Problem
A marketplace runs hundreds of promotions a day, and a blanket offer such as 30% off wastes margin on users who would have ordered anyway ("always buyers"). Only the "persuadables" create new orders, and a campaign's raw order count does not separate the two. The team has to decide who gets a fixed offer and, for flexible offers, how deep the discount should be, within a campaign budget.
Example from the article: DoorDash wanted restaurant-only customers to try grocery, retail and personal care, and previously sent the same coupon to every eligible user.
Solution in brief
Estimate each user's incremental effect of a promotion with causal learners (Double Machine Learning, DML, won offline), then pick the offer or discount that maximizes uplift minus a cost penalty under budget limits. Fixed offers are modeled as discrete treatments (who to target). Tunable discounts are modeled as a continuous treatment (how much to give).
Questions this case answers
- How to allocate a costly intervention across individuals so that incremental outcomes per unit of cost are maximized, when many recipients would respond anyway
- How to choose the intensity of an intervention per individual when historical intensities were assigned non-randomly and recorded imprecisely
- How to decide which users should get a fixed promotion so that coupons are not wasted on people who would order anyway?
- How to choose the discount depth for each user under a fixed budget?
Tools and methods
- Double Machine Learning (DML) — debiased uplift estimate for discrete and continuous treatments
- S-learner, T-learner — offline baselines for uplift
- Qini score and uplift@30 — offline model comparison
- Cost-aware optimizer — assigns the offer or discount per user under a budget
Google or a one-time email code. No password.