On the Assumptions of True Lift Models for Churn Prevention

Kurzfassung/Abstract

Preventing customer churn by subjecting carefully selected customers to customer relationship management activities is of crucial importance to many service industries. A promising selection of customers can be achieved using so called true lift or incremental models, which focus on customers at high churn risk, that are also likely to be persuadable through appropriate campaigns. In comparison to simpler models, true lift modeling however not only requires estimating churn probabilities of untreated customers but also their churn probabilities when treated. We argue that the estimation of the latter probabilities introduces a novel source of uncertainty not considered in state-of-the-art true lift models. In this paper, we assess the consequences of these uncertainties for true lift modeling. We identify assumptions regarding distribution of churn probabilities made by true lift models and argue that these assumptions are most likely not met in any practical setting. As a result, churn prevention campaigns can easily fail and even increase total churn rate, which might provide an explanation for the very few published empirical success stories on true lift models.