Corporations such as Fingerhut have found that, by combining years worth of business transaction data with U.S. Census Bureau and private gathered demographic information, they can predict consumer behavior we ll enough to profit from their knowledge.

Fingerhut is mining a data warehouse with information about more than 10 million current customers to find out which are most likely to buy certain types of products, and therefore should be mailed certain cat alogs. Fingerhut mails catalogs to demographic niches such as home textiles, outdoor, home handyman and holiday products.

"Data mining is a low-cost way for us to assess the buying behavior of groups of customers," Johnson said.

In a recent data mining effort, Fingerhut studied past purchases of customers who had changed residences, and found that they were three times more likely to buy items such as tables, fax machines, phones and decorative products, but were not more likely to purchase high-end consumer electronics, jewelry or footwear. Mining also showed that those buying patterns persist for 12 weeks after the consumer moves, but that purchases peak in about the first four week s.

As a result, Fingerhut has created a catalog aimed at people who have recently moved, and carefully tailored it to their purchasing patterns, Johnson said.

He hastened to say the data mining results don't mean that all consumers who move will avoid buying computers or shoes, but that "people will not buy at a rate that would justify our investment in printing cat alogs.'

Spanish connection

Another data mining effort discovered a subset of Fingerhut customers who responded more favorably to catalogs printed in Spanish. Billing, customer service and customer correspondence also are provided in Spa nish.

Aside from the language in which they were printed, the Spanish catalogs varied only slightly from a standard Fingerhut catalog of the same type: They contained a somewhat greater emphasis on fine jewelry, which data mining showed to be particularly appealing to people who like Spanish language catalogs, he said.

The result: The Spanish catalog generates about 40 percent more orders than would normally be expected from those consumers, Johnson said.

Data mining also helps Fingerhut identify customers who are not likely to purchase or not likely to purchase enough to make the catalog mailing profitable. This group, dubbed "potential attriters," are the tar get of special promotions to win them back, such as discount certificates. Customers identified as unlikely to be profitable enough to justify continued catalog mailings are likely to be dropped from Fingerhut's current mailing lists.

Privacy issues

Johnson acknowledges that consumers may be concerned that their privacy is being compromised by data mining. The consumer information studied includes which Fingerhut products customers do and don't buy, how o ften they return Fingerhut products, how they finance their Fingerhut purchases and whether they pay on time.

In addition, Fingerhut purchases demographic information from the U.S. Census Bureau and private companies. That research reveals information such as age, household income and number of people in the household.

Johnson's says Fingerhut has no interest in the behavior of individual consumers; target groups for telemarketing or catalog and coupon mailings must be very large to make it economical for Fingerhut to pursue them.

"The trick is to find consumers groups large enough that we can serve them as a group. It won't do any good to find three people because we can't economically print and distribute catalogs to them," Johnson said.

For example, the smallest target group Fingerhut's telemarketing would pursue is about 10,000 people. A single product offer would not be mailed to a group smaller than 300,000 consumers and a catalog would not be mailed to fewer than 1 million people. (H owever, separate demographic segments making up the total catalog mailing could total about 20,000 people each.)

Using data mining to predict which customers are likely to buy products enables Fingerhut to control its catalog printing costs by pruning its mailing lists of unlikely purchasers. The company pays $400 to $90 0 to print 1,000 copies of a catalog when the cost of paper, printing, binding and postage are included.

In some cases, data mining helps the company decide exactly how many catalogs to print. Mining has shown that 20 percent of Fingerhut's customers will move every year, so it prints a few million moving catalog s in advance.

Within the direct-mail catalog industry, data mining's biggest contribution to date has been in cutting costs but allowing companies to focus mailings on the customers most likely to buy, Johnson said.

"The big carrot is to develop a way to find additional revenue. That's what everyone is seeking." He declined to say how much Fingerhut has saved as a result of data mining, or how much additional revenue it h as obtained.

But, he adds, "We are the second-largest catalog company in the United States, and we would not be in business without the market segmentation data that data minin g produces."

Better mining

Modern data mining such as Fingerhut's is more precise and less expensive to use than the mathematical computer models that preceded it, and as a result is more important to business.

While the basic techniques have been available since the 1950s, DePlanque said mining is now being more widely used. That's because of the growing popularity of corporate data warehouses, which make it easier to search vast quantities of information, and the huge amount of historical business data now available in digital form rather than on paper. In addition, the necessary computer hardware cost $1 million five y ears compared with a few thousand dollars today, she said.

For example, Damark used to perform data mining on a 10 percent sample of its information because of the effort and cost involved. With improved and less expensive data mining tools it bought for about $500,000 early last year, the firm now sifts through all the information on its 5.5 million customers, or about 150 billion bytes of data.

The company is engaged in "a multimillion-dollar, multi-year effort that was started four years ago," Johnson said. The effort involves merging 31 Fingerhut databases of consumer information into a single data warehouse that contains 7.5 trillion bytes of data, and making it easier for nontechnical marketing people to mine the warehouse using models the statisticians cr eated.

Mary Lee Tesch-Stevson, marketing manager for Fingerhut's "Project Eureka!," which seeks to promote cooperation between marketers and data miners, says one of her goals is to make data mining readily available to nontechnical people.

"We have been able to create ways for the average marketing person to get at this information by going into the data warehouse, asking a question and getting an answer. It's a supplement to the more extensive data mining that goes on in the statistical analysis area," Tesch-Stevson said.

For example, a marketing person might tweak the data mining statistical model to make sure a catalog that offers coats is mailed to geographical areas where the temperature is cold enough to produce the highes t sales.

While the results are not dramatic, Damark's Voda says even a small increase in orders from the 100 million catalogs Damark mails each year can make data mining worthwhile. He said cost reductions and increase d sales attributed to data mining have allowed Damark to recover last year's data mining investment.

"We're not looking for the Holy Grail. It's a game of inches, and it does not take a huge improvement in orders to drop a lot of money to the bottom line," Voda says.