Problem
A bank offers a promotional renewal of a banking product: the client gets better terms for a new term, the bank keeps the volume longer. The promotion costs margin, so it pays off only for clients it actually changes. Offering it to clients who would have renewed anyway is an overpayment, and contacting a "sleeping" client may trigger churn. A model trained on the binary target "opened under the promotion" cannot tell these cases apart, so a campaign can lose money while the ML metrics look fine.
Example from the article: at VTB the model's top response clusters showed +146%, +100% and +70% response uplift, yet only the top two (+2.2% and +1.0% of control potential) added margin in the pilot, and the remaining eight were negative.
Solution in brief
Train the uplift model on randomized pilots at several promotion sizes, then judge it in money, not in response. Compute margin as margin × volume × term, take the break-even uplift threshold with a short-term correction, and extend the effect over a one-year renewal cascade. Select clusters above the threshold.
Questions this case answers
- How to choose which customers get a promotional offer so that the margin effect is positive when a model trained on response loses money
- How to compute a break-even uplift threshold for an offer that costs margin and extend the effect over repeated renewals
- How to choose which clients get a promotional offer so that the margin effect of the promotion is positive
- How to validate an uplift model by business metrics rather than by response uplift alone
Tools and methods
- Uplift models — class transformation (baseline), uplift trees and Causal Forest, S/T/X meta-learners
- Qini curve, AUUC, Qini in money — comparing models and choosing the deployment cut-off
- Break-even threshold and retention cascade — economic policy on top of the model
- PySpark / SQL, CatBoost — pipeline stack
Google or a one-time email code. No password.