Research On Prediction Of Auto Insurance Claim Severity Based On Bayesian Network

Auto insurance is the most important type of casualty insurance,its premium income directly affects the overall performance of property insurance companies.Therefore,providing more accurate auto insurance pricing is the goal of a lot of casualty insurance companies.The research of auto insurance premium pricing focuses on the discussion of rating factors,claim severity and claim frequency.The research of auto insurance claim severity is mainly divided into two aspects:model optimization and construction of new model factors This thesis analyzes auto insurance claim severity from the perspective of model optimization,and explores an effective predictive model solutionThe traditional method for predicting the severity of auto insurance claim is to establish a generalized linear model,in which Gamma distribution and Inverse Gaussian distribution can well fit the claim severity.In addition,the prediction effect of Lognormal distribution is also good,and these three models have their own more suitable application scenarios.We usually choose a model with the best performance for the data and estimate the parameters to obtain predicted results.However,a single model is difficult to predict large amounts of data in complex actual work,and we usually could not afford the loss of choosing the wrong models.To solve this problem,this thesis select two best models to fit the data firstly,and then improve the prediction accuracy by model averaging,in which the Bayesian Network is consifered to calculate the weights because of its advantages in classification.Firstly,this thesis analyzes and reduces the dimension of auto insurance data.Secondly,the generalized linear Gamma model,Inverse Gaussian model and Lognormal model are established to estimate the parameters and test the model of auto insurance claim severity,and two of the models with the best prediction effect are selected.The distance between the actual values and the predicted values of the two models are used as the basis of Bayesian Network classifier method for structure learning,and the optimal directed acyclic graph structure is obtained.When we input the information of each node,we can predict the probabilities that a policy follows the two models.Taking these two probabilities as the average weight of the model,we can get the final model prediction result by synthesizing all the information.The proposed method is applied to a set of real actual data,and the comparison results of MAE,RMSE and classification accuracy show that the comprehensive model based on Bayesian network has better prediction effect than the traditional models.The innovation of this thesis is to use Bayesian network to calculate the average weight of the models and take the results of various models into consideration to study the severity of auto insurance claim.Compared with some machine learning method such as ensemble learning and Adaboost methods,our model is more interpretative and could be a competitive method for auto insurance pricing practice.