Predict subscription churn

  • 5 minutes

You can create predictions to help predict missing values in your customer profile. This process is helpful because, with more complete profile data, you can better target your customers moving forward. However, performing predictions in Customer Insights - Data is more than predicting missing information; it's often about making predictions that relate to customer behavior, such as if a customer might discontinue using your services.

Customer Insights - Data includes a prebuilt customer churn prediction model that can help predict whether a customer is at risk for no longer using your company's subscription products or services. You can configure the customer churn prediction model from the predictions page.

Anyone who configures the model needs contributor permissions in their Customer Insights - Data environment. Additionally, they should have a strong business knowledge of what churn means for your business.

The process to configure the customer churn model is as follows:

  1. Name model - Specifies the name of your prediction model that displays in Customer Insights - Data and the name of the output table that the system creates to store data that relates to your prediction.
  2. Preferences - Defines what constitutes churn for your organization, such as how many days to wait after a subscription end date before it's considered churned, and the time period for predicting attrition before a subscription end date.
  3. Add required data - Defines the relevant fields that the model uses to predict which customers are at a higher risk of churning, including defining subscription details and activities that support the subscription.
  4. Update schedule - Defines how often to retain the model for your predictions.
  5. Review and run - Allows you to review the prediction details before you run the prediction for the first time.

Name your model

The first step in the process is to provide names for your model and the output table. Comparable to the process of creating the missing values of prediction models, you need to provide a model name and output table name. This information is important because the name is how you identify the model in the application. The system creates the output table automatically when the model runs for the first time. Then, it populates any relevant data, such as scoring details, to the table.

Screenshot of the model name.

Define preferences

You must define two key items when you're creating the model to ensure an accurate prediction of churn and attrition. The first item is to define how many days after a subscription ends that a customer must renew before they're considered churned. For example, your organization might not process a renewal until after the current subscription expires. If it takes seven days for that process to complete, you would want to provide a seven-day buffer. You would provide this value in the Days since subscription ended field.

Screenshot of the model preferences.

The second factor to consider is when the system should start to determine the risk of a customer churning. For example, you might set this value to 90 days to align with your organization's marketing retention efforts. Defining what this value should be is highly dependent on your organization's specific business requirements. Predicting churn risk for longer or shorter periods of time could make it more difficult to address the factors in your churn risk profile.

Add required data

To ensure that your model can make accurate predictions, you need to specify the subscription and customer activity data that the model considers when making the prediction. Before you begin this process, make sure that the data sources that contain customer subscription and activity information are already ingested into the application.

The first item that you configure is the subscription history information. The data you're using should include subscription identifiers that distinguish the individual subscriptions, including customer identifiers, so that the system can match the subscriptions to your customers. The data should include event dates, which define start dates, end dates, and the dates that the subscription events occurred on and the details that relate to recurring subscriptions and frequency.

When you first configure the subscription history data, you need to ensure that it matches with the customer table that represents your primary customer table. Customer Insights - Data prompts you to define the relationship between the two items if one doesn't exist.

After identifying the table that you want to use and defining the relationship, you need to map the semantic fields to attributes in your subscription history table. This approach helps the application better identify information, such as when a subscription ends, when it's determining churn.

When you're mapping history data, make sure that your subscription history data has fields that include the following information:

  • Subscription ID - A unique identifier of a subscription.
  • Subscription End Date - Date that the subscription expires for the customer.
  • Subscription Start Date - Date that the subscription starts for the customer.
  • Transaction Date - Date when a change happened, such as buying or canceling the subscription.
  • Is it a recurring subscription - A true/false field that determines if the subscription renews with the same subscription ID without customer intervention.
  • Recurrence Frequency (in months) - Subscription renewal period represented in months.
  • Subscription Amount (optional) - The amount of currency that a customer pays for the subscription renewal. It can help identify patterns for various levels of subscriptions.

Screenshot of the subscription history entity map.

After you define the subscription data, complete the same process to add customer activity data. During the process, you define an activity type that matches the type of customer activity that you're configuring.

When mapping your activity data, make sure that the data source includes the following information:

  • Primary key - A unique identifier for an activity, such as a website visit or a usage record showing that the customer viewed a TV show episode.
  • Timestamp - The date and time of the event, identified by the primary key.
  • Event - The name of the event that you want to use. For example, a field called UserAction in a streaming video service could have the value of Viewed.
  • Details - Detailed information about the event. For example, a field called ShowTitle in a streaming video service could have the value of a video that a customer watched.

Screenshot of the customer activity entity map.

Note

You need to have at least two activity records for 50% of the customers for whom you want to calculate churn. If customer activities are in multiple entities, you need to repeat this process for each table that contains customer activities.

Set schedule and review configuration

Before you can start consuming your prediction model, set a frequency to retrain your model. This setting is important to ensure the accuracy of predictions as new data is ingested in Customer Insights - Data. While every organization differs, most businesses can retrain once each month and receive good accuracy for their prediction.

Before completing your prediction model, review the configuration to ensure that it's correctly configured based on your needs and then make changes as necessary. After all values configure correctly, you can save and run to begin the prediction process.

Note

Depending on the volume of data that's used in the prediction, the prediction process can take several hours to complete.

Review prediction status and results

After you configure and run a prediction, you can view the prediction from Predictions under the Intelligence area and by selecting the My Predictions tab.

Each defined prediction model displays the following information:

  • Prediction name - The name of the prediction that's provided when you're creating it.
  • Prediction type - The type of model that's used for the prediction.
  • Output table - Name of the table to store the output of the prediction.
  • Predicted field - This field is populated only for some types of predictions and isn't used in subscription churn prediction.
  • Status - The status of the prediction's run.
    • Queued - The prediction is currently waiting for other processes to run.
    • Refreshing - The prediction is currently running the score stage of processing to produce results that flow into the output table.
    • Failed - The prediction has failed.
    • Succeeded - The prediction has succeeded.
  • Edited - The date when the configuration for the prediction was changed.
  • Last refreshed - The date when the prediction refreshed results in the output table.

You can perform actions, such as editing the prediction model, refreshing the data, viewing the details, or deleting the prediction, by selecting the vertical ellipsis.

When you open a prediction to view the results, three primary sections of data display on the results page: Training model performance, Likelihood to churn (number of customers), and Most influential factors.

Training model performance

The Training model performance section shows a score that indicates the performance of the prediction. This score can help you make the decision to use the results that the output table stores.

The three available score options are: A, B, or C.

Scores are determined based on the following rules:

  • A - Model accurately predicted at least 50% of the total predictions, and the percentage of accurate predictions for customers who churned is greater than the historical average churn rate by at least 10% of the historical average churn rate.
  • B - Model accurately predicted at least 50% of the total predictions, and the percentage of accurate predictions for customers who churned is up to 10% greater than the historical average churn rate of the historical average churn rate.
  • C - Model accurately predicted less than 50% of the total predictions or when the percentage of accurate predictions for customers who churned is less than the historical average churn rate.

Screenshot of the training model performance.

Likelihood to churn (number of customers)

The Likelihood to churn (number of customers) section shows groups of customers based on their predicted risk of churn. This data can be beneficial if you want to create a segment of customers with high churn risk. Such segments help to understand where the cutoff should be for segment membership.

Screenshot of the likelihood of churn.

Most influential factors

The application considers many factors when you create your prediction. Each factor has its importance calculated for the aggregated predictions that a model creates. You can use these factors to help validate your prediction results. You can also use this information later to create segments that could help influence churn risk for customers.

Screenshot of the most influential factors.