Claim dependence in credibility models

Existing credibility models have mostly allowed for one source of claim dependence only, that across time for an individual insured risk or a group of homogeneous insured risks. Numerous circumstances demonstrate that this may be inadequate and insufficient. In this dissertation, we developed a two-level common effects model, based loosely on the Bayesian model, which allows for two possible sources of dependence, that across time for the same individual risk and that between risks. For the case of Normal common effects, we are able to derive explicit formulas for the credibility premium. This takes the intuitive form of a weighted average between the individual risk's claims experience, the group's claims experience and the prior mean. We also consider the use of copulas, a tool widely used in other areas of work involving dependence, in constructing credibility premiums. Specifically, we utilise copulas to model the dependence across time for an individual risk or group of homogeneous risks. We develop the construction with several well-known families of copulas and are able to derive explicit formulas for their respective conditional expectations. Whilst some recent work has been done on constructing credibility models with copulas, explicit formulas for the conditional expectations have rarely been made available. Finally, we calibrate these copula credibility models using a real data set. This data set relates to the claims experience of workers' compensation insurance by occupation over a 7-year period for a particular state in the United States. Our results show that for each occupation, claims dependence across time is indeed present. Amongst the copulas considered in our empirical analysis, the Cook-Johnson copula model is found to be the best fit for the data set used. The calibrated copula models are then used for prediction of the next period's claims. We found that the Cook-Johnson copula model gives superior predictions. Furthermore, this calibration exercise allowed us to uncover the importance of examining the nature of the data and comparing it with the characteristics of the copulas we are calibrating to