Companies Boost Sales Efforts With Predictive Analysis
Tools from vendors mine data deeply, identifying trends and patterns
Lapsang souchong tea has a smoky, earthy taste that's more appealing to men--or so thought William Todd, co-owner of the Todd & Holland Tea Merchants specialty tea company. Imagine his surprise when an analysis of the company's customer list and a year's worth of sales data resulted in a recommendation that Todd & Holland market the tea to professional women between 25 and 35. "That was a revelation," Todd says.
The analysis was part of a test of predictive-analysis software from SPSS Inc. Like most retailers, the River Forest, Ill., company, which sells tea through the Web and a retail shop, derives a disproportionate share of revenue from the top 20% of its customers. "We know who our best customers are. Our goal is to move the customers who are one notch below that up into that best-customer category," Todd says.
To turn occasional customers into best customers, retailers must predict who will buy what products. In the past when shopkeepers knew customers by name, that was easy. But in these days of mass-market, multichannel retailing, even Todd & Holland, which prides itself on understanding its clientele's tastes, is finding it tough to make such predictions.
To do that, more businesses are turning to predictive analysis, a technique that models historical data with assumptive future conditions to predict outcomes or events. Predictive analysis includes forecasting and propensity analysis. Forecasting identifies trends and predicts future sales, for example. Propensity analysis uses data-mining algorithms such as regression analysis, decision trees, clustering, and neural networks to calculate consumers' predilection to buy a particular product, respond to an offer, or default on a loan.
Predictive analysis is a subset of data mining, technology used to uncover hidden trends and patterns within large amounts of data. A number of established data-mining software vendors, such as NCR's Teradata division, SAS Institute, SPSS, and several startups, including Genalytics, Magnify, Quadstone, and Sightward have developed predictive analysis tools. Prices range from $50,000 to several hundred thousand dollars. Sales of data-mining products worldwide are expected to grow from $597 million this year to $1.86 billion in 2005, according to International Data Corp.
One early adopter is the Body Shop International plc, the U.K. skin-and hair-care products vendor. The U.S. division in Burlingame, Calif., is testing Sightward's predictive-analysis offering to boost the efficiency of its mail-order businesses, cutting back on the number of catalogs mailed out while improving response rates.
"We knew it was going to be a tough economic year, and we wanted to be as profitable as possible," says Virginia Newman, the Body Shop's director of mail-order and new-business-development. The retailer cut back its catalog circulation by almost 50% in 2001. Normally, the Body Shop builds mailing lists using customer sales data from previous mailings. In September, it enlisted Sightward to analyze its database of catalog, Web, and store customers. Sightward used the results to build a mailing list for the 120,000 copies of the Body Shop's fall catalog.
The store is still quantifying the results, but it appears that revenue per catalog has increased 10% to 20%, Newman says. "By using Sightward, we were able to intelligently target consumers who we thought were more likely to buy and increase our revenue per catalog in a measurable way."
Retailers generate mailing lists using relatively simple recency, frequency, and monetary (RFM) formulas: How recently customers purchased something, how frequently do they buy, and how much do they spend. With predictive-analysis tools, they can build complex models incorporating hundreds of variables to make more-accurate assumptions.
Earlier this year, Sur la Table Inc., a Seattle supplier of kitchenware and cooking paraphernalia, also turned to Sightward to help generate a mailing list for an August catalog. Sur la Table's mail-order service bureau analyzed the retailer's customer list according to variables such as customer lifetime purchases, average order, type of products purchased, and response to Sur la Table's last marketing campaign.
"We wanted to increase the sophistication of our name selection for mailings," says Susan Ghilarducci, Sur la Table's marketing director. About 35% of the names on the list generated through the Sightward analysis were different from a list Sur la Table created using conventional RFM methods. Sightward "found us customers that we wouldn't have found using RFM segmentation," Ghilarducci says. She's still analyzing the results of the mailing.
Fingerhut Cos. in Minnetonka, Minn., has always used regression analysis to generate catalog mailing lists from its database of 6 million-customers. The mail-order company recently began using Quadstone software to fine-tune those efforts. Fingerhut eliminated from the list customers who were unlikely to respond. Business-intelligence director Randy Erdahl is confident the effort will improve return-on-investment through higher sales and lower mailing costs. "We know we will generate millions of dollars by refining our mailing lists," he says.
Financial-services companies may be the most sophisticated users of predictive analysis. Wells Fargo Home Mortgage, a subsidiary of Wells Fargo & Co. in San Francisco, uses data-warehousing and-analysis software from SAS Institute and internally developed applications to predict the performance of its mortgage loan portfolio and individual mortgage loans. "The goal here is risk avoidance," says Frank Eichorn, director of the subsidiary's credit-risk-management data group in Frederick, Md. Its data warehouse contains more than 200 million payment records on 12 million borrowers--one of every 12 U.S. mortgages.
To analyze all that data, the company built predictive models on the SAS platform. A credit-scoring model, for example, helps managers evaluate new loan applications and predict the likelihood of default. Models that predict the performance of the entire loan portfolio help managers assess potentially bad loans, says Patrick Homble, the division's VP of model analysis.
The system has produced quantitative benefits. Wells Fargo's own loan default rate predictions were one-half of what rating services such as Moody's and Standard & Poor's predicted. That let Wells Fargo Home Mortgage renegotiate the loans, assuming more risk and saving $250,000 in interest expenses. The more-accurate predictions also helped the bank form partnerships with mortgage-insurance companies that generate nearly $35 million a year in revenue.
Perhaps the most unusual application for predictive analysis is in Westminster, Calif., an Orange County community of 87,000, where police are using MetaEdge Corp.'s C-Insight software to develop a crime-analysis system. By analyzing data from computerized crime reports, the system is designed to detect and map local crime patterns. "That way, we can be more effective in our patrol efforts and use of the resources we have. We could curb crime trends before they get out of hand," says Westminster Lt. Derek Marsh. Police expect to use the system to predict crime trends, improve patrol assignments, and develop better crime-prevention programs.
Predictive analysis is finding its way into more mainstream applications for a variety of reasons. Just a few years ago, complex predictive models required supercomputers. Now, the latest Intel servers can do the job. What's more, the recession is forcing companies to squeeze more return out of their catalog mailings, telemarketing calls, and other sales and marketing efforts. And multichannel retailers want to leverage different customer databases.
Predictive analysis also could get a boost if the Predictive Model Markup Language catches on as an industry standard. The XML-based language for defining predictive-analysis models lets models built for one data-mining tool be used with other compliant software.
But the technology still faces hurdles. Building models is a complex process, assuming a business has a data warehouse with its customer data properly aggregated for predictive-analysis chores. The modeling and data-integration aspects of predictive analysis can be particularly complex. "Developing and testing predictive models isn't for the faint of heart; rather, it's still largely for the 'lab coat' set," Meta Group analyst Doug Laney says.
Few companies have the necessary expertise in-house to do the work. Startups such as Genalytics, which uses genetic algorithms to assemble its models, help bridge that expertise gap. But many companies aren't comfortable shipping their customer databases to outside contractors for running predictive analysis models, says Jeff Roster, a Gartner senior analyst. Another potential weakness is missed market niches or opportunities that aren't predictable, such as someone discovering a new use for a product, Meta's Laney says.
But Todd, of Todd & Holland Tea Merchants, is excited about the prospect of understanding his customers better. Now that SPSS has analyzed his customer data, he plans to install a recommendation engine on the company's Web site and a point-of-sale system, both of which will suggest additional products customers might like. "We want to provide information, but not shove it down their throats," he says. "The goal is to identify customers who can become very good customers."
About the Author
You May Also Like
2024 InformationWeek US IT Salary Report
May 29, 20242022 State of ITOps and SecOps
Jun 21, 2022