Population Mortality Study Based On Recurrent Neural Networks

With the improvement of the economy and medical technology,the mortality rate of our country's population has been declining year by year,and the average life expectancy has been increasing,and the risk of longevity will follow.Longevity is a sign of social progress,but it also poses huge challenges to the government's public financial expenditures and life insurance companies.Longevity risks increase the pressure on the future solvency of the pension insurance system and impact the stability of life insurance companies.Therefore,building a suitable mortality model to accurately predict the future trend of population mortality has become the basis for solving longevity risks.At present,there are relatively few statistics on population mortality in my country,and the quality of the data is still relatively poor.However,the research on population mortality in my country is mainly based on the traditional Lee-Carter model.In recent years,thanks to the improvement of computing power,deep learning technology has re-emerged on the stage.Deep learning has the characteristics of accurately learning data laws from samples.It is very useful in detecting unknown and unrecognizable patterns.Among them,cyclic neural network structure The main application areas of(RNN)are time series analysis and natural language processing.This article uses the two most popular structures of RNN neural network(LSTM and GRU)to explore its application in population mortality prediction.Considering that the sample size and data quality may have an impact on the prediction effect of the model,this article uses data from China and the United States to construct the model separately.Among them,China and the United States use age-specific mortality data from 1994-2018 and 1950-2018 respectively..The model is trained by using China's data from 1994-2014 as the training set to predict the mortality data from 2015-2018,and the mean square error is calculated based on the actual observations from 2015-2018 as the criterion for model evaluation.Similarly,the United States uses 1950-2010 as the training set and 2011-2018 as the test set.This paper constructs 4 cyclic neural network models,including gender-specific LSTM and GRU models and dual-sex joint modeling LSTM and GRU models,and uses the Lee-Cater model as the benchmark for comparison of cyclic neural network models.The forecast results are analyzed.The results show that whether it is on the Chinese data with a small sample size and poor quality or on the US data with a large sample size and good quality,the prediction results of the four recurrent neural network models constructed out of the sample are better than those of Lee-Carter model,and the recurrent neural network model has a better effect in capturing the mortality rate of a specific age group over time.Taken together,the best performance among the 4 cyclic neural network models is the LSTM model of dual-sex joint modeling,and on 50 different random seeds,it has better stability than the GRU model.Finally,the LSTM model of dual-sex joint modeling was used to predict the mortality rates of China and the United States by age and gender in the next 10 years.It was found that the mortality rate of both sexes in China will be significantly improved in the future,including females.The rate of improvement is greater than that of men,while the gender mortality rate in the United States has improved only in the youth population,and the overall rate of improvement is much smaller than that in China.