Subscriber churn prediction:
See the cancellation before it is written
Hardly any subscriber cancels on impulse - first they read less often, then not at all, then the cancellation arrives. An engagement score built on recency, frequency and volume spots this pattern weeks in advance and makes the rescue plannable.
This module is part of a six-part model calculation about a fictional regional media house (55,000 print and 25,000 digital subscriptions). All euro amounts are derived from the documented assumptions of the sample publisher; the levers follow published industry benchmarks and deliberately stay below their best documented values. The values shown are not results of real clients and not a performance promise. What values are achievable on your real data is established only in a pilot project.
01 The problem - Churn announces itself, but nobody is listening
Why the cancellation email is the wrong moment for the rescue
By the time the cancellation arrives, the decision was made long ago - discount offers work poorly then and train your base to haggle. The effective moment lies weeks earlier: where reading behaviour tips.
Exactly this tipping is in your data: gaps between visits grow longer, sessions shorter, the newsletter stays unopened. The Financial Times established the RFV score for this - Recency, Frequency, Volume - and steers its entire retention with it. No human watches 25,000 subscribers individually. A model does.
And the rescue costs a fraction of what winning the same subscriber back as new would cost.
02 The model - A risk score per subscriber, refreshed every week
Gradient boosting on engagement trajectories, calibrated against real past cancellations
The model delivers per subscriber a cancellation probability for the next 90 days - and, via SHAP values, the reason: dormant usage, an exhausted discount, a dead section. This becomes a weekly risk list with a matching action: content recommendations for some, benefit communication for others. The FT cuts cancellation rates by 10 percent through engagement steering along its RFV score alone.
03 Business impact - Rescued subscriptions instead of cancellation statistics
The lever: 15 percent fewer base cancellations through early, targeted intervention
No round number: every assumption comes from the sample publisher and is stored centrally. With your real figures only the input changes, not the method.
| Item | Value |
|---|---|
| Digital cancellations / year (4.5% × 12) | 13,500 |
| of which base from month 4 (60%) | 8,100 |
| Baseline: Addressable base cancellations / year | 8,100 |
| Lever: Reduction via early warning (15%) × 8 months × ARPU | 1,215 × 8 × €10.50 |
| Result: added revenue / year | €102,060 |
Assumptions of a sample publisher - in a real project your data replaces these values.
The 15 percent sits deliberately between the published benchmarks: the FT documents -10 percent via engagement steering, Mather Economics -17 percent in an A/B test of targeted interventions. Only late churn from month 4 onwards is counted - early churn belongs to Module 05, so nothing is counted twice. In the quarterly chart the decline refers to all cancellations (around -9 percent); the tile value of -15 percent refers to the addressable late churn from month 4 onwards.
04 Next steps in your publishing house
From the score to a weekly retention routine
We connect subscription system, web analytics and newsletter data - read-only, without touching your systems.
The model is validated against the real cancellations of the last 24 months. You see how accurately it would have warned in hindsight.
Weekly risk list into the CRM: "These 150 subscribers are tipping right now - recommended approach per segment attached."