10.1 Introduction
Uses Bayesian techniques to allow incorporation of expert opinion and also maintain the integrity of the prediction error
i.e. how expert opinion from sources other than the specific data set under consideration can be incorporated into the predictive distributions of the reserves
Take into account prior knowledge in setting reserves
e.g. Adjusting data to reflect changes in benefits or claims handling, or select L-year average LDFs
Calculating prediction error using prior knowledge
Use Bayesian to take into account our a-priori and the strength of that a-priori (credibility)
Paper focus on use of a-priori knowledge for LDF selection and BF Method
Development of MCMC techniques has made Bayesian methods much easier
Makes it easier to find the posterior distributions for future observations (defining the Bayesian model is always easy)
MCMC breaks down the simulation process into a number of simulations that are easy to carry out
\(\therefore\) This solves the common Bayesian problem of difficulty in finding the posterior distribution (as they can be multidimensional)
MCMC doesn’t simulate all the parameters at once, it use the conditional distribution of each parameter, given all the others
\(\therefore\) Reducing the simulation to a univariate distribution
Markov chain is formed because each parameter is considered in turn, and it is a simulation-based method
\(\therefore\) MCMC
10.1.1 Notation
For a \(n \times n\) triangle
Claims
- Incremental claims for AY \(i\) and age \(j\): \(c_{ij}\)
\[\{ c_{ij} \: : \: j=1,...,n-i+1 \: ; \: i = 1,...,n \}\]
- Cumulative claims
\[D_{ij} = \sum_{k=1}^j c_{ik}\]
- Ultimate losses
\[D_{in} = \sum_{k=1}^n c_{ik}\]
Remark. We only consider forecasting losses up to the latest development year \(n\)
- It is possible to extend this to allow a tail factor but not in this paper
Row Parameters
- Expected ultimate losses for AY \(i\): \(x_i\)
\[x_i = \mathrm{E}[D_{in}]\]
Column Parameters
- Expected % reported in each period: \(y_j\)
\[\sum y_i = 1\]
- Expected reported to date: \(p_j\)
\[p_j = \sum_{k=1}^j y_k\]
- Expected development from \(D_{ij-1}\) to \(D_{ij}\):
\[\lambda_j = \dfrac{p_j}{p_{j-1}}\]
\[\{\lambda_j \: : \: j = 2 ,...,n\}\]
- Weighted average LDFs: \(\hat{\lambda}_j\)
\[\hat{\lambda}_j = \dfrac{\sum_{i=1}^{n-j+1} D_{ij}}{\sum_{i=1}^{n-j+1} D_{ij-1}}\]