Creating a Customer Churn Indicator

The Churn Model is a predictive analytics tool within the AI Workbench. It uses your historical customer data to understand patterns and predict the likelihood of a customer, or a group of customers, discontinuing their engagement with your products or services. The model leverages sophisticated algorithms and allows for customized churn conditions to provide actionable insights that can inform your retention strategies.

Predict churn probabilities at both profile and group levels.

The model can predict churn risk for individual customers (the profile level), allowing for highly personalized interventions. It can also predict churn for a collection of customers (the group level), enabling broader retention campaigns.

Use advanced algorithms to train and analyze predictive behaviors.

The Churn Model utilizes sophisticated machine learning algorithms, for instance Gradient Boosted Trees (GBTs) and Multilayer Perceptron Classifier (MLPC), to identify complex patterns in customer behavior that indicate churn risk. These algorithms learn from your historical data and continuously improve their accuracy over time.

Enable business-specific customizations like churn conditions and observation/prediction windows.

You can tailor the Churn Model to your specific business needs by defining what constitutes churn for your business (e.g., no purchase in 90 days, unsubscribing from a newsletter). You also control the timeframe for the historical data analysis (observation period) and the timeframe for the churn prediction (prediction period).

Churn Condition

This defines the specific criteria that indicate a customer has churned. It's the trigger that tells the model a customer is no longer engaged. For example, for a subscription service, a churn condition might be "No subscription renewal" or "Account cancellation." For an e-commerce business, it could be "No purchase in the last 90 days."

Prediction Period

This is the timeframe into the future for which the model predicts churn. For instance, you might want to predict churn for the next 30, 60, or 90 days.

Observation Period

This is the timeframe of historical data the model uses to identify churn patterns. It's the period the algorithm ‘learns’ from. A longer observation period can provide more data for the model to learn from, but it also means older, and thus perhaps less relevant, data might influence predictions.

You can define churn conditions using the AI Workbench's flexible editor, choosing from four options:

1. Visual Editor. A user-friendly GUI interface for building formulas without coding.
2. Visual Editor and Text. Combines the visual editor with a read-only text view of the formula.
3. Advanced: Visual Editor and Editable Text. Allows you to edit the formula directly as text while also seeing the visual representation.
4. Advanced: Editable Text. For users comfortable with writing formulas directly as text.

The AI Workbench provides pre-built churn condition templates to get you started quickly. These templates cover common churn scenarios, such as customers not placing orders within a specific timeframe or exhibiting low spending patterns.

Example: "No Orders in the Last 90 Days" Template

This template defines a churn condition where a customer is considered churned if they haven't placed an order in the last 90 days. The formula looks like this:

count(Activities.Orders.OrderId, inLast (90 day)) = 0

This understanding of the Churn Model and its key components will allow you to effectively use the AI Workbench to predict and address customer churn. In the next section, we'll dive into the practical steps involved in configuring and running a Churn Model.

1. Go to Customers → Indicators from the left-hand side navigation menu.
2. Click Create new Indicator.
3. Choose Churn under Predictive Indicators on the Create New Indicator popup and click Create.
4. Choose the prediction time frame, segments, processing purposes and the churn criteria/condition.

This is a simplified view of the previous churn model settings. If you need more advanced or flexible settings, you still need to go to the AI Workbench and fine tune the model run settings.

Once you click Save, the new indicator will be activated and calculated. You will be able to see the new indicator values on customer profiles or groups when the calculation is completed.

1. Access the AI Workbench

   First, choose AI Workbench from the top Entity dropdown. After that, from the AI Workbench side panel, choose the Manage Models icon.

   
2. Choose the Churn Model Type

   Select Churn (Profile) to predict churn at the profile level, or

   Select Churn (Group) to predict churn for customer groups

   
3. Define the Model Details

   Predictive Indicator Name

   Give the churn model a name that reflects its purpose (e.g., "VIP Churn Prediction").

   Provide a description (optional).

   

   The name is considered to be the name of the predictive indicator, hence the field name. Once you complete the model run, a new predictive indicator will be created in the CDP, but it’s not activated by default.

   
4. Segment and Group Selection (optional)

   Choose the target population or a subset of your analytics data by selecting a segment. For example, a segment could be a set of profiles or groups from a particular city, e.g., "Customers from New York".

   For group-based predictions, specify the relevant group type (e.g., product category or subscription type).

5. Set Processing Purposes (optional)

   Ensure data processing permissions are configured under Processing Purposes.

   

   If you don't choose a processing purpose, the model will run on all profiles or groups and their orders, i.e., no matter whether or which processing purposes are granted.

6. Adjust Churn Conditions

   Define a new churn condition or choose one from the Templates (for instance "No Orders in the Last 90 Days").

   

   The churn conditions will be used to determine what constitutes a churned profile or group. You can use built-in templates or create a custom churn condition in the formula editor.

7. Customize Prediction Parameters.
   • Set Prediction period (for instance, 6 months).

     It is the window of time for which the model produces the dependent variables or predictions.

   • Define the Observation period (for instance, 12 months prior).

     It is the period of time that determines the independent variables of the model.

   • Configure Churn probability thresholds for high/low-risk categorization.

     Profiles or groups that are predicted to have a risk of churning at or above/below this probability will be labeled high/low risk.

8. Advanced Settings (Optional).

   Activate a specific algorithm and configure Evaluation metric(for instance, F1-score), Split ratio, and cross validation options.

   

   The advanced section is a powerful tool for data scientists. It allows you to fine-tune your model by selecting different algorithms and evaluation metrics. However, this section is optional and disabled by default. If you don't need to use it, simply move on to run the configured churn model.

   If you want to explore the advanced section, just toggle it on. Once it's open, you can further customize the algorithm you've chosen.