Research On Modeling Vehicle Insurance Claim Frequency From Classification Perspective

As the market reform of motor vehicle insurance rates continues to advance and competition in the auto insurance market intensifies,more scientific and reasonable auto insurance rate setting becomes an urgent need for insurance companies.The first step of auto insurance rate setting is claim frequency modeling.The traditional count-based claim frequency model using Poisson or negative binomial distribution can only predict the mean value of the number of losses,in order to improve the rationality and fairness of rate setting,this paper transforms the claim frequency from count-based to classification perspective,and transforms the claim frequency prediction into a multi-classification problem.Firstly,the paper discusses and introduces traditional count-based auto insurance claim count prediction models,including Poisson regression models and negative binomial regression models.This paper then turns to the transformation of the claim count problem into a multiclassification problem under the classification perspective.This paper compares the advantages and disadvantages of the "one-step classification strategy",which uses only one multi-classification model,and the "multi-step classification strategy",which transforms the multiclassification problem into multiple binary classification problems.The results show that the multi-step classification strategy performs better and is more adaptable to unbalanced auto insurance claims data.Secondly,this paper models the number of motor vehicle claims using traditional count-based claim count prediction models,as well as typical classification models dealing with classification problems,including logistic models,support vector machine models,decision tree models,and neural network models,in conjunction with actual auto insurance claim data.These models are also compared in terms of two evaluation metrics,the average Poisson deviation loss and the average absolute error,and the results show that each model from the classification perspective fits better than the traditional count-based claims count prediction model.Finally,the models are compared using the deviation of the predicted total premiums from the actual total claims by predicting the claim intensity through the gamma function,which enables the linkage with the insurance rates.The results show that the predicted total premium size of the model in the categorical perspective is lower than the actual total claims,while the predicted total premium size of the traditional count-based model is higher than the actual total claims.Therefore,in general,the categorical perspective has underestimated the overall risk,while the classical approach has overestimated the risk.When classifying high and low risk groups according to whether or not claims are made,the differentiation of risk by each model in the categorical perspective is better than that of the traditional count-based model.Also,in the low-risk group,the models from the classification perspective overestimate the predicted premium size to a lesser extent than the traditional counting models.However,in the high-risk group,the models in the categorical perspective do not perform as well as the traditional counting models.