Building its models on 100 years of historical data–which included high- and low-frequency periods–AIR worked to convince insurers not to base their decision-making on shorter-term views, because those would underestimate the risks.

“Now we're in an interesting position because we're in a warm phase with higher activity,” and companies are expecting AIR to “bump up” model frequency assumptions.

“We can't do that because our philosophy from the beginning was to give the long-term” perspective, she said, adding that “it would be a biased view” to shift gears during a high-frequency period. She said if insurers were to raise prices based on warm-period data, then when a cool period returned, “they would have to drop them significantly.”

“Our frequency assumptions aren't going to change,” she said–noting, however, that AIR, in addition to updating its long-term view in the next model release, will also provide a second near-term hurricane catalog. That catalog–a five-to-10-year view–will reflect the warm period of sea surface temperature characterized by more frequent and intense storms. Ms. Clark believes this view is appropriate for near-term decisions, such as how much reinsurance to buy this year.

A third view–an annual forecast based on the work of climatologists–will also be in the update, she said, noting that AIR has been providing this for the last few years.

Ms. Clark said the spring 2006 release will also reflect the fact that a very intense storm can have a larger radius than previously thought–a lesson of Katrina.

Finally, in the next version, AIR will be increasing the vulnerability factors (damage ratios) of two commercial construction types–light metal and wood-frame buildings. Detailed analysis of individual claims for clients from last year's storms revealed that these buildings perform poorly and are much more vulnerable than the models previously indicated. (Residential damage functions won't change, since the model already reasonably reflected damageability for those, she said.)

Ms. Clark, who spoke to NU here in Bermuda during the recent World Insurance Forum, was one of several experts at the conference to point out that the results from catastrophe models of different vendors vary widely. What causes such differences, given that all the firms start with essentially the same information?

While hurricane models, for example, all begin with databases of more than 100 years of hurricane activity, “the real differences are in implementing details and integrating the components,” she said, identifying them as hazard and engineering components, and a financial piece to apply policy terms and conditions. “There's no real blueprint as to how to do that. Every modeler will make different sets of assumptions that are equally valid,” she added.

Using earthquake models to illustrate the point, she said that to estimate ground motion, modelers use something called attenuation functions. With different attenuation functions available in the scientific literature, each modeler chooses which set to use, and how to weight them.

They then marry ground motion functions to engineering damageability functions, she said, noting that each modeler has its own proprietary damage functions, adding more possibilities for differences to emerge.

Even applying policy conditions–sublimits, deductibles, reinsurance–isn't a straightforward matter. It would be, if models weren't probabilistic. “If I know that the damage on this building is 10 percent, then it's easy to apply policy conditions. But even if the expected damage ratio is 10 percent, every building is not going to have 10 percent damage. Some will have zero; some 100 percent,” she said, explaining that becomes a complicated process to decide how to carry through policy terms.

Ms. Clark said the next version of AIR's model will be released in late May or early June. “It's not really a big learning curve to use the model,” she added–explaining that after installation, the information technology department and internal catastrophe modeling teams will likely work to test the models and see how results compare to prior results. “It could be a matter of a week or two,” she said, noting that a more typical timeframe from release to company use is six weeks.

With some companies announcing they have already internally adjusted catastrophe loss assumptions in anticipation of model updates from all the modelers, are they close to AIR at this point?

“It's hard to say. I have heard some companies say they are going to make some interim adjustments as to what they're going to price,” Ms. Clark said. “But at the end of the day, the pricing is what the market will bear…The models don't determine the pricing. The market does.”

Another modeler–RMS–gave some early indications late last year that its frequency assumptions would likely increase by one-third in the next release (see NU, Dec. 5, 2005, page 20). And during a December teleconference, Guy Carpenter representatives said that loss results in a near-term supplemental model from a third modeler–Eqecat–will more than double.

Ms. Clark didn't want to offer any such statistics, preferring to wait until AIR's research is complete.

Given the differences in results of different models, how do insurers select which ones to buy and use?

Ms. Clark said she believed that in the early days of modeling–15 years ago–senior decision-makers at companies were very concerned about the accuracy and reliability of the models. “There was a lot of scrutiny on the model results to make sure that they made sense.”

Over time, as the models have become more complex, she suggested that “it's become difficult for companies to see the differences in the science incorporated in the models, and somehow, they stopped scrutinizing the results and whether the changes actually made sense.”

In addition, she said that even where a company is convinced that one model gives more reliable results for a peril it is insuring, they might make decisions based on ease of use. “'My infrastructure is already set up to use this model,' or 'I typically get data in this format,' or 'This one imports data faster,'” are among the considerations Ms. Clark has heard expressed.

“There's some validity to that, but really what they're buying is the information. If the information is not credible, it doesn't matter how easy it is to use,” she said. “The market has gone from one extreme to the other, and hopefully now will go back to basing decisions on the quality of the information. That's what they're buying.”

“They're not just buying a piece of software…It's not just a data processing system. This is information that companies are using to make billion-dollar decisions,” she added.