Towards more robust uplift modeling for churn prevention in the presence of negatively correlated estimation errors

Abstract

The subscription economy is rapidly growing, boosting the importance of churn prevention. However, current true lift models often lead to poor outcomes in churn prevention campaigns. A vital problem seems to lie in instable estimations due to dynamic surrounding parameters such as price increases, product migrations, tariff launches of a competitor, or other events with uncertain consequences. The crucial challenge therefore is to make churn prevention measures more reliable in the presence of game-changing events. In this paper, we assume such events to be spatially finite in feature space, an assumption which leads to particularly bad churn prevention results if the selected customers lump in an affected region of the feature space. We then introduce novel methods which trade off uplift for reduced similarity in feature space when selecting customers for churn prevention campaigns and show that these methods can improve the robustness of uplift modeling.