Corporate Bond Credit Premium And Common Factors

With the development of corporate bond market,the credit risk premium of corporate bonds has attracted much attention.Common factors are important basis for describing the risk premium.Empirical studies show that factor model is helpful to identify risk premium,and the application of common factor model to corporate bond market is helpful to estimate credit premium reasonably.Foreign scholars have studied the common market factors of corporate bonds,and selected bond factors based on different characteristics of bonds and confirmed their role in corporate bond credit premium.In contrast,the existing research in China pays less attention to the common factors of corporate bonds,and mainly based on macroeconomic conditions and bond characteristics to explore the issue of bond credit premium,do not involve bond portfolio and common factors.In the research methods,linear regression and principal component analysis are mostly used,which can not effectively distinguish the explanatory ability of single factor with the presence of collinearity between factors.The application of machine learning algorithm in bond market is seldom studied.Therefore,We constructs bond factor portfolio by using portfolio spread and linear regression,and selects three machine learning algorithms,i.e.linear regression,sparse learning and ensemble learning,to investigate the effects of new factors and their combinations on the credit premium of corporate bonds,and to select the risk factors with strong explanatory power,then compare the overall explanatory power of different factor models.We choose corporate bonds in China's bond market as our research object,considering the small issuance of corporate bonds before 2010,the sample range is from January 2010 to December 2016.As the China Securities Regulatory Commission officially promulgated the 'Regulations on the Issue and Trading of Corporate Bonds' on January 15,2015,the issuer of corporate bonds will be expanded to all corporate entities,therefore,we also divides the sample interval by January 2015.Firstly,we select the relevant factors based on the bond characteristics and transaction characteristics,and construct the bond factor portfolio by using the method of portfolio spread.Then we compare the different bond factors' Sharp ratio,the excess yield based on market credit risk and the excess yield after considering the stock market factors in order to test the effectiveness of each bond factor.Machine learning algorithms,such as linear regression,sparse learning and ensemble learning,are used to compare the explanatory power of factor models to corporate bond credit premium,compare the difference of goodness of fit between different algorithms and models to corporate bond credit premium,choose LASSO family method to get the frequency of non-zero estimation of each bond factor,rank the importance of bond factors' power in-explaining the credit spread based on random forest algorithm,in order to compare and evaluate the relative role of bond factors in explaining credit premium.The results show that:size,downside risk,value and volatility factor can obtain the excess yield,which can not be explained by market credit risk,Fama-French three factors.The performance of 'size effect' and 'value effect' of corporate bond in the whole sample is robust,while the downside risk factor and volatility factor did not show a significant return after entering 2015.Size factor has the smallest annual volatility,low beta for market credit risk,therefore relatively strong anti-cyclical defense function.By comparing the explanatory power among these bond factors,we can find that downside risk factors and value factors play a decisive role in the credit premium of more than half of the corporate bonds,and have a high degree of variable importance;market credit risk has a weak explanatory role in the credit premium of a single bond credit spread.Comparing with the traditional bond factor model,the five factor bond model proposed in this paper maintains a relatively high degree of goodness of fit regardless of the fitting method.The new five-factor model is superior to the traditional factor model in explaining corporate bond credit premium,and the bond factor model based on stochastic forest ensemble algorithm is the best.The innovations of this paper are as follows:Firstly,we construct the credit premium factor of corporate bonds from the aspects of corporate bond characteristics,market transaction characteristics and stock market factors,especially considering the new market factors such as volatility and downside risk.Secondly,three machine learning algorithms,linear regression,sparse learning and ensemble learning,are synthetically used to compare different models.We choose LASSO,adaptive LASSO and random forestto select variables for the factor regression model,which can improve the accuracy and stability of the fitting model while taking into account the variable selection and variable importance ranking,we can not only compare the role of single factor in corporate bond credit premium,but also evaluate the overall goodness of fit of the model.This paper has the following shortcomings:Firstly,the bond factor model proposed in this paper can not fully explain corporate bonds' credit spread because we construct factors from a cross-sectional perspective,pay more attention to the analysis of transaction characteristics,donot consider the impact of macroeconomic factors,and we donot consider the factors that reflect the macroeconomic interest rate term structure.There are still many potential factors that affect credit premium.Secondly,there may be a certain degree of autocorrelation in corporate bond credit premium.Because different bond credit premium series may show different time patterns,and the sample period is short,this paper can not effectively extract all sample bond unexpected credit premium changes,so this paper does not discuss the factor model in the bond unexpected credit premium changes.But the sparse learning algorithm and the random forest algorithm used in this paper are based on random sampling,which can alleviate the impact of single bond credit premium sequence correlation to a certain extent.