Silicon And Software That Mine For Gold

Marketing companies have long used computers to sift through reams of supermarket scanner data, merge and purge mailing lists, and analyze market research reports. But now, ultracheap computing power and new software are making the process much more precise.

Gathering data isn't all that difficult: Point-of-sale terminals, teller machines, and 1-800 telemarketing all contribute to the flood. The big challenge lies in making sense of what's collected, rather than throwing most of it away. The quantities of data are potentially overwhelming: Fingerhut Cos., for example, is currently expanding its collection of mail-order customer data from about 600 billion characters today to about 2 trillion characters, or two terabytes--enough to choke any traditional mainframe. Using shoppers' clubs and other incentives, supermarket chains such as Vons Cos. and Safeway Inc. are starting to keep on hand more than a year's worth of detailed data about customer purchases.

SOUPED UP. Fortunately, powerful new technologies are at hand. Neural-network software, designed after the pattern of cells in the human brain, can automatically "learn" from large sets of data on its own. By scanning thousands of data records again and again, the software can build a strong statistical model describing important relationships and patterns in the data. All that's required is a standard, high-end PC equipped with a plug-in neural-net "accelerator" board. Customer Insight Co. has tailored HNC Software Inc.'s Database Mining Workstation software, based on neural-net techniques, just for database marketing.

Once a statistical model of the ideal customer is constructed, however, considerably more computing horsepower may be required to find all prospects matching the profile. So Fingerhut, Wal-Mart Stores, Delta Air Lines, American Airlines, and dozens of others are turning to so-called parallel-processing systems. These devote dozens or even hundreds of micro- processors to scouring a giant database for records that meet a complex set of criteria. The more criteria specified, the longer a search can take--but the more precisely aimed the resulting marketing efforts may be. Fingerhut has chosen a parallel Sun Microsystems Inc. machine that harnesses as many as 20 high-speed microprocessors. Wal-Mart Stores Inc. leads the way in retail databases using parallel database computers supplied by the former NCR/Teradata company, now a unit of AT&T.

Rapidly analyzing zillions of past business transactions is actually key to many "reengineering" efforts these days. Businesses ranging from airlines to banks and telephone companies are striving to reorganize based on a better understanding of their customers' buying patterns. The potential gains are big: tighter inventories, more compelling product displays, fewer out-of-stock items, and higher profits. Says Thomas Blischok, a vice-president at AT&T/NCR: "If I can guarantee that ev-

ery time you go to my establishment you find stuff to buy, you'll care a lot less about price."

BIG-IRON BRAWL. All of which is stirring sales in the otherwise moribund large-scale computer market. Sales of traditional mainframes have gone into long-term decline. But Gartner Group Inc., a computer market-research firm, reckons that today's $400 million commercial market for parallel computers will grow to $5 billion in 1998. IBM, the leader in mainframes, is adding parallel systems to its lineup in hopes of thwarting early advances by challengers such as AT&T, Tandem Computers, Sun, Hewlett-Packard, and even Cray Research.

What does the future hold? Lots more transaction data stored in much more powerful computers. Customers' thirst seems insatiable, says Jerre Stead, chief executive of AT&T/NCR. He was recently given a pair of hats when visiting Wal-Mart headquarters, where inspirational slogans abound. One, Stead recalls, demanded that AT&T computers be capable of storing "10 terabytes by the end of '95." The other demanded that "90% of all data [be] processed in 99 minutes." Says Stead: "We'll get there. We're working on it."


Use gangs of microprocessors to scan huge volumes of data in a flash. Sold by AT&T/NCR, IBM, Tandem, Meiko, and nCube.


Cross-indexes data records into giant matrixes, which makes finding specified records much easier. Sold by Oracle, Sybase, IBM, Informix.


Can automatically build a model of customer behavior based on analysis of previous transactions. Available from HNC Software, Customer Insight.


    Before it's here, it's on the Bloomberg Terminal.