Doing More with Less: Overcoming Ineffective Long-Term Targeting Using Short-Term Signals

Firms are increasingly interested in developing targeted interventions for customers with the best response, which requires identifying differences in customer sensitivity, typically through the conditional average treatment effect (CATE) estimation. In theory, to optimize long-term business performance, firms should design targeting policies based on CATE models constructed using long-term outcomes. However, we show that such an approach may fail to improve long-term performance, and can even harm it, when the outcome of interest (e.g. repeated purchases or CLV) accumulates unobserved individual differences over time. Our theoretical analysis demonstrates that unexplained variations in the outcome variable can lead to inaccurate CATE estimates and incorrect targeting policies. To address this issue, we propose using a surrogate index that leverages less noisy short-term purchases for long-term CATE estimation and policy learning. Furthermore, we introduce the separate imputation strategy to handle the non-separable nature of churn and purchase in marketing contexts. This involves constructing two distinct surrogate models, one for the observed last purchase time and the other for purchase frequency. Our simulation and real-world application show that (i) using short-term signals instead of the actual long-term outcome significantly improves long-run targeting performance, and (ii) the separate imputation technique outperforms existing imputation approaches.