Setting up Churn prediction and reading the results

This page explains how to enable Churn prediction on a Customer Intelligence model, where to find the results on the contact, on the database and on audiences, and what to do if the calculation fails. If you first want to understand what it is and what it is for, read Predictive and Churn prediction.

Prerequisites

Two conditions must be met to enable Churn prediction on a Customer Intelligence model:

1. The model data source must be a data table. Churn prediction needs the history of the individual orders, so it is not available on models that use Contact Fields, where all you have is the aggregate data of last purchase date, number of purchases and total spend. If you use one of the native eCommerce integrations (Shopify, Magento, Prestashop) that feeds a data table, you are already halfway there.
2. The model data source must already be configured, because Churn prediction reuses the same fields of the orders table.

You also need enough order history: if the data is not enough, the calculation stops with an explicit message, as explained in If the calculation fails.

Enable Churn prediction on a model

1. Go to Profile Studio, Customer Intelligence, Models list and open the model.
2. Open Settings and select the Enable Churn checkbox (if you do not see the checkbox, contact our Customer Care).
3. Open the Data source tab: Churn prediction needs two extra pieces of information, which you provide here:
   • Field containing product ID: identifies the product purchased in each order line.
   • Field containing the quantity ordered: the quantity purchased for each product
4. Select the Predictive tab and fill in the fields described below. The calculation writes five pieces of information to five contact fields (you can create them automatically with the plus button):
   • Churn date (date): the date by which churn is expected.
   • Churn probability (percentage): the probability of churn.
   • Churn segment (single choice): the risk segment of the contact, from Loyal to Churned.
   • Previous churn segment (single choice): the segment the contact was in before the last change.
   • Date of last segment change (date): when the segment last changed.

Every model must have its own output fields: if you try to use a field that is already assigned to the configuration of another model, saving reports an error.

It is worth knowing what each field is for, because you will use them in different ways. Churn segment is the field you build audiences and workflows with, and the one you will look at every day. Churn probability is what you need when you want to be more selective than the segment, for example to reserve the most expensive incentive for those above 90%. Churn date tells you by when to act, so it is the field to set the timing of a recovery sequence on.

The last two fields, Previous churn segment and Date of last segment change, are the ones that make it possible to react to segment transitions. They are updated only when the segment of the contact actually changes, so you can build workflows that start, for example, when a customer moves from Stable to At risk, or when they start buying again after having been Churned.

Advanced parameters

You can use Churn prediction without touching a single parameter: the default values are calculated from the thresholds of your Customer Intelligence model, and the Calculate automatically options derive the two most delicate values from your orders, how many days of history to observe and after how many days without purchases a customer is considered churned. This is why the same configuration works on businesses with very different repurchase rhythms, and as a rule it is best to leave it that way.

Work on the parameters, by turning on Advanced, when you know your repurchase cycle better than the model can infer it, for example if you sell products with strong seasonality, or when you want to move the threshold beyond which you consider a customer at risk.

• Analysis period: the number of days in the period the model uses to analyze the characteristics of each contact. With Calculate automatically, the default value, it is determined from the orders in the table.
• Days between last purchase and churn: after how many days without purchases a customer is considered churned. With Calculate automatically, the default value, it is derived from the purchasing habits of your customers, and it is the recommended option.
• Minimum probability for high risk: a number from 0 to 1, 0.75 by default. Above this probability the contact is considered at high risk of churn.
• Minimum probability for medium risk: a number from 0 to 1, 0.4 by default. Above this probability, but below the high risk one, the contact is considered at moderate risk.
• Upcoming churn window (days): if the expected churn date falls within this number of days from the calculation, churn is considered imminent. For example, with 15 days all the contacts with an expected churn date in the next 15 days are at risk of imminent churn. By default it matches the highest Recency threshold of the model.

When the calculation runs

The calculation runs once a day, at the time of the first available run of the schedule of the Customer Intelligence model. If you change the schedule of the model, the churn time falls into line at the next save, so you only have to configure the schedule in one place.

In the model home you find the Churn calculation card with the date of the last calculation performed, the date of the next scheduled calculation and the Start calculation link to run the activity right away, for example after changing the parameters.

In the Models list the Churn column shows at a glance which models have the calculation enabled.

As with RFM analysis, the values on the contacts are updated only when the calculation runs. After changing the advanced parameters, start the calculation to see the new segments.

Our advice is to run churn after the synchronization of the orders, so that the analysis always works on the most up to date data.

Where you see the results

Churn prevention on the contact record

On the contact home the Churn propensity card appears, with one panel for every model that has calculated data on that contact. It shows three indicators:

• the current churn segment, with the previous segment and the date of the change, and an arrow that tells you whether the situation has improved or got worse compared to the previous period;
• the churn probability as a percentage, colored green, yellow or red as the risk grows;
• the expected churn date, highlighted when it is close.

Churn prevention of an audience or a database

Two cards appear in the views that list contacts, that is the database dashboard and the audiences.

Churn propensity segments

It is a pie chart with the distribution of the contacts across the six segments, and the total of the customers taken into account in the center.

Churn propensity 

It is a dashboard that answers the question that counts when you have to justify an investment in retention: how much churn is costing you and how much of it you are recovering. 

It brings together six financial indicators, with a period selector (Last 30 days, Last 6 months, Last year):

• Recovered customers: contacts that were in the Churned segment and today are not any more.
• Recovered revenue: how much those customers purchased in the selected period.
• Lost revenue: how much the customers whose churn date fell in the selected period had purchased in the same period of the previous year.
• Customers at risk: contacts in the At risk and Imminent churn segments.
• Revenue at risk: how much the customers at risk purchased in the period.
• Recoverable revenue: an estimate of the share of revenue you can still recover from the customers who have just churned, weighted on their engagement, where those who click your communications count 100%, those who open them 50% and those who are no longer subscribed are not counted. Compare it with lost revenue: if the two numbers are close, the customers you have lost still read you and a reactivation campaign has a good chance of working. If the recoverable one is much lower, it is better to move the investment onto the customers at risk, before they leave.

Both cards respect the filters applied to the view: if you filter the list of contacts, the numbers update on that subset. It is the quickest way to compare churn propensity across different segments of your database, for example between those who purchased on sale and those who purchased at full price. If you have several models with churn enabled on the same database, every model has its own panel.

If the calculation fails

The prediction model needs a minimum amount of data. If it does not find it, the activity ends in error with a message that explains the reason:

• The orders data table configured in the activity is empty: no order found, often because of a filter on the orders that is too restrictive or because a synchronization has not run yet.
• The order history is not sufficient: the period covered by the orders is too short to train the model.
• The number of customers is too low: there are too few distinct customers in the table.
• The number of orders recorded in the last 30 days is not sufficient: recent orders are missing, so the model has no reference for the present.
• There are not enough customers with at least 3 orders: you need customers who have purchased again, otherwise there is no repeated behavior to learn from.

In general Churn prediction gives the best results on a database with an order history of at least one year and a significant share of customers who have purchased more than once. If your eCommerce has just started, it is better to wait until you have built up some history.

If the message is about parameters or validation instead, check the fields indicated in the Predictive tab, in particular the product ID field and the quantity field, and try again.

Frequently asked questions

Can I enable Churn prediction on a model whose data source is Contact Fields?

No. It needs the history of the individual orders, so the model has to use a data table. If you have the orders in a table but the model was created with contact fields, create a new model with Data Table as the data source.

Can I have more than one model with churn enabled?

Yes. Every model has its own output fields and its own scheduled activity, and a panel per model appears in the cards.

Why does a contact have no churn data?

Because the calculation has not run yet, because the contact does not fall within the filter of the model, or because they have no orders in the data table.

To see how to use the five output fields in audiences, workflows and personalizations, read Predictive and Churn prediction.