# Create a Predictive Model

Create a Predictive Model to improve your segment targeting.

Safari does not support some of the functions of predictive models.

# Define your Predictive Model

1. Open TD Console.
2. Navigate to **Audience Studio**.
3. Select a parent segment. For example:
![](/assets/image2021-6-14_13-25-22.81ad73efcae1deb4249e036183812042f3d8136a570121cee40cc4e156f807cd.cf3845cb.png)
4. Continue with the displayed parent segment or select a different one.
5. Select **Create New**.
6. Select **Predictive Model**.
![](/assets/image2021-6-14_10-49-52.efa1243525b0ec462eacd041e65c5b06e8b3f53554f7e8b3ca07a0574d5dc3d5.cf3845cb.png)
7. Name your model.
8. Select **Next**.
![](/assets/image2021-6-14_13-58-47.7e0973eaa78d48efbe1a9ca628534e78a03570f4520384b23624945cc5caac5c.cf3845cb.png)
9. Select your Training population.
This is your dataset of examples for use during the training process for Treasure Data's machine learning. The dataset is used to help produce sophisticated results. The dataset can be a segment or all profiles within a parent segment.
Funnel Stages are not supported as a Training Population, Scoring Target, or Positive Samples segment.


![](/assets/image2021-6-14_14-0-31.0715052573e26606aa58ae2ec8016dac90eb8b1cd0cb70d6fa6a09f8d02db5ef.cf3845cb.png)

1. Select your scoring target.
These profiles are used to help score the accuracy of the prediction. The scoring target is the group of profiles whose behavior you are trying to predict. It could be a segment or All Profiles within the Parent Segment.
2. Select your positive sample of data that fits the characteristics of what you are trying to predict.
Your choice of datasets for the Training population and Scoring target fields are excluded from the list of available values for the Positive sample field.
3. Select **Next**.
![](/assets/image2021-6-14_16-16-12.d91780364d13ba5371406afa352c3d00ef0ee62eddc4bdce11b5ed56273103d3.cf3845cb.png)
4. Optionally, select Use suggested features.
Predictive Scoring provides Feature Guess functionality. Treasure Data suggests highly relevant features (predictors) derived from attributes.


Feature Guess functionality does not support Hierarchical Groupings. Feature Guess will not suggest any Group Attributes.

![](/assets/image2021-6-14_16-26-0.5d064ce9f2e31af7e2db67f783198573d1be4c2ad4bdbf995f8ac847083a8c06.cf3845cb.png)

1. Optionally, instead of allowing predictions, you can define the following:


**Categorical features**
Use this field for attributes or group attributes whose values belong to discrete categories. Example: `_state_` with values such as CA, NY, or MA. Categorical attributes can be non-ordered (nominal) like state or gender, or ordered (ordinal) such as high, medium, or low temperatures. For instance, the feature “color” might take the values “purple,” “yellow,” and “blue.”

**Categorical array features**
Use this when the attribute stores an array of category values, such as `interest_words`.

**Quantitative features**
Use this for attributes or group attributes that contain numeric values, for example `price` or `frequency`.

Each feature field provides a drop-down list of attributes. Regular attributes appear at the top, while group attributes are listed toward the bottom with the `_ [Group]_` prefix.

![](/assets/predmodelgroupattrib.d0383233911c175292da04debf1c64d68555eeca1d3698a7eee5e276b32b4ed2.cf3845cb.png)

1. Select Save and train or **Save**.


![](/assets/image2021-6-15_10-33-19.bfcf09166c94f6a86b584155a17c5c13756fbfeaabde98d86be774b74cf56c09.cf3845cb.png)

You receive an error message if your training population, scoring target, or positive sample references a segment in a folder in which you don't have Full or View permissions.