Problem
A retail bank lets each customer pick monthly category cashback, and customers keep asking for the big categories: cafes, restaurants, supermarkets. Offering them to everyone would blow the loyalty budget, which runs to billions of rubles a month. The bank has to show each customer categories they like while the payout stays on target. One wrong optimizer run, such as supermarkets for everyone, could cost about a billion rubles extra per month.
Example from the article: Alfa-Bank picks 5 of 30 categories for each of 20 million customers every month. For a customer with a target effective rate of 1.3, a set at 1.42 interests them more, but the team picks a weaker set that hits 1.3.
Solution in brief
Forecast spend per customer and category, and score the customer's interest in each category. Then pick, for every customer, the set of 5 categories with the highest total interest whose forecast cashback cost matches that customer's target effective rate. The target rate itself is higher for customers with higher forecast CLTV.
Questions this case answers
- How to choose a personalized set of offer categories for each unit under a fixed payout budget, maximizing interest while hitting a target average rate
- How to make a per-unit payout target depend on the unit's forecast long-term value
- How to choose a personalized set of cashback categories for each customer within a loyalty budget
- How to hit a target effective cashback rate while maximizing customer interest
Tools and methods
- Spend forecast model — predicts spend per customer and category.
- Interest model — TF-IDF over spending shares.
- CLTV forecast — sets a personal target effective rate per customer.
- Optimizer — brute force over category sets with pruning.
Google or a one-time email code. No password.