Advanced Data Mining

Marketing secrets from the financial sector show how data mining charts a profitable course to customer management

Have you ever seen one of those posters that at first glance looks like a jumble of colored dots? Stare at it, and a three-dimensional picture will jump out from the pointillistic background. Now, think of those dots as the bits of information about your customers contained in your company's databases. If you look at the dots of information in the right light and at the right angle, they will reveal patterns that yield insight into customer behavior.

The banking industry has stared hard at its customer data "dots" to analyze customer behavior, and it has learned valuable lessons for other industries that use data mining. Although banks have employed statistical analysis tools with some success for several years, previously unseen patterns of customer behavior are now coming into clear focus with the aid of new data mining tools.

Data mining is the automated analysis of large data sets to find patterns and trends that might otherwise go undiscovered. By studying these patterns and trends, banking executives can predict with increasing precision how customers will react to interest rate adjustments, which customers will be most receptive to new product offers, which customers present the highest risk for defaulting on a loan and how to make each customer relationship more profitable.

"We're trying to understand the needs, preferences and behaviors of customers," says Russel Herz, senior vice president in charge of liability product management at The Chase Manhattan Bank in New York. Data mining has become an indispensable tool in that pursuit. "It permeates everything we do—pricing, promotion and product development," he says. Data mining led Chase to take the unusual step of reducing required minimum balances in customers' checking accounts for two consecutive years because the bank learned that customers who have difficulty maintaining a minimum balance may take their business to competitors with lower minimum balance requirements. Executives at the bank reasoned that customers who defected to another bank because of dissatisfaction with Chase's checking account terms might desert the bank for their other banking needs as well.

So while it was clear that for a certain segment of Chase's customer base, the minimum checking account balance was a key factor in their choice of banks, Herz still wanted to know if those customers were profitable for Chase. On the one hand, if Chase was losing profitable customers because its minimum balance requirement was too high, then lowering it would make sense. On the other hand, if the defecting customers weren't profitable, then the smart decision would have been to leave the minimum balance alone.

Chase analyzed the attributes and habits of its checking account customers for clues that might reveal how to set the minimum balance requirement to retain profitable customers. Herz's staff used data mining to develop profiles of customer groups whose members consistently had trouble maintaining minimum balances. They sought answers to questions like, How many checks do they draw per month? Do they use ATMs or conduct most of their business with tellers? and What other accounts and products do they hold? The answers helped Chase determine who their profitable customers were and predict the dollar floor at which to set the minimum balance to retain them. As a result of the two reductions of the minimum balance, Chase's percentage of profitable customers to overall customers went up. Herz doesn't want to tip his hand to competitors by specifying the percentage, but he says it was very significant.

Chase has possessed the customer data to alter strategy on such factors as minimum balances for many years, but recently the tools to study that data have improved significantly. Herz says, "It's been like [a doctor] going from having nothing to X-rays to an MRI." Herz can now predict the impact of strategic decisions with better precision and can make informed decisions much faster.

Tools to guide your way

Banks use an array of tools to analyze the customer data contained within their data warehouses and data marts. Data mining tools are still in the early stages of development, so practitioners may try several new tools each year. Boston-based Fleet Bank's data management and analysis group has tested at least 15 tools, including online analytical processing (OLAP) and other analytical products, in its two years of existence. "We're constantly looking at new software," says Victor Hoffman, vice president and manager of analytics at Fleet. Some vendors tailor their products for specific industries. San Francisco-based HyperParallel Corp., for example, offers a module with algorithms specific to the financial services industry in its data mining product. IBM Corp. will release a finance-industry-tailored version of its DecisionEdge software that incorporates the company's Intelligent Miner product later this year.

Of course, even before deciding which tools are best suited for a problem, data mining analysts must make sure they have consistent, accurate data. Banks historically have had separate databases for different areas: one for mortgages, another for auto loans, another for checking accounts and so on. But to obtain a full picture of customer relationships, an organization needs access to all customer data enterprisewide. Because of the legacy of multiple databases from several departments, data fields frequently contain different definitions and values in different databases, which can create misleading apples-to-oranges comparisons. Such inconsistencies have to be corrected before mining can begin. (See "Buried Treasure," CIO, Oct. 1, 1996.) For most companies, that's no simple task, and it slows down the work of data mining groups. "When we're building a data mining model, we spend the vast majority of our time—probably 75 percent of it—on data validation," says Chris Kelly, vice president and director of database marketing at Bank of America in San Francisco.

One of data mining's valuable traits is that results sometimes challenge commonly held assumptions. Other types of tools, such as OLAP, usually don't yield findings that directly challenge the status quo. Data analysts use OLAP tools to test the relevance of a theory, whereas they apply data mining tools to a problem in hopes that the findings will suggest an answer. Analysts use OLAP to answer, Is this true? but data mining can provide insight to answer, Why is this happening? and What might happen if...?

Managing Customers

Examining problems in a new light sometimes leads banks to abandon venerated strategies. Fleet subscribed to the long-held notion that the more cross-selling it did, the higher its profits would be. But the bank discovered through data mining that it doesn't always work that way. For instance, it doesn't pay for the bank to persuade customers content to leave their money in savings accounts to invest in certificates of deposit bearing a higher interest rate, says Fleet's Randall Grossman, senior vice president and director of customer data management and analysis. To do so decreases the profitability of such customers to the bank. To be sure, Fleet still puts a lot of effort into pitching additional products to its customers, but it now does so more selectively. "We target customers who are not only likely to buy the product but ones for whom the value of the relationship increases for us," Grossman says. In other words, Fleet wants to know not only which customers might buy a new product but also which customers will generate profits if they do buy it. Grossman says the bank's second goal for this sort of data mining is to find out how to "reconfigure, over time, [its] product and service offerings to enable [it] to profitably serve customers who today are unprofitable."

Many financial institutions use data mining to study the needs and habits of customer groups in the interest of "customer relationship management." The objective is to increase the amount of business with each customer. Banks regularly take aim at that goal with targeted promotional mailings and in the normal course of customer interactions; however, data mining helps them market more precisely, saving money on mailings and increasing the effectiveness of cross-selling efforts.

Fleet recently used data mining to identify the best prospects for its mutual fund offerings. Grossman's team mined customer demographics and account data including transaction activity and account balances along several product lines. From that analysis, they found customers who were likely to invest in mutual funds, and they used that information to help Fleet's Investment Services division target prospective clients.

Mining the Call Center

Some 13 million customers call Bank of America's West Coast customer service call center each month. It's an unbeatable marketing opportunity, but "rather than pitch the product of the week, we want to be as relevant as possible to each customer," says Kelly. So when a customer calls, a rep has a much better chance of cross-selling if he knows what accounts the customer holds and whether the customer is part of a group with a propensity to buy a particular product.

At Bank of America, customer service representatives equipped with customer profiles gleaned from data mining pitch new products and services that are the most relevant to callers. For example, a customer in a certain age group who has children and a home equity loan with the bank is a good candidate for taking out a student loan. Data mining helps the bank identify such customers.

Bank of Montreal mines its 1-terabyte "customer knowledge database" to develop profitability profiles of customers based on multiple factors such as the amount of money in particular accounts, demographic information, the number of monthly transactions and their choice of banking channels (teller, ATM or phone). The Toronto-based bank calculates current profitability profiles of households and produces models to predict the profitability of customers over a lifetime.

Managing Risk

In addition to helping increase the value of customer relationships, mining customer information databases aids banks in managing risk. Bank of Montreal, for example, analyzes mortgage customers' transactions in checking, savings and other accounts for insight into who is at risk of defaulting. The bank was surprised to find that some customers who consistently made their mortgage payments late were not necessarily at a high risk of defaulting. The bank found that a certain type of customer is in the habit of paying bills late but has the wherewithal to fulfill his or her obligations, says Jan Mrazek, chief specialist of the data mining group at Bank of Montreal. By analyzing the transactional behavior of customers across all their accounts, the bank can see which customers experience periodic cash flow crunches and which may truly be in danger of defaulting.

Some banks are currently testing data mining tools to manage their credit portfolios more efficiently. Data mining holds great promise in assessing the risk of a bank's entire portfolio of loans, says Jianmin Liu, vice president and project manager in credit risk management for Bank of America's mortgage division. By analyzing customer behaviors such as payment habits, data mining can provide answers to vital questions such as What percentage of loans will be refinanced next quarter? What percentage will go to foreclosure? and What percentage will be in serious delinquent status? Accurate answers to these questions allow credit risk managers to allocate optimal loan loss reserves—funds set aside to cover bad loans—which is important to profitability.

Credit risk assessment is a new area for data mining, Liu says. He is testing the beta version of a data mining product from SAS Institute Inc. of Cary, N.C., and is evaluating its various algorithms. His goal is to develop statistical models that accurately predict payment behavior, for example, how likely it is that a particular borrower will default on or, alternatively, prepay a loan. Accomplishing that will take patience and a lot of trial and error, Liu says, because data mining sometimes leads to dead ends and can yield misleading information.

For example, Bank of America built a data mining model to predict attrition of small business customers. One of the key factors was the length of time small businesses held accounts with the bank. That indicator proved to be misleading, however, because about 60 percent of small businesses go bankrupt within three years. Kelly says Bank of America's model was a better indicator of companies headed for bankruptcy rather than those headed to a rival bank. The bank subsequently revised the model to eliminate that factor.

Read It, but Don't Weep

Related:
1 2 Page 1
Page 1 of 2
SUBSCRIBE! Get the best of CIO delivered to your email inbox.