11.10 Conclusion

Goal of the paper was to test the predictive accuracy of various models, both mean and distribution of outcomes

• Not on the reserve estimate for individual insurers

Bayesian MCMC models can be developed to overcome shortcomings in existing models (e.g. how LCL and CCL loosens some of Mack Chainladder’s key assumptions)

11.10.1 Results Summary

Incurred Data

Mack understates variability as it assumes AYs are independent

CCL introduces AY correlation and does relatively well

Paid Data

Mack and ODP were biased high as well as CCL

There were change in environment that is not captured

• Calendar year trend: LIT and CIT still biased high

• CSR: significantly less bias than LIT and CIT (except for PA still failed)

Mack and ODP did better than CCL, LIT and CIT

11.10.2 Final Comments

Results were for specific annual statement year 1997

• Possible the speed up of claims settlement was specific to the period \(\Rightarrow\) CSR could potentially useless for another year

Could use more narrow priors to incorporate knowledge of insurer’s business operation and obtain superior results