Predict transaction churn

Transactional churn prediction helps predict if a customer will no longer purchase your products or services in a given time window.

You must have business knowledge to understand what churn means for your business. We support time-based churn definitions, meaning a customer is considered to have churned after a period of no purchases.

For environments based on business accounts, we can predict transactional churn for an account and also a combination of account and another level of information like product category. For example, adding a dimension can help determine how likely it is that the account "Contoso" will stop buying the product category "office stationery." In addition, for business accounts, we can also use AI to generate a list of potential reasons why an account is likely to churn for a category of secondary level information.

Tip

Try the transaction churn prediction using sample data: Transaction churn prediction sample guide.

Prerequisites

• At least Contributor permissions.
• At least 10 customer profiles, preferably more than 1,000 unique customers.
• Customer Identifier, a unique identifier to match transactions to your customers.
• Transaction data for at least double the selected time window such as two to three years of transaction history. Ideally at least two transactions per customer. Transaction history must include:
  • Transaction ID: Unique identifier of a purchase or transaction.
  • Transaction Date: Date of the purchase or transaction.
  • Value of the transaction: Currency or numerical value amount of the transaction.
  • Unique product ID: ID of the product or service purchased if your data is at a line item level.
  • Whether this transaction was a return: A true/false field that identifies if the transaction was a return or not. If the Value of the transaction is negative, we infer a return.
• Customer activity data:
  • Customer Identifier, a unique identifier to map activities to your customers.
  • Primary key: Unique identifier for an activity. For example, a website visit or a usage record showing the customer tried a sample of your product.
  • Timestamp: Date and time of the event identified by the primary key.
  • Event: Name of the event you want to use. For example, a field called "UserAction" in a grocery store might be a coupon use by the customer.
  • Details: Detailed information about the event. For example, a field called "CouponValue" in a grocery store might be the currency value of the coupon.
• Less than 20% of missing values in the data field of the table provided

For business accounts (B-to-B), add customer data aligned toward more static attributes to ensure the model performs best:

• CustomerID: Unique identifier for a customer.
• Created Date: Date the customer was initially added.
• State or Province: State or province location of a customer.
• Country: Country of a customer.
• Industry: Industry type of a customer. For example, a field called "Industry" in a coffee roaster might indicate if the customer was retail.
• Classification: Categorization of a customer for your business. For example, a field called "ValueSegment" in a coffee roaster might be the tier of customer based on the customer size.

Note

For a business with high customer purchase frequency (every few weeks) it's recommended to select a shorter prediction window and churn definition. For low purchase frequency (every few months or once a year), choose a longer prediction window and churn definition.

Create a transaction churn prediction

1. Go to Insights > Predictions.

2. On the Create tab, select Use model on the Customer churn model tile.

3. Select Transaction for the type of churn and then Get started.

4. Name this model and the Output table name to distinguish them from other models or tables.

5. Select Next.

Define customer churn

Select Save draft at any time to save the prediction as a draft. The draft prediction displays in the My predictions tab.

1. Set the Prediction window. For example, predict the risk of churn for your customers over the next 90 days to align to your marketing retention efforts. Predicting churn risk for a longer or shorter period of time can make it more difficult to address the factors in your churn risk profile, but it depends on your specific business requirements.

2. Enter the number of days to define churn in the Churn definition field. For example, if a customer hasn't made a purchase in the last 30 days, they might be considered as churned for your business.

3. Select Next.

Add purchase history

1. Select Add data for Customer transaction history.

2. Select the semantic activity type, SalesOrder or SalesOrderLine, that contains the transaction history information. If the activity has not been set up, select here and create it.

3. Under Activities, if the activity attributes were semantically mapped when the activity was created, choose the specific attributes or table you'd like the calculation to focus on. If semantic mapping did not occur, select Edit and map your data.

   

4. Select Next and review the attributes required for this model.

5. Select Save.

6. Add more activities or select Next.

Individual consumers (B-to-C)

Add additional data (optional)

1. Select Add data for Customer activities.

2. Select the semantic activity type that contains the data you would like to use. If the activity has not been set up, select here and create it.

3. Under Activities, if the activity attributes were semantically mapped when the activity was created, choose the specific attributes or table you'd like the calculation to focus on. If semantic mapping did not occur, select Edit and map your data.

4. Select Next and review the attributes required for this model.

5. Select Save.

6. Select Next

Business accounts (B-to-B)

Select prediction level

Most predictions are created at the customer level. In some situations, that may not be granular enough to address your business needs. Use this feature to predict churn for a branch of a customer, for example, rather than for the customer as a whole.

1. Select Select table for a secondary level. If the option is not available, ensure you have completed the previous section.

2. Expand the tables you would like to choose the secondary level from, or use the search filter box to filter the selected options.

3. Choose the attribute you would like used as a secondary level, then select Add.

4. Select Next.

Note

The tables available in this section are shown because they have a relationship to the table you chose in the previous section. If you don't see the table you want to add, ensure it has a valid relationship present in Relationships. Only one-to-one and many-to-one relationships are valid for this configuration.

Add additional data (optional)

1. Select Add data for Customer activities.

2. Select the semantic activity type that contains the data you would like to use. If the activity has not been set up, select here and create it.

3. Under Activities, if the activity attributes were semantically mapped when the activity was created, choose the specific attributes or table you'd like the calculation to focus on. If semantic mapping did not occur, select Edit and map your data.

4. Select Next and review the attributes required for this model.

5. Select Save.

6. Select Next

7. Optionally, select Add data for Customers data.

8. Map the semantic attributes to fields in your own customer data as identified. The data in the fields used should not change often to ensure the model's best performance. For example, selecting a field for "Classification" that changed every month would only have the last value used in the prediction, even though historically the same value might not apply to the customer when building the prediction patterns.

9. Select Next.

Provide an optional list of benchmark accounts

Add a list of your business customers and accounts that you want to use as benchmarks. The details for these benchmark accounts include their churn score and most influential features that impacted their churn prediction.

1. Select + Add customers.

2. Choose the customers that act as a benchmark.

3. Select Next.

Set update schedule

1. For the Data updates step, choose a frequency to retrain your model. This setting is important to update the accuracy of predictions as new data is ingested into Customer Insights. Most businesses can retrain once per month and get a good accuracy for their prediction.

2. Select Next.

Review and run the model configuration

The Review and run step shows a summary of the configuration and provides a chance to make changes before you create the prediction.

1. Select Edit on any of the steps to review and make any changes.

2. If you are satisfied with your selections, select Save and run to start running the model. Select Done. The My predictions tab displays while the prediction is being created. The process may take several hours to complete depending on the amount of data used in the prediction.

Tip

There are statuses for tasks and processes. Most processes depend on other upstream processes, such as data sources and data profiling refreshes.

Select the status to open the Progress details pane and view the progress of the tasks. To cancel the job, select Cancel job at the bottom of the pane.

Under each task, you can select See details for more progress information, such as processing time, the last processing date, and any applicable errors and warnings associated with the task or process. Select the View system status at the bottom of the panel to see other processes in the system.

View prediction results

1. Go to Insights > Predictions.

2. In the My predictions tab, select the prediction you want to view.

Individual consumers (B-to-C)

There are three primary sections of data within the results page:

• Training model performance: Grades A, B, or C indicate the performance of the prediction and can help you make the decision to use the results stored in the output table.

  Grades are determined based on the following rules:

  • A when the model accurately predicted at least 50% of the total predictions, and when the percentage of accurate predictions for customers who churned is greater than the baseline rate by at least 10%.
  • B when the model accurately predicted at least 50% of the total predictions, and when the percentage of accurate predictions for customers who churned is up to 10% greater than the baseline.
  • C when the 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 baseline.
  • Baseline takes the prediction time window input for the model (for example, one year), and creates different fractions of time by dividing it by 2 until it reaches one month or less. It uses these fractions to create a business rule for customers who have not purchased in this time frame. These customers are considered as churned. The time-based business rule with the highest ability to predict who is likely to churn is chosen as the baseline model.

• Likelihood to churn (number of customers): Groups of customers based on their predicted risk of churn. Optionally, create segments of customers with high churn risk. Such segments help to understand where your cutoff should be for segment membership.

• Most influential factors: There are many factors that are taken into account when creating your prediction. Each of the factors has its importance calculated for the aggregated predictions a model creates. Use these factors to help validate your prediction results. Or use this information later to create segments that could help influence churn risk for customers.

Business accounts (B-to-B)

There are three primary sections of data within the results page:

• Training model performance: Grades A, B, or C indicate the performance of the prediction and can help you make the decision to use the results stored in the output table.

  Grades are determined based on the following rules:

  • A when the model accurately predicted at least 50% of the total predictions, and when the percentage of accurate predictions for customers who churned is greater than the baseline rate by at least 10%.
  • B when the model accurately predicted at least 50% of the total predictions, and when the percentage of accurate predictions for customers who churned is up to 10% greater than the baseline.
  • C when the 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 baseline.
  • Baseline takes the prediction time window input for the model (for example, one year), and creates different fractions of time by dividing it by 2 until it reaches one month or less. It uses these fractions to create a business rule for customers who have not purchased in this time frame. These customers are considered as churned. The time-based business rule with the highest ability to predict who is likely to churn is chosen as the baseline model.

• Likelihood to churn (number of customers): Groups of customers based on their predicted risk of churn. Optionally, create segments of customers with high churn risk. Such segments help to understand where your cutoff should be for segment membership.

• Most influential factors: There are many factors that are taken into account when creating your prediction. Each of the factors has its importance calculated for the aggregated predictions a model creates. Use these factors to help validate your prediction results. Or use this information later to create segments that could help influence churn risk for customers.

An Influential feature analysis information page contains four sections of data:

• In the right pane, select an item from Top customers or Benchmark customers. Both lists are ordered by decreasing value of the churn score, whether the score is just for the customer or a combined score for customers and a secondary level like product category. The item selected determines the content on this page.

  • Top customers: List of 10 customers that are at highest risk of churn and lowest risk of churn based on their churn scores.
  • Benchmark customers: List of up to 10 customers that were selected while configuring the model.

• Churn score: Shows the churn score for the selected item in the right pane.

• Churn risk distribution: Shows the churn risk distribution across customers and the percentile the selected customer is in.

• Top features increasing and decreasing churn risk: Lists the top five features that increased and decreased the churn risk for the selected item in the right pane. Shows the value of the feature for that item and its influence on the churn score for every influential feature. The average value of each feature across low, medium, and high churn customer segments is also shown. It helps to better contextualize the values of the top influential features for the selected item and compare it with low, medium, and high churn customer segments.

  • Low: accounts or combinations of account and secondary level with a churn score between 0 and 0.33.
  • Medium: accounts or combinations of accounts and secondary levels with a churn score between 0.33 and 0.66.
  • High: accounts or combinations of accounts and secondary levels with a churn score greater than 0.66.

  When you predict churn at the account level, all accounts are considered in deriving the average feature values for churn segments. For churn predictions at the secondary level for every account, the derivation of churn segments depends on the secondary level of the item selected in the side pane. For example, if an item has a secondary level of product category (office supplies), then only the items having office supplies as the product category are considered when deriving the average feature values for churn segments. This logic is applied to ensure a fair comparison of the item's feature values with the average values across low, medium, and high churn segments.

  In some cases, the average value of low, medium, or high churn segments is empty or not available because there are no items that belong to the corresponding churn segments based on the above definition.

  The interpretation of values under the average low, medium, and high columns is different for categorical features like country or industry. Because the notion of "average" feature value doesn't apply to categorical features, the values in these columns are the proportion of customers in low, medium, or high churn segments that have the same value of the categorical feature as compared to the item selected in the side panel.

Note

In the output table for this model, ChurnScore shows the predicted probability of churn and IsChurn is a binary label based on ChurnScore with 0.5 threshold. If this default threshold doesn't work for your scenario, create a new segment with your preferred threshold. Not all customers are necessarily active customers. Some of them may not have had any activity for a long time and are considered as churned already, based on you churn definition. Predicting the churn risk for customers who already churned isn't useful because they are not the audience of interest.

To view the churn score, go to Data > Tables and view the data tab for the output table you defined for this model.