Credit ratings, viewed as CRAs’ reactions toward credit risk valuation, play an important role in the bond market because they correct the information asymmetry between bond issuers and financial markets (Bongaerts et al., 2012). Therefore, bond participants highly rely on credit ratings and view them as the license to the bond market (Partnoy, 2006). Abundant literature has studied the determining factors in credit ratings. For example, Altman et al. (1977), Xia (2014), and de Vries and de Haan (2016) pointed out that firms with better financial performance were more likely to receive high credit ratings. Bhandari and Golden (2021) found that CRAs had a stronger response to CEO’s political ideology. Camanho et al. (2020) revealed that CRAs tended to upgrade credit ratings when facing fierce competition in the credit rating market. However, there is little literature available for how CRAs or credit ratings react to financial news on the internet. Two representative studies conducted by Tsai et al. (2010, 2016) found that the text information from the newspapers contained incremental informational content for credit risk evaluation. Therefore, they concluded that CRAs had a significant reaction to the information from the newspapers. Lu et al. (2012) revealed that news was helpful to predict future credit ratings, implying that CRAs significantly reacted to the coverages in the newspaper. In this article, we will extend the previous literature to investigate the attitudes of CRAs to the financial news on the internet.

Abundant studies have pointed out that financial news plays an important role in the financial market, and those financial news coverages are widely used and analyzed by investors (Sabherwal et al., 2011; Narayan and Bannigidadmath, 2017; Du, 2020). Narayan and Bannigidadmath (2017) examined how investors reacted to the financial news on Islamic and conventional stock markets from 2005 to 2012. They found that both positive and negative news influenced stock returns, but positive news had a relatively larger impact on stock investors. However, Du (2020) got the opposite conclusion that stock investors in Japan had a stronger reaction to negative coverages. Based on 130 oil-related words, Loughran et al. (2019) investigated how investors responded to the media news on the oil market. Their findings showed that investors often overreacted to oil news. Corbet et al. (2020) examined the relationship between news coverages and Bitcoin investors’ reactions, and they found that bitcoin investors were not sensitive to CPI and GDP news but had a significant reaction to unemployment and durable goods news.

Recently, a handful of literature has focused on how the bond market reacted to financial news (Beber and Brandt, 2010; Defond and Zhang, 2014; Caporale et al., 2018). According to the contract theory, investors in the bond market were more conservative than investors in other financial markets (Kothari et al., 2010), implying that investor emotions in the bond market may be more sensitive to news coverages. Defond and Zhang (2014) tested how bond investors reacted to good and bad earning news. Their findings showed that bond investors were more likely to impound bad earning news on a timelier basis than stock investors. Beber and Brandt (2010) studied how investors responded to positive and negative macroeconomic news during economic expansions and recessions. They found that bond investors strongly responded to negative news associated with non-farm payrolls in expansions and positive news associated with inflation in recessions. Using a VAR-GARCH model, Caporale et al. (2018) investigated how bond investors reacted to macro news in the euro area from 1999 to 2014. Their findings showed that negative coverage associated with macro news positively affected bond investors in Greece, Ireland, Italy, Portugal, and Spain, and the responses became stronger in the 2008 financial crisis. However, there are several limitations in those studies. First, news used in existing literature mainly focused on the macroeconomic level, but seldom literature focused on the financial news on the internet. Second, previous literature mainly focused on the investor reactions in the secondary market rather than the investor reactions in the primary market. Therefore, this proposed article focuses on bond investors’ reactions toward the financial news on the internet at each enterprise-specific level in the primary bond market.

Previous literature has discussed the difference between central SOEs and other enterprises. For example, Zhao (2006) pointed out that central SOEs play an important role and enjoy many privileges in China’s economy. Lin et al. (2020) argued that central SOEs were viewed as a way to maintain social stability, without which the economy could not function properly. Ralston et al. (2006) analyzed different enterprise ownership types and concluded that SOEs contributed a significant share to the total output, particularly true for the sectors with strategic value. However, no literature associated the concept of enterprise ownership with the financial news on the internet. Considering that investor psychologies and stock market behaviors have different reactions to different news (Abreu and Mendes, 2012), we will extend previous literature and examine whether bond investors have a stronger emotion reaction to news about other enterprises than SOEs.

News timeliness has been studied extensively in previous literature. For example, Defond and Zhang (2014) examined the timeliness of the market reaction to earnings news, and they found that the timeliness of bond investors’ reactions to bad news was concentrated primarily among speculative-grade bonds. Matsubara et al. (2012) and Chen et al. (2021) concluded that the news influence would diminish over time. Eachempati and Srivastava (2021) argued that investors were inclined to ignore old news that may not stir the market sentiments, but they were sensitive to new news. In this article, our study will further analyze whether bond investors have different reactions to new news and old news.

If bond investors and CRAs have emotion reactions toward the bad (good) news, we will observe lower (higher) credit ratings or higher (lower) bond yield spreads. Thus, we employ credit ratings and bond yield spreads as the proxy variables of CRA and bond investor reactions.

The credit rating can be divided into two types. One is the bond project rating (BR), which can be regarded as the security of the bond itself. The other is corporate entity rating (CR), which can be viewed as the overall security of the bond issuer. Following Xia (2014) and Huang et al. (2020b), we employ the numerical value to quantify BR and CR as: AAA = 1, AA+ = 2, AA = 3, AA− = 4, A+ = 5, A = 6, A− = 7, BBB+ = 8, BBB = 9, BBB− = 10, BB+ = 11, BB = 12, BB− = 11, B+ = 12, B = 13, B− = 14, and lower than B− = 15.

Following Qian (2018) and Chen et al. (2021), the bond yield spread can be measured by the difference between the corporate bond yield and government bond yield, which is given as follow:

where _{yield_1} is the corporate bond yield, _{yield_2} is the government bond yield with the same maturity as the corporate bond.

To examine whether CRAs and bond investors have emotion reactions to financial news on the internet, we use following multiple regressions:

where X is the variable related to news sentiment indexes, including BadRatio, GoodRatio, DayBadRatio, and DayGoodRatio. _LnToalNews is the nature logarithm of the total number of news articles. CR is the corporate entity rating. Firm Control includes the natural logarithm of total assets, asset-liability ratio, total assets turnover, current ratio, return on equity, and enterprise ownership. Bond Control includes bond size, bond maturity, collateral clause, and bond type. RateControl includes two yields: 1-year government bond yield and government bond yield with the same maturity as the corporate bond. Year and industry fixed effects control specific time and industry factors. Table 1 describes these variables in detail.

We collect the data from several sources, ranging from 2010 to 2020. Data related to bond and firm characteristics are downloaded from Wind². 275,470 news articles are collected from China’s mainstream financial news websites using web-crawler technology. Deleting some observations with missing information, we finally select 2,510 bonds (central SOEs issue 690 bonds and other enterprises issue 1,820 bonds) issued by listed enterprises in China, excluding financial bonds, convertible bonds, notes, and asset-backed securities. Tables 2, 3 report the statistical results based on Stata 15.

Panel A of Table 2 reports the statistics about the whole news sample. The mean of TotalNews is 91.37, but the median is 19. BadRatio (mean of 0.181, median of 0.146) and DayBadRatio (mean of 0.209, median of 0.177) are relatively smaller than GoodRatio (mean of 0.683, median of 0.714) and DayGoodRatio (mean of 0.679, median of 0.700), indicating that more positive financial news is released on the internet.

Panel B of Table 2 provides the statistics about central SOEs. We find that central SOEs have relatively higher bond project ratings (mean of 1.268, median of 1), higher corporate entity ratings (mean of 1.362, median of 1), and lower bond yield spreads (mean of 224, median of 193). In addition, we also find that central SOEs have more positive news than negative news, which is consistent with Panel A.

Panel C of Table 2 presents the statistics about other enterprises. Consistent with Panel A and Panel B, other enterprises are also exposed to more good news. On the other hand, other enterprises have relatively lower bond project ratings (mean of 1.973, median of 2), lower corporate entity ratings (mean of 2.271, median of 2), and higher bond yield spreads (mean of 256.2, median of 239).

Table 3 reports the results of the Pearson correlation coefficients between the main variables. We can see that the correlation coefficient between LnTotalNews and CR is −0.3, indicating that bonds with higher credit ratings are exposed to more financial news. In addition, we find the negative/positive correlation between the bond yield spread and bad/good news, implying that bond investors may have negative/positive reactions to bad/good news.

Firstly, we study whether CRAs and bond investors have different reactions to financial news on the internet. The empirical results are shown in Table 4.

Columns (1)–(4) of Table 4 report the results of CRA reactions. Results show that the coefficients of BadRatio, GoodRatio, DayBadRatio, and DayGoodRatio are insignificant, indicating that financial news on the internet does not influence CRAs. This is inconsistent with Tsai et al. (2010), who found that news impacted CRAs.

The results may be interpreted as follows. First, CRAs extract the credit rating fees from bond issuers, whereas bond issuers can choose any CRAs they prefer. In this case, it is not strange that CRAs have motivations to underestimate financial news on the internet to please bond issuers and maintain the market share (Xia, 2014). Second, the financial news from the internet contains high noise and potential biases, which could mislead the market participants (Chen and Haga, 2021). Therefore, CRAs keep calm and tend to not adjust credit ratings before they see official announcements.

Columns (5)–(8) of Table 4 report the results of bond investor reactions. Results indicate that bond investors have significant reactions to both bad and good news. On average, 1% increase in bad news leads to an increase in the bond yield spread by 0.26 BP (column 5), while 1% increase in good news drives the bond yield spread down by 0.22 BP (column 6), indicating that the bad news has a more significant influence on bond investors. In addition, the coefficient of DayBadRatio (26.249, with a t-value of 3.01) and DayGoodRatio (−23.467, with a t-value of −3.31) in columns (7)–(8) show a similar conclusion that bond investors have an asymmetric reaction to bad and good news.

Overall, the empirical results indicate that bond investors have a stronger reaction to bad news than good news, which is similar to Du (2020) who found that negative news had a stronger influence on Japan’s stock market. Our results are also consistent with studies in psychology that negative information has a stronger impact on impressions than positive information (Baumeister et al., 2001). A number of explanations account for the asymmetric reaction have been put forward. The negative news bias may result from evolution as attention to negative information makes survival easier, and increases the possibility of genetic inheritance (Baumeister et al., 2001). The negative news bias also could be explained by loss aversion that investors care more about a loss of utility than a gain of equal magnitude (Kahneman and Tversky, 2013).

It is not amazing that central SOEs enjoy a tremendous news advantage in China. For example, central SOEs are more likely to get media attention and be exposed to positive news. Therefore, we divide the news into two groups: central SOE news and other enterprise news. Then we perform the interaction analysis to study the heterogeneous reactions under different enterprise ownerships. Particularly, we add four interaction terms, Dum_SOE × BadRatio, Dum_SOE × GoodRatio, Dum_SOE × DayBadRatio, and Dum_SOE × DayGoodRatio, to capture the differential reactions. Dum_SOE is a dummy variable, which is equal to 1 when the news is about central SOEs, but 0 otherwise. Table 5 shows the empirical results.

Column (1) shows that the coefficient of Dum_SOE × BadRatio is 30.919 with a t-value of 1.75, indicating that bond investors have stronger reactions to negative news about other enterprises.

Column (2) shows that the coefficient of Dum_SOE × GoodRatio is −32.737 with a significance at 10% level, indicating that bond investors strongly react to positive news about other enterprises. Columns (3)–(4) show a similar result that the coefficients of Dum_SOE × DayBadRatio and Dum_SOE × DayGoodRatio are 40.429 and −34.796, respectively, both significant at 5% level.

Next, we divide the sample into two subsamples: bonds issued by central SOEs and other enterprises. Then we conduct the regression analysis based on the two subsamples and compare the results with the interaction analysis. The empirical results are shown in Table 6.

In columns (1), (3), (5), and (7), the bonds are issued by central SOEs, and results show that the coefficients of GoodRatio, BadRatio, DayGoodRatio, and DayBadRatio are insignificant, indicating that bond investors are not influenced by the news about central SOEs.

In columns (2), (4), (6), and (8), the bonds are issued by other enterprises, and the results show that the coefficients of GoodRatio, BadRatio, DayGoodRatio, and DayBadRatio are all significant at 5% level, indicating that bond investors are more sensitive to the news about other enterprises than SOEs. This is also consistent with the results in Table 5.

Overall, our results confirm that there is heterogeneity in the psychological reaction where bond investors do not react to news about SOEs but to the news about other enterprises. Two reasons could explain our result. First, investors have adjusted their sensitivity to the news about central SOEs as too much news about central SOEs is released on the internet every day. Second, central SOEs could get strong support and implicit guarantees from governments (Huang et al., 2020a), implying that bonds issued by central SOEs are secure. Thus, market participants are not sensitive to the news about central SOEs.

Some news can affect the bond investors for only a long period, whereas some news has a short-term impact on the bond investors. To examine whether both new news and old news have different influences on bond investors, we divide the 6-month window into two 3-month windows, which is shown in Figure 1.

In Figure 1, T0 is the date when we issue the bond. We view the news in the first 3-month window as the past or old news, while news in the second 3-month window is viewed as the latest news or new news. Next, we construct the four new proxies to capture the news sentiment based on the two sub-windows, which is given as follow:

where i is equal to 1 if the news is released in the first 3-month window, but 2 in the second 3-month window. _{TND_i} is the total number of negative days in window i._TPDi is the total number of positive days in window i. _{D_i} is the total days in which the financial news is released on the internet for one specific enterprise in window i. D is the total days in which the financial news is released on the internet for one specific enterprise in the 6-month window.

Finally, we examine whether bond investors react to financial news based on the news in two sub-windows. The empirical results are shown in Table 7.

Columns (1)–(2) show that the coefficients of DayBadRatio₁ and DayGoodRatio₁ are both significant at 5% level, indicating bond investors react to bad and good news in the first 3-month window. Columns (3)–(4) show a similar result that news in the second 3-month window also exhibits a significant influence on bond investors. In column (5), the coefficient of CBadDayRatio₂ (29.251, with a t-value of 2.23) is large than the coefficient of CBadDayRatio₁ (23.641, with a t-value of 2.11), indicating that the latest news rather than past news has a stronger influence on bond investors. A similar pattern is observed about the coefficients of CGoodDayRatio₂ (−24.860, with a t-value of −2.92) and CGoodDayRatio₁ (−22.341, with a t-value of −2.74) in column (6).

Overall, our results confirm the asymmetric response driven by news timeliness that bond investors are more sensitive to new news than old news, which is within our expectations. Two explanations account for this phenomenon. First, old news contains gossip or uncertain information, which may fluctuate investor emotions in the past (Chen and Haga, 2021). However, as time goes by, the information in old news may have no fundamental values, suggesting that the influence of old news on bond investors will decline over time. Second, new news contains more risk and uncertainty, leading an emotional decision (Stracca, 2004).

Previous section in our paper calculates the bond yield spread based on the nominal interest rate when the bond is issued. In the first robust check, we use the average closing price in the secondary bond market to estimate the bond yield spread. Then we again perform the regressions to test investor reactions to financial news on the internet. Table 8 presents the results.

In columns (1)–(4), the bonds are issued by other enterprises. We observe that the coefficients of BadRatio, GoodRatio, DayBadRatio, and DayGoodRatio are significant at 1% level. In addition, the absolute value of the coefficients of BadRatio and DayBadRatio are relatively larger than the coefficients of GoodRatio and DayGoodRatio. Conversely, the bonds in columns (5)–(6) are issued by central SOEs, and we find that the coefficients of news sentiment variables are insignificant. Overall, the result confirms our previous conclusions.

The method in our empirical study to classify news may misclassify news type. In this subsection, we follow Chen et al. (2021) and measure the news sentiment as follow:

where N and P are the total number of negative and positive words for a specific enterprise.

Then we again perform the regression based new news sentiment variable. Table 9 presents the results. In columns (1)–(5), we calculate div based on the 6-month window. In column (6), div1 and div2 are calculated based on the first and second 3-month window, respectively.

Column (1) shows that the coefficient of div is insignificant, indicating that CRAs have no reactions to financial news on the internet. In columns (2)–(4), bonds are issued by all listed enterprises, central SOEs, and other enterprises, respectively. It is observed that the coefficients of div are negative in column (2) and column (4), both significant at 1% level, while the coefficient of div in column (3) is insignificant, indicating that bond investors are sensitive to the news about other enterprises but not to the news about central SOEs. Columns (5)–(6) show that the coefficients of div1 and div2 are 6.909 and 7.666, both significant at 1% level, suggesting that both old and new news influence bond investors, but bond investors have strong reactions to new news.

We also employ the instrumental variables (IVs) and conduct a two-stage least square (2SLS) estimation to avoid the endogenous problem. We introduce two variables that are theoretically non-related to the bond yield spread and bond rating but associated with DayBadRatio, DayGoodRatio, BadRatio, and GoodRatio as our IVs. These IVs are IV_BadRatio, IV_GoodRatioIV, IV_BadRatio, and IV_GoodRatio, which are the average of DayBadRatio, DayGoodRatio, BadRatio, and GoodRatio of other bond issuers with the same year, industry, corporate entity rating. A larger IV captures a higher likelihood of exposure to the medium, reflecting that bond issuers with similar performance and industry could get similar media attention. Next, we employ a first-stage probit model, as specified in Model (8), to estimate the predicted values of DayBadRatio, DayGoodRatio, BadRatio, and GoodRatio, respectively, which are then used as an instrument in the second-stage least squares estimation.

where X is DayBadRatio, DayGoodRatio, BadRatio, or GoodRatio, IVX is our IV candidates, which are IV_BadRatio, IV_GoodRatioIV, IV_BadRatio, or IV_GoodRatio.

Table 10 presents the 2SLS regression results. Columns (1), (3), (5), and (7) report the first-stage regression results, and we find the significant correlation between the four news indexes (DayBadRatio, DayGoodRatio, BadRatio, and GoodRatio) and our IV candidates (IV_BadRatio, IV_GoodRatioIV, IV_BadRatio, and IV_GoodRatio), thereby proving our instrumental variables’ relevancy. Columns (2), (4), (6), and (8) present the second-stage regression results, where DayBadRatio, DayGoodRatio, BadRatio, and GoodRatio in the baseline regression are replaced with the predicted value, Hat_DayBadRatio, Hat_DayGoodRatio, Hat_BadRatio, and Hat_GoodRatio, from the first-stage regression. Column (2) and column (6) show that the coefficient of Hat_DayBadRatio (coefficient: 68.436; t-value: 3.51) and Hat_BadRatio (coefficient: 67.202; t-value: 3.26) remains positive and significant at 1% level. Column (4) and column (8) show a similar results that the coefficients of Hat_DayGoodRatio (coefficient: −43.243; t-value: 2.75) and Hat_GoodRatio (coefficient: −38.93; t-value: −2.38) are significant. Overall, these findings confirm that bad (good) news can significantly affect the bond investor reactions, consistent with our previous results.