Predictors of adaptive models

Hi! I'm Kevin, a Data Scientist new to Pega and assigned to a Customer Decision Hub implementation project in financial services.

I will ask Bella, the Lead Decisioning Architect on the project, for recommendations on predictors that adaptive models use to calculate the likelihood of a positive outcome.

Hi Kevin! Nice to see you again!

So, Bella, could you guide me on how to select the predictors for the adaptive models?

A good approach is to start with data that is already in the application and then develop a roadmap to gradually incorporate and derive more data over time.

Keep in mind that when we talk about selecting predictors, we're actually talking about selecting predictor candidates.

You can have a wide range of candidate predictors at your disposal, sometimes even several hundred or more.

During the production phase of the project, Adaptive Decision Manager automatically identifies the best subset of these candidates to become active in the models.

It also groups correlated predictors together and selects the one with the strongest relationship to the outcome.

So I can throw in any data that I can get my hands on?

No, it is crucial to comply with all internal policy, legal, and regulatory requirements.

Also, you should avoid using customer identifiers as predictors for adaptive models, as they typically have a unique value for each customer and are not predictive.

Got it! So, where can I see what data sources are available in the application?

In the Customer Decision Hub portal, you can find a list of all the available data sources on the Profile Data Sources landing page.

Clear! So, do you have any general recommendations on the data that I should use?

Yes, I do. For the best results, it's advisable to use predictors that provide data from various data sources.

Start by incorporating customer profile information such as age, gender, and customer lifetime value.

Additionally, consider adding relevant fields like scores from externally generated predictive models as predictor candidates, and account attributes if you have account-level actions.

That is all static data. I see that event data sources cannot be associated with the customer context.

Correct. You need to summarize event data before you can use these summaries as predictors.

Event data contains information on past customer behavior.

The strongest predictors to predict future behavior typically contain data about past behavior.

How about the Financial services clickstream summary?

The Financial services clickstream is a summary included in the Customer Profile Designer component and used for aggregating clickstreams on financial services-based web pages.

Recent web browsing information can be highly relevant and, therefore, very predictive.

I'm currently working on the Predict Web Propensity prediction, that calculates the likelihood a customer clicks on a web banner.

Can you show me how to add candidate predictors to enhance its predictive power?

Sure! The Predict Web Propensity prediction uses the Web Click Through Rate adaptive model configuration.

There are three types of predictors based on data sources: regular predictors, parameterized predictors, and predictors based on interaction history summaries.

Regular predictors are input fields that are available on the primary page where the Adaptive Model rule is defined.

Parameterized predictors are input fields that are not available on the primary page where the adaptive model rule is defined, but which are on the Strategy Results page.

Predictors based on interaction history summaries are a predefined set of predictors that are based on interaction history summaries.

These predictors are enable by default.

Predictors can be one of two types: numeric or symbolic.

The system uses the property type as the default predictor type during the initial setup, but you can change the predictor type.

For example, when you know a numeric predictor has a small number of distinct values, change the predictor type from numeric to symbolic.

You can add a single or multiple candidate predictors.

Let's add all the Financial services clickstream to the model.

That is pretty straightforward!

So to summarize, I should use a wide range of data sources, that include the customer profile and customer behavior, as input for my models.

Yes, that sums it up!

Thanks, Bella! This has been very helpful.

My pleasure.