Quintile Test
Hi there,
Could you please elaborate on step (4) of the Quintile Test in the Battlewiki?
"Calculate Vquintile and Vquintile, holdout. Calculate v^quintile and vh^ as weighted averages of vi^ using the TT injury counts for the weights."
I guess this is related to the question that I asked in another thread, if the relativities in the "Predicted" column in the multi-dimensional table are calculated using vi (the estimated population injury type mean ratio to TT for class i), why do we need to multiply by the pure premium relativity for the HG containing the class?
Thanks!
Comments
A key idea of Couret & Venter's paper is the TT counts serve as an exposure base for the lower frequency but higher severity accidents. This is why we measure everything relative to TT claims because we can then scale.
You have the set of all classes in a hazard group. You divide the data into a training set and a test set. This is usually done by choosing a subset of accident years for the training and a different subset of years for the test (odd vs even years etc.) because you may not have enough risks in each class to split out randomly.
You order the classes in the training set according to the number of TT claims and then group into quintiles based on the TT claims.
Quintile 2 may contain three classes, 1,2,3 and these have the following Fatal claims experience in odd years (my training set): 2,4,6. TT exposures are 500, 2,000 and 1,000. This gives v_quintile2 as 4.286. Repeat this calculation on the holdout quintile 2.
These are then used as the actual experience for the quintiles and can be compared against the various methods to allow us to determine which method is best (lowest sum of squared errors).