It’s no secret that many catalogers use historical customer information to create elaborate predictive models. But Scott Nuernberger, lead statistician for Lanham, MD-based database marketing services firm Merkle Direct Marketing, says that without a prior plan to influence customer behavior, all the modeling wizardry in the world will fail.
“A lot of people get caught up in the technology without properly setting up the business problem first,” Nuernberger says. Say you have a problem with customer loyalty. “It will do you no good to identify customers about to leave if you have no plan to influence their behavior. Using a predictive model is not enough.” But having a plan in place to offer discounts, for instance, to those likely-to-stray customers can enable you to profit from the knowledge generated by the model.
Also be sure that the problem you’re fixing can be measured. Response rates and conversion rates are easily measured. But for more-ethereal factors, such as customer satisfaction, you will need to establish metrics, through customer surveys, for instance.
Another pitfall revolves around testing the model. Many times marketers will design a test without considering the necessary sample size to get statistically significant results, Nuernberger says. I you are interested in response rates, he says, it is really the number of responders that determines statistical significance, not the mail volume. So a marketer expecting a 0.1% response rate will need a much larger sample size than a marketer expecting a 5% response rate.