Forecast Mortality With Consideration Of Model Instability

The rapid economic development,gradual improvement of social environment,continuous enhancement of medical technology,and increasing awareness of health have resulted in an unprecedented improvement in life expectancy of populations worldwide,including China.According to the data from the 7th National Census,the number of people aged 60 and above will reach 264.02 million by 2020,accounting for 18.70% of the country’s total population,exceeding the internationally defined standard of 10% for aging.The increasing life expectancy,fueled by a declining mortality rate,has become a long-term influence on the severity of population aging,which has significant implications for government social security systems,insurance companies’ life insurance products,corporate pensions,and individual pension operations.Accurate and comprehensive population mortality projections are crucial to identify and measure longevity risk.Various stochastic frameworks have been developed to model mortality projections,but reliance on a particular model can be restrictive and lead to errors in model specification,parameter uncertainty,and overfitting.Furthermore,using a single model can be subject to significant uncertainty and may ignore useful aspects of other suboptimal models.To address these issues,this thesis proposes a Bayesian framework that embeds multiple mortality prediction models using finite mixture models.This approach jointly and coherently exploits different features of different model structures,leading to a more thorough description of potential mortality patterns.Three representative stochastic mortality models,the classical Lee-Carter model,the CBD model with curvature,and the RH model,are selected to construct Bayesian mixture models of Lee-Carter and CBD and Bayesian mixture models of Lee-Carter and RH.The residual distribution map and the deviation information criterion are used to evaluate model fit.Simulations based on male mortality values in Chinese mainland show that the fitting and prediction effects of the Bayesian mixture model are better than those of the individual models.In order to explore more realistic demographic characteristics of the mainland,this paper uses the male population mortality rates in the mainland and Taiwan as empirical data,first considers the correlation between the two data sets,incorporates the male population mortality data in Taiwan into an extended form of prediction model,and then applies error correction to the predicted values of male population mortality in the mainland with the help of co-integration theory,which helps to make more accurate predictions on the future mortality rates of men in the mainland.