Research On Reserve Estimation Method Based On Fusion Of Poisson Model And Neural Network

Insurance reserves refer to a certain amount of funds accrued from insurance premium income or surplus by insurance companies,its purpose is to perform insurance compensation liabilities as agreed.Insurance reserves can be divided into various types.This paper studies the estimation of outstanding claims reserves in non-life insurance business.Relevant regulatory agencies pay more attention to the estimation of outstanding claims reserves,because it is related to the sustainability of insurance companies’ operations and the vital interests of the insured.Relatively accurate estimation methods are of great significance.The traditional estimation methods of outstanding claims reserves mainly include chain ladder method,distribution-free stochastic Mack model,over-dispersed Poisson model and B-F method.In recent years,methods from the field of machine learning have also begun to be applied to estimates of reserves.The advantage of the traditional method is easy to be understood and used by business personnel,but its prediction deviation is large,while the machine learning model has a small prediction deviation,but the model interpretability is poor.First,this paper introduces the traditional methods in the field of reserves,and then attempts to integrate the over-dispersed Poisson model in the traditional method with the neural network model.The main purpose is to use the neural network model to improve the prediction accuracy of the traditional method,and can also retain part of the structure in the traditional model,which is convenient for business personnel to understand.The specific fusion process of these two models is achieved by embedding layers and skip connections.In order to fit the actual situation,compared with the previous single business line model,this paper will use the data simulator widely used in the field of reserves to simulate the claims development of different business lines of the insurance company.Then this paper will use the overdispersion Poisson model,the fusion model of a single business line and the fusion model across business lines to make predictions on the simulated datasets,and use the Bootstrap simulation method to compare the uncertainty of predictions of different models.The results show that the prediction bias of the two fusion models is relatively small,and the uncertainty of the model predictions does not increase significantly.This shows that the neural network can effectively improve the prediction accuracy of the over-dispersed Poisson model,and the methods in the field of machine learning can promote the development of reserve estimation methods.