Churn prediction vs uplift modelling: which customers should you target?

Which customers should you target? For a specific retention offer, target customers whose estimated additional value covers its costs. Churn prediction identifies risk; uplift modelling estimates the difference the offer makes. Use experimentally supported uplift estimates to guide recipients, then account for contact and incentive costs.

What is the difference between churn prediction and uplift modelling?

Churn prediction asks who is at risk of leaving. Uplift modelling asks how a particular action changes an outcome compared with a control. For retention, that means estimating the difference in retention probability over the same window for the same customer profile.

In their review of uplift modelling, Pierre Gutierrez and Jean-Yves Gérardy define the task around an action’s incremental effect. Their framework estimates an average effect conditional on customer characteristics. They also explain why this remains an estimate: the same person’s outcome under both treatment and control cannot be observed.

If your starting question is what the churn score actually measures, see what predictive churn means. Here, the decision is who should receive an already specified offer.

Why not start with the highest-risk customers?

Because being likely to leave does not establish sensitivity to an intervention. In research combining two field experiments with machine learning, Eva Ascarza found that the highest-risk customers were not necessarily the best retention targets. Ascarza recommends examining differences in response within randomised trials, beyond the average effect.

Consider an illustrative renewal offer. The following inputs are invented model estimates for three customer profiles, not paired observed outcomes for individuals. Control means withholding this particular offer, and the renewal window is the same in both conditions. These control estimates are explicitly defined for the example; do not assume an existing churn score measures risk without the offer.

Renewal probability is the complement of churn in this example: A’s 80% churn means 20% renewal. Subtract renewal without the offer from renewal with it: A gives 22% − 20% = 2 percentage points; B gives 75% − 60% = 15; C gives 90% − 90% = 0.

A risk-ranked audience puts A first. An uplift-ranked audience puts B first. C looks attractive if the team looks only at renewal after receiving the offer, yet its estimated renewal probability is unchanged by that offer.

When does positive uplift pay for the offer?

For a recipient in a given renewal window, positive uplift pays when expected additional renewal value exceeds expected incentive and contact costs. This calculation excludes campaign setup costs and later customer value.

In the illustrative example, with equal $40 renewal value before campaign costs in both conditions, $4 paid on every treated renewal, $1 per contact and no other variable effects, only B has positive expected incremental profit per recipient: $2.00. A loses $1.08 despite positive uplift. This covers one renewal window and excludes campaign setup costs.

With renewal probabilities pT under treatment and pC under control, renewal value V, incentive d and contact cost c, the simplified rule is:

Expected incremental profit per recipient = (pT − pC) × V − pT × d − c

This formula is a simplified special case of the rule in Johannes Haupt and Stefan Lessmann’s targeting analysis, with equal renewal value before campaign costs in both conditions.

Applying the formula:

The incentive term uses renewal probability with the offer, not uplift. In this setup, payment is due on every treated renewal, including those that would have happened without the offer. Under this every-renewal incentive basis, charging the incentive only against additional renewals would understate the offer’s cost.

For the stated goal of positive incremental profit per recipient, B qualifies; A and C do not. A is the useful warning: even its positive retention uplift leaves the offer losing money under these assumptions.

This calculation concerns one renewal window and recipient-level economics. It does not include campaign setup costs or later customer value. If renewal value changes under treatment, or the incentive is paid on a different basis, revise the inputs and formula before using this worksheet.

What should you do before replacing a risk-ranked audience?

Before replacing a risk-ranked audience, specify the offer, control, renewal outcome and observation window. Retain pre-treatment characteristics, assignment and outcomes, reserve evaluation data, and record the offer’s economics before choosing recipients.

1. Write the comparison. Name the offer, control, renewal outcome and observation window.
2. Keep the learning records. Retain pre-treatment characteristics, treatment assignment and subsequent outcomes.
3. Reserve evaluation data. Gutierrez and Gérardy describe evaluating uplift through aggregate treatment/control comparisons on held-out data, including uplift curves.
4. Add the economics. Record renewal value, contact cost and when the incentive becomes payable before choosing recipients.

If comparable treatment/control data are missing, a churn score cannot supply the missing effect. A randomised pilot is a defensible next step. Observational methods are possible, but Gutierrez and Gérardy warn that an observational comparison requires treatment assignment and potential outcomes to be independent given recorded characteristics.

The resulting decision is specific to this offer and its control. Leaving A out of the illustrative offer does not mean abandoning A as a customer. If the decision includes different interventions, use the broader next-best-action framework to define those choices before deciding who receives each one.