Cooking the Treats Right: How ML Helps Run Promotions Efficiently

Artemiy Sarantsev24 May 2025

Problem

A bank sends promo offers ("treats") to attract clients to a product or keep them in it. A promo notification beat a regular reminder more than 10 times in an A/B test, but conversion is the wrong success measure. Some responders collect the cashback and withdraw the money, and every send costs channel money. The team must find clients who respond and stay, and stop paying the others.

Example from the article: 2,000 rubles cashback for topping up an investment or savings account or opening a deposit, funds to stay about 30 days. Conversion rose from 0.1% to 1%, but about 40% of responders withdrew the funds right after the payout.

Solution in brief

Aim targeting at cutting two segments with negative economics: bonus hunters and the uninterested. Train a response model on everyone and a retention model on responders only, equalize the variance of both scores and multiply them. Choose the cutoff by profit per score bucket and roll out through canary deployment with A/B/C groups.

Questions this case answers
  • How to target offers so that spend goes to users who respond and keep using the product, and not to users who take the reward and leave
  • How to split a treatment-response target into two models trained on different populations and choose a cutoff by profit
  • How to target promo offers with ML so that bonus hunters and walk-ins receive less promo budget
  • How to choose the audience size of a promo campaign from model scores and unit economics

Tools and methods

  • Response model — probability of responding to the promo, trained on the whole audience
  • Retention model — probability of staying after the response, trained on responders only
  • QuantileTransformer — equalizes score variance before the scores are multiplied
  • Economic model — profit per score bucket
  • Canary deployment and A/B/C groups — safe rollout, monitoring and retraining
Continue reading — it's free

Google or a one-time email code. No password.