Research On The Evaluation Method Of Outstanding Claims Reserve From The Perspective Of Data Grouping

Reserves,as an important part of liabilities,are fund reserves reserved for fulfilling the liability for compensation of the insurance policies sold and related expenses.Therefore,whether it is in the analysis of the operating conditions of insurance companies or the evaluation of the supervisory effect of the higher-level competent authorities,reserves are in a relatively central position.The reserve for outstanding claims is the top priority in theory and practice because of its high uncertainty in its evaluation model,less known information,high evaluation difficulty,and high practical value.In this context,how to accurately evaluate has become a concern in recent years.This article mainly contains two parts,theoretical and empirical.The theoretical part mainly introduces entity embedding and cluster analysis into the current reserve theoretical framework,and makes appropriate adjustments to be more suitable for the reserve assessment problem to be studied,forming the basis of the reserve assessment model from the perspective of data grouping.Combining with the current research status of the related topics of outstanding claims reserves,the author comprehensively analyzes the theory and model of reserve assessment,and conducts corresponding theoretical explorations.Following the mainstream innovation trend,the neural network is introduced and used to explore the method of grouping the characteristics of important variables in the individual claim data.Then,the homogeneous data set when different variables are used as the grouping basis is obtained,and the individual claim data in each data set is sorted into the corresponding aggregated run-off triangle using the year of accident occurrence and the year that the accident has developed.With this process,it not only realizes the introduction of more individual claim information characteristics other than the type of business into the aggregate run-off triangle,but also provides solutions for multi-level variables as the basis for grouping,and effectively alleviate the inaccurate assessment problem caused by too little data when the grouping is too fine.The specific implementation method is to use the individual claim data of adjacent development years to model the chain-ladder factors using neural network.In the process of neural network model fitting,the embedding layer is used to learn the representation of the embedding vector of variable’s features,and then the dimension of the embedding vector is processed with the ideas of cluster analysis to obtain the conclusion about the feature grouping of each variable,and then the data can be homogenized grouping.Finally,the homogenized data set is used to get the aggregate run-off triangle to estimate the reserve.The empirical part is numerical research,which is based on the typical deterministic model chain ladder method,and stochastic model such as Mack model,generalized linear model（GLM）.First,analysis under the chain ladder method model.By comparing with the estimated value of reserves obtained in the case of heterogeneous data sets without data grouping processing,it is found that the result derived by the homogenous data sets grouped by the variables,especially the damaged parts,business lines,and quarters of accidents,is better,that is,it is closer to the real reserve value,which proves the validity of the reserve evaluation model based on data grouping.In addition,the random grouping of the data is realized by a simulation method,and then the estimation result of the total reserve under the grouping situation can be obtained.Repeat it about 1000 times,and get the quantiles and mean value of the total reserve estimate under these estimation results.It can be concluded that feature grouping can eliminate most of the heterogeneity is obtained by comparing the results of random grouping and variable feature grouping.Finally,further analysis under the stochastic model.We can see that the estimated value of reserve under the data grouped by the variable feature is less deviation than the reserve estimate by random grouping,which confirms that the claim reserving based on the data grouping proposed in this paper has good generality.In the conclusion of this article,the author puts forward supervisory recommendations based on the empirical results and prospects for the follow-up research content.