Synthetic base, 12,000 subscribers · calibrated to published churn bands
| Segment | Realised monthly churn | Published band |
|---|---|---|
| Consumer postpaid | 1.09% | 0.8 – 1.5% |
| Business | 0.72% | 0.5 – 1.0% |
| Prepaid | 4.52% | 3.0 – 6.0% |
Enforced in CI: a generator that drifts out of band fails the build. Risk model AUC 0.807, measured on 2,138 held-out control rows — a model fitted on the whole base would learn a rate last quarter's campaign already altered.
| Offer | Top 2,000 by risk | Top 2,000 by expected value | Expected value > 0 | Contacted |
|---|---|---|---|---|
| 18% × 6mo | €-36,223 | €-25,628 | €-8,521 | 442 |
| 10% × 6mo | €-22,025 | €-18,294 | €-7,013 | 593 |
| 10% × 3mo | €-13,152 | €-12,745 | €-6,314 | 880 |
| 5% × 3mo | €-8,715 | €-9,133 | €-5,261 | 1,122 |
| 3% × 3mo | €-6,940 | €-7,299 | €-4,703 | 1,297 |
| 2% × 3mo | €-6,053 | €-6,295 | €-4,365 | 1,410 |
Every row is negative. An 18% × 6-month offer on €23 ARPU costs about €29 including contact, against roughly €174 of margin at risk — so break-even needs about 17 percentage points of uplift. At 1.2% monthly churn the maximum possible uplift over 90 days is around 5pp. The arithmetic closes before targeting is discussed at all.
| Measure | Value |
|---|---|
| Correlation, estimated vs true individual uplift | 0.109 |
| Top 2,000 by expected value — estimated uplift | +0.1540 |
| Top 2,000 by expected value — realised uplift | +0.0231 |
| Shrinkage | 85% |
| Whole base — estimated / realised | +0.0121 / +0.0099 |
| Realised uplift, value-ranked vs risk-ranked | +0.0231 / +0.0245 |
The model is well calibrated in aggregate and nearly blind individually. Selecting on an estimate biases that estimate upward for whatever it selects — the optimiser's curse — so realised performance is effectively identical for both rankings. The quadrant guard removes 316 rows the model labels sleeping dogs and 29 it labels lost causes — but those are predicted labels, and at a correlation of 0.109 they should not be read as counts of real ones. In this base 13% of subscribers are genuinely harmed by outreach; the model cannot reliably say which.
The four-quadrant argument stays correct as a description of why propensity targeting misallocates. What it does not support is a claim that better ranking recovers a large sum on a conventional campaign.
In a 1%-monthly-churn market the binding constraint is offer economics. Most retention offers are unaffordable regardless of recipient, and no modelling fixes an offer that costs more than the margin it defends. The first question is not who do we contact but is this offer viable at all. And since aggregate uplift is measurable where individual uplift is not, treatment decisions belong at segment level, measured against holdouts — not as per-customer scores.