Predictions

Predictions, Segment’s artificial intelligence and machine learning feature, lets you predict the likelihood that users will perform any event tracked in Segment.

With Predictions, you can identify users with, for example, a high propensity to purchase, refer a friend, or use a promo code. Predictions also lets you predict a user’s lifetime value (LTV).

Segment saves predictions to user profiles, letting you build Audiences, trigger Journeys, and send data to downstream destinations.

For more details on AI usage and data, see Predictions Nutrition Facts Label.

On this page, you’ll learn how to build a prediction.

Build a prediction

Follow these steps to build a prediction:

1. In the Trait Builder, click Predictions, select the prediction you want to create, then click Next..
   • (For custom Predictive Goals) Add a condition(s) and event(s) to predict.
   • Select the event and (optional) property that you want to use to make a prediction.
2. Select a time period for the prediction.
3. (Optional) In Include all events, uncheck any events you don’t want Segment to factor into the prediction.
4. Click Calculate. If you’re satisfied with the available data, click Next.
5. (Optional) Connect a Destination, then click Next.
6. Add a name and description for the Trait, then click Create Trait.

Keep the following in mind when you build a prediction:

• Segment lets you predict the likelihood of a customer performing multiple events.
• You can choose a time period of 15, 30, 60, 90, or 120 days.
• You have granular control over the events Segment factors into the predictive model. By default, Segment’s model makes predictions on all events sent to Engage. Segment lets you exclude events you don’t want included by unselecting Include all events, then filtering out any events you want excluded from the model.

In the next section, you’ll learn more about the four available predictions.

Choosing a prediction

Segment offers four predictions: Custom Predictive Goals, Likelihood to Purchase, Predicted LTV, and Likelihood to Churn.

Custom Predictive Goals

Custom Predictive Goals require a starting cohort, target event, and quality data.

Starting cohort

When you build a Custom Predictive Goal, you’ll first need to select a cohort, or a group of users, for which you want to make a prediction. Traits with small cohorts compute faster and tend to be more accurate. If you want to predict for an entire audience, though, skip cohort selection and move to selecting a target event.

Target event

The target event is the Segment event that you want to predict. In creating a prediction, Segment determines the likelihood of the user performing the target event. Segment lets you include up to two target events and an event property in your prediction.

Access and data requirements

In machine learning, better data leads to better predictions. Because Segment prioritizes trust and performance, Segment has a number of data checks to ensure that each prediction is reliable and of high quality. Segment provides guidance in the UI before you create a trait, but some checks only occur during model training. If a trait fails, you’ll see an error message and description in the UI.

This sections lists Segment’s access and data requirements, service limits, and best practices for Predictions.

Definitions

• Feature Window: The past time period that contains the data used for model training.
• Target Window: The time horizon for which you want to make the prediction. You can select this in the UI for each trait.
• Target Event: The event predicting the likelihood of customer action.

For example, to predict a customer’s propensity to purchase over the next 30 days, set the Target Window to 30 days and the Target Event to Order Completed (or the relevant purchase event that you track).

Predictions access requirements

To access Predictions, you must:

• Track more than 1 event type, but fewer than 2,000 event types. An event type refers to the total number of distinct events seen across all users in an Engage Space within the past 15 days.
  • If you currently track more than 2,000 distinct events, reduce the number of tracked events below this limit and wait around 15 days before creating your first prediction.
  • Events become inactive if they’ve not been sent to an Engage Space within the past 15 days.
• To prevent events from reaching your Engage Space, modify your event payloads to set integrations.Personas to false.
  • For more information on using the integrations object, see Spec: Common Fields, Integrations, and Filtering with the Integrations object.
  • Analytics.js example: analytics.track("Button Clicked", {button:"submit form"}, {"integrations":{"Personas":false}})

Successful trait computation

This table lists the requirements for a trait to compute successfully:

Selecting events (optional)

Some customers want to specifically include or exclude events that get fed into the model. For example, if you track different events from an EU storefront compared to a US storefront and you only want to make predictions using data from the US, you could unselect the events from the EU space. This step is optional, Segment only recommends using it if you have a clear reason in mind for removing events from becoming a factor in the model.

Likelihood to Purchase

Likelihood to Purchase is identical to Custom Predictive Goals, but Segment prefills the Order Completed event, assuming it’s tracked in your Segment instance.

If you don’t track Order Completed, choose a target event that represents a customer making a purchase.

Predicted Lifetime Value

Predicted Lifetime Value predicts a customer’s future spend over the next 120 days. To create this prediction, select a purchase event, revenue property, and the currency (which defaults to USD). LTV is only calculated for customers that have performed the selected purchase events 2 or more times. The following table contains details for each property:

Likelihood to Churn

Likelihood to Churn proactively identifies customers likely to stop using your product. Segment builds this prediction by determining whether or not a customer will perform a certain action.

To use Likelihood to Churn, you’ll need to specify a customer event, a future time frame for which you want the prediction to occur, and if you want to know whether the customer will or won’t perform the event.

For example, suppose you wanted to predict whether or not a customer would view a page on your site over the next three months. You would select not perform, Page Viewed, and at least 1 time within 90 days.

Churn predictions are only made for eligible customers. In the previous example, only customers that have performed Page Viewed in the last 90 days would be eligible to recieve this prediction. The Segment app shows you which customers are eligibile to recieve this prediction.

Segment then uses this criteria to build the prediction and create specific percentile cohorts. You can then use these cohorts to target customers with retention flows, promo codes, or one-off email and SMS campaigns.

Use cases

For use cases and information on how Segment builds prediction, read Using Predictions.