abc

SFB 649 Discussion Paper 2013-003

Empirical Research on Corporate CreditRatings: A Literature Review
Alexander B. Matthies*

This research was supported by the Deutsche Forschungsgemeinschaft through the SFB 649 "Economic Risk".http://sfb649.wiwi.hu-berlin.de ISSN 1860-5664 SFB 649, Humboldt-Universität zu Berlin Spandauer Straße 1, D-10178 Berlin

SFB

* Christian Albrechts Universität zu Kiel, Germany

649

ECONOMIC RISK

BERLIN

Empirical Research on Corporate Credit-Ratings: A Literature Review?
Alexander B. Matthies†

Abstract We report on the current state and important older ?ndings of empirical studies on corporate credit ratings and their relationship to ratings of other entities. Speci?cally, we consider the results of three lines of research: The correlation of credit ratings and corporate default, the in?uence of ratings on capital markets, and the determinants of credit ratings and rating changes. Results from each individual line are important and relevant for the construction and interpretation of studies in the other two ?elds, e.g. the choice of statistical methods. Moreover, design and construct of credit ratings and the credit rating scale are essential to understand empirical ?ndings. Keywords: Rating agency; Credit Ratings; Through-the-cycle rating methodology; Corporate Governance JEL classi?cation: G20; G30; G32; G24; G34

1

Introduction

Credit ratings aim to measure the creditworthiness of an entity (e.g. a corporation). They represent an opinion of a rating agency that evaluates the fundamental credit strength of an issuer and his ability to fully and punctually meet his debt obligations (Gonzales et al. (2004), Berblinger (1996) p. 11, and Wappenschmidt (2009)). Credit ratings are produced by professional rating agencies. In the US and the EU regulatory norms exist that agencies must ful?l. For instance, in the US, agencies must be approved as nationally recognized statistical rating organizations (NRSRO) by the securities and exchange commission (SEC) (Hill (2004)). In an economic sense, rating agencies function as ?nancial intermediaries (Ramakrishan & Thakor (1984)) whose existence would not be justi?able under the hypothesis of strictly e?cient markets. The market for ratings has an oligopoly structure as the three leading agencies ‘Moody’s’, ‘Standard & Poor’s’ (S&P), and ‘Fitch Publishing Company’ control around 95% of the market (Asmussen (2005), Wappenschmidt (2009) p. 13). Ratings exist for a wide variety of issuers such as corporations, countries, and structured ?nance products. Banks approve loans and credits (Krahnen (2001) p. 1767) on the
This Research is supported by the German Research Foundation via the CRC 649 ‘Economic Risk’, Humboldt-University Berlin, and a stipend of ’F¨ orderung des wissenschaftlichen und k¨ unstlerischen Nachwuchses des Landes Schleswig-Holstein’, which is gratefully acknowledged. † Christian Albrechts Universt¨ at zu Kiel, e-mail: [email protected]
?

1

basis of internal ratings. These so called ‘shadow ratings’ are distinct from the ratings provided by professional rating agencies. The ratings of agencies can be further divided into solicited (commissioned and paid for by the issuer) and unsolicited ratings (Poon (2003), Poon & Firth (2005), Bannier et al. (2010)). Ratings are placed on a discrete ordinal scale, where for example under the S&P scale an AAA (the so called ‘triple A’) rating is the highest credit rating assigned to issuers with the highest credit-quality and a D rating is the lowest rating assigned to bonds or ?rms in default. Corporate, sovereign, and structured ?nance products have historically behaved differently (Benmelech & Dlugozs (2010)). The current crisis is largely due to bonds of sovereigns and structured ?nance products. In this context, empirical results of corporate ratings serve as a benchmark to understand those of other entities. Agencies insist that ratings are opinions and not recommendations to buy, sell, or hold. Nevertheless, ?nancial regulation in the US and the EU is often based on these ratings (Gonzales et al. (2004)). Overall, ratings are a highly aggregated classi?cation of an issuers debt that can inform a third party on the possibility that the issuers will not be able to meet his obligations (default) (Dilly & M¨ ahlmann (2010)). On the other hand, issuers can use ratings as a mechanism of corporate governance to signal low investment risk and transparency (Nordberg (2011) p. 60). Ratings can, in this sense, solve principal agent problems (Gonzales et al. (2004)). In return issuers achieve access to the debt market and can reduce capital costs, as a higher rating reduces the yield spread to a risk free investment (Gonzales et al. (2004)). Potential investors and banks can use ratings as benchmarks for comparisons with their own analysis and internal ratings (Erlenmaier (2006) p. 39). Credit ratings are produced by using public and con?dential information. Agencies employ a ?rm’s ?nancial statements, franchise value, management quality, and consider its competitive position under di?erent possible economic scenarios to form their judgement (Gonzales et al. (2004)). They attempt to asses the long term quality of a corporation. They therefore, take a through-the-cycle approach in contrast to a more point-in-time perspective (Amato & Fur?ne (2004) and Altman & Rijken (2006)). The through-thecycle approach means that agencies try to form their opinion independently of short term business cycle e?ects. Point-in-time estimations of default probabilities are based on Merton-type models (based on Merton (1974)) which present more of a short term risk assessments (Gonzales et al. (2004)). In the wake of numerous false assessments of credit quality in individual cases (e.g. Enron and Lehman Brothers, see Hill (2004), L¨ o?er (2005), G¨ uttler & Wahrenburg (2007), and Gopalan et al. (2009)), possible systematic errors in the market for structured ?nance products (Benmelech & Dlugosz (2010)), and the current sovereign rating crisis in Europe, the regulation and practices of agencies have come into the focus of public debate. In Europe the possible creation of an European rating agency is being considered. Additionally, agencies face further possible government reglementation and regulation (Stolper (2009)). Yet, ratings remain a voluntary institution of capital markets. Nevertheless, through regulation applicable for banks and fonds, ratings are e?ectively compulsory as a monitoring mechanisms (mostly in the US) (White (2007), Opp et al. (2010)), and form the basis of risk management (Nickel et al. (2000)) although there is no obligation to publish credit ratings. Therefore, it is striking what in?uence these privately o?ered measures, based on a censored calculation, have on capital markets. To illustrate, the downgrade of a corporation can signi?cantly increase its capital costs and in?uence its stock price (Kisgen & Strahan (2010)). 2

In the process of globalisation there are increasing ambitions to harmonise accounting standards. As from 2005 all EU corporations are obliged to publish their ?nancial statements under the international ?nancial accounting standards (IFRS). Similar steps have also been taken in other industrial and developing countries. These steps are motivated in principle by the conclusion that capital markets are not e?cient (Fama (1970)) and therefore regulatory measures ought to be taken to reduce information asymmetries. The IFRS is designed to help approximate the market value of equity and serve the market participants as a basis from which they can estimate the size and uncertainty of future cash ?ows. Empirical studies that measure the information content of ?nancial statements ?nd that they are a signi?cant factor of capital market dynamics (e.g. Kothari (2001)). In this sense, the assessment of creditworthiness via ratings and the publication of accounting data serve similar purposes. Credit ratings aim to re?ect the default probability of a ?rm in a single item and ?nancial statements should fairly present the economic situation of a ?rm. Therefore, ratings and ?nancial statements are alternative information instruments for market participants. Yet the obligational character of ?nancial statements and the non obligational character of ratings force one to re?ect to what degree ratings are determined by ?nancial statements. Ratings would be redundant if they could be reproduced with ?nancial statements, corporate governance characteristics, and macroeconomic variables. This paper reviews the empirical studies conducted on credit ratings in the context of international capital markets and their relationship to ?nancial statements. Moreover, we highlight the statistical properties of credit ratings and how these properties are essential to understanding credit ratings. The function and properties of ratings are crucial to formulating and imposing global and concrete regulation on ?nancial markets. Furthermore, market participants need to understand ratings and their relationship to other measures of default in order to apply them. Empirical studies on credit ratings can be roughly divided into three lines of research (Blume et al. (1998)). The ?rst two measure the information content of credit ratings in di?erent ways. The ?rst line analyses whether credit ratings measure what they claim to measure, i.e. an issuers creditworthiness. More speci?cally, the relationship of ratings and corporate default is measured (Zhou (2001) and Jorion & Zhang (2007)). The second line measures the information content of ratings on capital markets. Here, capital market reactions around rating changes are analysed to see if ratings contain additional information (for a review see Gonzales et al. (2004)). The third line of research investigates the determinants of credit ratings. Here ratings as independent variables are modeled on a number of ?nancial data (e.g. Ederington (1985)), corporate governance characteristics (Bhojraj & Sengupta (2003)), and macroeconomic factors (e.g. Amato & Fur?ne (2004)). These studies show that ratings can be replicated to a certain degree solely using publicly available data. Further studies in this context analyse the stability of credit ratings with respect to their ’through–the–cycle’ approach (Altman & Rijken (2006)) and the determining factors of rating changes and transition probabilities (e.g. Lando & Skødeberg (2002) and Koopman et al. (2008)). This paper continues as follows: in Section 2 we review the results of studies that investigate the relationship of ratings and default frequencies. The main ?nding is that ratings correlate on average negatively with future default rates, e.g. a ?rm with the highest rating AAA is less likely to default in the future than one with a B rating. In Section 3 we present the major ?ndings of studies investigating the information content of ratings on capital markets. The results show that the characteristic and structure 3

of the rating scale is crucial. Speci?cally, market reactions are asymmetrical, i.e. prices react di?erently to downgrades than to upgrades. Section 4 provides a review of empirical studies measuring the determinants of credit ratings. Ratings depend on ?nancial ratios, ?rm size, corporate governance mechanisms, and macroeconomic variables. Section 5 surveys the statistical methods used in the studies presented in Section 4 to estimate and predict credit ratings. As in Section 3 we ?nd that the character of the rating scale is an important feature in method selection. Section 6 reviews studies that analyse the empirical properties of rating transitions. We ?nd that transitions, like capital market reactions, have asymmetrical features. Section 7 concludes.

2

Credit ratings and defaults

The informativeness of credit ratings can be basically measured in two ways. First, one can relate ratings within the same class to the frequency of default. The second is to investigate the information content of changes in credit ratings on the adjustment process of prices and returns of bonds and stocks (Jorion & Zhang (2007)). The results of the latter line of inquiry are discussed later in Section 3. We ?rst analyse the results of the former line. Historic default frequencies can be used as estimates for default probabilities of individual rating categories. Agencies do not assign a speci?c default probability to individual rating categories but de?ne them as with regard to credit quality.

2.1

One year default rates

Rating agencies like Moody’s and Standard & Poor’s provide regular studies that relate their credit rating to the frequency of default in ?xed cohorts. Zhou (2001) provides an extensive study of yearly default rates for Moody’s rating from 1971 to 2000. Furthermore, studies estimating transition matrices (explained in detail below in Section 6) calculate transition frequencies to default for all S&P rating categories (e.g. Nickel et al. (2000) and Lando & Skødeberg (2002)), which are used as proxies for default probabilities. Rating scales are not cardinal but ordinal, i.e. the di?erence in credit quality between an AAA rated ?rm and an AA rated ?rm is not the same as between an AA and an A rated ?rm. The di?erence in quality increases with each category down the scale. This feature is re?ected in the measured default frequencies of the individual categories. For example, over an 18 year period from 1971 to 1988 there are no one year defaults in the Aaa and Aa categories (Zhou (2001)). R¨ osch (2005) compares default risk forecasts of di?erent rating philosophies. Speci?cally, the point-in-time and through-the-cycle approaches (more on that in Section 4.3). The through-the-cycle approach gives a certain amount of independence from the business cycle. Yet defaults are an objective event and therefore independent of the rating approach and thus default rates vary across the business cycle (Nickel et al. (2000)). Moreover, macroeconomic and sectorial shocks cause default rates to vary across bond market sectors for long time spans (Cantor & Falkenstein (2001)). Speci?cally, Cantor & Falkenstein (2001) ?nd that discrepant default rates in di?erent sectors and geographic regions are often caused by shocks in small samples. Short-term interval horizons can cause estimated default probabilities to increase nonuniformly. In contrast, longer horizons produce smoother patterns (Jorion & Zhang 4

(2007)). Furthermore, one year default rates correlate with the business cycle (Nickel et al. (2000) and Zhou (2001)) and may therefore be an inaccurate estimate of a categories default probability.

2.2

Multi year default rates

Multi year default rates correlate less with the business cycle and are therefore more likely to re?ect the agencies ’through–the–cycle’ approach. In order to present the default frequency of a particular rating class Jorion & Zhang (2007) and Zhou (2001) analyse multi year default rates for S&P and Moody’s respectively. Table 1 presents the average 8 year (Moody’s) and 10 year (S&P) cumulative default rates from 1983 to 1998 (Moody’s) and 1981 to 2002 (S&P). The default rates for S&P ratings increase smoother than those of Moody’s. This discrepancy could be caused by the lower horizon in the Moody’s set or it might indicate that S&P ratings are more accurate than Moody’s. Nevertheless there are common striking features worth noting. The statistics highlight a crucial feature of credit ratings. Di?erent downgrades can correspond to di?erent changes in default probability. For example, default rates for AAand Aa3 are about 1%, below 5% for A- and A3, and around 10% for BBB- and Baa3. This means, for instance, under the S&P scale a downgrade from AAA to A- over 6 notches presents a smaller change in default probability than a downgrade from BBB- to BB+ over 1 notch. The ordinality of the rating scale is an essential characteristic and a key component in understanding empirical observations of ratings and how to apply them, and crucial in evaluating credit rating related research. Rating Default frequency S&P AAA/Aaa 0.005 AA+/Aa1 0.004 AA/Aa2 0.007 AA-/Aa3 0.012 A+/A1 0.016 A/A2 0.017 A-/A3 0.023 BBB+/Baa1 0.047 BBB/Baa2 0.055 BBB-/Baa3 0.109 BB+/Ba1 0.140 BB/Ba2 0.187 BB-/Ba3 0.265 B+/B1 0.315 B/B2 0.396 B-/B3 0.492 CCC 0.572 Time period 1981 – 2002 Moody’s 0.005 0.004 0.009 0.006 0.011 0.014 0.010 0.024 0.024 0.054 0.143 0.138 0.324 0.397 0.343 0.483 – 1983 – 1998

Table 1: Default frequencies of rating categories of Moody’s and Standard & Poor’s (Sources Jorion & Zhang (2007) and Zouh (2001)).

5

In total, one year default rates correlate with the business cycle, whereas multi year default rates re?ect more the through-the-cycle approach of rating agencies. Both statistics nevertheless stress the ordinal character of the rating scale.

3

Capital market reactions

Ratings have an e?ect on capital markets, in?uencing them directly and/or indirectly through rating based regulation. The direct e?ect of ratings on yields implies that ratings contain information that is publicly unavailable, and that markets are therefore not e?cient. In this sense, empirical studies on market dynamics test the theoretical concept of market e?ciency. Further determinants of bond prices are taxation, systematic risk, volatility, supply and demand, and liquidity (Gonzales et al. (2004)).

3.1

Information content

The e?ect of credit ratings on capital markets has been analysed over a long time in highly aggregated studies. Empirically, the in?uence of ratings has been measured on the price or return of bonds (Katz (1974), Ederington et al. (1987), Goh & Ederington (1999)), stocks (Pinches & Singelton (1978), Holthausen & Leftwich (1986), Followill & Martell (1997), Jorion et al. (2005), Jorion & Zhang (2007)), and more recently credit default swaps (Micu et al. (2004), Carthart et al. (2010)). The focus in these research papers is on the big rating agencies and on the US in geographic terms. A few individual studies have also analysed the European market (Gropp & Richards (2001), Cesare (2006)) and single countries, e.g. the UK (Barron et al. (1997) and Batchelor & Manzoni (2006)), Germany (Steiner & Heinke (2001)), and Spain (Abad–Romero & Robles–Fern´ andez (2006) and (2007)). Speci?cally, Katz (1974) ?nds that investors react to ratings as they react to new information. He performs an event study to test the e?ciency of the bond market by analysing the price adjustment of bonds around rating changes. The expected yields of bonds are forecasted for their old and new rating classes. Katz ?nds no anticipation of the rating change prior to the announcement, implying that bond holders rely on the judgement of rating agencies and that the bond market is ine?cient. These results would justify the existence of rating agencies as ?nancial intermediaries. Ederington et al. (1987) show that investors use ratings and accounting data to form their decisions. Furthermore, they ?nd that ratings provide information beyond that of accounting statistics. Corporations often acquire two ratings. This is a way in which the rating market is informally regulated, the so called ‘two rating regime’ (White (2004)). Kish et al. (1999) test if markets value ratings of S&P and Moody’s di?erently. They ?nd diverging e?ects but can not show that one rating agency is more informative than the other. Beyond the bond market ratings also lead to price adjustment processes on other markets underscoring their importance for market participants. Pinches & Singelton (1978) analyse the movement of stock prices around rating changes. They conclude that agencies lag behind the market which implies that ratings add no additional information to stock markets. Rating agencies actions are not limited to rating changes, but also include placements on watch lists and review announcements. Holthausen & Leftwich (1986) measure an e?ect on the stock market when S&P places a ?rm on its credit watch 6

list. Similarly, Followill & Martell (1997) show that after announcements have an e?ect on the stock market, the e?ect of the actual rating change is negligible. The additional informational content of ratings depends also on how much more information agencies have than most market participants. Jorion et al. (2005) show that since the passing of the Fair Disclosure regulation in 2000 in the US, which gives agencies access to con?dential information, ratings have a stronger e?ect on stock markets than before. The choice of investing in either equity or debt is a central issue in corporate governance (Shleifer & Vishny (1997)). Barron et al. (1997) argues that, due to wealth redistribution e?ects from bond holders to stock holders, rating downgrades could increase stock returns in some cases. In this case, losses on the bond market would not lead to losses on stock markets. Abdad-Romero & Roles-Fern´ andez (2006) and (2007) argue that because stock prices on Spanish markets react negatively to both downgrades and upgrades there could be a wealth redistribution e?ect in the opposite direction from stock holders to bond holders. Alternatively, the discrepant market reactions might also be in?uenced by other factors. For instance, Gropp & Richards (2001) ?nd no announcement e?ects for rating changes of European banks on the bond market but strong e?ects on the stock market. They suggest that this could be due to insu?cient liquidity on the bond market. Micu et al. (2004) ?nd that rating changes cause dynamics on CDS markets. Cesare (2006) measures if market based indicators can predict rating changes. Consistent with other studies, he ?nds that ratings still add information to the market although measures based on bond prices, stock prices, and CDS anticipate downgrades. Speci?cally, CDS are good indicators for negative and stock for positive events. Rating changes of one ?rm can also in?uence other ?rms in the same sector. Cathcart et al. (2010) show that stock and CDS returns correlate through industry e?ects. Apart from the big rating agencies, there are many rating agencies that operate mostly in speci?c sectors or regions. Li et al. (2006) compare the in?uence of two local Japanese agencies with that of Moody’s and S&P on Japanese stock markets. They ?nd that the in?uence of Moody’s and S&P is greater. The lagging of actual rating changes to movements on markets and market based measures highlights the through-the-cycle approach of agencies. This gap to point-intime approaches could be closed by the use of outlooks and reviews (watch lists) (Altman & Rijken (2007)). Furthermore, the results of experimental analysis by Ferri & Morone (2008) suggest that agencies can prevent herding behaviour and increase price convergence. Overall, rating agencies seem to provide additional information to market participants. At least market participants, either through free will or forced by regulation, will act on ratings. This does not de?nitely answer if markets are not e?cient, but at least market participants act as if they are not.

3.2

Asymmetrical e?ects

Market reactions are asymmetrical, especially those of stock prices. Speci?cally, markets react stronger to downgrades than to upgrades (Gonzales et al. (2004)). Downgrades have an empirically signi?cant impact on stock prices in all studies, unlike upgrades that have no e?ect in most studies, e.g. Holthausen & Leftwich (1986) and Dichev & Pitroski (2001). This would imply that downgrades contain valuable information but upgrades do not.

7

Jorion & Zhang (2007) provide a theoretical basis for di?erent price e?ects as a nonlinear function of the rating prior to the announcement of the rating change. Here the change in default probability and the change in stock price are linked using a structural Merton-type model. Prior studies implicitly assumed an equal change in default probability for all rating changes, which is false (see Section 2). Jorion & Zhang (2007) show that, after including the prior rating to forecast price reactions, upgrades also have a signi?cant e?ect on stock prices. Yet this e?ect is still smaller than that of downgrades. This discrepancy in the e?ects of downgrades and upgrades might be due to several causes. It might be that rating agencies spend more resources to detect credit deterioration and ?rms are more willing to make positive information public (Jorion & Zhang (2007)). Downgrades may therefore — in contrast to upgrades — contain additional information that is not publicly available (Gonzales et al. (2004)). It could also be possible that stock markets overreact to downgrades (Dichev & Pitroski (2001)). A further reason may be the way in which rating agencies operate. As Altman & Rijken (2006) note, one of the consequences of the ‘through-the-cycle’ method of rating agencies is that agencies adjust their ratings upwards slower than their downgrades compared to more ‘pointin-time’ Merton-type methods. Another reason might be the observed correlation of ratings and the business cycle (Amato & Fur?ne (2004)). Bar-Isaac & Sharpio (2010) provide a theoretical model for the quality of ratings. It suggests that the quality of ratings decreases in boom phases of the business cycle. As upgrades are more likely to occur during economic expansions, the quality and therefore the information content of upgrades would be lower. Upgrades would then, on average, have a smaller impact on market reactions. In a similar line of research empirical studies test the information content of ?nancial statements. Financial statements can cause similar e?ects on capital markets as ratings do (for an empirical review see M¨ olls & Strauß (2007)). In the next section, we discuss that ?nancial statements are basic determinants of corporate credit ratings. Therefore accounting data and credit ratings can be seen as alternative and even competing mechanisms to reduce information asymmetries on capital markets. To conclude, it is found that capital markets react to credit rating changes, implying that markets are not e?cient. Moreover, reactions are asymmetrical, i.e. stronger for downgrades than for upgrades. This ?nding holds even after allowing for the actual change in default probability that di?ers over rating classes due to the ordinal structure of the rating scale.

4

Determinants of credit ratings

The determinants of credit ratings fall into three main categories. The ?rst are ?nancial ratios and ?nancial data. These variables proxy ?rm speci?c factors such as leverage, liquidity, and ?rm size (e.g. Ederington (1985), Kamstra et al. (2001), and Blume et al. (1998)). The second category are corporate governance mechanisms. Here, factors such as ownership structure and board independence are measured (Bhojraj & Sengupta (2003) and Ashbaugh-Skaife et al.(2006)). The third group comprises macroeconomic factors that could in?uence credit ratings like GDP growth measures (e.g. Amato & Fur?ne (2004)).

8

4.1

Financial statements and ?nancial ratios

Financial ratios are easily obtainable from ?nancial statements and are therefore publicly available information that determine corporate credit ratings. They are also traditionally applied in default prediction studies (e.g. Altman (1968)), which are similar to rating prediction and estimation studies. The ratios represent factors such as leverage, liquidity, interest coverage, and pro?tability that determine a ?rm’s creditworthiness. Similarly, the ?rm’s size also contributes to its default probability and creditworthiness. 4.1.1 Financial ratios and default probability

Altman (1968) uses ?ve ratios, similar to those used in rating studies, to predict bankruptcy: working capital/total assets, retained earnings/total assets, earnings before interest and taxes/total assets, market value of equity/book value of total assets, and sales/total assets. He highlights their possible impact on default probability, and therefore on ratings too. The working capital total assets ratio is a measure of net liquid assets of the ?rm relative to its capitalisation. A ?rm that experiences operational losses will have shrinking current assets in relation to its total assets. Without liquidity it will not be able to continue operations, or repay its debt. Therefore, this ratio should be positively related to credit ratings, i.e. the higher the ratio is ceteris paribus the better the ?rm’s rating. The retained earnings total assets ratio is a measure of cumulative pro?tability over time, implicitly this also measures the age of a ?rm. Younger ?rms are more likely to fail than older ?rms and usually have a low retained earnings total assets ratio. The retained earnings can be used in less pro?table times and may ensure continued operations. The ratio should therefore be positively correlated to credit ratings. The earnings before interests and taxes total assets ratio is a measure of the productivity of the ?rm’s assets. A ?rm’s existence, including its creditworthiness, is based on the earning power of its assets. The productivity of a ?rm should therefore be positively related to credit ratings. The market value of equity to book value of total debt ratio measures how much the ?rm’s assets may decline in value before its liabilities exceed its assets and the ?rm could become insolvent. The ratio is therefore positively related to credit ratings. The sales total assets ratio measures the sales generating ability of the ?rm’s assets, and thereby the capability of the management to deal in competitive situations. A good management should decrease the probability of default and therefore be positively related to credit ratings. Altman (1968) ?nds statistically signi?cant e?ects for all these variables in a default prediction exercise. Kim & Sohn (2008) employ these ratios to predict rating transitions, and also ?nd a signi?cant e?ect. 4.1.2 Determinants of ratings

The ratios of Altman (1968) are signi?cant determinants of corporate bankruptcy. Similar ratios and ?nancial variables are employed to estimate and forecast ratings. Speci?cally, Kaplan & Urwitz (1979) use interest coverage, the long term debt to total assets ratio, the long term debt to net worth ratio, the net income to total assets ratio, the coe?cient of variation of total assets, the coe?cient of variation of net income, and total assets. Ederington (1985) uses interest coverage, the long term debt to capital ratio, and total 9

assets. Kamstra et al. (2001) use net income plus interest expenses divided by interest expenses to represent interest coverage, a debt ratio measured by total debt divided by total assets, pro?tability captured by the net income total assets ratio, and ?rm size measured as book value of ?rm assets. The ?nancial ratios employed in these studies usually have a statistically signi?cant intuitive e?ect. Speci?cally, in Kamstra et al. (2001) the debt ratio is negatively related and return on assets is positively related to credit ratings. The ?rm’s size equally signi?cantly improves ratings, i.e. on average larger ?rms will have better ratings. In contrast, they are unable to ?nd a statistically signi?cant e?ect of the interest coverage, implying that it does not determine credit ratings. The stability of the determinant’s e?ects does change over the time span of the relationship between ?rms and the rating agency. M¨ ahlmann (2011) shows that corporate ratings improve the longer the relationship lasts although the default rates do not decrease. 4.1.3 Nonlinear e?ects and multicollinearity

Blume et al. (1998) provide a more comprehensive study by estimating ratings in a panel regression from 1978 to 1995. They examine whether the observed decline in the quality of US corporate debt is due to more stringent rating standards. They employ ratios for pre-tax interest coverage, operating income to sales, long term debt to assets, total debt to assets, and total assets. The ?ndings are that the long term ratio is signi?cantly related to credit ratings whereas the total debt ratio is insigni?cant. This is puzzling as the total debt ratio should be negatively related to credit ratings. An answer may lie in the high correlation of these two variables. More speci?cally, the insigni?cance might be due to multicollinearity as Amato & Fur?ne (2004) point out. Motivated by the strong skewness in the distribution of interest coverage (a much smaller average than median), Blume et al. (1998) allow for a nonlinear e?ect for interest coverage. They show that interest coverage has a signi?cantly positive e?ect for small values and an insigni?cant one for larger values. This result explains the empirical ?ndings of Kamstra et al. (2001). Amato & Fur?ne (2004) con?rm the basic ?nding that interest coverage has a nonlinear in?uence. Financial ratios are the basic determinants of credit quality. Studies that omit these variables are almost incomplete by de?nition. The strong signi?cance and e?ects that are measured emphasise the strong correlation and dependence credit ratings and ?nancial statements have and underscore the observation that credit ratings and ?nancial statements are alternative measures of corporate default. As Kamstra et al. (2001) point out, a problem of ?nancial ratios is that factors such as leadership quality, management ability, and technology changes are not captured or are hard to proxy. Blume et al. (1998) and Amato & Fur?ne (2004) use ? ’s and standard errors from market models to represent measures of management quality. Large ? and standard errors are associated with greater idiosyncratic risk and are therefore negatively related to credit ratings. Yet, as we will see below, this is problematic. Blume et al. (1998) omit measures of corporate governance to determine ratings that are crucial in assessing the credit worthiness of corporations and question the use of market based indicators.

10

4.2

Corporate governance

Bhojraj & Sengupta (2003) point out that a ?rm’s likelihood of default depends on the availability of credible information to evaluate the default risk and agency costs. Both of these are determined by governance mechanisms. Corporate governance is essentially the system by which ?rms are controlled and directed. Corporate governance is the focus of much research (Brown et al. (2011)). Here the choice of investing as a bondholder or a stockholder is one of the central issues (Shleifer & Vishny (1997)). The assessment of the rating agency re?ects the view of a debt owner. Some studies measure the in?uence of corporate governance on ?rm performance. A summary of the state of corporate governance research is provided by Bebchuck & Weisbach (2010). Bhagat & Bolton (2008) ?nd that governance as measured by the Gompers et al. (2003) index is positively correlated with better operating performance. Moreover, they argue that contrary to previous studies, governance measures are not correlated with future stock market performance, if endogeneity is adequately addressed1 . The relationship of bond yields and corporate governance measures is highly correlated to that of credit ratings and corporate governance. Bhojraj & Sengupta (2003) show that corporate governance measures that increase ratings also lower bond yields. Nevertheless, Moody’s and S&P ratings are not su?cient to explain spot rate curves and pricing relationships completely (Elton et al. (2004)). Liu & Jiraporn (2010) ?nd that higher decision making power of CEO’s lowers ratings and increases yield spreads. To value the e?ciency of corporate governance, there are also some corporate governance rating ?rms that seek to provide assistance to investors. The results of Daines et al. (2010) suggest that these commercially available corporate governance rankings provide no useful information to shareholders. Under the agency theory framework of Jensen & Meckling (1976), bondholders face two potential con?icts that can reduce the value of their claims by increasing the probability of default. The ?rst is that one between management and all stakeholders (equity and debt), and the second one is that between bondholders (providing debt) and shareholders (providing equity). 4.2.1 The principal agent problem

The separation of ownership and control leads to information asymmetry problems between external stakeholders and managers. Managers that act sel?shly can reduce expected cash ?ows to the ?rm and its external stakeholders. As the cash ?ows decline, the default risk increases and therefore e?ects a credit rating negatively. In this respect governance mechanisms that provide independent and e?ective monitoring of the management should improve ratings. Ashbaugh-Skaife et al. (2006) refer to this role as the management disciplining hypothesis. Bhojraj & Sengupta (2003) distinguish two mechanisms through which governance mechanisms e?ect credit ratings. The ?rst is agency risk. This is the risk that management acting in its self interest would take actions that deviate from ?rm value maximization, as well as the risk that the manger is incompetent. The second is information risk. This is the risk that managers have private information that would adversely a?ect the default risk of a loan. Governance mechanisms can reduce both these risks. Speci?cally with regard
For a more detailed discussion of the importance of endogeneity of governance measures see Brown et al. (2011).
1

11

to agency risk, ?rms with strong governance should receive a higher rating. Similarly mechanisms that induce ?rms to disclose information in a timely and transparent manner should reduce information risks and therefore improve a ?rm’s rating. Bhojraj & Sengupta (2003) list three factors to capture these mechanisms; institutional ownership, block holding, and board structure. Institutional owners may actively monitor the management actions and if necessary take the required steps to protect shareholder interests. In this case institutional investors have incentives to monitor and control management under the active monitoring hypothesis. On the other hand it could be argued that institutional investors are limited in their motivation to monitor management. Bad management under this passive monitoring hypothesis would then encourage institutional investors rather to sell their stock than to initiate corrective action. This e?ect could vary in a cross sectional analysis, as it might depend on the corporate governance codex of the relevant country. Under this hypothesis, the in?uence of institutional ownership on credit ratings would depend upon the investment strategy, i.e. voice or exit. Yet neither Bhojraj & Sengupta (2003) nor Ashbaugh-Skaife et al. (2006) investigate this e?ect. Concentrated ownership could allow block holders (holding more than ?ve percent of shares) to exercise undue in?uence over management to secure bene?ts that are detrimental to other providers of capital. This private bene?ts hypothesis is distinct from the wealth redistribution hypothesis discussed below. Under the private bene?ts hypothesis block holders are negatively related to ratings. Bhojraj & Sengupta (2003) focus on the con?ict between block holders and other stake holders. A contrary e?ect could be that block holders, similar to institutional investors, monitor management e?ectively. This is the shared bene?ts hypothesis, which postulates a positive e?ect of block holders on ratings. Corporate boards have the duty of monitoring management performance and protecting shareholder interests. Outside directors bear a reputation cost if a ?rm’s performance is poor. This should motivate them to monitor management more closely and thereby improve ratings. On the other hand, outside directors might be ine?cient as they are selected by management, or because board culture might discourage con?ict, thereby causing a negative relation to bond ratings. Bhojraj & Sengupta (2003) use the percentage of stock held by institutional investors and the percentage of outside board directors as proxies for institutional ownership and board independence respectively. They ?nd that both variables e?ect credit ratings positively. This supports both the monitoring hypothesis and the bene?ts of an independent board. To capture block ownership they separately use the percentage of stock held by block holders and the percentage of shares held by institutional investors. Both of these variables have a signi?cantly negative e?ect which supports the private bene?ts hypothesis. Yet without a control variable to represent a possible wealth redistribution e?ect, this result requires further analysis. Bhojraj & Sengupta (2003) argue, based on empirical work, that this e?ect is irrelevant. If this were the case, then the coe?cients of the control variables would be insigni?cant. Moreover, even if the argument from Bhojraj & Sengupta(2003) were correct and wealth transfer e?ects are irrelevant to determine the creditworthiness of a ?rm, rating agencies might still consider them relevant. It might then be similar to an anchor e?ect where a classi?cation decision is in?uenced by an irrelevant factor (Tversky & Kahneman (1974)). A further result of Bhajroi & Sengupta (2003) is that governance mechanisms are more critical for lower rated ?rms. In the light of the results of Blume et al. (1998), mentioned 12

above, who ?nd that ratios are more important for larger ?rms, this result should be further investigated. These results are logically compatible, as ?rm size is a positive determinant of credit ratings. Therefore larger ?rms on average have better ratings. Yet it is necessary to test if the credit rating in the Bhojraj & Sengupta (2003) analysis can be replaced by the independent variable ‘?rm size’ to discriminate for the importance of corporate governance mechanisms. 4.2.2 Wealth redistribution hypothesis

The con?ict between bond holders and stock holders is capsulated in the wealth redistribution hypothesis (Barron et al. (2001)). As shown in Section 3 rating changes can a?ect bond and stock prices di?erently. Gompers et al. (2003) ?nd that stronger shareholder rights correlate with higher ?rm value. Moreover, anti-takeover governance provisions are viewed favorably by bondholders yet not by stockholders (Klack et al. (2005)). Ashbaugh-Skaife et al. (2006) conduct a more comprehensive analysis and follow the framework of Standard & Poor’s (2004) to evaluate a ?rm’s governance structures and practices. Standard & Poor’s focuses on four major components of governance: ownership structure and in?uence, ?nancial shareholder rights and relations, ?nancial transparency, and board structure and process. Moreover, Ashbaugh-Skaife et al. (2006) test for possible wealth redistribution e?ects which Bhojraj & Sengupta (2003) intentionally ignore. They ?nd that the number of block holders has a negative impact on credit ratings this is consistent with Bhojraj & Sengupta (2003) results but also with the wealth redistribution hypothesis. Furthermore, their measure of shareholder rights — a score counting single shareholder rights — is negatively related to credit ratings. This suggests that greater shareholder rights are negative for credit ratings and supports the wealth redistribution hypothesis. Ashbaugh-Skaife et al. (2006) further ?nd a positive in?uence for their measures of ?nancial transparency and board independence and expertise. The results on board structure are also consistent with the results of Bhojraj & Sengupta (2003). In conclusion, Bhojraj & Sengupta (2003) and Ashbaugh-Skaife et al. (2006) provide empirical evidence that corporate default probability depends on governance mechanisms. 4.2.3 Measuring management quality

Blume et al. (1998) and Amato & Fur?ne (2004) employ the ? of market models to proxy for management quality. Moreover, Bhojraj & Sengupta (2003) use cumulative stock returns and the market model ? . Cumulative stock returns are almost unique as a determinant of credit ratings. Its use as a control variable may be considered unusual, as Bhojraj & Sengupta (2003) are uncertain about its relationship to credit ratings. The results of Bhojraj & Sengupta (2003) highlight the unclear in?uence of ? and cumulative stock returns. In a regression without measures of concentrated ownership both variables have an insigni?cant in?uence on credit ratings. This is in contrast to Blume et al. (1998) and Amato & Fur?ne (2004) who ?nd a signi?cantly negative relation for the ? . In the regression that includes block ownership the ? is signi?cantly negative and in the regression with the percentage of institutional owned shares it is signi?cantly positive (the coe?cient of cumulative stock returns shows the exact opposite behaviour). Variables derived from market models therefore seem inappropriate to capture management quality, as its e?ect is not understood. The results could also imply that governance 13

mechanisms that monitor management behaviour are more important for credit ratings than measures of management quality.

4.3

Macroeconomic determinants

Rating Agencies declare that they employ a rating through-the-cycle as opposed to a more point-in-time perspective (e.g. Merton-type models (Gonzales et al. (2004))), to avoid short term business cycle e?ects while assessing the creditworthiness of corporations (Standard & Poor’s (2002)). But it is empirically observed that agency ratings and the business cycle do correlate (Amato & Fur?ne (2004), Kim & Sohn (2008), Feng et al. (2008)). 4.3.1 Observed decline in US debt

Blume et al. (1998) argue that an — at that time — observed increase in downgrades by rating agencies is due to more stringent standards in rating assessment. Speci?cally they ?nd that in a panel regression intercepts for each year show a constant decline over time. This causes the panel model to underpredict ratings in the future, suggesting that rating standards have become more stringent. Amato & Fur?ne (2004) can reproduce this result and attribute the observed stronger trend (in comparison to Blume et al. (1998)) to the fact that they also employed speculative grade ratings. In order to tackle two possible criticisms, Blume et al. (1998) test the robustness of the results. The ?rst for them is their assumption that the slope coe?cients remain constant over time. Yet, in year by year regressions, the intercepts still decline over time and the slope coe?cients vary little and are similar to those in the panel model. Moreover, they ?nd no trend in the independent variables (except for market value, which increases). The second objection they consider is that there could be crucial omitted variables that caused the decline. They show that ?rms that had no rating change over time on average improved the factors that determine their credit ratings to counter the objection. Nevertheless, omitted variables might be found in the set of corporate governance variables used by Ashbaugh-Skaife et al. (2006) or the macroeconomic variables employed by Amato & Fur?ne (2004). Jorion et al. (2009) suggest that the observations of Blume et al. (1998) are due to a decline in accounting quality. Speci?cally, accounting information of investment grade ?rms has reduced in quality over time. In line with these results Hel?n et al. (2011) ?nd that higher annual report disclosures improve credit ratings. 4.3.2 Ratings through the cycle

Amato & Fur?ne (2004) show empirically that the observed decrease in credit quality might be due to macroeconomic e?ects of the business cycle. Nevertheless they claim to ?nd no evidence that ratings are unduly in?uenced by the business cycle, and agencies therefore achieve a form of rating stability. Speci?cally, they advance the Blume et al. (1998) analysis by performing three further regressions in which they replace the time dummies with a linear trend and alternatively with an indicator for expansion and recession (the NBER recession indicator), and two indicators for the state of the economy (the output growth gap and a discrete measure of the output gap). The linear trend is signi?cant which is consistent with the behaviour 14

of the time variables and therefore with the claim that rating standards have become more stringent. Furthermore, the macroeconomic measures are mostly insigni?cant which would con?rm the agencies’ claim that they see through the cycle. Yet, in a further regression with the time series of the ?nancial means, the linear trend is also insigni?cant. This result somewhat undermines the ?ndings of Blume et al. (1998). Moreover, they ?nd that the predictive power of the time dummy model is similar to that with a linear trend and cyclical variables. They subject their results to the possible criticism that putting together investment and speculative grade ratings might result in model misspeci?cation, by performing their analysis solely for investment grade ?rms. Their results imply that investment grade ?rm’s ratings are more cyclical, as the economic indicators have an signi?cant in?uence. Here, similar to the results of Ashbaugh-Skaife et al. (2006) which imply that ratings determine the importance of governance mechanisms, the rating classi?cation could be replaced with ?rm size. They then question the implicit assumption that each observation represents an active decision of the rating agency. Alternatively, it could be that due to resource constraints not every rating is accurate in time (compare with the information content of credit ratings). In a regression that solely uses new issues and rating changes, one can be certain that these observations re?ect active decisions based on recent research. For this sample, procyclicality is even stronger than for the investment grade ratings. Furthermore, the linear trend changes signs and implies that rating standards have become more lenient rather than more stringent, as Blume et al. (1998) suggest. Yet, as Amato & Fur?ne (2004) point out, these results are limited to a subset of the ratings universe, and therefore the results could be biased, as there are numerous instances of unchanged ratings where their presence on the agency’s watch list decisions suggests an active decision. L¨ o?er (2004) points out that a through-the-cycle approach requires a separation of permanent and cyclical components of default risk. In a Monte Carlo simulation he shows how this can cause the observed empirical irregularities of agency ratings. He (L¨ o?er (2005)) then goes on to show how the attempt of the agencies to avoid frequent rating reversals can lead to the empirically observed rating stability, serial dependence of rating transitions (G¨ uttler & Raupach (2010)), and the lag of ratings to changes in issuer’s default risk. Changes in ratings can empirically be captured by so-called transition-probability matrices. Feng et al. (2008) use a factor probit model to predict such credit rating matrices in order to show the e?ect of the economic cycle on corporate ratings. They argue that this e?ect provides evidence that rating agencies, contrary to their claims, apply the point-in-time perspective. Yet, as Amato & Fur?ne (2004) point out, the individual business risk factors of corporations can exhibit cyclical behaviour which may cause the cyclical behaviour of corporate ratings. The practical consequences of rating agencies’ through-the-cycle perspective are highlighted by Altman & Rijken (2006), arguing that the objectives of rating stability and default prediction performance can come into con?ict. They observe that rating agencies only partly adjust ratings to the actual credit quality, and less so for upgrades than for downgrades. This reduces the rating migration probability — con?rming the ?ndings of Amato & Fur?ne (2004) — and delays rating transitions. Furthermore, compared to their own point-in-time predictions, this causes the rating agencies’ one-year-horizon predictions to be less accurate, yet they turn out to be more accurate at larger forecast horizons. 15

More recently Bar-Isaac & Sharpio (2010) develop a model in which rating quality is dependent on factors a?ected by the business cycle. These include agencies’ income from fees and competition on the labour market for analysts. The model is set out to explain the inaccurate predictions of rating agencies during the 2008 ?nancial crisis, speci?cally those of structured ?nance products empirically observed by Benmelech & Dlugosz (2010). In summary, ?nancial ratios are essential to determine the creditworthiness of corporations. Corporate governance characteristics and macroeconomic variables supply additional relevant information for ratings. In particular, corporate governance mechanisms that can reduce principal agent problems between management and stakeholders and e?ect wealth redistribution from bondholders to shareholders are important. Agencies attempt to ?lter out macroeconomic e?ects by a through-the-cycle approach that seeks to be independent of the business cycle. Yet this is di?cult as it is often unclear which developments are fundamental and which are short-term. Unfortunately no study up to now has incorporated all three groups of determinants.

5

Statistical methods used to estimate credit ratings

There are a number of statistical applications to estimate and forecast credit ratings. They di?er in the underlying assumptions they make. Most studies employ a linear regression, logistic regressions, or discriminate analysis method. These are the standard approaches to estimate ratings. Furthermore, some studies for example use neuronal networks or duration and hazard models to forecast rating transitions.

5.1

Classical methods

The ordinary least squares (OLS) and the ordered probit model are ordered methods. They include the assumption that ratings are ranked. In contrast, the unordered logit and linear discriminant analysis are not ranked. Ederington (1985) provides a comprehensive theoretical comparison of these four methods and how they perform for in-sample estimation and for out-of-sample prediction. The OLS regression is an ordered method. An issue arises concerning the de?nition of the independent variable (ratings) which is addressed by assigning integers to the ordered groups. The problem with this approach is that it de?nes ratings as an interval scale on which the di?erence between each two rating classes are the same. The ordered probit model estimates an unobservable continuous variable (e.g. default probability), which falls into intervals that correspond to an observable discrete variable (rating class). Because the intervals can vary in size, the method addresses the scale problem more adequately than the OLS method. A problem of the two ordered methods is that variables may in?uence credit ratings di?erently across di?erent rating categories. The multinomial or unordered logit model allows the importance of variables to vary across ratings. Yet, as the name implies, it does not incorporate the ordered character of the ratings. An assumption of this model is that the error terms have a Weibull distribution, unlike an unordered probit model, where the disturbances are assumed to be normally distributed. A further unordered method is the multivariate discriminate analysis (MDA). It distinguishes itself from the unordered logit model by being a conjoint method. While logit 16

and probit model the rating as an independent variable, MDA considers the distribution parameters of the ?rm’s characteristics to be dependent on the bond ratings. Unordered logit and MDA have di?erent basic assumptions but use the same classi?cation equation. In an empirical application Ederington (1985) ?nds that the unordered logit and ordered probit outperform the OLS and MDA methods. Moreover, unordered logit achieves the best ?t for in-sample estimation and ordered logit performs best for out-ofsample prediction. Most newer studies that focus on business or economic issues apply the ordered logit method (e.g. Blume et al. (1998), Amato & Fur?ne (2004), and AshbaughSkaife et al. (2006)). Intuitively it is more appropriate due to its a priori assumptions to model ratings than the other methods listed above. It assumes the ordered structure of the ratings but it can also adjust to the speci?c features of the rating scale. Nevertheless, it assumes a constant in?uence of variables across all rating categories. Yet, as seen above in Section 4, the in?uence of individual factors vary across rating classes, in particular between investment and speculative grades. Furthermore, the ordered probit panel regression assumes an point-in-time perspective instead of the through-the-cycle approach, that is employed by rating agencies (Altman & Rijken (2004)). This is problematic if the probit method is used to forecast rating changes as in Amato & Fur?ne (2004). In other rating transition studies it is common to use forms of duration or hazard models (Du & Suo (2005) and Koopman et al. (2008)). Yet, due to its economic signi?cance, a large number of studies have employed further methods to predict ratings.

5.2

Learning methods

In principal the classi?cation of a credit rating to a given set of ?rm speci?c variables can be regarded as a mere categorisation problem. In this sense, methods can be ’trained’ on a sample of ratings and corresponding ?nancial data. These so called learning or arti?cial intelligence methods can then be used in forecasting exercises. For instance, within neuronal networks learning is de?ned as the search for the weights to produce the best ?t with the given training data (Kwon et al. (1997)). Most of these studies are not discussed within the business or economic research but in the context of the development and application of alternative methods in computer science. These are methods such as the already mentioned neuronal networks (Dutta & Shekhar (1988) and Kwon et al. (1997)), support vector machines (SVM) (Huang et al. (2004), H¨ ardle et al. (2005), Ahn et al. (2005), Cao et al. (2006), Chen & Shih (2006), Lee (2007), and Ye et al. (2008)), fuzzy logic (Shin et al. (2004), Liu & Liu (2005)), and ? -grammatical evolution (Brabazon & O’Neill (2008)). Overall, it is important to consider the a priori knowledge on ratings (ordered and ordinal scale) while selecting a statistical method to estimate or predict ratings. Yet the ordered probit model has two di?culties. One is the variation of in?uence of determinants over rating classes. The other is its implicit point-in-time perspective approach while forecasting rating transitions. Studies using so-called ‘learning methods’ are usually not discussed in an economic or business context.

17

6

Rating transitions

Similar to the estimation of credit ratings is the estimation and prediction of credit rating changes. A general di?erence in these two approaches is that studies that forecast rating changes usually condition their forecasts on prior rating. Some studies mentioned above (e.g. Amato & Fur?ne (2004)) incorporated rating transition predictions based on ?rm speci?c factors. Some studies solely rely on a set of more general predictors. Rating transition matrices (or some times migration matrices) are used, that estimate transition probabilities. Further studies investigate the empirical properties of transition probabilities and what determines them.

6.1

Rating transition matrices

Transition matrices are at the centre of modern risk management, as transition probabilities are a crucial component of many credit risk models (Lando & Skødeberg (2002)). Risk is measured using the distribution of rating transitions for the ?rms (bonds) in a given portfolio. In Nickel et al. (2002) the assumption is made that a constant transition probability pij exists that re?ects the likelihood in a given time period (usually a year) that a rating from class i will change to class j . They can, in the simplest way, be estimated by dividing the number of ?rms (or bonds) that change from i in time period t to j in t + 1 with the total number of ?rms in class i in at t. This is an unconditional estimation, where the transition matrix is made up by the pij . A common property of transition matrices is that they are diagonally dominated, i.e. most ratings do not change and those that do, do so only over one or a few notches (Nickel et al. (2000), Lando & Skødeberg (2002), Kim & Sohn (2008)). Nickel et al. (2000) estimate conditional and unconditional transition matrices based on S&P’s ratings from 1970 to 1997. They ?nd that the precision of probability estimation for lower rated bonds is reduced, as there are fewer speculative bonds and higher volatility in the unconditional transition matrix. Estimating conditioned transition matrices they further ?nd that transition probabilities are determined by the state of the business cycle, the regional origin of the issuer, and the issuers industry. Transition matrices are widely used in rating studies (e.g. Lando & Skødeberg (2002), Kim & Sohn (2008), and Koopman et al. (2008)). Nickel et al. (2000) implicitly assume that rating transitions probabilities are produced by a Markov–process, i.e. all ratings in a category have the same up and downgrade probabilities.

6.2

Duration and momentum e?ects

In order to produce more e?cient estimates of transition matrices, Lando & Skødeberg (2002), Du & Suo (2005), and Koopman et al. (2008) employ continuous-time estimations of transition probabilities. That is, instead of yearly transition rates they consider monthly (Lando & Skødeberg (2002) and Du & Suo (2005)) and even daily (Koopman et al. (2008)) observations. Lando & Skødeberg (2002) motivate this procedure by the fact that observations for large transitions are rare or do not occur at all. For example, AAA to D transitions are not observed in one year, but a bond or ?rm can still in the time of one year be downgraded from AAA to A and from A to D. In contrast to discrete-time methods, continuous-time methods capture e?ects like this one. Moreover, they test credit rating 18

transitions for so called non-Markov e?ects, in particular duration and momentum. In a Markov-chain a transition probability depends solely on the current state an object is in and not on how it reached it or how long it has been in that state. In terms of rating transitions that would mean that the probability of, for instance, an AA rated bond to be downgraded does not depend on how long it has been rated AA (duration) or if it reached its state through a down- or upgrade (momentum). Lando & Skødeberg (2002) and Du & Suo (2005) ?nd empirical evidence for both momentum and duration e?ects. Speci?cally, Lando & Skødeberg (2002) ?nd a strong downgrade momentum and only signi?cant upgrade momentum for lower rated bonds. This means that a previous downgrade increases the probability to be downgraded again, but the equivalent for upgrades only holds for lower rated bonds. Furthermore, with respect to duration, the longer a ?rm occupies a rating class the less likely it will be upor downgraded. The asymmetric e?ect measured for momentum ?ts together with the results of Jorion & Zhang (2007) with regard to the information content of credit ratings (see Section 3). Part of the price adjustment process around rating changes might include the expectations of market participants of further reclassi?cations. Furthermore, the reluctancy of agencies to issue an upgrade shortly after a downgrade (Altman & Rijken (2006)) and the higher possibility of a further downgrade might contribute to the observed asymmetric price adjustments that are empirically observed. In total, rating transition probabilities are often estimated using rating transition matrices. These di?er over the business cycle, industries, and the regional origin of the issuer. Rating transitions furthermore exhibit non-Markov e?ects, i.e. they do not solely depend on the current state of the rating.

7

Conclusion

This paper reviews the empirical work conducted on credit ratings in the context of international capital markets. We highlight three lines of research within this ?eld and how they interact with each other. Furthermore, we show how the implications of each individual line are crucial to understand the results in other ?elds. The ?rst line of research explores the information content of credit ratings and their relationship to corporate default. The results highlight the ordinal character of the rating scale. Di?erences in rating classes do not correspond to equivalent di?erences in default probability. This has important implications for the two second lines of research, the information content of ratings with regard to capital market reactions and the determinants of credit ratings. The relevance of credit ratings changes for capital markets, i.e. the e?cient market hypothesis, can only be measured e?ectively if they are conditioned on the respective change in default probability. Moreover, market reactions around rating changes are asymmetrical. Speci?cally, markets react stronger to downgrades than to upgrades (even after incorporating the corresponding default probability). This discrepancy might have multiple reasons. It might be caused by the way rating agencies operate or speci?c features of the market for ratings. On the other hand it might be caused by the behaviour of corporations and how they release relevant information. One important feature of the agencies’ approach that might cause the asymmetrical information content of credit ratings is the so 19

called through-the-cycle method. Rating agencies try to estimate the long term creditworthiness of a corporation independent of short-term business cycle e?ects. Nevertheless, ratings do correlate with the business cycle. Therefore macroeconomic variables along with ?nancial ratios and corporate governance characteristics are determinants of credit ratings. The third line of research highlights the relationship of ?nancial statements and credit ratings. Financial statements form the core element of determining credit ratings. Moreover, corporate governance mechanisms that can reduce principal agent problems between management and stakeholders and e?ect possible wealth redistribution from bondholders to shareholders are relevant to determine the creditworthiness of a corporation. The issue of choosing a statistical method again depends on the ordered and ordinal structure of the rating scale. Most recent studies therefore employ an ordered probit approach. It is ordered like the rating scale and can incorporate its ordinal character. Yet there are two potential de?ciencies. First, the method does not allow for the varying in?uence of variables over di?erent classes of ratings. This is problematic as the in?uence of corporate governance mechanisms and macroeconomic variables depends on the rating class. Second, in particular with respect to predicting rating changes, a panel regression assumes a point-in-time approach instead of the rating agencies’ through-thecycle method. Therefore employing the probit panel approach to credit rating change predictions might produce inaccurate results. Studies to estimate and predict credit rating transitions probabilities and rating transitions employ rating transition matrices and duration methods. They ?nd that rating probabilities depend on the state of the business cycle, its industry, and on the regional origin of the issuer. Moreover, rating transitions depend on the length a rating has spent in one rating class and how it reached that class (via up- or downgrade). More speci?cally, a ?rm that has recently been downgraded is more likely to be downgraded again. This ?nding might again be caused by the agencies’ rating procedures.

References
[1] Abad-Romero, P., D. Robles-Fern´ andez (2006), Risk and return around bond rating changes: New evidence from the Spanish stock market, Journal of Business Finance & Accounting 33(5): 885 – 908. [2] Abad-Romero, P., D. Robles-Fern´ andez (2007), Bond rating changes and stock returns: Evidence from the Spanish stock market, Spanish Economic Review 9: 79 – 103. [3] Ahn, H., K–j. Kim, I. Han (2005), Combining pairwise SVM classi?ers for bond rating, 2005 KMIS Internationazl Conferencehttp://dspace.kaist.ac.kr/bitstream/10203/5370/1/2005-104.pdf. [4] Altman, E. (1968), Financial ratios, discriminant analysis and the prediction of corporate bankruptcy, The Journal of Finance 23(4): 589 – 609. [5] Altman, E., H. Rijken (2004), A point-in-time perspective on through-the-cycle ratings, Financial Analysts Journal 62(1): 54 – 70. [6] Altman, E., H. Rijken (2006), How rating agencies achieve rating stability, Journal of Banking & Finance 28(11): 2679 – 2714. 20

[7] Altman, E., H. Rijken (2007), The added value of rating outlooks and rating reviews to corporate bond ratings, Financial Management Association Meeting, Barcelona. [8] Amato, J., C. Fur?ne (2004), Are credit ratings procyclical?, Journal of Banking & Finance 28: 2641 – 2677. [9] Ashbaugh-Skaife, H., D. Collins, R. LaFond (2006), The E?ects of Corporate Governance on Firms Credit Ratings, Journal of Accounting and Economics 42: 203 – 243. [10] Asmussen, J. (2005), Rating-Agenturen und Finanzaufsicht, Betriebswirtschaftliche Forschung und Praxis 57: 246 – 255. [11] Bhagat, S., B. Bolton (2008), Corporate governance and ?rm performance, Journal of Corporate Finance 14: 257 – 273. [12] Bar–Isaac, H., J. Sharpio (2010), Ratings quality over the business cycle, unpublished manuscript. [13] Barron, M., A. Clare, S. Thomas (1997), The e?ect of bond rating changes and new ratings on UK stock returns, Journal of Business Finance & Accounting 24 (3): 497 – 509. [14] Batchelor, R., K. Manzoni (2006), The dynamics of bond yield spread around rating revisions dates, The Journal of Financial Research 24(3): 405 – 420. [15] Bebchuk, L., M. Weisbach (2010), The state of corporate governance research, The Review of Financial Studies 23(3): 939 – 961. [16] Benmelech, E., J. Dlugozs (2010), The credit rating crisis, NBER 2010. [17] Berblinger, J. (1996), Marktakzeptanz des Ratings durch Qualit¨ at, in B¨ uschgen und Everling: Handbuch Rating Wiesbaden : 21 – 110. [18] Bhojraj, S., P. Sengupta (2003), E?ect of Corporate Governance on Bond Ratings and Yields: The Role of Institutional Investors and Outside Directors, Journal of Business 76(3): 455 – 475. [19] Blume, M., F. Lim, A. MacKinlay (1998), Thee declining quality of US corporate debt: Myth or reality?, Journal of Finance 53: 1389 – 1413. [20] Brabazon, A., M. O’Neill (2008), Bond rating with ? grammatical evolution, Studies in Computational Intelligence 102: 17 – 30. [21] Brown, P., W. Beekes, P. Verhoven (2011), Corporate governance, accounting and ?nance: A review, Accounting & Finance 51: 96 – 172. [22] Cantor, R., E. Falkenstein (2001), Testing for rating consistency in annual default rates, The Journal of Fixed Income 11(2): 36 – 51. [23] Cao, L., L. Guan, Z. Jingqing (2006), Bond rating using support vector machine, Intelligent Data Analysis 10(3): 285 – 296. [24] Cathcart, L., L. El-Jahel, L. Evans (2010), The correlation structure of the CDS market: An empirical investigation, Unpublished manuscript. 21

[25] Cesare, A. (2006), Do market–based indicators anticipate rating agencies? Evidence for international banks, Economic Notes by Banca Monte dei Paschi di Siena SpA 35(1): 121 – 150. [26] Chen, W–H., J–Y. Shih (2006), A study of Taiwan’s issuer credit rating systems using support vector machines, Expert Systems with Applications 30: 427 – 435. [27] Daines, R., I. Gow, D. Larcker (2010), Rating the ratings: How good are commercial governance ratings?, Journal of Financial Economics 98: 439 – 461. [28] Dichev, I., J. Piotroski (2001), The long–run stock returns following bond ratings changes, The Journal of Finance 56(1): 173 – 203. [29] Dilly, M., T. M¨ ahlmann (2010), Ableitung markt-impliziter Ratings aus Credit Default Swap-Spreads, Zeitschrift f¨ ur Bankrecht und Bankwirtschaft 22: 487 – 507. [30] Du, Y., W. Suo (2005), An empirical study on credit rating change behavior, unpublished manuscript. [31] Dutta, S., S. Shekhar (1988), Bond rating: A non-conservative application of neuronal networks, IEEE International Conference on Neuronal Networks 2: 443 – 450. [32] Ederington, L. (1985), Classi?cation models and bond ratings, The ?nancial review 20: 237 – 261. [33] Ederington, L., J. Yawitz, B. Roberts (1987), The informational content of bond ratings, The Journal of Financial Research 10(3): 211 – 226. [34] Elton, E., M. Gruber, D. Agrawal, C. Mann (2004), Factors a?ecting the valuation of corporate bonds, Journal of Banking & Finance 28: 2747 – 2767. [35] Erlenmaier, U. (2006), The shadow rating approach — Experience from bank practice, The Basel II Risk Parameters Berlin Heidelberg : 39 – 77. [36] Fama, E. (1970), E?cient capital markets: A review of theory and empirical work, The Journal of Finance 25: 383 – 417. [37] Feng, D., C. Gourieroux, J. Jasiak (2008), The ordered qualitative model for rating transitions, Journal of Empirical Finance 15: 111 – 130. [38] Ferri, G., A. Morone (2008), The e?ect of rating agencies on herd behaviour, Southern Europe Research in Economics Studies Working Paper No. 22. [39] Followill, R., T. Martell (1997), Bond review and rating change announcements: An examination of informational value and market e?ciency, The Journal of Economics and Finance 21(2): 75 – 82. [40] Goh, J., L. Ederington (1999), Cross–sectional variation in the stock market reaction to bond rating changes, The Quarterly Review of Economics and Finance 39(1): 101 – 112. [41] Gompers, P., J. Ishii, A. Metrick (2003) Corporate governance and equity prices, Quarterly Journal of Economics 118(1): 107 – 155.

22

[42] Gonzales, F., F. Haas, R. Johannes, M. Persson, L. Toledo, R. Violi, C. Zins, M. Wieland (2004), Market dynamics associated with credit ratings: a literature review, Banque de France Financial Stability Review 4: 53 – 76. [43] Gopalan, R., F. Song, V. Yerramilli (2010), Do credit rating agencies underestimates liquidity risk?, Working Paper Washington University, Pennsylvania State University, and Indiana University. [44] Gropp, R., A. Richards (2001), Rating agency actions and pricing of debt equity of European banks: What can we infer about private sector monitoring of bank soundness?, Economic Notes by Banca Monte dei Paschi di Siena SpA 30(3): 373 – 398. [45] G¨ uttler, A., P. Raupach (2010), The impact of downward rating momentum, Journal of Financial Services Research 37: 1 – 23. [46] G¨ uttler, A., M. Wahrenburg (2007), The adjustment of credit ratings in advance of default, Journal of Banking & Finance 31: 751 – 767. [47] H¨ ardle, W., R. Moro, D. Sch¨ afer (2005), Predicting bankruptcy with support vector machines, SFB 649 Discussion Paper 2005-009, Humbolt-Universit¨ at zu Berlin. [48] Hel?n, F., K. Shaw, J. Wild (2011), Credit ratings and disclosure channels, Research in Accounting Regulation 23: 20 – 33. [49] Hill, C. (2004), Regulating the rating agencies, Washington University Law Review 82: 43 – 95. [50] Holthausen, R., R. Leftwich (1986), The e?ect of bond rating changes on common stock prices, The Journal Financial Economics 17: 57 – 89. [51] Huang, Z., H. Chen, C-J. Hsu, W-H. Chen, S. Wu (2004), Credit ratings analysis with support vector machines and neural networks: a market comparative study, Decision Support Systems 37: 543 – 558. [52] Jensen, M., W. Meckling (1976), Theory of the ?rm: Managerial behaviour, agency costs, and ownership structure, Journal of Financial Economics 3: 305 – 360. [53] Jorion, P., Z. Liu, C. Shi (2005), Informational e?ects of regulation FD: Evidence from rating agencies, Journal of Financial Economics 76: 309 – 330. [54] Jorion, P., G. Zhang (2007), Information e?ects of bond rating changes: The role of the rating prior to the announcement, The Journal of Fixed Income 16(4): 45 – 59. [55] Jorion, P., C. Shi, S. Zhang (2009), Tightening credit standards: the role of accounting quality, Review of Accounting Studies 14: 123 – 160. [56] Kamstra, M., P. Kennedy, T-K. Suan (2001), Combining bond rating forecasts using logit, The Financial Review 37: 75 – 96. [57] Kaplan, R., G. Urwitz (1979), Statistical models of bond ratings: A methodological inquiry, Journal of Business 52: 231 – 261. [58] Katz, S. (1974), The price adjustment process of bonds to rating reclassi?cation: A test of bond market e?ciency, The Journal of Finance 29(2): 551 – 559. 23

[59] Kim, Y., S. Sohn (2008), Random e?ects model for credit rating transitions, European Journal of Operational Research 184: 561 – 573. [60] Kisgen, D., P. Strahan (2010), Do regulations based on credit ratings a?ect a ?rm’s cost of capital?, The Review of Financial Studies 23: 4325 – 4347. [61] Kish, R., K. Hogan, G. Olson (1999), Does the market perceive a di?erence in rating agencies?, The Quarterly Review of Economics and Finance 39: 363 – 377. [62] Klock, M., S. Mansi, W. Maxwell (2005), Does corporate governance matter to bondholders?, Journal of Financial and Quantitative Analysis 40(4): 693 – 719. [63] Koopman, S., A. Lucas, A. Monteiro (2008), The multi–state latent factor intensity model for credit rating transitions, Journal of Econometrics 142: 399 – 424. [64] Kothari, S. (2001), Capital markets research in accounting, Journal of Accounting & Economics 31(1–3): 105 – 231. [65] Krahnen, J. (2001), Stichwort ,,Rating Internes”, Gerke und Steiner: Handw¨ orterbuch des Bank- und Finanzwesens 3 Stuttgart : 1767 – 1775. [66] Kwon, Y., I. Han, K. Lee (1997), Ordinal pairwise partitioning (OPP) approach to neural networks training in bond raiting, Intelligent Systems in Accounting, Finance & Management 6: 23 – 40. [67] Lando, D., T. Skødeberg (2002), Analyzing rating transition and rating drift with continuous observations, Journal of Banking & Finance 26: 423 – 444. [68] Lee, Y–C. (2007), Application of support vector machines to corporate credit rating prediction, Expert Systems with Applications 33: 67 – 74. [69] Li, J., Y. Shin, W. Moore (2006), Reactions of Japanese markets to changes in credit ratings by global and local agencies, Journal of Banking & Finance 30: 1007 – 1021. [70] Liu, Y., P. Jiraporn (2010), The e?ect of CEO power on bond ratings and yields, Journal of Empirical Finance 17: 744 – 762. [71] Liu, X., W. Liu (2005), Credit rating analysis with AFS fuzzy logic, Lecture Notes in Computer Science 3612: 1198 – 1204. [72] L¨ o?er, G. (2004), An anatomy of rating through the cycle, Journal of Banking & Finance 28: 695 – 720. [73] L¨ o?er, G. (2005), Avoiding the rating bounce: Why rating agencies are slow to react to new information, Journal of Economic Behaviour & Organization 56: 365 – 381. [74] M¨ ahlmann, T. (2011), Is there a relationship bene?t in credit ratings?, Review of Finance 0: 1 – 36. [75] Micu, M., E. Remolona, P. Wooldridge (2004), The price impact of rating announcements: Evidence from the credit default swap market, BIS Quarterly Review June: 55 – 65. [76] M¨ olls, S., M. Strauß(2007) Zur Informationswirkung der Rechnungslegung — Stand der empirischen Forschung, Kapitalmarktorientierte Rechnungslegung 7(2): 72 – 92. 24

[77] Nickel, P., W. Perraudin, S. Varotto (2000), Stability of rating transitions, Journal of Banking & Finance 24: 203 – 227. [78] Nordberg, D. (2011), Corporate governance: Principles and issues, Sage Publication London. [79] Opp, C., M. Opp, M. Harris (2011), Rating agencies in the face of regulation: Rating in?ation and regulatory arbitrage, Working Paper of the University of California Berkley. [80] Pinches, G., J. Singleton (1978), The adjustment of stock prices to bond rating changes, The Journal of Finance 33(1): 29 – 44. [81] Poon, W. (2003), Are unsolicited credit ratings biased downward?, Journal of Banking & Finance 27(4): 593 – 614. [82] Poon, W., M. Firth (2005), Are unsolicited credit ratings lower? — International evidence from bank ratings, Journal of Business Finance & Accounting 32(9/10): 1741 – 1771. [83] Ramakrishnan, R., A. Thakor (1984), Information reliability and a theory of ?nancial intermediation, Review of Economic Studies 51: 415 – 432. [84] R¨ osch, D. (2005), An empirical comparison of default risk forecasts from alternative credit rating philosophies, International Journal of Forecasting 21: 37 – 51. [85] Shin, K–S., H–J. Kim, S–B. Kwon (2004), A GA–based fuzzy decision tree approach for corporate bond rating, Lecture Notes in Computer Science 3157: 505 – 514. [86] Shleifer, A., R. Vishny (1997), A survey of corporate governance, The Journal of Finance 52(2): 737 – 783. [87] Standard & Poor’s (2004), Standard & Poor’s Corporate Governance Scores and Evaluations: Criteria, Methodology and De?nitions, McGraw–Hill Companies, Inc., New York. [88] Standard & Poor’s (2004), Corporate Ratings Criteria, McGraw–Hill Companies, Inc., New York. [89] Steiner, M., V. Heinke (2001), Event study concerning international bond price e?ects of rating actions, International Journal of Finance & Economics 6(2): 139 – 157. [90] Stolper, A. (2009), Regulation of credit ratings agencies, Journal of Banking & Finance 44: 1266 – 1273. [91] Tversky, A., D. Kahneman (1974), Judgment under Uncertainty: Heuristics and Biases, Science 185 (4157): 1124 – 1131. [92] Wappenschmidt, C. (2009), Ratinganalyse durch internatinale Ratingagenturen, Frankfurt am Main usw. [93] White, L. (2007), A new law for the bond rating industry, Regulation : 30(1): 48 – 52.

25

[94] Ye, Y., S. Liu, J. Li (2008), A multiclass machine learning approach to credit rating prediction, 2008 International Symposiums on Information Processing (ISIP) : 57 – 61. [95] Zhou, C. (2001), credit rating and corporate default, Journal of Fixed Income 3(11): 30 – 40.

26

SFB 649 Discussion Paper Series 2013
For a complete list of Discussion Papers published by the SFB 649, please visithttp://sfb649.wiwi.hu-berlin.de.
001 002 003 "Functional Data Analysis of Generalized Quantile Regressions" by Mengmeng Guo, Lhan Zhou, Jianhua Z. Huang and Wolfgang Karl Härdle, January 2013. "Statistical properties and stability of ratings in a subset of US firms" by Alexander B. Matthies, January 2013. "Empirical Research on Corporate Credit-Ratings: A Literature Review" by Alexander B. Matthies, January 2013.

SFB 649, Spandauer Straße 1, D-10178 Berlinhttp://sfb649.wiwi.hu-berlin.de This research was supported by the Deutsche Forschungsgemeinschaft through the SFB 649 "Economic Risk".



doc_132893171.pdf