Knowledge of the tail shape of claim distributions provides important
actuarial information. This paper discusses how two techniques commonl
y used in assessing the most appropriate underlying distribution can b
e usefully combined. The maximum likelihood approach is theoretically
appealing since it is preferable to many other estimators in the sense
of best asymptotic normality. Likelihood based tests are, however, no
t always capable to discriminate among non-nested classes of distribut
ions. Extremal value theory offers an attractive tool to overcome this
problem. It shows that a much larger set of distributions is nested i
n their tails by the so-called tail parameter. This paper shows that b
oth estimation strategies can be usefully combined when the data gener
ating process is characterized by strong clustering in time and size.
We find that the extreme value theory is a useful starting point in de
tecting the appropriate distribution class. Once that has been achieve
d, the likelihood-based EM-algorithm is proposed to capture the cluste
ring phenomena. Clustering is particularly pervasive in actuarial data
. An empirical application to a four-year data set of Dutch automobile
collision claims is therefore used to illustrate the approach.