An alternative to projecting all-cause mortality directly is to project mortality rates for separate causes of death and then aggregate them to derive all-cause mortality rates. This approach has been advocated as a means of gaining forecasting accuracy in all-cause mortality projections and obtaining further insight into the drivers of mortality change. However, such approach is rarely used due to a number of theorical and practical challenges including, among others, the difficulty in modelling the dependence structure of the causes and the fact that aggregate mortality models are not necessarily appropriate for cause-specific rates. Moreover, while it is argued that cause-specific projections should result in more accurate projections of all-cause mortality, past studies tend to indicate that disaggregated forecasts do not result in significant gains in accuracy and tend to be more pessimistic than those without disaggregation. In this presentation, we will discuss two recent developments in the modelling of cause-of-death data which aim to address some of these limitations.

First, we will present a new approach for decomposing mortality improvement rate assumptions and projections into the additive contribution of different causes of death. Then, we will explore if combining this decomposition with forecast reconciliation methods has the potential to improve the accuracy of all-cause mortality projections.

Second, we will present neural network models to project all-cause mortality based on cause-of-death data. Recent studies have shown that neural networks can outperform traditional stochastic mortality models in both fitting and forecasting performance. Moreover, neural networks are ideally suited to model the complex dependence structure between causes of death without the need to specify the links among causes in advance.

We expect that the methods introduced in this presentation will allow practicing and academic actuaries to develop more accurate predictions for all-cause mortality and to better communicate mortality improvement assumptions to users and decision makers.

Find the Q&A here: Q&A on 'Mortality Modelling and Long Term Care'