Backtesting Recurrent Neural Networks with Gated Recurrent Unit: Probing with Chilean Mortality Data

Understanding the survival prospects of a given population is essential in multiple research and policy areas, including public and private health care and social care, demographic analysis, pension systems evaluation, the valuation of life insurance and retirement income contracts, and the pricing and risk management of novel longevity-linked capital market instruments. This paper conducts a backtesting analysis to assess the predictive performance of Recurrent Neural Networks (RNN) with Gated Recurrent Unit (GRU) architecture in modelling and multivariate time series forecasting of age-specific mortality rates on Chilean mortality data. We investigate the best specification for one, two, and three hidden layers GRU networks and compare the RNN’s forecasting accuracy with that produced by principal component methods, namely a Regularized Singular Value Decomposition (RSVD) model. The empirical results suggest that the forecasting accuracy of RNN models critically depends on hyperparameter calibration and that the two hidden layer RNN-GRU networks outperform the RSVD model. RNNs can generate mortality schedules that are biologically plausible and fit well the mortality schedules across age and time. However, further investigation is necessary to confirm the superiority of deep learning methods in forecasting human survival across different populations and periods.