Canadian Bank Mines for Data-Based Gold
ACTION ITEMS

ComputerWorld has reported that the bank of Montreal wants to offer its customers the right product, at the right price, at the right time. And it wants to make money doing so. The bank hopes sophisticated data mining techniques will help it do just that.

Using advanced techniques in mathematics and artificial intelligence, data mining uncovers complex patterns or models in data. Those models are then used to help solve business problems that come up in direct marketing, credit-risk evaluation, fraud detection and other areas.

At The Bank of Montreal, that means taking information about bank customers -- their assets, services they use and their history of account fees, for example -- and crunching it to determine what products and prices can be customized to fit their individual needs.

"The whole idea behind database marketing and data mining is to switch from a focus on products to customers, to view them as households and analyze their portfolios," said Jan Mrazek, manager of the data mining group at the Toronto bank. The bank was founded in 1817 and has $101 billion in assets.

The bank already has crunched the data on its 5.1 million customers and come up with a profitability figure for each - that is, how much revenue each customer generates for the bank in account fees and interest.

Those calculations took a staggering 1.5 years to crunch, primarily because the data-gathering process was time-consuming and complex, Mrazek said. Those calculations are processed and stored on an IBM SP2 server.

The bank then looked at each customer of its online banking service, Mbanx. It scored each customer and assigned each to a group of similar customers, such as highly educated people who are unprofitable in checking and savings accounts but profitable in mortgages and loans. That was accomplished by feeding to the data mining tools the profitability figures, demographics and behavioral variables.

The bank used clustering, available from IBM's Intelligent Miner data mining tool, to do that.

"Generically, what clustering identifies is how things are clumped together," said Herbert Edelstein, president of Two Crows Corp., a Potomac, Md.-based consultancy that specializes in data mining.

But he warned it won't reveal cause-and-effect relationships -- merely associations. "There may be cause and effect, but you need to go outside data mining to find it," such as through assumption or hypothesis based on experience and business knowledge, he said.

Return on investment for the project hasn't been calculated, Mrazek said. He declined to provide cost figures. The bank plans to update profitability calculations and customer clusters monthly, beginning next spring. The results will be stored in the Bank Information Warehouse, which resides in a DB2 database on an IBM 3090 mainframe that runs MVS. Information pertaining to Mbanx customers also will be stored on a data mart, an Oracle 7.3 database running on an IBM RS/6000.