Cause-of-death mortality forecasting using adaptive penalized tensor decompositions

Abstract

Cause-of-death mortality modeling and forecasting is an important topic in demography and actuarial science, as it can provide valuable insights into the risks and factors determining future mortality rates. In this paper, we propose a novel predictive approach for cause-of-death mortality forecasting, based on an adaptive penalized tensor decomposition (ADAPT). The new method jointly models the three dimensions (cause, age, and year) of the data, and uses adaptively weighted penalty matrices to overcome the computational burden of having to select a large number of tuning parameters when multiple factors are involved. ADAPT can be coupled with a variety of methods (e.g., linear extrapolation, and smoothing) for extrapolating the estimated year factors and hence for mortality forecasting. Based on an application to United States (US) male cause-of-death mortality data, we demonstrate that tensor decomposition methods such as ADAPT can offer strong out-of-sample predictive performance compared to several existing models, especially when it comes to mid- and long-term forecasting.