Manage Similar Customer Predictions

When you open a similar customer prediction, the details page helps you understand how the prediction is performing and how to use it in your campaigns. It summarizes what the prediction is measuring, shows key quality and refresh metadata, and helps you choose a similarity range to save as an audience.

Prediction recap

The Prediction Recap gives you a quick snapshot of what this prediction is measuring. What it displays depends on how the prediction was created:

• If the prediction was built from an existing audience, the recap shows the linked reference audience name as a clickable link.

  

• If the prediction was built from custom criteria, the recap shows the attribute criteria used to define the reference segment.

  

Long descriptions are truncated; hover over them to see the full text. Use this section to confirm that you’re viewing the right prediction, understand its overall scale (for example, seed size and eligible population), and get a quick sense of the opportunity before reviewing detailed results or saving a similarity range as an audience.

Prediction results

The Prediction Results section provides everything you need to understand how reliable your similar customer prediction is and how to use it for targeting. You can review key metadata such as prediction strength, when it was last refreshed, how often it updates (Refresh Frequency), and its current status.

The Target Range selector shows the estimated number of users and predicted similarity for different similarity ranges. The Likelihood Percentiles chart shows how users are distributed from least similar to most similar to your reference segment. Together, these views help you balance reach and precision, so you can decide which users to target and how narrow or broad your audience should be.

Prediction strength

Prediction strength reflects how accurately the model can identify similar customers. It is based on the quality of your data, the strength of behavioral patterns, and how well the prediction’s training process performed.

• Weak: The model strength is low. Results should be interpreted with caution and not used for campaign targeting.
• Moderate: The model strength is acceptable. There is a meaningful pattern, but still some uncertainty. You should use these results alongside other insights when planning a campaign.
• Strong: The model strength is high. Patterns are clear and consistent, so results are very reliable and should be used to inform campaign strategy.

If strength is lower than expected, verify that your dataset includes enough meaningful signals and that your seed segment is large and representative.

Last Refresh

Shows the last time the prediction was recalculated, according to the prediction’s refresh frequency.

Refresh Frequency

The refresh frequency is how often your prediction is recalculated using the latest available data. It tells you how frequently mParticle updates the prediction scores for your users, so you know how current the results are before using them to create and activate audiences.

Update a prediction’s refresh frequency

To update a prediction’s refresh frequency from the details page:

1. Click the three-dot action icon from the prediction’s detail view (or from the Predictive Attributes page).
2. Click Update frequency.
3. Use the left-hand dropdown to select one of the following intervals: Weekly, Monthly, or On demand.

4. Use the right-hand dropdown to specify when or how often (within the selected interval) the prediction should refresh.

   1. For example, if you select Weekly, choose which day of the week to refresh.
   2. If you select On demand, the prediction will only refresh when you click Run Now from the action menu.
5. Click Save.

Run now

You can recalculate a prediction manually by clicking Run Now from the prediction’s action menu, either on the detail view or from the Predictive Attributes page.

Running a prediction on demand updates its results immediately using the most recent data available.

Status

The status shows the current state of each prediction’s pipeline.

• Calculating: mParticle is in the process of generating predictions. Predictions that are currently being calculated cannot be used in audiences.
• Active: The prediction has been calculated and is ready to be used in an audience.
• Inactive: The prediction’s pipeline has not been run within the last 30 days.

• Failed: mParticle could not complete the prediction calculation. This can happen for the following reasons:

  • Data issue: The number of eligible users or reference users is too small (500 users or under) or there is an issue with the underlying data.
  • Incompatible criteria in the reference segment: The prediction detail page will display a message explaining the cause.
  • Development environment data: Similar Customer Predictions only evaluate production environment data. If the source audience was built from the development environment, the prediction will fail. Recreate the reference audience using production environment data.

Target range definition

The Target Range section helps you explore different slices of your scored users and choose the one you want to turn into an audience.

You can select a preset range or define your own, then compare how each option affects reach and similarity before you save it. For the selected range, mParticle shows:

• Number of Users: How many users fall within the selected percentile range.
• Predicted Similarity: How similar customers in the selected range are to your reference segment, displayed as a multiplier relative to the average similarity across all scored users.

Click the dropdown menu to select one of the preset ranges, or create a custom range:

• Most similar customers (0.82 to 1.00): Users with the highest predicted similarity to your reference segment. This is often the best option when you want maximum targeting precision.
• All similar customers (0.73 to 1.00): Users with predicted similarity above average. This option is commonly used when you want broader reach while still focusing on users who align with the reference segment.
• Custom top range: Lets you define your own top percentile range to reach a specific audience size or balance reach and similarity.

Save a target range as a new audience

Once you’ve identified the similarity range you want to target, you can create an audience directly from the prediction’s details page:

1. From the prediction details page, after you’ve selected a target range under Define Your Target Range, click Save as New Audience.

2. In Save as New Audience, choose how you want the audience to be composed:

   1. Predictive Audience: Only include predicted users.
   2. Expanded Audience: Include both the predicted and seed users.
3. Click Create Audience. The audience builder opens with your selection prepopulated, using the prediction’s percentile range as part of the audience membership criteria.
4. Save and activate the audience to use it in downstream campaigns the same way you activate other audiences.

As the prediction refreshes on its schedule, audiences built from similar customer predictions stay aligned with the latest results.

Learn more about creating audiences from similar customer predictions in Build an Audience with Similar Customer Predictions.

Likelihood percentiles chart

The Likelihood Percentiles chart shows how your scored users are distributed from least similar to most similar to your reference segment, and how similarity changes across that spectrum.

Users are grouped into 10 equally sized percentile ranges (deciles) along the x-axis. The predicted similarity line shows how similar each decile is to the reference segment relative to the average similarity across all scored users.

You can use this chart to understand how similarity increases as you narrow to higher percentiles, then use that insight to choose a range that balances reach and similarity.

The chart includes several visual elements:

• Predicted Similarity line: Shows predicted similarity for each percentile group relative to average.
• Selected range: Highlights the percentile range you have selected in the Target Range controls, so you can see where your chosen audience sits on the similarity spectrum.
• Average similarity line: Represents the average similarity across all scored users. You can compare your selected range to this baseline to understand whether your audience is more or less similar than average.
• Similarity bands: The background shading groups users into dissimilar customers, moderately similar customers, and most similar customers based on similarity thresholds.

Model quality details

If your prediction’s strength is Weak, or if its status becomes Inactive or Failed, check the following:

• Confirm that your eligible users and reference segment contains at least 500 users and is based on meaningful signals. Segments smaller than this cannot be used to train the prediction, and overly specific seed definitions can reduce model reliability.
• Check that the prediction has been refreshed in the past 30 days. Predictions that have not run recently can become inactive.
• Verify that data ingestion is active and that relevant user activity is still flowing into mParticle. The model relies on behavioral patterns to identify similarity.
• Review your seed definition and any criteria that may be too restrictive. Narrow criteria can limit the training base and reduce the model’s ability to distinguish similar customers.

Adjusting your seed definition and ensuring consistent data flow can help improve prediction strength and maintain reliable results over time.

What happens when you edit a reference audience

If you used an existing audience as your reference segment, mParticle keeps the prediction in sync with that audience. When you save changes to the reference audience, a confirmation modal appears:

• The modal warns “This audience is linked to a prediction.” Clicking Proceed anyway automatically triggers a prediction recalculation using the updated definition. Clicking Cancel discards your changes.
• If your changes introduced incompatible criteria: the modal displays an additional warning in red: “This edit introduces incompatible criteria and will cause the prediction to fail and stop updating.” Clicking Proceed anyway saves your changes and sets the prediction’s status to Failed. Clicking Cancel discards your changes.

Incompatible criteria include:

• Event-based attributes (non-user attributes)
• Computed or derived user attributes, such as predictive attributes, group attributes, or segment attributes
• Changes that cause the classification pipeline to no longer receive negative labels

If the reference audience is a child audience in an audience group, edits to a parent audience in the group trigger the same confirmation flow.

Deleting a similar customer prediction

When you delete a similar customer prediction, mParticle checks whether the attribute is used in the definition of any audience. If it is, deletion is blocked until you remove the prediction from those audiences.

If the prediction was created from an existing audience, deleting the predictive attribute removes the connection between the prediction and the audience, but does not delete the audience itself.

The deletion check only validates whether the predictive attribute is used in audience definitions. Any internal system records associated with this prediction are cleaned up automatically and do not require any action from you.

The reverse also applies: if you try to delete an audience that is being used as the reference segment for a prediction, deletion is blocked until you delete the associated prediction first.