents
claiming one or two lives. It is therefore reasonable to believe that the POT model can
work in the life catastrophe setting with an m as low as three or four. This would be handy,
since as mentioned above, the Cat XL parameter M is often chosen to be three to five.
Since the number of deaths is a discrete random variable, it could be argued that it is
logical to use the discrete counterpart of the Pareto distribution, the Zeta distribution.
354 E. Ekheden & O. Ho¨ssjer
Downloaded by