9.5 Variations on the ODP Bootstrap
Reason to use paid data:
For insurance risk it is best to focus on the claim payment stream:
It measures the variability of the actual cash flows that directly affect the bottom line
Case reserves temper the volatility
Changes in case reserves and IBNR reserves will also impact the bottom line, but to a considerable extent the changes in IBNR are intended to counter the impact of the changes in case reserves
To some degree, then, the total reserve movements can act to mask the underlying changes due to cash flows
Reason to include case reserves:
- Case reserves contain valuable information about potential future payments
9.5.1 Bootstrapping the Incurred Loss Triangle
2 approaches to model the unpaid loss distribution using incurred loss triangle
Method 1: Modeling the incurred data and convert the ultimate values to a payment pattern
Run the paid and incurred data model in parallel
For each iteration and each AY individually:
Use the payment pattern (from paid model) to convert the ultimate values (from incurred model) to a payment stream
Method 1 Advantages:
We improve the ultimate estimates by incorporating the case reserves while still focusing on the payment stream for measuring risk
Which effectively allows a distribution of IBNR to become a distribution of IBNR and case reserves
Can make it more sophisticated by correlating some part of the paid and incurred models (e.g. the residual sampling and/or process variance portions)
So that if we have large payment @ an older age, the incurred should be large as well
Method 2: Applying the ODP bootstrap to the Munich chain ladder model
- See Liu and Verrall (2010)
Method 2 Advantages:
Don’t have to model the paid loss twice
Explicitly measuring and imposing a framework around the correlation of the paid and outstanding losses
9.5.2 Bootstrapping the BF and Cape Cod Method
ODP issue: Distribution for the most recent AYs can produce results with more variance than you would expect when compared to earlier AYs in the actual data
Due More LDFs are used to extrapolate the sampled values for the most recent accident years and the random samples of incremental values
Similar to the leverage effect of the deterministic chainladder
Solution: Incorporate BF or Cape Cod
Have the a-priori be stochastic e.g. draw the BF a-priori from a distribution or apply Cape Cod to each simulated triangle
More complicated approach is to modify the underlying assumptions of the GLM framework which would results in a completely different set of residuals (this is beyond scope)