Most marketers and analysts agree that data mining plays a critical role in assuring a successful customer acquisition or retention campaign. Most also agree that even minor technical errors can destroy a campaign, But most of the problems that emerge in a modeling project have little to do with technical issues, says Sam Koslowsky, vice president of modeling solutions for New York-based marketing services provider Harte-Hanks. Here are a few of what he views as the most common—but preventable—errors:
* Setting the wrong objective. Take the case of a cataloger whose model identified 34% of customers as frequent buyers. While this appeared excellent at first glance, further analysis demonstrated that more than 30% of the merchandise purchased by these top responders was returned! This program achieved its objective of finding the top responders. The only problem is that it was the wrong objective.
* Using the wrong analysis. Some marketers, in trying to cut corners, skimp on their analytical tools. For instance, as part of aggressive growth plans, an apparel merchant with a primarily midwestern customer base decided to expand along the West Coast. Management opted to use an existing model for acquisition that had worked well for the company in the past. But the company soon discovered that what works fine in one geographic area may not perform appropriately in another region.
* Failing to “freeze” the file. “There is an old rule that I learned many years ago,” Koslowsky says. “I still invoke it today: ‘If it’s too good to be true, it isn’t.’” Such was the case with one modeling exercise. It seems that more than 60% of all potential responders were identified as fitting in the top segment. Upon further examination, however, a problem was identified with the analysis file used to crunch the numbers. Typically, previous marketing programs with known results are used as a source for model development. Two periods are defined. One of these includes the time before (“pre”) the mailing. The other, of course, is “after the solicitation,” the response period.
Predictors come only from the “pre” period. In this example, the most critical predictor emerging from the model was an attribute called “lifetime services.” When results using the model did not materialize, an analyst discovered that “lifetime services” as used by the modeler included services established after or as a result of the mailing. The file preparer did not freeze the file at the time of the mailing.