I’m excited to share our latest research paper on using the LocalGLMnet, an explainable deep learning model, to forecast mortality rates for multiple populations. This paper is joint work with Francesca Perla, Salvatore Scognamiglio and Mario Wüthrich.
Mortality forecasting is crucial for actuarial applications in life insurance and pensions, as well as for demographic planning. However, most existing machine learning models for this task are not transparent. We wanted to bridge this gap by adapting the LocalGLMnet, which preserves the interpretable structure of generalized linear models while allowing for variable selection and interaction identification.
We applied our model to data from the Human Mortality Database (HMD) and the United States Mortality Database, and found that it produced highly accurate forecasts that can be explained by autoregressive time-series models of mortality rates. We also showed how regularizing and denoising our model can improve its performance even further.
The image below shows a comparison of forecasting results between different models for the HMD.
The full paper can be found here: