Hybrid Lee-Carter Model: An Analysis of Mortality Rate Prediction Using ARIMA and Deep Learning Techniques

Abstract

e Lee-Carter model, is a fundamental approach in mortality forecasting that estimates and projects mortality rates based on historical data. While deep learning techniques have been extensively explored in mortality prediction, most of the studies have focused on Long Short-term Memory within the family of recurrent neural networks. This thesis investigates the application of the four different Deep Learning models, namely Long Short-Term Memory, Gated Recurrent Unit, Bidirectional Long Short-Term Memory and Bidirectional Gated Recurrent Unit in mortality rate prediction, both independently and in hybrid combinations with Lee-carter model. The proposed models are evaluated using yearly mortality data categorized by genders from nine census divisions in the United States. Through the out-of-sample testing, our finding indicates that all the model outperform the traditional Lee-Carter model in terms of two selected error metrics, mean absolute error (MAE) and root mean square error (RMSE). The Gated Recurrent Unit hybrid model produces the most accurate estimations among the hybrid models, while the Bidirectional Gated Recurrent Unit model has the best overall performance.