Churn Prediction and Retention Modelling
Churn models built around a question most projects skip: not who is likely to leave, but which of them would stay if you did something, and what that something is worth.
A churn model that identifies customers who are about to leave is easy to build and frequently useless. Many of them have already decided, and contacting them wastes the retention budget or, worse, reminds them to cancel. The valuable question is narrower and harder.
Churn prediction is the modelling of which customers are likely to stop buying or subscribing within a defined window. Retention modelling goes further by estimating which of those customers would respond to a specific intervention, so that limited retention effort is spent on the persuadable rather than on the highest-risk.
Risk is not the same as opportunity
Standard churn scoring ranks customers by probability of leaving. Uplift modelling ranks them by how much an intervention would change that probability, which splits your customer base into four groups that need different treatment:
| Group | Behaviour | Correct action |
|---|---|---|
| Persuadable | Stays if contacted, leaves if not | Spend the retention budget here |
| Sure thing | Stays either way | Contact wastes budget and margin on discounts |
| Lost cause | Leaves either way | Contact wastes budget and effort |
| Do not disturb | Leaves if contacted, stays if not | Contact actively causes the churn |
That last group is real, particularly in subscription businesses where a retention call reminds a dormant customer that they are paying. A pure risk model targets it enthusiastically, because those customers do look like they are about to leave.
How we build it
Define churn precisely
For contractual businesses it is a cancellation event. For non-contractual ones it is an absence, and someone has to decide how long an absence counts, which differs by segment and by product. This definition decides everything downstream and is usually contested until it is written down.
Pick a window that leaves time to act
Predicting churn within thirty days is accurate and useless if your intervention takes six weeks to have an effect. The prediction window is set from how long the retention action needs, not from what produces the best model score.
Model uplift where you have the data
Uplift modelling needs historical experiments: customers who were contacted and comparable customers who were not. Where that exists, it is the highest-value part of the engagement. Where it does not, we build the risk model and design the experiment that will produce the data, so the second version can do it properly.
Explain in terms retention can use
Usage down forty per cent quarter on quarter, two unresolved support cases, no login for eighteen days. A retention agent needs a reason to open the conversation with, and an unexplained risk score gives them nothing to say.
Match the volume to the capacity
The threshold is set by how many customers your team can genuinely reach with a meaningful intervention, not by the model's optimal cut-off. A list of eight thousand at-risk accounts handed to a team of five is a report, not a programme.
Measure with a holdout
A portion of high-scoring customers is deliberately left untreated so the effect of the programme can be measured rather than assumed. Without it, every retained customer gets credited to the intervention and the programme's value is unprovable in either direction.
Discounting is not the only intervention
The default retention action is a discount, which is measurable, immediate and margin-destroying. Onboarding help, a feature people are not using, a support case actually closed, or a different plan are frequently more effective and considerably cheaper. Uplift modelling per intervention is what tells you which.
What tends to predict churn
- Engagement trajectory rather than level. A decline from heavy usage matters more than consistently light usage.
- Support experience. Unresolved cases, repeat contacts and long resolution times are among the strongest signals in most businesses.
- Onboarding completion. Whether the customer ever reached the point of value, often months earlier.
- Breadth of use. Single-feature or single-user accounts churn far more readily than embedded ones.
- Commercial events. Renewal dates, price changes, contract end and payment failures.
- Relationship changes. The champion leaving is frequently the single strongest predictor in B2B.
How the engagement runs
The churn definition and the action window come first; they determine whether the model is usable.
Definition and data audit
Churn defined per segment, prediction window set from intervention lead time, historical experiment data assessed.
Risk and uplift modelling
Risk model built with time-based validation, uplift model where historical treatment data supports it.
Explanations and thresholds
Business-language drivers per account, volume matched to retention capacity, intervention mapping agreed.
Delivery and holdout
Scores into the CRM or retention tool with reasons, holdout group configured before any outreach begins.
Measurement and handover
First measured results against the holdout, with the experiment design for improving uplift modelling next.
What you receive
A retention list your team can work, and proof of whether working it helped.
Churn definition
Per segment, with the prediction window justified by intervention lead time.
Risk model
Deployed scoring with time-based validation and drivers per account.
Uplift model or experiment design
Persuadability estimates where data allows, or the test that will generate it.
Intervention mapping
Which action suits which driver, with cost per intervention.
Holdout design
Untreated control group and the measurement plan agreed before outreach.
Impact report
Retention effect against the holdout, by segment and by intervention type.
Is this the right engagement?
Worth being direct. Churn Prediction and Retention Modelling is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- Churn is material and a retention team or automated programme exists.
- You have usage, support and commercial history per customer.
- Interventions take a known amount of time to have an effect.
- You will accept a holdout so impact can be measured.
- Someone can decide what to offer and what it costs.
Choose something else if
- No one acts on the scores, in which case this is a report.
- Churn events are too few to learn from.
- The only planned intervention is a discount to everyone flagged.
- You need broad segmentation rather than churn specifically. See segmentation and propensity.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
What is the difference between churn prediction and uplift modelling?
Churn prediction estimates who is likely to leave. Uplift modelling estimates who would stay because of an intervention. They rank customers differently, and targeting on risk alone spends budget on people who were leaving regardless and on people who were never going anywhere.
How far ahead can you predict churn?
Far enough for your intervention to work, which is the constraint that matters. Shorter windows produce better model scores and less useful lists. We set the window from how long the retention action takes to have an effect, then accept the accuracy that comes with it.
Can contacting at-risk customers make churn worse?
Yes, and it is a documented effect. In subscription businesses a retention approach can remind a dormant customer that they are paying, and a risk-only model targets exactly those people. Uplift modelling exists to identify and avoid them.
What if we have never run a retention experiment?
Then uplift modelling is not yet possible, and we build the risk model while designing the experiment that will produce the data. That experiment is usually the highest-value thing the engagement leaves behind, because it makes the next version considerably better.
How do we prove the programme worked?
With a holdout group of high-scoring customers left untreated, agreed before any outreach starts. It feels wasteful and it is the only way to distinguish retention you caused from retention that would have happened anyway.
Often paired with this
Most clients combine two or three engagements from the Machine Learning & Predictive Analytics pillar. These are the ones that most often run immediately before or after.
Customer Segmentation and Propensity Scoring
Segments that map to actions and propensity scores per offer, activated in your marketing platform.
Read more →Predictive Analytics
Predictions embedded in the decisions they are meant to improve, measured on the business outcome rather than on AUC.
Read more →Recommendation Engine Development
Recommenders tuned to a business objective, with cold start, diversity and rules handled, proven by A/B test.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.