Empirical Research On Credit Risk Prediction Of Commercial Banks Based On Deep Belief Network

The credit risk is the first of the three risks of commercial banks defined by the Basel Committee which include credit risk,market risk and operational risk.It is the most important risk confronted by modern commercial banks,and it is also one of the most frequent causes which lead to bankruptcy of commercial banks.The financial crisis in the United States in 2008 and the rapid growth of non-performing assets of commercial banks in China in recent years both warn us to take notice of the credit risk.Managing the credit risk and improving the ability of credit risk measurement and warning are the main means of enhancing the ability of risk prevention and control.In 2011,the China Banking Regulatory Commission introduced Guiding Opinions on the Implementation of New Regulatory Standards in China's Banking Sector and Administrative Measures for the Capital of Commercial Banks(for Trial Implementation)based on Basel III,which provide that qualified commercial banks can calculate risk weighted assets by employing internal evaluation,which requires commercial banks to measure credit risk accurately.The domestic researches on the management,measurement and prediction of the credit risk are in the ascendant,but there are a lot of gaps between the credit risk measurement capacity of the Chinese commercial banks and that of developed countries on account of many limitations.These factors urgently need to increase the level of credit risk measurement and predictive research in commercial banks.Based on our national conditions,this thesis aims to build up a credit risk measurement system which is suitable for the reality of commercial banks by means of studying and making reference home and abroad to the advanced models of credit risk measurement and warning and the latest artificial intelligence technology.The research is conducted from such perspectives as theories,practice,technology and service operations.The thesis gives a comprehensive overview about the warning systems,the measurement models and the implement tools of the credit risk and investigates its basic theories and its kernel ideas.By empirically analyzing the major cause of the rapid assets quality decline of China's joint-stock commercial banks,the present study puts forward a credit risk warning model for commercial banks which is urgently needed.Based on the characteristics of data mining in the age of information explosion,the idea is selected that artificial intelligence is a powerful tool for enhancing credit risk measurement capacity.By researching into the algorithm of the Deep Belief Network(DBN),this thesis points out that in solving labelled problems,the algorithm based on the Restricted Boltzman Machine(RBM)underutilizes the labelled data so that relevant information is discarded.To fully develop the characteristics of the labelled data,the DBN based on the classified and partitioned RBM is put forward to solve the problem of supervised classification.Classified and partitioned penalty terms are added to the layer unit parameters of the RBM,and the vectors of the penalty terms are based on the labelled values of the training sample,which produce corresponding vectors that conform to the Gaussian distribution for each group of labels.The vectorare determined at the time of system initialization and remain unchanged in the course of training.The penalty vectors employed are chosen according to the classification of labels for each time of training.The penalty terms are canceled when the system is optimized in the second stage of the DBN.Adding classified penalty terms may increase the uncertainty of the weight in the course of training and tendentious alter the influence of the weighting values according to the classification of the labels.The improved algorithm has an excellent effect on preventing over-fitting and under-fitting.Based on the financial data of the single enterprise,this thesis builds a financial crisis warning model.Guided by the idea of data mining and considering the difference between the artificial intelligence system and the traditional statistical model,the present study selects large quantities of financial indices as the research object and sets whether the enterprise makes profit as the aim predicted to establish a empirical research system which is based on the classified and partitioned RBM and the DBN.The predicting empirical research is conducted on the three time nodes: T1,T2,and T3——the first three years of the time node T.The predicting accuracy rates of the first type of risk of three nodes are respectively 90.28%,88.24% and 84.20%.The classification accuracy rate of the first type of sample which is relatively small is higher than that of the second type of sample which is relatively large.Compared to relevant work,the predicting accuracy ratesare higher,which also shows that the capacity of the improved algorithm's processing small sample data is greatly enhanced.The empirical research preliminarily builds up a systematic framework of the credit risk measurement and warning of commercial banks,which is of great practical use.In view of the defects of the model,a hierarchical network algorithm for semi supervised learning is proposed,which can effectively optimize the depth confidence network to solve the problem of supervised learning with different output of the same input.Using the a-share listed companies financial statement the empirical study has been carried on.The prediction of financial statements of the absolute error is small with high reliability and great breakthrough in the prediction of specific indicators.The results of this study will not only help commercial banks and other institutions to predict the future business situation and to provide reliable data for the analysis of experts in various fields but also provide reliable intermediate data for other degree model and improve the prediction depth.The conclusion of this thesis enables the system developers of commercial banks to reprocess and readjust the model according to their own data and build up directly their risk warning system that can be applied to the management of their banks after simple individualized modification.Adhering to the combination of theory and practice the thesis strives to proceed from the operational reality of financial institutions and build up the bridge of science and technology and business to enable technology to step out of the ivory tower and become the power source of economic development.