Abstract:

Classical actuarial modeling is largely based on generalized linear models (GLMs). We revisit the theory of GLMs and, in particular, we discuss the exponential dispersion family of distributions. This family of distributions has many good properties within a GLM approach, for instance, it fulfills the important balance property.

Based on the exponential dispersion family we illustrate how classical GLMs can be integrated into neural network architectures. This integration approach is universal in the sense that it can be applied to any parametric regression model that can be brought into neural network form. We call the resulting blended model a combined actuarial neural network (CANN) model. If this CANN model is calibrated appropriately, it will lead to an enhancement of the classical actuarial regression model with neural network features.

Topics considered:

- exponential dispersion family

- generalized linear models

- feedforward neural network

- combined actuarial and neural network models

Bio: Mario Wüthrich is Professor in the Department of Mathematics at ETH Zurich, Honorary Visiting Professor at City, University of London (2011-2019) and Honorary Professor at University College London (2013-2019). He holds a Ph.D. in Mathematics from ETH Zurich (1999). From 2000 to 2005, he held an actuarial position at Winterthur Insurance, Switzerland. He is fully qualified actuary of the Swiss Association of Actuaries, served on the board of the Swiss Association of Actuaries (2006-2018), and is Editor-in-Chief of ASTIN Bulletin.