Joint research initiative (2020-23): Longevity model selection and change detection in the covid-19 era

Bayesian model averaging for mortality forecasting using leave-future-out validation

Relying on one specific model can be too restrictive and lead to some well documented drawbacks including model misspecification, parameter uncertainty and overfitting. In this paper, we consider a Bayesian Negative-Binomial model to account for parameter uncertainty and overdispersion. Moreover, model averaging based on out-of-sample is considered as a response to model misspecifications and overfitting. Overall, we found that our approach outperforms the standard Bayesian model averaging in terms of prediction performance and robustness.

Parsimonious Predictive Mortality Modeling by Regularization and Cross-Validation with and without Covid-Type Effect

Most mortality models are very sensitive to the sample size or perturbations in the data. In this paper, we show how regularization and cross-validation can be used to smooth and forecast the mortality surface. In particular, our approach outperforms the P-spline model in terms of prediction and is much more robust when including Covid-type effect.

StanMoMo: An R package for Bayesian Mortality Modelling with Stan

Glad to share a new R package for Bayesian Mortality Modelling. The package provides functions to fit and forecast standard mortality models (Lee-Carter, APC, CBD, etc) in a full Bayesian setting. The package also includes functions for model selection and model averaging based on leave-future-out validation.