A decision-oriented empirical comparison of predictive and uplift-based scoring under budget constraints

Decision makers often operate under budget constraints and must allocate limited interventions across large populations. A common approach ranks individuals by predicted response probability and selects the top- users. However, predictive scoring targets outcome likelihood rather than incremental impact and may allocate resources to individuals who would respond even without intervention. Uplift-based scoring, in contrast, ranks individuals by estimated treatment-induced response and is therefore more directly aligned with intervention allocation. This paper presents a decision-oriented empirical comparison of predictive and uplift-based scoring under a unified budget-constrained top- targeting framework. Using a large-scale observational digital-lending dataset, we construct a temporally ordered evaluation design in which current-month features and treatment are mapped to next-month drawdown behavior. Predictive models estimate next-month response probability, while uplift-based models estimate incremental treatment effects. All scores are converted into the same top- policy and evaluated using inverse propensity scoring (IPS), self-normalized weighting (WIS/SNIPS), and doubly robust/augmented inverse probability weighting (DR/AIPW) policy-value estimators. The results show that predictive models provide reasonable response-prediction performance, indicating that they are not weak baselines. Nevertheless, under the unified budget protocol, uplift-based policies achieve higher DR/AIPW policy value than predictive policies across the evaluated budget levels, including when compared against the strongest predictive baseline. IPS estimates are more conservative and sometimes negative, reflecting high weighted control benchmarks among selected users. Supporting analyses using WIS/SNIPS, bootstrap uncertainty, weight trimming, hidden-confounding sensitivity analysis, placebo tests, selected-group decomposition, and response-type diagnostics are consistent with the DR/AIPW-centered findings. These findings provide dataset-specific empirical evidence that uplift-based scoring is better aligned with budget-constrained intervention allocation when the operational objective is incremental impact. The study highlights the importance of evaluating scoring models by the policy value of the decisions they induce, rather than by predictive accuracy alone.

Text 126743.pdf - Published Version Available under License Creative Commons Attribution. Download (4MB) Official URL or Download Paper: https://link.springer.com/10.1007/s10791-026-10251...

Actions (login required)

View Item