Mean reversion in stochastic mortality : why and how?

Life insurance companies use stochastic models to forecast mortality. We know from literature that non mean reversion models seem to be more suitable for mortality in comparison with mean reversion models having a fixed long term target. In this paper, we adopt stochastic a ne processes for the force of mortality, and study the impact of adding a time dependant long term mean reversion level to two non-mean reverting processes. We calibrate the models to different generations in the Belgian population, and assess their ability to predict mortality by using different statistical methodologies. The backtest shows that the survival curves provided by the mean reverting processes are closer to reality. We conclude that incorporating a time-dependant target to the considered models improves significantly their performance.