Newsletter Sign-up

Senior Vice President
Karen Clark & Co.

Fleshing Out Models

The problem was not the models themselves, just how they were being used.


Most models are beautiful, but too thin and fragile to be meaningful outside their own rarefied world. What they need is more substance, more meat on their bones. That has long been the case in fashion, but also in models used for forecasting catastrophic risk, especially hurricane and earthquake exposure.

Responding to inquiries by underwriters, Glen Daraskevich, senior vice president at Karen Clark & Co. in Boston, led the development of characteristic events (CEs) that do not replace prevailing cat models, but significantly enhances the practical application of those models.

Specifically, Daraskevich and his colleagues found that the major challenge in managing cat risk is the lack of stable and robust risk metrics for managing and monitoring the loss potential over time. Existing commercial models -- the primary tool -- provide a lot of numbers, but those numbers are highly volatile from model to model and, over time, through updates. That volatility makes it difficult to develop and implement consistent underwriting and pricing strategies to account for and to control risk.

"The problem is not with the models themselves," said Karen Clark, founder and CEO of the company, "but with the way they have been used. They were never designed to predict specific events, just to predict overall losses. But many companies were relying on those models for their PML 1/100 and 1/250 events.

"That is because the model output shows a false precision, down to the pennies. But that is an illusion of accuracy," she said. "The models are highly volatile, and when they would be updated every year, the output numbers would change dramatically."

Clark explained that the problem has been around for years because some people do not realize that the models were not designed for the demands being put on them, while others do realize the problem but cannot find an alternative.

"We heard that over and over, until we decided to do something about it," she said.

As a practical example, Daraskevich noted that the northeastern United States has not suffered a full hurricane in about 20 years (Irene, which roughed up the Eastern Seaboard last summer, was a tropical storm by the time she made landfall in New Jersey). "So there has been no new storm data in two decades. But every year, the models get updated and the output changes. It is very frustrating for the companies trying to write windstorm coverage in the region."

Instead of random simulations, Daraskevich said, the CEs pay more attention to the key return periods for perils and regions. "We asked, 'What is the 1/100 event for that region?' " he explained. "Once that is established, we hold it steady so that underwriters and risk managers have a consistent yardstick."

One early adopter of the model said "it makes us feel more confident about managing our enterprise risk because we are looking at that and saying, 'What is our maximum probability of loss? Should we be adding business or subtracting business?' The analysis went beyond the cost of reinsurance. It got to questions about where we have too much business and where we have not enough business. Once we establish that, then can we grow without increasing our total exposure."

Another user added, "We have been trying to figure out how to get some answers that could give us a good feel for what our exposure spikes are and what causes those spikes. CEs take 10,000 storms and give us averages for every 20 years and every 100 years. We could have a loss like this or that. That enables us to determine where should we write less business and where we should write more."

Clark said there was no eureka moment for CEs: "It was an evolutionary development to address a long-standing need in the market. The only quantum change was going outside the models for the CEs. Then we had to develop the software to use it, and that took some time."

She stressed that "CEs do not replace the existing models; we use the same data. It simply enhances the output so that risk managers can use it as a meaningful planning tool. They can also use it to see the benefits of risk mitigation. Without CEs, using only the model, that is a moving target."

The first CEs were released early in January and covers U.S. hurricanes. Daraskevich said the response has been very enthusiastic, and the next steps are to create a full software package with multiple functions. After that, the program will be expanded to other perils, including U.S. earthquakes and European windstorms.

--By Gregory DL Morris