- A churn score tells you who is likely to lapse; a lifetime value score tells you what that lapse would cost, and together they rank the retention list.
- Uplift models send an offer only to shoppers it changes, so you stop paying discounts to people who would buy anyway.
- The same churn, value and propensity scores work per point of sale, so a consumer brand can rank the stores and outlets it serves.
- H2O is the scoring layer under the marketing systems you already run.
Your churn model is a list, not a decision
Most retailers already have a churn model. It produces a list of shoppers at risk, and the list gets the same offer, whether the shopper spends ten dollars a year or ten thousand.
That costs you twice. High-value customers still leave because the offer was not worth their attention, and loyal customers who were never going to leave take a discount they did not need.
How retail customer churn and lifetime value work
Churn and lifetime value are predictive AI models trained on your own history, running where your customer data already sits.
Step 1: build one view per customer
Join transactions, loyalty activity, returns, service contacts, browsing and campaign responses, across stores and online.
Step 2: score churn
H2O Driverless AI trains a classification model that predicts who will lapse in the window you choose, with automatic feature engineering on recency, frequency, basket and channel mix.
Step 3: score lifetime value
A regression model predicts each customer's spend or margin over the next period, from the same features.
Step 4: rank the list
Combine the two scores, so the top of the list is the customers who are both likely to leave and worth keeping.
Step 5: explain every score
Reason codes from explainable AI, such as fewer visits, a return or a service complaint, show your team why each customer is on the list.
Step 6: monitor and retrain
Your team watches for model drift as seasons, prices and assortment change, and retrains when the scores move away from what happens.
Uplift: skip the offer that would not change anything
A churn score says who might leave; it does not say who an offer will change. Uplift modeling answers that question by comparing customers who received an offer with a hold-out group who did not, so an offer goes only to shoppers a positive uplift shows it would help. Customers who would buy anyway keep full price, and customers an offer would annoy are left alone.
Scoring by point of sale, not only by shopper
Consumer brands and distributors often sell through thousands of stores, kiosks and outlets rather than to shoppers directly. The same models work at that level: churn risk, lifetime value and product propensity for each point of sale, so a sales rep can see which outlets are slipping and what changed before the visit.