A few best practices can make developing predictive models achievable for any carrier, regardless of the level of predictive modeling expertise. As I explained in my previous post, following a structured development process that aligns with the four-stage product development lifecycle is the key to successfully applying predictive modeling within your underwriting workflow.
Now I’d like to share key insights about how to bring a predictive model to life through ideation, design and development, and implementation ― the first three stages of the product development lifecycle.
Step 1: Ideation
The basis for any predictive model is a problem that needs solving and determining whether or not a predictive model is the right tool for doing so. However, there are two prerequisites for this process:
- Strong executive sponsorship for the effort (to ensure all the right resources will be applied)
- A committed cross-functional team that can help bring the idea to reality
In the ideation phase, this team will:
- Showcase the benefits of predictive modeling and establish buy-in across the organization
- Identify, validate, select and prioritize problems to solve through predictive modeling
- Determine costs and justify the Return on Investment (ROI) for any predictive models to be built
- Explore how the predictive model will integrate into the operational workflow, including benchmarks and success measures
Step 2: Design and development
While predictive models can be used for risk selection, pricing, claims fraud detection, claims subrogation potential, and so on, within small commercial there’s a growing movement to use predictive modeling for risk assessment and pricing by building insurance scores that rank order risks in terms of loss propensity. Designing and developing this type of model is a very iterative process, which begins with data exploration, followed by creating and validating the model, and finally, ensuring it complies with any applicable regulatory requirements.
Data exploration is a joint effort bringing together business analysts, statistical modelers, IT resources and regulatory experts. There are a number of third-party data sources that can be tapped for building a predictive model for risk assessment, including commercial credit, consumer credit, and public records. For risk assessment, a large amount of premium and loss data is required to “train” the model to predict the target as well as to test and validate that it works. Data is partitioned as either training data or testing data for the project. Of course, all data sources and attributes used in the model to predict the target must comply with any applicable regulatory requirements.
Step 3: Implementation
Once the model has been designed and proven, it’s ready to be implemented within the workflow. Because implementation impacts so many parts of the operation, there are a number of key considerations, including:
- Identifying and documenting impacts to existing business rules and procedures, such as rating and underwriting
- Determining IT system requirements for building the model, application workflow changes, and score storage and tracking
- Ensuring support for any applicable customer dispute processes
- Training all stakeholders and impacted parties
- Creating a rollout plan
At this juncture, the hard work is done. The last step is ensuring your model works for you the way you want it to. In my next and final post, I’ll share how to ensure you get the best results from your predictive modeling efforts through the best practice of monitoring.
For additional information about predictive modeling best practices, please see our whitepaper, Making Predictive Modeling Work for Small Commercial Insurance Risk Assessment.