Telecom Upgrade Campaign Targeting

An operator had several upgrade offers in market and a contact capacity far smaller than its customer base. Campaigns were built by ranking everyone on a single propensity score and contacting the top of the list until the budget ran out. The ranking included customers who would have upgraded anyway, and it sometimes reached customers who were already likely to leave — for whom the wrong offer made the situation worse.

The campaign decision

The weekly decision was not "who is likely to buy?" It was "which customer should receive which offer this week, given contact capacity, margin per offer and the risk of triggering churn?" A single propensity ranking cannot answer that, because it mixes three different outcomes and ignores the cost of the intervention.

What we worked with

• Customer snapshot: around 50,000 records and 230 anonymised attributes, numeric and categorical, covering usage, tenure, equipment, billing and service history.
• Three labelled outcomes: churn (about 7% positive), appetency (about 2%) and up-selling (about 7%) — each imbalanced in a different way.
• Offer economics: margin, contact cost and eligibility rules for each upgrade path.
• Contact capacity: the number of outbound contacts the team could make per week and the frequency limit per customer.

The feature names were anonymised, so the analysis could not lean on intuition about individual variables. Stability across time and consistency between the three targets carried more weight than a plausible-sounding name.

The traps we designed around

• Propensity is not incrementality. A high score includes customers who would have bought without contact. Ranking purely by propensity overstates campaign value; only a holdout shows the difference between what happened and what would have happened anyway.
• Three targets, one budget. A customer with high up-sell propensity and high churn risk is not simply a good lead. The offer, timing or channel may need to change — or the contact may need to be suppressed.
• Missing values carried signal. Blank fields often reflected how a customer was acquired or served rather than random gaps. Imputing them silently would have erased part of the pattern.
• Accuracy was meaningless. With positive rates between 2% and 7%, predicting "no" for everyone scores above 90%. Evaluation used lift at the capacity the team could actually work, plus calibration and expected margin.
• Contact fatigue. Repeated campaigns to the same segment train customers to wait for discounts. Frequency limits were part of the operating policy from the start, not a later correction.

How the targeting was built

1. Define the offer paths. Each upgrade option was described in terms of eligibility, margin, contact cost and the customer behaviour it is meant to change.
2. Separate the three targets. Churn, appetency and up-selling were modelled independently. Missingness indicators were kept as features because the pattern of blanks was informative.
3. Evaluate at the operating point. Models were compared on lift within the weekly contact capacity, precision-recall trade-offs and calibration — not on overall accuracy or AUC alone.
4. Convert scores into money. Expected margin combined offer value, conversion probability, contact cost and a penalty for customers whose churn risk would rise with the contact.
5. Allocate under constraints. The weekly list was produced by solving a simple allocation problem: eligibility, capacity, frequency caps and offer inventory, ranked by expected incremental margin.
6. Build the test before launch. A holdout group and a pre-agreed success metric were part of the delivery, so increments could be measured against a counterfactual rather than against last month.
7. Monitor what decays. Calibration, contact frequency and incremental conversion were tracked after launch, with a review cadence agreed with the campaign team.

What this changes

• The list is built around margin, not score. The most contactable customers are no longer automatically the most valuable ones to contact.
• Offers are matched to intent. Upgrade, retention and suppression become different actions rather than different names for the same campaign.
• Capacity is explicit. The team knows how many contacts it can make and what the next contact is expected to return.
• Success is measurable. With a holdout in place, the campaign can separate genuine incremental sales from customers who were going to buy anyway.

What we delivered

• Calibrated propensity models for churn, appetency and up-selling, with stability checks across time periods.
• An expected-margin calculation that combines offer value, contact cost and churn risk.
• A weekly allocation worklist sized to contact capacity and offer availability.
• A holdout test plan and a monitoring view for calibration drift and incremental conversion.

Where this applies

This fits telecom, banking, insurance, utilities and loyalty programmes where several offers compete for the same contact budget. The minimum inputs are a stable customer identifier, labelled outcomes for each offer, offer economics and a known contact capacity. The first deliverable is usually the offer economics and the evaluation design — because without them, a better model only produces a better-looking list that nobody can prove is worth calling.

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 campaign targeting under limited contact capacity.