Using Mixture Density Networks in Modeling Pure Premium

Contributed Talk | Day 2 | 3:25 pm | 40 Minute Duration | Grand Gallery D
  • Drew Betz
    Farmers Insurance Group Assistant Vice President/Head of Statistical Analytics - Personal Lines

Using Mixture Density Networks in Modeling Pure Premium

Contributed Talk | Day 2 | 3:25 pm | 40 Minute Duration | Grand Gallery D

Using Mixture Density Networks in Modeling Pure Premium. Insurance rating models are subject to regulatory and business scrutiny during which model coefficients are manually reviewed and approved. As a result, model improvement hinges on using new predictors, new interactivity, or changing the functional form of the model. Typical model forms used in insurance center on the Poisson, Gamma, and Tweedie distributions. Advances in computational power now enable far more powerful approaches to modeling insurance data, even when faced with extremely large datasets. This paper discusses using Mixture Density Networks (Bishop, 1994) to model the PDF of
loss dollars, and demonstrates that Double GLMs are a special case of Mixture Density Networks. The approach effectively bypasses the need to assume a given distribution when estimating a regression model. A hybridization of Shaul’s (2011) and Zeiler’s (2012) stochastic gradient descent algorithms is presented as a means to estimate these models at scale.