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

Predicting the evolution of mortality rates plays a central role for life insurance and pension funds. Various stochastic frameworks have been developed to model patterns by taking into account main stylized facts driving these patterns. However, relying on prediction one specific can be too restrictive lead some well-documented drawbacks, including misspecification, parameter uncertainty, overfitting. To address issues we first consider modeling in Bayesian negative-binomial framework overdispersion uncertainty about estimates natural coherent way. Model averaging techniques are then considered as response misspecifications. In this paper, propose two methods based leave-future-out validation compare them standard (BMA) marginal likelihood. An intensive numerical study is carried out over large range simulation setups performances proposed methodologies. illustration real-life datasets, along with sensitivity analysis Covid-type scenario. Overall, found that both an out-of-sample criterion outperform BMA approach terms performance robustness.