Neural Network-based Models And Methods For Multi-Population Mortality Forecasting

Mortality prediction is an important research in the fields of life insurance actuarial and demographics,and is a necessary research for life insurance rate determination,longevity risk management,and demographic analysis.In recent years,two innovations in mortality modeling have received much attention and are highly effective:on the one hand,innovation in the use of data,and on the other hand,innovation in the use of methods.Data use innovation:from single population modeling to multiple population modeling,related models do not study a certain population in isolation,but jointly model the mortality data of related populations.Method use innovation:from classic statistical modeling to modern algorithm modeling,in the context of rapid development of artificial intelligence technology and statistical models encountering big data processing bottlenecks,algorithm models are tried as a new technology for automatic extraction of high-dimensional nonlinear complex features,used for mortality modeling.The core work of this paper is:a cross-type study of these two innovations.This paper studies the multi-population mortality modeling based on three types of neural networks(general neural network,recurrent neural network and convolutional neural network),aiming to improve the effect of mortality forecasting,make up for the modeling disadvantages of statistical models,and promote Deep Learning application.The research idea of this article is:take the neural network technology to expand the method of mortality prediction as the core,from the four aspects of data,statistical model,algorithm model and application in turn,the use methods and modeling effects of different neural networks are studied.The research content and conclusions of each aspect are as follows.In terms of data.First,summarize the mortality database and explain the concept and calculation method of mortality.Secondly,the mortality in mainland China,especially the data status of the country,cities,towns and villages,and 31 provinces are explained,and the quality problems are analyzed by means of comparing data and quantitative errors.Finally,from the three perspectives of advanced old ages data correction,year interpolation and combination with micro data,the method of mortality correction is summarized,and the data is appropriately revised.The study found that:mortality in China is still different from international calculation methods;male data quality is usually better than female data quality,national and rural data quality is usually better than towns and cities;SVD-Comp model is suitable for interpolating provincial mortality.In terms of statistical models.First,review the development of the mortality model.Secondly,summarize the existing multiple population mortality models based on statistical models,explain from the aspects of model structure,parameter estimation,and model evaluation methods,and propose a parameter estimation method based on the combination of singular value decomposition method and least square method.It is suitable for solving most of the multi-population model parameters.Finally,using actual data,using a unified model evaluation standard,comparing several multiple population stochastic mortality models from the perspectives of fitting effect and prediction effect,including residual graphs,model robustness and rationality of results,etc.The study found that:the goodness of fit of the multiple population model is weaker than the single population model,but the prediction accuracy will be better than the single population model.Among them,the ACF1 model and the CAE model have better modeling effects;The multipopulation model has a coherent forecast,and the results are more in line with demographic theory;the multi-population model has a large difference in the prediction effect of different data in the sample.Therefore,when researching on a specific population,not only should pay attention to the average prediction accuracy,but also the prediction accuracy in a specific population;The multipopulation model has good model robustness,and the expanded sample has a positive impact on the model’s goodness of fit and prediction accuracy.In terms of algorithm models.First,the system summarizes the working principle of the algorithm model in mortality prediction,and classifies related algorithm models from three perspectives.Second,it focuses on three types of Neural Networks(NN),Recurrent Neural Networks(RNN),and Convolutional Neural Networks(CNN),including NN models,simple RNN models,Long-Short Term Memory(LSTM),Gated Recurrent Unit(GRU)and CNN models,a total of five models,using diagrams and codes to discuss the precautions.Finally,use actual data to construct a number of specific models to judge and compare the pros and cons of models from multiple perspectives such as statistics and demography.The study found that:the actual modeling effect of NN model and simple RNN model is better,and the calculation cost is lower.The modeling effect of NN6 model and RNN1 model designed in this paper is better than most statistical models;In the CNN model,LSTM model and GRU model,there are often situations where model fitting accuracy is high,prediction accuracy is low,or unpredictable.It has higher requirements for network structure design and activation function selection,and is more difficult to use;the effect of neural network modeling is relatively more affected by data quality,and the better the data quality Higher,the better the model effect;the neural network model is easier to expand in the face of higher-dimensional data,and can be adapted to different data,which solves the modeling disadvantages of statistical models.In terms of application.Starting with the law of life expectancy growth and longevity risk measurement,it mainly studies the differences in the application of different models to specific problems.On the one hand,it studies the growth law of life expectancy of the population by gender across the country,cities,towns and villages.On the other hand,it studies the longevity risks faced by provincial pension system.The study found that:Compared with statistical models,neural network models can effectively avoid the problem of underestimating the improvement in mortality,neural network modeling can more adequately deal with the problem of population aging;In the longevity risk measurement,the value of the neural network has lower capital requirements.Although it will weaken the ability of the institution to resist risks,it can enhance the investment and operation capabilities of the pension system.