Guidewire’s Approach to Predictive Analytics, Part Three: Building Predictive Models

The Big Picture

Business intelligence looks at historical data to provide information about a company’s operations and performance. This can provide valuable insight but has limitations for predicting the future. Simply looking at historical patterns does not tell us how much of what we see in the data is a reliable pattern that can be depended on to continue – what we will refer to as “Signal” – and how much is due to random chance or unknown variables – what we will refer to as “Noise”. 

Predictive analytics brings mathematical techniques to the data to distinguish the Signal from the Noise, the systemic information from the random happenstance. Various techniques are available and different approaches lend themselves to different situations. The specialized knowledge of predictive modelers entails an understanding of the mathematics of each approach so that sensible choices can be made, and the results can be properly understood.

In this installment we consider what is required for the creation of predictive models.

Modeling Requirements for Predictive Analytics

The actual creation of predictive models is done by software and guided by the predictive modeler. The following considerations should be made for this software:

  • Ease of importing and transforming data
  • Learning curve and array of modeling techniques
  • Automation, model output, and transparency

While we discussed data in the previous installment of this series, there are additional aspects of data worth mentioning. First, any software product will have specific requirements for data to be ingested. These need to be kept track of for each software package used. Second, building a predictive model starts with exploratory data analysis to evaluate how the data should be treated for the given predictive model to be built. This may include data transformations, groupings, and others.

The more analytical techniques a software package supports, the greater the learning curve in using it. This is partly because of the greater array of techniques, but also due to the necessary interfaces. Open source approaches, which allow for nearly any conceivable mathematical technique, are functionally equivalent to programming languages, requiring significant expertise to use. A larger learning curve means a larger cost, either in the time delay for actual learning, or in paying for a predictive modeler who has already acquired the required skillset.

Automation is key for efficiently building predictive models. At the same time, automation should not be used to replace human evaluation. The proper use of predictive modelers’ expertise is in evaluating output and interpreting results, rather than in running multiple iterations. In addition, the array of output provided should not only meet the needs of the modeler, but also support the needs of subsequent audiences – supervisors, executives, agents, and insurance regulators. Many of these audiences who will evaluate proposed models will need a level of transparency that must be met.

Guidewire’s Approach to Building Predictive Models

Guidewire’s BUILD application allows for predictive modelers to import, evaluate, and transform data, and provides an array of modeling techniques that are standard methodologies used in insurance. This application provides a user interface making advanced analytics accessible to a wide array of predictive modelers, as well as facilitating the sharing of information with a team and keeping records for model management.

However, the real advantage of the BUILD application is that it is part of the platform approach to analytics. Easy importation of Guidewire data and automatic passing of model information to the next steps that deploy and embed models into business processes can transform the possibilities for efficient and flexible models.

Predictive models built in Guidewire’s BUILD application will integrate most seamlessly with the entire approach, but Guidewire’s platform also allows for models built elsewhere to be imported so that they can take advantage of the deployment capabilities. This will be discussed more in part four of this series on deploying and embedding predictive models.  Also, don’t miss the previous installment discussing the data needs of predictive analytics.

Get started