A Multi-population Mortality Model Based On Fuzzy Clustering And Bi-GRU

At the level of life table theory and risk management,population mortality affects life table parameters and estimates of life expectancy.Errors in estimation can impact the stability of pension systems and insurance companies.Early researchers used classical models to study individual population mortality,which generally underestimated mortality,had adverse effects on the measurement and management of longevity risk in later years,leading to financial deficits or reserve shortfalls.As population mortality declines and population longevity rises worldwide,models that fit individual populations are poorly and incomplete to other populations,there is an urgent need for multi-population models that can fit most populations.There is some commonality in population mortality across populations,especially in populations with similar economic levels or geographic regions.In this thesis,we focus on multi-population common factor mortality models,combining machine learning methods such as data characterization and clustering to improve the prediction accuracy by neural networks of the original model.Firstly,this thesis applies a common age effect model to male and female mortality in China for the first time.A common age effect model is a multi-population mortality model that extracts the common age effect shared by all data subjects in multiple population data.The effect of the model is tested based on the overall mortality data of five European countries,and then the common age effect model is applied to Chinese male and female mortality data,the standard residual plots and error indicators are compared with the Lee-Carter model.Secondly,four new models are constructed based on the common factor mortalit y model by combining machine learning methods.Firstly,the age and period effects of 14 countries and regions are extracted from the ILC model,clustered and visualized separately.The common age effect and Jiont-k model are combined with fuzzy k-Means clustering and cohort effect respectively to construct the new models.The numerical results from the new model dealing with the data of countries and regions with high similarity show that the combination with fuzzy clustering shows good results.The model with the addition of cohort effects is less effective due to model and birth year limitations.This is a useful attempt to study multi-population common factor mortality by using machine learning.Finally,a bidirectional gated recurrent neural network(Bi-GRU)is constructed based on mortality data to make prediction on overall、male and female mortality data in Japan.The Bi-GRU for mortality is an effective attempt to extend our mortality prediction method with neural networks by comparing the error metrics with the classical mortality model.