Analysis readout
What predicts churn
A gradient-boosted model on 18 months of subscription data, and what it says to do next.
AnalyticsSynthetic data, sample deckOctober 2026
The strongest signal is not price. It is the first two weeks.
Feature importance
Early activity outweighs plan price and support tickets.
| Feature | Share of importance |
|---|---|
| Sessions in days 1 to 14 | 31 |
| Setup completed | 22 |
| Seats invited | 17 |
| Plan price | 9 |
| Support tickets | 6 |
Retention by onboarding group
Teams that finish setup keep 2x as many users at month six.
| Month | Setup done | Not done |
|---|---|---|
| M1 | 96 | 88 |
| M2 | 90 | 71 |
| M3 | 85 | 58 |
| M6 | 76 | 38 |
Model comparison
| Logistic baseline | Gradient boosting | |
|---|---|---|
| AUC, hold-out | 0.74 | 0.86 |
| Top-decile lift | 2.1x | 4.3x |
| Explainability | Coefficients | SHAP values |
Caveats before anyone acts
What we know
- Early activity predicts churn well out of sample
- The effect holds across all three plans
What we do not
- Whether nudging setup causes retention
- Behaviour after the March pricing change
Recommendation
Test guided setup
Ask two-week A/B on new signupsanalytics@example.com