Despite technical and scientific advances in catastrophe models during the past 20 years, observers say the most significant developments appear to be in the form of human capabilities.

That's because a shift is under way among insurers and reinsurers to better understand and more effectively interpret the results of models that simulate the effects of storms and other perils on their risk portfolios.

In addition, many reinsurers who use cat models to help price and structure reinsurance contracts have developed in-house experts to analyze the models and apply their own expertise to the results.

Experts say cat models have become a vital tool for the industry, although events of the past several years have shown they are not perfect predictors of a reinsurer's loss potential from catastrophes.

“In our view, cat models are very valuable because they offer a consistent measurement of risk,” said Carl Hedde, head of risk accumulation at Munich Re America in Princeton, N.J. “But it takes a sophisticated user to adjust that measurement and come up with what they ultimately think is the risk.”

Like many reinsurers, Munich Re America uses a multiple model approach, combining “vendor models” offered by the three main catastrophe modeling firms—AIR Worldwide Corp., Risk Management Solutions Inc. and EQECAT Inc.—with proprietary models developed in-house to better understand its book of business, Mr. Hedde said. The company also relies on a staff of scientists, meteorologists and engineers to help evaluate scientific data and apply its own quantitative techniques “so we can match up our view of risk with what the models are telling us,” he said.

Cat models were criticized widely for underestimating losses from hurricanes Katrina and Rita in 2005, underscoring the need for more detailed “input” data, also called “exposure data.” This includes details about the properties—including exact address, construction materials and type of business—that is required to produce reliable model results of insured losses.

High catastrophe losses in 2004 and 2005 also revealed what some observers said was the industry's over reliance on catastrophe models. “Too many companies have used them as a way to arrive at an exact answer, but they are not intended to be so precise. They can only give you an indication of what might happen,” said Andrew Castaldi, senior vp and head of catastrophe perils, Americas, at Swiss Reinsurance Co. in New York.

But now, users are beginning to put the model results in the proper context, experts say.

“In the past, many users just ran the models once and accepted one result. Now, top reinsurers have built up huge networks of computing resources. They can run the models in parallel, applying numerous stress tests to obtain 20 or 30 alternative perspectives to the model output,” said Robert Muir-Wood, London-based chief research officer at Newark, Calif.-based RMS.

Potential stress tests include running the models with variations on the “secondary modifiers,” a technical term used to categorize properties included in a book of business. This could include building characteristics, such as whether it has hurricane shutters or the age of its roof, that affect its vulnerability to loss. “They're turning on and off the various switches to see how it might shift the results,” Mr. Muir-Wood said.

In addition, “reinsurers are able to run the models in real time so they can see how every new piece of business they are writing fits alongside with the business they already have in their portfolio, and expand or contract accordingly” to avoid concentration of risk, said Mr. Muir-Wood.

“Reinsurers are very interested in building out a more robust system to proactively manage their catastrophe risk,” said Chris McKeown, CEO of global analytical and specialty practices at Guy Carpenter & Co. L.L.C. in New York.

“We're spending a lot of time with clients emphasizing the underwriting systems and asking, "How do you house that vendor model output in your underwriting system? And how does that interact with your capital model?'” It's a more dynamic approach that allows users “to match up the asset strategy with the liabilities as (they) actively write the business,” Mr. McKeown said.

Reinsurers that smooth out the model outputs will be better equipped to evaluate the overall level of capital required to pay potential claims and allocate capital more effectively across their business, experts say.

But data quality continues to be an issue, some reinsurers say. Extra precautions are necessary because reinsurers rely on exposure data from cedents. Although improving, exposure data often is viewed as inadequate due to the absence of a standard approach in capturing and cataloging the data, experts say.

In addition, “everyone makes different assumptions on how they portray their exposures,” said Munich Re America's Mr. Hedde.

To address the issue, reinsurers have been taking an active role in devising methods to measure the quality of the data and then adjust the models, Mr. Muir-Wood said. “It's a big concern because, post-event, many companies end up discovering that there were a lot of undervaluations on the properties, which meant they weren't collecting the correct premium for that business,” he said.

Mr. Hedde said reinsurers also engage in “data quality reviews” with many cedents to better understand their approach to collecting exposure data.

The models also are limited by insufficient scientific data. While modelers may have a good handle on catastrophe-prone Florida, other regions like the U.S. Northeast have little historical data. In such cases, modelers build on catastrophe experience globally, experts say.

Efforts also are under way to help users understand and adjust for inherent uncertainties in models.

“Models are based on many assumptions and there is quite a bit of uncertainty inherent in the models and, hence, the model loss estimates,” said Karen Clark, president and CEO of Boston-based Karen Clark & Co. and developer of the first hurricane catastrophe model.

Catastrophe models don't always account for other sources of loss caused by an event, experts say. For example, some previous hurricane models didn't account for or didn't adequately account for factors such as storm surge, demand surge, inland flooding and business interruption losses. Although many of these factors are being addressed in updated models, users have recognized the limitations and started incorporating their own view of these risks into the modeling process, experts say.