A Study Of Auto Insurance Rating Problem Under Two-stage Neural Network Algorithm

In recent years,China’s motor vehicle ownership has been rising rapidly,and the proportion of emerging factors such as intelligent network-connected vehicles,self-driving cars,new energy vehicles and online ride-hailing cars is further increasing.At the same time,the new round of comprehensive auto insurance reform has started,and the marketization level of auto insurance premium rate will be further improved.The above factors will have a significant impact on auto insurance rating,and it is increasingly important to establish a more accurate and reasonable auto insurance pricing model.The current standard method of auto insurance actuarial practice is the GLM method,which mainly includes the twostage pure premium method and the one-stage aggregated pure premium method.The two-stage pure premium method models claim frequency and claim severity separately and assumes that they are independent of each other.This approach simplifies the model complexity,but ignores the possible dependency between claim frequency and claim severity in the claim data.At the same time,with the development of insurance technologies such as big data for telematics,the shortcomings of generalized linear models have continued to emerge,so the application of machine learning methods represented by artificial neural networks in auto insurance rating has gained attention.There are few studies on modeling the accumulated claims using artificial neural network algorithms,and the dependency of auto insurance data is rarely considered.Based on this background,this paper explores a two-stage neural network with a dependency structure from the perspective of improving the accuracy and rationality of auto insurance pricing models,which predicts pure risk premiums through a preliminary estimate and then revise approach.Due to the presence of a large number of categorical variables in auto insurance data and the lack of interpretability of general neural networks,this paper uses an embedding layer approach and the t-SNE algorithm for optimization.In the simulation study section,a set of simulated auto insurance loss data with significant zero-inflated,over-discrete and thick-tailed characteristics is randomly generated and fitted using a two-stage neural network.The results show that the two-stage neural network performs better than one-stage neural network.And the second-stage output value in the two-stage model reduces the error of the first stage.In the empirical study section,two sets of auto insurance data are used for the empirical analysis,namely,French third-party insurance data and auto insurance data of a domestic property and casualty insurance company.The empirical results show that the two-stage neural network has more accurate prediction and better risk differentiation ability than the one-stage neural network and traditional generalized linear model.Meanwhile,based on the embedding layer method and t-SNE algorithm,the model also realizes the downscaling and information metric of auto insurance pricing factors.In the context of the market-oriented auto insurance rating reform,this paper also introduces a simulated market competition mechanism to test the pricing ability and market competition ability of the model.The simulated competition results show that the two-stage model has strong price competitiveness in the above-mentioned simulated competition.It can identify high and low risks more accurately,and maintain profitability under the condition of obtaining a certain premium scale.Meanwhile,the price sensitivity test shows that the two-stage model also has operational stability.Therefore,the structure of the two-stage neural network,the embedding layer method and the t-SNE algorithm have positive implications for premium rate determination.In summary,for auto insurance claims data,the two-stage neural network can reduce prediction errors,improve prediction accuracy,and increase the explanatory power of the model.The method provides a reference for the application of artificial neural networks in auto insurance rating.