While catastrophe model vendors’ focus is to build deterministic and probabilistic models, they have also provided inputs for real-time event loss estimates for many years. Model vendors usually release their loss estimate tools within one or two weeks of an event’s occurrence. At that time, the hazard parameters are sparse, unverified and conflicting over multiple sources, according to Imelda Powers, Senior Cat Management Advisor – Asia Pacific, Guy Carpenter.
Often the information on the damage is patchy from news, social media, satellite and aerial images. The model vendors will likely deploy reconnaissance teams to survey the damage shortly after the occurrence. However, these surveys offer only a sample of the damages – some sites may be inaccessible or the surveyors may not obtain permission to enter damaged properties. The severe damages are visible, but there are many hidden exterior and interior damages. These factors contribute to the underestimation. In some cases, the surveyors may find damage in areas where there is no instrument recording the hazard parameters, leading them to impute the hazard value based on observed damage and nearby instrument readings. Based on the derived hazard footprint, the model vendors deliver several complementary products to help insurers assess their liabilities:
- Extent maps with intensity bands map the affected areas with hazard intensities, enabling users to analyze the accumulation, support its claims triage process and estimate losses. The larger the map extent, the more claims it captures, but it may also capture false claims. Guy Carpenter sources these maps from specialists who can provide hazard footprints more promptly than the catastrophe vendors. The table below offers comparisons from multiple sources for the Australian Townsville Flood event that took place on February 4, 2019.
- Custom made hazard scenarios allow users to obtain estimated losses by running a vendor cat model.
- Similar events from a model’s catalog of stochastic events enable users to estimate modelled event losses quickly. The catalog is optimized for computational efficiency, so it is not uncommon for the actual event to be quite different from the available stochastic events.
All of these products rely on robust hazard data and damage reports. However, the observed damage is partial. It can only be compensated by experienced surveyors/claims adjusters who can extrapolate based on the observations with great expertise, by taking the incurred but not reported loss into account.
The challenge of getting good hazard and damage data in a real-time event differs from that of building the deterministic and probabilistic models. With this understanding, insurers should seek additional inputs to complement the vendors’ estimates. As an example, Guy Carpenter’s GIS platform, GC AdvantagePoint®, hosts real-time hazard footprints from multiple sources. Overlaying them with the exposures provides further inputs. Additional analytics to estimate non-unmodeled risks and subperils will fill out the picture. Integrating all these inputs and taking internal damage surveys and initial claims into account will help an insurer form its view of the event loss.
Appreciating the way the real-time event products are derived, we can see that sometimes underestimations may be more reflective of the limitations in the imperfect inputs than of the model fitness. These limitations are persistent and they cannot be compensated by model updates to include missed subperils, risk types or more extreme hazards.