General Insurance Spring Webinar Series 2025: Why are ML Predictions so Leaky?

Wed 23 Apr 2025 -
12:30 - 13:30 (BST)

Despite showing significant promise, ML predictive models in personal lines pricing have proved to be less robust than traditional GLMs. The accuracy of predictions made quickly degrade, meaning the models leak value after being placed in market. This problem is much more prevalent to insurance use cases than other fields. In this talk we explore the causes of the phenomenon and demonstrate from real-world examples how changes to the underlying data can cause the effect. 
  
Topics we will cover include: 

• What is different about insurance data than successful use cases in other industries
• The dangers of using too much data
• How governance mitigated these challenges for GLMs, and why that it is harder for ML models
• How smoothing of predictions to avoid cliff-edges can help mitigate the challenges

The talk is designed for intermediate to advanced users. Specific details of the underlying mathematical constructs of the models are not required, but we do assume a general familiarity with the data, processes, and challenges of personal lines pricing. 

This will be delivered virtually and will cover technical findings. It should be of interest to actuaries at all levels with an interest in ML including senior actuaries and team leaders who are already using, or interested in using ML in the pricing process. 

Speakers

Martin Cairns and Ben Gaby, FTI Consulting