2023 Australasian Actuarial Education and Research Symposium

Jianjie Shi

Monash University

Beyond linearity: a Bayesian nonparametric VAR approach to mortality modelling

This is joint work with Yunyun Wang

This study introduces a novel nonparametric approach to mortality modelling, addressing the limitations of traditional Vector Autoregressive (VAR) models. Although VAR models are prevalent, their inherent assumption of linearity between endogenous variables and their lags can sometimes be restrictive. This limitation becomes particularly evident when dealing with extreme observations, such as those arising from the COVID-19 pandemic. To address these challenges, we introduce the Bayesian Additive Vector Autoregressive Tree (BAVART) model into mortality modelling. This innovative model, a fusion of VAR and Bayesian additive regression tree (BART) techniques, provides a flexible framework capable of capturing intricate nonlinear relationships. Furthermore, we incorporate a sparsity-inducing Dirichlet hyperprior on the tree's splitting proportions, enhancing adaptability to sparsity in the BAVART model. Our empirical analysis, drawing from mortality datasets across multiple countries, highlights the BAVART model's proficiency in accurately capturing these dynamics. Owing to its flexibility and precision, the BAVART model stands as a promising tool for future mortality modelling research and related fields.