• Home
  • Resources
  • Blog
  • Guidewire’s Approach to Predictive Analytics, Part Four: Implementation

Guidewire’s Approach to Predictive Analytics, Part Four: Implementation

  • Chris Cooksey

October 15, 2020

Untitled

The Big Picture

Collecting and analyzing data does not provide business value. Improved business processes that are more accurate and efficient through the use of analytics provide business value. Without embedding a predictive model into a relevant business process, all the work and cost put into predictive analytics will be for nothing. Yet the implementation of predictive models has been an afterthought for many insurers.

Guidewire believes that the process for implementing predictive models should be considered from the beginning and tied to the rest of the process to promote efficiency and transparency. We will consider “implementation” of a predictive model to have two parts – the “deployment” of the model and the “embedding” of that model into core systems.

In this installment we consider the unique needs of implementing predictive models into business processes with predictive analytics.

Implementation Requirements for Predictive Analytics

Insurance companies have struggled with realizing the value of existing predictive models for two related reasons. First is underestimating what it means to deploy a predictive model. Here our language does us a disservice. Literally, a predictive model is simply a collection of rules which transform existing known information, such as choice or limit or cause of loss, into predictions of an unknown outcome. However, “implementing the predictive model” itself is not enough. Insurers must also consider business rules which guide and enhance the model’s behavior in various real-life situations for a true implementation. For example:

  • When should a predictive model return a prediction and when should it return an error?

  • How is exceptional business that was not considered in the model build to be identified and handled in production?

  • How are bounds on the minimum and maximum predictions addressed?

  • What if the output of two or more predictive models are required to guide the new business process?

Second is not successfully embedding the predictive information into the processes and systems that the insurer is already using. With legacy core systems, many insurers find that making modifications is a costly and lengthy project, and the output from predictive models has too often been provided on the side or in separate applications to the underwriters, adjusters, and other end-users who should be benefiting from the information. In these situations, the realized value from the predictive model is at best muted and in many cases, undermined entirely.

Guidewire’s Approach to Implementation

Deploying and embedding predictive models are separated in Guidewire’s approach to allow for different emphases. While a business owner or product manager will want to own and guide the entire process, deployment focuses on making sure that the entire solution provides correct and complete information in all circumstances, whereas embedding focuses on a seamless and useful experience for the end-user.

To facilitate each of these steps, Guidewire provides the following:

  • A DEPLOY application within Guidewire Predictive Analytics that allows users to create what is called a “Solution”. A single Solution may include one or more predictive models, look-up tables, free-form scripting, controls over allowed values, calls to third-party data, and other functionality required by many real business problems. Each Solution can be pushed to a cloud-hosted environment and called by Guidewire’s core systems.

  • Pre-built screens and workflows within InsuranceSuite and InsuranceNow to surface the information provided by Solutions for the most common applications of predictive analytics. This functionality includes a coherent framework for customizing exactly what is shown, and to whom, so that insurers can meet their particular business needs.

This approach to model implementation also allows for the importation of predictive models built in other applications. Models already created may still need an implementation path, and the variety of business problems requires the ability to be flexible in what modeling techniques are required.

Guidewire appreciates that predictive analytics only produces value through improved business processes, and if predictive models never get implemented, then the process has been a waste. What is needed is a streamlined implementation process that provides efficiency and flexibility. Stay tuned for installment five of this series on monitoring predictive models. Also, don’t miss the previous installment on building predictive models.