Research On Credit Risk Prediction Based On GRU-FCM Fusion Model

Society’s growth has made bank credit a critical element in maintaining financial stability.Credit risk is a complex systemic problem,so it is necessary to solve the problem from a macro perspective.It has become the mainstream to combine credit business with machine learning methods to build intelligent models with high prediction accuracy.Most of the current forecasting models can effectively forecast the credit business,but lack certain interpretability,which is not conducive to a better understanding of the system.Fuzzy Cognitive Map(FCM)is a graph structure network that combines fuzzy logic with three-value cognitive map.It has the characteristics of neural network.It can not only predict complex system effectively,but also has certain interpretability.Fuzzy cognitive graph is widely used in various fields because of its powerful logical reasoning ability and knowledge representation ability.Due to the influence of its structure and optimization algorithm,fuzzy cognitive graph will have a large error when studying a certain problem.Therefore,this paper starts from the structure of fuzzy cognitive graph and selects important influencing factors as the concept nodes of fuzzy cognitive graph for the credit data set,and deletes irrelevant factors to reduce the redundancy of the concept nodes of fuzzy cognitive graph.An optimization method based on Tabu-genetic algorithm is proposed for weight matrix optimization of fuzzy cognitive graph.Finally,fuzzy cognitive map is combined with deep learning model GRU to improve the predictive performance of model,and the effectiveness of the proposed model is verified by comparative experiments in credit risk prediction.The main innovations of this paper are as follows:(1)Structure optimization of fuzzy cognitive graphFor complex system modeling,fuzzy cognitive graph contains a large number of concept nodes and directed edges.Irrelevant concept nodes will reduce the computational efficiency of fuzzy cognitive graph.Therefore,a method based on Pearson correlation coefficient is proposed in this paper.By calculating the correlation between variables and target variables,irrelevant nodes are eliminated to optimize their structure.(2)We present an optimization algorithm of fuzzy cognitive graph weight matrixWhether weight matrix fits the system accurately affects the prediction ability of model.If weight matrix has a large error,it can not be predicted accurately.Therefore,on the basis of genetic algorithm,tabu genetic algorithm is proposed to optimize weight matrix.Tabu-genetic algorithm enhances local search ability by introducing tabu table and amnesty rule,so that it can search global optimal solution quickly.The error of the optimized weight matrix is ensured to be small so as to improve the prediction performance of the model.(3)We propose a credit risk prediction model combining GRU and fuzzy cognitive graphBanks’ primary source of revenue is the credit industry.Reducing non-performing loans is of great significance to the stable development of banks.Fuzzy cognitive map is used to model and forecast and help bank decision makers to make decisions.Fuzzy cognitive map prediction model shows good prediction performance,but there is still a certain gap compared with the most mainstream deep learning prediction model.Therefore,this paper proposes a credit risk prediction model combining GRU and fuzzy cognitive graph.The accuracy of credit risk prediction is improved by combining the two models while maintaining the interpretability of fuzzy cognitive map.For the above research objectives,the effectiveness of the proposed Tabu genetic optimization algorithm is verified by comparing the fitness of genetic algorithm with Tabu genetic algorithm and the predictive performance of corresponding models.For the credit risk prediction,Logistic,BP neural network,LSTM,XGBoost-RF and a single fuzzy cognitive map prediction model were compared to verify the effectiveness of the proposed FMC and GRU credit risk prediction model and the effect of improving the prediction performance after the fusion model.