Research On Automobile Insurance Claim Counts Based On Clustering Algorithm

In the auto insurance market,consumers’ willingness to renew their policies is low due to poor purchase experience and tortuous insurance payout,which has impacted the revenues of auto insurance companies.Only by setting reasonable premiums and improving the quality of subsequent coverage can we increase consumers’ intention to enroll and renew their policies.The technology of Ratemaking has a direct impact on the quality of auto insurance products,and the selection of rate factors and predicted number of claims as the core steps of rate setting have been hot research topics in auto insurance pricing.In the context of the expansion of insurers’ autonomous pricing power,technological innovation and data growth,the traditional actuarial models are slightly incapable of pricing,ignoring the non-linear effects of explanatory variables on the number of claims,and unable to deal with the lack of empirical data due to the high dimensionality of insurance data.To optimize auto insurance products,this paper applies an improved actuarial model with neural network clustering method to enhance the fairness of product rates.This paper compiles domestic and international literature from two perspectives:claim count models and clustering algorithms,summarizes the research materials of extended generalized linear model and neural network in auto insurance,and composes the source and improved algorithm of SOM neural network.Then the model principles of the generalized additive model with random effects and the two-stage clustering algorithms: The Self-Organizing Map(SOM)neural network and K-means,are described to provide theoretical support for the empirical process.The empirical evidence starts with visualizing the proportion of claims and the number of risks for each risk factor,showing its inter-group variability in the number of claims.Then,in order to investigate whether the selected count actuarial model and clustering algorithms are effective for rate determination,a generalized linear Poisson model and a generalized additive Poisson model are first fitted to the number of claims,and then a multi-category composite factor is obtained by two-stage clustering and is added to the generalized additive model as a random variable to regress the number of claims based on the comparison results of the evaluation indexes of these models.Through empirical research,it is found that: Compared with the traditional actuarial model,the generalized additive model is more accurate for the nonlinear factor "car price";When the number of risk factors is too large,the model is further improved in terms of fitting effect and parameter significance by downscaling some significant factors(car age,common accessory burden index,damageability score,repairability score)into one random variable through SOM neural network and KMeans,which also prevents that the number of samples within the class is insufficient;from the results of the optimal model,it is found that age and NCD have negative and positive effects on the number of claims,respectively,and the reason for the differentiation of gender on the number of claims is that the proportion of age and NCD is different among males and females.The high NCD and low age in females lead to a higher number of claims than in males.