Study On Intellectualized Evaluation For Non-life Insurance Loss Reserving Based On Machine Learning

In practice,regular evaluation of reserves is a thing which needs to be done lots of work by professional evaluatior,it's certainly significant for insurance companies to realize inte-lligent and automatic evaluation through the machine learning algorithm.In this paper,the algorithms and historical data are combined to realize the idea of automatic evaluation.and the actuarial knowledge,algorithms' knowledge,data and realization of automation are mainly introduced.In the aspect of actuarial science,this paper mainly introduces the run-off triangle and the deterministic evaluation method,and the deterministic evaluation method includes chain ladder method,reserve progress method,case mean method and the B-F method,and compares them from three aspects of data,hypothesis and applicable conditions.The run-off triangle is a powerful tool for evaluation,and the above methods must be completed with the aid of this tool.In the aspect of algorithm,the development of artificial intelligence and the relationship between artificial intelligence and machine learning are summarized.Four kinds of machine learning algorithms are introduced in detail,which are probabilistic neural network,decision tree,support vector machine and generalized regression neural network.Where probabilistic neural network,decision tree and support vector machine are used for pattern recognition,and the generalized regression neural network is used for numerical prediction.Four algorithms are used to study the same problem.On the one hand,it is better to find out the possible problems of the study itself,on the other hand,to compare the matching of the test method and the problem.Then,the independent programming is carried out based on the theory of probabilistic neural network and generalized regression neural network through MATLAB.The results show that the factors of the probabilistic neural network selection are exactly the same as those of the artificial selection,and the factors of the generalized regression neural network are bigger.The research shows that the algorithm combined with historical data can solve the problem of automatic evaluation to some extent,but it is not the most ideal method.The main reasons are: 1)the data is not consistent and is not easy to collect,the inconsistency between the data will greatly reduce the generalization ability of the model;2)Although learning the historical data directly with the algorithm will improve the learning efficiency,the algorithm only imitates people,the ideal situation reaching the level of imitator,and can not overcome the shortcomings of the traditional evaluation method.Only in this way of studying further and creating theory,can we realize automatic evaluation in practice.