# Predictive segments

Predictions rank the users of your app every day by how likely they are to stop opening the app (churn) or to do a specific event, such as a purchase. Each prediction writes the score and category to two user tags, so you can build segments and journeys for the users most at risk or most likely to convert.

<Aside type="note">
Predictions are in beta and available on selected plans. To turn them on for your account, contact your account manager.
</Aside>

## Prediction types

When you create a prediction, choose what it predicts:

| Type | Who gets a high score |
| --- | --- |
| **Churn** | Users who will stop opening the app within the prediction window |
| **Event likelihood** | Users who will do the selected event within the prediction window |

The **prediction window** is the period the prediction looks ahead. For churn, choose 7, 14, 21, or 28 days, or **Auto**: Pushwoosh picks a window of 7 to 42 days, in whole weeks, from how long your users usually take to come back to the app. For an event, choose 7, 14, or 30 days.

## Prediction tags

Each prediction writes two tags to every scored user:

| Tag | Type | Values |
| --- | --- | --- |
| `PW Predict · <prediction name> Score` | Integer | 0 to 100, in steps of 5 |
| `PW Predict · <prediction name> Category` | String | `low`, `medium`, `high`, `very_high` |

The score is the user's rank among all scored users of the app, as a percentile. A score of 90 means the user is more likely to churn (or to do the event) than about 90% of the scored users. The score isn't the probability that the user churns. A user with a score of 90 can still be unlikely to churn if few users of your app churn at all. To see how many users in a score range actually churned or did the event, check the [prediction details page](#check-prediction-quality).

The category groups the same ranks into four bands:

| Category | Score |
| --- | --- |
| `low` | Below 40 |
| `medium` | 40 to 69 |
| `high` | 70 to 89 |
| `very_high` | 90 and above |

The score is a share of users, so the bands always split the scored users the same way: about 10% of them are `very_high`, about 20% are `high`, about 30% are `medium`, and about 40% are `low`. A user close to a band boundary keeps the previous score and category until they move clearly past it, so tags don't change back and forth every day.

Pushwoosh scores only users who opened the app within the period set in **Score users active in the last**. When a user drops out of that period, Pushwoosh removes both tags from the user.

## Create a prediction

1. Go to **Audience** → **Predictions**.

<img src="/segmentation-predictive-segments-1.webp" alt="Predictions page with two predictions in Ready status and the Create prediction button" width="876" />

2. Click **Create prediction**.
3. Enter a **Name**. The name becomes part of the tag names, for example `PW Predict · Churn in 2 weeks Score`.
4. In **What to predict**, choose **Churn** or **Event likelihood**.
5. For **Event likelihood**, select the **Event** to predict.
6. Choose the **Prediction window**.
7. In **Score users active in the last**, choose 30, 60, or 90 days.

<img src="/segmentation-predictive-segments-2.webp" alt="Create prediction dialog with a churn prediction named Churn in 2 weeks, Auto prediction window and 30 days of activity" width="624" />

8. Click **Create**.

The prediction runs once a night. The first scores and tags appear after the next nightly run. After that, Pushwoosh updates scores every day and retrains the model every week.

<Aside type="caution">
The model learns from your app's history of the predicted event: app opens for churn, the selected event for event likelihood. Track the event for at least four months before you create the prediction. A longer prediction window needs a longer history, and **Auto** can pick a window of up to 42 days. Until there's enough history, the prediction shows **Waiting for data** with the date the history must start from, and writes no tags.
</Aside>

## Prediction statuses

The **Status** column on the **Predictions** page shows whether the prediction writes tags:

| Status | What it means |
| --- | --- |
| **Training** | The model hasn't been built yet. It waits for the next nightly run. |
| **Waiting for data** | The app hasn't tracked the predicted event long enough. The reason under the status says from which date the history is needed. |
| **Not enough signal** | The model doesn't predict noticeably better than a simple rule based on when the user was last active, or too few users churned or did the event. The reason is shown under the status and on the prediction details page. |
| **Ready** | The model passed the quality check. Pushwoosh writes the score and category tags every day. |

Pushwoosh writes tags only while the prediction is **Ready**. A **Paused** badge next to the status means the prediction isn't scored until you resume it.

## Check prediction quality

To decide how many users to target, click the prediction name on the **Predictions** page to open the prediction details page. For a **Ready** prediction, it shows how well the model performed on a past test period:

<img src="/segmentation-predictive-segments-3.webp" alt="Quality cards of a churn prediction: model accuracy, churn rate, users scored, and last scored date" width="876" />

- **Model accuracy (ROC AUC)**: how well the model separates users who churned (or did the event) from those who didn't. A value close to 0.5 means the model is no better than a random guess, and 1 means a perfect ranking. The line under the value shows the accuracy of ranking users only by their last activity, so you can compare the two.
- **Churn within N days** or **Do event within N days**: the share of all scored users who churned or did the event. This is the result you get if you target users without the prediction.
- **Users scored**: how many users received the tags.
- **Last scored**: the date of the latest nightly run.

## Create a segment from a prediction

The **Build a segment** table on the prediction details page shows what you get if you target the top 5%, 10%, 20%, 30%, or 50% of users by score:

- **Score**: the threshold. The segment keeps users whose score tag is at or above it.
- **Users**: about how many users fall into the segment today.
- **Precision**: the share of users in the segment who churned or did the event in the test period. Compare it with the share of all scored users in the quality cards above.
- **Coverage**: the share of all users who churned or did the event that the segment includes.

A smaller segment is more precise but includes fewer of the users you want to reach. To create a segment:

1. Click **Create segment** in the row with the threshold you want.

<img src="/segmentation-predictive-segments-4.webp" alt="Build a segment table with Create segment highlighted in the Top 10% row" width="962" />

2. Keep the suggested name, for example `Churn in 2 weeks · score 90+`, or enter your own.
3. Click **Create**.

The segment is a regular segment with one condition on the score tag. Use it in campaigns and journeys like any other segment.

## Use prediction tags in segments and journeys

The prediction tags work like any other [user tags](/product/audience-data-and-segmentation/user-data-tags/tags/). To target users by prediction without the **Build a segment** table, [create a segment by tags](/product/audience-data-and-segmentation/segmentation/create-segments/by-tags/):

- **By score**: compare the `… Score` tag with a number, for example users with a score of 80 or higher.
- **By category**: select users whose `… Category` tag is `high` or `very_high`.

In Customer Journey, use these segments in an [Audience-based entry](/product/customer-journey/journey-elements/entry-elements/audience-based-entry/) to start a journey for at-risk users, or in a [Segment split](/product/customer-journey/journey-elements/flow-controls/segment-split/) to send at-risk users down a separate branch. Tags update every night, so a user can enter or leave the segment from one day to the next.

## Pause or delete a prediction

To manage a prediction, click the menu icon in its row on the **Predictions** page:

- **Pause** stops daily scoring. **Resume** starts it again.
- **Delete** removes the prediction. Pushwoosh removes its two tags from all users overnight.

<Aside type="caution">
You can't delete a prediction while its tags are used in a segment, a campaign, or a journey. Remove the tags from them first.
</Aside>

## Plan and limits

Predictions are available when your plan includes them. Without them, the **Predictions** page shows this banner and no **Create prediction** button:

<img src="/segmentation-predictive-segments-5.webp" alt="Banner on the Predictions page: Predictions are not available on your plan" width="880" />

An account can have up to 3 predictions across all its apps, unless your plan sets a different limit. When the account reaches the limit, the page shows how many predictions are in use and hides **Create prediction**. To create a new prediction, delete an existing one. To raise the limit, contact your account manager.