Abstract

I will discuss several interrelated projects on the use of Gaussian Process (GP) models for longevity analysis. The underlying Age-Period-Cohort structure is well-suited for capturing by a GP in order to address the common actuarial tasks of nowcasting the latest mortality rates and probabilistically projecting them into the future. I will review the GP spatial covariance framework in the context of mortality surfaces and the key steps of kernel and prior mean selection, improvement factor computation, and posterior sampling. Among the various GP implementations we have developed, I will highlight: (i)multi-output GPs for joint analysis of several dozen populations, hierarchically arranged along nationalities, genders and causes-of-death; (ii) compositional GP kernel search to identify the fittest kernels matching the spatio-temporal mortality dynamics in different countries; (iii) deflator GP models to capture the relative mortality of a small pension fund population vis-a-vis a national mortality table. Plentiful illustrations using Human Mortality Database datasets and corresponding insights into evolving mortality patterns will be given. Co-authors include Nhan Huynh, Jimmy Risk, Rodrigo Targino and Eduardo de Melo.