MoEngage Predictions help predict the actions users are going to perform in the next few days. For example, predictions of users who are going to make a purchase, or uninstall the app or become dormant. All this information is very useful for the growth of your product or service and you can perform necessary and required actions.


With MoEngage Predictions you can: 

  • Predict users who are going to convert, or uninstall, or become dormant
  • Predict users who are going to perform an action defined by you
  • Predict for a specific action like predict users who are going to purchase from category fashion, predict users who are going to uninstall from Android platform, etc.
  • Create a campaign for the predicted users with few clicks.


MoEngage provide three out of the box prediction

  1. Conversion Prediction - In the next 7 days, the user executes the Conversion Event
    (Here conversion event is taken from the general Settings.)
  2. Uninstall Prediction - In the next 7 days, the user executes Device Uninstall.
  3. Dormancy Prediction - In the next 7 days, the user does not execute App/Site Opened.

With MoEngage prediction, you can have five live predictions at any time. You can archive an older prediction and create new predictions as well.

Sherpa, the machine learning module of MoEngage, uses all available data points and generates a propensity score of performing/not performing an action for each user. You can choose the propensity category (High, Medium, Low) and target users accordingly.

MoEngage prediction updates the propensity category for all users with a frequency for your choice. This way you can reach out to the user right before they uninstall or become dormant or do any purchase.

Create Prediction

To create a prediction Click on the + Create new in the sidebar and click on Prediction. You can also click on Predict > All Predictions and then click Create Prediction button.




Define Prediction

You can define:

  • What do you want to predict for
  • What conversion (uninstall, dormancy, and so on) means for your business and get the required predicted users.



On the create page you can create the desired prediction provide the following:


Field Description
Prediction name Type the name of the prediction
Prediction type

Choose one of the following:

  • Predict dormancy
  • Predict Uninstall
  • Predict Conversion
  • Predict Custom
Define Prediction

Define the prediction time only in multiples of 7 days.


Only one prediction for each of the default prediction types can be running at a time. For example, you can create uninstall prediction, archive it and create again.


Prediction Settings

You can modify the prediction definition and prediction settings of the default predictions as required. MoEngage prediction also lets you define the prediction of your choice, with custom prediction you can create a prediction with desired definition and settings.

Both action and inaction of users can be predicted for example users who are going to purchase or users who are going to not open the app. Also, you can apply an event attribute filter to predict for specific actions such as predict only for iOS platform or predict for category fruits, etc.


You can define how much data MoEngage sherpa should take to generate the prediction and set the frequency at which the prediction should be refreshed and updated propensity categories to be available. You can also define on which day of the week you want the prediction to be available.

For this phase, the last 60 day's data taken to generate the prediction and prediction is refreshed every 7th day.

Computation for the prediction may take up to 24-48 hours for the first prediction instance. From the second prediction instance onwards, it will be available on the selected day.


Prediction Result

Prediction results (Propensity categories) are saved in user properties and in custom segments (one for each propensity type), which can be used in segmentation and campaigns to reach out to desired users. MoEngage also creates custom segments using these user properties.




For the default predictions, the user properties and custom segment names are predefined and those can not be modified.

Prediction Type User Property and Custom Segment Example Actions for Predictions
Conversion Prediction

User Property

  • Propensity to Convert

Custom Segments

  • Propensity to Convert (High)
  • Propensity to Convert (Mid)
  • Propensity to Convert (Low)

When the propensity is 

  • High - No action taken
  • Mid - Nudge the users towards conversion
  • Low - Inform users of benefits of conversion
Uninstall Prediction

User Property

  • Propensity to Uninstall

Custom Segments

  • Propensity to Uninstall (High)
  • Propensity to Uninstall (Mid)
  • Propensity to Uninstall (Low)

When the propensity is 

  • High - Send communications with the value proposition of retaining the app.
  • Mid - Re-engage the users by providing details about best performance and advantages.
  • Low - Send communications of features and what's new
Dormancy Prediction

User Property

  • Propensity of Dormancy

Custom Segments

  • Propensity to Dormancy (High)
  • Propensity to Dormancy (Mid)
  • Propensity to Dormancy (Low)

When the propensity is 

  • High - Send communications with the value proposition.
  • Mid - Send communications with benefits and advantages of coming back to the web or app.
  • Low - Send communications of features and what's new.


For the custom prediction, you can define the name and description of the user property. Custom segment names are automatically created by appending the propensity category at the end.




Click on the Create Prediction button to create a prediction.


These user properties and custom segments are available by default in segmentation dropdowns.





All Predictions Page

All of the created predictions will get listed on the all prediction page.




All prediction page displays the following:

  • Prediction name
  • Prediction type
  • Status of prediction
  • Prediction created date 
  • Prediction last refreshed
  • Prediction quality
  • User distribution between different propensity buckets. 

You can archive any prediction by clicking on the 3 dot menu and then clicking on Archive.

Prediction quality is a measure of accuracy for the prediction. It tells how much the prediction is capable of distinguishing between users performing an action and not performing an action. Higher the quality, the better the prediction. A good prediction should have a quality score greater than 60. This is calculated using the standard method of Area Under the ROC curve.

You can filter the prediction by prediction name, created date, type, and status.

Click on any of the predictions to view the prediction details page.


Prediction Details


The prediction detail page shows prediction details such as ID, Type, Created Time, and Last Run Time.

You can click on Prediction Details for instance and choose the particular instance to display the details.




Prediction overview provides the count of users present in different propensity buckets along with Prediction quality.

You can click on Recommendation to view the recommended actions for the respective prediction.

You can create a campaign with the respective propensity bucket, just click on the +Create campaign button and chose the channel and campaign type to create a campaign from the prediction.




Propensity distribution

Propensity Distribution is the distribution of users in the decile buckets and also in categories of high medium and low with respect to the propensity.

Decile buckets (1-10%, 11-20% etc.) of propensity shows the user count in the respective propensity range. Distribution by category shows the high (propensity > 70%), mid (propensity 40-70%), and low (propensity < 30%) propensity respectively.

For example, if a user purchases an item the propensity is 70%, if a user has items in the carts then propensity is 40% and if the user has no items in the cart or not viewed any products propensity is 10%

Click on the download button to download this information in a chart or CSV file.




Correlation parameters

Correlation is the measure of the strength of the linear relationship between correlation parameters and predictions. An increase in the positive correlation parameter increases the probability of predicted action or no action. An increase in the negative correlation parameter decreases the probability of predicted action or no action.

For example, when items are added to the cart, the action is positive.

Click on the download button to download this information in a chart or CSV file.




 Instance Details

Select Prediction Details for instance to display specific details about the prediction instance. Details such as

  • User segment for which the prediction is performed
  • Definition of prediction
  • User attribute and custom segments where the results are saved
  • Forecast period
  • Lookback duration (data analysis duration)
  • Prediction start and end time.





Was this article helpful?
2 out of 2 found this helpful