M-3 - Building Machine Learning Algorithms for Insurance

Much of the development of machine learning algorithms has been focused on improvements to predictive power. However, insurance practitioners know they need to balance other priorities, such as transparency and stability. This session will discuss alterations to traditional machine learning algorithms that more appropriately meet the needs of those in the insurance space. For example, GBMs suffer from both poor transparency – it is impossible to discern which factors are main effects and what interactions may exist – and instability – small changes in the predictors can lead to surprisingly significant changes in model predictions. We will demonstrate how adjustments can be made to the GBM algorithm to address both of these issues and others. In addition to discussing existing approaches, the session will challenge attendees to think about future innovations.