IN THE PROPERTY and casualty (P&C) insurance business, nothing's more important, or challenging, than knowing your customers and the items they want to insure. P&C insurers figure the pricing of their policies based on the calculated risk of the property being insured -- a type of car, for instance -- and the owner. Normal insurance practices try to identify various groups and their associated risk, such as the infamous male drivers under age 25 who drive sports cars: high risk.

But since insurance companies are also subject to competitive pressures, it's not as simple, once the risk is calculated, as "spreading the risk." If a company overcharges low-risk customers in an effort to balance higher-risk ones, it may lose those low- risk -- and highly profitable -- customers to competitors who more accurately price their policies. And if a company undercharges high-risk customers, it may end up attracting more high-risk customers away from competitors, at once damaging its bottom line while improving the profitability of the competition!

Fortunately, insurers collect reams of client data that can help identify various risk groups, not so much to avoid certain groups, but to fairly and profitably price policies for various types of risk. Unfortunately, identifying risk groups has been largely a speculative process: develop a hypothesis, then check to see if the data supports it.

The IBM team set about determining the best approach to mining this data. By combining its own expertise with the domain knowledge of Farmers' actuaries and market analysts, the team was able to focus the mining attempts on a specific set of policy types in a certain region (known as books of business in industry parlance). They developed the necessary algorithms to search through the data and confirm known sets of "rules" (such as "male drivers under age 25 who drive a sports car have a claim frequency of 25% and an average claim amount of US$3200"). As part of the process, they also hoped to discover previously unidentified rules, and hence, risk groups.

One particular challenge the group faced was in designing algorithms that allowed for simultaneous modeling of both claim frequency and average claim amounts. To mine for either statistic separately, and then combine the results in figuring the cost to insure, would lead to erroneous conclusions since each initial result might be based on different sub-populations of customers -- the equivalent of polling the hungriest group of people in a room, then the thirstiest, and trying to base a lunch order for the combined group based on that information. Only by mining for risk groups while modeling both features simultaneously could accurate and verifiable results be guaranteed.

The Underwriting Profitability Analysis (UPA) solution developed by the team, which was run at IBM Research in Yorktown and also on Farmers' RS/6000 machine at the company's headquarters, provided some incredible "nuggets" of knowledge: the company had far more market segments and sub-segments than previously known, including a few that were counterintuitive. Farmers found that covering a certain type of "high-risk" sports car was not always so risky -- in fact, it could be quite profitable, provided the owner had at least one other vehicle.

In all, some 43 individual pieces of essential business information were found, with one of those pieces alone capable of generating over $2 million in a combination of higher revenues and lower claims. By extending data mining to other "books of business," Farmers and other e-business insurers will be able to turn the data already being collected on clients into real competitive advantage.

FUTURE APPLICATIONS: It is difficult to predict specifically what "nuggets" will be discovered by data mining , except that they will almost always be pieces of insight and knowledge previously unobtainable. Current efforts include helping customers develop advanced -- and more focused -- marketing, intelligent customer relationship management, and analysis of Web usage data to make Web surfing more enjoyable and useful. Data mining of network access patterns will also likely reveal patterns of illicit use, enabling early "hacker detection."