Predefined Predictions

Firebase Predictions predefines four user groups that are dynamically generated based on your analytics data.

Churn predictions

The churn group contains users who have been active during the last 7 days but are predicted to become inactive within the next 7 days. Note that churn is different from uninstalling your app. Users may churn long before they uninstall an app; in fact, many users will never uninstall an app short of replacing their device. However, Predictions uses app uninstalls as one of the signals to train its model.

The not_churn group contains users who have been active during the last 7 days and are predicted to continue to be active over the next 7 days.

Note that membership to the not_churn group is different from not being a member of the churn group. Users who are not members of the churn group include both members of the not_churn group and users for whom no prediction could be made with confidence.

Consequently, you should not target the negative of a prediction. When you do so, you also target users the model is unable to confidently predict. Instead, target users that have been affirmatively predicted not to perform an action, such as users in the not_churn group.

Note also that membership to the churn and not_churn groups is not mutually exclusive. All predictions make some errors, and it's possible, for example, that the churn model erroneously categorizes a user as likely to churn, while at the same time, the not_churn model correctly classifies the user as a non-churner. In a situation like this, both models likely have low confidence for the user, so by reducing the risk tolerance level, you can also reduce the number of users that get assigned to multiple groups.

Spend predictions

The spend group contains users who are predicted to make in-app purchases or ecommerce purchases (that is, users expected to log the in_app_purchase or ecommerce_purchase event) in the next 7 days. Subscriptions are not included in the spend (and not_spend) models.

The not_spend group contains users who are predicted not to spend in the next 7 days. As with the not_churn group, membership in the not_spend group is different from not being a member in the spend group, and you should generally not target users based on them not being a member of the spend group.