A holistic approach to interpretable modelling and forecasting of human mortality by gender and country

Studies from many countries find that gender differences in mortality rates and life expectancy vary by country. The multipopulation Lee-Carter family of models, a widely-used methodology, decompose mortality rates into age, time, and country components, offering valuable insights into mortality trends. We delve into the interpretability of the Lee-Carter multipopulation model, elucidating its ability to capture underlying mortality patterns and project future trajectories. Moreover, we extend our analysis by incorporating machine learning techniques to model the residuals of the Lee-Carter framework. The main contribution of the paper is to introduce these techniques in the context of the multiple population mortality models. Specifically, we employ Random Forest to refine joint mortality forecasts by country, effectively capturing complex nonlinear relationships in residuals and improving predictive performance. In this paper, we revisit these models using new statistical techniques and data sets from the Human Mortality Database. By leveraging advanced computational algorithms, we aim to enhance the accuracy of mortality rate predictions and account for residual patterns that may not be captured by the traditional Lee-Carter approach alone. Through empirical validation and comparative analyses, we demonstrate the efficacy of integrating machine learning into multiple population mortality forecasting, thereby contributing to the refinement and improvement of mortality modeling methodologies.