Data-adaptive Dimension Reduction for US Mortality Forecasting

Forecasting accuracy of mortality data is important for the management of pension funds and pricing of life insurance in actuarial science. Age-specific mortality forecasting in the US poses a challenging problem in high dimensional time series analysis. Prior attempts utilize traditional dimension reduction techniques to avoid the curse of dimensionality, and then mortality forecasting is achieved through features' forecasting. However, a method of reducing dimension pertinent to ideal forecasting is elusive. To address this, we propose a novel approach to pursue features that are not only capable of representing original data well but also capturing time-serial dependence as most as possible. The proposed method is adaptive for the US mortality data and enjoys good statistical performance. As a comparison, our method performs better than existing approaches, especially in regard to the Lee-Carter Model as a benchmark in mortality analysis. Based on forecasting results, we generate more accurate estimates of future life expectancies and prices of life annuities, which can have great financial impact on life insurers and social securities compared with using Lee-Carter Model. Furthermore, various simulations illustrate scenarios under which our method has advantages, as well as interpretation of the good performance on mortality data.