1. Catastrophe Reinsurance Ratemaking Midwestern Actuarial Forum Sean Devlin March 7, 2008

2. Slide 2 Agenda Model Selection Climate and Hurricane Prediction Model Adjustments Unmodeled Exposure Summary Q&A

3. Slide 3 Model Selection

4. Slide 4 Model Selection Major modeling firms  AIR  EQE  RMS  Other models, including proprietary Options in using the models  Use one model exclusively  Use one model by “territory”  Use multiple models for each account

5. Slide 5 Model Selection Use One Model Exclusively Benefits  Simplify process for each deal  Consistency of rating  Lower cost of license  Accumulation easier  Running one model for each deal involves less time Drawbacks  Can’t see differences by deal and in general  Conversion of data to your model format

6. Slide 6 Model Selection Use One Model By “Territory” Detailed review of each model by “territory” Territory examples (EU wind, CA EQ, FL wind) Select adjustment factors for the chosen model Benefits  Simplify process for each deal  Consistency of rating  Accumulation easier  Running one model involves less time Drawbacks  Can’t see differences by deal  Conversion of data to your model format

7. Slide 7 Model Selection Use One Model By “Territory” – An Example

8. Slide 8 Model Selection Use Multiple Models Benefits  Can see differences by deal and in general Drawbacks  Consistency of rating?  Conversion of data to each model format  Simplify process for each deal  High cost of licenses  Accumulation difficult  Running one model for each deal is time consuming

9. Slide 9 Climate and Hurricane Prediction

10. Slide 10 TCNA Adjustments - Climate 20052006 IntensityActualF’castVarActualF’castVar Named Storms 2712-15100%1013-17-33% All Hurricanes 157-988%58-10-44% Major Hurricanes 73-575%24-6-60% 20072005-7Cimate IntensityActualF’castVarActualF’castAverage Named Storms 1513-170%17.312-179.9 All Hurricanes 67-10-29%8.77-106.0 Major Hurricanes 23-5-25%3.73-62.6 Despite impressive science, the individual season predictions, the last few years was off the mark. However, actual and forecast are both above avg in total

11. Slide 11 TCNA Adjustments - Climate Option 1 - Find no credibility in the forecasts  Use a vendor model based on long term climate  Adjust the loss curve down of a vendor model that has increased frequency/severity  Use own model  A blend of the above

12. Slide 12 TCNA Adjustments - Climate Option 2- Believe that the forecasts are directionally correct  Credibility weighting between models in option 1 and a model with frequency adjustments  Adjust a long-term model for frequency/severity  Adjust long-term version of a vendor model  Adjust own model for frequency/severity  Combination of the above

13. Slide 13 TCNA Adjustments - Climate Option 3 - Believe completely in the multi-year forecasts  Implement a vendor model with a multi-year view  Make frequency/severity adjustments to a long term vendor model  Adjust own model  Blend of the above

14. Slide 14 TCNA Adjustments - Climate Option 4 - Believe completely in the single year forecasts  Implement seasonal forecast version for a vendor model  Adjust vendor model for frequency/severity  Adjust internal model for frequency/severity  Combination of the above

15. Slide 15 Model Adjustments

16. Slide 16 TCNA Adjustments – Frequency/Severity Adjust whole curve equally  Ignores shape change  Treats all regions equally Adjust whole curve by return period/region

17. Slide 17 Modeled Perils – Other Adjustments Actual vs. Modeled – look for biases (Macro/Micro)  Model recent events with actual portfolio  More confidence on gross results, but some insight may be gained on per risk basis  One or two events may show a material upward miss. Key is to understand why. Exposure Changes / Missing Exposure/ITV Issues  TIV checks/audits  Scope of data – international, all states & perils  Changes in exposure, important for specialty writers

18. Slide 18 Modeled Perils – Actual vs Modeled

19. Slide 19 Modeled Perils – Other Adjustments Other Biases in modeling LAE Fair plans/pools/assessments – know what is covered by client and treaty prospectively FHCF – Reflect all probable outcomes of recovery Storm Surge Demand Surge  Pre Event  Post Event

20. Slide 20 “Unmodeled” Exposures

21. Slide 21 “Unmodeled” Exposures Tornado/Hail Winter Storm Wildfire Flood Terrorism Fire Following Other

22. Slide 22 Unmodeled Perils Tornado Hail  National writers tend not to include TO exposures  Models are improving, but not quite there yet  Significant exposure  Frequency: TX  Severity: 2003: 3.2B - #12 all time 2001: 2.2B - #15 2002: 1.7B - #18  Methodology  Experience and exposure ate  Compare to peer companies with more data  Compare experience data to ISO wind history  Weight methods  Percentile Matching with model

23. Slide 23 Unmodeled Perils

24. Slide 24 Unmodeled Perils

25. Slide 25 Unmodeled Perils Winter storm  Not insignificant peril in some areas, esp. low layers  1993: 1.75B - #19 all time  1994: 100M, 175M, 800M, 130M, 455M  1996: 600M, 90M, 395M, 735M  1999: 775M, 575M  2003: 1.6B  # of occurrences in a cluster?????  Possible Understatement of PCS data  Methodology  Degree considered in models  Evaluate past event return period(s)  Adjust loss for today’s exposure  Fit curve to events  Aggregate Cover?????

26. Slide 26 Unmodeled Perils Wildfire  Not just CA  Oakland Fires: 1.7B - #20 All time  2003 Fires: 2B  Development of land should increase freq/severity  Two main loss drivers  Brush clearance – mandated by code  Roof type (wood shake vs. tiled)  Methodology  Degree considered in models  Evaluate past event return period(s), if possible  Incorporate Risk management, esp. changes  No loss history - not necessarily no exposure

27. Slide 27 Unmodeled Perils Flood  Less frequent  Development of land should increase frequency  Methodology  Degree considered in models  Evaluate past event return period(s),if possible  No loss history – not necessarily no exposure Terrorism  Modeled by vendor model? Scope?  Adjustments needed  Take-up rate – current/future  Post TRIA extension issues  Other – depends on data

28. Slide 28 Unmodeled Perils Fire Following  No EQ coverage = No loss potential? NO!!!!!  Model reflective of FF exposure on EQ policies?  Severity adjustment of event needed, if  Some policies are EQ, some are FF only  Only EQ was modeled  Methodology  Degree considered in models  Compare to peer companies for FF only  Default Loadings for unmodeled FF  Multiplicative Loadings on EQ runs  Reflect difference in policy T&Cs

29. Slide 29 Unmodeled Perils Other Perils  Expected the unexpected  Examples: Blackout caused unexpected losses  Methodology  Blanket load  Exclusions, Named Perils in contract  Develop default loads/methodology for an complete list of perils

30. Slide 30 Summary Don’t trust the Black Box  Understand the weakness/strengths of model  Know which perils/losses were modeled  Perform reasonability checks  Add in loads to include ALL perils  Reflect the prospective exposure