Research Of Claim Frequency On Auto-insurance Based On Neural Network Models

Selected topic background of this paper is the car insurance market share of the domestic property company's overall size of more than sixty percent,and domestic property management comprehensive cost rate is high,so the car insurance product pricing is very important to our country property of the company's stable operation,and claim frequency research is vital for car insurance product pricing.Traditional car insurance claim frequency problem research mainly based on the generalized linear model,this method ignores the explain the interaction between variables,and larger scale,car insurance data samples are a big data science application category,allowing for large data algorithm of machine learning model has very extensive application value.This article selects machine learning algorithms of neural network model for research of car insurance claim frequency,because of the neural network model theory is derived from the abroad study,therefore,this article first reviewed the overseas about theoretical research and practical application of neural network algorithm model,based on the theory of neural network algorithm model abroad study process analysis,can help us better understand theory,neural network model based on the foreign research,we reviewed the domestic problems about neural network is applied in car insurance claims frequency of related research,The general idea of the neural network model for the regression analysis of claim frequency problem and the theoretical reason for the superiority of the neural network model over the traditional generalized linear model are summarized.Before and put forward for the problem of applying neural network algorithms claim frequency is insufficient,is before the deep theory study did not explore the feasibility of the neural network model,and no more shallow the merits of the neural network structure and deep structure of neural network and was given about the optimal neural network model of the structure of the research conclusion Because of the complexity of the structure of neural network algorithm model and the weak explanatory theory model,this paper analyzed in detail in chapter 2,the neural network type,and respectively with illustrations and formula of single layer neural network described the perceptron,double layer or multilayer perceptron neural network,the multilayer neural network that is deep learning of the three main feedforward neural network model theory and application of the method.Monolayer neural network is the most widely used neural network model structure with the most mature relevant theories.Monolayer neural network is equivalent to the logistic regression model and can be used for linear regression and linear classification problems.Monolayer neural network is also the neural network structure selected by empirical research in this paper.The structure of double layer neural network is the key point of neural network algorithm.However,perceptrons do not solve the xor problem.However under the premise of adding a computing layer,the two-layer neural network has very good nonlinear classification and regression effect.The multi-layer neural network is the deep learning we discuss.Since deep learning is widely used at present,this paper also briefly introduces the relevant theoretical basis of deep learning.Chapter in the third chapter is the focus of this article,first to build the neural network model for reference contrast of generalized linear poisson regression model fitting effect,in the construction of a neural network model,the input layer of neural network elements must be compatible with the dimension of the feature vector,and the output layer neural network unit number must be explained with the variable dimension to match.The number of neural network elements in the middle layer is determined by itself.Therefore,the model structure has great flexibility.However,the setting of the number of neural network elements will directly affect the fitting effect of the whole model.At present there is no perfect theory to determine the number of hidden layer neural network units.The general practice is to set based on experience.Feasible way is to preset several optional value,by calculating these values to evaluate the model prediction effect,choose the best prediction effect the choice of value as the final model Numbers of hidden layer neurons,this method is called the grid search method,and this chapter introduces a model structure complexity metric biggest by discussing the maximum complexity index respectively from the Angle of theory and practice to explore the neural network model is applied to the data regression analysis of the optimal structure and finally chose a contains 20 of shallow neural network as the number of neurons in hidden layer of shallow neural network used in the empirical modelChapter iv based on the theoretical model proposed in this paper,the commercial motor vehicle third party liability insurance policy to France claim DATA has carried on the empirical research,first introduced in this paper,the selection of DATA source,the DATA from the Swiss actuarial Association(Swiss Association of Actuaries,SAV)DATA SCIENCE working group,openly in Switzerland actuarial Association's official web site,the DATA can be directly invoked in the R software.Then through the analysis of sample data and simple after pretreatment with above build poisson generalized linear regression model and to simulate the shallow layer neural network model respectively,before the application of shallow neural network fitting,because of the characteristic of the neural network model,advance the sample data are normalized processing,unified and a dummy variable Settings for disordered explanatory variables.On the analysis of the model fitting effect,this paper introduces performance measurement model calibration,set it to the average deviation loss,poisson deviation loss by average poisson statistic model to evaluate the two different types of sample data,the fitting effect of the scheme by the sample data 9:1 ratio is set to validate and training sets,it is concluded that the sample evaluation index of the deviation and sample deviation.Neural network model has greater flexibility,don't need to advance to the structure of the neural network model to make certain assumptions,the natural advantages of neural network model can automatically identify the interactions between variables and automatically include them in the model,neural network model is more complex than the generalized linear model of the model structure,the number of neural network parameters are also far bigger than GLM parameter number,generalized linear model,so compared with traditional neural network model has better fitting effect.From the empirical results,the neural network model has better goodness of fitting,which further verifies the rationality of the theoretical model.At the same time,this paper compares with the traditional generalized linear model.The results show that the neural network model considering the interaction between explanatory variables can better reflect the influence of explanatory variables on claim frequency,improve the prediction accuracy of claim frequency,and lay a theoretical foundation for further pricing of non-life.