Predictive modeling—a brave new world

True impact is yet to be known

By Tom Wetzel

Predictive modeling has become a well-established, widely implemented practice within the industry, though pitched battles in various states over its use will continue this year. The debate, however, will change.

Predictive modeling is a process by which a statistical model of future behavior is created. Predictive analytics, of which predictive modeling is part, involves the mining of data to forecast probabilities and trends. Modeling is used in dozens of ways—to qualify sales leads, target agents, pinpoint prospects, improve hit ratios, decrease acquisition costs, screen applicants, identify fraudulent claims, and better manage catastrophes. Until recently, the data collected has been largely the structured variety, or what one finds in databases in which each bit of information has an assigned format and significance.

Unstructured data is what one finds in e-mails, loss control reports, PowerPoint presentations, voice mail, phone notes, and photographs. Unstructured data comprises at least three-quarters of an organization’s knowledge stores. The mining of such data is far more complex and expensive; however, experts say it will yield more accuracy in specific functions such as underwriting and fraud detection and more unders-tanding of risk as a whole.

“It’s too early to know what the true impact of this information will be,” says Dale Halon, vice president of ISO Innovative Analytics. New sources of data are being found and mined every day, some structured and some unstructured. They will be used for many business purposes of predictive models, including perhaps to replace credit scoring one day, which many companies would like to do. The ultimate objective, however, is to make better and more predictable decisions using historical data in order to improve pricing, marketing, claims and sales decisions.

“Today, predictive modeling is being used more holistically,” says Halon. “There is a trend toward companies taking all the data and looking at all the pieces together to understand the entire product lifecycle. It’s not just about establishing the right price; it involves customer satisfaction and retention and many other concerns.”

Halon adds that insurance company databases are also designed primarily to capture transactions, not just data.

“The big question is, what will this new information be used for?” he says. “Most of the unstructured data is in claims files. Right now, we are doing loss cost projections by peril. It’s really just the tip of iceberg.”

“It will be a slow integration,” says Stuart Rose, global insurance marketing manager for SAS, one of a handful of companies that create soft-ware that can mine unstructured data.

“The only companies that are starting to mine unstructured data are some of the larger companies so far,” Rose says. “The software is expensive and is still not widely used. Companies can ease into the process by studying one line or function, such as call centers or e-mails.”

Some companies do not divulge many details about how they use predictive modeling in underwriting for fear of giving up a competitive advantage. Many, however, appear to be open about discussing their use with producers.

“Our carriers held meetings with producers long before they started using insurance scoring,” says Cappy Stults, president of Allen & Stults, an all-lines general lines insurance agency in Heightstown, New Jersey. Stults adds that he has to make sure to keep track of the different scoring systems. “One will have a AA—ZZ scale; another will have 0-40 scale.”

The propriety element

Robert Whitlock, senior vice president and chief underwriting officer of Pennsylvania-based Harleysville Insurance, acknowledges that some aspects of predictive modeling are proprietary, but he stresses the importance of maintain-ing open lines of communication with agents.

“We do not have a formal policy but we have been very forthcoming about our use of predictive modeling, Whitlock says. “We do treat the details as proprietary, but we give our underwriters a lot of discretion in discussing the types of characteristics and encourage a dialogue with agents.”

He adds that when his company does field questions about the practice, “It is less about modeling and more about how did you come up with this price?”

Whitlock says the predictive models his company uses are embedded in the system. “It gives us the ability to match price with quality,” he says.

Stults says many agents were at first skeptical of predictive modeling. “We fear what we don’t know,” he says, but he adds that modeling has proved to be of great benefit to his agency’s competition with direct writers. “It has kept us in the ball game, enabling us to offer a more competitive rate.”

Even with accolades, there remains some tension about keeping agents fully informed about how companies use the data.

“The issue of insurer communica-tion with agents is still a problem,” says Patricia A. Borowski, CPIW, CAE, senior vice president of Government/Regulatory Affairs for the National Association of Professional Insurance Agents. “As a business partnership, there is an obligation to make it work for both sides. When a carrier issues a termination of coverage 30 days into the risk, an agent is lucky if they are notified.”

With respect to insurance scoring, Borowski says that agents need to know what factors are to be used and what financial impact it will have. “We need to know that the process is handled in a fair, reasonable, and empirically demonstrated manner. It is our business. If the person we place is not satisfied, we need to know.”

The author
Tom Wetzel is a seasoned public relations counselor and principal of Thomas H. Wetzel & Associates, Inc. He has nearly 30 years of insurance communications experience. Prior to forming his marketing communications firm, he held positions with the Michigan Association of Insurance Companies and the Insurance Information Institute.