After you set up Firebase Predictions, your users are dynamically segmented into groups based on their predicted behavior. Firebase Predictions has a broad spectrum of signals it can potentially use to make these predictions, including Analytics events that are automatically logged and that you explicitly logged, the user's device configuration, and basic attributes of your app, such as its binary size. However, for any given prediction, only some of the available signals are relevant.
As a hypothetical example, when predicting whether a user is likely to spend, a user's recent spending behavior is probably very relevant, whereas its binary size is probably not. However, in practice, the signals that actually have predictive relevance are often surprising and can change over time.
To gain insight into exactly what data went into making your predictions, you can explore your predictions in the Firebase console.
Open a prediction's details page
To explore a prediction:
Open the Predictions section of the Firebase console.
Ensure the app and platform you're interested in is selected from the drop-down menu at the top of the console.
If you have predictions set up for the app you selected, you will see a page that contains cards with information about the predefined predictions and any custom predictions you might have created for the app.
On the card for the prediction you want to explore, click Select an action > Explore this prediction.
The prediction details page opens.
What went into making this prediction?
Expand the What went into making this prediction section of the prediction details page to see what data went into making the selected prediction over the most recent 28-day period predictions are available (the exact timeframe is indicated at the top-right of the page).
The signals shown in this section were all used to determine whether each of your 28-day active users were predicted to perform the action in question. Also shown are the average values and value range of these signals over all users that were predicted to perform the action, given the risk tolerance you selected. By selecting different risk tolerance levels, you can see how the values of each signal affected prediction confidence.
While this page lists most of the signals the prediction models use, the models are frequently updated and improved, so some signals might not yet be included.
The data that contribute to a prediction belong to four high-level categories: user engagement metrics, event frequency, user properties, and app attributes. Note that some predictions might not use any signals in a category, in which event the category will have no items under it.
The following sections describe the signals that can appear under each category.
User engagement metrics
User engagement metrics measure how much users used your app over several timeframes.
Whether a user was active during a given timeframe.
Users are considered active if they trigger an Analytics event that
indicates user activity at least once in a given timeframe. The most
common event that indicates activity is
Click Active users to see the percentage of users that were active over each of several day-long and week-long intervals, out of all users that were predicted, with the selected risk tolerance, to perform the action.
For example, if you select "not_churn—low risk tolerance" and the histogram indicates 97% active users one week ago, that means that of all the users who are confidently predicted not to churn, 97% were active last week.
|Continuous active days in last 7 days||The number of continuous days a user was active in the last 7 days.|
|Continuous active weeks in last 4 weeks||The number of continuous weeks a user was active in the last 4 weeks.|
|Active days in last 7 days||The number of days in the last 7 days a user was active.|
|Active weeks in last 4 weeks||The number of weeks in the last 4 weeks a user was active.|
|Continuous inactive days in last 7 days||The number of continuous days a user was inactive in the last 7 days.|
|Continuous inactive weeks in last 4 weeks||The number of continuous weeks a user was inactive in the last 4 weeks.|
|Inactive days in last 7 days||The number of days in the last 7 days a user was inactive.|
|Inactive weeks in last 4 weeks||The number of weeks in the last 4 weeks a user was inactive.|
|Days since earliest active day||The number of days since a user first logged activity in your app.|
Firebase Predictions can potentially use any of the Analytics events your app logs to make predictions, whether the Firebase SDK automatically logs the event or you explicitly log the event.
Each of the Analytics events shown contributed to the prediction. You can click them to see the most common parameter values logged with the event and the percentage of users that triggered the event over each of several day-long and week-long intervals, out of all users that were predicted, with the selected risk tolerance, to perform the action. If one of the intervals is grayed out, for that time interval, too few users triggered the event for the event to be relevant.
For example, if you select "not_churn—low risk tolerance" and the
histogram indicates 77% users triggered the
spend_virtual_currency event one
week ago, that means that of all the users who are confidently predicted not to
churn, 77% triggered the
spend_virtual_currency event last week.
Firebase Predictions uses properties of the user's device to make predictions about the user's future behavior.
How recently the operating system on the user's device was updated.
The prediction details page shows the average OS freshness of all users in the selected prediction and risk tolerance segment, measured on a scale from 0.0 to 1.0.
How recently the user upgraded to a new version of your app.
The prediction details page shows the average app freshness of all users in the selected prediction and risk tolerance segment, measured on a scale from 0.0 to 1.0.
|User default language||The default language configured on the user's device.|
Firebase Predictions uses static attributes of your app to make predictions about a user's future behavior.
|App size||The size of your app on the device.|
|App ID||Your app's unique ID.|