Transformer Self-Attention Network for Forecasting Mortality Rates

The transformer network is a deep learning architecture that uses self-attention mechanisms to capture the long-term dependencies of a sequential data. The Poisson-Lee-Carter model, introduced to predict mortality rate, includes the factors of age and the calendar year, which is a time-dependent component. In this paper, we use the transformer to predict the time-dependent component in the Poisson-Lee-Carter model. We use the real mortality data set of some countries to compare the mortality rate prediction performance of the transformer with that of the long short-term memory (LSTM) neural network, the classic ARIMA time series model and simple exponential smoothing method. The results show that the transformer dominates or is comparable to the LSTM, ARIMA and simple exponential smoothing method.