Predicting customer behavior using predictive models

Predictive analytics uses past data to find patterns and uses those patterns to predict what will likely happen in the future.

There are two approaches to predictive analytics, and we use them both at Pega.

In one approach, called adaptive analytics, models are created in real-time without human intervention. Pega primarily uses adaptive models to capture customer responses in near or real-time. These adaptive models are automatically updated after new responses have been received and can start making predictions without any historical information because they learn on the fly.

Adaptive analytics automates everything that can be automated in the predictive model development and execution process. No human intervention is required in the generation of adaptive models. Pega’s adaptive modeling tool is called Adaptive Decision Manager. As a data scientist, all you have to do is configure a set of potential predictors the models can use. Adaptive Decision Manager then uses this definition to create what we call a ‘container’ in which customer and behavior data, or evidence, is captured in real-time. The software analyzes this evidence, and at frequent intervals generates new predictive models based on it. Because this is an automated process, model generation can scale rapidly. With real-time predictive modeling, you can quickly have 500 adaptive models running in the background learning from every customer interaction and generating new models on a daily basis. Because it requires no existing behavior data to get started, adaptive analytics is most often used to predict how new and unique actions will perform.

The other approach creates what are known as predictive models. Predictive models are created offline using historical data by people working with a predictive modeling tool. Pega provides the Prediction Studio portal to do this type of modeling. But there are many other vendors who provide predictive modeling tools, such as “SAS Predictive Analytics” or the free “R Statistical Software.” Any model developed using the PMML standard can be executed by Pega Decision Management.

It normally takes a person much longer to create a new model using offline predictive modeling than it takes for a new model to be created by other adaptive models.

A data scientist creates adaptive and predictive analytics models in Pega Prediction Studio.

There are three options for creating predictive models:

1. Using Pega Machine Learning. You can build a new predictive model using the proprietary Pega machine learning wizard. Import a file containing historical data and build the model in Prediction Studio.
2. Importing models. You can import PMML models that were built in third-party tools like R or Python. Similarly, you can import model files that have been generated in H2O.ai. H2O.ai is a modelling platform, and the procedure for using the model is similar to PMML.
3. Referencing external models. In the Pega platform, you can reference a model on an external platform like GoogleML or Amazon SageMaker. In this case, the model itself is executed on the third-party platform, and the outcome is sent back to Pega.

When you create a predictive model, the input fields that you select as predictor data play a crucial role in the predictive performance of that model.

To achieve the best results, use predictors that provide data from many different sources, including:

Customer profile. This includes information such as age, income, gender, and current product subscriptions. This is usually part of the Customer Analytic Record (CAR) and is refreshed regularly.

Interaction context. This includes information such as recent web browsing records, support call reasons, or input that is gathered during a conversation with the customer. This data can be highly relevant and, therefore, very predictive.

Customer behavior. This includes data such as product usage or transaction history. The strongest predictors of future behavior typically contain data about past behavior.

Scores. These could be credit scores or other results from the off-line execution of external models.

To use a predictive model in a decision strategy, use a predictive model component.

The decision strategy provides a customer's information as input to the predictive model. The output of the model is available to the other components of the strategy.

For example, you can use a predictive model component in a decision strategy to predict a customer’s churn risk. If the churn risk is high, the strategy selects a retention offer for the customer.