My view is that we should not fit Calendar Year in a CAT model unless we are aware of a genuine calendar year effect such as a change to policy terms and conditions or some other change we know will affect frequency and severity. My reasoning is as follows: • The data is likelyy to consist of a small number of CAT events over, say, a ten year period corresponding to a particular peril (e.g. flood) and number of events (e.g. 4 events) • Say the CATs will have occurred in 4 out of the 10 years • The differences in the Cat events are likely to be down to a difference in the risk profile of policies where the CAT events took place. For example, the first CAT event may have occurred in a wealthy area consisting of owner occupied, detached homes whereas another event might effect a city centre area consisting of apartments, terraced houses in a lower income area • By fitting calendar year, these differences in risk profile will be absorbed by the calendar year factor and will effectively dampen down the predictiveness of the other factors (and the variation of the predicted values) • We should fit calendar year only if it is not correlated with other factors we are fitting in the model Am I incorrect in my thought process?
I'm a bit confused by your question but I think if you look at the structure of the cat models as given in the notes you'll see that they answer all the points your making. The structure of the models with their different modules (event, hazard, etc.) allows the different moving parts that you've identified to be considered separately so hopefully we shouldn't end up with the kind of spuriously overfitted model you're trying to avoid.
