It takes time to clarify the hazard measurements at locations with instruments. Hurricane Michael hit Florida on October 10, 2018 as a Category 4 event. Six months later, on April 19, 2019, the National Hurricane Center upgraded Michael to Category 5 because reanalysis showed that its landfall wind speed was 160 MPH instead of 155 MPH.
After some time, the vendor entered a historical event hazard into its model based on
- Hazard measurement from locations with instruments.
- Filled hazard imputed from observed damage for areas without any measurements, taking into account nearby measurements, if available.
- Claims data from contributing insurers. These are limited to losses above the deductibles.
An error in filling out the hazard of a historical event does not imply model shortcoming unless there is evidence that historical hazard reconstructions errors are systematic and therefore influence the proper development of the model, according to Imelda Powers, Senior Cat Management Advisor – Asia Pacific.
When a model user applies the reconstructed hazard footprint to a portfolio, the modeled loss may be higher or lower than the actual claims, indicating a possible difference of the portfolio’s vulnerability to that of the model. Or, the hazard footprint may not be very good for locations for this particular portfolio. Unless the insurer’s claims form a majority of the loss validation, it is often necessary to do some customization in using a model.
We should be mindful that there is a difference in the way hazard footprints of historical and stochastic events are constructed. The former is individually customized, while the latter is generated systematically. Therefore, if we discover a tendency for a stronger wind field on one side of historical storm tracks, as an example, we cannot assume this characteristic is also in the model’s stochastic events.
In summary, a vendor’s real-time event loss tools are built with limited observation of hazard and damage. The difference between the reconstructed historical event loss estimates and actual claims can provide hints of model fitness. However, additional analyses are necessary to make a final conclusion due to the different ways hazards of historical and stochastic events are constructed.
Guy Carpenter’s Model Suitability Analysis (MSA)® has a rich set of tests to validate historical and stochastic event hazard and vulnerability to allow users to reach conclusions on model fitness. An example of validating the hazard in the stochastic model is a test of the return period of the intensity for locations where historical records are available. Figure 1 shows that the vendor intensities (in red) at the weather station at Heathrow, U.K. are within bounds of the reference scientific curves. Depending on the outcome of these assessments, companies will be able to make adjustments to model outputs and their risk management strategies, going forward.