《保险研究》20200707-《自编码器在死亡率预测中的应用研究》（张连增、申晴、丁宁）

［摘   要］在全球人口预期寿命不断提高的背景下,死亡率下降导致的长寿风险会直接影响到寿险公司的稳健经营和偿付能力充足率,因此,量化管理长寿风险至关重要。长寿风险量化管理的基础是建立动态死亡率模型,对死亡率进行准确预测。传统的Lee-Carter模型基于确定的模型形式,获得了良好的预测效果。本文应用自编码器,建立死亡率的神经网络模型,使模型通过训练能够自行学习到死亡率的潜在特征,并将模型的拟合结果和预测结果与传统Lee-Carter模型进行比较分析。结果显示,自编码器对死亡率的拟合效果与Lee-Carter模型相近,而预测效果显著优于Lee-Carter模型,说明自编码器显著提高了对死亡率的预测性能,能够为长寿风险的量化管理提供精确的数据基础。

［关键词］死亡率预测；Lee-Carter模型；自编码器

［基金项目］本文得到国家自然科学基金(61673225)、教育部重点研究基地(16JJD910001)的资助。

［作者简介］张连增,南开大学金融学院精算学系教授,博士生导师,研究方向:精算与风险管理；申晴,南开大学金融学院精算学系博士研究生,研究方向:精算与风险管理；丁宁,中国人寿保险股份有限公司精算部,研究方向:精算与风险管理。

The Application of Autoencoder in Mortality Prediction

ZHANG Lian-zeng,SHEN Qing,DING Ning

Abstract：In the background of the increasing life expectancy of the global population,the longevity risk caused by the decline of mortality will directly affect the steady operation and solvency adequacy ratio of life insurance companies,so it is very important to quantify the management of longevity risk.The quantitative management of longevity risk is based on the establishment of a dynamic mortality model to accurately predict mortality.The traditional Lee-Carter model is based on the determined model form,and has a good prediction effect.This paper used autoencoder to build a neural network model of mortality,which enabled the model to learn the potential characteristics of mortality through training,and compared the fitting results and prediction results of the model with the traditional Lee-Carter model.The results show that the fitting effect of autoencoder on mortality is similar to Lee-Carter model,but the prediction effect is significantly better than the model,which shows that autoencoder can significantly improve the mortality prediction performance and provide accurate data basis for the quantitative management of longevity risk.

Key words：mortality prediction;Lee-Carter model;autoencoder