Health scoring model

The model behind the health score, risk score, health levels and risk drivers

What it predicts

For every customer with a subscription, contract or recurring agreement, the probability that they churn in the coming period. The probability is transformed into a risk score from 1 to 10 and mirrored into the health score (11 minus risk), then bucketed into the three health levels Healthy, Neutral and Unhealthy. See Health, risk and engagement scores for the definitions and bands.

Data it needs

Required: subscriptions with start and end dates, payments, customer profile. Recommended: support tickets, communication and campaign engagement, product usage, account-manager activity, NPS or sentiment. Each connected source becomes a health breakdown group with its own sub-score, so more sources mean a more complete explanation.

How it works

The model learns from your own history: every customer-month in the past, whether the customer churned afterwards, and what their data looked like at that moment. Several model families (gradient boosting such as XGBoost and LightGBM, and regularised logistic regression) are trained and evaluated per organisation; the best performing one is deployed. Retraining happens on a schedule and after data changes, and a new model only replaces the current one when it passes an acceptance check. A separate engagement model trained on engagement features only produces the engagement score.

How often it runs

Scores are recomputed daily for every active customer. Yesterday's score, the previous level and the previous week's average and maximum are kept so that risk changes are visible.

How to read the output

• Health score and level tell you how urgent a customer is.

• Top risk drivers tell you why: the features that push the prediction up or down, whether the value is low or high, and whether you can influence it.

• Risk change drivers tell you what recently moved.

• Health breakdown tells you which part of the relationship is weak.

• Model Insights shows feature importance for the whole base and the realised churn rate per health score, so you can check that the score separates churners from stayers.

Where it surfaces

Home dashboard tiles, the customer and CSM pages, the Technical Dashboard, Model Insights, audience conditions and workflow triggers such as Customer health changed.