February 28th, 2019

Building for Resilience: How to Avoid a Catastrophe Model Failure: Part III

Posted at 1:00 AM ET


Imelda Powers, Global Chief Catastrophe Modeler


If resources permit, an insurer may carry out additional, micro level model suitability analyses - including a review of model hazards and vulnerabilities using the latest scientific literature and engineering studies. Guy Carpenter’s Model Suitability Analysis (MSA)® framework is designed to guide users through such an analysis.

It is not sufficient that a model work well for known risk locations, valuations and property characteristics. In practice, every portfolio contains uncertainties regarding these and other elements. Insurers should not underestimate the importance of a detailed consideration of these uncertainties.

(1) Location

(a)     Precise identification is often achieved via the model’s geocoding engine or national address databases, which use the longitude and latitude of a risk to identify its location. Loss estimates can be greatly impacted by the placement of the risks.

(b)     A disaggregation methodology approach is taken when a location cannot be tied to a specific longitude and latitude, but only a general geographic area (postcode, county or state, for example). There are several methods for assigning risk location when traditional tools do not provide clarity, but insurers must recognize that each produces different loss estimates.

Insurers will place greater focus on this component of model suitability under the European Union’s new General Data Protection Regulation (GDPR), which may cause them to use less specific location information in their analyses. This is especially important for flood modeling, where results are very sensitive to risk locations.

(2) Building characteristics - When a property’s physical characteristics are unknown, the model estimates a range of possibilities. For example, if the year of construction is unavailable, the model will assume a distribution over a range of years. The exhibit below illustrates the differences between an insurer’s actual data and regional census data in Portugal. In this case, the model’s assumed distribution, based on census data, is older than the insurer’s, so losses on risks with unknown construction years may be overestimated.

Link to Part I >>

Link to Part II >>

Link to Part IV >>

Click here to register to receive e-mail updates >>

AddThis Feed Button
Bookmark and Share

Related Posts