Model Selection For Vehicle Insurance Data With Excess Numbers Of Zero-and-one Counts

Usually when we deal with some data problems,we will encounter data that cannot be solved by traditional statistical statistical models such as Poisson model,binomial model or negative binomial model.For example data samples inflated at sample points 0,1.Such data appear from time to time in insurance,genetics,psychology and biomedicine.In the non-life insurance actuarial field,the claim frequency data of auto insurance is typical such data.If the analysis does not fully consider too many zero points and one points,there will be a shortage of information to get accurate rates for non-life portfolios.The zero-inflated Poisson(ZIP)regression model is often used for zero-inflated count data,while the zeroand-one-inflated Poisson(ZOIP)regression model is considered a better model choice for count data that is inflated at both zero and one point..However,recent studies have shown that artificial neural network(ANN)models can also handle inflation count data well.In order to compare the differences between different models when dealing with inflation counting data,this thesis uses a variety of model evaluation criteria to analyze the inflation counting of the ZOIP regression model and ANN model in the case of small samples and large samples through simulation studies and case studies.The simulation study results found that in the case of small samples,the ZOIP regression model is the competitive model for dealing with zero-inflated and zero-and-one-inflated count data,while in the case of large samples,the ANN model gives better performance.In the case study,this thesis firstly tests the hypothesis of inflation data on the case data,and illustrates the advantages and disadvantages of the ZOIP regression model and the ANN model according to the test results.Furthermore,the comparison of related evaluation criteria show that when the sample size is small,the fitting and predictive ability of the ZOIP regression model is better than that of the ANN model,and when the sample size is large,the ANN model works better.