9.11 Model Testing
Important to test the model results against actual (see Meyers)
Challenges:
Don’t have underlying distribution to compare with
Can’t wait for 10 years to see how it forecast
General Insurance Reserving Oversight Committee test:
GIROC created triangles that met the assumptions of Mack and ODP
30K 10 \(\times\) 10 triangles were created
Mack: losses were above the 99th percentile about 8-13% of the time \(\Rightarrow\) Mack underestimates tail events
ODP from England & Verrall: losses were above the 99th percentile about 3% of the time
- This does not include the residual adjustments
Future testing:
- Create datasets from claims transaction data and use them for model testing
9.11.1 Future Research
Test ODP bootstrap on realistic data from CAS loss simulation model
See how adjustments discussed here improve predictive power
Expand ODP bootstrap with
Munich Chainladder for incurred/paid
Claim counts and average severity
Different residuals (e.g. deviance or Anscombe residuals)
Select weights using Bayesian methods by AYs
Other risk analysis measures and use for ERM
SII requirements
Research in correlation matrix (difficult to estimate)