The myth of the ML model
Everyone wants a churn prediction model. The reality is that most B2B SaaS companies don't have clean enough data to train one, and wouldn't act on the outputs even if they did.
The better question is: what behavior predicts churn in your product?
Start with cohort analysis
Pull a list of customers who churned in the last 12 months. For each one, look at their activity 30, 60, and 90 days before they left. You will find a pattern. You always do.
Common leading indicators in B2B SaaS:
- Login frequency drops below once per week
- The key user who championed the purchase stops logging in
- Support ticket volume spikes with no resolution
- Usage of the core feature drops by more than 50%
Build a health score in a spreadsheet
Assign points to each indicator. Weight them by how strongly each one correlates with churn in your historical data. Review the list weekly with your customer success team. Call the accounts in the red zone.
That's your churn prediction model.
| Indicator | Weight | Trigger |
|---|---|---|
| Login frequency | 30 | < 1/week |
| Champion active | 40 | Last login > 14 days |
| Support spikes | 20 | > 3 unresolved tickets |
| Core feature usage | 10 | < 50% of baseline |
Why this works better than you expect
A model you understand and act on beats a black-box model you don't trust. Customer success teams engage more consistently with a score they helped define. The human in the loop is a feature, not a limitation.
It won't win a Kaggle competition. It will save customers.