Consensus credit ratings: a view from banks

Abstract

While the production of credit ratings has long been limited mainly to rating agencies (CRAs), recent years have seen the growing popularity of consensus credit ratings crowdsourced from banks (i.e., bank ratings). We provide the first comprehensive examination of the properties and informativeness of bank ratings relative to CRA ratings. We find that bank ratings often deviate from CRA ratings, with over 60% of firm-months having different bank and CRA ratings. These deviations contain useful information. Bank ratings improve out-of-sample prediction of defaults and CRA rating revisions and explain the cross-section of credit spreads. However, bank ratings do not improve out-of-sample prediction of credit excess returns, indicating that current prices incorporate bank rating information. Overall our findings suggest that bank ratings are a useful supplement to traditional credit ratings.

1 Introduction

Credit ratings play a critical role in capital markets by reducing information asymmetry and facilitating capital allocation. A burgeoning literature examines the ratings produced by credit rating agencies (CRAs) (e.g., Beaver et al. 2006; Kraft 2015a; Bonsall et al. 2018; Beatty et al. 2019), which are key market gatekeepers (Roychowdhury and Srinivasan 2019). Recently, increasing investor demand and advances in financial technology have led to the rising popularity of credit ratings crowdsourced from another key player in the market: banks.¹ These consensus credit ratings (hereafter bank ratings) are used by various market participants, including institutional investors, insurers, creditors, and underwriters. We provide the first comprehensive analysis of bank ratings to shed light on their properties and informativeness.

Four distinguishing features of bank ratings motivate our study. First, banks collect detailed proprietary information about their borrowers, and, as a result, their collective opinion about the creditworthiness of borrowers may help other investors. Second, in part due to the widespread criticism of CRAs regarding rating quality, conflicts of interest, and perceived role in the great financial crisis (e.g., Securities and Exchange Commission 2003, 2008; Partnoy 2017), the demand for alternatives to CRA ratings has grown (e.g., Financial Stability Board 2014; European Securities and Markets Authority 2015). As a result, consensus bank ratings, which are crowdsourced from leading banks, have become increasingly important in credit risk assessments, especially among institutional investors. Third, due to information frictions and transaction costs, the debt market is segmented (e.g., Diamond 1991; Faulkender and Petersen 2006; Leary 2009; Morellec et al. 2015). Bank lending is concentrated among firms that do not have easy access to the bond market, such as those that are informationally opaque, smaller, or in the middle of the creditworthiness spectrum. Consequently, bank ratings can reveal information to investors for a new and different set of firms. Last, studies show that lenders shape firms’ information environments by influencing their disclosure and financial reporting incentives (e.g., Beatty and Weber 2003; Lo 2014; Kim et al. 2018). If bank ratings inform investors, they could represent an additional and previously unrecognized way that lenders shape firms’ information environments.

Our empirical analyses use data on 55,686 firm-month consensus bank ratings for 857 publicly traded firms from June 2015 to December 2022. We obtain the bank rating data from Credit Benchmark, an analytics company specializing in credit risk assessment through crowdsourced information.² Credit Benchmark collects, anonymizes, and aggregates risk estimates from contributing banks to create consensus ratings, which it provides to the contributing banks and data subscribers. We discuss Credit Benchmark’s methodology and data collection and aggregation process in Section II. Because bank ratings reflect the default risk of a firm rather than a specific debt obligation, we compare them to firm credit ratings from the largest three CRAs (i.e., S&P, Moody’s, and Fitch), which collectively represent approximately 95% of all agency ratings (e.g., White 2010; Securities and Exchange Commission 2022).³

Our first finding is that bank and CRA ratings complement each other in terms of their coverage. Among firms with bank ratings, 29.15% are not rated by major CRAs.⁴ Conversely, among firms rated by CRAs, 42.33% do not have bank ratings. Compared to firms exclusively rated by CRAs, firms only rated by banks tend to be smaller, less profitable, more volatile, and less leveraged. Notably, firms rated by both banks and CRAs are, on average, rated by six banks and two CRAs and are three times more likely to receive rating revisions from banks than CRAs, highlighting the potential value of crowdsourced bank ratings. Our second finding is that bank ratings often deviate from CRA ratings. The majority (64%) of bank ratings differ from CRA ratings, and 17% deviate by two or more notches. These deviations exist in all industries, years, and rating levels.

We consider two explanations for the differences between bank ratings and CRA ratings. First, disagreements between banks and CRAs may naturally arise if banks and CRAs use different inputs. To test this explanation, we examine the potential financial determinants of bank ratings and compare them with those of CRA ratings. We consider a broad set of credit risk determinants from prior research. Our results indicate that bank-CRA rating disagreements are influenced by several nonmutually exclusive factors, including rating difficulty, differences in how banks and CRAs weigh publicly available financial data, and the availability of other sources of information, such as traded CDS contracts.

The second potential reason behind rating disagreements is that bank ratings contain distinct information about firms.⁵ While CRAs can receive nonpublic information from their meetings with firm management (e.g., Jorion et al. 2005; Bonsall et al. 2017), lender-borrower relationships provide banks with unique and timely private information about their borrowers (e.g., Beatty et al. 2010; Plumlee et al. 2015; Chen 2016). Furthermore, research suggests that CRA ratings may not fully capture the likelihood of default but provide value in other ways. For instance, Hilscher and Wilson (2017) find that CRA ratings poorly predict raw default probabilities yet do measure the exposure to common (and hence undiversifiable) variation in default probability, which affects credit spreads and risk premia. Cornaggia and Cornaggia (2013) observe that CRAs are sluggish in responding to credit risk changes, as they offer rating stability, which may reduce the costs associated with frequent portfolio rebalancing and recontracting. These considerations suggest that bank ratings may provide credit risk information not fully reflected in CRA ratings.

Consistent with this conjecture, we find that bank ratings possess significant explanatory power for future defaults that is incremental to CRA ratings and firm fundamentals. In contrast, the explanatory power of CRA ratings is subsumed by bank ratings and firm fundamentals. Notably, adding bank ratings to CRA ratings and firm fundamentals significantly increases the accuracy of out-of-sample default forecasts, an improvement not observed when adding CRA ratings to bank ratings and firm fundamentals. Collectively, these findings underscore the usefulness of bank ratings in forecasting defaults.

Next we examine bank ratings’ incremental informativeness in predicting CRA rating revisions. We find that bank ratings have significant incremental explanatory power for future CRA rating upgrades and downgrades, after controlling for current CRA ratings and other credit risk predictors. Furthermore, bank ratings significantly improve the accuracy of out-of-sample predictions of CRA rating upgrades and downgrades. Overall the results support the prediction that bank ratings contain information that is not fully incorporated into CRA ratings.

We further illuminate the information contained in bank ratings by examining their ability to explain current credit spreads and predict future credit excess returns. We find that bank ratings explain the cross-sectional variation in credit spread levels and that this explanatory power is incremental to CRA ratings and firm fundamentals. These results suggest that bank ratings contain information not fully captured by CRA ratings and that this information is priced in credit spreads. However, we find that adding bank ratings to CRA ratings and firm fundamentals does not significantly improve the performance of out-of-sample prediction of future credit excess returns. These findings suggest that the market incorporates bank rating information into current prices.

Finally, we examine bank ratings for firms that are not rated by CRAs. We find significant differences in the determinants of bank ratings between CRA-rated and non-CRA-rated firms, suggesting that known determinants of ratings may not accurately portray credit risk for firms without CRA ratings. We also find that bank ratings of non-CRA-rated firms are significantly associated with proxies for credit risk, such as expected default frequency, Altman’s Z-score, and stock return volatility. Collectively, these results highlight the role of bank ratings in providing relevant credit risk information when traditional credit ratings are unavailable.

Our study contributes to several streams of literature. First, it contributes to the literature on credit ratings. Several studies examine the properties of CRAs’ ratings (e.g., Beaver et al. 2006; Cheng and Neamtiu 2009; Kraft 2015b; deHaan 2017; Bonsall et al. 2018; Beatty et al. 2019; Even-Tov and Ozel 2021). Others examine the information role of sell-side debt analysts (e.g., Johnston et al. 2009) and firms’ fixed-income conference calls (De Franco et al. 2022). We extend this research by providing the first comprehensive examination of credit ratings produced by another key group of capital market participants, banks. Our study improves the understanding of the properties and informativeness of bank ratings and the factors influencing their informativeness relative to CRAs. Our results suggest that bank ratings complement traditional credit ratings.

Second, our paper adds to the literature examining how banks shape firms’ information environments through their monitoring and information production. Research shows bank monitoring influences borrowers’ financial reporting and disclosures (e.g., Beatty and Weber 2003; Lo 2014; Kim et al. 2018). However, banks also produce private information about borrowers through internal loan risk assessments, which are typically not publicly disseminated. Notably, Beyhaghi, Howes, and Weitzner (2024) show that banks have an information advantage over markets by using supervisory data on banks’ expected loss estimates. Importantly, the nature of this private information can be shaped by factors like regulatory capital management incentives that may lead to strategic bias in risk assessments (Plosser and Santos 2018; Berg and Koziol 2017; Behn et al. 2022). By examining publicly available consensus bank ratings, our study advances knowledge of this new and important yet underexplored way banks transmit their private assessments of borrowers’ creditworthiness to capital markets.

Finally, our paper adds to the nascent literature on the informational role of crowdsourced forecasts in capital markets. Studies examine aggregated opinions of the online investor community (e.g., Jame et al. 2016; Bartov et al. 2018), employees (e.g., Hales et al. 2018; Huang et al. 2020), and customers (e.g., Huang 2018; Tang 2018). Our study extends this literature by examining aggregated opinions of lenders. While data from individual banks are not publicly available due to privacy concerns, our study illuminates the usefulness of anonymized aggregated bank ratings.

Our paper introduces the phenomenon of consensus bank ratings to the literature and lays the groundwork for future research. Widespread criticism of CRAs has generated a considerable debate among researchers about what motivates CRAs to improve their ratings (e.g., Cheng and Neamtiu 2009; Cornaggia and Cornaggia 2013; Beatty et al. 2019; Bonsall et al. 2022). Our findings that bank ratings reflect meaningful information about firms’ creditworthiness that is not yet incorporated by CRA ratings raise the question of whether banks can discipline CRAs by exposing their limitations. Second, given the importance of information transmission in the banking industry (e.g., Bushman et al. 2010; Ertan et al. 2017; Balakrishnan and Ertan 2021), examining whether the information transmitted by bank ratings influences decisions of other lenders is worth undertaking. Finally, given the significant interest in understanding factors affecting private firm financing (e.g., Minnis 2011; Badertscher et al. 2019), examining whether bank ratings influence private firms’ access to external financing is a fruitful area of future research.

2 Institutional background

Credit Benchmark is a private analytics company founded in 2012 with the goal of collecting credit risk views from lenders to provide an alternative to or complement to traditional credit ratings. Credit Benchmark contributors include more than 40 large banks, including 15 global systemically important banks in the United States, Continental Europe, the United Kingdom, Japan, Canada, Australia, and South Africa. These banks are Internal Ratings-Based (IRB) banks, employing their own rating models to assess borrower credit risk for business and regulatory purposes.⁶ Importantly, Credit Benchmark requires IRB banks to contribute their risk estimates and pay a subscription fee to gain access to Credit Benchmark’s consensus data. This ensures stable and credible participation from these banks. In recent years, out of approximately 60 IRB banks globally, more than 40 find the data valuable enough to contribute their estimates.

Credit Benchmark collects, anonymizes, and aggregates this information and provides consensus bank ratings to the contributors and data subscribers. Credit Benchmark and these contributing banks enter into a contractual agreement. This contract requires Credit Benchmark to provide aggregated data and requires the banks to contribute all their internal risk estimates, establishing mutual responsibility for the quality and completeness of the data. According to Credit Benchmark, bank ratings represent the wisdom of crowds by capturing risk views from more than 20,000 credit analysts from the contributing banks. Bank ratings cover more than 30,000 entities, including public and private corporations, financial firms, and sovereign entities, located in more than 100 countries. Credit Benchmark’s bank ratings date back to May 2015. As evidence of growing investor interest in bank ratings, beginning in November 2020, Bloomberg made Credit Benchmark’s bank ratings available through its terminals. Figure 1 plots the change in the number and market capitalization of bank-rated firms over time. It shows that the number of bank-rated firms has grown significantly over time, reaching 1,200 by the end of 2022. The plot also shows that bank-rated firms comprise a growing share of the total market capitalization of all Compustat/CRSP firms with positive debt, surpassing 85% by the end of the sample period.

Fig. 1

Credit Benchmark’s Coverage Over Time. This figure illustrates the coverage of Credit Benchmark from June 2015 to December 2022. The navy bars, plotted against the right axis, show the number of Credit Benchmark-covered firms that can be matched to Compustat-CRSP and have nonmissing values for the control variables used in our analyses. The red line, plotted against the left axis, shows the proportion of the total market capitalization of Credit Benchmark-covered firms relative to all Compustat-CRSP firms with nonmissing values for the same set of control variables used in our analyses

Credit Benchmark collects from contributing banks their estimates of borrowers’ probability of default, which is the likelihood that a borrower will default on one or more of its debt obligations within a specified period.⁷ Banks estimate probabilities of default using a hybrid through-the-cycle approach that combines elements of the through-the-cycle approach, which aims to reduce rating volatility by averaging probabilities of default over the credit cycle, and the point-in-time approach, which aims to obtain the most accurate risk estimates, given current market conditions. Because different banks have different assumptions about the length of the credit cycle, to ensure comparability, banks convert probabilities of default to an annual metric and provide these annualized hybrid through-the-cycle estimates to Credit Benchmark.⁸ After validating probability of default estimates collected from contributing banks, Credit Benchmark calculates and reports the mean, standard deviation, skewness, and the number of estimates of probability of default across contributors.

Credit Benchmark maps mean probabilities of default to the 21-letter credit risk scale commonly used by CRAs for ease of interpretation and to facilitate comparison with CRA ratings. The breakpoints for this mapping have been collaboratively established by Credit Benchmark and contributing banks. Appendix A shows the detailed mapping between probabilities of default and these 21-letter ratings. Specifically, for each bank rating level, the appendix delineates the lower bound of probabilities of default, expressed in basis points. For instance, probabilities of default exceeding 2.25 basis points but falling below 3.25 basis points correspond to an AA rating. We use these 21-letter ratings to examine differences between bank and CRA ratings.

Furthermore, to provide more detail, Credit Benchmark also offers 100-notch bank ratings, which add a finer level of sub-categories within each of the 21 notches. For example, BBB- is divided into four sub-categories. The last column of Appendix A shows the mapping between the 100-notch and 21-notch bank ratings. Our information content analyses focus on the 100-notch ratings, given their potential to reveal more information.

While bank and CRA ratings share several key features,⁹ they differ in at least two ways. First, bank ratings are explicitly tied to absolute levels of probabilities of default, as detailed in Appendix A and discussed above. In contrast, CRA ratings reflect relative probabilities of default (Standard & Poor’s 2022). Second, CRA ratings may, in some cases, consider not only the likelihood of default but also the expected loss rate in the event of default (Moody’s 2009).¹⁰ Also, even though banks’ probabilities of default incorporate the risk of a borrower’s default on any of its debt obligations, banks may be more concerned about and thus put more weight on the risk of default on their debt instruments. These differences may restrict the usefulness of bank ratings.

Banks have several incentives to contribute to the Credit Benchmark database. Contributing banks gain access (for a fee) to anonymized consensus data for all entities in the database. As mentioned earlier, IRB banks are not permitted to purchase the data (including through Bloomberg) without contributing. This policy reduces potential free-riding and incentivizes banks to participate in the database. The cost of information sharing is that making borrower performance information available to rivals can increase competition and reduce borrower retention (Liberti et al. 2022). The benefit is that the bank gains access to information provided by other banks, which reduces monitoring and screening costs, the problems with verifying borrower-reported credit history with each rival lender before lending to a new borrower (e.g., Padilla and Pagano 1997), and increases willingness to lend to new borrowers. Along these lines, Credit Benchmark states that its service is valuable for banks in assessing public and private entities that are not yet covered by their internal systems or lack traditional credit risk sources, such as agency ratings or market information. Banks also use Credit Benchmark data to benchmark their risk estimates against the consensus, which may help them improve their models by identifying potential inaccuracies in their risk estimates. While information sharing increases competition, it reduces information gathering and processing costs and risks of lending and can discipline the borrower’s behavior.

Credit Benchmark takes several steps to ensure the quality of the bank rating data. First, its contract obligates contributing banks to provide risk estimates for all their corporate clients, mitigating the concern of potential free-riding through partial data submission.¹¹ Second, all contributing banks complete a detailed methodology questionnaire, which allows Credit Benchmark to identify and adjust for possible differences across banks to ensure that different estimates reflect differences in credit risk views rather than differences in bank methodologies. Third, Credit Benchmark uses multiple checks to detect and quarantine unusual estimates until the contributor verifies the estimates.¹² Fourth, Credit Benchmark requires at least three estimates for the same entity, which helps preserve contributor anonymity and enhances its ability to detect unusual estimates. Fifth, Credit Benchmark periodically reviews the methodology in consultation with contributing banks. Still, contributing banks may provide incomplete or inaccurate data, which may reduce the usefulness of the consensus ratings. Also, if banks with relatively less private information are more likely to benefit from participation, this may reduce the informativeness of the bank rating.

3 Sample data and descriptive statistics

3.1 Data sources

Our analyses use data from various sources. We obtain consensus bank rating data from Credit Benchmark. These data include the consensus (i.e., mean), the standard deviation, and the number of bank ratings at the firm-month level. To match the Credit Benchmark data to Compustat/CRSP, we conduct fuzzy matching based on company name and manually check each potential match using Google search to decide whether it is a good match based on company name, industry, and country.

We obtain ratings produced by the three biggest CRAs—S&P, Moody’s, and Fitch— from Bloomberg terminals. We link these data to Compustat/CRSP using stock ticker symbols and confirm that each ticker corresponds to only one unique firm during our sample period.¹³ We restrict our attention to entity-level CRA ratings because we are interested in comparing CRA ratings with bank ratings, which are at the entity level. We follow prior research and convert CRA and bank ratings to the 21-point numeric scale, as shown in Appendix B. When more than one CRA rates a firm, we calculate the average rating rounded to the nearest notch, following prior research (e.g., Bongaerts et al. 2012).

We collect data on corporate defaults and bankruptcies from five sources: bankruptcydata.com, Mergent FISD, the UCLA-LoPucki Bankruptcy Research Database, S&P annual default and rating transition study, and Moody’s annual default reports. Defaults include missed interest or principal payments and distressed exchanges. We link the data from Bankruptcydata.com to Compustat/CRSP using EIN and CIK identifiers when they are available. We match the data from Mergent FISD to Compustat/CRSP using historical CUSIP identifiers. We exclude covenant violations (default_type = ”C”). We match the data from Bankruptcy Research Database to Compustat/CRSP using GVKEY. For the default data from S&P and Moody’s, we conduct fuzzy matching based on company names and manually verify each case to determine the best match.

We obtain bond transaction data from the TRACE Enhanced database and bond characteristics from Mergent FISD. We match these data to Compustat/CRSP using the Bond CRSP Link built by WRDS. We clean bond transaction data following the procedure outlined by Dick-Nielsen (2009, 2014) and compute monthly bond returns using the codes provided by WRDS.¹⁴

We source credit default spread (CDS) data from Markit and match them to Compustat/CRSP using CUSIP. We focus on five-year CDS contracts referencing senior debt (tier in (SNRFOR, SNRLAC)) with no-restructuring (doclause = XR14, i.e., no restructuring for data referencing the 2014 ISDA Definitions) as the restructuring clause and denominated in US dollars (i.e., currency = USD).

3.2 Sample construction

We provide details of our sample construction in Table 1 Panel A. The sample period spans from June 2015 (the start of Credit Benchmark’s coverage) to December 2022. The sample ends in 2022 to accommodate our analyses of one-year-ahead credit market outcomes. Our sample construction begins with 441,163 firm-month observations for 7,645 unique firms in the intersection of CRSP and Compustat. We then exclude banks and insurers and require nonmissing control variables for the baseline model. We further restrict the sample to firms with both bank and CRA ratings. Our main sample comprises 55,686 firm-month observations for 857 unique firms.

Table 1 Sample construction

To provide further insights, Panel A delineates the sample composition based on the availability of CRA and banks and shows their respective default rates. This breakdown shows that CRA ratings are available for more firms than bank ratings (96,567 and 78,596 firm-month observations, respectively). Among firms rated by CRAs, 57.67% have bank ratings. In comparison, among firms rated by bank ratings, 70.85% have CRA ratings. These statistics point to significant complementarities between bank and CRA ratings in terms of coverage. The last column reports one-year default rates. The average default rate in the initial sample is 0.71%, which aligns with prior research (e.g., Gutierrez et al. 2020). The average default rate in our final sample is somewhat lower (0.52%), reflecting the lower risk of firms rated by both CRAs and at least three contributing banks (the Credit Benchmark’s requirement for inclusion in the bank rating dataset).

Panel B of Table 1 details the steps of our bond sample selection process. The initial sample comprises 2,039,745 bond-month observations for 222,220 bonds covered by the TRACE Enhanced database from June 2015 to December 2022. We apply the following filters to this sample: we remove potential reporting errors in TRACE following Dick-Nielsen (2009, 2014); exclude observations with missing bond characteristics; restrict the sample to index-included bonds following Andreani, Palhares, and Richardson (2023); require issuing firms to be on CRSP; select representative bonds following Correia et al. (2018); and require nonmissing control variables for the baseline model. The sample is further refined to include only firms that are rated by both banks and CRAs. This process yields a final bond sample of 31,225 bond-month observations for 2,040 unique bonds.

3.3 Comparison of bank and CRA ratings

Table 2 Panel A reports the descriptive statistics for bank and CRA ratings. We use different samples to study rating properties and coverage. Accordingly, the table reports rating characteristics for firms rated by banks, firms rated by banks only, firms rated by CRAs only, and finally firms rated by both. The average bank rating in both the full sample of firms with bank ratings and the sample with both bank and CRA ratings is around 12, which corresponds to BBB-. Among firms with both ratings, the mean bank and CRA ratings are comparable (12.25 and 11.94, respectively). However, the majority (64%) of bank ratings differ from CRA ratings, with an average unsigned difference between bank and CRA ratings of 0.91. In 17% of cases, bank ratings deviate from CRA ratings by two or more notches. Sixty percent of firms are rated by five or more banks, and 82% are rated by two or all three major credit rating agencies.

Table 2 Characteristics of bank and CRA ratings

To take a closer look at the distribution of bank ratings, Fig. 2 plots the distribution along each notch of bank rating for all firms with such ratings as well as CRA-rated and non-CRA-rated subsamples, respectively. Not surprisingly, there are fewer observations at both ends of the rating spectrum. Most firms have a bank rating ranging from 9 (BB-) to 15 (A-). There is a kink at the rating of 11 (i.e., BB + , the rating immediately below the investment grade), which is driven by non-CRA-rated firms.

Fig. 2

Bank Rating Distribution. These figures plot the distribution of bank ratings for all firms rated by banks, firms rated by banks and CRAs, and firms rated by banks but not CRAs, respectively. The sample is a firm-month panel from June 2015 to December 2022. The horizontal axis indicates bank ratings with a higher value representing higher credit quality (lower credit risk). Appendix B reports the mapping between rating categories and the numeric scale

Figure 3 shows the distribution of CRA ratings against bank ratings. The horizontal axis indicates bank ratings, and the vertical axis indicates CRA ratings. Orange bubbles above (navy bubbles below) the 45-degree line represent firm-months with CRA ratings above (below) bank ratings. Green bubbles represent firm-months with CRA ratings equal to bank ratings. Bubble size indicates the relative prevalence of the corresponding firm-months, with a larger size indicating more firm-months. Overall the two ratings are highly correlated as most firm-months are close to the 45-degree line. Consistent with the statistics reported in Table 2, bank and CRA ratings disagree for a significant portion of the sample. The disagreement is especially commonplace at the mid- and low-range rating levels.

Fig. 3

Comparison of Bank and CRA ratings. This figure plots the scatter distribution of CRA ratings against bank ratings. The sample includes all observations for firms with both bank and CRA ratings from June 2015 to December 2022. The horizontal axis indicates bank ratings with a higher value representing higher credit quality (lower credit risk). The vertical axis indicates CRA ratings. Appendix A reports the mapping between rating categories and the numeric scale. Orange bubbles above (navy bubbles below) the 45-degree line represent firm-months with CRA ratings above (below) bank ratings. Green bubbles represent firm-months with CRA ratings equal to bank ratings. Bubble size indicates the relative prevalence of the corresponding firm-months, with larger size indicating more firm-months

Table 2 Panel B illuminates the time-series variation in bank and CRA ratings within firm. Notably, consensus bank ratings, benefiting from more contributors, are updated more frequently than consensus CRA ratings. In a given month, the proportions of bank and CRA rating revisions are 8.663% and 2.52%, respectively, for firms rated by banks and CRAs. Similarly, the frequency of upgrades and downgrades is higher for bank than CRA ratings. The more frequent updates for bank ratings highlight the potential value of crowdsourced ratings in more swiftly capturing changes in credit risk. The lower part of Panel B shows that, while the pooled standard deviation is similar for bank and CRA ratings (2.953 and 3.067, respectively), the within-firm standard deviation of bank ratings exceeds that of CRA ratings (1.032 and 0.757, respectively). This higher within-firm variation is consistent with more frequent updates of bank ratings, suggesting a higher responsiveness to changes in firm conditions.

Table 2 Panel C presents the number of observations and average values of key variables by year for the sample of firms with both bank and CRA ratings. The number of observations is lower in 2015, as the bank rating data became available in June 2015, and monotonically increases throughout the sample period. The credit quality of sample firms slightly deteriorates over time, gauged by both CRA and bank ratings. On average, bank ratings remain higher than CRA ratings throughout the sample period, with 33% to 52% of observations having bank ratings above CRA ratings and 16%–29% having bank ratings below CRA ratings.

Table 2 Panel D shows the distribution of observations across the Fama–French 12 industries. The energy and telecommunication industries have the worst credit quality, according to both CRA and bank ratings. Interestingly, the two ratings disagree the most for these two industries, although in different directions. Banks assign lower ratings in the energy industry, with a signed (unsigned) difference of -0.53 (1.206), and higher ratings in the telecommunication industry, with a signed (unsigned) difference of 0.555 (0.915). Utilities have the best credit quality as a whole, with both ratings almost reaching 14 (equivalent to BBB +). They also have the lowest rating disagreement with a signed (unsigned) difference of 0.145 (0.674).

3.4 Descriptive statistics

Table 3 presents the summary statistics for all variables, except bank and CRA ratings. Similar to Table 2, we report summary statistics for firms rated by banks, firms rated by banks only, firms rated by CRA only, and finally firms rated by both banks and CRAs. A comparison of firms exclusively rated by banks with those only rated by CRAs reveals that bank-rated firms are smaller (Size: 7.969 and 8.360, respectively), hold more cash (Cash: 0.137 and 0.094, respectively), are less leveraged (Market_Leverage: 0.219 and 0.470, respectively), are less profitable (Profit: 0.011 and 0.030, respectively), have more volatile profit (Fundamental_Vol: 0.059 and 0.032, respectively), and have fewer tangible assets (Tangibility: 0.280 and 0.383, respectively). The mean default rate for firms exclusively rated by banks is 0.0004. While firms rated solely by banks tend to be smaller, potentially raising concerns about incomplete default data, our study’s focus on firms rated by both banks and CRAs alleviates this issue. By requiring the availability of both bank and CRA ratings, our sample tilts toward larger, more profitable firms with lower fundamental volatility, where default events are less likely to be missed or underreported. Furthermore, in additional analyses reported in Table 9, we employ alternative measures of financial distress beyond defaults to further mitigate any concerns about data completeness.

Table 3 Summary statistics

4 Empirical results

4.1 Determinants of bank ratings and differences between bank and CRA ratings

We begin our analyses by examining factors influencing bank ratings, CRA ratings, and their differences. We consider commonly used determinants of credit ratings and credit risk models from the literature (e.g., Correia et al. 2012; Correia et al. 2018; Baghai et al. 2014; Bonsall et al. 2018; Ahmed, Wang and Xu 2024). Specifically, we consider the ratio of total cash and cash equivalents to total assets (Cash), interest coverage ratio (Interest_Coverage), profit-to-total sales ratio (Profit), the ratio of total debt to the market value total assets (Leverage), the natural logarithm of the market value of total assets (Size), debt-to-profit ratio (Debt_Profit), a loss indicator (Neg_Profit), the ratio of convertible debt to total assets (Convertible_Debt), the ratio of rental payments to total assets (Rent), the ratio of net property, plant, and equipment to total assets (Tangibility), the ratio of capital expenditures to total assets (CapEx), fundamental asset volatility (Fundamental_Vol), and market-based volatility (Market_Vol). We follow prior research and include industry and month fixed effects to control for industry and time-specific factors.

The first two columns in Panel A of Table 4 present the results of estimating the regressions of bank and CRA ratings on their predicted determinants for the sample of firms with CRA ratings. We find that bank and CRA ratings are associated with the predicted determinants in the same direction. These results suggest that bank and CRA ratings incorporate publicly observable accounting and market data in a broadly consistent manner. The relatively high adjusted R² values indicate that our model explains a significant portion of the observed variability in both types of ratings. This high explanatory power is consistent with prior research (e.g., Baghai et al. 2014), which finds substantial explanatory power of publicly available variables for CRA ratings.

Table 4 Determinants of bank and CRA ratings

The last column of Panel A presents the results for the determinants of bank ratings for firms not covered by CRAs. Although the coefficients are qualitatively similar for most variables, several determinants become insignificant. These results suggest that, despite some common factors that affect both bank and CRA ratings, previously documented associations between firm characteristics and credit ratings may not capture credit risk for firms not covered by CRAs.

Next we examine the determinants of disagreement between bank and CRA ratings. First, we consider the potential effects of other publicly available information sources. Because CDS spreads contain useful information about firms’ credit risk (e.g., Longstaff et al. 2005; Even-Tov 2017; Lee et al. 2018; Bonsall et al. 2022), we include an indicator of whether the firm has traded CDS contracts (CDS_Trading) to capture the availability of relevant information from the CDS market. We also include media coverage (Ln_News) because the media is an important source of information in capital markets (e.g., Bushee et al. 2010; Bushman et al. 2017). Furthermore, CRA ratings may be more rigorous for firms with greater media coverage (Bonsall et al. 2018). We include an indicator of recent restatements (Restatement) to capture the potential effect of the quality of accounting information, an important input to evaluating credit risk (e.g., Costello and Wittenberg-Moerman 2011; Alissa et al. 2013; Jung et al. 2013; Kraft 2015a; Bonsall and Miller 2017; Akins 2018).

As additional factors that may affect the differences between bank and CRA ratings, we include the number of banks that rate the firm (Bank_Raters), an indicator that the firm is rated by fewer than five banks (Bank_Raters < 5), an indicator that the firm is rated by two or more CRAs (CRA_Raters), the standard deviation of bank ratings among raters (Bank_Disagreement), and an indicator of split CRA ratings (CRA_Split). The number of banks or CRAs that rate the firm may influence the quality of the consensus ratings and the differences between them. Bank_Disagreement and CRA_Split likely reflect credit risk uncertainty and rating difficulty (Akins 2018), which could lead to greater disagreement between banks and CRAs.

Table 4 Panel B presents the results of the analysis focusing on the disagreement between bank and CRA ratings. Columns 1–3 (4–6) present the results for the unsigned (signed) differences between bank and CRA ratings. We present results for the specification that includes the base set of credit rating determinants (Columns 1 and 4), the specification that adds variables related to other sources of public information and the quality of accounting information (Columns 2 and 5), and the specification that includes all variables (Columns 3 and 6).

The results reveal several factors influencing disagreements between bank and CRA ratings. First, the negative coefficient on CDS_Trading in Columns 2 and 3 is consistent with the conjecture that bank and CRA ratings are more likely to agree when the CDS market provides relevant credit risk information. Second, the positive coefficient on Bank_Disagreement suggests that banks and CRAs are more likely to disagree about firms with greater disagreement among banks, consistent with these firms being harder to rate. Relatedly, the results in Column 6 suggest that bank ratings are more conservative than CRA ratings when firms are more difficult to rate, as evidenced by the negative coefficients on Bank_Disagreement and CRA_Split. When compared to Panel A, the results in Panel B columns 4 through 6 indicate that banks put relatively more weight on debt-to-profit, negative profit, and capex, while CRAs put greater weight on cash holdings and firm size.

Overall the results reveal that several factors that are not mutually exclusive influence the differences between bank and CRA ratings. These factors include rating difficulty, the use of different weights on financial metrics, and the availability of additional information on borrowers’ creditworthiness from traded CDS contracts.

4.2 Future defaults

Research suggests that banks obtain unique private information about client firms due to their lender-borrower relationships and access to management (e.g., Beatty et al. 2010; Plumlee et al. 2015; Chen 2016). Furthermore, banks obtain nonpublic information because they often process client firms’ financial transactions (e.g., Mester et al. 2007; Norden and Weber 2010).¹⁵ These considerations suggest that bank ratings may capture information that CRA ratings do not.

However, bank ratings may not contain significant incremental information. First, because the estimates of firms’ credit risk are inputs in banks’ calculation of regulatory capital, banks have incentives to manage these estimates, thereby reducing the informativeness of bank ratings (e.g., Begley et al. 2017; Plosser and Santos 2018; Behn et al. 2022). Second, while banks may have close relationships with borrowing firms, CRAs also have access to nonpublic information (e.g., Jorion et al. 2005; Bonsall et al. 2017). Thus whether and to what extent bank ratings contain information beyond that in CRA ratings is an open empirical question.

To investigate this question, we conduct both in-sample regression analysis and out-of-sample prediction. We focus on the 100-notch ratings, Bank_Ratings[100], as they are likely to contain more information. Our dependent variable, Default, is an indicator that equals one if the firm files for bankruptcy or experiences a default in the next 12 months and zero otherwise. For firms with multiple defaults, we retain only the first bankruptcy or default event.¹⁶

For a depiction of the distribution of default rates, Fig. 4 plots one-year-ahead default rates for each bank and CRA rating, using 21-notch bank ratings for ease of comparison to CRA ratings. Consistent with lower ratings indicating higher credit risk, we observe that default rates tend to be higher for firms with relatively low ratings and lower for firms with relatively high ratings for both bank and CRA ratings. Among riskier firms (ratings below B-), bank ratings appear to have a slightly better ability to discriminate firms with a higher default probability.

Fig. 4

Default Rates by Bank and CRA Ratings. This figure plots the average one-year-ahead default rates by bank ratings (black bars, using 21-notch bank ratings for ease of comparison to CRA ratings) and CRA ratings (gray bars) for each rating notch. The horizontal axis indicates bank and CRA ratings with a higher value representing high credit quality (low credit risk). Appendix B reports the mapping between rating categories and the numeric scale

Next we formally assess bank ratings’ predictive ability relative to two benchmark models using logit regressions. Motivated by Correia et al. (2018), our first benchmark model includes the distance to default barrier (LnV_X), recent excess equity returns (ExRet), the natural logarithm of the market capitalization (LnE), fundamental asset volatility (Fundamental_Vol), and market asset volatility (Market_Vol). Unlike Correia et al. (2018), we do not include the skewness, kurtosis, and the fifth percentile of the distribution of historical return on net operating assets because including these variables significantly reduces the sample size and Correia et al. find these variables to be generally insignificant default predictors.¹⁷ Our second benchmark model includes CRA ratings in addition to the default predictors in the first model.

Panel A of Table 5 reports the results of the in-sample regression analysis. Column 1 shows the results for the initial benchmark model that includes the market and accounting variables while excluding CRA and bank ratings. Supporting expectations, the results show that the probability of default decreases with the distance to the default barrier and recent stock returns and increases with market volatility. In Columns 2 and 3, we find that CRA and bank ratings each have significant incremental explanatory power for future defaults relative to the initial benchmark model in Column 1, with higher ratings corresponding to a lower likelihood of default. Column 4 shows that CRA ratings lose significance when bank ratings are added to the model. In contrast, bank ratings remain highly significant when both CRA ratings and the other default predictors are included. Overall the results in Panel A show that bank ratings have significant incremental in-sample explanatory power for future defaults.

Table 5 Predicting future defaults

Next we examine the out-of-sample predictive ability of bank ratings for future defaults. Following Correia et al. (2012, 2018), we employ an expanding window approach to ensure that all estimates are available at the time of default prediction and avoid overfitting of the data. This approach entails two steps. First, we estimate coefficients for each prediction model using past data. Specifically, for each month t starting in January 2017, we estimate the model coefficients using market, accounting, and rating data from the beginning of our sample period (i.e., June 2015) to month t-12 to predict one-year-ahead defaults. Second, we apply these estimated coefficients to the current market, accounting, and rating data in month t to obtain out-of-sample forecasts of the probability of default for months t + 1 to t + 12. We gauge the out-of-sample predictive ability using the area under the curve (AUC). An AUC of 0.5 (1) indicates no (perfect) predictive power.

Panel B of Table 5 presents the results of the out-of-sample prediction analysis. Each column corresponds to the respective model presented in Panel A. In Columns 2 and 3, we find that CRA and bank ratings each improve the out-of-sample predictive ability relative to the initial benchmark model: AUC exhibits a significant increase of 0.0243 and 0.0542, respectively. These improvements are economically meaningful and comparable to those observed in prior research (e.g., Correia et al. 2018). Specifically, these increases in AUC represent improvements in the predictive power of (0.0243/0.8536 =) 2.8% and (0.0542/0.8536 =) 6.3%, respectively. Notably, the model with bank ratings has a significantly higher out-of-sample predictive ability than that with CRA ratings. Column 4 reveals that incorporating both bank and CRA ratings improves the predictive ability relative to the model augmented solely with CRA ratings. However, it does not offer additional improvement over the model that only adds bank ratings (AUC is 0.9076 and 0.9078, respectively). This result aligns with the findings from Column 4 of Panel A, where the explanatory power of CRA ratings is subsumed by bank ratings. Collectively, the results in Panel B indicate that bank ratings have significant incremental out-of-sample predictive ability for defaults.

4.3 Future CRA rating revisions

In this section, we further illuminate the informativeness of bank ratings by investigating whether bank ratings contain credit risk information not yet reflected in the current CRA ratings. We do so by conducting in-sample regression and out-of-sample prediction analyses of whether bank ratings predict upgrades and downgrades in CRA ratings.

Panel A of Table 6 reports the results of estimating in-sample regressions. The dependent variable CRA_Up (CRA_Down) equals one if the CRA rating at the end of the examined window (i.e., the next three months, six months, and 12 months) is higher (lower) than the current rating. If the rating changes and then reverts to the original level, we record it as neither an upgrade nor a downgrade. The regressions examine the incremental explanatory power relative to the baseline model that includes the current CRA ratings and the market and accounting default predictors from Table 5.

Table 6 Predicting future CRA rating revisions

Columns 1 through 6 show the results for CRA rating upgrades. Consistent with bank ratings containing incremental information about future CRA rating revisions, we find positive and significant coefficients on bank ratings across all time windows, indicating that better bank ratings predict a higher likelihood of CRA rating upgrades. The results for downgrades yield similar inferences. In Columns 7 through 12, the coefficients on bank ratings are negative and significant, consistent with higher bank ratings predicting a lower likelihood of CRA rating downgrades. Taken together, the results for both upgrades and downgrades indicate that bank ratings contain information that is not fully reflected in current CRA ratings.

Next we conduct an out-of-sample prediction analysis and report the results in Panel B of Table 6. Like the default prediction tests, we use an expanded window approach to ensure the forecasts are made out-of-sample and gauge the prediction accuracy using AUC. Columns 1 through 3 report the results for CRA rating upgrades. We find that adding bank ratings to the baseline model significantly improves the prediction accuracy for upgrades in the next three and six months (AUC increases by 0.0063 and 0.0031, respectively), while there is no significant improvement for upgrades in the next 12 months. Turning to downgrades, the results in Columns 4 through 6 show that adding bank ratings to the baseline model significantly improves the prediction accuracy for downgrades in the next six and 12 months (AUC increases by 0.0053 and 0.0056, respectively), although the improvement for three-month-ahead downgrade prediction is not statistically significant at the conventional level. Note that these improvements in AUC are considerably smaller than those for default prediction, suggesting that changes in CRA ratings are relatively more difficult to predict. Overall the evidence suggests that bank ratings improve out-of-sample prediction of future CRA ratings revisions.

4.4 Credit Spreads

In this section, we examine whether bank ratings explain cross-sectional variation in credit spreads. Our approach follows the literature that employs credit spreads to gauge the relevance of the information to credit-risk assessment (e.g., Correia et al. 2012, 2018; Arora et al. 2014; Kraft 2015a; Beaver et al. 2019). If bank ratings capture relevant credit risk information not reflected in CRA ratings and other fundamental or market variables, assuming credit markets are reasonably efficient, we expect them to incrementally explain the cross-sectional variation in credit spreads.

We begin our analysis by plotting the average credit spread for firms with bank ratings above, equal to, and below the CRA ratings. Figure 5 presents this plot, using 21-notch bank ratings to align with CRA ratings for ease of comparison. Consistent with higher credit ratings indicating lower credit risk, the average credit spreads decline as CRA ratings increase. More importantly, within the same CRA rating notch, the average credit spread generally decreases with the relative level of bank ratings. Specifically, the average credit spread tends to be lower (higher) when the bank rating is above (below) the CRA rating for each level of CRA ratings. This univariate evidence suggests that bank ratings capture information about firms’ credit risk that is incremental to CRA ratings.

Fig. 5

Bank-CRA Rating Disagreements and Bond Spread. This figure plots the average bond spread for firm-months with bank ratings below (black bars), equal to (gray bars), or above (white bars) the CRA rating for each CRA rating notch. The horizontal axis indicates CRA ratings with a higher value representing high credit quality (low credit risk). Appendix B reports the mapping between rating categories and the numeric scale

Next we conduct regression analyses and report the results in Table 7. The dependent variable, Bond_Spread, is the contemporaneous bond credit spread, calculated as the difference between the bond’s yield and the yield of a modified-duration-matched Treasury security (Correia et al. 2018; Andreani et al. 2023). Columns 1 and 2 report the results from estimating Fama–MacBeth cross-sectional regressions, and Columns 3 and 4 show the results from estimating pooled OLS regressions with month fixed effects. Across both specifications, the results indicate that bank ratings have significant explanatory power after controlling for concurrent CRA ratings and the market and accounting default predictors. The coefficient on bank ratings is significantly negative, consistent with higher bank ratings indicating lower credit risk. Overall the results are consistent with bank ratings containing credit risk information that is priced in bond spreads but not reflected in CRA ratings.

Table 7 Credit spreads

4.5 Future credit excess returns

The results thus far suggest that bank ratings contain relevant credit risk information and that this information is priced in credit spreads. However, an interesting remaining question is whether the market fully incorporates the information from bank ratings in credit spreads. If investors do not, there may be mispricing in the credit markets, rendering future returns predictable.

We examine this possibility using an approach that follows Correia et al. (2012, 2018). First, we calculate a measure of potential mispricing as the difference between the observed and predicted spreads from the cross-sectional models in Table 7. If the predicted spreads contain useful information not fully incorporated in actual spreads, the differences between actual and predicted spreads would predict credit excess returns. Furthermore, if the market is not fully attentive to bank ratings, we would expect stronger return predictability when the prediction model includes bank ratings.

To this end, we conduct return predictability analysis by estimating cross-sectional Fama–MacBeth regressions. To assess the incremental predictive value of bank ratings, we compare the prediction performance for credit spread residuals based on (1) the baseline model from Column 1 of Table 7, CS_Res_Base, and (2) the baseline model augmented by bank ratings from Column 2 of Table 7, CS_Res_CB. We calculate these residuals using an expanded window approach to ensure the predictions are made out-of-sample. Given that returns are a continuous variable, we gauge the prediction performance using the adjusted R².

Table 8 Panel A presents the estimates from this analysis. The dependent variables are future excess credit returns for months t + 1, t + 3, t + 6, and t + 12. Excess credit returns are calculated as total bond return (percentage change in the bond price on the last day of the month, including accrued interest) minus the return of a modified-duration-matched Treasury issue.¹⁸ Following Correia et al. (2018), we control for short-term momentum (MOMS), long-term momentum (MOML), market capitalization (LnE), market-to-book ratio (MTB), earnings-to-price ratio (E/P), and equity market beta (Beta). Consistent with expectations and prior findings, we observe that both short- and long-term momentum positively predict excess returns in the next month. Turning to our main variables of interest, we find that both CS_Res_Base and CS_Res_CB are positive and significant for future excess bond return in months t + 1 and t + 3. These findings support the prediction that deviations of actual spreads from predicted spreads predict returns. However, the inclusion of bank ratings does not improve the performance. The adjusted R² values are very similar across the models without and with bank ratings for months t + 1 (0.213 and 0.212, respectively) and t + 3 (0.195 and 0.195, respectively), with the difference between the two models being statistically insignificant at the conventional level. The lack of improvement suggests that the market largely incorporates bank rating information into current prices.¹⁹

Table 8 Future bond excess returns

To corroborate our findings, we conduct portfolio analyses. At the beginning of each month, we construct five portfolios based on the quintiles of CS_Res_Base and CS_Res_CB. We report the time-series average excess credit returns for these portfolios in months t + 1, t + 3, t + 6, and t + 12. We also report returns for hedge portfolios that take a long (short) position in the top (bottom) quintile portfolio. The results, presented in Panel B of Table 8, show that both CS_Res_Base and CS_Res_CB exhibit significant return predictability. For example, the hedge portfolios earn excess returns of 0.549 and 0.543 percentage points in the next month, significant at the 5% level, for CS_Res_Base and CS_Res_CB, respectively. However, the inclusion of bank ratings does not improve the return predictability. The difference in hedge returns between CS_Res_Base and CS_Res_CB is statistically insignificant across all horizons. These results align with our findings from the Fama–MacBeth tests, confirming our conclusion that the market largely incorporates bank ratings into current prices.

4.6 Firms not rated by major CRAs

By design, the evidence in the previous sections is limited to firms rated by CRAs. One appealing feature of bank ratings, however, is that they often cover firms that are not rated by major CRAs. In this section, we extend our evidence on whether bank ratings capture relevant credit risk information to these non-CRA-rated firms. We follow prior research (e.g., Kedia et al. 2014; Bonsall et al. 2017) and use the expected default probability, EDF, based on the Merton (1974) model to gauge credit risk. We corroborate our findings using Altman’s Z-score as an alternative measure of credit risk as well as stock return volatility as a model-free proxy for risk.

We report the results of this analysis in Table 9. Panel A presents the results for EDF from the Merton model, calculated following the approach of Bharath and Shumway (2008), for months t + 1, t + 3, t + 6, and t + 12. Consistent with expectations, we find that bank ratings are incrementally negatively associated with EDF across all specifications, suggesting that bank ratings capture relevant credit risk information for firms not rated by CRAs. Panel B shows the results for Altman’s Z-score. Given that higher Z-scores imply lower credit risk, we expect a positive relationship with bank ratings. We find consistent results. The association between bank ratings and Altman’s Z-scores is significantly positive across all time horizons. Finally, the results in Panel C show that bank ratings are incrementally negatively associated with the volatility of daily stock returns across all specifications. Overall the collective results in Table 9 suggest that bank ratings continue to capture relevant information when firms are not rated by CRAs. These findings suggest that bank ratings can be a useful alternative to traditional credit ratings for firms outside the CRA’s coverage.

Table 9 Firms not Rated by Major CRAs

5 Conclusion

In recent years, advances in financial technology have significantly affected the sourcing and nature of credit risk information. Many leading banks contribute opinions about their borrowers to the anonymized crowdsourced bank ratings that are now widely disseminated to the investment community. We enhance the understanding of this phenomenon by examining the informativeness of bank ratings relative to CRA ratings.

We find that bank ratings often disagree with CRA ratings and that banks frequently rate firms that major CRAs do not. The disagreement between bank and CRA ratings is greater when firms are more difficult to rate. We find strong evidence that bank ratings contain information that is not fully reflected in CRA ratings. Controlling for CRA ratings and market and accounting determinants of credit risk, bank ratings improve out-of-sample predictions of future defaults and CRA rating revisions and explain cross-sectional variation in credit spreads. However, the addition of bank ratings does not improve out-of-sample prediction of credit excess returns, implying that bank rating information is already impounded in current bond prices. Overall we conclude that bank ratings complement credit ratings.

Our paper makes several contributions. First, by examining the properties and informativeness of bank ratings, it expands the understanding of a new phenomenon: consensus credit ratings produced by lenders. Second, our findings that bank ratings convey new information to the market highlight a new way that lenders impact the information environments of their borrowers. Finally, by examining crowdsourced lender opinions, our paper adds to the emerging research on how changes in technology affect the information environment of capital markets.

Appendices

Appendix 1. The mapping between the probability of default and bank ratings

This table reports the mappings between the probability of default, 21-notch bank ratings, and 100-notch bank ratings. For each 21-notch rating category, the second column presents the lower bound of probabilities of default denominated in basis points. For example, probabilities of default above 0 and below 1.25 correspond to an AAA rating, probabilities of default above 1.25 basis points and below 2.25 basis points correspond to an AA + rating, and so on. The third column presents the range of 100-notch bank ratings corresponding to the 21-notch ratings

	Bank_Rating	Probability of default lower bound	Bank_Rating[100]
Investment grade	AAA	0	>  = 98
	AA +	1.25	95–97
	AA	2.25	92–94
	AA-	3.25	89–91
	A +	5	86–88
	A	6.75	83–85
	A-	10	79–82
	BBB +	15	75–78
	BBB	22	71–74
	BBB-	33	67–70
High yield	BB +	48	62–66
	BB	78	57–61
	BB-	130	52–56
	B +	255	46–51
	B	400	40–45
	B-	650	34–49
	CCC +	1,000	28–33
	CCC	1,700	22–27
	CCC-	2,500	15–21
	CC	3,700	8–14
	C	7,000	<  = 7

Appendix 2. The rating scheme

This table reports the mapping between rating categories and the 21-point numeric scale for S&P, Moody’s, Fitch, and bank ratings

Numeric Rating	S&P	Moody’s	Fitch	Bank Rating
21	AAA	Aaa	AAA	AAA
20	AA +	Aa1	AA +	AA +
19	AA	Aa2	AA	AA
18	AA-	Aa3	AA-	AA-
17	A +	A1	A +	A +
16	A	A2	A	A
15	A-	A3	A-	A-
14	BBB +	Baa1	BBB +	BBB +
13	BBB	Baa2	BBB	BBB
12	BBB-	Baa3	BBB-	BBB-
11	BB +	Ba1	BB +	BB +
10	BB	Ba2	BB	BB
9	BB-	Ba3	BB-	BB-
8	B +	B1	B +	B +
7	B	B2	B	B
6	B-	B3	B-	B-
5	CCC +	Caa1	CCC +	CCC +
4	CCC	Caa2	CCC	CCC
3	CCC-	Caa3	CCC-	CCC-
2	CC	Ca	CC	CC
1	C	C	C	C

Appendix 3. Variable definitions

Variable	Definition
Rating Characteristics
Bank_Rating	21-notch consensus bank ratings for a firm in a given month
Bank_Rating[100]	100-notch consensus bank ratings for a firm in a given month
CRA_Rating	The mean of a firm’s credit ratings provided by S&P, Moody’s, and Fitch rounded to the nearest notch
Bank-CRA	Bank_Rating minus CRA_Rating
|Bank-CRA|	The absolute difference between Bank_Rating and CRA_Rating
Bank_Raters	The number of banks contributing to the consensus bank rating. Credit Benchmark does not provide the number of contributing banks when it is 3 or 4. For this reason, we set Bank_Raters to 3 when it is unavailable
Bank_Raters < 5	Dummy variable, equal to 1 if Bank_Raters is less than five
Bank ≠ CRA	Dummy variable, equal to 1 if Bank_Rating is not equal to CRA_Rating
Bank ≠ CRA[> 1Notch]	Dummy variable, equal to 1 if Bank_Rating differs from CRA_Rating by more than one notch
Bank > CRA	Dummy variable, equal to 1 if Bank_Rating exceeds CRA_Rating
Bank < CRA	Dummy variable, equal to 1 if Bank_Rating is less than CRA_Rating
Bank_Disagreement	The standard deviation of bank ratings
CRA_Raters	Dummy variable, equal to 1 if the firm is rated by two or more CRAs
CRA_Split	Dummy variable, equal to 1 if CRA ratings differ from each other, zero otherwise. The variable is set to zero when only one CRA rates the firm
Bank_Revision	Dummy variable, equal to 1 when Bank_Rating of this month differs from its value last month
CRA_Revision	Dummy variable, equal to 1 when CRA_Rating of this month differs from its value last month
Firm Characteristics
Default	Dummy variable, equal to one when the firm defaults or goes into bankruptcy within one year. We collect default and bankruptcy events from bankruptcydata.com, Mergent FISD (excluding covenant violations), the LoPucki Bankruptcy Research Database, S&P annual default and rating transition study, and Moody’s annual default reports. We only take the first default or bankruptcy event for each sample firm, following Correia et al. (2018)
CRA_Up(t + x)	Dummy variable, equal to one when the CRA rating as of month t + x improves relative to the current month t
CRA_Down(t + x)	Dummy variable, equal to one when the CRA rating as of month t + x deteriorates relative to the current month t
EDF	Expected default frequency, calculated based on the Merton (1974) model using the method described by Bharath and Shumway (2008). Specifically, EDF = N (− DTD) where N (·) is the cumulative standard normal distribution and DTD is the distance to default, defined as {ln[ (E + D)/D] + (rit−1 − 0.5σV2)T}/(σV√T) with E representing the market value of firm i’s equity in millions of dollars (shrout multiplied by the absolute value of prc from CRSP monthly), D the face value of debt in millions of dollars (dlcq + 0.5*dlttq from Compustat quarterly), rit−1 the firm’s stock return over the previous year, estimated by cumulating monthly delisting adjusted returns and requiring have at least six months’ data per year, σV the asset volatility, estimated as σV = E/(E + D)*σE + D/(E + D)* (0.05 + 0.25σE), σE the annualized volatility estimated over a moving 12-month window from daily stock returns for each month, and T equal to 1. We winsorize E, F, σE, and rit−1 at the first and 99th percentiles
Zscore	Altman’s Z-score, calculated as 1.2(actq-lctq)/atq + 1.4(req/atq) + 3.3(earnings/atq) + 0.6cshoq*prccq/atq + 1.0sales/atq where earnings is the sum of the most recent four quarters’ operating income after depreciation (oiadpq) and sales is the sum of the most recent four quarters’ sales (saleq)
Stock_Vol	Monthly stock volatility, calculated as the annualized standard deviation of daily stock returns in a given month
Firm Characteristics
Cash	Cash holding (cheq/atq)
Interest_Coverage	Interest coverage ratio (oibdpq/xintq), replaced by the 99th percentile of its distribution when xintq is 0, following Baghai et al. (2014)
Profit	Profitability, oibdpq/saleq
Market_Leverage	Market leverage, calculated as book value of debt (dlttq + dlcq) divided by the firm’s market value (market value of equity, (abs(prc)*shrout/1000), plus book value of debt)
Size	Market firm size, calculated as the natural logarithm of the market value of equity (|prc|*shrout/1000, in millions, at the month-end from CRSP monthly) plus book value of debt (dlttq + dlcq from Compustat quarterly)
Debt_Profit	Debt to profit ratio, (dlttq + dlcq)/oibdpq
Neg_Profit	Indicator variable that equals one if Debt_Profit is negative and zero otherwise
Convertible_Debt	Convertible debt divided by total assets (dcvt/at), using data from the Compustat annual file
Rent	Rental payments divided by total assets (xrent/at), using data from the Compustat annual file
Tangibility	Net property, plant, and equipment divided by total assets at the end of the fiscal year (ppentq/atq)
Capex	Quarterly capital expenditures (adjust cumulative year-to-date capital expenditure capxy to point-in-time) divided by total assets (atq)
Fundamental_Vol	Fundamental asset volatility, calculated as the standard deviation of return on net operating assets over the past five years, requiring at least four quarters of data. Following Correia et al. (2018), return on operating assets is the ratio of operating income after depreciation (oiadpq) to the average of net operating assets at the beginning and end of the quarter. We compute net operating assets as the sum of common equity, preferred stock, long-term debt, debt in current liabilities, and minority interests minus cash and short-term investments, ceqq + pstkq + dlttq + dlcq + mibq—cheq
Market_Vol	Market-based asset volatility, calculated as an annualized three-month rolling sample standard deviation of daily stock returns (requiring at least five observations, replaced with the cross-sectional mean when it is missing), multiplied by the ratio of market value of equity to total firm market value (market value of equity plus book value of debt) to adjust the equity volatility to asset volatility, following Campbell et al. (2008)
CDS_Trading	Dummy variable, equal to one when the firm has CDS trading
Restatement	Dummy variable, equal to one when the firm has a restatement issued in the past 12 months
Ln_News	Natural logarithm of news count, following Bonsall et al. (2018)
LnV_X	Dollar distance to default barrier in natural logarithm, where V is the market value of the firm and X is the sum of short-term debt (dlcq) and half of the long-term debt (dlttq), assuming half of the long-term debt will be due within one year, following the literature (Bharath and Shumway 2008; Campbell et al. 2008; Correia et al. 2018)
ExRet	Excess equity return, calculated as the difference between individual stock returns over the past 12 months and value-weighted market returns over the same window
LnE	Natural logarithm of market value of equity at the month-end (|prc|*shrout/1000)
MOMS	Short-term momentum, calculated as the stock return of the current month t
MOML	Long-term momentum, calculated as 11-month rolling exponentially weighted (a half-life of three months) average stock return, ending in month t-1
MTB	Market to book ratio, calculated as the ratio of the market value of equity (|prc|*shrout/1000) to the book value as of the quarter-end. Specifically, book equity is defined following Davis, Fama, and French (2000) and outlined in detail by Cohen, Polk, and Vuolteenaho (2003). Following Campbell et al. (2008), we adjust book equity by adding 0.1*the difference between market and book equity to deal with outliers and negative book equity
E/P	E/P ratio, following Campbell et al. (2008), calculated as net income (niq) divided by market value of the firm (market value of equity (|prc|*shrout/1000) in each month-end plus total liability (ltq) from the most recently released quarterly financial data). We then take 12-month rolling exponentially weighted moving average, ending in month t, to give larger weights to more recent data
Beta	Equity market beta, estimated from a rolling regression of one-year data of daily individual stock returns on value-weighted market returns (requiring at least 100 observations in each regression)
Bond Characteristics
Bond_Spread	The difference between a bond’s yield and the yield of a modified-duration-matched Treasury issue, multiplied by 100 for ease of presentation
BondExRet_LDM	Bond excess return in percentage, calculated as total bond return (percentage change in bond price on the last trading day of the month, including accrued interest) minus the return of a modified-duration-matched Treasury issue
BondExRet_L5M	Bond excess return in percentage, calculated as total bond return (percentage change in the last price of the month as long as it is traded within the last five days before the month-end, including accrued interest) minus the return of a modified-duration-matched treasury issue
BondExRet_EOM	Bond excess return in percentage, calculated as total bond return (percentage change in the last bond price of the month (irrespective of which day in the month), including accrued interest) minus the return of a modified-duration-matched Treasury issue
CS_Res_Base	Residuals from the baseline model of bond spreads (the same as the specification in Column 1 of Table 7), estimated using an expanding window approach
CS_Res_CB	Residuals from the expanded model of bond spreads (baseline controls plus consensus bank ratings, the same as the specification in Column 2 of Table 7), estimated using an expanding window approach
Bond_Age	Bond age, calculated as time passed from bond issuance to the current month t (in years)
Bond_Duration	Bond duration