The death rate of the world’s population has been continuously reduced, and the life expectancy has become unpredictable. Therefore, it is of great significance to improve the mortality prediction method to accurately predict the future population changes. The traditional Lee-Carter model describes the change of mortality through three parameters: the average mortality rate of the age group, the time term and the sensitivity of the age factor to change with time. The time term in the model is predicted by ARIMA method. However, the method can not solve the problem of long memory of mortality data, and the existing research rarely combines traditional demographic methods with machine learning methods in the context of today’ s big data. Therefore, this paper introduces LSTM ( long-term and short-term memory deep learning neural network) and fractional Brownian motion driven O-U process to improve the prediction of mortality. Due to the small sample size and incomplete data on mortality mainland China, this paper selects the mortality data of male age groups in Hong Kong, China, and used time series ARIMA method, the ARIMA-LSTM method combined with time series and machine learning, and the fractional O-U process to fit and predict the time items in the model, and compares the short-term prediction effects of the three methods through residual diagram and three evaluation index values. The results show that ARIMA-LSTM method has the best short-term prediction effect, which proves the feasibility of introducing machine learning method to improve the mortality prediction method, provides a new idea for the government to predict future mortality, and also provides a certain basis for relevant institutions to study the risk of longevity.
