Targeting ‘persuadables’ with Net Lift modeling

Optimization of movie rental discount campaigns

Is Net Lift modeling a practical technique for the tuning of target groups for promotional campaigns? Egwin Avau and Maarten Baeten, two Management Informatics students at KU Leuven, tested it out for their master graduate thesis in 2014. They analyzed the responses on a promotional campaign of a major American movie and game rental company. Using SAS software, they demonstrated that Net Lift modeling techniques indeed enable a more precise identification of the most promising target group for discount offers.

Making the most from physical movie rental

This major US based movie retailer is a household name for American movie and game lovers, well-known for their eye-catching self-service retail kiosks placed in thousands of supermarkets, convenience stores, fast-food restaurants and pharmacies across the United States and parts of Canada. The company commands nearly a 50% share of the physical video rental market.

Despite the growing importance of online media streaming—for which they are setting up a parallel business—the company believes that physical rentals will remain an important and profitable business in the upcoming years. Improved marketing campaigns should help make the most out of the available retail network.

Re-awakening dormant customers

One such marketing campaign aims at re-awakening dormant customers by offering them a substantial discount for their next purchase. “The initial target group for the discount offer was all customers in the database who had not rented anything for the last 150 days,” explains Maarten Baeten. “It was our mission to refine the target group in order to improve the success rate of the discount campaign. We were able to analyze a large sample of their database, containing contact and transaction information of 400,000 customers. With this wealth of information, we were able to discern nearly 270 customer variables potentially indicative of customer behavior.”

Taking the challenge to a level higher

A limited email campaign was launched to trigger customer response and learn from the results. Egwin Avau explains: “By analyzing responses against customer profiles using regression techniques, we were able to build an initial predictive model that indicated the dormant customers that would most likely respond positively to the discount and become active again. In addition, we also took into account post-campaign customer behavior in order to identify the segment of customers that not only grab the discount but remain active afterwards.”

However, thesis advisors Professor Bart Baesens and Véronique Van Vlasselaer were firm on their insistence that they take the challenge to a level higher. “Straightforward profile-based response modeling falls short in at least one important respect,” elaborates Avau. “We cannot be certain with this analysis whether the reactivated customers came back as a result of the discount or some other factor; some of them might have come back anyway. And that’s where Net Lift modeling comes in.”

Avoiding targeting sure things

Net Lift modeling—also known as True Lift Modeling or Incremental Response Modeling—is a response modeling technique first proposed around the turn of the century by Professor Victor Lo, among others. It aims at reducing the size of the target group by eliminating customers that would purchase again even if not targeted by a campaign. “We call these customers the sure things,” says Baeten. “Offering them a discount would be a waste of money, even when they have been dormant for quite a while. Our real target group is those we call the persuadables, the customers that must be triggered with a bonus.”For this reason, Avau and Baeten launched a second campaign, omitting the discount to build a control group. Then the control group responses were analyzed using the same regression technique that had been applied to the test group. Net Lift modeling then involves subtracting the positive results of the control group from the positive results of the test group.

Implemented exactly as specified

The results were very satisfying. “Using this technique, we were able to reduce the target group significantly without impacting the response rate,” says Avau. “And we’re quite proud to be among the first to successfully apply this technique on real-life data from customers.”

The students made extensive use of SAS Enterprise Miner to carry out all of the analyses, including the Net Lift modeling. In fact, they even coded a new Net Lift node for that purpose. “Our research required that we code the node ourselves,” says Baeten. “This enabled us to implement the technique exactly according to the theoretical framework provided by Victor Lo. SAS Enterprise Miner enabled us to do that. It automated routine tasks and enabled us to concentrate on specifics such as the correct implementation of sampling techniques.”