An interpretable neural network approach to cause-of-death mortality forecasting

Cause-of-death mortality forecasting, a key topic in public health and actuarial science, is a challenging task due to the difficulty of modeling that accounts for dependencies among causes of death. While several cause-of-death mortality models have been proposed to address this difficulty, little attention has been paid to improving their predictive performance. Recently, purely data-driven approaches using tensor decomposition methods have been introduced to cause-of-death mortality modeling, demonstrating strong out-of-sample predictive performance compared to existing models. However, these methods have difficulties in the interpretability of multi-rank tensor components to achieve strong predictive performance. In response, we propose a novel tensor-based cause-of-death mortality model by replacing the tensor decomposition with a convolutional autoencoder with a one-dimensional latent layer that provides a Lee-Carter-like time-series factor; the model also provides the age sensitivity of cause-specific log mortality to the time-series factor. Due to the representational capability of the neural network, our model achieves better predictive performance compared to the existing tensor decomposition-based models, despite the simplified latent layer for model interpretability.