Uplift Modelling and Doubly Robust Causal Learning for Bank Marketing Targeting: Optimizing ROI with Coverage–Incremental Profit Curves

Abstract

Marketing decisions should optimize incremental business value rather than predictive accuracy alone. A response classifier that ranks customers by subscription probability may inefficiently allocate budget by targeting individuals who would subscribe regardless of marketing intervention. Uplift modeling addresses this limitation by estimating the incremental causal impact of an action and enabling ROI-oriented targeting. This study formulates the “who to contact” problem as a causal decision task using the UCI Bank Marketing Dataset, specifically the bank-additional.csv subset (n = 4,119). Since the dataset includes only contacted clients, the treatment is operationalized as channel assignment, where cellular represents treatment and telephone represents control, as recorded in the contact variable. The leakage feature duration is removed following the dataset documentation. Two causal estimators are implemented: a two-model uplift estimator (T-learner) and a doubly robust learner that regresses an augmented inverse probability weighted (AIPW) pseudo-outcome with cross-fitting. Targeting policies are evaluated on a held-out test set using doubly robust policy evaluation with logistic-regression propensity scores clipped to [0.01, 0.99], and performance is assessed through AUUC, Qini, and coverage–incremental profit curves. On the test set (n = 1,236; seed = 42), the two-model uplift method achieves the best uplift ranking (AUUC = 0.0491, Qini = 0.0246). Outcome prediction AUC is 0.726 for a covariate-only model and 0.722 for an S-learner including the treatment indicator, demonstrating that predictive accuracy on outcomes does not identify incremental impact. Under a profit model with revenue (r = 100) and channel cost (c = 1), the two-model uplift generates 5,627.8 incremental profit at 10% coverage and peaks at 7,517.7 around 40% coverage. Sensitivity analysis further indicates that uplift-based targeting becomes increasingly advantageous as intervention costs rise. Overall, the study demonstrates a reproducible framework for translating causal uplift estimates into ROI-optimized bank marketing decisions.