Credit rating accuracy: Ratings inflation
for issuers on the boundary of investment
grade?
Finance Master Thesis
Abstract:
___________________________________________________________________________ This thesis assesses the quality of credit rating information of firms from the USA after passage of the ‘Dodd-Frank Wall Street Reform and Consumer Protection Act’. Credit rating agencies publish higher ratings for firms with lower implied default probability by several models. More specifically, this thesis addresses the strictness of the rating standard when firms are on the boundary of investment grade. There are indications that credit rating agencies are more stringent when ratings are near the boundary between investment grade and speculative grade.
___________________________________________________________________________
Studentnr: s2054027 Name: Wouter Mulder Supervisor: Dr. M. Zaouras Study Program: MSc Finance
1 Introduction
Since the first publicly available bond ratings in 1909 (White, 2010), credit rating agencies have become important to capital markets (Rhee, 2015). Credit rating agencies provide professional opinions about the credit risk of a wide range of securities or their issuers (Standard & Poor’s, 2011). The three main credit rating agencies are Standard & Poor’s, Moody’s and Fitch who together dominate the credit rating market.
Higher credit ratings are assigned to firms that are perceived to have higher creditworthiness by credit rating agencies. High credit ratings are favourable for the firms as they are perceived to be at lower financial distress. It has been confirmed empirically that an increase in credit rating is followed by an increase in price whereas a decrease in credit rating is followed by a decrease in price for both bonds and stocks (Hand et al., 1992). An additional advantage of high credit ratings is that credit ratings are incorporated in regulations for many financial institutions (White, 2010). Higher credit ratings can be associated with lower capital requirements and as such firms might attract more investors when their credit rating increases.
It has been suggested that credit rating agencies did not provide correct credit ratings as the bankruptcy of some major companies had not been predicted. Prior research indicate that the quality of credit ratings in the United States was lower in the period before 2002 than in the period after 2002 (Cheng and Neamtiu, 2009). Despite this increase in quality, the ‘National Commission on the Causes of the Financial and Economic Crisis in the United States’ has claimed that credit rating agencies failures were the “cogs in the wheel of financial
destruction” (Financial Crisis Inquiry Commission, 2011) during the 2008 financial crisis.
This thesis addresses the question if there is inflation of credit ratings for issuer credit ratings that are on the boundary of the investment grade. First the reliance of credit ratings on both accounting and market based measures will be analysed. This thesis will directly compare credit ratings and alternative default probability estimators. Prior research often uses several explanatory variables to estimate credit ratings but few papers use alternative default probability estimators directly.
addresses the pre-crisis period or the period around the financial crisis, few address the period after the enactment of the Dodd-Frank act. Furthermore, the focus of this thesis will be the credit rating standard around the investment grade bound. The credit rating standard around the investment grade boundary will be assessed by performing pooled OLS and ordered logit regression analysis.
The results indicate that the Z-score, O-score and a default probability implied by a market model are incorporated in long-term credit ratings on the issuer level. Using these default estimators, a bias indicating stricter rating standards is found for firms just above the speculative grade boundary.
This thesis will be structured as follows: section 2 provides a review of the relevant literature and the hypotheses will be developed, section 3 contains a description of the methodology and describes how the hypotheses will be tested, in section 4 the sources and basic description of the data will be given, in section 5 the results will be provided and, section 6 will provide a conclusion inferred from these results and a discussion regarding the research and potential future work.
2 Literature Review
2.1 Origins of credit rating agencies
In the first half of the 20th century the three main credit rating agencies or ‘nationally
recognized statistical rating organizations’ (NRSRO) were founded: Moody’s, Standard & Poor’s, and Fitch (White, 2010). These agencies used to rate bonds only (Partnoy, 1999) but extended its business to providing rating on more complicated products (Wojtowicz, 2014) and issuers (Johnson, 2004). The definition of an issuer-level credit rating of S&P is as follows:
“Credit ratings are opinions about credit risk. Standard & Poor’s ratings express the agency’s opinion about the ability and willingness of an issuer, such as a corporation or state or city government, to meet its financial obligations in full and on time.” (Standard & Poor’s, 2011)
To understand the importance of credit rating agencies a brief summation of regulatory events is required. White (2010) describes that in 1936 it was decided that banks were only allowed to hold ‘investment grade’ securities. Only securities with a sufficiently high credit rating by one of the main rating agencies were considered ‘investment grade’. In the years following this decision up until the 1970s capital requirement restrictions were linked to credit ratings for insurance companies, pension funds and broker dealers (White, 2010). In the 1990s the ‘Securities and Exchange Commission’ (SEC) decided that only credit ratings by NRSRO’s would be recognized. In addition to this the credit rating agencies have altered their business model, switching form ‘investor pay’ to ‘issuer pay’ (White, 2010). The investor no longer paid for a credit rating on firms but the firm itself had to pay the credit rating agency for its services.
2.2 Credit rating quality and criticism
The quality of the credit ratings has been an ongoing research topic for years. Blume et al. (1998) summarize the research into credit ratings up to 1998 and find research into quality ratings dating back to 1958 (Hickman, 1958). Cheng & Neamtiu (2009) define three aspects of credit rating quality: accuracy, timeliness and volatility. Accuracy refers to how accurate a credit rating predicts defaults, timeliness measures if a default was predicted well ahead, and volatility measures how often a credit rating is updated. High quality is a combination of high accuracy and timeliness and low volatility.
In 2002 credit rating agencies came under scrutiny because none were able to correctly assess the creditworthiness of Enron days before its bankruptcy (Alp, 2013; Hill, 2002; White, 2010). Criticism on credit rating agencies further increased during the 2008 global financial crisis. According to the financial crisis report, collateralized debt obligations (CDOs) containing mortgage backed securities were often rated triple A by credit rating agencies (Financial Crisis Inquiry Commission, 2011). A main concern is that credit rating agencies assign inflated ratings (Efing and Hau, 2015; Stolper, 2009). Inflated ratings are ratings that underestimate the default risk, which could be categorized as low accuracy.
theory, the ‘issuer pay’ model can incentivize agencies to give favourable ratings since payment is only received for published ratings (Mählmann, 2011). Another implication of this business model is that issuers might theoretically start shopping for ratings and publish the highest (Skreta and Veldkamp, 2009), but no empirical evidence of this has been found (Morkoetter et al., 2017). These conflicts of interest might result in inflated credit ratings. It is argued that competition might influence the amount of ‘conflict of interest’. There are competing views on how competition influences the degree of ‘conflict of interest’ and as such the quality of credit ratings (Bae et al., 2015). Xia (2014) finds that the introduction of a new ‘investor paid’ rating agency led to an increase in quality in S&P’s ratings. This suggests that rating quality increases when competition intensifies. Interestingly, S&P’s ratings were mainly adjusted when this new agency provided lower credit ratings, suggesting that reputational costs for inflated ratings are higher than for deflated ratings. Even more recent Morkoetter et al. (2017) have found empirical evidence that competition increases both the amount of information provided and the quality of this information. Opposing this are mainly theoretic results that increased competition facilitates issuers to shop for ratings incentivizing credit rating agencies to inflate ratings in order to increase their market share (Bolton et al., 2012). The effect of competition on this conflict of interest is a particularly interesting topic to regulators as there currently are potential barriers to entry (White, 2010).
An additional argument for inflated credit ratings is their legislative value. As mentioned in section 2.1, certain institutions had their capital requirements tied to credit ratings (Partnoy, 1999; Stolper, 2009). As such these institutions strongly depend on credit ratings in their investing behaviour, preferring higher credit ratings. Due to these regulations, a favourable credit rating is crucial to attracting investors. Partnoy (1999) argues that credit rating agencies sell “regulatory licenses” by giving bond issuers ratings on which regulators rely. Although the credit ratings have this regulatory value, credit rating agencies are barely liable for the credit ratings they issue. Credit rating agencies claim to publish opinions and are hence protected against legal challenges (Partnoy, 1999; White, 2010).
states that the high-end bankruptcies allowed credit rating agencies to refine their models (Kolasinski, 2009).
In 2010, as a response to the criticism on credit rating agencies, the ‘Dodd-Frank Wall Street Reform and Consumer Protection Act’ was enacted. This act, commonly referred to as ‘Dodd-Frank’, makes credit rating agencies increasingly vulnerable to claims of inaccurate ratings (Dimitrov et al., 2015). Furthermore, the act aims to undo the regulatory reliance on credit ratings and to improve the governance and control of credit rating agencies (Dimitrov et al., 2015). Dimitrov et al. (2015), have looked into the effect of this act on credit ratings for specific debt instruments. They find that credit rating agencies issue lower ratings but do not find evidence of improvement in accuracy. This thesis covers years following the enactment of Dodd-Frank. As such, the research will assess the effects of Dodd-Frank on issue level credit ratings.
2.3 Reliance on quantitative information
A difficulty in assessing the performance of credit rating agencies is that no uniform way to measure credit risk exists. By 1996, S&P claimed to use no fixed rules and use high amounts of subjectivity (Blume et al., 1998). There is evidence of ‘subjectivity’ in CDO credit ratings prior to the global financial crisis (Griffin and Tang, 2012) as simulation results were only partly able to explain credit ratings. Subjectivity in this sense, refers to looking beyond financial models. Although the credit rating agencies have access to private information, possess experience and, apply subjectivity in their ratings, there are certain measures that give an indication of credit risk. It is described by Estrella et al. (2000) that there are several alternative sources of default risk. Two of these complementary sources are accounting-based and market-accounting-based scores (Agarwal and Taffler, 2008; Estrella, 2000). Whereas these scores were initially constructed to predict bankruptcies (Altman, 1968; Hillegeist et al., 2004; Ohlson, 1980), they are often used to get an indication of default risk (Bharath and Shumway, 2004; Cheng and Neamtiu, 2009).
bankruptcy in a period up to two years ahead of a bankruptcy (Altman, 2000) by 2000. The O-score by Ohlson (1980) is an alternative measure using accounting data and can in contrast to the Z-score be constructed without any market information but requires data from previous periods. A detailed description of the Z-score and O-score will be provided in the methodology.
Alp (2013) and Blume et al. (1998) used accounting information to discover trends in credit rating standards, using accounting information as an indicator of default risk. The Z-score and O-score are often used accounting measures of default risk (Altman, 2000; Cheng and Neamtiu, 2009) that have shown to be accurate (Altman, 2000). Although it is known that credit ratings use a wide range of inputs, it could logically be expected that accurate ratings would include accounting ratios that proved successful. As such, the first hypothesis will be:
H1: “Default probabilities implied by accounting ratios have a negative effect on credit ratings.”
A different approach to estimating default risk in the form of probability of bankruptcy is by using a market based model. These market based models are inspired by the Merton (1974) and Black-Scholes (1973) models. Two popular market based models are the KMV-Merton model (Bharath and Shumway, 2004) and the BSM-Prob model (Hillegeist et al., 2004). In line with Merton, these similar models regard equity as a call option on a firms value with a strike price equal to the book value of its debt (Bharath and Shumway, 2004). These market models estimate the probability that the face value of debt becomes larger than the market value of the firm (Bharath and Shumway, 2004).
The default estimates using these models are more difficult to calculate. The KMV-Merton model requires the market value of equity, its volatility and a risk-free rate which can be observed. The model requires that the market value of the firm and its volatility are determined using iterative processes.1 A ‘naïve’ alternative to the KMV-Merton model that can be determined without iterations is able to give accurate default forecasts (Bharath and Shumway, 2004) as well. This model is called the naïve market model because instead of simultaneously solving some equations, rough simplifications are such that a default probability can be determined linearly. These simplifications include using the book value of
debt instead of estimating the market value of debt and making the volatility of debt linearly dependant on the volatility of equity.
Hillegeist et al. claims that BSM-Prob provides higher information than the conventional accounting measures based approach (Hillegeist et al., 2004). This is supported by Agarwal and Taffler who find that the market model does not have superior predictive power but does add unique information (Agarwal and Taffler, 2008). Since these market models estimate default probability it could be expected that accurate credit ratings include information from empirically verified market models. As such the second hypothesis is constructed as follows:
H2: “Default probabilities implied by market models have a negative effect on credit ratings”
It should be noted that accounting- and market-based scores are not the only predictors of credit risk as there is an ongoing research in this direction. Other alternatives include a model covering market and accounting data (Campbell et al., 2011) and models using artificial intelligence (Huang et al., 2004). These models will not be used since they are not used as often as the chosen models and are not as common in the field of finance respectively.
2.4 Investment grade rating
Since an investment grade rating is a legal requirement for investments of certain investors, receiving an investment grade rating is important to firms as it potentially attracts new investors. As such the difference between the lowest investment grade rating (BBB-) and the highest speculative grade rating (BB+) will have bigger implications than any other two adjacent credit rating categories. This might enhance the adverse incentives for credit rating agencies and stimulate rating shopping by issuers, both leading to an inflation of credit ratings.
different standards for investment and speculative grade issuers. Evidence of this has been found until 2002 by Alp (2013) and Jorion et al. (2009). Whereas Alp made a distinction between investment and speculative grade, the implications of being just above the boundary was not discussed.
Johnson (2004), did address issuers just above the speculative grade separately. He finds evidence that a downgrade from the lowest investment grade to a speculative grade is often followed by further downgrades. This might be caused by the loss of contracts due to the legislative advantage of investment grade but Johnson (2004) also poses a different explanation. This alternative explanation is that rating agencies tolerate high levels of default risk in their definition of investment grade and hence a downgrade justifies a downgrade of multiple notches. Whereas Johnson (2004) does look at the implications of a downgrade from investment to speculative grade, this study only studies the momentum and does not justify these ratings using creditworthiness estimators or actual default rates.
There is a gap in literature regarding the potential inflation of credit ratings just above the boundary of investment and speculative grade.
Due to the additional legislative benefits (Partnoy, 1999; White, 2010) of receiving an investment grade rating, some firms in the lowest investment rating category (BBB-) could be expected to have successfully used all means necessary to receive an inflated credit rating. These means include rating shopping (Skreta and Veldkamp, 2009) and abusing relationship benefits (Mählmann, 2011) to prevent ending up in the highest speculative grade category (BB+) instead. Although there are theoretical indications that rating inflation might occur, the Dodd-Frank act aims to diminish the legislative advantage of an investment grade rating (Skeel, 2010). Therefore, the following hypothesis is formulated.
H3: “Credit ratings for issuers just above the boundary of investment grade are inflated.”
3 Methodology
This section covers the methods that are used to perform research. This thesis compares issuer level credit ratings to default risk proxies. This will be done by estimating models by using the following baseline equation.
𝑦𝑖𝑡 = 𝛼 + 𝛽1𝑥𝑖𝑡+ 𝛽2𝑧𝑖𝑡+ 𝑢𝑖𝑡 (1)
In this thesis pooled OLS regressions is applied in line with Baghai et al. (2014) and Xia (2014) to estimate this equation. The advantage of using OLS is that it allows the use of firm-fixed effects (Baghai et al., 2014). The errors are estimated as robust standard errors such that the results are robust against heteroskedasticity, as is done in prior literature (Hillegeist et al., 2004; Kisgen, 2006).
In the baseline equation, the dependent variable 𝑦𝑖𝑡 will be the credit rating for all models where 𝑖 is the identifier for the firm and 𝑡 for the year of observation. The independent variable 𝑥𝑖𝑡 will be one or multiple default risk proxies. Finally, the control variables are
contained in the 𝑧𝑖𝑡 term, the error term is given by 𝑢𝑖𝑡 and 𝛼 is a constant.
Besides OLS, ordered logit models will be included in this thesis as it allows a deeper analysis into the differences between rating categories as will be further explained in section 5.2. The baseline equation of the ordered logit model will be as follows:
𝑦𝑖𝑡∗ = 𝛽1𝑥𝑖𝑡+ 𝛽2𝑧𝑖𝑡+ 𝑢𝑖𝑡 (2)
The baseline equation contains a latent variable 𝑦𝑖𝑡∗ which cannot be observed and linearly depends on the explanatory variables. In the ordered logit model with 𝑘 possible values for the dependent variable, this latent variable is translated to the dependent variable 𝑦𝑖𝑡. In this
model the value of the dependent variable is determined by the latent variable as follows.
𝑦𝑖𝑡 = { 1 if 𝑦𝑖𝑡∗ ≤ 𝛾2 2 if 𝛾2 < 𝑦𝑖𝑡∗ ≤ 𝛾3 ⋮ ⋮ 𝑘 − 1 if 𝛾𝑘−1 < 𝑦𝑖𝑡∗ ≤ 𝛾𝑘 𝑘 if 𝛾𝑘 < 𝑦𝑖𝑡∗ (3)
The dependent variable 𝑦𝑖𝑡 is the issuer-level credit ratings assigned by S&P. In line with prior
research, the long-term issuer level credit rating is used as the dependent variable. These credit ratings are converted to a numerical score as is shown in appendix Appendix A to allow further analysis. The ratings are converted in line with Morkoetter et al. (2017) and Cantor & Packer (1997) such that the numerical values increase with the perceived default risk. The explicit conversion of actual rating to numerical rating is found in appendix Appendix A. One should keep in mind that an increase in this numerical rating corresponds to a decrease in actual credit rating.
The independent variable is one of the following default risk proxies: Z-Score, O-score, and naïve distance to default. To accurately compare credit ratings to other estimators, all creditworthiness judgements must cover creditworthiness over the same period. For the accounting-based creditworthiness measures, the value of the coefficients depends on the period in which defaults are predicted. In this thesis, I chose to use coefficients linked to default prediction in the timeframe of one year. This is the most often applied form of these models and links closely to the use of annual data. Re-estimation of model coefficients was not possible due to the lack of default data. As such the original coefficients (Altman, 1968) are used and the Z-score is composed as follows:
𝑍 = 0.012 𝑋1+ 0.014𝑋2+ 0.033𝑋3+ 0.006𝑋4+ 0.999𝑋5 (4) The variables used in this equation are: working capital/total assets (𝑋1), retained earnings/total assets (𝑋2), EBIT/total assets (𝑋3), market value equity/book value total
liabilities (𝑋4), and sales/total assets (𝑋5).
The second default risk proxy is the O-score by Altman (1968) which is also based on accounting ratios. The coefficients of the O-score are often re-estimated using actual default rates (Hillegeist et al., 2004; Joy Begley et al., 1996). In the absence of actual default data, the original coefficients (Ohlson, 1980) will be used as these are still widely applied (Hillegeist et al., 2004; Joy Begley et al., 1996). The original O-score is determined as follows:
𝑂 = −1.32 − 0.407 𝑆𝐼𝑍𝐸 + 6.03 𝑇𝐿𝑇𝐴 − 1.43 𝑊𝐶𝑇𝐴 + 0.0757 𝐶𝐿𝐶𝐴 − 1.72 𝑂𝐸𝑁𝐸𝐺
−2.37 𝑁𝐼𝑇𝐴 − 1.83 𝐹𝑈𝑇𝐿 + 0.285 𝐼𝑁𝑇𝑊𝑂 − 0.521𝐶𝐻𝐼𝑁 (5)
liabilities/current assets (CLCA), a binary that is one if total liabilities is larger than total assets (OENEG), net income/total assets (NITA), funds from operations/total liabilities (FUTL), a binary that is one if net income was negative for the last two years (INTWO), and a measure of change in net income (CHIN).
Several control variables 𝑧𝑖𝑡 are used only when the independent variable 𝑥𝑖𝑡 is the Z-score
or the O-score. Many papers use the accounting ratios used by Blume et al. (1998) as independent variables when a credit rating is the dependent variable (Alp, 2013; Dimitrov et al., 2015). The accounting based information such as firm size, is already included in the accounting based default proxies however. Besides accounting ratios, Blume et al. (1998) use measures of idiosyncratic and systematic risk when assessing the credit rating standard. These are not included in any of the accounting-based ratios and will be included as control variables. Idiosyncratic risk and systematic risk will be determined similar to as is done in Alp (2013) and Dimitrov et al. (2015). A detailed description is provided in appendix B.3.
The final default risk proxy that is used as explanatory variable 𝑥𝑖𝑡 the naïve distance to
default estimated by a market model. The KMV-Merton model provides a way to estimate defaults by using market information (Bharath and Shumway, 2004). The original KMV-Merton model requires iteratively solving equations. In this thesis, an alternative of the same functional form using the same basic inputs but without iterative procedu res will be used. This naïve alternative has been shown to give accurate default predictions (Bharath and Shumway, 2004). In this model the naïve distance to default (𝐷𝐷) is calculated using the following equation.
𝐷𝐷 =ln((𝐸 + 𝐹) /𝐹) + (𝑟𝑖𝑡−1 − 0.5𝜎𝑉
2)𝑇
𝜎𝑉√𝑇
(6) Subsequently, this number can be transformed into a default probability. This is done by using the cumulative normal distribution. The definition of the naïve default probability is given by the following equation.
𝜋naïve = 𝒩 (−ln((𝐸 + 𝐹) /𝐹) + (𝑟𝑖𝑡−1− 0.5𝜎𝑉
2)𝑇
𝜎𝑉√𝑇 ) (7)
probability is mainly used for comparison with actual defaults (Hillegeist et al., 2004). Furthermore, using relative scores instead of estimated default probabilities is of the same form as using accounting based scores. The variables contained in the naïve market model are the market value of equity (E), face value of debt (F), return over the previous year (𝑟𝑖𝑡−1),
the volatility of the firm value (𝜎𝑉), and the considered time horizon (T). How these variables
will be estimated is covered in section 4.
When the independent variable 𝑥𝑖𝑡 is the distance to default certain specific control variables
are used to complement this information. Xia (2014) specified these control variables that provide additional information to a market based default risk proxy. These control variables are: leverage, profitability, M/B ratio, tangibility, sales, leverage volatility, and profitability volatility. The construction of these variables is presented in appendix B.2.
Besides control variables linked to the independent variable, some control variables 𝑧𝑖𝑡 will
4 Data
4.1 VariablesThis thesis requires a large panel data set with company and year as dimensions. To construct this, multiple sources are consulted. The obtained data will be processed and merged such that statistical analysis can be carried out.
Credit rating data is retrieved from Compustat North America which is accessed through WRDS. The dataset contains three different ratings from S&P: ‘S&P Domestic Long Term Issuer Credit Rating’, ‘S&P Domestic Long Term Issuer Credit Rating’, and ‘S&P Quality Ranking – Current’. These data are provided on an annual basis. The quality ranking encompasses more than credit risk and is hence not included in this analysis. The issuer credit ratings represent the general creditworthiness of the issuing firm and contrasts with bond credit ratings as it is not limited to a specific debt instrument. The long-term issuer rating deals with general creditworthiness for periods over 1 year whereas the short-term issuer rating deals with general creditworthiness within a year. In line with prior literature (Alp, 2013; Baghai et al., 2014; Xia, 2014), only the long-term issuer level credit ratings will be used.
Furthermore Compustat Capital IQ - North America provides access to annual financial data through the ‘Fundamentals annually’ database. To prevent double entries only the industrial format is included. Furthermore, since the Dodd-Frank act was passed in the USA, only entries from the USA are used.
These financial data are used for calculating the Z-scores and O-scores as described in section 3. The fourth financial ratio in the Z-score contains information about the market value of equity. This data is constructed by using CRSP monthly stock data. The market equity is determined as the product of the stock price and number of shares outstanding. For stocks whose closing price is unknown, the stock price is set to the bid ask average of that trading day.
The data for this is collected from the world bank website.2 Furthermore, the O-score requires information about the funds from operations. The funds from operations information is absent for most company year observations and is constructed in these sit uations as is described in appendix B.1.
The market models require additional information. The advantage of using the naïve KMV-Merton model is that al variables can be determined without iterative procedures. In line with Vassalou & Xing (2004) and Bharath & Shumway (2004), the face value of debt is constructed as debt in current liabilities plus 0.5 times the long term debt. The return over the previous year is determined as the sum of the natural logarithm returns of the twelve months before the fiscal year end month. The volatility of equity is estimated as the annualized historical standard deviation over these monthly returns.
Whereas Alp (2013) and Dimitrov et al. (2015) relied on daily stock returns to approximate systematic and idiosyncratic risk, this thesis uses monthly stock returns. The reason for doing so is that the daily stock file only provided data on only 748 firms whereas the monthly stock file contains information on 11,550 firms.
Due to the limited number of observations for some industries when using industry dummies for every unique two digit SIC code, larger industry groups are used. Jorion et al. (2009) showed that using even larger categories does not influence the results. The categories that are used can be found in appendix B.4. A summary of all gathered data is provided in Table 1: Data Requirements. The first column describes what information is retrieved. The purpose of this information is stated in the second column. The third column contains the source through which the data was accessed and in the final column, a specification is given about which dataset was consulted.
2 GDP deflator is constructed using data from the World Bank:
Table 1: Data Requirements
Data: Application: Source: Dataset:
Issuer Credit Ratings Credit rating Compustat Capital IQ Ratings Company financials Z-score, O-score Compustat Capital IQ Fundamentals
Annually Additional financials Control variables Compustat Capital IQ Fundamentals
Monthly Market data Z-score, naïve distance
to default
Compustat – CRSP Monthly Stock File Monthly stock returns Control variables Compustat – CRSP Monthly Stock File
GDP deflator O-score The World Bank -
4.2 Data description
All gathered data is merged using a combination of the ticker code to represent the company and the date as unique identifiers. All data are acquired at the end of the fiscal year of the concerning firm such that there is no time discrepancy within data of a firm in a certain year. Using annual data from the fiscal year end is in line with previous work (Agarwal and Taffler, 2008; Hillegeist et al., 2004). The use of quarterly data is not feasible since the quarterly data do not include income statement items. The EBIT is an income statement item that is needed for constructing a Z-score.
Table 2: Observations per rating, industry and year Observations
This table contains the number of observations specified per rating category, industry category and year respectively.
Observations per rating category Observations per industry
AAA 12 Industry 1 0 AA+ 4 Industry 2 204 AA 32 Industry 3 31 AA- 50 Industry 4 1147 A+ 73 Industry 5 501 A 125 Industry 6 60 A- 158 Industry 7 180 BBB+ 282 Industry 8 114 BBB 367 Industry 9 354 BBB- 292 Industry 10 14
BB+ 216 Observations per year
BB 257 BB- 267 Year 2011 520 B+ 199 B 174 Year 2012 509 B- 74 CCC+ 13 Year 2013 514 CCC 7 CCC- 2 Year 2014 532 CC 0 SD (selective default) 1 Year 2015 530 D (default) 0
Table 2 displays the lack of observations in the first or ‘Agriculture, Forestry & Fishing’ industry. These industries will not be redefined despite the irregular distribution of observations to maintain consistency with prior literature (Jorion et al., 2009; Xia, 2014) and theoretic relevance. The observations per rating category have a shape similar to a normal distribution. The outer categories contain a lower number of observations than the inner categories. Furthermore, all fiscal years have a similar and high number of observations.
4.3 Descriptive Statistics
The dependent variable, the credit rating does not contain outliers as it ranges from 1 to 22 by definition and hence, is not winsorized. For most of the explanatory variables, winsorizing strongly decreases the maximum and strongly increases the minimum.
Table 3: Summary Statistics
Long-term credit rating
Short-term
credit rating Z-score O-score
Default probability (Naïve) Distance to default (Naïve) Systematic risk Idiosyncratic
risk Leverage Profitability
5 Results
5.1 Regression results
The first results of this thesis contain the OLS regression results for several models in which the long-term credit rating is regressed against explanatory variables. Depending on the default risk estimator, the control variables will be as specified in section 3. To prevent perfect collinearity some dummies will not be included. More specifically, the year dummy for the first year observation and the industry dummies for the first two industry categories are excluded. The reason for removing two industry dummies is that no observations match with the ‘Agriculture, Forestry & Fishing’ industry. By choosing to remove these industries, dummy coefficients will be relative to the second industry group and the fiscal year 2011. The most recent year observations are from 2015 as the O-score could not be constructed in later years due to the absence of the GDP deflator data.
In line with Baghai et al. (2014) several models will be estimated. A group of six models is estimated for several combinations of independent variable and control variables. Table 4 contains the regression model for the long-term credit rating with the Z-score while Table 5 and Table 6 contain the regression model with the long-term credit rating with the O-score and distance to default respectively. For every set of six models the first model will contain the dependent variable, the independent variable and the year- and industry dummies. Subsequently, alternative models using additional control variables and using firm fixed effects instead of industry fixed effects will be estimated. Each model will contain either firm fixed or firm fixed effects as these effects are perfectly collinear to each other. Additionally, some models will include a dummy variable for speculative grade firms. 3
Table 4
Long-term credit rating estimations using Z-score
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the Z-score. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5 and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 use the market beta and standard error retrieved from estimating the market beta as proxies for systematic and idiosyncratic risk respectively. The significance is shown by asteri sks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 11.109 *** (0.277) 14.733 *** (0.477) 8.445 *** (0.144) 10.002 *** (0.289) 9.641 *** (0.113) 9.684 *** (0.169) Z-score 0.415 *** (0.122) 0.322 *** (0.108) 0.178 *** (0.067) 0.157 ** (0.065) 0.023 (0.09) 0.03 (0.091) Systematic risk -7.261 *** (0.407) -2.455 *** (0.262) -0.241 ** (0.098) Idiosyncratic risk -40.41 *** (9.796) -19.326 *** (5.986) 1.107 (2.737) Speculative dummy 5.196 *** (0.068) 4.952 *** (0.068) 1.577 *** (0.113) 1.566 *** (0.114) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.032 0.165 0.691 0.705 0.967 0.967 Table 5
Long-term credit rating estimations using O-score
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the O-score. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5 and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 use the market beta and standard error retrieved from estimating the market beta as proxies for systematic and idiosyncratic risk respectively. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Table 6
Long-term credit rating estimations using distance to default
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the naïve distance to default. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5 and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 include control variables based on accounting information. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 12.259 *** (0.234) 20.362 *** (0.311) 9.208 *** (0.137) 14.5 *** (0.277) 9.813 *** (0.07) 14.711 *** (0.704) Distance to default -0.266 *** (0.009) -0.137 *** (0.009) -0.102 *** (0.007) -0.074 *** (0.007) -0.021 *** (0.004) -0.007 ** (0.003) Leverage 6.497 *** (0.514) 3.478 *** (0.381) 3.79 *** (0.543) Profitability -8.484 *** (2.534) -7.641 *** (1.672) -1.008 (1.257) Market to book ratio -0.635 *** (0.076) -0.289 *** (0.054) -0.34 *** (0.045) Tangibility -0.442 *** (0.039) -0.287 *** (0.029) -0.517 *** (0.08) Sales -1.241 *** (0.035) -0.681 *** (0.031) -0.713 *** (0.1) Leverage volatility -3.113 (2.833) 0.002 (2.074) 0.636 (1.101) Profitability volatility 23.565 *** (4.499) 13.066 *** (3.233) 1.427 (2.87) Speculative dummy 4.614 *** (0.071) 3.475 *** (0.072) 1.56 *** (0.113) 1.359 *** (0.104) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.309 0.622 0.724 0.795 0.968 0.972
The other accounting score, the O-score, gives somewhat different results in Table 5. In contrast to the Z-score, the O-score is expected to decrease with lower perceived financial distress and as such a positive coefficient is expected when this score is used to estimate the credit rating. OLS regression coefficients are not only positive but also statistically significant at the 1% level for all eight models involving the O-score. Based on these observations it can be stated that the O-score does influence the credit rating and in contrast to the Z-score the coefficients have the expected sign.
The final group of models that is estimated, is the group using the naïve distance to defa ult as an indicator for default risk. Since this naïve distance to default is based on market data instead of accounting data, these models will include different control variables. As would be expected, the sign on the coefficient of the distance to default is negative since a higher distance to default would lead to a higher numerical credit rating (lower credit rating). Whereas including firm-fixed effects made the coefficient on the Z-score and O-score insignificant, this is not the case for the naïve distance to default. Significance at the 5% level is even found for model 6 which contains the highest number of control variables and firm-fixed effects. The models with firm-firm-fixed effects do however give suspiciously high R-squared values which makes that results from model 4 are preferred for interpretation. It can be concluded that that the distance to default implied by the naïve market model has an influence on the credit rating and that its sign is in line with expectations.
Finally, the speculative dummy is statistically significant for all models that include this dummy in Table 4, Table 5 and Table 6. However, the size of the coefficient strongly depends on the inclusion of firm or industry effects. When firm fixed effects are applied, the coefficient for the speculative dummy is significantly smaller in size. Using these firm fixed effects leads to extremely high R-squared values however. This suggests an overfitted model as it is generally known that credit ratings do not solely rely on the chosen explanatory variables. Consequently, model 4 is expected to give the most reliable results.
The specific focus of this thesis is on the ratings on firms that are rated just above the boundary of the speculative grade. From the results in Table 4, Table 5, and Table 6 it can be seen that the speculative grade dummy is significant at the 1% for all 24 estimated models. It is unclear however if the ratings that are on the boundary of investment grade are inflated. If ratings were inflated a statistically significant positive coefficient would be expected on a dummy that has a value of 1 for firms on this boundary and 0 otherwise. As such the Model 4 for the Z-score, O-score and distance to default of are re-estimated with the inclusion of a dummy that checks if the rating is on the investment grade boundary. Model 4 is chosen because of the high number of control variables and the use of industry fixed effects instead of entity fixed effects.
When interpreting Table 7, one should recall that an increase in the numerical credit rating value corresponds to a decrease in credit rating as can be seen in Appendix A. Positive and significant coefficients are acquired for the boundary dummy, which implies that firms with a BBB- rating get assigned lower credit ratings (equivalent of a higher numerical rating) than other investment grade firms, having controlled for other variables. The higher coefficient on the speculative dummy indicates that firms with a speculative grade rating however are rated even lower on average than those on the boundary of the investment grade.
Table 7
Investment boundary regressions
These models are the same models as Model 4 of the previous regressions. In these models however, a dummy is added that equals 1 for firms whose rating equals BBB-, which is the lowest investment grade rating. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Z-score Model O-score Model
Distance to default Model Constant 9.004 *** (0.262) 9.384 *** (0.251) 13.112 *** (0.264) Z-score 0.111 * (0.06) O-score 0.27 *** (0.021)
Naïve distance to default -0.062 ***
Market to book ratio -0.257 *** (0.05) Tangibility -0.243 *** (0.026) Sales -0.579 *** (0.029) Leverage volatility 0.015 (1.941) Profitability volatility 13.607 *** (3.123) Boundary dummy 2.59 *** (0.062) 2.488 *** (0.063) 2 *** (0.061) Speculative dummy 5.561 *** (0.07) 5.329 *** (0.067) 4.135 *** (0.074) Year dummies Y Y Y Industry dummies Y Y Y Firm dummies N N N Observations 2605 2605 2605 Adjusted R-squared 0.762 0.780 0.828
These results imply that credit rating standards become stricter in lower rating categories as the boundary dummy and speculative dummy have positive coefficients of which the speculative dummy has the largest magnitude. The boundary dummy indicates that firms on the boundary are rated stricter than other investment grade firms. Appendix D.1 provides additional regression results from models without the speculative dummy, such that the credit rating standard on firms on the boundary are compared to all other firms (investment and speculative grade). Caution is required when inferences are made from Table 7 as the boundary and speculative dummy are constructed by using the dependent variable, long-term issuer level credit ratings, which is discussed further in appendix D.2.
5.2 Further analysis
Table 8
Long-term credit rating estimations using multiple default risk proxies
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the Z-score. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5, and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 use the additional control variables for the default risk proxies. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 12.306 *** (0.233) 20.771 *** (0.391) 9.343 *** (0.137) 15.175 *** (0.344) 9.61 *** (0.105) 14.777 *** (0.72) Z-score 0.338 *** (0.094) 1.011 *** (0.07) 0.167 *** (0.06) 0.586 *** (0.059) 0.116 (0.094) 0.093 (0.103) O-score 0.43 *** (0.038) 0.147 *** (0.033) 0.25 *** (0.025) 0.155 *** (0.024) 0.13 *** (0.019) 0.026 (0.027) Distance to default -0.234 *** (0.009) -0.115 *** (0.01) -0.089 *** (0.007) -0.066 *** (0.007) -0.015 *** (0.003) -0.006 (0.004) Systematic risk -1.927 *** (0.289) -0.999 *** (0.216) -0.176 * (0.093) Idiosyncratic risk -8.911 (6.389) -8.294 * (4.79) -1.001 (2.609) Leverage 4.835 *** (0.632) 1.879 *** (0.438) 3.536 *** (0.594) Profitability -8.369 *** (2.303) -6.446 *** (1.596) -0.808 (1.326)
Market to book ratio -0.583 *** (0.071) -0.262 *** (0.053) -0.336 *** (0.046) Tangibility -0.195 *** (0.042) -0.126 *** (0.031) -0.499 *** (0.083) Sales -1.265 *** (0.035) -0.721 *** (0.031) -0.719 *** (0.099) Leverage volatility -0.809 (2.704) 1.174 (2.035) 0.648 (1.121) Profitability volatility 14.116 *** (4.1) 8.02 *** (3.07) 1.053 (2.874) Speculative dummy 4.474 *** (0.067) 3.313 *** (0.073) 1.562 *** (0.092) 1.359 *** (0.104) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.359 0.658 0.740 0.808 0.968 0.972
Furthermore, the results from 5.1 indicate that the rating standard becomes more stringent on lower rated firms. There are reasons to question the results however, since the speculative and boundary dummy directly depend on the value of the dependent variable. To provide additional information about which credit ratings are assigned and more specifically on the rating standard differences between credit rating groups, a different regression technique is applied, namely ordered logit. This model is chosen since it does not assume the distance between rating categories to be equal and it is used in prior literature to estimate credit scores (Baghai et al., 2014; Dimitrov et al., 2015). Moreover, ordered logit provides means to analyse rating categories without causality issues as was the case with the dependent variable based dummies.
The estimated ordered logit models will use the same variables as model 2 from Table 4, Table 5, and Table 6. This second model is chosen since the use of firm-fixed effects would induce the incidental parameter problem (Baghai et al., 2014) and the speculative dummy should be excluded. This model 2 is estimated for the Z-score, the O-score, the distance to default and the complete model using ordered logit. Its estimates can be found in Table 9 which contains coefficient and limit point estimates. In this ordered logit model, only the sign of the coefficients has a direct economic interpretation. The limit points however provide new insights into the cut-off points of all rating categories, not only the cut-off between speculative and investment grade. In Table 9, the lower rating categories (14 and higher) have been grouped together as without grouping, insignificant coefficients would be obtained for these categories due to the low number of observations. 45
4 The same model has been estimated without grouping rating categories together as well. This led to similar
results.
5 This is in line with Alp (2013), who pooled some of the lower rating categories with few observations
Table 9
Logit model estimations
This table contains the results of estimating ordered logit models using the Z-score, O-score, distance to default, and a combination of those respectively. This table does not only provide coefficients but also provides limit point estimates for each model. Limit 2 is the implied threshold value of the boundary between a credit rating of AAA and AA+, Limit 3 is the implied threshold value of the boundary between a credit rating of AA+ and AA etcetera. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Z-score O-score Distance to default Complete
Variables Z-score 0.195 *** (0.067) 1.012 *** (0.081) O-score 0.388 *** (0.023) 0.139 *** (0.03)
Naïve distance to default -0.118 ***
(0.008) -0.1 *** (0.008) Systematic risk -4.469 *** (0.242) -4.343 *** (0.244) -1.824 *** (0.259) Idiosyncratic risk -20.181 *** (5.903) -21.452 *** (5.958) -8.669 (6.26) Leverage 6.401 *** (0.499) 5.216 *** (0.619) Profitability -6.465 *** (2.158) -6.68 *** (2.231)
Market to book ratio -0.648 ***
Limit 8: A-/BBB+ -4.27 *** (0.301) -4.776 *** (0.303) -12.726 *** (0.371) -13.807 *** (0.46) Limit 9: BBB+/BBB -3.586 *** (0.298) -4.06 *** (0.3) -11.681 *** (0.36) -12.713 *** (0.45) Limit 10: BBB/BBB- -2.887 *** (0.296) -3.324 *** (0.297) -10.597 *** (0.35) -11.562 *** (0.439) Limit 11: BBB-/BB+ -2.377 *** (0.294) -2.783 *** (0.295) -9.778 *** (0.342) -10.691 *** (0.432) Limit 12: BB+/BB -1.992 *** (0.293) -2.371 *** (0.294) -9.133 *** (0.337) -10.008 *** (0.427) Limit 13: BB/BB- -1.489 *** (0.293) -1.824 *** (0.293) -8.293 *** (0.33) -9.111 *** (0.421) Limit 14: BB-/B+ -0.839 *** (0.293) -1.117 *** (0.293) -7.243 *** (0.323) -7.989 *** (0.415)
Figure 1
The plots show that on first sight the distances between the limit points seem evenly distributed. The limits enclosing the boundary of the investment grade (BBB-) are limit 10 and limit 11. The distance between two adjacent limit points gives an indication of the size of the rating category which can be used to infer information about the strictness of the rating standard for certain rating categories. Further details about the implied rating category size and the credit rating standard are provided in appendix D.3.
If there would be inflation for firms in this category, the distance from limit 10 to limit 11 would be larger than other distances. In that scenario, some firms that would receive a speculative rating when all distances between adjacent limits were equal, would now receive an investment grade rating. The distance between limit 10 and limit 11 however, is smaller than most other distances between adjacent limit points. This indicates that some firms that would be rated investment grade when all distances would be equal, are now rated speculative grade. This can be interpreted as a more stringent rating standard for firms in the lowest investment grade category.
-25.000 -20.000 -15.000 -10.000 -5.000 0.000 2 3 4 5 6 7 8 9 10 11 12 13 Limit Points
Ordered Logit: Limit Points
Table 10 Rating category size
This table contains the relative sizes of each rating category implied by the logit model. This is constructed by the absolute distance between two adjacent limit points. The rating category size could not be established for AAA rated firms since there is no upper boundary for the highest category by definition of the ordered logit model. Ratings up to and including BBB- are categorized as investment grade.
Z-score Model O-score Model Distance to default Model Complete Model AAA AA+ 0.292 0.295 0.343 0.355 AA 1.127 1.154 1.330 1.401 AA- 0.752 0.779 0.964 1.016 A+ 0.609 0.632 0.859 0.903 A 0.636 0.661 0.960 1.006 A- 0.536 0.556 0.835 0.866 BBB+ 0.684 0.716 1.045 1.094 BBB 0.699 0.736 1.084 1.151 BBB- 0.510 0.541 0.819 0.871 BB+ 0.385 0.412 0.645 0.683 BB 0.503 0.547 0.840 0.897 BB- 0.650 0.707 1.050 1.122
Mean Investment grade 0.649 0.674 0.915 0.963
Median Investment grade 0.636 0.661 0.960 1.006
ratings and the similar reputation hypothesis (Dimitrov et al., 2015; Morris, 2001) that firms will not assign inflated ratings when they are suspected to do so . This is also in line with empirical results that credit rating agencies have become more protective of their reputation after the enactment of Dodd-Frank (Dimitrov et al., 2015).
5.3 Robustness tests
The robustness of the results is tested by several robustness tests. First, it is verified that there is no multicollinearity by calculating variance inflation factors. These is calculated for the models from Table 7 and the complete model, which are the models that contain the most explanatory variables and are as such the most likely to give high variance inflation factors. The variance inflation vectors provide a measure of the goodness of fit of a given explanatory variable regressed by the other explanatory variables. The acquired variance inflation vectors are provided in appendix D.4. The variance inflation factors of the default proxies and specific control variables are all smaller than 5 which implies that there will be no multicollinearity. This test validates that it is appropriate to use the complete model.
Recall that this thesis uses default probability proxies using a combination of accounting ratios in the form of Z-score and O-score instead of separate accounting ratios as is done in Blume et al. (1998). It will be checked if using the accounting ratios from these scores separately gives new results. The motivation behind using the variables separately is that using a Z-score or O-score imposes a fixed ratio between the involved variables. By using these variables as explanatory variables themselves, these ratios between these variables are no longer fixed and a better fit can be created. To circumvent the effect of extreme outliers, all explanatory variables are winsorized at the 1st and 99th percentile.
Based on the observations from
Table 11
Long-term credit rating estimations using variables from the Z-score
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the variables of the Z-score. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5, and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 use the market beta and standard error retrieved from estimating the market beta as proxies for systematic and idiosyncratic risk respectively. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 12.033 *** (0.221) 13.903 *** (0.425) 9.269 *** (0.137) 10.334 *** (0.276) 10.101 *** (0.113) 10.183 *** (0.163) X1 (WC/TA) 4.118 *** (0.422) 3.593 *** (0.412) 0.969 *** (0.295) 0.827 *** (0.291) 0.505 (0.311) 0.496 (0.311) X2 (RE/TA) -3.825 *** (0.159) -3.451 *** (0.154) -1.789 *** (0.102) -1.693 *** (0.1) -1.495 *** (0.234) -1.488 *** (0.235) X3 (EBIT/TA) -6.883 *** (0.981) -6.054 *** (0.918) -3.41 *** (0.62) -3.193 *** (0.606) -1.929 *** (0.598) -1.89 *** (0.598) X4 (ME/TL) -0.297 *** (0.042) -0.27 *** (0.04) -0.074 ** (0.029) -0.068 ** (0.029) -0.128 *** (0.028) -0.127 *** (0.028) X5 (S/TA) 0.314 *** (0.103) 0.273 *** (0.095) 0.211 *** (0.064) 0.196 *** (0.063) 0.142 (0.097) 0.139 (0.098) Systematic risk -4.069 *** (0.356) -1.698 *** (0.244) -0.148 (0.094) Idiosyncratic risk -20.605 ** (8.632) -14.276 ** (5.67) -0.812 (2.615) Speculative dummy 4.367 *** (0.067) 4.248 *** (0.067) 1.533 *** (0.109) 1.529 *** (0.11) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.395 0.434 0.745 0.751 0.971 0.971
Table 12
Long-term credit rating estimations using variables from the O-score
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the variables of the O-score. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5, and 6 use entity fixed effects instead to approach effects on the firm level. Model 5, and 6 use the market beta and standard error retrieved from estimating the market beta as proxies for systematic and idiosyncratic ris k respectively. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the additi on of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 15.292 *** (0.364) 17.006 *** (0.425) 11.483 *** (0.27) 12.637 *** (0.333) 9.436 *** (0.659) 9.542 *** (0.673) SIZE -3.018 *** (0.088) -2.862 *** (0.087) -1.665 *** (0.077) -1.628 *** (0.075) -0.735 ** (0.287) -0.749 *** (0.289) TLTA 3.471 *** (0.284) 3.275 *** (0.274) 1.814 *** (0.205) 1.757 *** (0.202) 2.399 *** (0.353) 2.384 *** (0.354) WCTA 0.653 (0.52) 0.31 (0.509) 0.623 * (0.351) 0.446 (0.35) -0.016 (0.393) -0.024 (0.394) CLCA -0.209 (0.21) -0.254 (0.207) 0.274 ** (0.132) 0.241 * (0.132) -0.054 (0.121) -0.055 (0.122) OENEG -1.137 *** (0.231) -1.022 *** (0.224) -0.522 *** (0.167) -0.477 *** (0.165) -0.139 (0.139) -0.136 (0.138) NITA -16.366 *** (0.923) -15.017 *** (0.91) -8.211 *** (0.65) -7.797 *** (0.642) -2.438 *** (0.557) -2.393 *** (0.56) FUTL -0.087 (0.107) -0.104 (0.104) -0.221 *** (0.079) -0.224 *** (0.078) 0.428 ** (0.169) 0.417 ** (0.17) INTWO 0.748 *** (0.171) 0.656 *** (0.164) 0.482 *** (0.119) 0.445 *** (0.117) 0.244 *** (0.091) 0.242 *** (0.091) CHIN 0.944 *** (0.093) 0.883 *** (0.091) 0.485 *** (0.063) 0.467 *** (0.062) 0.152 *** (0.036) 0.151 *** (0.036) Systematic risk -3.137 *** (0.287) -1.582 *** (0.215) -0.168 * (0.094) Idiosyncratic risk -25.042 *** (6.298) -17.425 *** (4.71) -0.053 (2.538) Speculative dummy 3.608 *** (0.075) 3.506 *** (0.075) 1.418 *** (0.096) 1.412 *** (0.097) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.601 0.623 0.797 0.802 0.972 0.972
Table 13
Long-term credit rating estimations using the naïve default probability
This table contains the estimated OLS coefficients of models in which the long-term credit rating is estimated using the naïve default probability. All models include dummies for the observed fiscal years. Model 1, 2, 3, and 4 use industry dummies and model 5, and 6 use entity fixed effects instead to approach effects on the firm level. Model 2, 4, and 6 include control variables based on accounting information. The significance is shown by asterisks. The addition of *** to a coefficient indicates significance at the 0.01 level, the addition of ** indicates significance at the 0.05 level, and * denotes significance at the 0.10 level. Robust standard errors are reported in brackets below the coefficient.
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Constant 10.777 *** (0.258) 20.302 *** (0.324) 8.394 *** (0.136) 14.142 *** (0.29) 9.618 *** (0.07) 14.629 *** (0.707) Naïve default probability 4.297 *** (0.278) 1.419 *** (0.214) 1.613 *** (0.172) 0.911 *** (0.148) 0.453 *** (0.131) 0.266 ** (0.111) Leverage 9.159 *** (0.516) 4.589 *** (0.373) 3.847 *** (0.521) Profitability -8.941 *** (2.785) -7.766 *** (1.725) -1.034 (1.263) Market to book ratio -1.107 *** (0.075) -0.504 *** (0.053) -0.337 *** (0.044) Tangibility -0.475 *** (0.04) -0.295 *** (0.029) -0.523 *** (0.081) Sales -1.322 *** (0.035) -0.693 *** (0.032) -0.713 *** (0.1) Leverage volatility -0.437 (2.957) 1.518 (2.06) 0.699 (1.097) Profitability volatility 26.971 *** (4.771) 13.973 *** (3.244) 0.907 (2.834) Speculative dummy 5.08 *** (0.068) 3.658 *** (0.074) 1.595 *** (0.113) 1.375 *** (0.105) Year dummies Y Y Y Y Y Y Industry dummies Y Y Y Y N N Firm dummies N N N N Y Y Observations 2605 2605 2605 2605 2605 2605 Adjusted R-squared 0.093 0.585 0.700 0.785 0.967 0.972
6 Conclusion
This thesis complements the existing literature involving credit rating standards. (Alp, 2013; Bae et al., 2015; Baghai et al., 2014; Blume et al., 1998; Jorion et al., 2009). Whereas most papers use a wide range of explanatory variables, this thesis uses knowledge from the default and financial distress prediction literature (Altman, 2000, 1968; Bharath and Shumway, 2004; Campbell et al., 2008; Hillegeist et al., 2004; Ohlson, 1980).
Based on the first part of the results, it can be concluded that credit rating standards include information of the Z-score, O-score and market model. The market model and O-score had the coefficients that would be expected whereas the Z-score had the opposite sign for which no explanation could be found. That the Z-score is the one deviating from expectations is not completely surprising as the market model is more recent than the accounting scores and the O-score is more recent and based on a larger sample than the Z-score.
A further analysis into the credit rating standard on firms that are in the lowest region of investment grade ratings (BBB-), provides no indication that there is any inflation of credit ratings. Even more so, the results indicate that rating standards are more stringent around the investment grade. This contradicts the expectation that firms would successfully get higher ratings by abusing their relationship with rating agencies (Mählmann, 2011) or ratings shopping behaviour (Skreta and Veldkamp, 2009). These results however, do support the reputation hypothesis on credit rating agencies (Dimitrov et al., 2015; Morris, 2001) that states that rating agencies will not inflate ratings when they are suspected to do so as this will severely damage their reputation. Dodd-Frank provides further incentives to not inflate ratings by penalizing inflated ratings without penalizing deflated ratings (Dimitrov et al., 2015; Goel and Thakor, 2011).
subsamples of industries and rating categories. Third and finally, the estimation methods come with several limitations such as that OLS is not optimal for ordinal numbers and the use of dummies that depend on the dependent variable.
7 Bibliography
Agarwal, V., Taffler, R., 2008. Comparing the performance of market-based and accounting-based bankruptcy prediction models. J. Bank. Financ. 32, 1541–1551.
Alp, A., 2013. Structural shifts in credit rating standards. J. Finance 68, 2435–2470. doi:10.1111/jofi.12070
Altman, E.I., 2000. Predicting financial distress of companies: revisiting the z-score and zeta models background.
Altman, E.I., 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J. Finance 23, 589–609.
Bae, K.-H., Kang, J.-K., Wang, J., 2015. Does increased competition affect credit ratings? A reexamination of the effect of Fitch’s market share on credit ratings in the corporate bond market. J. Financ. Quant. Anal. 50, 1011–1035. doi:10.1017/S0022109015000472 Baghai, R.P., Servaes, H., Tamayo, A., 2014. Have rating agencies become more
conservative? Implications for capital structure and debt pricing. J. Finance 69, 1961– 2005. doi:10.1111/jofi.12153
Beaver, W.H., 1966. Financial ratios as predictors of failure. J. Account. Res. Empir. Res. Account. Sel. Stud. 4, 71–111.
Becker, B., Milbourn, T., 2010. How did increased competition affect credit ratings? doi:10.1016/j.jfineco.2011.03.012
Bharath, S.T., Shumway, T., 2004. Forecasting default with the KMV-Merton model.
Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities. J. Polit. Econ. 81, 637–654.
Blume, M.E., Lim, F., MacKinlay, A.C., 1998. The declining credit quality of U.S. corporate debt: myth or reality? J. Finance 53, 1389–1413.
Bolton, P., Freixas, X., Shapiro, J., 2012. The credit ratings game. J. Finance 67, 85–112. doi:10.1111/j.1540-6261.2011.01708.x
Campbell, C., John, Y., Hilscher, D., Szilagyi, J., 2011. Predicting financial distress and the performance of distressed stocks. J. Invest. Manag. 9, 14–34.
Campbell, J.Y., Hilscher, J., Szilagyi, J., 2008. In search of distress risk. J. Finance 63, 2899– 2939.
Cantor, R., Packer, F., 1997. Differences of opinion and selection bias in the credit rating industry. J. Bank. Financ. 21, 1395–1417. doi:10.1016/S0378-4266(97)00024-1
Cheng, M., Neamtiu, M., 2009. An empirical analysis of changes in credit rating properties: Timeliness, accuracy and volatility. J. Account. Econ. 47, 108–130.
doi:10.1016/j.jacceco.2008.11.001
