A New Method for Identifying Bank Credit Risk: SVM-KNN Combined Model

Since 2019, the Baoshang Bank incident has attracted widespread attention from both academia and industry, sparking renewed discussions on credit risk management in commercial banks. Since the 2008 global financial crisis, risk management in commercial banks, particularly credit risk management, has gained increasing importance; however, the complexity brought about by economic globalization and the continuous emergence of innovative financial derivatives has made it increasingly difficult for commercial banks to identify and prevent credit risk.

In this context, using scientific credit risk models to detect, measure, analyze, and manage credit risk plays a positive role in the operational development of commercial banks and the healthy development of the entire socio-economic environment. Currently, the modern credit risk models widely used internationally include: KMV model from KMV Corporation, Creditmetrics Model from JP Morgan, Credit Portfolio View from McKinsey, Cridetrisk+ from Credit Suisse, and Mortality Rates Model. These modern credit risk measurement models have been widely applied in countries with highly developed financial markets, but their application effects in the credit risk management of commercial banks in China are not ideal. On one hand, due to the late start and imperfect system of China’s financial market, the asset value of listed companies cannot fully reflect on stock market value, affecting the accuracy of these models in application; on the other hand, most of these models require a large amount of sample data for support. Although China’s banking system has increased its efforts in data research in recent years, there still exist issues such as a relatively small overall number of credit data, missing historical data, non-standard statistical methods for some data, and relatively poor data reliability. How to construct a credit risk model suitable for the actual situation of Chinese commercial banks has been a key topic of discussion in both industry and academia.

Chinese commercial banks find it difficult to directly use internationally accepted modern credit risk models in the process of credit risk identification. This paper explores credit risk measurement models suitable for China’s national conditions. The SVM model only achieves high classification accuracy when the sample sizes are comparable; conversely, if the sample data is imbalanced, its classification accuracy is poor, especially in binary classification where the classification accuracy for unknown samples near the optimal hyperplane is not ideal. In comparison, the K-Nearest Neighbor (KNN) method has certain advantages in processing high-dimensional multi-class samples. Considering that the financial risk-related data indicators in China are characterized by a small amount of historical data, non-linearity, and high dimensionality, both SVM and KNN models demonstrate good classification accuracy when dealing with such data. The SVM-KNN combined model has achieved good application results in fields such as computer science and transportation. This paper attempts to construct an SVM-KNN combined model based on SVM and KNN models for credit risk identification in Chinese commercial banks.

The empirical results indicate that there are significant differences in the effectiveness of the SVM model, KNN model, and SVM-KNN combined model in classifying credit risk in commercial banks. Among them, the classification accuracy of both the SVM model and KNN model is 80%, which is lower than the 87.5% classification accuracy of the SVM-KNN combined model, indicating that the overall classification effect of the SVM-KNN combined model is significantly better than that of the SVM model and KNN model. Other classification effect indicators also partially explain this result. Firstly, the recall rate of the SVM-KNN combined model is 87.8%, slightly lower than that of the other two models, indicating that the proportion of normal enterprises determined by this model is slightly lower than that of the other two models. At the same time, the first error rate of the SVM-KNN combined model is slightly higher than that of the other two models, but it is within a reasonable range. Moreover, the precision rate of the SVM-KNN combined model is 96.7%, far exceeding that of the other two models, indicating that the proportion of truly normal enterprises among those determined by this model is quite high. Additionally, the second error rate of the SVM-KNN combined model is lower than that of the other two models. In the actual credit process of commercial banks, the second type of error in model classification results poses greater risks and losses to banks than the first type of error. Therefore, both the first error rate and second error rate are within reasonable ranges, and models with lower second error rates should be prioritized.

In summary, the classification effect of the SVM-KNN combined model is superior to that of using the SVM model and KNN model alone, and this combined model enhances and improves the application effects of the SVM model and KNN model in credit risk identification in commercial banks.

Accurately assessing the credit level of corporate clients and lending to high-quality corporate clients with good credit levels to reduce potential credit risks is not only an important aspect of bank credit risk management but also plays a significant role in improving the bank’s operational status. The SVM-KNN combined model proposed in this paper can effectively address the identification of credit risks in banks. Based on the above research content and conclusions, this paper offers the following recommendations:

First, apply scientific pre-loan risk management models to strengthen credit access management. It is understood that many banks still rely solely on experience and credit guarantees to assess loan projects in the actual pre-loan risk assessment and approval processes, making it difficult to accurately and scientifically judge the actual credit risk situation of enterprises, which is not conducive to managing and preventing credit risk. Therefore, it is recommended to conduct qualitative and quantitative analyses of factors that may lead to credit risk using the combined model proposed in this paper or other credit assessment models before making loan decisions, strengthening credit access management, and providing a basis for loan decision-making.

Second, continue to improve and enhance China’s banking credit reporting system and corporate database system. The modern credit risk models commonly used internationally and the combined model mentioned in this paper require substantial data support in practical applications. In recent years, China’s banking system has intensified its efforts in data research, achieving good results, but there are still issues such as a generally small amount of credit data and non-standard statistical methods for some data. Therefore, it is recommended that relevant departments continue to improve and standardize China’s banking credit reporting system and corporate database system, making the financial and non-financial data required for credit risk analysis by Chinese commercial banks more complete and standardized. This will positively promote the improvement of banks’ credit risk management levels.

Third, strengthen the integration of finance and technology to optimize the bank’s risk management system. The combined model proposed in this paper is an exploration of the application of machine learning algorithms in the field of credit risk identification in banks. In the current context, combining artificial intelligence, deep learning, and big data for credit risk identification is a trend. This will make credit risk identification more convenient and scientific but also requires a continuous learning and exploratory process. Therefore, it is recommended that banks deepen the application of computer science and technology in the financial sector, build specialized credit risk management systems, and achieve deep integration of finance and technology.

This is a refined version of the paper, representing only the author’s personal academic thoughts. For the full text, please refer to the Journal of Financial Regulation Research, Issue 7, 2020.

Credit Risk Stress Testing for Large Group Clients of Commercial Banks: Based on Monte Carlo Simulation Method

Research on Credit Risk of Credit Asset Securitization in Commercial Banks—Based on Modified KMV Model

Deconstruction, Quantitative Measurement, and Prevention Mechanism of Credit Risk Causes