《保险研究》20190907-《基于机器学习算法的个体索赔准备金评估模型》（孟生旺、王海淘）

［摘   要］非寿险赔款准备金对保险公司的风险管理和财务决策具有重要影响。传统的准备金评估方法通常基于汇总的流量三角形数据进行建模,没有充分利用个体索赔案件的信息,且存在参数过度化、难以处理大额赔款和负增量赔款等问题。本文基于每份保单的个体索赔信息,使用随机森林和XGBoost等机器学习算法对案件的赔付状态、赔付金额分别建立了预测模型,改进了传统准备金评估模型的预测效果。实证研究结果表明,影响赔付状态的因素主要是结案状态、报案延迟等跟案件相关的信息,而影响赔付金额的因素则主要是历史赔付金额等反映出险事故严重程度的信息。本文最后还给出了RBNS准备金的预测分布,其结果更加接近准备金的真实值且方差更小,表明在非寿险RBNS准备金评估中,基于机器学习算法的个体索赔准备金评估模型优于传统的准备金评估模型。

［关键词］个体索赔信息；准备金；机器学习；赔付状态；赔付金额

［基金项目］本文获教育部人文社会科学重点研究基地重大项目“基于大数据的精算统计模型与风险管理问题研究”(16JJD910001),国家社科基金重大项目“巨灾保险的精算统计模型及其应用研究”(16ZDA052)和中央高校建设世界一流大学(学科)专项资金资助。

［作者简介］孟生旺,兰州财经大学“飞天学者”讲座教授,中国人民大学应用统计科学研究中心研究员,研究方向:风险管理与精算,E-mail:mengshw@ruc.edu.cn；王海淘,中国人民大学统计学院硕士研究生,研究方向:风险管理与精算。

Individual Claims Reserving Based on Machine Learning Algorithms

MENG Sheng-wang,WANG Hai-tao

Abstract：Non-life claim reserving has an important influence on the risk management and financial decisions of an insurance company.Traditional claim reserving models usually use aggregated run-off triangle data for modeling.These models do not make full use of individual claim information,and has the problems of over parameterization and inadequacy for dealing with large claims and negative claims properly.This paper developed an individual loss reserving model,using machine learning algorithms such as Random Forest and XGBoost to predict claim status and claim amount which improved on the projection results of traditional models.The empirical result shows that factors related to the case itself like settlement status and report delay have an important influence on claim status,while factors related to the accident seriousness like historical payments impact claim amount more.A predictive distribution of RBNS claim reserving was given by simulation at last which was more accurate and had lower variance.The result suggests that in the evaluation of non-life RBNS claim reserving,machine learning algorithms perform better than traditional claim reserving models.

Key words：individual claim information;claim reserving;machine learning;claim status;claim amount