Credit Risk Assessment Based On GP-GA Hybrid Algorithm

Credit risk is one of the most important risks encountered by the financial institutions. With the expansion of credit transaction, credit risk is increasing. As Line of banks open to the world and foreign banks come into China, domestic banks pay more attention to the entity quality as a result of competition. demand the bank sufficient rate of the capital amounts to 8%. Banks must decrease credit risk weighted entity. The research on the credit risk has profound affection on the banks even on the financial line. It is contributed to the country economic development healthily especially at the stage of transition of economic system. The research of our commercial banks on the credit risk is a current and later long-term work. Because of the lack of credit data in our country, the paper aims at finding a model idea based on the GP-GA hybrid algorithm to solve the problem of little sample. GP and GA are the branches of the evolution algorithm. They have the same idea that through encoding the problem, initiating original solution at random and then evolving the best solution generation by generation. Genetic programming is automatic match strategy, need no preliminary knowledge, the model could reflect the problem virtually and objectively.The character of this paper is employing variability and random of GP-GA to make up for the lack of the data and conquer the problem of little sample. Given the cost of errors of two types in the default event, we do not use the difference between predict solution and actual solution as fitness function to control genetic operations as usual in the part of the genetic programming, but employ a fitness function combining predict correct percentage and the probability of the errors and their costs from some credit risk materials. The bigger the solution is, the better the function is. However, in the part of genetic algorithm, we use two fitness functions to evaluate individual in order to gain the best parameters series.In chapter one, the paper summarize the situation of the existing credit risk analyzed methods from the view of nationality, quality and quantity, amount of samples. In chapter two, the paper from the view of corporation purposed behavior, credit system construction lagging, credit data availability, and models itself analyzes the reasons of the lack of credit data. The paper introduces two methods developed to solve the problem. One is Wang Chunfeng's little sample repeated use method, he suggested a model techniques enhancing limited sample utility efficiency to a large extent based on statistics principles. The other is Wu Desheng and Liangliang's propose of a strategy of generating samples based on genetic algorithms in 2004. In order to gain rational sample, they combined genetic algorithms and neural networks to establish a model for credit scoring. We find that there is little study on the little sample. Because of the good properties of GP in the model establishment, the paper attempts to solve little sample based on the GP-GA hybrid algorithm idea.In chapter three, the paper introduces genetic programming and genetic algorithm. Genetic programming is good at optimizing models structure while genetic algorithm is good at optimizing parameters of a model. We give the idea of the GP-GA hybrid algorithm combining the strengths of the two methods. And we process the data preliminarily. The paper attempts to optimize the structure of the model through genetic programming and optimize the parameters of the model advanced in last step and control the individual function by using the sample data.In chapter four, the paper provides genetic programming design process in JAVA language and parts of the source code including a series of function and method. List the applicable steps of genetic programming and genetic algorithm.The last part, we summarize the whole paper and analyze the limitation of the method. Stress the strength of genetic programming and summarize data process before the model.There are some genetic programming systems in some of the foreign lib. But there is little research on the genetic programming, so there is no such software in the Market. Individual wants to realize it only by designing program. But the model's feasibility is in no doubt. Genetic programming is used in all kinds of problems, provided that we encode the problem properly, supply the evaluation standard of the individual, then the computer will compute the automatic result for us. Genetic programming is promising comprehensively in practice. The characters of the data in our country is high dimensions, little sample. Much space leaved to be studied in little sample field later. In the problem of little sample, GP-GA is a good method to establish models, because GP supposes no certain structure of a model in advance. So it entitles the model much variability and random. We regard the data simulation as symbol regression problem. Genetic programming is good at solving sign regression problem. I hope more and more people attach importance to and study the little sample problem in establishing models.