Research On Auto Insurance Pricing Based On Generalized Regression Neural Network Model

Motor vehicle insurance,namely car insurance,it is with motor vehicle itself and the liability of the third party that cause is a kind of transport vehicle insurance of insurance mark.After nearly 8 years of development since the insurance business resumed in 1980,automobile insurance has emerged from many other kinds of insurance and officially become the largest insurance in China's non-life insurance market.From the relative proportion of the original premium income of motor vehicle insurance,the proportion of the original premium income in the non-life insurance premium income has been stable at about 70%since 2007.At the same time,our country has carried out the reform of the commercial car insurance rate since 2015,on the one hand,gradually expand the autonomy in the pricing of insurance company,on the other hand also to the insurance company to risk identification ability and car insurance rates reasonable science put forward new and higher level of requirements,how to build a precise frequency probability model is insurance company car insurance claim is the important subject to be solved.At present,the mainstream model advocated by scholars at home and abroad is still the generalized linear model.Aiming at the problem that the assumed distribution does not accord with the actual auto insurance claim data,on the one hand,scholars put forward many generalized linear models with pertinence.The empirical results show that the generalized linear model is more stable than the generalized linear model itself.On the other hand,scholars also pay attention to the hot spot in recent years--machine learning method,and try to apply the auto insurance claim data to the machine learning method represented by neural network model,so as to provide new ideas for auto insurance pricing model.Combined with this background,this paper mainly does the following work:A set of actual auto insurance claim data is used to construct the auto insurance cumulative claim prediction model by using the traditional generalized linear model and the neural network model.Among them,Tweedie distribution hypothesis is used in the generalized linear model,and the neural network model USES BP neural network,which is relatively common,generalization ability and generalized regression neural network,which is less difficult to model.On the one hand,the machine learning method represented by neural network can better fit auto insurance claim data,and its performance is better than the traditional generalized linear model.On the other hand,the generalized regression neural network modeling effect is better than a BP neural network,this is mainly due to the generalized regression neural network has the following advantages:first,the generalized regression neural network has the advantages of sensitivity to noise is not strong,so has the very strong nonlinear mapping ability and the learning speed,on the fitting ability and generalization ability is often better than BP neural network;Secondly,generalized regression neural network breaks through the local convergence characteristic of traditional machine learning methods represented by BP neural network,and has global convergence characteristic.Thirdly,the parameter of generalized regression neural network only has a smoothing factor,which is much less difficult than other machine learning methods in practical modeling and application.The empirical study shows that the generalized regression neural network performs better than BP neural network in the application of the model of forecasting the cumulative claim amount of auto insurance.In addition,the original data are divided into training set,verification set and test set in a ratio of 7.2:1.8:1 when the model is built.The main purpose of partitioning data sets is to prevent the over-fitting of models from leading to the generalization ability of models,i.e.the ability to adapt to new data is poor.Due to the small amount of data used in this paper,the division of original data will seriously affect the effect of modeling to a large extent.How to make the division of data sets uniform and independent is more critical.So when testing the independence of the three data sets,respectively using the Python software to map the training set,validation set and test set of car insurance claim amount of kernel density distribution curve,the graphic shows the trend of the three curves roughly coincides,shows three groups of similar to the distribution of data sets,namely the dataset partition results is effective.The innovation of this paper is mainly reflected in the following two aspects:(1)this paper abandons the framework of traditional machine learning method and USES generalized regression neural network(GRNN),which is seldom studied and applied at present,to establish loss prediction model.In the empirical study,the generalized regression neural network has advantages over the BP neural network,which makes it perform well in the application of the auto insurance cumulative claim prediction model.(2)Considering that most of the data in the practical application are changeable and dynamic,this paper takes generalization ability as the standard for evaluating the accuracy of the model,so that the model constructed can better adapt to the new data through the learning of historical data and play a certain role in risk prevention in practice.