Sensitivity Of Decision Tree Algorithm To Class-imbalanced Bank Credit Risk Early Warning

For a long time, an important feature of financial markets is widespread credit risk, as a global problem, it is already integrated with the banking system, and an inevitable reality in the process of banking business. With the deepening of economic system reform in our country, gradually increased the number of corporate bankruptcies led to serious low quality of bank credit assets, and compared with other countries, the non-performing loan amount and ratio are in the forefront, bank credit risk has become a huge hidden trouble in the financial sector. Therefore, how to establish a bank credit risk early warning mechanism to reduce credit risk and improve the quality of credit assets is currently a very pressing issue.This article reviews the large number of domestic and foreign bank credit risk management literature, on this basis, the existing risk early warning management of bank creditproblems are analyzed, and focus on research of bank credit risk warning more realistic, more practical significance. In the processing method is mainly used methods of artificial intelligence, data mining and other related subjects on the basis of financial analysis theory, the category of non-equilibrium theory and decision theory on early warning of bank credit risk early warning system to study the theory and method to study early warning theory and method system of bank credit risk.First, for the previous column in a 1:1 ratio to select the number of positive and negative samples, this paper based on the ratio of 1:3 to collect ST and the ST companies, in a class-imbalanced perspective to study the problem, more realistic and rational.Then, combine with the Decision Tree algorithm which is more classic in the classification method of class-imbalanced data, the synthetic minority over-sampling technique (SMOTE) sampling methods, differences in sampling rate resampling technique (DSRA) and Bagging technology to construct the DSB-ID3 model, and the DSB-ID3 model and a single Decision Tree ID3 model, based oversampling decision tree ID3 model (OS-ID3), and based on the over-under double sampling decision tree ID3 model (OUS-ID3) to compare, analyze their strengths and weaknesses.Finally, at the empirical research process,select 138 ST companies and 414 non-ST companies, and select seven significant financial indicators, initial experimental data obtained by the relevant pretreatment. After analyzing the results of the empirical data, we can find that, G and F mean and overall accuracy rate of DSB-ID3 model obtained are the highest, and stability is the best. Therefore, DSB-ID3 model proposed in the above warning accuracy rate compared to the previous several models has been significantly improved, and the model has more practical value.