Research On Enterprise Risk Prediction Based On Cerebellar Model Articulation Controller And Generative Adversarial Network

How to construct the business environment and promote the sustainable development of all walks of life is a crucial topic for the construction of an orderly and prosperous economic and financial society in our country,among which the effective control and prevention of enterprise risks is an integral part of the problem.The premise of efficient control of enterprise risk is accurate prediction of enterprise risk.However,enterprise data has the characteristics of high feature dimension,large differences between features and unbalanced samples,which are great challenges for financial institutions to predict enterprise risk.In order to solve the above difficulties,this thesis proposes a method based on deep learning,which combines cerebellar neural network and generative adversarial network to predict enterprise risk,and realizes its application in the field of economics and finance.In this thesis,the long-term credit risk prediction dataset of small and micro enterprises from Data Castle,the SBA enterprise fraud risk dataset from Kaggle,and the Taiwan enterprise bankruptcy risk dataset are used as experimental data.In the stage of data exploratory analysis,this thesis analyzes the sample imbalance of each dataset,and preliminarily explores the data distribution of some characteristic variables and their association with the risk dependent variable.In the stage of data cleaning and preprocessing,this thesis removes the default and irrelevant features,normalizes the data type,normalizes the data and logarithms.After processing,datasets A,B,and C have 98,33,and 94 feature variables,respectively.Deep learning is a branch of machine learning.It is an algorithm for feature learning of data based on artificial neural network architecture,which is widely used in image processing,speech recognition,and financial risk control.In this thesis,a predictive model of enterprise risk control is proposed by combining cerebellar neural network and generative adversarial network,which are commonly used in deep learning.In this model,the cerebellar neural network CMAC is used to reduce the dimension of the data with high-dimensional feature variables,and the generative adversarial network GAN is used to generate enough pseudo-high-risk samples to realize sample rebalancing.Finally,the fully connected network MLP is used to predict the risk label of enterprise samples.In the comparison experiment,this thesis selected six comparison models in the form of combining data augmentation method and prediction model.The experimental results show that the CMAC+GAN+MLP model proposed in this thesis has the recall rate increased by 7.57% on average and the FPR decreased by 10.9% on average compared with the ADASYN+MLP prediction model,which is the best performing model in the comparison model.The AUC is increased by 7.06% on average.In the ablation experiment,compared with the ablation experimental model,the recall rate of CMAC,GAN and MLP modules is increased by 67%,51% and 18% on average,the FPR is decreased by 59%,50% and 18% on average,and the AUC is increased by74%,58% and 17% on average.The comparison experiment and ablation experiment verify that the model proposed in this thesis has good performance in prediction accuracy and generalization performance,and both cerebellar neural network and generative adversarial network have a great role in improving the performance.In addition,the model proposed has relatively stable and consistent good performance on three different datasets,which reflects the model has relatively excellent generalization performance.Finally,we summarize the contents and conclusions of our research,and analyze the weakness of the research and the potential enhancements of future work.