Mortality rate forecasting: can recurrent neural networks beat the Lee-Carter model?

Human mortality rates form a particularly challenging task for time series forecasting. Forecasts should be produced separately for different ages, preferably for multiple years ahead into the future, having just a relatively small amount of historical data available. It is pretty difficult to create and evaluate forecasts under such circumstances. In this paper, we apply a recurrent neural network to mortality rate forecasting. RNNs are usually used in data rich environments.