Research On Auto Insurance Claim Frequency Based On Wide&deep Neural Network And Its Application In Reserve Evaluation

In China,the insurance industry is still operated separately,and auto insurance can only be operated in property insurance companies.Auto insurance has always been the largest type of insurance in the property insurance company,and the profit of auto insurance is also the largest in the overall profit of the property insurance company,so auto insurance has been playing a pivotal role in the property insurance company.In recent years,insurance regulatory authorities and insurance industry associations have paid more attention to regulating the auto insurance market.In order to regulate the auto insurance market and benefit consumers,insurance regulatory authorities and industry associations have carried out comprehensive reforms on auto insurance for many times.In September2020,China Banking And Insurance Regulatory Commission(CBRC)issued guidance on Comprehensive Reform of Auto Insurance,representing the beginning of a new round of auto insurance reform.In this reform,not only the automobile commercial insurance was reformed,but also the insurance amount of compulsory traffic insurance was modified to a certain extent.In this reform,almost all property insurance companies are affected.With the increase of auto insurance coverage,auto insurance customers need to pay lower premiums than before the reform,which undoubtedly requires insurance companies to improve their risk management level in auto insurance.In non-life insurance,the expected loss of the insured subject matter is the product of the claim frequency and the average claim intensity.There are many researches on the number of auto insurance claims.In this paper,the classical prediction model,the generalized linear model,the neural network based on machine learning,and the wide&deep neural network model are firstly introduced,and then the data set of auto insurance claims published by the north American non-life actuarial society(CAS)is fitted.In the actual claim data,the users who do not get out of insurance account for more than 95%of the total sample,and the unbalanced state of the sample tends to lead to the model judging all the insurance data as no claim.In this way,although the accuracy of the model can reach more than 95%,the model will be meaningless at this time.Therefore,this paper first uses SMOTE method to clean the auto insurance claim data,making the positive and negative proportion of sample in the training process is1:1.After data cleaning,wi DE&deep neural network is used to predict whether auto insurance customers have claims and the occurrence of 1or 2 claims.Meanwhile,the prediction results are compared with generalized linear Poisson regression,zero expansion Poisson regression and feedforward neural network.Finally,it analyzes the deficiency of the traditional flow triangle in the reserve evaluation,and how to realize the application of auto insurance claim frequency in the auto insurance reserve for example.