A subscription telecom operator had a retention budget and one blunt instrument: offer a discount to customers who called to cancel. By the time the call happened, the decision was often already made. The alternative — treating every customer as high risk — was worse. Blanket offers trained customers to threaten cancellation and gave away margin to people who were never going to leave.
The retention decision
The team could contact a limited number of customers each week, and its offers had different costs. The decision was not "who will churn?" on its own. It was "who is likely to leave, what are they worth, and which action is justified for this customer?" A score that ranks risk without value sends the most expensive offers to the wrong people.
What we worked with
Around one hundred features per customer, spanning:
- Usage: voice and data volume, peak and off-peak minutes, roaming, dropped and blocked calls, and changes in usage over time.
- Revenue and billing: average monthly revenue, overage charges, payment behaviour and handset or equipment age.
- Service quality: customer-care contacts, dropped-call rates and network experience indicators.
- Customer context: tenure, demographics, household composition, credit category and income band.
- Outcome: a churn flag.
Many variables were correlated with each other, a few were only available after churn, and the churn rate was low enough that accuracy would have been a misleading headline.
The traps we designed around
- Leakage at the cancellation month. Final bills, closure-related care contacts and post-churn usage predict churn perfectly and are useless in production. Everything was checked against the date the score is produced.
- Risk is not value. A high-risk, low-value customer and a high-risk, high-value customer deserve different treatment. Collapsing them into a single ranking makes the retention budget look efficient while it is being spent in the wrong place.
- Raw usage counts mislead. Minutes and megabytes are correlated with plan size more than with intent. Usage change, drop-offs and service friction carried more signal than absolute volumes.
- Contact fatigue. A retention programme that calls the same customers every month creates the churn it is trying to prevent. Frequency limits belong in the operating policy, not in a later fix.
How the worklist was built
- Define churn and the prediction window. We fixed what counts as churn, when the score is produced and how far ahead it must predict. A model that uses the cancellation month to predict cancellation is not a model.
- Audit for leakage. Usage and billing fields were tested against the prediction date. Post-churn activity, final bills and closure-related service contacts were excluded.
- Reduce the feature set. Correlation pruning and variable importance cut the hundred-column input to a smaller, defensible set. Lower usage, more care contacts and equipment age consistently carried signal; many raw counts did not.
- Compare interpretation with performance. A logistic baseline provided reason codes and a sanity check. A gradient-boosted challenger captured interactions. Both were evaluated on AUC, KS, lift and calibration, not accuracy.
- Translate lift into capacity. The threshold came from the number of customers the team could contact each week. We reported how many churners would be captured in the slice the team could actually work.
- Add value to risk. Expected margin over the next twelve months was combined with churn probability to produce a value-versus-risk matrix. The score ranked; the matrix decided where to spend.
- Design the test before the launch. A holdout group and a pre-agreed success metric were part of the delivery, so the programme could be measured against a counterfactual instead of against last quarter.
What this changes
- Retention moved from reactive to planned. The team worked a weekly list instead of waiting for cancellation calls, and could see why each customer was on it.
- Offers were sized to value. The most expensive interventions stopped going to customers whose expected lifetime value did not justify them.
- Someone owned the threshold. Capacity, contact frequency and offer cost became an explicit operating policy rather than an implicit habit.
- Success became measurable. With a holdout in place, the team could separate genuine retention from customers who would have stayed anyway.
What we delivered
- A calibrated churn score with customer-level reason codes.
- A weekly retention worklist sized to team capacity.
- A value-versus-risk matrix for intervention design and offer selection.
- A holdout test plan and a monitoring dashboard for score drift and retention outcomes.
Where this applies
This fits telecom, insurance, banking, utilities and any subscription business with a retention budget. The minimum inputs are a stable customer identifier, a churn definition with a date, at least a year of usage and billing history, and a known contact capacity. The first deliverable is usually the leakage review and the value model — because a churn score without a value dimension tends to optimise for the wrong customer.
This is an illustrative case study built on an anonymized dataset; it does not report results from a specific client engagement. It reflects our standard method for subscription churn analytics.
