Research On Credit Rating Adjustment Of Corporate Bonds Based On Machine Learning

Since the reform and opening up,the bond market has played an increasingly im-portant role in Chinese economy.The issuance volume of corporate bonds,as a critical part of the bond market,has been increasing year by year in recent years.However,bond defaults are also becoming the norm.And often when the default is approach-ing,corporate bonds are still concentrated in the high grade of AA and above,and the grade adjustment is very lagging,resulting in poor early risk warning capabilities.The business pain point faced by investment institutions is how to efficiently track and monitor the massive amount of corporate bond ratings and make timely and accurate adjustments.This paper use machine learning algorithms to prospectively adjust bond rating forecasts from the perspective of investment institutions.The method intro-duced in this paper can improve investment institutions’ efficiency of adjusting the corporate bonds’ rating in spite of investment institution human capital.In addition,the model,which can be immune from the influence of the personal subjective factors of analysts,is more objective than the traditional way such as scoring method.This paper mainly includes the following aspects:firstly,the purpose and signifi-cance of this paper are introduced through the background of the study,and theories of BP neural network and classification tree are introduced to prepare the theoretical pavement for the later empirical evidence.Secondly,BP neural network and classifi-cation tree models were constructed respectively to predict the direction of corporate bonds rating adjustment through the steps of data collection and cleaning,multiple covariance analysis and index system establishment.Due to the situation that the proportion of samples labeled "Keep"is too large and the number of samples labeled“Decrease”is much less than that labeled“Increase”,the samples are optimized to balance the data set using over sampling,under sampling,bivariate sampling,and SMOTE sampling.In the existing research on using BP neural network to predict corporate bonds rating adjustment,the original BP neural network was used.In this paper,genetic algorithm is utilized to optimize the BP neural network,getting GA-BP neural network model.As to the classification tree model,the best model result is gained by optimizing parameters of model.Finally,the article compares the effects of the two models and analyzes the importance of the selected metrics.The article concludes that,in terms of model classification performance and pre-diction accuracy,the prediction accuracy of the classification tree for the adjustment of direction of corporate bonds’ rating is overall higher than that of the BP neural network.The pruning classification tree model trained with over-sampling samples shows the best classification performance.The accuracy rates of the best-performing classification tree and BP neural network models are 87.68%and 67.90%,respectively.Compared with the original neural network,the optimization effect of GA-BP neural network is not obvious,and the accuracy rate increase is less than 10%.In terms of the importance of indicators,the article illustrates that the three indicators of return on net assets,current ratio and cash ratio play the most important role in predicting the trend of corporate bond rating adjustment,and the three indicators of whether it is a listed company,equity ratio and company attributes have the lowest importance,so profitability and solvency should be used as the main monitoring points of corporate bond risk.Combined with the conclusions obtained,the article points out that the issuer should keep up-to-date information on all dimensions of corporate bonds,so that investment institutions can track the quality of bonds and adjust the rating in time.From the perspective of regulators,they should supervise issuers to disclose information on bonds in time and remove unhealthy and high-risk bonds from the bond market to prevent false information from deceiving investors and damaging their interests.In summary,the research has some practical value in corporate bond risk warning.