The underwriters at Farmers Insurance Group know a few things about drivers. They know young people crash their automobiles more than older drivers, and that sports cars are involved in more accident claims than station wagons. But with 10 million auto policies in its database, Farmers suspected there was a lot more to be learned about its customers. It's easy to run statistics by a driver's age or the type of vehicle, says Tom Boardman, assistant actuary for personal lines pricing at Farmers in Los Angeles, but you can't define a customer by any one characteristic. "It's trickier than that," he says. "We wanted to see the interaction of five or six variables."
Not an easy task, given that each Farmers policy has about 200 individual pieces of information tied to it-everything from the type of car to the number of kids in the policy holder's household. Enter data mining. With help from the specialists at IBM, Farmers pulled 2 million policies from its database to run a pilot test. Then Boardman and his crew stepped aside, letting IBM's DecisionEdge software go to work, detecting interesting patterns among the records on its own.
Early findings, including the fact that young people file more claims than older drivers, left Farmers unimpressed. Then, Boardman says, came the "good stuff." Exhibit A: sports car owners. Think of one and you probably imagine a twentysomething single guy flaming down the highway in his hot rod. In fact, there were plenty of just those types in Farmers' database, but there was another, previously unnoticed, niche of sports car enthusiasts: married boomers with a couple of kids and a second family car, maybe a minivan, parked in the driveway. Claim rates among these customers were much lower than other sports car drivers, yet they were paying the same surcharges. Armed with this information, Farmers relaxed its underwriting rules and cut rates on certain sports cars for people who fit the profile. Today the company is planning another data mining expedition, this time to identify car insurance customers who are good prospects for Farmers homeowner policies. "We're sitting on a gold mine," Boardman says of the company's database. "It can do much more than just send out the bills."
Farmers Insurance isn't the only company prowling in the mining pits. >From banks to e-commerce players, health care providers to telecoms, many companies today are plunging into their databases to determine who their best customers are and how better to market products and services to them. Most are wading into their data, hoping to substantiate hypotheses or hunches with hard numbers. But others, like Farmers, are venturing into the realm of automated discovery, relying on sophisticated, computer-driven methods that use artificial intelligence-such as neural networks, association rules, and genetic algorithms-to flesh out meaningful, actionable information that can make a difference to the bottom line. "It's the same algorithm, whether you're trying to predict volcanoes on Venus or a customer's propensity to buy high heels," says Usama Fayyad, senior researcher in the decision theory and adaptive systems group of Microsoft Research.
In a new survey of corporations with data warehouses conducted by consulting firm META Group, 54 percent of respondents said they plan to purchase data mining and other knowledge discovery tools this year. That's a 20 percent jump since 1996. But there's still a long way to go: Only 8 percent of respondents currently use data mining software. Meanwhile, two Internet powerhouses that have been prescient in the ways of the Web have gobbled up entire companies with data mining expertise. In April, Amazon announced it would buy Alexa Internet, a three-year-old ad-supported Web navigation service that tracks Internet users as they surf, provides information about the pages they view, and suggests other sites they might like to visit. Alexa now boasts 13 terabytes of clickstream data -roughly equivalent to 722 million copies of this 18-kilobyte article-and the ability to detect patterns buried in it. And in March, Yahoo purchased HyperParallel, a developer of data mining software. The mission at both Amazon and Yahoo is clear: to earn value from the bits and bytes of data collected on Netizens every day. By 2002, companies will spend $113 billion to analyze such customer data, including mining it, estimates Palo Alto Management Group, a California-based research firm.
What's driving the growth? For one thing, the sheer amount of data in the back room. According to the META Group, data warehouses with more than 1 terabyte will increase from 19 percent of all installations to 30 percent, becoming the largest segment this year. As a result, companies will be able to capture more customer information and retain it longer.
Catalog retailer Fingerhut Corp. in Minnetonka, Minnesota, began scrutinizing its database when its first customer placed an order back in 1948. Today, the direct marketer puts out some 130 different catalogs and touts a 6-terabyte data warehouse with information on more than 65 million customers. Hundreds of in-house users from merchandising, marketing, and analytics query the database daily and can crunch more than 3,000 variables on the company's most active 12 million customers (those who've purchased in the last four years). That list of variables includes everything from specific product transactions to demographic data collected from customer surveys and outside research vendors like The Polk Company.
More than 300 predictive models are now used to scour the terabytes at Fingerhut. One model predicts the likelihood of someone responding to a targeted electronics catalog, while another scores the chances of a customer returning her merchandise. One or more models is run on each potential recipient of every catalog that goes out the door, says Bill Flach, Fingerhut's director of corporate research and analysis.
Data mining has also led to the creation of new catalogs. In one analysis, researchers found that customers who changed their residence tripled their purchasing in the three months after their move. Their product choices also followed a pattern, with furniture and decorations topping the shopping list. That may seem like a no-brainer, but to Fingerhut, it was a valuable nugget to capitalize on. The company developed a new "mover's catalog" filled with targeted products for this consumer segment. At the same time, it saved money by not mailing other catalogs to these folks right after they relocated. Flach declines to give numbers on the mover's catalog, but says it's one of the most successful products from their data mining efforts.
Clustering customers into segments is a top business objective in data mining, says Mark Brown, global product manager at SAS Institute in Cary, North Carolina. Others include customer retention, acquisition, and cross-selling opportunities. "You need to get that customer-centric view into your data warehouse," he says. "Many companies don't have a scope that's narrow enough when they get started."
In other words, don't expect the analysts in IT to zoom in on the business strategy. "Data mining projects that start on the IT side are a recipe for disaster," says Gregory Piatetsky-Shapiro, director and chief scientist of Boston-based Knowledge Stream Partners, a data mining consulting firm for the financial services industry. "The problems they're trying to solve may not be the important ones."
That's not a concern at Eddie Bauer in Redmond, Washington. No longer do the analysts in IT hold the magic keys to unlocking hidden patterns in the retailer's database. Director of circulation Kevin Hillstrom and others in the company's marketing department now run queries right from their desktops to determine which promotions to offer in their stores and which catalog to send to a customer who hasn't bought in a while. Simplicity is key to end users, says Microsoft's Fayyad. "The user wants a solution, they want something that talks their language," he says.
For such a user-friendly desktop system to work, however, the data has to be ready to be mined. Data preparation, analysts contend, can eat up the most time-and create the most headaches-in any mining project. SAS Institute's Brown estimates that in any given analysis, companies spend 80 percent of their time just readying the data.
Stephen Coles, assistant director of research and development at American Century Investments, a financial-services firm in Kansas City, Missouri, knows this firsthand. To start data mining, ACI pulled 2 million customer records out of its 25 million master file, each with roughly 800 variables attached. Then came a snag: Not every record contained all 800 pieces of information. Some lacked occupational data; others had empty blanks on valuable credit card information. Throwing out these records before mining the data would have eliminated too much of the base-and skewed the results, Coles says. To fix the problem, the company used SPSS Missing Value Analysis to analyze patterns within the large data sets and impute the missing values. Estimated time to fill in the blanks? Thirty minutes. ACI's data mining has led to better-targeted direct mail campaigns. One mailed to high-end customers drew a 7 percent response rate, in part because the company could identify the right recipients in its database.
But when you have hundreds-or thousands-of demographic and other variables to choose from, how do you know which ones will help you understand customers better? Many analysts agree that past behavior is more predictive than age, sex, and income, but that broad definition doesn't narrow the number of fields very much. Isolating just the right handful of variables for making forecasts is still a work in progress, says Dr. Ashok Srivastava, chief technologist of IBM's knowledge discovery consulting group. "Let's say you're trying to predict whether the price of IBM stock is going to go up or down in the next ten minutes," he says. "There are a huge number of variables that you could use to make that forecast, including recent news, past behavior of IBM stock, and the behavior of similar stocks, to name a few. Data mining gives you the ability to sift through thousands of potential variables to isolate a few key variables that are highly predictive." In some cases, Srivastava says, the most predictive variables may not even exist in the database but need to be created using data that is already available. "This is a significant research activity which I think holds the key to making good predictions," he says.
Still missing from data mining models is the element of time, adds Srivastava. "There's a lot of emphasis on analyzing static databases," he says. "But the real world has many dynamic aspects to it, just like the stock market moves minute by minute. You need to build models that characterize things that change in time." A large bank, he suggests, has many outlets to reach customers-the telephone, ATMs, the Internet, branches inside grocery stores, and so on. The bank may run an advertising campaign to promote its branches in grocery stores, but then discover a few months later that business in the branches is still slow. Meanwhile, its Internet site is booming. Srivastava believes that one day, methods will be able to forecast such business shifts. If the bank can predict that business in the grocery stores will shrivel up in a few months, it can invest its ad dollars in promoting outlets with more potential, like the Internet.
Some companies hope data mining can help them anticipate customer needs. Bank of America recently completed a proof-of-concept project with IBM's data mining tools to delve into its database of corporate accounts. "Each corporate relationship can involve potentially millions of transactions," says Fritz Offensend, vice president of financial engineering systems at Bank of America. "There's probably no one in the bank who knows every single deal we've cut with large clients. We want to figure out what products they might need next, what deals we should start to discuss in the coming months." In a pilot test, Bank of America pulled together the records of thousands of corporate clients, each with roughly 400 variables. Offensend declined to comment on the trends and correlations uncovered in the study but says they definitely proved the worthiness of data mining. "The real payoff will be when we're chalking up profits based on our insights," he adds.
Still, not everyone is convinced that finding less-than-obvious patterns in a database breeds business value. Skeptics love to tell the story of a grocery chain that found a correlation between purchases of diapers and beer. An interesting discovery that probably would have gone undetected without data mining, but ultimately, it was deemed unactionable. For one thing, the company decided it would be too expensive to move inventory around in its stores. "Data mining produces too many answers that may not be causal. How do you know which are actionable and which are not?" asks Mike Duffy, director of analytics and database services at Kraft Foods. "There is no judgment factor."
Of course, data mining requires a kind of explorer's mentality-you never know what you're going to find and what changes that might lead to in other areas of the business. Consider retail: In 1974, a pack of Wrigley's gum was waved over a glass panel and became the first product ever scanned at a supermarket checkout. Thanks to frequent-shopper programs and ongoing advances in scanner technology, grocery chains have been swimming in customer data for several years. Trouble is, real-world applications are only now catching up with the innovations. One analyst recalls a grocer who threw out reams of customer data because the information was outdated. The grocer could have used that data when it was fresh off the scanner to offer targeted promotions to customers. Instead, the information gathered dust in the back office. Today, a few companies are truly leveraging that knowledge to boost business. Dick's Supermarkets, an eight-store chain in Wisconsin, uses transaction data culled from its loyalty-card program to personalize shopping lists it mails to nearly 30,000 members. "Many retailers don't have the infrastructure to mine their data," says Steven Kingsbury, president of Promotion Decisions, a Cincinnati-based company that analyzes coupon promotions based on household scanner panel data. "There's a real opportunity for manufacturers to help them build those techniques. That's going to be part of the overall competitive situation."
But even its strongest advocates believe data mining is only part of the solution when it comes to building customer relationships. Another critical ingredient: people with solid knowledge of the marketplace. "The effective competitors in the future will be the ones who consolidate theiranalytical insights with business judgment that's not captured in any database," says Offensend from Bank of America. "You need to tap the knowledge of the computer-and the knowledge of the human."