LUND UNIVERSITY

PO Box 117221 00 Lund+46 46-222 00 00

Essays on Credit Ratings

Cociorva, Anamaria

2018

Document Version:Publisher's PDF, also known as Version of record

Link to publication

Citation for published version (APA):Cociorva, A. (2018). Essays on Credit Ratings. Lund University (Media-Tryck).

Total number of authors:1

General rightsUnless other specific re-use rights are stated the following general rights apply:Copyright and moral rights for the publications made accessible in the public portal are retained by the authorsand/or other copyright owners and it is a condition of accessing publications that users recognise and abide by thelegal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private studyor research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal

Read more about Creative commons licenses: https://creativecommons.org/licenses/Take down policyIf you believe that this document breaches copyright please contact us providing details, and we will removeaccess to the work immediately and investigate your claim.

AN

AM

AR

IA C

OC

IOR

VA

Essays on C

redit Ratings

Department of Business AdministrationLund University, School of Economics,

Business and Management Lund Studies in Economics and Management 144

ISBN 978-91-7753-956-8ISSN 1652-8220

Essays on Credit RatingsANAMARIA COCIORVA

DEPARTMENT OF BUSINESS ADMINISTRATION| LUND UNIVERSITY

Essays on Credit RatingsThis book investigates the role, as well as the accuracy of credit ratings, both for companies and countries. In the first article, I find that credit ratings capture financial constraints more effectively relative to other popular proxies. The second paper focuses on a change in the rules of a widely-followed bond index, and finds that there is a significant market response for the bonds affected by this change. This suggests that credit ratings also have a significant role for bond market segmentation. In the third paper, I find that sovereign ratings have become increasingly conservative over time. However, in my final paper, I focus on the method used for estimating rating ”conservatism”, and , by using a new, two-step model, I find that previous research has overestimated its magnitude.

978

9177

5395

68

144

1

Essays on Credit Ratings

Anamaria Cociorva

DOCTORAL DISSERTATION

by due permission of the School of Economics and Management, Lund University, Sweden.

To be defended at EC3:101. Date 22.01.2019 and time 13:15.

Faculty opponent Martin Holmén

2

I, the undersigned, being the copyright owner of the abstract of the above-mentioned dissertation, hereby grant to all reference sources permission to publish and disseminate the abstract of the above-mentioned dissertation.

Signature Date 2018-12-12

Organization LUND UNIVERSITY School of Economics and Management Department of Business Administration P.O. Box 7080, SE 220-07 Lund, Sweden

Document name Doctoral Dissertation

Date of issue: 2019-01-22

Author(s):Anamaria Cociorva Sponsoring organization

Title and subtitle: Essays on Credit Ratings

Abstract This thesis consists of four self-contained articles, all of which contribute to the empirical research on credit ratings. Broadly speaking, the first two papers highlight two less ordinary “uses” of credit ratings, in the context of (1) measuring financial constraints and (2) bond market segmentation. The last two papers focus on a relatively new concept in the credit ratings’ literature, specifically the time variation in rating standards, by looking at the sovereign credit ratings, and at how the time variation in rating standards has been measured and interpreted in prior literature. By using an extensive set of tests and corporate investment decisions indicating various degrees of financial constraints, the first paper provides first-hand evidence of the ability of credit ratings to capture corporate responses consistent with various levels of financial constraints. We find that they outperform six common financial constraints proxies in each of our tests and shocks. In the second paper, I use an exogenous change in the index inclusion criteria for a large, widely followed bond index, to show that bonds also exhibit an “index” effect following either addition or deletion to the bond index. Furthermore, this exogenous change leads to different types of index effects, allowing me to compare, for the first time, the relative effect of positive (addition) and negative (deletion) changes in a bond index composition. In this comparison, I find that index addition triggers a market reaction twice as large as index deletion, suggesting that investor awareness may be an important channel for bond pricing. The third paper is the first to provide evidence of rating conservatism for sovereign (i.e. country) credit ratings. I find that a sovereign that received an AA rating in 1993 would only get an A+ twenty years later in spite of maintaining the same economic environment, and this change in conservatism is not justified by an increase in the default rates or by a riskier global environment. In the final paper, I revisit the method commonly used for measuring time variation (i.e. conservatism) of rating standards over time, and I find that the decrease in the year coefficients (i.e. the measure for conservatism according to prior research) is largely a result of common secular trends in the financial variables included in the model. While my findings do not disprove the existence of conservatism, they suggest that its true magnitude may be significantly lower, and that other empirical methods might be more suitable for its assessment.

Key words: Credit ratings, financial constraints, bond index, credit rating standards, rating conservatism, time variation in credit ratings, secular trends, Credit Rating Agencies.

Classification system and/or index terms (if any): JEL:G01, G10, G14,G15, G24

Supplementary bibliographical information Language: English

ISSN and key title: 1652-8220 ISBN:978-91-7753-956-8 (print) 978-91-7753-957-5 (pdf)

Recipient’s notes Number of pages: 220 Price

Security classification

3

Essays on Credit Ratings

Anamaria Cociorva

4

Copyright Anamaria Cociorva

School of Economics and Management Department of Business Administration ISBN 978-91-7753-956-8 (print) ISBN 978-91-7753-957-5 (pdf) ISSN 1652-8220 Printed in Sweden by Media-Tryck, Lund University Lund 2018

Media-Tryck is an environmentallycertified and ISO 14001 certifiedprovider of printed material.Read more about our environmentalwork at www.mediatryck.lu.se

NO

RDIC

SWAN ECOLABE

L

1234 5678

5

To all those who believed in me…

6

Table of contents

Chapter 1. Introduction .............................................................................. 13

1.1. Foreword ...................................................................................... 13

1.2. Credit ratings – a beginner’s guide ............................................... 14

1.3. Why should we care about CRAs? ............................................... 18

1.4. A “tricky” business ....................................................................... 23

1.5. Ratings’ accuracy ......................................................................... 27

1.6. Summary of essays ....................................................................... 28

References ........................................................................................... 33

Chapter 2. Do credit ratings measure financial constraints? .................. 37

2.1. Introduction .................................................................................. 37

2.2. Credit ratings and financial constraints ........................................ 41

2.3. Sample and empirical method ...................................................... 48

2.4. Results .......................................................................................... 57

2.5. Discussion and concluding remarks ............................................. 77

References ........................................................................................... 81

Appendices .......................................................................................... 88

Chapter 3. The indexing effect: evidence from the bond market ............ 99

3.1. Introduction .................................................................................. 99

3.2. Institutional background ............................................................. 102

3.3. Data and sample selection .......................................................... 104

3.4. Empirical strategy ...................................................................... 108

3.5. Index effect results ..................................................................... 111

3.6. Good news versus bad news: market response (a)symmetry ..... 116

3.7. Matching returns ........................................................................ 117

3.8. Conclusion .................................................................................. 120

References ......................................................................................... 122

Appendices ........................................................................................ 124

7

Table of contents

Chapter 4. Credit rating agencies’ conservatism revisited: the sovereign market ............................................................. 129

4.1. Introduction ................................................................................ 129

4.2. Rating standards and the sovereign market ................................ 131

4.3. Data ............................................................................................ 133

4.4. Empirical results ......................................................................... 137

4.5. Robustness tests .......................................................................... 140

4.6. Sample splits .............................................................................. 148

4.7. Is conservatism justified? ........................................................... 153

4.8. Conclusion .................................................................................. 158

References ......................................................................................... 160

Appendices ........................................................................................ 162

Chapter 5. Tightening credit standards: the importance of common trends ....................................................................... 171

5.1. Introduction ................................................................................ 171

5.2. Literature review and conceptual framework ............................. 173

5.3. Data ............................................................................................ 176

5.4. Empirical framework .................................................................. 190

5.5. Results ........................................................................................ 193

5.6. Conclusion .................................................................................. 209

References ......................................................................................... 212

Appendices ........................................................................................ 216

8

Acknowledgements

They said, “A good thesis is a done thesis.” I guess, at this point, I can definitely agree! Here comes the day I have been dreaming about for some time… The end of the greatest challenge I have ever taken; and, as it feels right now, ever will. It was a long and sometimes difficult, but also an interesting, fun, exciting and adventurous journey that I got through thanks to all of you. I would need the size of another doctoral thesis to properly thank each and every single person who contributed, directly or indirectly, to this moment. Therefore, I would like to start by saying that, even if your name is not mentioned here, I do remember exactly when and how I met you (although I might not remember your name…), and that I’m thankful for whatever I’ve learned, heard, experienced, or gotten help with.

First, I would like to thank Niclas Andrén, the one person who inspired me to pursue an academic career (when, in fact, all I wanted to do was become a teacher as good as him!), who encouraged me to apply for the position, and throughout the years, continued to help and support me, even if during some of these years I might not have been the perfect PhD student. Thanks as well for deciding to write a paper together with me. I did learn a lot from this (long) process, it feels like I should have started with our paper! Also, without you, Niclas, I wouldn’t be here now. Tusen tack! Även om det finns inga ord som kan beskriva hur tacksam jag är…

I am also very grateful to Hossein Asgharian, who agreed on short notice to step in as my second supervisor, in spite of the large number of commitments he already had. I really wish this had happened earlier. Thank you so much for your support, extremely quick help, encouragement, and helpful feedback and suggestions!

I was lucky to have not only great supervisors, but also helpful and understanding colleagues. Wherever I end up, I will surely miss the great atmosphere from the department.

A very special “teşekkürler” here goes to Naciye, who quickly became not only my colleague, but also a good friend. Sharing the office and traveling with you were definitely two of the highlights of my PhD!

I am also extremely grateful to Maria Gårdängen, because of whom I discovered how fun it is to supervise students, and whose encouraging words motivated me to continue doing that. Especially at a time when I was losing my motivation for doing research, having the opportunity to supervise and teach reminded me why I started this journey. Maybe, one day, I will get back

9

to research. But until then, I’m very thankful to Maria, Niclas and all the others who gave me the opportunity to discover how much I enjoy the teaching side. Not least, thanks to all our students who contributed to making this job so fun!

Many thanks go as well to Anders Vilhelmsson, Jens Forssbaeck, Håkan Jankensgård, Lars Oxelheim, and many others from the NEK department who gave me valuable comments, advice, or were simply there during the more challenging times of my PhD, when I needed to find the motivation and confidence to finish my thesis. Thanks as well to those at FEK who made it possible for me to keep an office, computer, access to databases and the registration as a PhD student.

Going beyond Lund University, I would first like to thank Avri Ravid, with whom I had fun, insightful and motivating talks each time he came to Lund as a visiting professor, even if he was busy. Thanks as well for the feedback on my research!

I am also grateful for the three months spent as a visiting student at the Swedish House of Finance, where I took my most useful PhD courses and got extremely detailed and valuable research feedback. I’m especially thankful to Mike Burkart and Anders Anderson, who made it possible for me to be there in spite of all odds, to Ramin Baghai for being my “supervisor” during my visiting time, and to Lars Nordén for the very constructive feedback and talk at the PhD Workshop. Also, Fatima, I’m so glad I ended up sharing the office with you. You inspired me and gave me the right motivation during those late work evenings! My visiting time at SHOF gave me a valuable insight into how “top-notch” research is done and, at the time, the extra kick of motivation I sorely needed for continuing my thesis work.

And now beyond Sweden. I consider myself extremely lucky to have discovered IAFDS conferences - I attended only two, but I really wished I could have gone back to see all of you again each year! It truly felt more like a family of constructive researchers and colleagues and less like a bunch of critics who are there to get you (as most conferences turn out to be...). Special thanks to Robert Faff who took this initiative, and, more than that, agreed to be my discussant for the final seminar. Thank you for the support, time, and all the constructive comments! And, of course, thanks to all IAFDS participants who made it fun and interesting – I will never forget how fun that one-hour pitching competition was (perhaps because we won?)!

Even if the dreaded “you-should-be-writing-now” demon was always lingering somewhere in the back of my head, taking a longer time to finish my PhD also

10

gave me the opportunity to do what ultimately matters: enjoy my life. For me, this meant traveling, meeting new people, trying new things. Granted, with a 9-to-5 job, I could have never experienced all the wonderful things that happened during these years. At the same time, doing these enjoyable activities gave me the mental strength to go back to my thesis, and to keep on writing.

Thank you all for your appearances (long or short) in my life during these years. The travelers, the hitchhikers, the hippies, the geniuses, the crazy ones, the philosophers, the sport freaks, the dancers, the academics, the perfectionists, the artists, but also the negative ones, the critics, the deceivers. Meeting all of you made me who I am today, and the amount of insights and knowledge about the world and about myself I got from you, cannot be gained while sitting in an office.

A couple of amazing experiences and activities gave me the strength, positivity and a feeling of safety when my PhD journey seemed more like a roller-coaster. Because of them, I could enjoy the ride, and I knew that every downhill would have its own uphill. And that, once the ride is over, I can be happy and proud I dared to hop in and enjoy the thrill. Therefore, I am extremely grateful to the people who introduced me to those activities, or motivated me to try them.

First, my former student association, AEGEE. The years as an active member were undoubtedly the best ones of my life, and this is why it became quite hard for me to let it go, even after I started my PhD. A huge thanks to Ciprian Cordun (yes, I still remember it was you!) who introduced me to AEGEE , and to Annika Johansson who struggled along with me to keep AEGEE Lund active and to find somebody to replace us. Thanks of course to all of you, former AEGEE members, who made those years unforgettable and who were actually the reason I started to travel and ended up doing a master’s and living abroad. Without AEGEE, I would not have been here today, either.

Second, couch-surfing. The one thing I’m grateful for, Ilker Bildircin, is that you introduced me to this amazing community. In all these years, couch-surfing made it possible to meet the most interesting (and sometimes weird) people, stay in the most diverse environments and homes, and learn about quite amazing things, such as Liberland, Bitcoin (before it got mainstream…), Rainbow Gatherings, the underground life of Tehran, wilderness survival tips, and much more. Traveling would surely not be the same without couch-surfing!

Third, dancing. Thanks to my prior roommate Felipe and, more recently, to two great friends, Florido and Jared, for introducing me to salsa and lindy-hop.

11

Dancing was my best “medicine” during my toughest times. Thanks to all who danced with me, and to the dance schools Sabrosa and HepTown, which made courses and social events possible. Now, I honestly cannot imagine a life without dancing!

Fourth, cycling. Thanks to a great and motivating neighbor, Maria Jacky, who introduced me to this amazing sport. Because of you and Tjejklungan I own four bikes now! And I also discovered how awesome it is to travel by bike.

Fifth, other sports. Because being active really helped me throughout these years: nothing is better than enjoying nature while moving, or than having a nice and supportive group to train with. Speaking of which: Team Turtles! Thanks to all of you, who kept me running on a regular basis for the last three years. I’m especially grateful to Anne Säfström Lanner, an inspiring and amazing (mom-I-wish-I-had) woman, who created our running group and keeps motivating all of us to believe in ourselves and to keep going no matter what!

And now some thanks to my dearest and closest. A downside of Lund is that people don’t stay here for long, and I’ve met many people who meant a lot to me but left, and distance kills most relationships.

This did not happen with Astrid. A decade of friendship, can you believe it? Thank you for being an amazing and caring friend, and for being there for me every time I needed you, even if there are many hundred km between us. I guess real friendships can beat any distance!

Mehdi, even if we did not talk for a while, I want you to know that I miss the times when me, you and Astrid were sharing the apartment. When you all encouraged me to persist in applying for the PhD, when all hope seemed lost. But most of all, I am thankful for your understanding in the moments where I when I was truly lost and I did mistakes, both towards you and towards other people that I cared for.

Shakil…in case you didn’t realize that already, you’re the brother I wish I had. Even if we only met two years ago, you helped me more than anybody else to get through the final parts of my PhD journey and to keep believing in myself. Huge thanks for listening to me and for all the endless talks (and laughs) we’ve had during these two years!

Vera Radu...I guess you win for being my “oldest” friend. Even if far apart, thanks for everything and for always being there when I needed you!

12

Another Vera I should thank, too: you know who you are. Thanks for helping me survive during our first year in Sweden, being a great master (and learning) companion and for making it possible to stay in Sweden and to apply for the PhD when I was completely broke. I truly wish that I could have had your self-discipline during my PhD journey as well!

Woranouch - thanks as well for being so nice and caring, I promise I will come to visit you soon now that it’s finally over!

And of course thanks to the Cozy Crew - without having you around me, I would probably still not be done yet. Thank you all for being a shoulder to cry on, somebody to laugh with (or at), and for simply being who you are!

Last but not least, thanks to my family. I am truly indebted to my grandma, because of whom I had the amazing opportunity to come to Sweden. I pretty much owe you everything, I wish you were still around... Thanks as well to my caring mom; I am sorry I cannot visit as often as you might like and that I have been rather distant all these years. But regardless, I want you to know I love you. And well, thank you dad and Alex for trying to reach out and keep in touch, I really appreciate it!

Looking back at all these years, it really feels that I have done a lot more than just write this thesis. Apart from experiencing new things and meeting inspiring people, I was privileged to learn about fun and interesting research, and to be part of really exciting experiments ranging from diabetes research to building firemen equipment and designing emergency exits. I learnt new languages, improved the ones I knew, and, due to various reasons (among which, yes, I must admit, my never-ending curiosity and fascination for trying out new things), I ended up studying a range of really interesting subjects such as molecular biology, psychology, programming, and much more. I talked about science (and economics) in front of high school students, and I truly enjoyed explaining to them the importance of science and research and motivating them to continue their education. I went from not being able to talk in front of an audience to giving lectures in front of more than 100 students; from first failing a course to actually teaching it to master’s students. But, ultimately, it boils down to the tremendous amount of knowledge I gained: about finance, research, the world, and, not least, about myself. Today, I am proud not only of finishing this thesis, but also of achieving all these things, which, at some point in time, seemed impossible. At the same time, I wouldn’t have accomplished these by myself, so, one more time, thanks to all of you!

13

Chapter 1. Introduction

“There are two superpowers in the world today in my opinion. There’s the United States and there’s Moody’s Bond Rating Service. The United States can destroy you by dropping bombs, and Moody’s can destroy you by downgrading your bonds. And believe me, it’s not clear sometimes who’s more powerful.”

Thomas Friedman, 19961

1.1. Foreword

Two decades, two financial crises and one controversial US election later, the answer to the above statement became clear: it is Moody’s, together with its main competitors S&P and Fitch, also known as the “Big Three” credit rating agencies (hereafter referred to as CRAs).

What happened during recent years proved that CRAs have indeed the power to “make or break” companies, investors, governments and entire capital markets. With an ever-growing number of investment strategies, restrictions, financial instruments and regulations being “hardwired” to credit ratings, they have become much more than the mere “opinions on creditworthiness,” as CRAs define them (Gonzalez et al., 2004).

Since the writing of this dissertation began relatively shortly after the financial crisis of 2009, for which CRAs were considered to be one of the main culprits (e.g., Griffin and Tang, 2012), credit rating research seemed an obvious choice. While there is already a vast amount of research focused on various aspects of credit ratings, the controversies surrounding CRA incentives, the oligopolistic nature of the industry, the complexity of the rating process, the sheer number

1 Public Broadcasting Service, 1996. Free Market Society, Transcript interview between David

Gergen and Thomas L. Friedman.

14

of regulations and not least the dramatic impact on various market participants suggest that research possibilities are far from being exhausted. Therefore, through this dissertation, my purpose is to extend and, at the same time, challenge part of the existing research on credit ratings.

The introductory part of this thesis includes a brief definition of ratings and of the rating process, followed by an overview of the role of credit ratings and a description of the rating industry, along with its inherent conflicts of interest and specific regulatory framework. The following four chapters will each represent one article.

1.2. Credit ratings – a beginner’s guide

CRAs emergence dates back to the early 20th century, when railroad building in U.S. created the need for external capital lending (Partnoy, 1999) and therefore a demand for third-party assessment of borrowers’ creditworthiness. To fulfil this demand, John Moody published the first bond rating in 1909, and began selling manuals and detailed information to the prospective lenders, as well as monitoring the borrowers on a regular basis. These services were particularly useful for smaller investors, who lacked the time and knowledge, and created a prerequisite for the debt market to develop. Over the next decade, the parent companies of the other two biggest rating companies, S&P and Fitch, started to provide rating services as well. Today, rating industry has developed into a multi-billion dollar industry, being an important component of financial markets and overall economies (Güttler and Wahrenburg, 2007). This is not surprising, given that debt is, by far, the main source of capital for both corporates and governments alike. More specifically, according to the “Securities Industry and Financial Markets Association” (SIFMA, 2018), international debt issues were more than double the amount of equity offerings in 2016, with a market value of 92.000 billions USD ( as compared to only 60.000 billions USD for equity).

Interestingly, even if it has been one century since their emergence, the “big three” (S&P, Moody’s and Fitch) still dominate the rating industry, with a market share of 96% by the end of 2016 (SIFMA, 2017). The market share distribution in 2016 for the three CRAs is displayed in Figure 1.

15

Figure 1. US Market share in 2016 The figure displays the market share for US companies by the end of 2016 (SIFMA, 2017).

As CRAs state in the headline of their websites, credit ratings represent an “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 and Poor’s, 2018).

CRAs provide two types of ratings: issue-specific, which relate to one financial instrument, and issuer-specific, concerning the overall credit quality of an obligor. The issuers can be corporations (corporate credit ratings), governments (sovereign ratings) or municipalities (municipal ratings). While the basic definition remains the same, the particular characteristics of each type of issuer result in a very different credit risk assessment. For instance, while corporate issuer ratings assess the “ability” to repay debt, sovereign ratings refer rather to the “willingness” of a government to do so. By the same token, the various features of debt instruments, such as collateral, maturity, coupon, etc., are additional factors that are considered when assessing issue-specific credit risk, meaning that the same issuer can have different ratings for its debt issues (Moody’s Investor Service, 2002). While there are many types of rated instruments and issuers, it is corporate and sovereign debt that have the largest share in terms of market value, accounting for more than half of the total amount outstanding in 2017, as displayed in Figure 2. These are also the two types of debt this thesis will focus on.

16

Figure 2.Outstanding US bond market debt (BUSD) The figure displays the market value over time for each types of public debt instruments issued in US.

Following a lengthy rating process, CRAs assess both quantitatively and qualitatively internal factors (such as financials for companies or macro variables for sovereigns), and external ones (for companies, those would be industry- or country-specific factors; for sovereigns this would mean the regional /global environment). To justify their fees, CRAs claim to have access to a significant amount of proprietary information, which, at least in theory, would allow them to provide a more accurate assessment relative to what an investor could extract from public information, and therefore reduce information asymmetry (Standard and Poor’s, 2018). To exemplify, Figure 3 displays S&P’s rating process for corporate issuers as in Standard and Poor’s, (2013). Figure 3. S&P corporate rating criteria The figure shows a simplified scheme of the corporate rating criteria used by S&P (S&P, 2013)

17

The outcome of this complex and rather opaque process is an alphabet-oriented ordinal scale, ranging from AAA (highest degree of creditworthiness) to D (default). Table 1 describes in more detail the rating scales for S&P, Moody’s and Fitch, as well as the corresponding credit risk assessment (as defined by S&P).

Table 1. "Big Three" long-term credit rating scales for corporate issuers

S&P Fitch Moody's Definition (S&P)

AAA AAA Aaa AAA is the highest rating. Capacity to meet financial commitments is extremely strong.

AA+ AA+ Aa1

'AA' differs from the highest-rated issuers only to a small degree. Capacity to meet financial commitments is very strong.

AA AA Aa2

AA- AA- Aa3

A+ A+ A1 Somewhat more susceptible to the adverse effects of changes in circumstances and economic conditions than obligations in higher-rated categories. However, the capacity to meet financial commitments is still strong.

A A A2

A- A- A3

BBB+

BBB

BBB-

BBB+

BBB

BBB-

Baa1

Baa2

Baa3

Exhibits adequate protection parameters. However, adverse economic conditions or changing circumstances are more likely to weaken the capacity to meet financial commitments.

BB+ BB+ Ba1 Less vulnerable to nonpayment than other speculative issues. However, the issuer faces major ongoing uncertainties or exposure to adverse business, financial, or economic conditions that could lead to an inadequate capacity to meet its financial commitments.

BB BB Ba2

BB- BB- Ba3

B+ B+ B1 More vulnerable to nonpayment than the issuers rated 'BB', but has the capacity to meet its financial commitments. Adverse business, financial, or economic conditions will likely impair the issuer's capacity or willingness to meet its financial commitments.

B B B2

B- B- B3

CCC+ CCC+ Caa1 Vulnerable to nonpayment and is dependent upon favorable business, financial, and economic conditions for the issuer to meet its financial commitments. In the event of adverse business, financial, or economic conditions, the issuer is not likely to have the capacity to meet its financial commitments.

CCC CCC Caa2

CCC- CCC- Caa3

CC CC Ca Highly vulnerable to nonpayment. The 'CC' rating is used when a default has not yet occurred but the default is expected to be a virtual certainty, regardless of the anticipated time to default.

C C C Highly vulnerable to nonpayment, and the obligation is expected to have lower relative seniority or lower ultimate recovery compared with obligations that are rated higher.

D D -

In default or in breach of an imputed promise. The 'D' rating also will be used upon the filing of a bankruptcy petition or the taking of similar action and where default on an obligation is a virtual certainty, for example due to automatic stay provisions.

18

Furthermore, ratings can also be classified according to whether they assess long- or short-term credit risk, as well as whether they incorporate the foreign currency risk (this is particularly important for sovereign ratings). Besides ratings, CRAs also offer various “signaling” assessments, such as outlooks or watch-lists. These serve as a signal to investors that an issuer is under scrutiny and a rating change is possible, but not certain. In a sense, they are similar to a warning for the market participants.

Since this dissertation will focus on S&P’s long-term domestic (corporate or sovereign) issuer ratings, I will not discuss further the other types of ratings and rating services. In the second paper, where Moody’s issue ratings will also be employed, these will be discussed in more detail.

1.3. Why should we care about CRAs?

Originally, ratings were created to reduce the large amount of information asymmetry faced in the past by investors, which led to a decrease in the cost of debt and to a dramatic development of the debt instruments and capital markets. Presently, ratings are much more than a key to accessing the debt markets and an information source for all market participants. The various channels through which ratings affect various market participants is one of the major foci of empirical research on ratings. At the same time, disentangling these different channels is empirically challenging, given the endogenous nature and complexity of the rating process, as well as the scarcity of “natural experiment”-like settings, such as, for example, assigning the wrong rating. In the following section, I will discuss in more detail the importance and impact of credit ratings, focusing particularly on the aspects directly related to my research.

1.3.1. Information asymmetry

When CRAs emerged, in the absence of internet or publicly available financial data and virtually no active capital markets, there was an obvious need for third-party assessment of the riskiness of potential debtors. While banks had to invest a significant amount of time and effort before giving loans to their customers, smaller or private investors had no such resources. This is where CRAs found their niche.

19

Presently, most rated companies are listed, meaning that their financial data is publicly available. Furthermore, since their stocks and bonds are traded on the market, there is an unprecedented amount of “shared market data.” For instance, even information that is not publicly available can be “leaked” to the market either by various media and corporate announcements, or by the signals given by, for example, insider traders. The sheer magnitude and ease of access of data contained online, including on social networks such as LinkedIn or even Facebook, means that today there is less of an information asymmetry than ever before.

Yet, with a multi-billion dollar industry, CRAs seem to thrive. Do they really have as much private information as they claim to? Given that a vast amount of research has shown that ratings can be predicted with a high degree of accuracy with only a few financial variables (e.g., Jones et al., 2015), this seems difficult to believe. Yet, empirical evidence has also proven that markets react to rating announcements (Hull et al., 2004). This can happen either because of the fact that there is indeed new information, or because markets react in anticipation of the consequences of a specific rating change.2

While it is very difficult to gauge the amount of proprietary information embedded in credit ratings, what ultimately matters is that the investors believe in its existence – or, at the very least, that they believe that CRAs have more information than they do. Furthermore, using ratings is also a matter of convenience, since they provide an easy and accessible tool for comparing the credit risk of various issuers. 3

1.3.2. Monitoring

Along with reducing information asymmetry, monitoring mitigates various moral hazard problems. Therefore, in that respect, CRAs perform a similar function to banks. However, the major difference is that, while banks assess the riskiness of their customers with the sole interest of minimizing their own losses, CRAs get paid by their customers to rate them, in order to get debt from other parties. Since, in this case, CRAs are an intermediary, a potential 2 For example, a downgrade to speculative grade, even if completely devoid of new information,

might trigger a market reaction because there are investment restrictions, which will increase the cost of debt for the issuer.

3 CRAs claim to strive to make the ratings as comparable as possible, even within different asset classes. Empirical evidence, however, proved that in practice, this is not always the case (e.g., Cornaggia et al., 2017)

20

principal-agent problem is created instead (Tirole, 2010), raising the question as to whether CRAs can be effective monitors.

Finally, achieving or maintaining a specific rating has become an important managerial target (Graham and Harvey, 2001). This, in turn, leads to companies actively changing their capital structure and financial policy prior to being monitored for a rating change (Kisgen, 2006; Begley, 2015). When this happens, the causal effect between ratings and firm-specific factors blurs even further, making it difficult for empirical researchers to draw causal inferences on the impact or accuracy of ratings.

1.3.3. Access to debt markets

Given the arguments discussed in previous sections, having a rating significantly reduces cost of debt. Furthermore, there is an ever-growing market for rated debt instruments, meaning that any (private) investor can trade. Ratings are, in a sense, the building block of public debt markets, which currently have a much larger market value relative to stocks as discussed in the introduction (SIFMA, 2017). Furthermore, as debt is the preferred source of financing worldwide, reducing the cost of debt is extremely important for companies.

Ratings have also become a useful tool for empirical researchers, since they can be used as a proxy for bond market access or financial constraints, which cannot be directly measured otherwise (e.g., Farre-Mensa and Ljungkvist, 2016; Harford and Uysal, 2014). Sufi (2007) shows that introduction of bank loan ratings affects the financial and investment policy of the companies obtaining the rating, meaning that third-party certification can also have a direct effect on the “end-user.”

In the first paper, we complement this area of research by using credit quality (measured by the rating level), as opposed to access to public debt markets (measured by having a rating), as a proxy for various degrees of financial constraints. By investigating several investment financing behavior patterns and supply shocks consistent with the existence of financial constraints, we find that ratings consistently outperform several established financial constraints scores based on both qualitative and quantitative data. Our results suggest therefore that CRAs do indeed have an important role as monitors and in reducing information asymmetry, with ratings capturing information above and beyond pure credit risk.

21

1.3.4. Rating-based regulations and restrictions

Since 1931, when banks were required to “mark-to-market” lower rated bonds, rating-based regulations and investment restrictions turned ratings from mere “opinions” to almost law-like assessments (e.g., Partnoy, 1999; Kisgen and Strahan, 2010). In 1989, financial institutions were prohibited from investing in speculative bonds. In 1993, the Basel Committee on Banking Supervision proposed that all international commercial banks should hold extra capital for their non-investment grade investments. In time, these restrictions extended beyond speculative-graded bonds, and to more than just banks. For instance, SEC implemented rules where credit ratings became the basis for calculating broker-dealer “haircut” requirements and mutual funds couldn’t invest in bonds rated less than A+. Pension funds, Eurobond and Asset-backed securities markets also faced several rating-based investment restrictions.

But the most far-reaching rating-based regulation was the so-called Standardized Approach in Basel II, where the risk-weights for various assets classes, needed as an input for calculating banks’ and insurance companies minimum capital requirements, were dependent on the ratings of those assets. While Basel II was never mandatory, it has been implemented by 122 countries since 2008, when it was created, and therefore many financial institutions worldwide are affected (Basel Committee on Banking Supervision, 2016).

By placing a restriction on the potential bond investors, these regulations ultimately affect issuers’ cost of debt and capital structure (see e.g., Kisgen and Strahan, 2010). Moreover, as I will discuss in more detail in section 1.4.3., these regulations increased the power of CRAs, and may lead to a deterioration in ratings accuracy, as several theoretical and empirical papers have shown.

Following the financial crisis, regulations and governments worldwide have realized that rating-based restrictions need to be counteracted with other regulations limiting the power of CRAs. These will be discussed as well in section 1.4.3.

1.3.5. Signaling / labeling effect

A subtler effect of various rating events – even of the ones that can be perceived as unrelated to changes in credit risk – is the one of being a “signal” to the market. The “signaling” effect can explain why the market reacts to seemingly non-informative rating changes. For example, Cornaggia et al. (2017) show that credit spreads change significantly for the bonds affected by

22

a re-calibration in Moody’s municipal rating scale, which was completely exogenous to issuer fundamentals.

However, it is difficult to claim that such events are purely non-informative. Most likely, the market reacts in anticipation of what might happen after the rating change, for instance the restrictions several investors would face, or the consequent changes in capital structure. Back to the previous example using the re-calibration of Moody’s municipal bond rating scale, Adelino et al. (2017) show that municipalities that were mechanically upgraded as a result of the re-calibration increased their expenditures. Therefore, even if a rating change had been made by mistake (which would be a “natural experiment” type of setting), its consequences will deem it informative to the investors, and this will be reflected in the bond yields or CDS spreads.

In Chapter 3, I investigate a more unusual, and scarcely researched, such setting: the addition or exclusion from a major bond index as a result of a mechanical change in the way the ratings for index constituents are taken into account. More specifically, I find that the market reacts significantly when a bond is moved, added or deleted from the index – even if there is virtually no change in its ratings. In this case, there is a so-called “indexing effect,” akin to the one thoroughly researched in the stock index literature (e.g., Chen et al., 2004; Chang et al., 2014). This effect can be explained by the fact that many institutional investors use the index for their investing, and therefore the movement of a specific bond within or out of the index is important to their investing strategies, which will be further discussed in the following section. At the same time, my results also highlight the fact that index membership is, in fact, an important signal for the investors.

1.3.6. Investing strategies

Another consequence of the widespread use of ratings is the emergence of rating-based financial instruments and investment strategies. While rating-based covenants have been used for a long time, there are tradable instruments, such as options, which have a specific rating group as an underlying asset. Furthermore, “cross-over” bonds (being borderline between investment and speculative, typically split rated) are gaining popularity since they provide a way to get higher yields while still being “investable,” especially for financial institutions that have restrictions for junk bonds (Chen et al., 2014).4 This is 4 Similarly to bond index rating rules, the rating-based regulations need to have as a starting

point a common rating for the split-rated bonds, and that could be ”having at least one

23

yet another channel that can explain the findings in Chapter 3: the market reacts to bonds being added or deleted from an index not because of their original ratings (which do not change), but because the inclusion to this particular index allows them to invest in these bonds, even if they are speculative-rated by two out of three CRAs.

1.4. A “tricky” business

As if the complexity of the rating process and its multi-layered impact on financial markets were not enough, the rating industry structure has been criticized for its alleged conflicts of interest. This is the area where most of the theoretical work within rating research has been done, in search for the “optimal” business model or industry structure – which, not surprisingly, hasn’t been found yet. In the following sections, each of the main characteristics of the rating industry and their business model will be discussed in more detail. Lastly, I provide a brief overview of how these characteristics could alter ratings accuracy, according to existing empirical and theoretical evidence.

1.4.1. Money versus reputation

As CRA critics would be quick to point out, the fact that CRAs get paid by their customers5 can be a major threat to ratings accuracy, since ratings may be inflated to ensure returning customers. But why would CRAs risk their reputation by doing this? For a long time, nobody questioned the accuracy of ratings, given that their reputation was crucial for CRA expansion. However, what happened during recent decades revealed a darker side of the rating business. Morris (2001) developed a model of “political correctness,” where an informed advisor will be unbiased only as long as the uninformed decision maker believes in their honesty. Another explanation for why ratings seem to have become increasingly inaccurate in more recent years is advanced in

investment-grade rating.” In this case, if a bond has speculative rating from Fitch and Moody’s, but is rated investment by S&P, then the institutional investor can still have it in their portfolio, while the bond will have relatively higher yields.

5 The “Big Three” assign mostly issuer-paid ratings, with a few exceptions, such as several governmental entities; however, more recently, there have been quite a few smaller CRAs that are investor-paid;

24

Skreta and Veldcamp (2009), who show how increased asset complexity (which implies that other market participants cannot easily detect biases in ratings) leads to an incentive for CRAs to inflate ratings in order to please (particularly) high-paying customers, such as mortgage-backed securities (MBS) issuers. Opp et al. (2013) find that rating-based regulations, which favor highly rated securities, exacerbate the incentive to inflate ratings. Finally, Bar-Isaac and Shapiro (2013) suggest that reputation is countercyclical, i.e., that CRAs care less about reputation (and therefore issue less accurate ratings) in good times, when fee-income is high and default probabilities are low.

1.4.2. Competition

In spite of numerous attempts and regulations aimed at making the rating industry more accessible to smaller CRAs, the “Big Three” still dominate the rating market, as discussed in Section 1.2.

Before 2010, there was a significant barrier to entry in the rating industries, since CRAs had to meet a rather long and ambiguous list of criteria in order to become a NRSRO (“Nationally Recognized Statistical Rating Organization”), given that NRSRO ratings were used in all rating-based rules and regulations. Concerned that the oligopolistic nature of the rating industry may lead to biased ratings and prohibitive fees, regulators have sought to simplify NRSRO criteria, and more recent regulations aiming to reduce the power of the “Big Three”, such as the Dodd-Franck Act, imposed that the rating industry should be open for more competition (Becker and Milbourn, 2011). However, as shown in Figure 2, the remaining CRAs only have a 4% market share combined, so the regulatory efforts didn’t seem to change the industry structure.

In a regular industry, competition would be regarded as a positive factor. However, given the conflicts of interest already existing within the rating industry, it is not clear whether competition would actually improve ratings. Firstly, having the option to choose among several CRAs means that companies will choose the CRA that offers the best rating, a concept known as “rating shopping” (e.g., Sangiorgi and Spatt, 2009). Secondly, if a company has multiple ratings, another CRA willing to rate this company can become a “tiebreaker”. Bongaerts et al. (2012), show that Fitch ratings for split-rated bonds are biased upwards, since Fitch would have an interest to be the “chosen” CRA. Thirdly, the higher the number of ratings for a same issue or issuer, the more complicated and ambiguous it will be to decide which rating

25

to be used for regulations, bond indexes, mutual funds, etc.6 Lastly, empirical evidence on the effects of competition reveal that the quality of the existing ratings’ actually deteriorates after a new CRA rates an already-rated issuer (Becker and Milbourn, 2011).

1.4.3. Regulating CRAs

During the last two decades, two severe sovereign crises and last century’s most far-reaching financial downturn have markedly damaged the reputation and credibility of CRAs. As a response to the avalanche of critiques and worldwide concerns, regulators have sought to make the CRA industry and the rating process more transparent, more open to competition and, not least, attempted to reduce the excessive reliance on credit ratings.

Internationally, the most known CRA-specific such regulation, The Code of Conduct Fundamentals for CRAs, had been already published in 2004 by IOSCO (The International Organization of Securities Commissions).7 In response to the financial crisis, however, a special CRA committee was established, and new evaluations and special reports were issued, such as the ones for structured finance products, as well as changes to the original document requiring increasing transparency and disclosure from CRAs (IOSCO, 2009).

Another international entity, the Bank of International Settlements (BIS), which sets the standard framework for banks’ capital requirements and risk assessment through the Basel Accords, made significant changes to its latest Basel III following the crisis (Basel Committee on Banking Supervision, 2017). The so-called standardized approach (SA) used in Basel I and II, which allowed banks to use external credit ratings for risk-weighting their asset portfolios (and therefore calculating minimum capital requirements), was heavily criticized. Consequently, Basel III suggested a more flexible approach, where banks can alternatively use their internal risk assessments (IRB approach). By doing this, BIS tried to reduce the over-reliance on external credit ratings, and to encourage the use of rating models more specific to each financial institution’s asset portfolio.

6 This happens when ratings are different, since, for index or regulatory purposes, only one rating

is taken into account.

7 IOSCO is an international body that brings together most of the world’s securities regulators and it is recognized as the global standard setter for securities

26

However, neither IOSCO nor BIS can override regional or country-specific CRA regulations, if they do exist. Therefore, it is ultimately the responsibility of the local law makers to make sure that there are clear regulations regarding CRA activity and their use for investment purposes.

In the US, The Security Exchange Commission (SEC) performs this role. SEC’s most important regulation concerning CRAs was the Dodd-Frank Wall Street Reform and Consumer Protection Act, which was passed by the Congress in July 2010, in response to the financial crisis. The Dodd-Frank has set a comprehensive number of new rules aimed at increasing transparency of the rating process, NRSRO requirements, disclosure of rating fees and possible conflicts of interest, as well as imposing legal liability for inaccurate ratings and requiring federal agencies to remove rating-based requirements from existing regulations (Dimitrov et al., 2015).

In Europe, the same post-crisis concerns lead to the founding of ESMA (European Securities and Markets Authority), which has a similar role to SEC: “safeguarding the stability of the European Union’s financial system.” However, ESMA’s tasks are more challenging, since they also involve the convergence of country-specific financial systems and regulations with an integrated financial market. Similarly to SEC’s designations of CRAs as NRSRO’s, European CRAs also have to meet several criteria in order to be ESMA-approved. In a sense, ESMA’s rating-specific regulations followed the Dodd-Frank, by emphasizing excluding ratings from binding regulations/investment restrictions, as well as increasing transparency, reporting conflicts of interests and being held liable for inaccurate ratings (Masera, 2011).

While it might be too early to assess whether these attempts to decrease CRAs’ power and to find a better alternative to ratings had positive effects, there is some empirical evidence that this is not the case. For instance, Dimitrov et al. (2015) show that ratings have actually become less accurate after the Dodd-Frank Act. While this might seem counterintuitive, it is consistent with the rationale in Morris (2001): if feeling threatened, CRAs might become overly conservative, since issuing relatively lower ratings has no regulatory cost.

27

1.5. Ratings’ accuracy

What ultimately matters for all market participants is the extent to which ratings accurately measure what they are supposed to: relative default probability. However, even assuming that CRAs have access to all the information required to issue an accurate rating, the regulatory incentives/pressure, conflicts of interest and the alleged stability of ratings make it difficult to assign completely unbiased ratings. Table 2 summarizes possible factors which may impact ratings accuracy, along with a relevant reference and the expected direction of the resulting bias. Table 2. Overview of various factors relevant for ratings' accuracy

Rather than assuming that ratings are completely accurate, it seems more reasonable to investigate whether there is a systematic change in rating standards, either in (1) cross section (i.e. among different types of issuers), or (2) across time (i.e. an issuer with the same creditworthiness would have a different rating today than, say, two decades ago).

Prior theoretical and empirical research has shown that, indeed, there is a significant change in rating standards in both dimensions. Skreta and Veldkamp (2009) show how asset complexity can lead to rating inflation, which explains why ratings seemed to be overly conservative for sovereigns (i.e. the downgrade of Asian and Eurozone countries during the two sovereign crises), but too lenient for the mortgage-backed securities (the crash of the

Factor Brief description Direction of bias Key reference

Issuer-pay CRAs get paid by their customers Rating inflation Jiang et al. (2012)

Rating shopping Issuers can pick the CRA that offers the best rating

Rating inflation Sangiorgi et al. (2009)

Increased competition

This increases the risk of rating shopping since there is more choice

Rating inflation Becker and Millbourn (2011)

Asset complexity For complex securities, there is higher information asymmetry, but also higher revenues (ex. MBS)

Rating inflation Skreta and Veldkamp (2009)

Rating-based regulation

Investment restrictions based on specific ratings e generally become harsher once the rating falls below a specific threshold

Rating inflation Opp et al. (2013)

Regulatory pressure/criticism

CRAs are held liable/criticized only for assigning worse ratings, so there is a higher cost to being "too nice"

Rating conservatism

Dimitrov et al. (2015)

Reputation There is a higher reputational cost if the ratings are higher than they should be

Rating conservatism

Morris (2001)

28

highly rated MBS is what, according to many, eventually led to the financial crisis). In a more extensive comparison of various asset classes, Cornaggia et al. (2017) show that ratings are not comparable, as CRAs claim.

Across time, a change in rating standards may be triggered by different types of regulations: rating-based restrictions might lead to rating inflation (Opp et al., 2013), while stricter CRA regulations can lead to an opposite incentive (Dimitrov et al., 2015). Furthermore, Bar-Isaac and Shapiro (2013) develop a theoretical model where time variation in rating standards is counter-cyclical, meaning that ratings are less accurate in good times, when fee-income is high and issuers’ default probability is low. However, empirical evidence on corporate ratings suggest that there is a monotonic increase in conservatism over time for US non-financial companies (Blume et al., 1998; Alp, 2013; Baghai et al., 2014).

By using the same measure for conservatism as the aforementioned papers, while replacing financials with macroeconomic variables (both country-specific and global), I find that sovereign ratings follow the same trend in my third paper (Chapter 4).

However, in Chapter 5, I question the interpretation of this measure as indicative of conservatism, and I propose a new, two-step approach, which allows me to disentangle the common trend in the existing variables from other unexplained variation, which may be conservatism. I also suggest an alternative view on conservatism as consistent with a secular trend in the weights of the rating determinants over time: more conservative ratings would imply an increasing weight for e g, leverage or beta, while size or investment would, conversely, become less important over time. My overall findings confirm the fact that previous results in e.g., Baghai et al. (2014) cannot be interpreted as conservatism, due to other confounding trends existent in the explanatory variables.

1.6. Summary of essays

1.6.1. Do credit ratings measure financial constraints?

This paper is a joint effort between Niclas Andrén and myself. While my main responsibilities included the data collection, managing, and modeling, Niclas wrote the first draft, and we took joint responsibility for finalizing the paper.

29

Being able to measure the severity of financial constraints, loosely defined as the wedge between internal and external cost of capital, is central to appreciating the importance of market frictions to corporate investments and financial policies (Fazzari, Hubbard, and Petersen, 1988), asset pricing (Whited and Wu, 2006), and monetary policymaking (Gertler and Gilchrist, 1994). Since financial constraints are not directly observable, researchers rely on indirect proxies. These proxies are generally estimated as a linear combination of company-specific characteristics, and are subsequently extrapolated to very different samples or time periods. Perhaps not surprisingly, several papers (e.g., Andrén and Jankensgård, 2018; Farre-Mensa and Ljungqvist, 2016; Bodnaruk et al., 2015; Hoberg and Maksimovic, 2014), have shown that these proxies fail to identify firms that behave in a manner consistent with being financially constrained.

Given the proprietary and complex nature of credit rating assessment, we expect ratings to be more informative about a firm’s financial constraints, relative to the aforementioned proxies. By using an extensive set of tests and corporate investment decisions indicating various degrees of financial constraints, we find that ratings indeed outperform six commonly used proxies for financial constraints.

More specifically, we find that ratings accurately predict corporate behavior indicative of improvement or deterioration in external financial constraints: omitting or increasing dividends, recycling equity, and having underfunded pension plans. Ratings also predict investment financing behavior, with worse-rated firms relying more on internal financing when financing investment.

Finally, ratings accurately predict corporate responses to two very different shocks to the supply of capital. The reduction in investment after the global Financial Crisis varies predictably with the level of credit rating, while companies benefiting from The Jobs and Growth Tax Relief Reconciliation Act (JGTRRA) increased investment proportionally with their ratings.

1.6.2. The indexing effect: evidence from the bond market

Numerous empirical papers have documented the so-called indexing effect, i.e., the positive long-term risk-adjusted returns following a widely followed stock index addition (e.g., Shleifer, 1986; Wurgler and Zhuravskaya, 2002; Chang et al., 2014). However, the reason why this happens, especially given

30

the efficient market hypothesis,8 is more difficult to assess empirically. Chen et al. (2004) suggest that an effective empirical strategy in this respect would be comparing index additions with other events, such as deletions or re-calibrations, since various channels may differentially influence the effect of “negative” versus positive events. Since stock index deletions are generally a consequence of bankruptcies or de-listings, this comparison is difficult to do for stock indexes.

I circumvent this issue by turning instead to the bond market, while at the same time investigating the indexing effect in an entirely different setting.9 This paper is the first to compare the bond market response to additions with deletions from a widely followed bond index. I do this by exploiting an exogenous, unexpected change in the inclusion criteria of Lehman Brothers Corporate Bond Index, which affected a significant amount of bonds both negatively and positively. By matching the affected bonds with their otherwise similar counterparts, I find that the group of bonds that were added to the index following the change experienced a long-term yield decrease of 0.53%, while the bonds removed from the index had a yield increase of up to 0.30%. These changes are statistically and economically significant, being comparable in absolute terms with yield changes following a bond rating downgrade from investment to speculative grade. My findings are complementary to previous research on stock indexing effects, suggesting that investor awareness is relatively more important compared to other investment-related channels (such as benchmarking, monitoring, or changes in expectations), since the index addition effect is almost twice as large as the effect of deletion.

1.6.3. Credit Rating Agencies’ conservatism revisited: the sovereign market

Sovereign credit ratings have been the target of much criticism in recent years. After the Asian crisis, Credit Rating Agencies were criticized for failing to predict the crisis (Mora, 2006). During the more recent Eurozone sovereign debt crisis, however, CRAs were instead blamed for being too aggressive (Barroso, 2010). Given that there is empirical evidence suggesting increasing conservatism for corporate credit ratings (Blume et al., 1998; Alp, 2013; 8 According to the efficient market hypothesis, stocks have many substitutes and therefore their

demand curves should be flat.

9 To the best of my knowledge, there is only one paper, Chen et al. (2014), looking at the effect of bond index addition (but not deletion)

31

Baghai et al., 2014), I investigate whether there is a similar variation in sovereign credit rating standards for a large sample of countries over a period of two decades.

By following the method suggested by Baghai et al. (2014), who interpret year indicators (in a panel estimation) and the difference between actual and predicted ratings as proxies for conservatism, I find that, after controlling for country-specific characteristics (I build on the sovereign rating model proposed initially by Cantor and Packer, 1996) and global risk factors, governments experience in 2013 a drop of up to 2.2 notches due to tightened standards relative to the base year 1993. This implies that a sovereign which received an AA rating in 1993 would only get A+ twenty years later in spite of maintaining the same fundamentals. This finding is robust to including additional factors, using alternative estimation methods and different subsamples. Furthermore, my result is similar to both Baghai et al. (2014) and Alp (2013), who find an increase in conservatism of 2.9 and respectively 1.6 notches over earlier time periods of similar length.

I also find that this conservatism is not fully justified. First, sovereign bond markets take it into account when pricing debt by “undoing” about 20% of its effect. Second, by using Bank of Canada’s detailed sovereign default dataset and S&P’s rating-specific predicted default frequencies, I show that default rates have been decreasing over time. Third, I find that other “global” factors such as credit scarcity, business cycles or global economic growth only explain up to 15.5% of the year indicators, i.e., my measure of conservatism.

1.6.4. Tightening credit standards: the importance of common trends

Between 1980 and 2009, the average corporate credit rating for US non-financial companies has been steadily declining (see e.g. Alp, 2013). While this trend has been initially attributed to a deterioration in the rated issuers’ credit quality (e.g., Lucas and Lonski, 1992), more recent research suggests that this may be the outcome of increasing conservatism over time (Blume et al., 1998; Alp, 2013; Baghai et al., 2014).

For measuring conservatism, the aforementioned papers model ratings as a function of several company-specific variables, and interpret the unexplained time variation (captured by year fixed effects) as indicative of the level of conservatism in each period. While this variation is monotonically decreasing

32

over time, interpreting this trend as conservatism seems far-fetched, from both an economic and an econometric perspective.

I therefore revisit their results by using two alternative methodological approaches, and I find that the magnitude of conservatism is significantly overstated due to the common secular trends exhibited by other variables.

I first use a two-step approach, which allows me to separate the trend component related to creditworthiness from the one related to other unobservable factors, such as conservatism. Since the unexplained time trend in the credit rating may also depend on omitted variables, I also investigate whether adding other variables that have been changing over time, such as various measures of accounting quality (Givoly et al., 2017; Jorion et al., 2009) or macroeconomics (Baghai et al., 2014), can explain the remaining time trend. I find that this remaining trend, while statistically significant, is less than 20% of the one in e.g. Baghai et al. (2014), corresponding to a decrease of less than half a rating notch over a period of 32 years.

Lastly, I propose an alternative approach to assess conservatism by testing whether the weights of the relevant accounting variables have been changing over time by running yearly cross-sectional regressions (similarly to the first step in Fama and McBeth, 1973). I find that the coefficients for most of the variables used in Baghai et al. (2014) do not exhibit any time trend. This finding doesn’t support the conservatism argument, since more conservative ratings would imply assigning a lower weight to factors negatively correlated with credit risk (e.g., investment, size), and a relatively larger weight for the ones increasing credit risk (e.g., leverage, risk).

While the remaining variation in the time trend may be indeed caused by increased conservatism, my findings suggest that other methods may be more suitable for assessing its true magnitude. Given that Blume et al. (1998) alone have been cited more than 1000 times,10 and their views seem to have influenced the views on CRA behavior of academics, practitioners and regulators (Jorion et al., 2009), this can be an important contribution to the literature.

10 According to Google Scholar

33

References

Adelino, M., Cunha, I., & Ferreira, M. A. (2017). The economic effects of public financing: Evidence from municipal bond ratings recalibration. The Review of Financial Studies, 30(9), 3223-3268.

Alp, A. (2013). Structural shifts in credit rating standards. The Journal of Finance, 68(6), 2435-2470.

Andrén, N., & Jankensgård, H. (2018). Reappearing investment-cash flow sensitivities. Knut Wicksell Center working paper, Lund University School of Economics and Management.

Baghai, R. P., Servaes, H., & Tamayo, A. (2014). Have rating agencies become more conservative? Implications for capital structure and debt pricing. The Journal of Finance, 69(5), 1961-2005.

Bar-Isaac, H., & Shapiro, J. (2013). Ratings quality over the business cycle. Journal of Financial Economics, 108(1), 62-78.

Barroso, J.M.D. (2010). Statement to the European Parliament prior to the meeting of the Heads of State and Government of the Euro Area. European Parliament Plenary, Brussels.

Basel Committee on Banking Supervision (2016). Implementation of Basel standards, Bank for International Settlements.

Basel Committee on Banking Supervision (2017). Basel III: Finalizing post-crisis reforms, Bank for International Settlements.

Becker, B., & Milbourn, T. (2011). How did increased competition affect credit ratings? Journal of Financial Economics, 101(3), 493-514.

Begley, T. (2015). The real costs of corporate credit ratings. Working Paper.

Blume, M. E., Lim, F., & MacKinlay, A. C. (1998). The declining credit quality of US corporate debt: Myth or reality? The Journal of Finance, 53(4), 1389-1413.

Bodnaruk, A., Loughran, T., & McDonald, B. (2015). Using 10-k text to gauge financial constraints. Journal of Financial and Quantitative Analysis, 50(4), 623-646.

Bongaerts, D., Cremers, K. M., & Goetzmann, W. N. (2012). Tiebreaker: Certification and multiple credit ratings. The Journal of Finance, 67(1), 113-152.

Cantor, R., & Packer, F. (1996). Determinants and impact of sovereign credit ratings. The Journal of Fixed Income 6.3: 76-91.

Chang, Y. C., Hong, H., & Liskovich, I. (2014). Regression discontinuity and the price effects of stock market indexing. The Review of Financial Studies, 28(1), 212-246.

Chen, Z., Lookman, A. A., Schürhoff, N., & Seppi, D. J. (2014). Rating-based investment practices and bond market segmentation. The Review of Asset Pricing Studies, 4(2), 162-205.

34

Chen, H., Noronha, G., & Singal, V. (2004). The price response to S&P 500 index additions and deletions: Evidence of asymmetry and a new explanation. The Journal of Finance, 59(4), 1901-1930.

Cornaggia, J. N., Cornaggia, K. J., & Hund, J. E. (2017). Credit ratings across asset classes: a long-term perspective. Review of Finance, 21(2), 465-509.

Cornaggia, J., Cornaggia, K. J., & Israelsen, R. D. (2017). Credit ratings and the cost of municipal financing. The Review of Financial Studies, 31(6), 2038-2079.

Dimitrov, V., Palia, D., & Tang, L. (2015). Impact of the Dodd-Frank act on credit ratings. Journal of Financial Economics, 115(3), 505-520.

Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal of Political Economy, 81(3), 607-636.

Farre-Mensa, J., & Ljungqvist, A. (2016). Do measures of financial constraints measure financial constraints? The Review of Financial Studies, 29(2), 271-308.

Fazzari, S., Hubbard, R. G., & Petersen, B. (1988). Investment, financing decisions, and tax policy. The American Economic Review, 78(2), 200-205.

Gertler, M., & Gilchrist, S. (1994). Monetary policy, business cycles, and the behavior of small manufacturing firms. The Quarterly Journal of Economics, 109(2), 309-340.

Givoly, D., Hayn, C., & Katz, S. (2017). The changing relevance of accounting information to debt holders over time. Review of Accounting Studies, 22(1), 64-108.

Gonzalez, F., Haas, F., Persson, M., Toledo, L., Violi, R., Wieland, M., & Zins, C. (2004). Market dynamics associated with credit ratings: a literature review. Working Paper.

Graham, J. R., & Harvey, C. R. (2001). The theory and practice of corporate finance: Evidence from the field. Journal of Financial Economics, 60(2-3), 187-243.

Griffin, J. M., & Tang, D. Y. (2012). Did subjectivity play a role in CDO credit ratings? The Journal of Finance, 67(4), 1293-1328.

Güttler, A. & Wahrenburg, M. (2007). The adjustment of credit ratings in advance of defaults. Journal of Banking & Finance 31(3), 751-767.

Harford, J., & Uysal, V. B. (2014). Bond market access and investment. Journal of Financial Economics, 112(2), 147-163.

Hoberg, G., & Maksimovic, V. (2014). Redefining financial constraints: A text-based analysis. The Review of Financial Studies, 28(5), 1312-1352.

Hull, J., Predescu, M., & White, A. (2004). The relationship between credit default swap spreads, bond yields, and credit rating announcements. Journal of Banking & Finance, 28(11), 2789-2811.

International Organization of Securities Commission. (2004) IOSCO Statement of Principles Regarding the Activities of Credit Rating Agencies. International Organization of Securities Commissions.

35

International Organization of Securities Commissions. (2009). Final Report on the Subprime Crisis. International Organization of Securities Commissions.

Jiang, J. X., Stanford, M. H., & Xie, Y. (2012). Does it matter who pays for bond ratings? Historical evidence. Journal of Financial Economics, 105(3), 607-621.

Jones, S., Johnstone, D., & Wilson, R. (2015). An empirical evaluation of the performance of binary classifiers in the prediction of credit ratings changes. Journal of Banking & Finance, 56, 72-85.

Jorion, P., Shi, C., & Zhang, S. (2009). Tightening credit standards: the role of accounting quality. Review of Accounting Studies, 14(1), 123-160.

Kisgen, D. J. (2006). Credit ratings and capital structure. The Journal of Finance, 61(3), 1035-1072.

Kisgen, D. J., & Strahan, P. E. (2010). Do regulations based on credit ratings affect a firm's cost of capital? The Review of Financial Studies, 23(12), 4324-4347.

Lucas, D. J., & Lonski, J. G. (1992). Changes in corporate credit quality 1970–1990. The Journal of Fixed Income, 1(4), 7-14

Masera, R. (2011). Reforming financial systems after the crisis: a comparison of EU and USA. Working Paper.

Moody’s Investor Service, (2002).Understanding Moody’s Corporate Bond Ratings and Rating Process. Moody’s Global Credit Policy.

Mora, N. (2006). Sovereign credit ratings: Guilty beyond reasonable doubt? Journal of Banking & Finance, 30(7), 2041-2062.

Morris, S. (2001). Political correctness. Journal of Political Economy, 109(2), 231-265.

Opp, C. C., Opp, M. M., & Harris, M. (2013). Rating agencies in the face of regulation. Journal of Financial Economics, 108(1), 46-61.

Partnoy, Frank. (1999). The Siskel and Ebert of financial markets: Two thumbs down for the credit rating agencies. Wash. U.L.Q. 77, 619.

Public Broadcasting Service (1996). Free Market Society, Transcript interview between David Gergen and Thomas L.Friedman.

Sangiorgi, F., Sokobin, J., & Spatt, C. (2009). Credit-rating shopping, selection and the equilibrium structure of ratings. Working Paper, Stockholm School of Economics and Carnegie Mellon University.

Securities Industry and Financial Markets Association. (2018). SIFMA Fact Book. New York.

Securities Industry and Financial Markets Association. (2017). Annual Report on NRSROs. New York.

Shleifer, A. (1986). Do demand curves for stocks slope down? The Journal of Finance, 41(3), 579-590.

Skreta, V., & Veldkamp, L. (2009). Ratings shopping and asset complexity: A theory of ratings inflation. Journal of Monetary Economics, 56(5), 678-695.

36

Standard and Poor’s (2018). Understanding Ratings. RatingsDirect.

Standard and Poor’s (2013).Corporate Methodology. RatingsDirect.

Sufi, A. (2007). The real effects of debt certification: Evidence from the introduction of bank loan ratings. The Review of Financial Studies, 22(4), 1659-1691.

Tirole, J. (2010). The theory of corporate finance. Princeton University Press.

Whited, T. M., & Wu, G. (2006). Financial constraints risk. The Review of Financial Studies, 19(2), 531-559.

Wurgler, J., & Zhuravskaya, E. (2002). Does arbitrage flatten demand curves for stocks? The Journal of Business, 75(4), 583-608.

37

Chapter 2. Do credit ratings measure financial constraints? *with Niclas Andrén

2.1. Introduction

Financial constraints refer to frictions in the supply of capital; the greater the wedge between the internal and external costs of capital, the more financially constrained a firm is. However, capital-supply elasticity is unavailable to the researcher. The typical solution is to infer elasticity by sorting firms on proxies for vulnerability to capital-market imperfections, especially information costs (see Hadlock and Pierce, 2010, for an overview). However, Andrén and Jankensgård (2018), Farre-Mensa and Ljungqvist (2016), Bodnaruk, Loughran, and McDonald (2015), and Hoberg and Maksimovic (2015) show that commonly used proxies fail to identify firms that behave in a manner consistent with being financially constrained. Instead, Farre-Mensa and Ljungqvist (2016) and Andrén and Jankensgård (2018) show that the severity of financial constraints seems to be related to firms’ financial status. A credit rating summarizes this status, and we evaluate its informativeness in this respect, in addition to financial distress. The credit rating contains a great deal of information of relevance for the cost of credit, and we propose that it may be used as a more general proxy for the wedge between the internal and external costs of capital.

Credit ratings fill an important role in the public bond markets by bridging information asymmetries between issuers and investors concerning issuers’

creditworthiness (Santos, 2006). Faulkender and Petersen (2006), Kisgen (2006), and Sufi (2007) show that rated firms have access to more options of debt financing and have higher leverage ratios than unrated firms. Yet it is not the market-access role of credit ratings that interests us. Farre-Mensa and Ljungqvist (2016) and Andrén and Jankensgård (2018) show that being rated is not informative on whether or not a firm is financially constrained; unrated

38

firms do not respond any differently to increases in corporate income taxes (Farre-Mensa and Ljungqvist, 2016), nor are they less likely to initiate or increase dividend payments or to recycle equity. They are also not more likely to reduce or omit dividend payments or have an underfunded pension plan (Andrén and Jankensgård, 2018). Our interest rather revolves around the informativeness of the rating level.

We normally relate ratings to financial distress rather than to financial constraints as they are intended to measure the relative likelihood that a corporation may default, and financial distress costs are just one market friction that may constrain a firm’s access to external financing. Empirically, the default likelihood seems to vary monotonically with the credit rating (e.g., Cheng and Neamtiu, 2009), and ratings also convey information on recovery rates of defaulted bonds (Jankowitsch, Nagler, and Subrahmanyam, 2014). But financial distress is a continuum, not a state. It can range from zero probability of default to bankruptcy, and ratings are intended to reflect the entire continuum. Financial constraints similarly represent a continuum ranging from zero cost wedge between internal and external financing to complete capital rationing where the firm lacks access to external financing. The extent to which these continua overlap has, as far as we know, never been thoroughly investigated, but it is plausible that they at least partly overlap. This is where ratings come in. They may proxy for vulnerability to capital-market imperfections beyond default risk, not least information-related problems. Odders-White and Ready (2006), Cheng and Subramanyam (2008), and He, Wang, and Wei (2011) show that rating levels are inversely related to several measures of equity-market adverse selection, while Harford and Uysal (2014) and Sufi (2007) find that ratings influence investment propensity. Sufi (2007) further shows that the introduction of syndicated-loan ratings by Moody’s and Standard & Poor’s in 1995 increased the supply of available debt financing, while Tang (2009) finds that Moody’s introduction in 1982 of numerical modifiers (+ and –) as refinements of their broad letter-rating classes significantly influenced bond yields. Friewald, Wagner, and Zechner (2014) and Ivashina and Scharfstein (2010) show that credit market access can vary substantially across the rating scale, seen in particular in the global financial crisis. The monitoring by ratings analysts and the threat of rating revision may also mitigate corporate moral hazard problems associated with risk shifting and shirking (Diamond, 1984; Boot, Milbourn, and Schmeits, 2006). Ashbaugh-Skaife, Collins, and LaFond (2006) show that credit ratings are positively associated with accrual quality and earnings timeliness, as well as with

39

multiple governance characteristics, not least board independence, ownership, and expertise, number of blockholders, CEO power, and shareholder rights.

We find strong support for using credit ratings as measures of financial constraints. Corporate characteristics vary systematically across the rating scale, in ways that are reasonable for differences in financial constraints. Hadlock and Pierce (2010) and others emphasize size and age as constraints’ indicators, since they are relatively exogenous firm characteristics and since transparency tends to increase when firms grow larger and older. But size and age may also be used as rating indicators since size and age change near-monotonically across the rating scale, as do many other characteristics, with worse-rated firms being not only smaller and younger, but also less profitable and generating smaller operating cash flows, paying smaller dividends, and investing less in capital expenditures and research and development. The range of characteristics that varies systematically across the rating scale underscores one of the benefits of credit ratings: their multivariate nature. Previous attempts at creating financial constraints indexes (e.g., Mulier, Schoors, and Merlevede, 2016; Hadlock and Pierce, 2010; Whited and Wu, 2006; Cleary, 1999; Kaplan and Zingales, 1997) agree that financial constraints are a multivariate construct.

We show that ratings accurately predict a range of corporate financial behaviors: omitting or increasing dividend payments, recycling equity, and having underfunded pension plans. We employ the tests proposed by Bodnaruk et al. (2015) and find that the worse the rating, the more likely the firm is to omit dividend payments and have underfunded pension plans, and the less likely the firm is to increase dividends and recycle equity. Ratings also predict corporate investment financing, with worse rated firms relying more on internally generated cash flow to finance their investments. In addition, they predict corporate responses to shocks to the supply of capital. We investigate corporate investments in the face of two distinctly different shocks. Our first shock is the global financial crisis of 2007-09. This was the most severe crisis in the US credit system since the great Depression. During this period, credit spreads rose sharply, access to bank loans and credit markets contracted, and credit standards tightened. Firms with weaker ratings reduced their capital expenditures substantially more than better-rated firms in the global financial crisis, consistent with these firms facing greater financial constraints. The second shock is the enactment of the Jobs and Growth Tax Relief Reconciliation Act in 2003. The Act reduced shareholder-level tax rates on dividends and capital gains substantially, and was explicitly intended to

40

encourage corporate investment by reducing the cost of equity capital. Corporate investment responses to the Act line up neatly with the rating scale, with worse-rated firms augmenting investments more strongly than better-rated firms. Again, this is what we would expect if ratings measure financial constraints.

Our paper is most strongly related to Andrén and Jankensgård (2018) and Farre-Mensa and Ljungqvist (2016), who show that financial constraints are inversely related with the firm’s financial status. Our work is also related to that of Bodnaruk et al. (2015) and Hoberg and Maksimovic (2015) who, together with the aforementioned papers by Andrén and Jankensgård (2018) and Farre-Mensa and Ljungqvist (2016) find that existing constraints measures are inadequate. In line with these studies, we also find that the constraints indexes developed by Kaplan and Zingales (1997), Whited and Wu (2006), and Hadlock and Pierce (2010) fail to predict constrained firm behavior. We extend on these studies by evaluating the ability of the constraints scores derived by Hoberg and Maksimovic (2015) to measure financial constraints. In contrast to Hoberg and Maksimovic’s findings, we find that these scores fail to identify severity of financial constraints for our sample of rated US firms. Hoberg and Maksimovic employ automated text-based analysis of 10-Ks to identify constraints from verbal statements in the Management’s Discussion and Analysis section. One reason for the better performance of credit ratings may be that ratings analysts, in contrast to textual analysis and constraints scoring models have the advantage of being able to speak directly to managers, and to do so in confidence and with judgment.

We also contribute to a strand of research that investigates the usefulness of the sensitivity of corporate investment to internal cash flow as a measure of financial constraints. Following Fazzari, Hubbard, and Petersen (1988), a range of studies debate whether constrained firms’ investments are more sensitive to cash flow. Investment-cash flow sensitivity has been criticized on theoretical (e.g., Kaplan and Zingales, 1997), statistical (e.g., Erickson and Whited, 2000), and empirical (e.g., Chen and Chen, 2012) grounds, while proponents defend it on every account (e.g., Fazzari, Hubbard, and Petersen, 2000; Lewellen and Lewellen, 2016; Aǧca and Mozumdar, 2017). We join the proponents by showing that investment-cash flow sensitivity increases monotonically with credit ratings. Moshirian, Nanda, Vadilyev, and Zhang (2017) suggest that investment-cash flow sensitivity is a function of investment intensity rather than financial constraints. In accordance with Andrén and

41

Jankensgård (2018), we find that sensitivity increases monotonically with ratings even after controlling for investment intensity.

In the next section, we summarize theoretical and empirical evidence on the ability of ratings to capture information-related market imperfections. In Section 2.3, we present our sample and methodology. Section 2.4 contains our empirical findings on the ability of credit ratings to measure financial constraints. Section 2.5 summarizes our findings and concludes.

2.2. Credit ratings and financial constraints

A firm is considered more financially constrained when it faces more severe difficulties financing profitable investment opportunities due to the wedge between the costs of internal and external financing. Under a perfect capital market assumption, this cost wedge would be zero and external and internal funds would be perfect substitutes. The zero-cost-wedge hypothesis fails when capital markets are imperfect, not least due to information-related problems such as adverse selection and agency problems (Fazzari et al., 1988) and deadweight costs related to financial distress (Almeida, Campello, and Weisbach, 2011).

A range of financial-constraints proxies have been proposed over the years, beginning with Fazzari, Hubbard, and Peterson (1998), who classified firms on whether or not they had a history of paying dividends. Other popular classifiers include size, age, or whether the firm is rated (see, for example, Gilchrist and Himmelberg, 1995; Almeida, Campello, and Weisbach, 2004; Brown, Fazzari, and Petersen, 2009). Kaplan and Zingales (1997), Hadlock and Pierce (2010), Bodnaruk et al. (2015), and Hoberg and Maksimovic (2015) derive financial constraints classifications by scoring 10-K text based on constraining words. Whited and Wu (1996) alternatively estimate a financial constraints score using a structural model, where they relate investment to the cost of external financing.

Constraints proxies are traditionally evaluated based on patterns in descriptive statistics (e.g., Kaplan and Zingales, 1997; Hadlock and Pierce, 2010); more systematic evaluations of the adequacy of constraints proxies are more recent. Farre-Mensa and Ljungqvist (2016) evaluate five constraints measures (having a history of paying dividends, having a credit rating, and the Kaplan and Zingales (KZ), Hadlock and Pierce (SA), and Whited and Wu (WW) scores).

42

They find no difference in how public firms classified as constrained vs unconstrained respond to financing demand and supply shocks. In contrast, they find privately held firms and public firms close to default to respond to these shocks in the expected, constrained manner. Hoberg and Maksimovic (2015) compare corporate investment in the face of economic shocks. They find that the KZ and WW scores, as well as size fail to predict changes in investment consistent with being financially constrained; age performs better, but inconsistently. In contrast, firms that score higher on their “delay investment score” curtail investments more across the three investigated economic shocks. Bodnaruk et al. (2015) also evaluate the KZ, SA, and WW scores, showing that they do not predict subsequent, informative corporate liquidity-related events, such as changes in dividends, equity recycling, or funding of underfunded pension plans. Their own “percent constraining words” score seems to identify constrained firms more accurately. Andrén and Jankensgård (2018) evaluate eight commonly used proxies (dividend payout ratio, size, age, being rated, and the Cleary (1999), SA, WW, and Mulier et al. (2016) scores) using Bodnaruk et al. (2015) type tests, finding no proxy to accuratelypredict corporate behavior. In contrast, they find that financing status, specifically leverage and interest coverage, accurately predicts corporate behavior correctly.

Farre-Mensa and Ljungqvist (2016) show that rated and unrated firms respond similarly to exogenous shocks to supply of debt and equity financing. This suggests that differences in access to public debt markets are uninformative about how financially constrained a firm is. Instead, they show that public firms close to default appear to be financially constrained. We interpret this as implying that the rating level may matter to financial constraints. In support of our interpretation, Tang (2009) finds that firms that received a higher refined rating in 1982, when Moody’s added numerical modifiers (+ and –) to their broad letter rating classes, increased their use of long-term debt, invested more, and accumulated less cash following the rating refinement than firms that received a lower refined rating. This is consistent with firms that received a higher refined rating becoming less financially constrained, even from such a small rating change as adding a numerical modifier. These findings point towards the rating level, rather than whether or not the firm is rated, being informative on severity of financial constraints.

Financial constraints and financial distress are distinct but related concepts. Distress is often used to refer to a state where a firm has difficulties meeting its principal and interest obligations (e.g., Fazzari et al., 1988), but this is

43

overly restrictive. Distress rather represents a continuum; a firm begins to incur distress costs the second the probability of default is greater than zero. Financial constraints also refer to a continuum, which turns positive as soon as the firm faces investment disruptions due to a cost wedge between internal and external financing. If creditors expect to experience losses in case of default, then any non-zero probability of default will translate into a cost wedge between internal and external financing, and thus distress and constraints are bound to overlap. We emphasize this, because credit ratings are intended to classify firms along the distress continuum. The credit rating represents the credit rating agency’s (CRA) opinion about the issuer’s ability and willingness to meet its obligations on time and in full. CRAs divide credit risk into discrete risk classes, expressing ratings in letterform, from AAA (highest credit quality/lowest credit risk) to C (lowest credit quality/highest credit risk) and D for default. For credit ratings to be informative on financial constraints, they should be informative about information-related problems, in addition to financial distress. In the next section, we ask whether this is a reasonable expectation.

2.2.1. Credit ratings and financial constraints

CRAs are in the business of mitigating information asymmetry. They exploit scale economies in information collection, processing, and production, thereby relieving investors of (duplication of) the cost of evaluating credit risk (Benston and Smith, 1976). CRAs may mitigate information asymmetry, as in the models of Leland and Pyle (1977) and Ramakrishnan and Thakor (1984), through the ratings analyst’s ability to verify the quality of private information while keeping such information confidential. As one example of the informativeness of the credit analysis, Kraft (2015) shows that the adjustments to financial statements that ratings analysts make as part of their company analysis are informative in that they are significantly associated with credit spreads. In line with an information-verification explanation, Sufi (2007) finds that the introduction of syndicated bank loan ratings by CRAs in 1995 allowed borrowers to approach less-informed lenders, such as foreign banks and nonbank institutional investors. Consistent with the idea that ratings incorporate private information, Czarnitzki and Kraft (2007) add credit rating to a bankruptcy-prediction model containing accounting information, thereby increasing model-R2 six-fold. Batta (2011) shows that the explanatory power of accounting information declines drastically when adding credit rating to a CDS-spread prediction model.

44

Agency costs arise from the limited liability feature of the corporate form of business organization that creates incentives for managers to sometimes act counter to the interests of creditors. Hence, CRAs are also in the business of monitoring creditors. They reappraise creditors regularly, as well as in the face of major corporate and external events. Accordingly, besides information verification, CRAs may fill a role as monitors of creditor behavior. The monitoring by ratings analysts and the threat of rating revision may mitigate moral hazard problems associated with risk shifting and shirking, as in the model of Boot, Milbourn, and Schmeits (2006). CRAs are not the only intermediaries monitoring managers, but Langohr and Langohr (2008) suggest that they are the main monitors of firms’ risk shifting at the expense of bondholders. Odders-White and Ready (2006) suggest, and find, that credit ratings mitigate equity-market adverse selection. In particular, firms at greater risk of experiencing unobservable (private) rather than observable (common) shocks will experience higher levels of equity-market adverse selection. Due to beneficial access to confidential information, ratings analysts can incorporate the risk of both private and common shocks into the rating, making ratings informative about equity-market adverse selection. Supporting this, Jory, Ngo, and Wang (2016) show that M&A premiums are lower for rated targets. On the same accord, Cheng and Subramanyam (2008) find an inverse relationship between number of equity analysts following a firm and the credit rating, while Ederington and Goh (1998) find that equity analysts react to rating changes by revising their earnings forecasts. He, Wang, and Wei (2011) show that the rating level is inversely related to several measures of equity-market adverse selection: privately informed trading, bid-ask spreads, dispersion of analysts’ earnings forecasts, and institutional equity holdings. They further find that as ratings improve, so do each of these information-asymmetry measures. Avramov, Chordia, Jostova, and Philipov (2013) similarly find that the worst rated firms have considerably weaker equity market liquidity, fewer analysts following, much larger analysts’ earnings revisions, and greater dispersion in analysts’ earnings per share forecasts. These results are in agreement with the findings of several recent studies such as Liu and Jiraporn (2010) and Qi, Roth, and Wald (2010).

Financial distress costs and adverse selection may have complementary effects on the cost of borrowing. Adverse selection can lead to a decline in the average creditworthiness of the pool of firms willing to borrow, as more creditworthy firms are driven out of market by larger credit spreads. Bond prices tend to be more information-sensitive when credit risk increases. For example, announcement effects of rating changes are stronger for speculative-grade than

45

investment-grade (Hand,Holthausen, and Leftwich,1992; Goh and Ederington, 1993; Dichev and Piotroski, 2001; Steiner and Heincke, 2001; see Jorion and Zhang, 2007, and Bannier and Hirsch, 2010, for exceptions). Han and Zhou (2014) show that increases in information asymmetry significantly influence credit spreads, and more so for high-yield than investment-grade bonds.

2.2.2. Criticism of credit ratings

Credit ratings are not without objection, which may explain why they have not previously been used as measures of financial constraints (with the exception of Hennessy (2004), who distinguishes between investment- and speculative-grade rated firms). CRAs have been accused of lack of timeliness.11 This is of relevance to us only insofar as it hurts the ability of ratings to identify constraints in a timely manner. Empirical evidence suggests that rating agencies are sometimes slow to respond to new information (Holthausen and Leftwich, 1986; Goh and Ederington, 1999; Steiner and Heinke, 2001; Hull, Predescu, and White, 2004; Norden and Weber, 2004; and Finnerty, Miller, and Chen, 2013; for exceptions, see Wansley, Glascock, and Clauretie, 1992; Goh and Ederington, 1993; and Hite and Warga, 1997).12 Still, it seems rating changes and, in particular, downgrades are informative to equity (Holthausen and Leftwich 1986; Hand, Holthausen, and Leftwich, 1992; Goh and Ederington, 1993, 1999; Dichev and Pietroski, 2001; Norden and Weber, 2004; Jorion Liu, and Shi, 2005; Jorion and Zhang, 2007; Bannier and Hirsch, 2010), 11 Measures of credit quality should balance three somewhat conflicting objectives: stability,

timeliness, and accuracy. Stability is a desirable trait in that more frequent rating changes would increase the frequency of bond-portfolio rebalancing and could exacerbate financial crises by forcing banks and investors to liquidate their positions as ratings decline to adhere to capital requirements, and in that rating reversals within a short period would hurt an agency’s reputation. Rating agencies use the through-the-cycle methodology. Through-the-cycle ratings are explicitly long-term oriented; rating agencies focus their rating assessment on the permanent credit-risk component by assessing expected performance at the bottom of a credit quality cycle. This stands in stark contrast to the point-in-time perspective that tends to be the focus of bankers and in credit scoring, where credit quality is judged on the basis of the actual position in the corporate credit cycle. The objective of through-the-cycle ratings is to measure default risk over long investment horizons, whereas point-in-time scoring is focused on current credit risk. In the through-the-cycle methodology ratings are changed only when agencies are confident that observed changes in a company’s credit quality are likely to be permanent. This will, by necessity, mean that ratings are perceived as less timely when evaluated relative to current credit quality. Also, credit risk is continuous, while ratings are ordinal, so we would expect some delay in ratings to changes in credit risk.

12 It should be noted that all studies that find significant pre-announcement effects also find significant announcement effects.

46

bond (Hand, Holthausen, and Leftwich, 1992; Wansley et al., 1992; Hite and Warga, 1997; Kliger and Sarig, 2000; Steiner and Heincke, 2001; May, 2010), CDS (Hull, Predescu, and White, 2004; Norden and Weber, 2004; Finnerty et al., 2013), and options (Kliger and Sarig, 2000) markets.13

Several studies find that bond ratings can be efficiently predicted using accounting and market information (e.g., Doumpos, Niklis, Zopounidis, and Andriosopoulos, 2015; Altman and Rijken, 2006), casting doubt on the information value of ratings. Hilscher and Wilson (2016) further show that agency ratings are a relatively poor predictor of corporate defaults compared to a credit-scoring model based on accounting data and stock market prices. The rating-agency response is that ratings should change only when fundamental credit risk changes, which, CRAs argue, generally happens quite slowly (see Cantor and Mann (2003) for further discussion). In short, the relative informativeness of credit ratings is, and continues to be a source of academic controversy. Our interest is absolute rather than relative informativeness, meaning that ratings may still be useful as constraints indicators, though not necessarily the best indicators.

A potentially more important weakness, for us, is that rated firms may be structurally different from unrated firms. Faulkender and Petersen (2006) and Lemmon and Roberts (2010) find that unrated firms are smaller, riskier, less profitable, and have lower proportions of tangible assets and higher market-to-book ratios than investment- and speculative-grade rated firms. They are also less levered and choose bank debt over bond financing. As a result, firms that lack a rating may face a more restricted set of alternative sources of financing and may be more sensitive to shocks to the banking system. Focusing on ratings means that we may miss the most constrained firms: those that lack debt capacity, those facing the greatest adverse selection, or those that cannot 13 In contrast to downgrades, research finds either an insignificant or a weak market reaction to

rating upgrades. Ederington and Goh (1998) argue that the lack of market impact of rating upgrades could be due to companies voluntarily releasing good news to the market but being reluctant to release unfavorable information. This would create a bias towards negative information content for credit rating changes. This could also explain the stronger market response to rating changes for speculative-grade than to investment-grade firms. Alternatively, rating agencies could be more attentive to deteriorations in credit quality. Jorion and Zhang (2007) instead refer to the ordinal character of the rating scale in comparison to the exponentially increasing default frequency across the rating scale, where a one-notch rating change represents a larger change in probability of default the lower the initial rating. They find that controlling for the initial rating significantly reduces the difference in market response to up- and downgrades.

47

afford a rating. On the other hand, lack of a rating is not synonymous with being financially constrained. Faulkender and Petersen (2006) report that public debt is not a major source of capital for many public firms, so lack of rating could also be a result of lack of need for public debt financing, such as due to managerial conservatism, limited investment needs, or abundant access to alternative sources of financing. In addition, Farre-Mensa and Ljungqvist (2016) show that rated and unrated firms respond similarly to exogenous shocks to supply of debt and equity financing.

A second more potent risk is that of endogeneity: that financial constraints drive the rating decision rather than ratings informing us about constraints. A financially constrained firm may see benefits in targeting a better-than-optimal rating to mitigate the costs of financial constraints. Alternatively, financial constraints may force the firm to underlever, underinvest, and rely more on equity-based financing solutions, all of which may influence its rating favorably. Firms seem to attach a lot of value to the rating in their risk taking. Graham and Harvey (2001) find that the credit rating, next to financial flexibility, is the most important determinant of a firm’s financing policy. Kisgen (2006) finds that firms close to a rating up- or downgrade tend to issue less debt than other firms; Kemper and Rao (2013) show that this is particularly the case for firms with low (≤ B-) ratings. Kisgen (2009) finds evidence suggesting that firms reduce leverage following rating changes to protect the newly assigned rating, with a more pronounced effect for downgrades from investment to speculative grade, while Khieu and Pyles (2012) find that firms increase their holdings of cash and, in particular, excess cash following rating downgrades. Kuang and Qin (2013) show that downgraded firms reduce managerial risk-taking incentives by lowering performance sensitivity (vega) of new executive options grants. Bannier and Hirsch (2010) find that speculative-grade issuers, in particular, reduce risk-taking following a credit warning. Yet another possibility is that financially constrained firms are overrepresented among those firms who decide not to have their rating made public (the issuer-pay rating model). The validity of this possibility is reduced by the fact that rated and unrated firms seem to respond similarly to exogenous shocks to supply of debt and equity financing (Farre-Mensa and Ljungqvist, 2016). Endogeneity is shared by all constraints proxies. Following Farre-Mensa and Ljungqvist (2016), we take any endogeneity as given and focus on assessing whether firms classified as constrained based on the rating seem to behave as if they are constrained. In other words, the proof of the pudding is in the eating.

48

2.3. Sample and empirical method

In this section, we present our sample and the different tests we employ to evaluate the ability of ratings to capture financial constraints.

2.3.1. Sample and variables

Our data consists of all US-incorporated industrial firms that have an issuer credit rating from Standard & Poor’s (S&P) at some point between 1987 and 2014.14 We collect long-term domestic issuer credit ratings (Compustat code splticrm) from the Compustat monthly file. Since we use yearly data for the other variables, we keep the last rating for each fiscal year. Financial data are retrieved from Compustat’s annual North American Datafile. To avoid mismatches between calendar and fiscal years, we match the Compustat datafiles based on fiscal year-end. All variables are defined in detail in the Appendix. We exclude utilities (SIC 4900s) and financial firms (SIC 6000s). We use customary sample restrictions (e.g., Chen and Chen, 2012). To avoid potential business discontinuities caused by mergers and acquisitions, we exclude firm-years with asset or sales growth exceeding 100%. To mitigate outliers, we require the firm’s total assets and sales to be greater than $1 million in every year and we winsorize observations at the 1st and 99th percentiles in each year. The final sample consists of 944 firms and 17,950 firm-year observations (though lagging certain variables as well as gaps in the panel structure of many firms reduce the sample size used in our modelling).

Columns 1 and 2 in Table 1 contain descriptive statistics for our sample firms. Faulkender and Petersen (2006) show that firms with a rating are fundamentally different from firms without a rating, not least by being much larger, older, and more profitable. So too for our sample. We compare the characteristics of our sample to those of a Compustat sample covering the same sample period and using the same sample restrictions, but without a credit rating (Column 3, Table 1).

14 Issuer credit ratings reflect the rating agency’s assessment of a firm’s ability and willingness

to meet its senior unsecured obligations on a timely basis. Strictly speaking, the empirical results refer only to the ratings of Standard and Poor’s. We are not aware of any reason why the empirical results and the conclusions presented here for Standard and Poor’s ratings should not apply for the ratings of Moody’s and Fitch. Güttler (2005) compare the ratings of S&P, Moody’s, and Fitch and concludes that assigned ratings on average differ by less than one notch.

49

Our sample firms are sizably larger and older. They are more profitable, pay larger dividends, and generate on average almost three times as large cash flows. The average sales growth rate is twice as high, whereas the R&D investment rate of rated firms is 40% that of nonrated ones. There is hardly any difference in capital expenditures, while the average Tobin’s Q is 70% of the average nonrated firm-year. Further, our sample firms hold more tangible assets (34% of total assets being property, plant, and equipment, compared to 27% for the non-rated sample) and are more levered (debt-to-assets of 34% vs. 23%), but hold less cash (cash-to-assets of 8% vs. 18%). In conclusion, our sample is not representative for nonrated firms. On the other hand, it is not expected nor intended to be. Our focus is the information content of ratings, which will by necessity reduce the representativeness.

We compare the ability of ratings to identify financial constraints to six other constraints classifiers suggested in the literature: the KZ, SA, and WW scores and Hoberg and Maksimovic’s (2015) Delay Investment (DI), Equity-Focused (EF), and Debt-Focused (DF) scores. Kaplan and Zingales (1997) classify 10-Ks into five discrete categories of financial constraints and then use ordered logit regression to relate their categorization to accounting variables. In Lamont et al. (2001), Kaplan and Zingales re-estimate their model and relate their categorization to five publicly available variables: cash flow, Tobin’s Q, leverage, dividends, and cash holdings. In line with other studies, we employ this re-estimated KZ score. Hadlock and Pierce (2010) extend the sample size and qualitative input relative to Kaplan and Zingales (1997). They also apply ordered logit regression to derive a financial constraints score, the SA score, where they relate their categorization to size and age. They show that the SA score outperforms the KZ score in identifying companies that show signs of being financially constrained. Whited and Wu (2006) criticize the qualitative approach of Kaplan and Zingales (1997) and Hadlock and Pierce (2010) and construct their WW score by using a structural model in which they relate investment to the cost of external financing. Whited and Wu (2006) criticize the qualitative approach of Kaplan and Zingales (1997) and Hadlock and Pierce (2010) and construct their WW score by using a structural model in which they relate investment to the cost of external financing.

50

Tab

le 1

. F

irm

ch

arac

teri

stic

s an

d f

inan

cial

co

nst

rain

ts c

lass

ific

atio

n

The

tabl

e re

port

s de

scrip

tive

stat

istic

s fo

r th

e fu

ll sa

mpl

e an

d th

e sa

mpl

e pa

rtiti

oned

into

fina

ncia

l con

stra

ints

cla

sses

. The

firs

t tw

o co

lum

ns r

epor

t ful

l-sam

ple

mea

ns a

nd

stan

dard

dev

iatio

ns.

The

third

col

umn

repo

rts

mea

ns fo

r a

sim

ilarly

def

ined

sam

ple

of n

on-r

ated

firm

-yea

rs.

The

rem

aini

ng c

olum

ns r

epor

t di

ffere

nces

bet

wee

n in

dust

ry-

adju

sted

gro

up m

eans

for

the

mos

t (w

orst

rat

ings

/hig

hest

sco

res)

min

us t

he le

ast

(str

onge

st r

atin

gs/lo

wes

t sc

ores

) co

nstr

aine

d fir

m-y

ears

. W

e so

rt f

irms

into

fou

r ra

ting

clas

ses

(A o

r hi

gher

, B

BB

, B

B,

and

B o

r lo

wer

), w

hile

the

com

paris

on s

core

s ea

ch y

ear

are

split

into

qua

rtile

s w

ith h

igh

scor

es r

epre

sent

ing

mor

e fin

anci

al c

onst

rain

ts.

All

varia

bles

exc

ept

indu

stry

Q a

re c

ompu

ted

as a

firm

’s r

aw v

alue

for

the

spe

cific

var

iabl

e m

inus

the

indu

stry

val

ue

with

indu

stry

def

ined

by

two-

digi

t S

ICs.

Rea

l siz

e is

ex

pres

sed

in m

illio

n U

SD

for

the

full

sam

ple

mea

n an

d st

anda

rd d

evia

tion,

and

in lo

gs fo

r di

ffere

nces

in in

dust

ry-a

djus

ted

grou

p m

eans

.

F

ull s

ampl

e N

o ra

ting

sam

ple

(3)

Rat

ing

KZ

sco

re

SA

sco

re

WW

sco

re

DI s

core

E

F s

core

D

F s

core

Dep

ende

nt v

aria

ble

Mea

n(1)

S

td d

ev (

2)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Rea

l siz

e (m

n/lo

g)

8,39

6 18

,299

58

1 -1

.744

***

-0.5

34**

* -2

.225

***

-3.0

89**

* -0

.033

0.

173**

* -0

.437

***

Age

(ye

ars)

26

.656

23

.138

13

.450

-1

4.68

2**

* -5

.949

***

-46.

557**

* -2

3.49

3***

-4.7

80**

* -1

.960

**

-4.0

63**

*

Q

1.36

2 0.

845

1.89

6 -0

.585

***

-0.5

85**

* 0.

195**

* 0.

004

0.06

0**

0.13

4***

-0.3

58**

*

Indu

stry

Q

1.51

4 0.

415

1.93

8 -0

.151

***

-0.3

36**

* 0.

023**

-0

.023

**

-0.1

02**

* -0

.060

***

-0.1

58**

*

Sal

es g

row

th

0.10

2 0.

200

0.02

9 -0

.392

***

-0.3

97**

* -0

.218

***

-0.2

73**

* -0

.236

***

-0.1

21**

* -0

.174

***

Cap

ital e

xpen

ditu

res

0.06

8 0.

070

0.06

1 -0

.032

***

-0.0

04*

0.00

0 0.

000

-0.0

12**

* -0

.010

***

0.00

5**

R&

D

0.03

7 0.

051

0.08

9 -0

.007

***

-0.0

13**

* 0.

008**

* 0.

006**

* 0.

011**

* 0.

014**

* -0

.017

***

Ret

urn

on a

sset

s 0.

162

0.09

6 -0

.062

-0

.069

***

-0.0

53**

* 0.

016**

* -0

.013

***

-0.0

05

-0.0

06*

-0.0

17**

*

Ope

ratin

g ca

sh fl

ow

0.10

7 0.

072

0.02

2 -0

.063

***

-0.0

45**

* 0.

003

-0.0

21**

* 0.

001

0.00

0 -0

.024

***

Leve

rage

0.

342

0.23

8 0.

251

0.26

2**

* 0.

213**

* 0.

044**

* 0.

109**

* 0.

038**

* 0.

004

0.08

6***

Cas

h ho

ldin

gs

0.08

2 0.

090

0.15

3 0.

009**

* -0

.079

***

0.01

7***

0.02

0***

0.00

5* 0.

017**

* -0

.058

***

Div

iden

ds

0.01

5 0.

025

0.00

7 -0

.020

***

-0.0

28**

* -0

.009

***

-0.0

18**

* -0

.001

0.

000

-0.0

04**

*

51

The predictive variables in their model are cash flow, a dividend payment dummy, leverage, size, and firm and industry sales growth. Following convention in the literature, we calculate the KZ, SA, and WW scores using the originally estimated coefficients. Hoberg and Maksimovic (2015) also score 10-K text, but extend the sample size 26-fold relative to Hadlock and Pierce (2010) to cover the entire Compustat universe by employing automated textual analysis. Hoberg and Maksimovic additionally differentiate between various sources of financial constraints, where the DI score measures each firm’s degree of overall investment constraints, while the EF and DF scores measure financial constraints uniquely faced by firms intending to issue equity or debt. The text processing software converts the number of constraint-related words used into a continuous score, where a higher score translates to a higher degree of financial constraints. We obtain the raw DI, EF, and DF scores directly from the authors.

The original rating scale is converted into numbers, ranging from 1 (AAA) to 20 (CC). Since we have few observations with ratings below CC, we exclude all firm-years with such ratings. We also group the rating scale into four broader groups. Rating4 is a four-way sort where we sort investment- and speculative-grade rated firms into two classes each, separating between firms rated A or higher, BBB, BB, and B or lower.15 We resort the sample annually based on the beginning-of-year rating. The sample distribution across rating categories is found in Table 2. We compare the ability of the 20-step rating scale to capture financial constraints to the continuously distributed KZ, SZ, WW, DI, EF, and DF comparison scores. We also each year and for each comparison score distribute the firms equally into four groups from least (score of one) to most (score of four) constrained. This sorting procedure deviates slightly from the ratings distribution. In our full sample, 56% (44%) of firm-years carry an investment grade (speculative grade) rating, while Rating4 is distributed 27% in group 1, 29% in group 2, 26% in group 3, and 18% in group 4.

15 The general interpretations of the issuer-rating categories provided by Standard and Poor’s

are: AAA/AA: extremely/very strong capacity to meet financial obligations; A: strong capacity to meet financial obligations; BBB: adequate capacity to meet financial obligations; BB: less vulnerable in the near term than lower-rated obligors, but faces major ongoing uncertainties and exposure to adverse conditions; B: current capacity, but adverse conditions will likely impair capacity or willingness to meet financial commitments; CCC: currently vulnerable and dependent on favorable conditions to meet financial commitments; CC/C: currently highly/very highly vulnerable.

52

Table 2. Classification and distribution of ratings The table shows our two rating classification scales: Rating, which uses the full notch-level rating scale, and Rating4, which is a four-way sort, where we sort the investment- and speculative-grade firms into two classes each (A or higher, BBB, BB, and B or lower). The table also shows the frequency of firm-year observations in each rating category.

Rating Rating4 Frequency (%) Rating Rating4 Frequency (%)

AAA 1 1 1.54 BB+ 11 3 6.75

AA+ 2 1 0.51 BB 12 3 9.04

AA 3 1 2.21 BB- 13 3 10.17

AA- 4 1 2.59 B+ 14 4 8.85

A+ 5 1 4.78 B 15 4 6.17

A 6 1 9.17 B- 16 4 2.20

A- 7 1 5.96 CCC+ 17 4 0.61

BBB+ 8 2 8.32 CCC 18 4 0.22

BBB 9 2 11.48 CCC- 19 4 0.05

BBB- 10 2 9.33 CC 20 4 0.06

2.3.2. Empirical framework

We evaluate the ability of credit ratings, as well as the comparison scores, to separate more from less financially constrained firms in four ways. First, we apply descriptive statistics on operational and financial health to assess the success of credit ratings relative to the comparison scores in separating corporate characteristics in ways that are reasonable for constrained vs. unconstrained firms.

Second, we employ the tests suggested by Bodnaruk et al. (2015) to evaluate if different constraints classifiers predict corporate behavior indicative of improvement or deterioration in external financial constraints, what Bodnaruk et al. refer to as liquidity events. Bodnaruk et al. consider four liquidity events where constrained and unconstrained firms would be expected to behave differently: dividend omissions, dividend increases, equity recycling, and underfunded pension plans. We define the dependent variables in these tests as follows:

o Dividend omission: a dummy variable set to one if the firm stops paying dividends. Only firms that paid dividends in the prior year are assigned a value. Both theoretical (e.g., Miller and Rock, 1985) and empirical (e.g., Michaely, Thaler, and Womack, 1995) studies find that dividend omissions are very important, costly signals about firm value. Following Bodnaruk et al. (2015), we expect that the willingness to

53

pay the cost of dividend omissions would increase in financial constraints.

o Dividend increase: a dummy variable set to one if the firm increases its dividend per share relative to the previous fiscal year. Only firms that paid dividends in the prior year are assigned a value. Since dividend reductions are costly, managers are reluctant to increase dividend payments if there is a risk that the increase will have to be reversed in coming years (e.g., Brav, Graham, Harvey, and Michaely, 2005), and the more financially constrained the firm is, the greater the reluctance to increase dividends.

o Equity recycling: the ratio of cash dividends plus purchases of common and preferred stock to the sale of common and preferred stock and normalized by beginning-of-period total assets. As pointed out by Farre-Mensa and Ljungkvist (2016), the cost of simultaneously raising and paying out equity could increase for financially constrained firms.

o Underfunded pension plan: a dummy variable set to one if the firm has an underfunded pension plan (projected pension benefit obligations greater than pension plan assets). Only firms that report pension obligations and pension plan assets are assigned a value. Rauh (2006) finds that mandatory contributions to defined benefit pension plans deprive firms of funds required to finance investments, and Franzoni (2009) shows that the stock market response to such contributions is inversely related to financial constraints.

We estimate logit models for dividend omissions, dividend increases, and underfunded pension plans and use panel regressions for equity recycling to investigate if ratings and the comparison scores predict each of these liquidity events. Following Bodnaruk et al. (2015), we in all regressions include as control variables the natural logarithm of market capitalization, Tobin’s Q, the excess prior year buy-and-hold stock return, and period and industry fixed effects.

Third, following Fazzari et al. (1988) we estimate investment-cash flow (ICF) sensitivity using an extended Q-model: = + + + + + × + ′ + (1)where Ii is the firm’s capital expenditures deflated by its beginning-of-period total assets, Qit-1 is Tobin’s Q, which proxies for investment opportunities, Cit is the firm’s cash flow from operations deflated by beginning-of-period total

54

assets, Fit-1 is the firm’s credit rating (or comparison score), and X are control variables. We calculate Qit-1 as the market value of equity plus the liquidation value of preferred stock plus long- and short-term debt minus deferred taxes and investment tax credits and divided by total assets. To remove time-invariant firm characteristics, we include firm-fixed effects, αi, and to weed out firm-invariant macro shocks, we include period-fixed effects, αt. We address the concern that cash flow may be picking up on investment opportunities when used in conjunction with Q by employing Lewellen and Lewellen’s (2016) IV estimator, where Q is instrumented by contemporaneous, lagged, and squared cash flow and the first four lags of stock returns. To avoid the distortionary effects of negative cash flows (e.g., Allayannis and Mozumdar, 2004; Cleary Povel, and Raith, 2007), we only include firm-years with positive cash flow.

The underlying idea of ICF sensitivity as a financial constraints test is that constrained firms are expected to more systematically prioritize financing investments with internally generated cash flow so as to avoid the excess cost of external financing (Fazzari et al., 1988). Unconstrained firms, on the other hand, could substitute internal for external financing at no or lower cost, and so would not be expected to systematically prioritize internal financing. ICF

sensitivity is given by the expression = + . The main parameter

of interest is β4, which captures the impact of ratings or the comparison scores on ICF sensitivity. If a constraints’ classifier measures financial constraints, then β4 > 0.

The critical question is if ICF sensitivity is a useful measure of financial constraints. This has been one of the dominant debates in corporate finance over the last three decades. Kaplan and Zingales (1997) argue that there is no strong theoretical argument for expecting ICF sensitivity to increase monotonically with financial constraints, a position that Fazzari, Hubbard, and Petersen (2000) forcefully refute. Chen and Chen (2012) criticize the usefulness of ICF sensitivity by showing that it has disappeared in recent years and even during the financial crisis of 2007-2009. In their words, “if one believes that financial constraints have not completely disappeared, then investment-cash flow sensitivity cannot be a good measure of financial constraints.” (ibid: 394) Lewellen and Lewellen (2016) show that Chen and Chen’s results are sample-specific; using a cleaner measure of cash flow (operating cash flow) and extending the sample to manufacturing and non-manufacturing industrial firms, they find ICF sensitivity to be alive and kicking. Moshirian et al. (2017) show that ICF sensitivity seems to be a

55

function of investment intensity rather than financial constraints. Andrén and Jankensgård (2018) reinterpret their results as suggesting that one needs to distinguish between the need for investment financing (investment intensity) and the cost of external financing (financial constraints). They show that ICF sensitivity is a valid measure of financial constraints once you control for need for investment financing.

Fourth, we evaluate how firms respond to exogenous capital-supply shocks. Here, we ask if ratings and the comparison constraints classifiers discriminate firms in ways consistent with how we expect firms with varying levels of financial constraints to respond to capital-supply shocks. The basic premise for evaluating responses to shocks to capital supply is that the corporate responses will reflect managerial perceptions of the elasticity of capital supply. An unconstrained firm should not be affected by shocks to capital supply since it could easily shift to other sources of capital. A constrained firm would face an inelastic supply curve, thus finding it more costly, or even impossible to substitute across sources of capital and would, as a consequence, curtail investments more in the face of unexpected shocks. The critical challenge is finding shocks to capital that are not accompanied by shocks to investment opportunities.

Our first shock is the global financial crisis (GFC) of 2007-09. The crisis began in the summer of 2007 when the asset-backed commercial paper market began to unravel (Acharya and Schnabl, 2010). Contrary to other financial disruptions, the financial crisis primarily had its roots outside of the real sector, but transferred to the production side of the economy in the second half of 2008. The financial part of the crisis truly abated only in the spring of 2009 when the stress tests of the large US banks brought private capital back into the financial system. There is no possibility to fully isolate supply from demand effects, but by concentrating on the financial part of the crisis, we at least try to reduce the impact of the g FC on the demand for investments (also see Hoberg and Maksimovic, 2015). The crisis hurt bank lending as well as bond markets. Corporate bond yield spreads began climbing in July 2007, especially for speculative grade bonds (Friewald, Jankowitsch, and Subramanyam, 2013), and for industrial and financial firms alike (Dick-Nielsen, Feldhütter, and Lando, 2012). We model changes in capex between 2006 (the last year prior to the shock) and 2009 (after the financial part of the crisis abated) as a function of lagged financial constraints, with lags measured in 2006. Our expectation is that constrained firms cut back on investments

56

more aggressively, as financial constraints become more binding in times of crisis.

Our second shock is the introduction of the Jobs and Growth Tax Relief Reconciliation Act (JGTRRA), which was signed into law on May 28, 2003. JGTRRA reduced shareholder-level tax rates on dividends and capital gains.16 It represents a strong experiment, for several reasons (Brown, Liang, and Weisbenner, 2007). First, it was the first major reduction in dividend and capital gains taxation in the US in 17 years. Second, there were no other major confounding tax law changes surrounding the introduction of JGTRRA. Third, the tax cuts were largely exogenous. The legislative timeline was short, with the initial proposal floated by the Bush administration in late December 2002 and a modified proposal signed into law on May 28, 2003. The uncertainty of whether the bill would pass was high; the modified version of the proposal passed both the House of Representatives (231-200) and the Senate (50-50 with Vice President Cheney breaking the tie with a “yea” vote) on May 23, 2003. This gave firms little time or incentive to alter their financial and investment policies beforehand. The tax rate cuts on dividends were effective retroactively to January 1, 2003, and the rate cuts on capital gains were effective on sales of stock after May 6, 2003.

Inspired by Campbell, Chyz, Dhaliwal, and Schwartz (2013), we investigate changes in capex after the JGTRRA, with the dependent variable being the change in capex between 2002 and 2004. One of the stated goals of JGTRRA was to encourage corporate investment by, in particular, reducing the cost of equity capital. This should be more favorable to firms facing greater cost wedges between internal and external financing and, in particular, equity-dependent firms. Consistent with this, Dai, Shackelford, Zhang, and Chen (2013) argue and show that the reduction in cost of equity was greater for financially constrained firms.

16 The individual tax rate on dividends decreased from a maximum rate of 38.6 percent to 15

percent, the individual tax rate on long-term capital gains decreased from 20 percent to 15 percent, and the individual tax rate on interest income decreased from a maximum rate of 38.6 percent to 35 percent. For individuals, interest income is taxed at the ordinary income tax rate. The maximum statutory individual income tax rate bracket pre-JGTRRA was 38.6 percent, and JGTRRA reduced the top bracket to 35 percent. In addition, the second highest bracket was reduced from 35 percent to 33 percent, the third highest bracket from 30 percent to 28 percent, and the fourth highest bracket from 27 percent to 25 percent.

57

2.4. Results

2.4.1. Corporate characteristics and constraints classification

We begin our analysis by assessing the association between financial constraints classifications and firm characteristics. This is the dominant approach to assessing constraints classifications; still, it is imprecise in the sense that many characteristics have unclear connections with financial constraints due to endogeneity, and so will only be indicative. In Table 1, we present descriptive statistics for ratings and the six comparison scores sorted into four groups. We sort ratings into four classes (A or higher, BBB, BB, and B or lower), while the comparison scores each year are split into quartiles with high scores representing more financial constraints. The table reports differences between industry-adjusted group means for the most (worst ratings/highest scores) minus the least (strongest ratings/lowest scores) constrained firm-years.

Firms with worse ratings are on average smaller, younger, and less profitable, and they generate smaller operating cash flows and pay smaller dividends. They invest less, both in capital expenditures (capex) and research and development (R&D). These patterns are consistent with less creditworthy companies being more financially constrained. Worse-rated firms also grow at a slower pace. This goes against Hoberg and Maksimovic (2015) and Farre-Mensa and Ljungkvist (2016), who find that firms classified as most constrained are high-growth firms (it should be emphasized that Farre-Mensa and Ljungkvist strongly refute the usefulness of these classifiers for identifying constrained firms). On the other hand, growth subsumes investment, and as argued by Hoberg and Maksimovic (2015), constrained firms would be expected to underinvest relative to what would be the case if they were unconstrained. This, in turn, would suggest relatively slower (industry-adjusted) growth, which is what we find.

Average Tobin’s Q is lower for worse-rated firms. Whether or not this is consistent with worse rated firms being constrained is open to debate. On the one hand, as constrained firms face higher marginal costs of external financing, they would be expected to leave better investments on the table. In line with this, Hadlock and Pierce (2010), Hoberg and Maksimovic (2015), and Farre-Mensa and Ljungkvist (2016) find firms classified as constrained to have higher Q. On the other hand, more costly external financing should translate into higher marginal costs of capital, which would tend to reduce the present

58

value of all expected future cash flows, including those from assets-in-place. This would tend to depress market valuations of constrained firms, and thus be suggestive of relatively lower Q. Ascioglu, Hegde, and McDermott (2008) use measures of asymmetric information from the market microstructure literature (bid-ask spread, relative price impact, and probability of informed trading) to classify firms on financial constraints. Such measures capture more narrow aspects of equity-market information constraints, in contrast to more general-purpose type classifications such as ratings and the comparison scores, but have the advantage of more unambiguously capturing constraints. Using their narrow but more direct constraints proxies, they find constrained firms to have lower Q.

Firms with worse ratings are more levered. Since leverage is a choice variable, constrained firms could be argued to be either more or less levered than unconstrained ones. Leverage may be constraining in that it commits internal cash flows, reduces creditworthiness, and gives rise to agency conflicts with creditors (risk shifting and debt overhang problems). Accordingly, leverage weighs positively in the KZ and WW scores. On the other hand, a financially constrained firm may prioritize maintaining precautionary debt capacity or may face constrained access to debt financing, which would suggest that constrained firms are less levered than unconstrained ones. Using several commonly used constraints classifiers, Farre-Mensa and Ljungkvist (2016) find firms classified as constrained to be less levered.

Firms with worse ratings also hold more cash. Again, expectations are unclear. Kaplan and Zingales (1997) consider small cash holdings an indication of financial constraints, while Hadlock and Pierce (2010) and Almeida et al. (2004) argue that constrained firms could be expected to hold precautionary excess cash reserves.

We consider ratings to generate a reasonable partitioning of the sample firms (perhaps with the exception of Q), and even more so than the comparison scores. The two classifications that generate partitionings most resembling those of credit ratings are the KZ and DF scores, which only differ from ratings on capex and cash holdings. It is reasonable that these two classifiers resemble ratings; Kaplan and Zingales (1997) emphasize financial status when classifying firms, and Hoberg and Maksimovic (2015) find firms that score high on the DF score to show indications of being distressed. Still, we consider the partitioning of ratings the better one, since we believe it is more reasonable to expect more constrained firms to invest less than unconstrained ones.

59

Farre-Mensa and Ljungkvist (2016), using a sample of more than 91,000 public US firm-years, find the SA and WW scores to classify similar kinds of firms as constrained: smaller and younger, less profitable but faster growing, more liquid but less levered, and investing more in R&D and having significantly higher Q ratios. Rather than capturing financial constraints, Farre-Mensa and Ljungkvist suggest that the SA and WW scores identify firms in the growth phase of their life cycle. We, using our sample of public US firms with a credit rating, also find the SA and WW scores to identify as constrained firms that are smaller, younger, and more liquid, but in contrast to Farre-Mensa and Ljungkvist, they grow at a slower pace and are more levered. This, to us, would be more consistent with these firms actually being financially constrained. However, we also find the SA score to identify as constrained firms that are more profitable, invest more in R&D, and have higher Q, which we (perhaps with exception for Q) find more difficult to reconcile with financial constraints. The WW score performs slightly better in that it finds firms identified as constrained to be less profitable and generate smaller operating cash flows, though they invest more in R&D.

Hoberg and Maksimovic (2015), using a sample of 44,000 public US firm-years, find the DI score to identify as constrained firms that are less profitable and have higher Q, invest less in capex and R&D, and are more levered and pay smaller dividends. These are all (again with a possible exception for Q) reasonable constraints traits. Our sample generates a partly different partitioning for the DI score. We also find DI-constrained firms to be more levered and to invest less in capex, but they show no difference in terms of size, profitability, and dividends, and they invest more in R&D. Hoberg and Maksimovic find the EF score to result in a similar, but more extreme identification as the DI score. Our results support this. Like Hoberg and Maksimovic, we find firms classified as constrained by the EF score to have higher Q, be more levered, and invest less in capex, but we also find them to invest more rather than less in R&D and to generate operating cash flows as large as those of unconstrained firms. Furthermore, they are larger than unconstrained firms.

Credit ratings yield more marked separations of the sample firms than the comparison scores on most characteristics. Ratings yield the largest mean spreads on Q, investment in capex, profitability, operating cash flows, and leverage, and are second only to the KZ score on sales growth and dividends. This suggests – given that the direction of the separation is reasonable – that

60

ratings have the potential to achieve a clearer separation of constrained vs. unconstrained than the comparison scores.

The previous assessment is based on comparisons of extremes. This should allow us to more clearly reveal any patterns in the data, but at the expense of neglecting all the information in the middle of the sample distribution. As an alternative, we in Table 3 present a correlation matrix between the full credit rating/continuous comparison scores and the corporate characteristics. Correlations between ratings and characteristics verify what we see in Table 2: worse rated firms tend to be smaller, younger, less profitable, pay lower dividends, have lower Q, and be more levered. Correlations are small for capex and R&D investments and for cash holdings. Again, ratings seem to yield more marked separations, as correlations for the comparison scores are typically much lower.

Table 3. Correlation matrix between constraints classifiers and corporate characteristics The table reports correlations between industry-adjusted corporate characteristics and financial constraints measures. All variables except industry Q are computed as a firm’s log raw value for the specific variable minus the log industry value with industry defined by two-digit SICs.

Dependent variable Rating KZ score SA score WW score DI score

EF score

DF score

Real size -0.517 -0.036 -0.455 -0.809 -0.022 0.050 -0.162

Age -0.312 0.023 -0.829 -0.472 -0.120 -0.050 -0.081

Q -0.312 -0.143 0.032 -0.043 -0.007 0.031 -0.152

Sales growth -0.143 -0.073 -0.094 -0.066 -0.136 -0.110 -0.039

Capital expenditures -0.031 0.083 0.094 0.092 -0.019 0.000 0.013

R&D -0.042 -0.023 -0.001 -0.004 0.033 0.060 -0.139

Return on assets -0.388 -0.198 0.030 -0.118 -0.036 -0.025 -0.124

Operating cash flow -0.306 -0.146 0.034 -0.112 -0.004 0.021 -0.165

Leverage 0.378 0.092 0.221 0.295 0.123 0.049 0.132

Cash holdings 0.077 -0.265 0.046 0.052 -0.013 0.004 -0.160

Dividends -0.383 -0.466 -0.227 -0.319 -0.083 -0.033 -0.067

Credit ratings are highly correlated with the SA and WW scores (ρSA = 0.45 and ρWW = 0.65), but less so with the remaining scores (ρ ranging from 0.06 (EF) to 0.23 (DF)). These correlations suggest that ratings should share characteristics with the SA and WW scores rather than with the KZ, DI, EF,

61

and DF scores. The main, shared characteristics seem to be size and, to a lesser extent, age. Size is an input to both the SA and WW scores, and is also a strong predictor of credit ratings, whereas the correlation with size is clearly more limited for the KZ, DI, EF, and DF scores. As a further step in assessing the overlap in sample partitioning between credit ratings and the comparison classifiers, we in Figure 1 report average (standardized) comparison scores across the rating scale. There is a positive relationship between ratings and each of the comparison scores, with the scores increasing by 0.10 to 0.17 standard deviations per rating notch. The SA, WW, and DF scores increase near-monotonically across most of the rating scale. The pattern is slightly less monotonic for the KZ score, where the relationship is almost flat for the AA to B classes, and even less so for the DI and EF scores.

Figure 1. Credit ratings vs. alternative financial constraints classifications The diagram shows the average score for each rating class, with ratings classified from AAA to CC. Scores are standardized by subtracting the mean score and dividing by the standard deviation. KZ is the Kaplan and Zingales score as estimated in Lamont et al. (2001). SA is the Hadlock and Pierce (2010) score. WW is the Whited and Wu (2006) score. DI (EF, DF) score is the delay investment (equity financing, debt financing) score from Hoberg and Maksimovic (2014).

2.4.2. Predicting corporate liquidity events

Ratings seem to generate a reasonable partitioning of firms, but because of endogeneity a comparison of constraints classifiers on corporate characteristics is primarily indicative. To derive direct evidence concerning the ability of the investigated classifiers to separate constrained from unconstrained, we in this

62

section turn to assessing the ability of the classifiers to predict future corporate behavior consistent with improvement or deterioration in financial constraints. Following Bodnaruk et al. (2015), we investigate four informative liquidity events: dividend omissions, dividend increases, equity recycling, and underfunded pension plans. We estimate logit models for dividend omissions, dividend increases, and underfunded pension plans and use panel regressions for equity recycling. The main independent variable in each regression is each firm’s rating/constraints score for the previous fiscal year. Results are presented in Table 4. Panel A presents full results for models using credit ratings as the independent variable, whereas we in Panel B compress results by only presenting estimates for the regression coefficients on the comparison scores.

Credit ratings are significant predictors for all liquidity events, and with the expected signs. The worse the rating, the more likely the firm is to omit dividends. Weaker ratings also imply a smaller likelihood to recycle equity and a greater likelihood to have an underfunded pension plan. These results deviate markedly from those of the six comparison scores. None of the comparison scores predict all four liquidity events. The KZ, WW, and DF scores accurately predict one of the liquidity events, while the SA, DI, and EF scores predict two of them accurately. The EF score accurately predicts dividend omissions and having an underfunded pension plan, but generates a counterintuitive result for dividend increases by suggesting that more constrained firms are more likely to increase dividend payments.

In Table 4, we use as independent variables the full rating scale and continuous comparison scores. As an alternative, we sort firms into four categories and replace the full rating scale/continuous comparison scores with dummy variables for each of the four classes. Results are presented in Appendix 2. We focus our analysis on differences between the least and most constrained classes. Ratings still come out significant and with the expected signs for all four liquidity events. The KZ score is also significant for all four liquidity events and with the correct signs. Remaining comparison scores perform less well. The other scores predict two liquidity events accurately. The SA score also predicts two events accurately, but yields counterintuitive predictions on dividend increases and equity recycling.

63

Table 4. Corporate liquidity events The table reports results from logit regressions of dividend omissions (dummy variable set to 1 if the firm stops paying dividends), dividend increases (dummy variable set to 1 if the firm increases its dividend per share relative to the previous fiscal year), and whether the firm has an underfunded pension plan (dummy variable set to 1 if the firm has an underfunded pension plan), and panel regressions of if the firm recycles equity (the ratio of cash dividends plus purchases of common and preferred stock to the sale of common and preferred stock and normalized by beginning-of-period total assets). The main test variable in each regression is the score on a financial constraints classifier. In Panel A, we report full model results for credit ratings scored from 1 (AAA) to 20 (CC). In panel B, we report estimated coefficients for the KZ, SA, WW, DI, EF, and DF scores. We in all regressions include as control variables the natural logarithm of market capitalization, Q, the excess prior year buy-and-hold stock return, a dummy variable set to one if the firm reported negative earnings in the previous fiscal year, and period and industry fixed effects. In the model for underfunded pension plan we additionally include a dummy variable set to one in the firm had an underfunded pension plan in the previous fiscal year. T-statistics in parenthesis. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Panel A Dividend omission

Dividend increase

Underfunded pension plan

Equity recycling

Rating 0.439*** -0.189*** 0.064*** -0.001***

(11.59) (-10.24) (2.77) (-4.87)

Log(market capitalization) -0.062 0.039 -0.004 -0.006***

(-0.72) (1.22) (-0.07) (-11.11)

Log(Q) -0.150 0.597*** -0.065 0.007***

(-0.53) (6.65) (-0.63) (7.52)

Excess prior return -0.503*** -0.186 0.223 -0.005***

(-2.59) (-1.20) (1.50) (-5.55)

Negative earnings dummy -0.588*** 1.332*** -0.025 0.004***

(-2.65) (12.12) (-0.15) (3.09)

Lagged underfunded pension plan 3.666***

(15.92)

Industry fixed effects Yes Yes Yes Yes

Period fixed effects Yes Yes Yes Yes

No observations 7,346 7,008 6,791 7,983

McFadden R2/Adj. R2 0.265 0.152 0.592 0.072

64

Table 4. Corporate liquidity events (continued)

Bodnaruk et al. (2015) investigate the predictability of corporate liquidity events on a sample of almost 52,000 firm-year observations for public US firms. They find their text-based constraints measure to accurately predict all liquidity events, whereas the KZ, SA, and WW scores only predict one or two liquidity events. Farre-Mensa and Ljungqvist (2016) also evaluate the ability of the KZ, SA, and WW scores to predict equity recycling and find the KZ score to predict accurately, while the SA and WW scores do not. Our results are consistent with Andrén and Jankensgård (2018), who find the firm’s financial status to predict corporate liquidity events, while the eight alternative classifiers that they evaluate do not. Overall, our results on the ability of the different constraints classifiers to predict corporate liquidity events suggest that credit ratings outperform the other measures. The KZ score also performs well, but only when comparing extremes (top vs. bottom quartiles). Credit ratings predict accurately irrespective of whether we compare extremes or use the entire rating scale.

2.4.3. Investment-cash flow sensitivities

Our next step in evaluating the ability of ratings to capture financial constraints is to assess if ratings allow us to separate ICF sensitivities. ICF sensitivity is one of the most studied, but also one of the most debated concepts in corporate

Panel B

Dividend omission

Dividend increase

Underfunded pension plan

Equity recycling

KZ score -0.005 0.002 0.003* -0.000***

(-1.63) (0.85) (1.78) (-3.86)

SA score 0.473*** 0.094 0.264*** 0.000

(2.87) (1.28) (2.78) (0.17)

WW score -0.063 -0.077 4.117*** -0.000

(-0.14) (-0.34) (2.61) (-1.44)

DI score 3.611** 0.264 1.180* -0.002

(2.43) (0.88) (1.90) (-0.61)

EF score 3.547** 0.932** 1.438* -0.000

(2.45) (2.36) (1.84) (-0.07)

DF score 2.605 -0.161 2.167** 0.030**

(1.56) (-0.26) (2.25) (2.09)

65

finance. The underlying notion is that constrained firms’ investments ought to be more sensitive to internal cash flows, since external financing is more costly and/or less accessible for these firms. In Table 5, we report results from estimations of the extended Q model in Eq. 1. The main variable of interest is the interaction term between each firm’s rating/constraints score for the previous fiscal year and cash flow. The interaction term is significantly positive for ratings (Column 1 in Panel A), so worse-rated firms tend to exhibit higher ICF sensitivity. If we accept the assumption that ICF sensitivity measures financial constraints, this result supports ratings as a constraints measure. Alternatively, in light of the ability of ratings to predict corporate liquidity events, the significant interaction term could be interpreted as providing support for the use of ICF sensitivity as a constraints measure. The interaction term remains significantly positive if we exclude leverage and cash holdings or replace or add net debt and equity issuance and add beginning-of-period tangibility as a proxy for pledged collateral (see Columns 2-3). In Appendix 3, we show that our results are robust to adding sales growth as an additional investment opportunity proxy, to adding log beginning-of-period real size and log age, and to allowing for time-to-build by adding lagged capex. In the previous regressions, we instrument Q using the Lewellen and Lewellen (2016) IV estimator. As an alternative, we follow the recommendations of Almeida, Campello, and Galvao (2010) and alternatively instrument Q with the second lag of first-differenced Q and with third and fourth lags of first-differenced Q, cash flow, leverage, and cash holdings, without qualitatively changing the results (results tabulated in Appendix 3).

In Column 4, we extend Eq. 1 by adding interaction terms between ratings, leverage, and cash holdings. Although most of the literature focuses on ICF sensitivity, some authors (e.g., Kashyap, Lamont and Stein, 1994; Kaplan and Zingales, 1997) evaluate the importance of cash holdings as an alternative source of liquidity. Our results for investment-cash holdings sensitivity are similar to those for ICF sensitivity: worse-rated firms exhibit higher sensitivity, which is consistent with more constrained firms holding excess cash for precautionary purposes. The concordance in sensitivities is comforting, since theory makes no distinction between cash flow and cash holdings as a source of investment funding. Results (in Appendix 3) remain unchanged if we replace cash flow and cash holdings by the sum of the two liquidity sources.

66

Tab

le 5

. In

vest

men

t-ca

sh f

low

sen

siti

vity

an

d f

inan

cial

co

nst

rain

ts

Pan

el A

R

atin

g

(1)

Rat

ing

(2)

Rat

ing

(3)

KZ

sco

re

(4)

SA

sco

re

(5)

WW

sco

re

(6)

DI s

core

(7)

E

F s

core

(8

) D

F s

core

(9

)

Ope

ratin

g ca

sh fl

ow (

C)

0.09

1***

0.04

1 0.

087**

0.

180**

* 0.

457**

* 0.

295**

* 0.

187**

* 0.

191**

* 0.

188**

*

(2

.29)

(1

.09)

(2

.19)

(1

0.56

) (3

.97)

(3

.59)

(8

.77)

(8

.52)

(8

.92)

Con

stra

ints

cla

ssifi

er (

F)

-0.0

01

-0.0

03**

* -0

.002

***

-0.0

00**

* -0

.010

0.

027

-0.0

23

-0.0

29

0.01

4

(-

1.52

) (-

4.36

) (-

2.72

) (-

3.00

) (-

1.54

) (1

.01)

(-

1.19

) (-

1.33

) (0

.50)

F ×

C

0.00

8**

0.01

3***

0.00

9**

0.00

1***

0.07

0**

0.31

1 0.

164

0.26

2 -0

.111

(2

.17)

(3

.48)

(2

.31)

(2

.65)

(2

.53)

(1

.53)

(0

.93)

(1

.34)

(-

0.46

)

Q

0.02

2***

0.01

9***

0.02

2***

0.02

2***

0.02

3***

0.02

2***

0.02

4***

0.02

4***

0.02

4***

(8

.83)

(7

.26)

(8

.51)

(8

.61)

(8

.84)

(8

.67)

(7

.01)

(7

.03)

(7

.00)

Leve

rage

-0

.050

***

-0

.074

***

-0.0

50**

* -0

.051

***

-0.0

54**

* -0

.048

***

-0.0

49**

* -0

.049

***

(-

6.71

)

(-3.

57)

(-6.

43)

(-6.

65)

(-7.

04)

(-4.

92)

(-4.

92)

(-4.

98)

Cas

h ho

ldin

gs

-0.0

06

-0

.077

***

-0.0

10

-0.0

07

-0.0

13

0.00

3 0.

003

0.00

3

(-

0.45

)

(-2.

76)

(-0.

77)

(-0.

59)

(-0.

95)

(0.1

5)

(0.1

6)

(0.1

7)

Deb

t is

sues

0.02

6***

(6.1

8)

Equ

ity is

sues

-0.1

32**

*

(5.3

5)

Tan

gibi

lity

0.

131**

*

(11.

82)

F ×

Lev

erag

e

0.

002

(1

.20)

F ×

Cas

h ho

ldin

gs

0.00

7**

(2

.47)

N

9.15

1 8.

483

9.15

1 8.

675

8.65

3 8.

685

5.21

6 5.

216

5.21

6

Adj

R2

0.21

6 0.

276

0.22

1 0.

220

0.22

0 0.

222

0.18

4 0.

184

0.18

4

The

tabl

e re

port

s es

timat

es fr

om p

anel

reg

ress

ions

of c

apex

on

cash

flow

, a c

onst

rain

ts c

lass

ifier

(w

ith th

e in

clud

ed c

onst

rain

ts c

lass

ifier

giv

en b

y th

e co

lum

n he

adin

g), a

n in

tera

ctio

n te

rm b

etw

een

cash

flo

w a

nd t

he c

onst

rain

ts c

lass

ifier

, an

d be

ginn

ing-

of-p

erio

d Q

, pl

us c

ontr

ol v

aria

bles

(le

vera

ge,

cash

hol

ding

s, n

et d

ebt

issu

es,

net

equi

ty

issu

es,

and

tang

ibili

ty).

Pan

el A

sho

ws

full-

sam

ple

resu

lts.

Pan

el B

giv

es r

esul

ts f

or f

irm-y

ears

sor

ted

on b

egin

ning

-of-

year

inve

stm

ent

inte

nsity

, w

ith in

vest

men

t in

tens

ity

mea

sure

d by

tan

gibi

lity.

Hig

h (lo

w)

tang

ibili

ty m

odel

s in

clud

e fir

m-y

ears

in

the

top

(bot

tom

) te

rcile

of

the

tang

ibili

ty d

istr

ibut

ion.

All

mod

els

incl

ude

firm

and

per

iod

fixed

ef

fect

s. S

tand

ard

erro

rs a

re h

eter

osce

dast

icity

-con

sist

ent

and

clus

tere

d at

the

firm

leve

l. T

-sta

tistic

s in

par

enth

esis

.

67

Tab

le 5

. In

vest

men

t-ca

sh f

low

sen

siti

vity

an

d f

inan

cial

co

nst

rain

ts (

con

tin

ued

)

Pan

el B

Hig

h ta

ngib

ility

Lo

w t

angi

bilit

y

C

ash

flow

C

onst

rain

ts ×

Cas

h flo

w

Cas

h flo

w

Con

stra

ints

× C

ash

flow

Rat

ing

0.12

2 0.

010

**

0.02

0***

-0.0

02

(1

.00)

(2

.07)

(3

.50)

(-

1.61

)

KZ

sco

re

0.30

9***

0.04

8***

0.04

3***

0.00

1*

(5

.96)

(2

.91)

(4

.63)

(1

.68)

SA

sco

re

0.63

2* 0.

075

0.06

1 0.

005

(1

.83)

(0

.90)

(0

.88)

(0

.32)

WW

sco

re

0.74

8**

* 1.

04**

0.05

3 0.

049

(3

.97)

(2

.19)

(1

.12)

(0

.44)

DI s

core

0.

372

***

-0.0

69

0.02

9***

-0.0

42

(5

.01)

(-

0.15

) (3

.23)

(-

0.51

)

EF

sco

re

0.37

2***

0.16

1 0.

028**

* -0

.051

(4

.96)

(0

.31)

(3

.20)

(-

0.54

)

DF

sco

re

0.37

2***

-0.0

33

0.03

1***

-0.1

41

(5

.14)

(-

0.05

) (3

.44)

(-

1.12

)

68

Moshirian et al. (2017) show that ICF sensitivity increases in investment intensity and suggest that ICF sensitivity measures investment intensity rather than financial constraints. Andrén and Jankensgård (2018) extend on these results by distinguishing between the need for and the cost of external investment financing, using investment intensity as a proxy for the need for financing. They show that ICF sensitivity remains a valid measure of financial constraints after controlling for investment intensity. We evaluate this in Panel B of Table 5. Following Moshirian et al., we first sort our sample into terciles on tangibility and then re-estimate Eq. 1 separately for the high- and low-investment-intensity groups. Our results are consistent with Andrén and Jankensgård (2018). Investment is an increasing function of the Rating×Cash flow interaction term in the high-investment-intensity group, whereas the interaction term is insignificant for the low-investment-intensity group.

Overall ICF sensitivity, , is also significantly higher in the high-investment-

intensity group.

The KZ and SA scores also generate significantly positive coefficients on the interaction term between constraints and cash flow, suggesting that ICF sensitivity increases with these classifiers as well. Remaining comparison scores yield insignificant interaction terms. These results are robust to alternative control variables (see Appendix 3). Our finding that ICF sensitivity increases in the KZ score is consistent with Aǧca and Mozumdar (2017), but not with Kaplan and Zingales (1997), Hovakimian and Hovakimian (2009), and Grullon, Hund, and Weston (2018). The result is similar in character to Chen and Wang (2012), who find high-KZ firms to experience significantly poorer abnormal returns and operating performance after announcing share repurchases, and to Baker, Stein, and Wurgler (2003), who find investment in high-KZ firms (where they use the KZ score as a measure of equity dependence) to be more sensitive to movements in stock prices. They are also in line with Almeida et al. (2004) and Hadlock and Pierce (2010), who show that firms with lower KZ scores tend to save more of their cash flow. The positive coefficient on the interaction term for the SA score is consistent with Aǧca and Mozumdar (2017) and Hovakimian and Hovakimian (2009), but not with Kadapakkam, Kumar, and Riddick (1998), Cleary (2006), Andrén and Jankensgård (2015), Guney, Karpuz, and Ozkan (2017), Moshirian et al. (2017), and Grullon et al. (2018). The lack of a systematic pattern in ICF sensitivity for the WW score is inconsistent with Aǧca and Mozumdar (2017), whereas it is in line with Hadlock and Pierce (2010), who find greater tendencies to save cash flows among high-WW firms, and Hennessy and

69

Whited (2007), who find the cost of external financing to be higher for high-WW firms. We cannot compare our findings on ICF sensitivity for the DI, EF, and DF scores, since their ability to classify ICF sensitivity has not been evaluated before, but our results suggest that they do not succeed in informatively separating firms with high vs low ICF sensitivity.

In Table 5, we use as independent variables the full rating scale and continuous comparison scores. As an alternative, we again sort firms into four categories and estimate separate regressions for the most and least unconstrained classes. Results are presented in Appendix 4. The difference in ICF sensitivities between firms with weak vs. strong ratings is weakly significant and with the expected sign; the difference turns strongly significant if we instead control for net debt and equity issues and tangibility, or add controls for sales growth, age, size, and lagged capex. Irrespective of whether we use the entire rating scale or compare extremes, ratings suggest that firms with weaker ratings exhibit greater sensitivity of investment to cash flow. The difference in constraints is significant and of the right sign for the KZ, DI, and EF scores as well, whereas the other comparison scores yield insignificant differences (but with the expected sign).

2.4.4. Responses to shocks to demand and supply for capital

The final step in our evaluation of the ability of ratings to capture financial constraints is to evaluate the ability of ratings (and the comparison scores) to predict corporate investment in the face of exogenous shocks to capital supply. We investigate two qualitatively different shocks: the global financial crisis (GFC) and the introduction of the Jobs and Growth Tax Relief Reconciliation Act (JGTRRA). The GFC was primarily a negative shock to credit markets and our hypothesis is that financially constrained firms experienced greater curtailment of investment than unconstrained ones. The JGTRRA was instead a positive shock to equity financing, where we expect financially constrained firms to benefit more from reductions in the cost of equity. We model changes in capex in the year after vs the year before each shock; for the GFC, the dependent variable is the difference in capex in 2009 vs. 2006, while the dependent variable in the JGTRRA models is the difference in capex in 2004 vs 2002. Investment fell abruptly during the GFC, from a sample average of 7.0% in 2006 to 4.9% in 2009. In contrast, capex increased following the JGTRRA, from 5.3% in 2002 to 5.7% in 2004. This means that a fundamental

70

prerequisite, that investment overall moved in the expected direction, ismet. To evaluate if capex changed predictably with financial constraints, we employ a difference-in-difference setup where we regress the change in capex on the pre-shock level of financial constraints (and control variables).

We report results from our shock-response models for the GFC in Panel A of Table 6. Our main specification (first column) follows Hoberg and Maksimovic (2015) in using log age, log real size, Q, and firm and industry sales growth as control variables. All control variables are measured in the year before the shock.

Table 6. Investment in shock periods The table reports results from regressions of the change in capex on beginning-of-period Q, log age, log real size, sales growth, and industry sales growth (in column 2 also operating cash flow, leverage, and cash holdings). Panel A contains models in which the change in capex is measured as the difference in capex in 2009 vs. 2006 and control variables are held constant at their 2006 level. Panel B contains models in which the change in capex is measured as the difference in capex in 2004 vs. 2002 and control variables are held constant at their 2002 level. T-statistics in parenthesis. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Panel A Rating Rating

KZ score

SA score WW score DI score EF score

DF score

Constraints -0.003** -0.003*** -0.000** -0.018 -0.275*** -0.018 -0.012 -0.063

classifier (F) (-2.05) (-2.59) (-2.27) (-1.02) (2.53) (-0.41) (-0.24) (-0.71)

Q 0.000 0.020 0.009** 0.012*** 0.009** 0.005 0.005 0.004

(0.03) (1.46) (2.07) (2.70) (2.17) (0.66) (0.66) (0.56)

Log(age) 0.007* 0.002 -0.007 -0.001 0.012** 0.012** 0.012**

(1.76) (0.59) (-0.70) (-0.16) (2.24) (2.26) (2.25)

Log(real size) -0.005 0.001 0.000 -0.013** -0.004 -0.004 -0.004

(-1.43) (0.29) (0.14) (-2.38) (-1.09) (-1.08) (-1.09)

Sales growth -0.080*** -0.081** -0.080** -0.088*** -0.091*** -0.091*** -0.093***

(-3.32) (-2.55) (-2.53) (-2.82) (-3.40) (-3.42) (-3.41)

Industry sales growth

0.042 0.051 0.045 0.083** 0.012 0.012 0.012

(1.32) (1.48) (1.37) (2.17) (0.32) (0.33) (0.34)

Operating cash flow (C) -0.302***

(-4.53)

Leverage 0.002

(0.11)

Cash holdings 0.118***

(3.46)

N 514 482 465 467 466 401 401 401

71

Table 6. Investment in shock periods (continued) Panel B

Rating Rating KZ score

SA score

WW score DI score EF score

DF score

Constraints classifier (F)

0.003** 0.002** 0.000 -0.002 -0.107 0.077** 0.093** -0.037

(1.98) (2.07) (1.43) (-0.13) (-1.14) (2.20) (2.42) (-0.57)

Q 0.001 -0.002 -0.001 -0.001 -0.001 0.000 -0.000 -0.001

(0.51) (-0.57) (-0.48) (-0.50) (-0.82) (0.06) (-0.04) (-0.29)

Log(age) -0.003 -0.004* -0.005 -0.005* -0.003 -0.003 -0.004

(1.21) (-1.79) (-0.67) (-1.77) (-1.10) (-1.14) (-1.55)

Log(real size) 0.002 -0.001 -0.001 -0.006 0.000 -0.000 0.001

(0.68) (-0.16) (-0.24) (-1.60) (0.09) (-0.04) (0.19)

Sales growth -0.029*** -0.040** -0.039** -0.043** -0.025*** -0.025*** -0.027***

(-3.01) (-2.17) (-2.15) (-2.22) (-2.73) (-2.73) (-2.79)

Industry sales growth

-0.019 -0.015 -0.014 -0.005 -0.028 -0.031 -0.025

(-1.10) -0.85) (-0.79) (-0.26) (-1.29) (-1.38) (-1.11)

Operating cash flow (C) 0.038

(0.77)

Leverage 0.012

(0.79)

Cash holdings 0.022

(0.98)

N 442 364 382 385 383 330 330 330

Firms that were constrained prior to the GFC, as measured by the credit rating, curtailed capex more the worse the level of constraints, even after controlling for age, size, Q, and sales growth. This result supports the validity of the credit rating as a financial constraints measure. To verify the robustness of the greater curtailment of investment in worse rated firms, we in the second column estimate a variant of Eq. 1, where we include as controls lagged (and instrumented) Q, cash flow, leverage, and cash holdings. Again, the coefficient on credit ratings is significantly negative. When considering the robustness of our results, we also emphasize that our difference-in-difference setup implicitly accounts for firm fixed effects, since they are differenced out of the models.

Results for the comparison scores are mixed. The KZ and WW scores also suggest that firms classified as constrained curtailed investment more than less

72

constrained firms did.17 It is reasonable that these two scores succeed in differentiating firm behavior in the GFC, considering that both scores, in slightly different ways, emphasize financial status. The result is consistent with Duchin, Ozbas, and Sensoy (2010), but in contrast to Hoberg and Maksimovic (2015), who did not find any of these scores to predict capex during the GFC. In contrast to our lack of result for the SA score, Duchin et al. (2010) find size to predict investment patterns in the crisis.18 None of the Hoberg and Maksimovic scores come out significant in our testing. Hoberg and Maksimovic find the DI and EF scores to strongly predict curtailment of capex for their more dispersed sample of firms. Apparently, these scores work less well for rated companies. All results for the comparison scores are robust to controlling for cash flow, leverage, and cash holdings (tabulated in Appendix 5).

In Panel B, we turn to the JGTRRA. We employ the same model specifications as for the GFC. Ratings come out significant for this shock as well. The worse the rating, the more firms augmented investment following the introduction of the JGTRRA. This result applies after controlling for age, size, Q, firm and industry sales growth, cash flow, leverage, and cash holdings. As for the GFC, the credit rating accurately predicts corporate behavior consistent with being financially constrained. The results for the comparison scores are again mixed, with only the KZ, DI, and EF scores accurately predicting corporate behavior. The EF score measures equity dependence, so is a reasonable predictor of investment responses to the JGTRRA. The result for the DI score is also reasonable, considering that the DI score de facto also seems to be a measure of equity dependence, as it is very highly correlated (ρ = 0.90) with the EF score.

Campbell et al. (2013) find the increase in investment to be larger for firms that are more likely to fund investment from new equity rather than internal funds. This is consistent with our results for the DI and EF scores. Interestingly, credit ratings also seem to capture this equity dependence, even though the correlations between ratings and the DI and EF scores are limited (ρDI = 0.17

17 As Q (firm and industry sales growth) is an input in the calculation of the KZ (WW) index,

we re-estimate the models excluding these controls to avoid spurious multicollinearity. The KZ index is robust to this alteration, whereas the WW index is not (see Appendix 5).

18 To account for possible spurious multicollinearity, we also estimate a model excluding size and age, since these are inputs into the calculation of the SA index, but without influencing the result (see the Appendix 5).

73

and ρEF = 0.06). One explanation may be Dai et al.’s (2013) finding that the reduction in the cost of equity following the JGTRRA was greater for firms classified as constrained using the SA and WW scores. As we saw previously, these are the two comparison scores most correlated with credit ratings.

The critical challenge in our shock response modelling is that the changes in capex should reflect financing supply effects rather than changes in demand. We follow Hoberg and Maksimovic (2015) in controlling for this by adding contemporaneous changes in sales at the firm and industry levels. As pointed out by Hoberg and Maksimovic, this is a stringent test for demand effects. Still, including these controls does not qualitatively influence our results (Columns 3 in Panels A and B).

In Table 6, we use as independent variables the full rating scale and continuous comparison scores. As an alternative, we again sort firms into four categories and replace the full rating scale/continuous comparison scores with dummy variables for each of the four classes. Results are presented in Appendix 6. As for the liquidity events we focus our analysis on differences between the least and most constrained classes. Ratings exhibit significant differences between firm-years with strongest and weakest ratings and with the expected signs for both GFC and JGTRRA. The KZ score is the only comparison score that exhibits significance for GFC, while the DI and EF scores are significant for JGTRRA.

Agha and Faff (2014) show that the cost of capital increases (decreases), while capital expenditures decrease (increase) following a rating downgrade (upgrade). This is what we would expect if ratings capture financial constraints. To conclude our testing of the information content of credit ratings, we evaluate the impact of credit re-ratings on capex. In Columns 1-3 of Table 7, we regress the change in capex in the year after a re-rating relative the year before on up- and downgrade dummies. The modelling is similar to the other shock-response models, except that we have treated firms in every year, so estimate panel difference-in-difference regressions where treated firms see their ratings improve or deteriorate, whereas the control group contains all firms with no change in their credit ratings. Consistent with our results for GFC and JGTRRA, we find that firms decrease (increase) capex following a rating downgrade (upgrade).

74

Table 7. Investment responses to credit re-ratings The table reports results from panel regressions of the change in capex on up- and downgrade dummies (lagged one period), beginning-of-period instrumented Q, credit rating, log age, log real size, sales growth, industry sales growth, operating cash flow, leverage, and cash holdings. Change in capex is measured as capex in the year after minus the year before a credit re-rating; control variables are held constant at their year-before-re-rating level. All models include firm and period fixed effects. Standard errors are heteroscedasticity-consistent and clustered at the firm level. T-statistics in parenthesis. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Rating (1)

Rating (2)

Rating (3)

Rating upgrade 0.009*** 0.008*** 0.010***

(4.21) (3.95) (4.90)

Rating downgrade -0.012*** -0.011*** -0.014***

(-6.35) (-5.82) (-7.04)

Credit rating 0.001*** 0.002*** 0.001

(3.33) (3.04) (1.02)

Q 0.008*** 0.010*** 0.005*

(2.99) (3.69) (1.82)

Log(age) 0.008*** 0.009***

(2.73) (3.32)

Log(real size) -0.011*** -0.012***

(-6.74) (-7.47)

Sales growth -0.048*** -0.043***

(-9.18) (-8.29)

Industry sales growth 0.005 0.005*

(1.62) (1.65)

Operating cash flow -0.126*** -0.106***

(-7.58) (-6.99)

Leverage -0.009 -0.011

(-1.26) (-1.54)

Cash holdings 0.070*** 0.050***

(4.73) (3.27)

N 8,816 8,873 8,813

R2 0.115 0.088 0.126

75

2.4.5. Financially constrained equals financially distressed?

The distinction between constraints and distress is a concern in the literature. Credit ratings are intended to classify firms along the distress continuum, but we show that they also allow classifying firms along the constraints continuum, and that ratings are more valid constraints measures than other constraints classifiers. A possible objection is that we are measuring distress rather than constraints. We argue to the contrary. We find that ratings predict liquidity events, investments, and responses to shocks to the debt and equity markets using both the full rating scale and by comparing strongest and weakest ratings. Had we been capturing distress rather than constraints, this would have shown up as significant when comparing extremes, but not when using the full rating scale. In this context, we reiterate that we assume the rating scale to be linear, and the two extreme rating groups (A or higher and B or lower) make up 27% and 18% of the full sample. Still, to further corroborate that our results are not caused by a subset of distressed firms, we in this section show that our results are robust to sample restrictions aimed at excluding potentially distressed firm-years. Results are presented in Table 8. Each row of the table reflects a unique sample restriction, with results for the different tests presented in the columns. We only present coefficients on the key test variables, so each cell represents a separate regression, except for Columns 8-9, which are estimated in the same model.

76

Tab

le 8

. R

ob

ust

nes

s to

exc

lud

ing

po

ssib

ly d

istr

esse

d f

irm

-yea

rs

The

tabl

e re

port

s re

sults

from

re-

estim

atio

ns o

f our

var

ious

test

s gi

ven

diffe

rent

sam

ple

rest

rictio

ns, w

ith th

e in

clud

ed s

ampl

e sp

ecifi

ed in

the

head

ing

on e

ach

row

. Eac

h co

lum

n re

pres

ents

a s

epar

ate

test

; Col

umns

1-4

are

sim

ilar

to C

olu

mns

1-4

in T

able

4; C

olum

n 5

is s

imila

r to

Col

umn

3 in

Pan

el A

of T

able

5; C

olum

ns 6

-7 a

re s

imila

r to

C

olum

n 1

in P

anel

s A

and

B o

f T

able

6;

and

Col

umns

8-9

are

sim

ilar

to C

olum

n 1

in T

able

7.

We

only

pre

sent

est

imat

es o

n th

e ke

y te

st p

aram

eter

s, s

o ea

ch c

ell

repr

esen

ts o

ne m

odel

, ex

cept

Col

umns

8-9

, w

hich

are

est

imat

ed i

n th

e sa

me

mod

el.

Col

umns

1-4

con

tain

est

imat

es f

or l

agge

d ra

ting,

Col

umn

5 fo

r th

e in

tera

ctio

n be

twee

n la

gged

rat

ing

and

oper

atin

g ca

sh fl

ow, C

olum

n 6

(7)

for

the

third

(se

cond

) la

g of

rat

ing,

and

Col

umn

8 (9

) fo

r ra

ting

upg

rade

(do

wng

rade

). w

ith te

sts

for

liqui

dity

ev

ents

in C

olum

ns 1

-4, I

CF

sen

sitiv

ity in

Col

umn

5, re

spon

ses

to G

FC

and

JG

TR

RA

in C

olum

ns 6

-7, a

nd re

spon

ses

to re

-rat

ings

in C

olum

ns 8

-9.T

-sta

tistic

s in

par

enth

esis

.

D

ivid

end

omis

sion

(1

)

Div

iden

d in

crea

se

(2)

Und

erfu

nded

pe

nsio

n pl

an

(3)

Equ

ity

recy

clin

g (4

)

ICF

se

nsiti

vity

(5

)

GF

C

(6)

JGT

RR

A

(7)

Re-

ratin

g U

pgra

de

(8)

Re-

ratin

g D

owng

rade

(9

)

(1)

0.50

9***

-0.1

68**

* 0.

076**

* -0

.001

***

0.01

3***

-0.0

04**

0.00

1* 0.

011**

* -0

.017

***

Sal

es g

row

tht-

1 >

0

(8.6

3)

(-7.

64)

(3.6

1)

(-2.

98)

(2.8

2)

(-2.

01)

(1.8

6)

(4.9

4)

(-6.

40)

(2)

0.45

9***

-0.1

88**

* 0.

090**

* -0

.001

***

0.01

2***

-0.0

03**

0.00

2***

0.01

0***

-0.0

15**

*

Ope

ratin

g ca

sh fl

owt-

1 >

0

(10.

02)

(-9.

78)

(5.0

7)

(-4.

32)

(3.1

0)

(-2.

14)

(2.6

1)

(5.1

5)

(-7.

63)

(3)

0.54

9***

-0.1

90**

* 0.

093**

* -0

.001

***

0.01

3***

-0.0

04**

0.00

1**

0.01

1***

-0.0

15**

*

No

divi

dend

red

uctio

n t-1

(9

.96)

(-

8.12

) (5

.32)

(-

4.68

) (3

.51)

(-

2.25

) (2

.11)

(5

.41)

(-

7.19

)

(4)

0.49

0***

-0.1

91**

* 0.

094**

* -0

.001

***

0.01

2***

-0.0

04**

0.00

2**

0.01

1***

-0.0

16**

*

Rat

ing t

-1 >

CC

C+

(1

1.65

) (-

9.95

) (5

.93)

(-

4.61

) (2

.99)

(-

2.35

) (2

.33)

(5

.51)

(-

7.94

)

(5)

0.64

5***

-0.1

95**

* 0.

105**

* -0

.001

***

0.01

2**

-0.0

02*

0.00

1 0.

010**

* -0

.013

***

Rat

ing t

-1 >

B+

(8

.36)

(-

9.20

) (5

.99)

(-

3.91

) (2

.31)

(-

1.76

) (0

.87)

(4

.64)

(-

6.66

)

(6)

0.49

0***

-0.2

06**

* 0.

075**

* -0

.001

***

0.01

4**

-0.0

04**

0.00

2**

0.01

1***

-0.0

16**

*

AA

- >

Rat

ing t

-1 >

CC

C+

(1

1.92

) (-

11.3

5)

(4.4

2)

(-4.

40)

(2.6

8)

(-2.

39)

(2.4

2)

(5.5

8)

(-7.

94)

(7)

0.66

6***

-0.1

91**

* 0.

108**

* -0

.001

0.

017

-0.0

04*

0.00

1 0.

011**

* -0

.016

***

A-

> R

atin

g t-1

> B

+

(8.6

8)

(-7.

44)

(4.2

3)

(-1.

44)

(1.4

9)

(-1.

79)

(0.6

0)

(4.2

9)

(-6.

17)

77

Since distress is a continuum, there is no one generally accepted measure of when a firm is in distress. Following Fazzari et al. (1988) and Lamont et al. (2001), among others, we in the first row restrict the sample to firms in year t with positive sales growth in year t-1. Again, ratings accurately predict liquidity events (coefficients on lagged ratings in Columns 1-4). ICF sensitivity increases with the credit rating (coefficient on the interaction term between the lagged rating and operating cash flow in Column 5); credit ratings predict investment responses to GFC and JGTRRA (coefficients in third and second lags of Ratings in Columns 6-7), and the investment response to credit upgrades (downgrades) is positive (negative). In Row 2, we follow Allayannis and Mozumdar (2004) and Cleary et al. (2007) and restrict the sample to firm-years with positive cash flows in year t-1. Again, results remain robust across the boards. Row 3 instead report results where we exclude firm-years with dividend reductions in year t-1, with unchanged results. In Rows 4-5, we exclude the worst-rated firms; Row 4 excludes firm-years with a rating below B- , while Row 5 excludes firm-years rated below BB-. In Rows 6-7, we additionally exclude the best-rated firms by excluding firm-years with a rating above A+ (Row 6) and BBB+ (Row 7). The reason for this is to evaluate if firms with strong ratings distort our results. With minor exceptions, results remain robust to all these variations, so it seems neither weak, nor strong ratings seem to distort our results. In conclusion, we feel confident in concluding that our results are not caused by a portion of our sample consisting of distressed rather than constrained firms.

2.5. Discussion and concluding remarks

Kashyap et al. (1994), Gilchrist and Himmelberg (1995), and Almeida et al. (2004) investigate if access to public debt markets influences capital investments, finding investment in unrated firms to be more sensitive to internal cash flow. Contrary to these findings, Farre-Mensa and Ljungqvist (2016) show that rated and unrated firms respond similarly to exogenous financing demand and supply shocks. We, in contrast, ask if credit quality (measured by the rating level), as opposed to access to public debt markets (measured by having a rating), allows us to meaningfully discriminate more from less financially constrained firms.

78

There are strong reasons to expect ratings to be informative about a firm’s financial constraints status. To begin with, ratings line up monotonically with corporate default rates as far as up to 20 years into the future (Moody’s, 2018; Standard and Poor’s, 2018). Further, CRAs mitigate information asymmetries by providing investors with an independent assessment of corporate risk profiles. They can also convey inside information that investors cannot obtain from public sources. For example, firms often provide CRAs with strategic plans and management’s forward-looking projections, allowing credit ratings to verify and incorporate private information without disclosing such information. CRAs also mitigate agency problems by assessing not only the firm’s ability to meet its financial obligations, but also the willingness of the management and owners to meet those obligations.

Credit ratings find empirical support in every test we run. Ratings yield an intuitive partitioning of sample firm characteristics, with worse-rated firms being smaller, younger, and less profitable, as well as generating smaller cash flows, paying smaller dividends, investing less in capex and R&D, and growing sales at a slower pace than industry peers. Not only do we consider these characteristics reasonable for financially constrained firms, but this combination of characteristics is not exhibited by any of the six comparison scores.

Ratings accurately predict corporate behavior indicative of improvement or deterioration in external financial constraints: omitting or increasing dividends, recycling equity, and having underfunded pension plans. The worse the rating, the more likely the firm is to omit dividend payments and have underfunded pension plans, and less likely to (increase dividends and recycle equity. Not only do these findings support using ratings as a measure of financial constraints, but they also suggest that the credit rating is a more reliable constraints measure than the comparison scores, as none of the comparison scores accurately predict all four liquidity events. Ratings also predict investment financing behavior, with worse-rated firms relying more on internal financing when financing investment, irrespective of whether we control for investment intensity or not. The usefulness of ICF sensitivity as a constraints measure is debated, but if we make the conjecture that ICF sensitivity is valid, then ratings would not only be a valid constraints measure, but a more valid constraints measure than the comparison scores, with the exception of the KZ score. Finally, ratings accurately predict corporate responses to two very different shocks to the supply of capital. The reduction in investment in the

79

GFC varies predictably with the credit rating, with worse-rated firms curtailing investments more. The increase in investment following the introduction of the JGTRRA also varies predictably with the credit rating, with worse-rated firms augmenting investments more. Not only do these findings support using ratings as a measure of financial constraints, but they also suggest that the credit rating is a more reliable constrains measure than the comparison scores, as none of the comparison scores succeed in predicting responses to both events. Rather, the two comparison scores that most directly relate to credit market conditions – the KZ and WW scores – predict investment behavior in the GFC, whereas the two scores that most directly capture dependence on equity financing – the DI and EF scores – predict investment behavior in the JGTRRA. What is striking with our results is that the credit rating succeeds in predicting corporate responses to shocks to both debt and equity financing (as well as to negative and positive capital supply shocks alike).

None of the comparison scores seem to be reliable measures of financial constraints. This is in line with Farre-Mensa and Ljungqvist (2016), Bodnaruk et al. (2015), and Andrén and Jankensgård (2018), who sharply criticize existing measures of financial constraints. Our results suggest that the criticism of the KZ, SA, and WW scores in these studies also applies to the Hoberg and Maksimovic (2014) constraints classifiers. The DI, EF, and DF scores receive strong support in Hoberg and Maksimovic’s own testing, but fail to reliably identify constrained firms when applied to our more constrained sample of rated firms. A possible reason for ratings performing better than the Hoberg and Maksimovic scores is the superior access of information that ratings analysts have. While Hoberg and Maksimovic rely on textual analysis of the Management’s Discussion and Analysis section of the 10-K, ratings analysts have direct management access, giving them access to private information and allowing them to make judgment calls on that information.

We believe our findings are of significant relevance to the research area. Being able to measure the severity of financial constraints is key to appreciating the importance of market frictions to corporate investments and financial policies (Fazzari et al., 1988), asset pricing (Whited and Wu, 2006), and monetary policymaking (Gertler and Gilchrist, 1994). Most pre-existing constraints measures are cast in doubt by recent research, but as we show, credit ratings outperform other measures regarding their ability to predict corporate liquidity events, investment financing behavior, and responses to shocks to supply of external debt and equity.

80

Kaplan and Zingales (1997) called for measures of financially constrained firm-years rather than constrained firms. The intuition is that the severity of financial constraints should vary predictably over time as the firm and its environment change. Such time variation is difficult to capture with corporate characteristics; as the global financial crisis and other financial disruptions have shown, even large and mature firms can quickly become financially constrained. In this regard, credit ratings benefit from being monitored on an ongoing basis and from being reviewed both regularly and when corporate or industry developments call for it. Periodic and event-driven reviews allow for ratings to be sensitive to changes in financial constraints in, for example, large and mature companies that corporate characteristics such as size and age may not reflect, which would mean that credit ratings would have the added benefit of measuring dynamics in financial constraints.

81

References

Acharya, V. V., & Schnabl, P. (2010). Do global banks spread global imbalances? Asset-backed commercial paper during the financial crisis of 2007–09. IMF Economic Review, 58(1), 37-73.

Ağca, Ş., & Mozumdar, A. (2017). Investment–Cash Flow Sensitivity: Fact or Fiction? Journal of Financial and Quantitative Analysis, 52(3), 1111-1141.

Agha, M., & Faff, R. (2014). An investigation of the asymmetric link between credit re-ratings and corporate financial decisions: “Flicking the switch” with financial flexibility. Journal of Corporate Finance, 29, 37-57.

Allayannis, G., & Mozumdar, A. (2004). The impact of negative cash flow and influential observations on investment–cash flow sensitivity estimates. Journal of Banking & Finance, 28(5), 901-930.

Almeida, H., Campello, M., & Galvao Jr, A.F. (2010). Measurement errors in investment equations. Review of Financial Studies, 23(9), 3280-3328.

Almeida, H., Campello, M., & Weisbach, M. S. (2004). The cash flow sensitivity of cash. The Journal of Finance, 59(4), 1777-1804.

Almeida, H., Campello, M., & Weisbach, M. S. (2011). Corporate financial and investment policies when future financing is not frictionless. Journal of Corporate Finance, 17(3), 675-693.

Altman, E. I., & Rijken, H. A., 2006. A point-in-time perspective on through-the-cycle ratings. Financial Analysts Journal, 52(1), 54-70.

Andrén, N., & Jankensgård, H. (2018). Reappearing investment-cash flow sensitivities. Knut Wicksell Center working paper, Lund University School of Economics and Management.

Ascioglu, A., Hegde, S. P., & McDermott, J. B. (2008). Information asymmetry and investment–cash flow sensitivity. Journal of Banking & Finance, 32(6), 1036-1048.

Ashbaugh-Skaife, H., Collins, D. W., & LaFond, R. (2006). The effects of corporate governance on firms’ credit ratings. Journal of Accounting and Economics, 42(1-2), 203-243.

Avramov, D., Chordia, T., Jostova, G., & Philipov, A. (2013). Anomalies and financial distress. Journal of Financial Economics, 108(1), 139-159.

Baker, M., Stein, J. C., & Wurgler, J. (2003). When does the market matter? Stock prices and the investment of equity-dependent firms. The Quarterly Journal of Economics, 118(3), 969-1005.

Bannier, C. E., & Hirsch, C. W. (2010). The economic function of credit rating agencies–What does the watchlist tell us?Journal of Banking & Finance, 34(12), 3037-3049.

82

Batta, G. (2011). The direct relevance of accounting information for credit default swap pricing. Journal of Business Finance & Accounting, 38(9‐10), 1096-1122.

Scholes, M., Benston, G. J., & Smith Jr, C. W. (1976). A transactions cost approach to the theory of financial intermediation. The Journal of finance, 31(2), 215-231.

Bodnaruk, A., Loughran, T., & McDonald, B. (2015). Using 10-k text to gauge financial constraints. Journal of Financial and Quantitative Analysis, 50(4), 623-646.

Boot, A. W. A., Milbourn, T. T., & Schmeits, A. (2006). Credit ratings as coordinating mechanisms. Review of Financial Studies, 19(1), 81-118.

Brav, A., Graham, J. R., Harvey, C. R., & Michaely, R. (2005). Payout policy in the 21st century. Journal of Financial Economics, 77(3), 483-527.

Brown, J. R., Liang, N., & Weisbenner, S. (2007). Executive financial incentives and payout policy: Firm responses to the 2003 dividend tax cut. The Journal of Finance, 62(4), 1935-1965.

Brown, J. R., Fazzari, S. M., & Petersen, B. C. (2009). Financing innovation and growth: Cash flow, external equity, and the 1990s R&D boom. The Journal of Finance, 64(1), 151-185.

Campbell, J. L., Chyz, J. A., Dhaliwal, D. S., & Schwartz Jr, W. C. (2013). Did the 2003 tax act increase capital investments by corporations? The Journal of the American Taxation Association, 35(2), 33-63.

Cantor, R. & Mann, C. (2003). Are corporate bond ratings procyclical? Special comment, Moody’s Investor Services.

Chen, H. J., & Chen, S. J. (2012). Investment-cash flow sensitivity cannot be a good measure of financial constraints: Evidence from the time series. Journal of Financial Economics, 103(2), 393-410.

Chen, S. S., & Wang, Y. (2012). Financial constraints and share repurchases. Journal of Financial Economics, 105(2), 311-331.

Cheng, M., & Neamtiu, M. (2009). An empirical analysis of changes in credit rating properties: Timeliness, accuracy and volatility. Journal of Accounting and Economics, 47(1-2), 108-130.

Cheng, M., & Subramanyam, K. R. (2008). Analyst following and credit ratings. Contemporary Accounting Research, 25(4), 1007-1044.

Cleary, S. (1999). The relationship between firm investment and financial status. Journal of Finance, 54(2), 673-692.

Cleary, S. (2006). International corporate investment and the relationships between financial constraint measures. Journal of Banking & Finance, 30(5), 1559-1580.

Cleary, S., Povel, P., & Raith, M. (2007). The U-shaped investment curve: Theory and evidence. Journal of Financial and Quantitative Analysis, 42(1), 1-39.

83

Czarnitzki, D., & Kraft, K. (2007). Are credit ratings valuable information? Applied Financial Economics, 17(13), 1061-1070.

Dai, Z., Shackelford, D. A., Zhang, H. H., & Chen, C. (2013). Does financial constraint affect the relation between shareholder taxes and the cost of equity capital? The Accounting Review, 88(5), 1603-1627.

Diamond, D. W. (1984). Financial intermediation and delegated monitoring. The Review of Economic Studies, 51(3), 393-414.

Dichev, I. D., & Piotroski, J. D. (2001). The long‐run stock returns following bond ratings changes. The Journal of Finance, 56(1), 173-203.

Dick-Nielsen, J., Feldhütter, P., & Lando, D. (2012). Corporate bond liquidity before and after the onset of the subprime crisis. Journal of Financial Economics, 103(3), 471-492.

Doumpos, M., Niklis, D., Zopounidis, C., & Andriosopoulos, K. (2015). Combining accounting data and a structural model for predicting credit ratings: Empirical evidence from European listed firms. Journal of Banking & Finance, 50, 599-607.

Duchin, R., Ozbas, O., & Sensoy, B. A. (2010). Costly external finance, corporate investment, and the subprime mortgage credit crisis. Journal of Financial Economics, 97(3), 418-435.

Ederington, L. H., & Goh, J. C. (1998). Bond rating agencies and stock analysts: who knows what when? Journal of Financial and Quantitative Analysis, 33(4), 569-585.

Erickson, T., & Whited, T. M. (2000). Measurement error and the relationship between investment and q. Journal of Political Economy, 108(5), 1027-1057.

Farre-Mensa, J., & Ljungqvist, A. (2016). Do measures of financial constraints measure financial constraints? The Review of Financial Studies, 29(2), 271-308.

Faulkender, M., & Petersen, M. A. (2005). Does the source of capital affect capital structure? The Review of Financial Studies, 19(1), 45-79.

Fazzari, S. M., Hubbard, R. G., and Petersen, B. C. (1988), Financing constraints and corporate investment. Brookings Papers on Economic Activity, 1988(1), 141-206.

Fazzari, S. M., Hubbard, R. G., & Petersen, B. C. (2000). Investment-cash flow sensitivities are useful: A comment on Kaplan and Zingales. The Quarterly Journal of Economics, 115(2), 695-705.

Finnerty, J. D., Miller, C. D., & Chen, R. R. (2013). The impact of credit rating announcements on credit default swap spreads. Journal of Banking & Finance, 37(6), 2011-2030.

Franzoni, F. (2009). Underinvestment vs. overinvestment: Evidence from price reactions to pension contributions. Journal of Financial Economics, 92(3), 491-518.

84

Friewald, N., Wagner, C., & Zechner, J. (2014). The cross‐section of credit risk premia and equity returns. The Journal of Finance, 69(6), 2419-2469.

Friewald, N., Jankowitsch, R., & Subrahmanyam, M. G. (2013). Illiquidity or credit deterioration: A study of liquidity in the US corporate bond market during financial crises. In Managing and Measuring Risk: Emerging Global Standards and Regulations after the Financial Crisis (pp. 159-200).

Gertler, M., & Gilchrist, S. (1994). Monetary policy, business cycles, and the behavior of small manufacturing firms. Quarterly Journal of Economics, 109(2), 309–340.

Gilchrist, S., & Himmelberg, C. P. (1995). Evidence on the role of cash flow for investment. Journal of Monetary Economics, 36(3), 541-572.

Goh, J. C., & Ederington, L. H. (1993). Is a bond rating downgrade bad news, good news, or no news for stockholders? The Journal of Finance, 48(5), 2001-2008.

Goh, J. C., & Ederington, L. H. (1999). Cross-sectional variation in the stock market reaction to bond rating changes. The Quarterly Review of Economics and Finance, 39(1), 101-112.

Graham, J. R., & Harvey, C. R. (2001). The theory and practice of corporate finance: Evidence from the field. Journal of Financial Economics, 60(2-3), 187-243.

Grullon, G., Hund, J., & Weston, J. P. (2018). Concentrating on q and cash flow. Journal of Financial Intermediation, 33, 1-15.

Guney, Y., Karpuz, A., & Ozkan, N. (2017). R&D investments and credit lines. Journal of Corporate Finance, 46, 261-283.

Güttler, A., 2005. Using a bootstrap approach to rate the raters. Financial Markets and Portfolio Management, 19(3), 277-295.

Hadlock, C. J., & Pierce, J. R. (2010). New evidence on measuring financial constraints: Moving beyond the KZ index. The Review of Financial Studies, 23(5), 1909-1940.

Han, S., & Zhou, X. (2013). Informed bond trading, corporate yield spreads, and corporate default prediction. Management Science, 60(3), 675-694.

Hand, J. R., Holthausen, R. W., & Leftwich, R. W. (1992). The effect of bond rating agency announcements on bond and stock prices. The Journal of Finance, 47(2), 733-752.

Harford, J., & Uysal, V. B. (2014). Bond market access and investment. Journal of Financial Economics, 112(2), 147-163.

He, Y., Wang, J., & Wei, K. J. (2011). Do bond rating changes affect the information asymmetry of stock trading? Journal of Empirical Finance, 18(1), 103-116.

Hennessy, C. A. (2004). Tobin's Q, debt overhang, and investment. The Journal of Finance, 59(4), 1717-1742.

Hennessy, C. A., & Whited, T. M. (2007). How costly is external financing? Evidence from a structural estimation. The Journal of Finance, 62(4), 1705-1745.

85

Hilscher, J., & Wilson, M. (2016). Credit ratings and credit risk: Is one measure enough? Management Science, 63(10), 3414-3437.

Hite, G., & Warga, A. (1997). The effect of bond-rating changes on bond price performance. Financial Analysts Journal, 35-51.

Hoberg, G., & Maksimovic, V. (2014). Redefining financial constraints: A text-based analysis. The Review of Financial Studies, 28(5), 1312-1352.

Holthausen, R. W., & Leftwich, R. W. (1986). The effect of bond rating changes on common stock prices. Journal of Financial Economics, 17(1), 57-89.

Hovakimian, A., & Hovakimian, G. (2009). Cash flow sensitivity of investment. European Financial Management, 15(1), 47-65.

Hull, J., Predescu, M., & White, A. (2004). The relationship between credit default swap spreads, bond yields, and credit rating announcements. Journal of Banking & Finance, 28(11), 2789-2811.

Ivashina, V., & Scharfstein, D. (2010). Bank lending during the financial crisis of 2008. Journal of Financial Economics, 97(3), 319-338.

Jankowitsch, R., Nagler, F., & Subrahmanyam, M. G. (2014). The determinants of recovery rates in the US corporate bond market. Journal of Financial Economics, 114(1), 155-177.

Jorion, P., Liu, Z., & Shi, C. (2005). Informational effects of regulation FD: evidence from rating agencies. Journal of Financial Economics, 76(2), 309-330.

Jorion, P., & Zhang, G. (2007). Good and bad credit contagion: Evidence from credit default swaps. Journal of Financial Economics, 84(3), 860-883.

Jory, S. R., Ngo, T. N., & Wang, D. (2016). Credit ratings and the premiums paid in mergers and acquisitions. Journal of Empirical Finance, 39, 93-104.

Kadapakkam, P. R., Kumar, P. C., & Riddick, L. A. (1998). The impact of cash flows and firm size on investment: The international evidence. Journal of Banking & Finance, 22(3), 293-320.

Kaplan, S. N., & Zingales, L. (1997). Do investment-cash flow sensitivities provide useful measures of financing constraints? The Quarterly Journal of Economics, 112(1), 169-215.

Kashyap, A. K., Lamont, O. A., & Stein, J. C. (1994). Credit conditions and the cyclical behavior of inventories. The Quarterly Journal of Economics, 109(3), 565-592.

Kemper, K. J., & Rao, R. P. (2013). Do credit ratings really affect capital structure? Financial Review, 48(4), 573-595.

Khieu, H. D., & Pyles, M. K. (2012). The influence of a credit rating change on corporate cash holdings and their marginal value. Financial Review, 47(2), 351-373.

86

Kisgen, D. J. (2006). Credit ratings and capital structure. The Journal of Finance, 61(3), 1035-1072.

Kisgen, D. J. (2009). Do firms target credit ratings or leverage levels? Journal of Financial and Quantitative Analysis, 44(6), 1323-1344.

Kliger, D., & Sarig, O. (2000). The information value of bond ratings. The Journal of Finance, 55(6), 2879-2902.

Kraft, P. (2015). Do rating agencies cater? Evidence from rating-based contracts. Journal of Accounting and Economics, 59(2-3), 264-283.

Kuang, Y. F., & Qin, B. (2013). Credit ratings and CEO risk‐taking incentives. Contemporary Accounting Research, 30(4), 1524-1559.

Lamont, O., Polk, C., & Saá-Requejo, J. (2001). Financial constraints and stock returns. The Review of Financial Studies, 14(2), 529-554.

Langohr, H., & Langohr, P. (2008). The Ratings Agencies and Their Credit Rating. Wiley.

Leland, H. E., & Pyle, D. H. (1977). Informational asymmetries, financial structure, and financial intermediation. The Journal of Finance, 32(2), 371-387.

Lemmon, M., & Roberts, M. R. (2010). The response of corporate financing and investment to changes in the supply of credit. Journal of Financial and Quantitative Analysis, 45(3), 555-587.

Lewellen, J., & Lewellen, K. (2016). Investment and cash flow: New evidence. Journal of Financial and Quantitative Analysis, 51(4), 1135-1164.

Liu, Y., & Jiraporn, P. (2010). The effect of CEO power on bond ratings and yields. Journal of Empirical Finance, 17(4), 744-762.

Michaely, R., Thaler, R. H., & Womack, K. L. (1995). Price reactions to dividend initiations and omissions: Overreaction or drift? The Journal of Finance, 50(2), 573-608.

Miller, M. H., & Rock, K. (1985). Dividend policy under asymmetric information. The Journal of Finance, 40(4), 1031-1051.

Moody’s (2018). Annual Default Study: Corporate Default and Recovery Rates, 1920 – 2017. Moody’s Investor Services.

Moshirian, F., Nanda, V., Vadilyev, A., & Zhang, B. (2017). What drives investment–cash flow sensitivity around the World? An asset tangibility Perspective. Journal of Banking & Finance, 77, 1-17.

Mulier, K., Schoors, K., & Merlevede, B. (2016). Investment-cash flow sensitivity and financial constraints: Evidence from unquoted European SMEs. Journal of Banking & Finance, 73, 182-197.

Norden, L., & Weber, M. (2004). Informational efficiency of credit default swap and stock markets: The impact of credit rating announcements. Journal of Banking & Finance, 28(11), 2813-2843.

87

Odders-White, E. R., & Ready, M. J. (2005). Credit ratings and stock liquidity. The Review of Financial Studies, 19(1), 119-157.

Qi, Y., Roth, L., & Wald, J. K. (2010). Political rights and the cost of debt. Journal of Financial Economics, 95(2), 202-226.

Ramakrishnan, R. T., & Thakor, A. V. (1984). Information reliability and a theory of financial intermediation. The Review of Economic Studies, 51(3), 415-432.

Rauh, J. D. (2006). Investment and financing constraints: Evidence from the funding of corporate pension plans. The Journal of Finance, 61(1), 33-71.

Santos, T. (2006). Comment on: Credit risk transfer and contagion. Journal of Monetary Economics, 53(1), 113-121.

Schaefer, S. M., & Strebulaev, I. A. (2008). Structural models of credit risk are useful: Evidence from hedge ratios on corporate bonds. Journal of Financial Economics, 90(1), 1-19.

Standard & Poor’s (2018). 2017 Annual Global Corporate Default Study and Rating Transitions. Standard & Poor’s Financial Services.

Steiner, M., & Heinke, V. G. (2001). Event study concerning international bond price effects of credit rating actions. International Journal of Finance & Economics, 6(2), 139-157.

Sufi, A. (2007). The real effects of debt certification: Evidence from the introduction of bank loan ratings. The Review of Financial Studies, 22(4), 1659-1691.

Tang, T. T. (2009). Information asymmetry and firms’ credit market access: Evidence from Moody's credit rating format refinement. Journal of Financial Economics, 93(2), 325-351.

Wansley, J. W., Glascock, J. L., & Clauretie, T. M. (1992). Institutional bond pricing and information arrival: The case of bond rating changes. Journal of Business Finance & Accounting, 19(5), 733-750.

88

Appendices

Appendix 1. Variable definitions Variable Description (Compustat code)

Real size Total assets (at) deflated using the 2005 year CPI (cpi)

Age

Current year minus date of IPO (year-ipodate) (data for companies listed before 1980 collected from CRSP and company websites)

Q

[Market value of equity (prcc_f*cshpri) + liquidation value of preferred stock (pstkl) + total debt (dltt+dlc) – deferred taxes (txditc)]/Total assets (at)

Industry Q Average SIC3 Q

Sales growth Net sales growth (sale/L.sale)-1

Industry sales growth Average SIC3 sales growth

Capital expenditures Capital expenditures (capx)/Beginning-of-year total assets (L.at)

R&D R&D expense (xrd)/Beginning-of-year total assets (L.at)

Return on assets Earnings before interest, depreciation and amortization (oibdp)/ Beginning-of-year total assets (L.at)

Operating cash flow Operating cash flow (oancf)/Beginning-of-year total assets (L.at)

Leverage Total debt (dltt+dlc)/Assets (at)

Cash holdings Cash and short term investments (che)/Assets (at)

Dividends Common dividends (dvc)/Beginning-of-year total assets (L.at)

Market capitalization Market capitalization (fiscal year end) (prcc_f*csho)

Excess prior return

Residuals from a regression of Q on contemporaneous, lagged, and squared cash flow and the first four lags of stock returns [log( prcc_f/l.prcc_f)]

Negative earnings dummy Positive income before extraordinary items (ib) dummy =1 if ib>0, else 0

Debt issues [Long-term debt issuance (dltis) – long-term debt reductions (dltr)]/Beginning-of-year assets (L.at)

Equity issues

[Sale of common and preferred stock (sstk) - purchase of common and preferred stock (prstkc)]/Beginning-of-year assets (L.at)

Tangibility Net property, plant, and equipment (ppent)/Beginning-of-year assets (L.at)

Underfunded pension plan Pension-projected benefit obligation (pplao) - Pension plan assets (pbpro)

Equity recycling

[Cash dividends (dv) + purchase of common and preferred stock (prstkc)]/Sale of common and preferred stock (sstk)/Beginning-of-year assets (L.at)

Rating Long-term domestic issuer rating (splticrm)

KZ score

–1.001909*[ (ib+dp)/L.ppent] + 0.2826389* [(at+prcc_f*csho-ceq-txdb)/at] + 3.139193*[(dltt+dlc)/(dltt+ dlc+seq)] – 39.3678*[(dvc+dvp)/L.ppent] – 1.314759*(che/L.ppent)

SA score –0.737*log(real size) + 0.043*log(real size)2 – 0.040*age, with real size capped at and age capex at 35 years

WW score

–0.091* [(ib+dp)/at] – 0.062*[positive dividends dummy] + 0.021*(dltt/at) – 0.044*log(real size) + 0.102*(ind_sale_g) – 0.035*(sale_g)

DI score Obtained from Hoberg and Maksimovic (2014)

EF score Obtained from Hoberg and Maksimovic (2014)

DF score Obtained from Hoberg and Maksimovic (2014)

89

Appendix 2. Corporate liquidity events: 4-way sort The table reports results from logit regressions of dividend omissions (dummy variable set to 1 if the firm stops paying dividends), dividend increases (dummy variable set to 1 if the firm increases its dividend per share relative to the previous fiscal year), and if the firm has an underfunded pension plan (dummy variable set to 1 if the firm has an underfunded pension plan), and panel regressions of if the firm recycles equity (the ratio of cash dividends plus purchases of common and preferred stock to the sale of common and preferred stock and normalized by beginning-of-period total assets). Ratings are sorted into A and higher, BBB, BB, and B and lower, whereas the comparison scores are sorted into quartiles. The main test variables are the dummies for each of the four groups. “Low” (“High”) refers to firm-years with the strongest ratings/lowest scores (weakest ratings/highest scores). In Panel A, we report results for credit ratings. In panel B, we report estimated coefficients for the KZ, SA, WW, DI, EF, and DF scores. We in all regressions include as control variables the natural logarithm of market capitalization, Q, the excess prior year buy-and-hold stock return, a dummy variable set to one if the firm reported negative earnings in the previous fiscal year, and period and industry fixed effects. In the model for underfunded pension plan we additionally include a dummy variable set to one if the firm had an underfunded pension plan in the previous fiscal year. T-statistics in parenthesis. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Panel A Dividend omission

Dividend increase

Underfunded pension plan

Equity recycling

Low rating -4.908*** -1.775*** -1.705*** 0.053***

(-5.48) (-5.24) (-5.67) (12.48)

2nd to lowest -3.213*** -2.424*** -1.472*** 0.049***

(-4.95) (-7.87) (-5.02) (12.63)

2nd to highest -1.489** -2.957*** -1.309*** 0.048***

(-2.46) (-10.15) (-5.09) (12.57)

High rating -0.780 -3.077*** -1.214*** 0.047***

(-1.38) (-13.49) (-4.19) (12.64)

Log(market capitalization) -0.121 0.100*** -0.094** -0.006***

(-1.37) (3.13) (-2.37) (-11.21)

Log(Q) -0.110 0.630*** -0.078 0.007***

(-0.37) (6.79) (-0.93) (7.26)

Excess prior return -0.514** -0.210 0.094 -0.005***

(-2.48) (-1.35) (0.84) (-5.66)

Negative earnings dummy -0.720*** 1.374*** -0.031 0.004***

(-3.57) (11.99) (-0.28) (3.85)

High ≠ Low? Yes (***) Yes (***) Yes (***) Yes (**)

Industry fixed effects Yes Yes Yes Yes

Period fixed effects Yes Yes Yes Yes

No observations 7,346 7,008 6,968 7,983

McFadden R2/R2 --- --- --- 0.131

90

Appendix 2. Corporate liquidity events: 4-way sort (continued)

Panel B

Dividend omission

Dividend increase

Underfunded pension plan

Equity recycling

KZ score = 1 0.124 -2.510*** -1.243*** 0.058***

(0.19) (-9.25) (-3.48) (11.81)

KZ score = 4 1.009* -2.985 -0.600** -0.048***

(1.76) (-13.39) (-1.96) (11.87)

Accurate prediction? Yes (***) Yes (***) Yes (***) Yes (***)

SA score = 1 -0.106 -3.077*** -1.109*** 0.047***

(-0.13) (-10.37) (-3.77) (10.72)

SA score = 4 0.602 -2.852*** -0.627** 0.050**

(0.95) (11.77) (2.38) (11.22)

Accurate prediction? Yes (**) No Yes (***) No

WW score = 1 0.714 -3.162*** -1.220*** 0.057***

(0.48) (-8.19) (-3.04) (8.69)

WW score = 4 0.917 -3.059*** -0.669** 0.049***

(1.11) (-9.98) (-2.13) (10.16)

Accurate prediction? No No Yes (**) Yes (***)

DI score = 1 -0.938** -0.300 0.021 0.011**

(-2.57) (-1.11) (0.12) (2.92)

DI score = 4 -0.135 -0.309 0.447** 0.010**

(-0.42) (-1.18) (2.42) (2.30)

Accurate prediction? Yes (**) No Yes (***) No

EF score = 1 -0.846** -0.374 0.067 0.009**

(-2.03) (-1.31) (0.28) (2.49)

EF score = 4 -0.053 -0.323 0.464** 0.010**

(-0.19) (-1.13) (2.44) (2.41)

Accurate prediction? Yes (**) No Yes (***) No

DF score = 1 -0.669** -0.189 -0.334 0.008**

(-1.97) (-0.81) (-1.62) (2.70)

DF score = 4 -0.389 -0.483 0.420* 0.012**

(-0.91) (-1.51) (1.84) (2.91)

Accurate prediction? No Yes (**) Yes (***) No

91

Ap

pen

dix

3. I

nve

stm

ent-

cash

flo

w s

ensi

tivi

ty a

nd

fin

anci

al c

on

stra

ints

: ad

dit

ion

al c

on

tro

l var

iab

les

The

tab

le r

epor

ts e

stim

ates

fro

m p

anel

reg

ress

ions

of

cape

x on

cas

h f

low

, a

co

nstr

ain

ts c

lass

ifier

(w

ith t

he in

clud

ed c

onst

rain

ts c

lass

ifier

giv

en b

y th

e co

lum

n he

adin

g),

an in

tera

ctio

n te

rm b

etw

een

cash

flow

and

the

con

stra

ints

cla

ssifi

er,

and

begi

nnin

g-of

-per

iod

Q, p

lus

cont

rol v

aria

bles

as

spec

ified

in C

olum

n 1.

Q is

inst

rum

ente

d us

ing

the

Lew

elle

n an

d Le

wel

len

(201

6) I

V e

stim

ator

, ex

cept

for

Col

umn

2 (3

), w

here

Q is

inst

rum

ente

d by

the

sec

ond

lag

of f

irst-

diffe

renc

ed Q

(th

ird a

nd f

ourt

h la

gs o

f fir

st-

diffe

renc

ed Q

, ca

sh f

low

, le

vera

ge,

and

cash

hol

ding

s).

All

mod

els

incl

ude

firm

and

per

iod

fixed

effe

cts.

Sta

ndar

d er

rors

are

het

eros

ceda

stic

ity-c

onsi

sten

t an

d cl

uste

red

at t

he

firm

leve

l. T

-sta

tistic

s in

pa

ren

thes

is. *

,**

an

d **

* de

note

sta

tistic

al s

igni

fican

ce a

t the

10%

, 5%

and

1%

leve

ls.

Pan

el A

R

atin

g(1)

R

atin

g(2)

R

atin

g(3)

R

atin

g(4)

K

Z (

5)

SA

(6)

W

W (

7)

DI

(8)

EF

(9)

D

F (

10)

Ope

ratin

g ca

sh fl

ow (

C)

0.00

8 -0

.050

-0

.043

0.13

1**

* 0.

430**

* 0.

314**

* 0.

138**

* 0.

143**

* 0.

137**

*

(0

.29)

(-

0.99

) (-

0.74

)

(9.8

0)

(3.3

8)

(4.7

0)

(7.9

6)

(7.7

7)

(7.7

2)

C +

Cas

h ho

ldin

gs

-0

.004

(-0.

15)

Con

stra

ints

cla

ssifi

er (

F)

-0.0

01**

* -0

.002

**

-0.0

01**

-0.0

03**

* -0

.000

***

-0.0

16**

* -0

.036

-0

.033

**

-0.0

42**

-0.0

11

(-

2.73

) (-

2.39

) (-

2.11

) (-

3.40

) (-

2.74

) (-

2.93

) (-

1.51

) (-

2.15

) (-

2.48

) (-

0.44

)

F ×

C

0.01

2**

* 0.

018**

* 0.

016**

*

0.00

1***

0.07

5***

0.46

8***

0.24

2* 0.

311**

0.

050

(3

.97)

(3

.92)

(3

.30)

(3.1

7)

(3.3

8)

(2.9

8)

(1.8

1)

(2.0

3)

(0.2

2)

F ×

(C

+ C

ash

hold

ings

)

0.00

8***

(3.2

7)

Q

0.01

6***

0.02

3***

0.02

0***

0.02

4***

0.01

3***

0.01

4***

0.01

4***

0.01

9***

0.01

9***

0.01

9***

(8

.83)

(6

.20)

(5

.06)

(9

.29)

(7

.61)

(7

.96)

(7

.97)

(7

.01)

(7

.04)

(6

.98)

Leve

rage

-0

.041

***

-0.0

48**

* -0

.050

***

-0.0

64**

* -0

.040

***

-0.0

41**

* -0

.042

***

-0.0

42**

* -0

.042

***

-0.0

43**

*

(-

7.20

) (-

6.08

) (-

5.79

) (-

4.04

) (-

7.26

) (-

7.36

) (-

7.41

) (-

5.40

) (-

5.39

) (-

5.48

)

Cas

h ho

ldin

gs

0.00

8 -0

.023

-0

.013

0.01

0 0.

011

0.01

2 0.

014

0.01

4 0.

014

(0

.80)

(-

1.58

) (-

0.92

)

(0.9

8)

(1.0

5)

(1.1

2)

(0.9

4)

(0.9

4)

(0.9

4)

F ×

Lev

erag

e

0.00

1

(1.1

4)

Sal

es g

row

th

0.02

3***

0.

030**

* 0.

029**

* 0.

029**

* 0.

024**

* 0.

024**

* 0.

024**

*

(6

.95)

(7.9

5)

(7.8

2)

(7.8

9)

(10.

25)

(5.9

2)

(5.9

1)

92

Ap

pen

dix

3. I

nve

stm

ent-

cash

flo

w s

ensi

tivi

ty a

nd

fin

anci

al c

on

stra

ints

: ad

dit

ion

al c

on

tro

l var

iab

les

(co

nti

nu

ed)

Log(

real

siz

e)

-0.0

09**

*

-0.0

06**

* -0

.007

***

-0.0

06**

* -0

.009

***

-0.0

09**

* -0

.009

***

(-

6.51

)

(-5.

63)

(-5.

45)

(-4.

01)

(-3.

99)

(-3.

97)

(-4.

01)

Log(

age)

0.

004

0.

003

0.00

0 0.

003

0.00

7*

0.00

7* 0.

007*

(1

.44)

(1.0

3)

(0.0

4)

(1.1

7)

(1.7

2)

(1.7

1)

(1.7

5)

Cap

ext-

1

0.38

3***

0.

403**

* 0.

404**

* 0.

405**

* 0.

353**

* 0.

353**

* 0.

353**

*

(1

4.74

)

(15.

09)

(15.

21)

(15.

22)

(10.

25)

(10.

24)

(10.

25)

N

9.07

9 8.

502

6.60

6 9.

151

8.61

0 8.

635

8.61

9 5.

180

5.18

0 5.

180

R2

0.38

7 0.

212

0.22

7 0.

197

0.38

8 0.

389

0.38

9 0.

334

0.33

5 0.

333

Pan

el B

K

Z s

core

(1)

S

A s

core

(2)

W

W s

core

(3)

D

I sco

re (

4)

EF

sco

re (

5)

DF

sco

re (

6)

Ope

ratin

g ca

sh fl

ow (

C)

0.18

6***

0.41

9***

0.28

5***

0.19

4***

0.19

9***

0.19

3***

(1

0.30

) (3

.33)

(3

.23)

(8

.37)

(8

.21)

(8

.39)

Con

stra

ints

cla

ssifi

er (

F)

-0.0

00**

* -0

.010

-0

.013

-0

.032

-0

.039

* -0

.012

(-

3.63

) (-

1.51

) (-

0.48

) (-

1.61

) (-

1.76

) (-

0.43

)

F ×

C

0.00

2**

* 0.

059*

0.26

5 0.

214

0.32

1 0.

055

(3

.19)

(1

.96)

(1

.23)

(1

.19)

(1

.64)

(0

.22)

Q

0.01

9***

0.02

0***

0.01

9***

0.02

1***

0.02

1***

0.02

1***

(7

.19)

(7

.44)

(7

.27)

(6

.27)

(6

.29)

(6

.26)

Deb

t is

sues

0.

025**

* 0.

025**

* 0.

025**

* 0.

030**

* 0.

031**

* 0.

030**

*

(5

.83)

(5

.78)

(5

.76)

(5

.55)

(5

.60)

(5

.58)

Equ

ity is

sues

0.

109**

* 0.

111**

* 0.

108**

* 0.

104**

* 0.

103**

* 0.

103**

*

(4

.12)

(4

.17)

(4

.05)

(3

.01)

(2

.99)

(2

.99)

Tan

gibi

lity

0.04

9***

0.04

8***

0.04

9***

0.01

1 0.

011

0.01

0

(4

.10)

(4

.00)

(4

.16)

(0

.58)

(0

.57)

(0

.53)

N

8.05

2 8.

027

8.05

8 4.

832

4.83

2 4.

832

R2

0.23

3 0.

232

0.23

3 0.

200

0.20

1 0.

199

93

Ap

pen

dix

4. I

nve

stm

ent-

cash

flo

w s

ensi

tivi

ty a

nd

fin

anci

al c

on

stra

ints

: 4-

wa

y so

rt

Pan

el A

rep

orts

est

imat

es f

rom

pan

el r

egre

ssio

ns o

f ca

pex

on o

pera

ting

cash

flo

w (

C),

beg

inni

ng-o

f-pe

riod

inst

rum

ente

d Q

, le

vera

ge,

and

cash

hol

ding

s on

firm

-yea

rs

with

hig

h an

d lo

w r

atin

gs/c

ompa

rison

sco

res.

Rat

ings

are

sor

ted

into

A a

nd h

ighe

r, B

BB

, BB

, an

d B

and

low

er,

whe

reas

the

com

paris

on s

core

s ar

e so

rted

into

qua

rtile

s.

Mod

els

deno

ted

“Low

” (“

Hig

h”)

are

estim

ated

on

firm

-yea

rs w

ith s

tron

g ra

tings

/low

sco

res

(wea

k ra

tings

/hig

h sc

ores

). P

anel

B c

onta

ins

robu

stne

ss t

ests

for

the

rat

ings

m

odel

s in

whi

ch w

e va

ry th

e ra

nge

of c

ontr

ol v

aria

bles

by

repl

acin

g le

vera

ge a

nd c

ash

hold

ings

by

net d

ebt a

nd e

quity

issu

es a

nd b

egin

ning

-of-

perio

d ta

ngib

ility

or

addi

ng

sale

s gr

owth

, ag

e an

d be

ginn

ing-

of-p

erio

d re

al s

ize

and

cape

x (m

odel

s in

row

s). A

ll m

odel

s in

clud

e fir

m a

nd p

erio

d fix

ed e

ffect

s. S

tand

ard

erro

rs a

re h

eter

osce

dast

icity

-co

nsis

tent

and

clu

ster

ed a

t the

firm

leve

l. T

-sta

tistic

s in

par

enth

esis

. *,

** a

nd *

** d

enot

e st

atis

tical

sig

nific

ance

at

the

10%

, 5%

and

1%

leve

ls.

Pan

el A

Lo

w

ratin

g (1

)

Hig

h ra

ting

(2)

Low

KZ

(3

) H

igh

KZ

(4

) Lo

w S

A

(5)

Hig

h S

A

(6)

Low

WW

(7

)

Hig

h W

W

(8)

Low

DI

(9

) H

igh

DI

(10)

Lo

w E

F

(11)

H

igh

EF

(1

2)

Low

DF

(1

3)

Hig

h D

F

(14)

C

0.12

7**

* 0.

194**

* 0.

046**

* 0.

339**

* 0.

149**

* 0.

210**

* 0.

166**

* 0.

225**

* 0.

092**

* 0.

215**

* 0.

091**

* 0.

167**

* 0.

155**

* 0.

194**

*

(4

.50)

(5

.52)

(3

.63)

(9

.52)

(6

.02)

(4

.33)

(5

.05)

(6

.12)

(4

.02)

(4

.09)

(3

.73)

(3

.67)

(3

.96)

(4

.70)

Q

0.01

8***

0.04

1***

0.01

6***

0.03

5***

0.01

6***

0.03

4***

0.02

1***

0.03

1***

0.02

4***

0.02

5**

0.02

8***

0.02

2**

0.02

6***

0.03

1***

(3

.96)

(5

.80)

(5

.86)

(5

.28)

(4

.80)

(3

.03)

(5

.29)

(6

.02)

(6

.49)

(2

.47)

(4

.92)

(2

.32)

(3

.74)

(3

.95)

Leve

rage

-0

.038

***

-0.0

53**

* -0

.016

**

-0.1

05**

* -0

.072

***

-0.0

34

-0.0

73**

* -0

.037

***

-0.0

41**

* -0

.042

* -0

.042

***

-0.0

51**

-0.0

38**

-0.0

84**

(-

2.59

) (-

3.23

) (-

2.19

) (-

6.47

) (-

6.68

) (-

1.49

) (-

4.41

) (-

2.65

) (-

3.45

) (-

1.81

) (-

3.01

) (-

2.10

) (-

2.45

) (-

3.87

)

Cas

h -0

.044

**

-0.0

46

-0.0

02

0.07

2 -0

.031

0.

076

-0.0

50**

0.07

9**

-0.0

25**

0.07

0 -0

.032

* 0.

069

-0.0

24

0.02

9

(-

2.25

) (-

1.48

) (-

0.20

) (1

.43)

(-

1.34

) (0

.83)

(-

2.08

) (2

.08)

(-

2.01

) (1

.20)

(-

1.74

) (1

.17)

(-

0.85

) (0

.97)

Hig

h >

Low

?

Yes

(* )

Y

es (

*** )

N

o N

o Y

es (

**)

Yes

(* )

N

o

N

2,67

2 1,

304

2,17

8 2,

193

3,06

5 1,

016

3,14

7 1,

283

1,26

4 1,

179

1,19

3 1,

195

1,20

2 1,

167

R2

0.30

8 0.

200

0.30

2 0.

271

0.33

3 0.

093

0.31

7 0.

231

0.22

4 0.

170

0.22

3 0.

175

0.20

5 0.

263

94

Ap

pen

dix

4. I

nve

stm

ent-

cash

flo

w s

ensi

tivi

ty a

nd

fin

anci

al c

on

stra

ints

: 4-

wa

y so

rt (

con

tin

ued

)

Pan

el B

C

Q

Le

vera

ge

Cas

h ho

ldin

gs

Deb

t is

sues

E

quity

is

sues

T

angi

bilit

y S

ales

gr

owth

A

ge

Rea

l si

ze

Cap

ext-

1

N

R2

Low

rat

ing

0.09

8***

0.01

7***

0.04

4***

0.14

0**

0.09

0***

2,45

8 0.

337

(3

.37)

(3

.58)

(3

.93)

(2

.15)

(3

.35)

Hig

h ra

ting

0.20

0***

0.03

1***

0.02

4***

0.10

2***

-0.0

30

1,26

6 0.

217

(6

.02)

(4

.77)

(3

.89)

(3

.32)

(-

1.12

)

Hig

h >

Low

? Y

es (

*** )

Low

rat

ing

0.06

0***

0.00

9***

-0.0

47**

* -0

.018

*

0.04

2***

0.00

3 -0

.008

***

0.43

7***

2,66

1 0.

522

(3

.70)

(2

.98)

(-

4.68

) (-

1.65

)

(5.3

6)

(0.7

3)

(-3.

41)

(8.5

8)

Hig

h ra

ting

0.18

4***

0.03

3***

-0.0

50**

* 0.

054*

0.

022**

-0

.009

-0

.015

***

0.17

6***

1,29

1 0.

257

(5

.45)

(5

.38)

(-

3.21

) (1

.90)

(2.5

1)

(-0.

81)

(-2.

95)

(3.4

7)

Hig

h >

Low

? Y

es (

*** )

95

Ap

pen

dix

5. I

nve

stm

ent

in s

ho

ck p

erio

ds

and

fin

an

cial

co

nst

rain

ts:

add

itio

nal

co

ntr

ol v

aria

ble

s T

he t

able

con

tain

s m

odel

s in

whi

ch t

he c

hang

e in

cap

ex is

mea

sure

d as

the

diff

eren

ce in

cap

ex in

200

9 vs

. 20

06 a

nd c

ontr

ol v

aria

bles

are

hel

d co

nsta

nt a

t th

eir

2006

le

vel.

Col

umns

1-3

rep

ort

re-e

stim

atio

ns o

f C

olum

ns 3

-5 in

Pan

el A

of

Tab

le 6

whe

re w

e e

xclu

de c

ontr

ol v

aria

bles

tha

t ar

e in

clud

ed in

the

cal

cula

tion

of th

e K

Z,

SA

, an

d W

W in

dexe

s, a

s th

ey m

ay

caus

e m

ultic

ollin

earit

y pr

oble

ms.

In C

olum

ns 4

-9, w

e in

stea

d in

clud

e be

ginn

ing-

of-p

erio

d Q

, ope

ratin

g ca

sh fl

ow, l

ever

age,

and

cas

h ho

ldin

gs

as c

ontr

ol v

aria

bles

. T-

stat

istic

s in

par

enth

esis

. *,

** a

nd *

** d

enot

e st

atis

tical

sig

nific

ance

at t

he 1

0%, 5

% a

nd 1

% le

vels

.

K

Z s

core

(1

) S

A s

core

(2

) W

W s

core

(3

) K

Z s

core

(4

) S

A s

core

(5

) W

W s

core

(6

) D

I sco

re

(7)

EF

sco

re

(8)

DF

sco

re

(9)

Con

stra

ints

cla

ssifi

er

-0.0

01**

* -0

.003

-0

.091

-0

.001

***

-0.0

02

-0.0

91**

-0.0

01

-0.0

01

-0.0

92

(-

4.33

) (-

0.57

) (-

0.78

) (-

3.71

) (-

0.36

) (-

2.19

) (-

0.02

) (-

0.04

) (-

1.18

)

Q

0.

035**

* 0.

036**

* 0.

043**

* 0.

045**

* 0.

045**

* 0.

025*

0.02

5* 0.

024

(4.0

4)

(4.0

2)

(5.1

2)

(5.2

3)

(5.2

9)

(1.6

9)

(1.6

9)

(1.6

2)

Log(

age)

-0

.000

0.00

1

(-

0.04

)

(0.2

8)

Log(

real

siz

e)

0.00

3

-0.0

03

(0

.92)

(-0.

42)

Sal

es g

row

th

-0.0

90**

* -0

.072

**

(-

2.69

) (-

2.27

)

Indu

stry

sal

es g

row

th

0.09

1***

0.03

8

(2

.86)

(1

.08)

Ope

ratin

g ca

sh fl

ow

-0

.397

***

-0.3

87**

* -0

.394

***

-0.2

86**

* -0

.286

***

-0.2

89**

*

(-5.

81)

(-5.

70)

(-5.

78)

(-3.

90)

(-3.

89)

(-3.

96)

Leve

rage

-0.0

24

-0.0

24

-0.0

10

-0.0

23

-0.0

23

-0.0

22

(-1.

34)

(-1.

43)

(-0.

57)

(-1.

08)

(-1.

08)

(-1.

05)

Cas

h ho

ldin

gs

0.

092**

* 0.

101**

* 0.

112**

* 0.

100**

* 0.

100**

* 0.

093**

*

(2.8

7)

(3.2

1)

(3.4

8)

(3.0

4)

(3.0

3)

(2.7

4)

N

573

448

446

447

447

447

372

372

372

Adj

R2

0.07

3 0.

094

0.06

6 0.

239

0.22

9 0.

239

0.14

4 0.

144

0.14

7

96

Ap

pen

dix

6. I

nve

stm

ent

in s

ho

ck p

erio

ds

and

fin

anci

al c

on

stra

ints

: 4-

wa

y so

rt

The

tab

le c

onta

ins

mod

els

in w

hich

the

cha

nge

in c

apex

is m

easu

red

as t

he d

iffer

ence

in c

apex

in 2

009

vs.

2006

(P

anel

A)

or in

200

4 vs

. 20

02 (

Pan

el B

) an

d co

ntro

l va

riabl

es a

re h

eld

cons

tant

at

thei

r 20

06 le

vel.

The

mai

n te

st v

aria

bles

are

dum

mie

s se

t to

1 w

hen

the

4-gr

oupe

d fin

anci

al c

onst

rain

ts c

lass

ifier

tak

es t

he v

alue

1 o

r 4.

“L

ow”

(“H

igh”

) re

fers

to

firm

-yea

rs w

ith t

he s

tron

gest

rat

ings

/low

est

scor

es (

wea

kest

rat

ings

/hig

hest

sco

res)

. T

-sta

tistic

s in

par

enth

esis

. *,

** a

nd *

** d

enot

e st

atis

tical

si

gnifi

canc

e at

the

10%

, 5%

and

1%

leve

ls.

Pan

el A

R

atin

g(1)

K

Z s

core

(2)

SA

sco

re(3

) W

W s

core

(4)

DI s

core

(5)

EF

sco

re(6

) D

F s

core

(7)

Low

sco

re

-0.0

19

-0.0

05

-0.0

79**

-0.0

08

-0.0

07

-0.0

12

-0.0

06

(-

0.81

) (-

0.40

) (-

2.32

) (-

0.39

) (-

0.62

) (-

0.96

) (-

0.44

)

2nd to

low

est s

core

-0

.021

-0

.002

-0

.074

**

-0.0

19

-0.0

25*

-0.0

25*

-0.0

13

(-

1.12

) (-

0.19

) (-

2.52

) (-

1.23

) (-

1.77

) (-

1.71

) (-

0.94

)

2nd to

hig

hest

sco

re

-0.0

35**

-0.0

28**

-0.0

77**

* -0

.029

* -0

.016

-0

.018

-0

.019

(-

2.28

) (-

2.30

) (-

3.22

) (-

1.85

) (-

1.13

) (-

1.27

) (-

1.34

)

Hig

h sc

ore

-0.0

47**

-0.0

48**

* -0

.054

***

-0.0

29**

-0.0

14

-0.0

08

-0.0

17

(-

2.41

) (-

3.43

) (-

3.16

) (-

2.41

) (-

0.96

) (-

0.60

) (-

1.33

)

Q

-0.0

01

0.00

1 0.

010

***

0.00

8**

-0.0

02

-0.0

02

-0.0

03

(-

0.11

) (0

.36)

(2

.96)

(2

.51)

(-

0.28

) (-

0.29

) (-

0.34

)

Log(

age)

0.

008*

0.00

1 0.

009

0.00

1 0.

008

**

0.00

8**

0.00

8**

(1

.88)

(0

.28)

(1

.25)

(0

.28)

(2

.00)

(1

.97)

(2

.06)

Log(

real

siz

e)

-0.0

04

-0.0

01

0.00

2 -0

.005

-0

.006

* -0

.006

* -0

.007

**

(-

1.36

) (-

0.22

) (0

.81)

(-

1.34

) (-

1.90

) (-

1.87

) (-

1.99

)

Sal

es g

row

th

-0.0

80**

* -0

.079

***

-0.0

75**

-0.0

82**

* -0

.106

***

-0.1

07**

* -0

.106

***

(-

3.31

) (-

2.69

) (-

2.41

) (-

2.64

) (-

3.69

) (-

3.71

) (-

3.68

)

Indu

stry

sal

es g

row

th

0.05

4**

0.06

6**

0.06

2**

0.08

0***

0.02

2 0.

025

0.02

2

(2

.07)

(2

.44)

(2

.33)

(2

.63)

(0

.74)

(0

.86)

(0

.76)

Hig

h <

Low

? Y

es (

**)

Yes

(**

* )

No

No

No

No

No

N

514

465

467

466

401

401

401

R2

0.24

2 0.

272

0.21

3 0.

216

0.24

3 0.

243

0.12

4

97

Ap

pen

dix

6. I

nve

stm

ent

in s

ho

ck p

erio

ds

and

fin

anci

al c

on

stra

ints

: 4-

wa

y so

rt (

con

tin

ued

)

Pan

el B

R

atin

g (1

) K

Z s

core

(2

) S

A s

core

(3

) W

W s

core

(4

) D

I sco

re

(5)

EF

sco

re

(6)

DF

sco

re

(7)

Low

sco

re

0.00

6 0.

021**

* 0.

058**

* 0.

016

0.00

7 0.

008

0.01

1*

(0

.70)

(3

.44)

(3

.30)

(1

.56)

(1

.24)

(1

.40)

(1

.81)

2nd to

low

est s

core

0.

009

0.01

4**

0.04

9***

0.01

9**

0.00

9 0.

008

0.01

3**

(1

.30)

(2

.33)

(3

.14)

(2

.51)

(1

.39)

(1

.46)

(2

.19)

2nd to

hig

hest

sco

re

0.01

0 0.

021

***

0.04

9***

0.02

1***

0.00

9 0.

009

0.01

6**

(1

.54)

(3

.19)

(3

.82)

(2

.85)

(1

.39)

(1

.37)

(2

.53)

Hig

h sc

ore

0.02

6***

0.02

1***

0.02

9***

0.01

7**

0.02

2***

0.02

3***

0.01

1

(3

.23)

(3

.00)

(4

.12)

(2

.30)

(3

.28)

(3

.52)

(1

.64)

Q

-0.0

01

-0.0

02

-0.0

02

-0.0

02

-0.0

01

-0.0

01

-0.0

01

(-

0.75

) (-

1.22

) (-

1.58

) (-

1.33

) (-

0.77

) (-

0.83

) (-

0.85

)

Log(

age)

-0

.001

-0

.001

-0

.010

**

-0.0

02

0.00

0 -0

.000

-0

.001

(-

0.61

) (-

0.84

) (-

2.38

) (-

1.15

) (0

.08)

(-

0.01

) (-

0.47

)

Log(

real

siz

e)

-0.0

01

-0.0

03*

-0.0

04**

-0.0

02

-0.0

02

-0.0

02

-0.0

01

(-

0.71

) (-

1.72

) (-

2.36

) (-

0.97

) (-

0.97

) (-

1.00

) (-

0.75

)

Sal

es g

row

th

-0.0

23**

* -0

.029

**

-0.0

28**

-0.0

29**

-0.0

16**

-0.0

15**

-0.0

14**

(-

3.53

) (-

2.22

) (-

2.10

) (-

2.18

) (-

2.14

) (-

2.07

) (-

1.91

)

Indu

stry

sal

es g

row

th

-0.0

15

-0.0

10

-0.0

10

-0.0

11

-0.0

18

-0.0

20

-0.0

19

(-

1.22

1 (-

0.75

) (-

0.71

) (-

0.81

) (-

1.27

) (-

1.38

) (-

1.25

)

Hig

h >

Low

? Y

es (

**)

No

No

No

Yes

(**)

Yes

(**)

No

N

412

371

374

372

320

320

320

R2

0.07

0 0.

053

0.06

2 0.

047

0.06

8 0.

072

0.04

6

98

99

Chapter 3. The indexing effect: evidence from the bond market

3.1. Introduction

Numerous empirical papers have shown that stocks being added to widely followed indexes experience positive long-term risk-adjusted returns (e.g., Shleifer, 1986; Wurgler and Zhuravskaya, 2002; Chang et al., 2014). While this might seem at odds with the central tenet of the efficient market hypothesis,19 the preferred argument for the “index addition premium” relates to the excess demand by passive index funds or institutional investors which benchmark to these indexes (Basak and Pavlova, 2013). On the other hand, recent studies reveal a more nuanced story, challenging the long-held assumption that index additions are information-free events. Elliott et al. (2006) find that factors such as downward-sloping demand curves, liquidity constraints, changes in monitoring or information asymmetry, expectations on improved performance, media coverage and investor awareness can trigger a significant market response following the addition to the S&P index.

However, investigating the relative importance of these factors is a more challenging empirical task, since most of them have a simultaneous effect on the final price change following an index addition. As suggested in Chen et al. (2004), an empirical strategy, which may be effective for disentangling some of the confounding factors, compares index additions with other events, such as deletions or recalibrations. While most channels discussed in Elliott et al. (2006) lead to a symmetric response to additions and deletions, investor awareness or

19 According to the efficient market hypothesis, stocks have many substitutes and therefore their

demand curves should be flat.

100

recognition does not disappear when a bond is removed.20 More specifically, investors become aware of the addition of a specific bond or stock to a widely followed index, but they do not typically know when it is removed. This means that by comparing the index effect of additions with that of deletions, the “awareness” channel can be ruled out given that there is a symmetric response.

Yet due to the lack of sufficient observations and to stock index deletions generally being a consequence of bankruptcies or de-listings, there are few papers looking at changes in stock index composition changes involving both deletions and additions. The study with the largest sample of stock index deletions, Chen et al. (2004), finds an asymmetric effect: significant price increases for additions, but no price response for deletions. On the other hand, Chang et al. (2014), after controlling for endogeneity by using a regression discontinuity design, find a symmetric response.

The current paper contributes to the empirical evidence on the effect of additions relative to deletions, by looking instead at the changes in a widely followed bond index, formerly known as Lehman Brothers’ Investment-Grade (IG) Corporate Index. Debt instruments are a largely unexplored but equally interesting avenue for investigating index effects. Firstly, debt is the main source of financing for US firms (Securities Industry and Financial Markets Association, 2018). Secondly, there are important structural and pricing differences between equity and debt instruments; stocks are relatively more sensitive to changes in the underlying asset value, meaning that the results from the stock market may not necessarily translate to bonds (Bessembinder et al., 2008). Third, there is a significantly higher percentage of institutional investors and passive index replicators in the bond market,21 and therefore, if index effects are driven by e.g. benchmarking, price changes would be comparatively larger for bonds than for

20 For simplicity, I will refer to the factors which have a symmetric effect on additions and

deletions as “investment”-related (such as downward-sloping demand curves, liquidity constraints, changes in monitoring/information asymmetry, expectations on improved performance, benchmarking), while the ones having an asymmetric impact will be referred to as “awareness”-related (investor awareness/ attention/ recognition, media coverage).

21 In a Goldman Sachs report, approximately 66% of corporate equity is held by institutional investors, whereas the proportion raises to 89% for the bond market (Goldman Sachs, 2016).

101

stocks. Last but not least, major changes in bond indexes are less common and do not receive as much media coverage as additions to stock indexes.22

Yet to the best of my knowledge, there is only one paper focusing specifically on indexing effects for bond markets: Chen et al. (2014). By using an exogenous change in the inclusion criteria for Lehman Brothers’ corporate bond indexes announced in 2005, they find that bonds that were mechanically moved from the high-risk index (consisting of speculative- rated bonds, the HY index) to the larger, low-risk one (consisting of investment-rated bonds, the IG index), experienced a significant drop in their yield and a corresponding increase in their returns, consistent with a positive market reaction. However, since the 2005 announcement resulted in only one type of change involving the addition of certain bonds to the IG index, the authors could not test whether there is a symmetric effect for index deletions. Furthermore, the announcement date was in January 2005, when another change in Merrill Lynch’s Corporate Bond Index became effective.23 Differently from the Lehman Brothers change, the changes that Merrill Lynch made to their bond index inclusion criteria resulted in additions and deletions in the same index. Since there is partial overlap between the Lehman Brothers IG index and the Merrill Lynch Corporate Bond Index constituent lists, there is a risk that Chen et al.’s (2014) results are picking up Merrill Lynch index deletions rather than Lehman Brothers IG index additions.

In contrast, Lehman Brothers in 2003 announced a more substantial change to its IG index inclusion criteria by increasing the minimum offering amount and changing how the index rating was calculated. This resulted in index additions as well as deletions, allowing me to test for symmetry, and without any risk of confounding effects from changes to the inclusion criteria of other bond indexes.

By investigating the market response following the 2003 change for the various groups of affected bonds, this paper extends the literature focused on comparing stock index additions and deletions by providing further evidence on the mechanisms leading to price changes. Consistent with the investment channel and with Chang et al. (2014), I find a significant effect for all the bonds affected by the change in index constituent rules. However, the effect of index addition

22 For instance, in 2004, the first available year on Google Trends, the Google SVI (Search

Volume Index) showed a daily average of 53 searches in the US for “stock index” and 24 searches for “bond index.”

23 https://www.businesswire.com/news/home/20051013005328/en/Merrill-Lynch-Announces-Global-Bond-Index-Rules

102

is twice as large as the effect of index deletion, suggesting that investor awareness is a relatively more important factor, as in Chen et al. (2004).

More specifically, the bonds that were added to the IG index following the change in 2003 experienced a yield drop of 0.53% following the announcement, an effect that persisted throughout the next six months.24 Compared to Chen et al. (2014), the initial response in the period immediately following the announcement is roughly twice as large, whereas the long-term impact is similar. The bonds excluded from the index, as a result of the change in the size and rating criteria for the index, experienced a yield increase of 0.20-0.30% for up to 100 days after the announcement date. These changes are economically significant, translating into abnormal bond returns of approximately 2-3% over the longest event window.25 Putting this into perspective, Hand et al. (1992) find an average excess return of up to 2.25% in the period immediately following a rating change announcement. This suggests that index membership can have a price impact comparable with a fundamental change in the riskiness of a bond.

3.2. Institutional background

Since 1973, Lehman Brothers’ (now Barclays Capital) bond indices have been the most widely used benchmarks for fixed income investors (Chen et al., 2014). This paper focuses on the largest US corporate bond index, which had a market value of USD 150 million in 2016, a volume comparable with one of the major stock indexes, Russell 3000.26 Furthermore, Lehman Brothers’ bond indexes are “the most widely accepted benchmarks in the asset management industry, used by over 90% of the US institutional investors” (PR Newswire, 2003). Therefore, the inclusion criteria are clear and transparent, with detailed data for facilitating its replication by investors. The index is constructed based on a set of publicly available criteria related to various bond characteristics (Barclays, 2015). Lehman Brothers has redefined its index rating methodology three times since its creation, and this paper will focus on the second and most far-reaching change, which took place in 2003. In contrast with the 2005 re-definition used

24 The change was effective approximately three months after the announcement date.

25 These numbers are based on the CAR differences between the treated and matched group within the [-10; +100] window in Table 5.

26 https://www.bloomberg.com/markets/rates-bonds/bloomberg-barclays-indices

103

by Chen et al. (2014), this event affected a larger fraction of bonds (approximately 15% of the bonds included) both negatively (index deletion/transition to the HY index), and positively (index addition). Moreover, to the best of my knowledge, there was no other index-related announcement in the same period and therefore there is a lower risk of confounding effects.

On June 17, 2003, Lehman Brothers announced that, effective October 1, 2003, the liquidity constraint (which refers to the offering amount) would be raised from USD 150 million to 200 million. This change reduced the number of constituent bonds by 13% and decreased the market value of the index by 2%.27 A second major change concerned the index ratings. Previously, only Moody’s ratings were taken into account. Effective October 1, S&P ratings would also be considered, with the most conservative rating among the two used as a criterion for index inclusion.28

These changes provide an excellent quasi-natural experimental setting for testing whether there is an “indexing” effect for the affected bonds, since the re-definition was exogenous to company fundamentals, and not anticipated by the market.29 While Chen et al. (2014) find that bonds that moved from the high-yield (HY) to the investment-grade (IG) index experienced a significant drop in yields following the announcement date, the 2003 change allows me to test how the bond market reacted to three different types of actions.

Firstly, a significant number of bonds, which were only rated by S&P and had an investment-grade rating, were added to the index from October 1. If the market views index addition as a positive signal, the yields of the affected bonds would decrease. I will refer to this as the “addition effect,” and to the group of bonds affected by this change as the “SP group.”

Secondly, investment-grade bonds with a size between USD 150 million and 200 million would be removed from the index as a result of the change in the liquidity criterion. If index membership is valued by the investors, one would expect an increase in the yields following the announcement – a “deletion

27 https://www.plansponsor.com/lehman-shakes-up-bond-indices/

28 https://www.plansponsor.com/lehman-shakes-up-bond-indices/

29 In the press release, it is mentioned that the changes are unrelated to the constituent bonds, but rather to make the indexes more comparable with the competing indexes; the former change in the rules for index constituents took place in 1999 (PR Newswire, 2003)

104

effect.” Bonds with an offering amount within this range are referred to as the “Size group.”

Lastly, the bonds which have an IG Moody’s rating and a HY S&P rating would be also removed from the index, since the most conservative rating would be HY. Since the size criterion for the HY corporate index is still USD 150 million, all the bonds in this group were moved to the HY bond index.30 I refer to this as a “transition effect,” which should also result in a yield increase, since the HY index is associated with a greater risk, and fewer investors are allowed or willing to invest.31 The bonds affected by this change will be included in the “HY group.”32

By comparing the statistical and economic significance of the indicator variables for each of these groups, this paper is the first to test whether the bond market reacts in a similar manner to “good” (i.e. index additions) and “bad” news (index deletion or transition to HY). If the investment-related factors discussed in the introduction trigger a market response, then one would expect a similar effect for all three groups, consistent with the results in Chang et al. (2014). Conversely, if awareness is a relatively more important factor, the price effect would be relatively smaller for the events involving removal from the IG index, as in Elliott et al. (2006) and Chen et al. (2004).

3.3. Data and sample selection

For both bond- and company-specific data, I use S&P’s Capital IQ database. This database has several important advantages relative to TRACE, which has been used in previous research. Firstly, as opposed to TRACE, it has bond pricing data prior to 2005. Secondly, Capital IQ contains both bond- and issuer-specific data, including the respective ratings from Moody’s and S&P. In previous work, bond- and issuer-specific data had to be merged from different datasets (such as TRACE - FISD - COMPUSTAT), resulting in a decrease in

30 All other criteria for HY and IG indexes are very similar, and thus the differences are now

limited to size (200 for IG and 150 for HY) and index rating requirements (Barclays, 2015)

31 For instance, most financial and insurance institutions are not allowed to invest in securities below investment grade.

32 A schematic clarification of how bonds were assigned to the SP, Size or HY groups, used as controls or excluded from the sample is shown in Appendix 3.

105

observations and possible matching errors. Thirdly, Capital IQ has readily available price and yield data for frequently traded bonds.

For the initial bond screening, I collect all the corporate bonds that are active during the time period of my analysis and meet the criteria for inclusion in the Lehman Brothers Corporate Index (Barclays, 2015): (1) offering date before 2003; (2) denominated in USD; (3) fixed-rate coupon; (4) having at least one year until maturity (thus maturity date after 2005); (5) not convertible, redeemable or having any other special features.

Concerning credit ratings, it is mentioned in the index description that “unrated (senior) securities may use an issuer rating for index classification purposes if available.” Since issuer ratings are more prevalent than issue ratings, I exploit this information by including in my sample unrated securities, which are issued by rated companies. However, the “same-rating” assumption implies that the bond is “representative” in the sense that it is reflecting its issuer’s credit risk and does not have any characteristics that would lead to “notching.” As described in the rating methodologies of both Moody’s and S&P, the securities that typically fall into this category are “senior unsecured corporate debentures” (Moody’s, 2008). Consequently, I further filter my search results to match this description, while having an issue credit rating from either Moody’s or S&P or, if the respective issue is not rated, an issuer credit rating from at least one of the two CRAs. 33

Furthermore, by focusing only on one type of bond without any special features, I rule out potential confounding effects due to the smaller changes announced on the same date, which affected only a small number of specific bonds, such as, for example, step-ups (Vanderbilt Avenue Asset Management, 2003).

There are 1360 bonds corresponding to 1120 issuers matching these criteria. For these bonds, I collect data on maturity, coupon, offering amount, daily yield and price values. As mentioned in Chen et al. (2014), many of these bonds are rather illiquid, so the sample size drops to about 450 bonds with available yield data during 2003. While this number might seem very small in comparison to Chen et al. (2014), the sample of treated bonds is significantly larger in my case, since the 2003 change involved three changes as opposed to only one in 2005. More

33 Before 2005, Lehman Brothers did not take Fitch into account for the index rating, and therefore

these ratings are not included in the analysis.

106

specifically, 57 bonds were included in the treated group in Chen et al. (2014), while the total number of treated bonds in the current paper is 207.

In order to control for issuer-specific variables that could influence bond yields, I follow Chen et al. (2014) and collect data on various leverage, profitability, Tobin’s Q and tangibility measures. The variables are described in Appendix 1. Table 1 displays summary statistics for the various bond groups used in the main empirical analysis. The three treated groups, Size, HY and SP, have been defined in Section 3.2. The remaining bonds with similar features from the final sample that haven’t changed their index status as a result of the announcement are included in the control group (see Appendix 3). The number of observations vary, since not all issuers with available financial data have complete information on their bonds, and thus the number of observations used in the empirical analysis is somewhat lower.

In terms of both bond and issuer characteristics, the Size group is the most similar to the control group, with no statistically significant differences besides the offering amount. The HY group is, by construction, very different, since the bonds included have a lower creditworthiness: smaller size, higher coupons and leverage. Importantly for the main empirical analysis, there is no significant difference between issuer and issue rating, with minor differences between S&P and Moody’s (except for the HY group).

With few exceptions, the average values are similar to Chen et al. (2014). In terms of bond characteristics, a notable difference is that the bonds in my sample have significantly larger numbers for the years-to-maturity variable (4 years in Chen et al. (2014) compared to about 18 in my full sample). Since their dataset includes all existing bonds with available data, including non-index bonds and those with special features, which are generally issued by smaller and younger companies, this difference is not surprising.

107

Table 1. Sample averages The table summarizes the average values of relevant bond and issuer characteristics for all groups included in the analysis with non-missing yield data on the announcement date. *,** and *** denote statistical significance at the 10%, 5% and 1% levels for the equal mean t-test relative to the control group. Here, the control group consists of all comparable index bonds (senior unsecured, fixed-coupon vanilla bonds) that were not included in the other samples, i.e. were not affected by the changes announced on June 17. The rating averages were approximated to the nearest integer value and written using S&P's letter notation for easier comparability. Financial variables were collected in the quarter prior to the announcement date. All variables are defined in Appendix 1.

Panel A. Bond characteristics

Control group Size group High-Yield Group S&P Group

Offering amount (MUSD) 571.18 177.99*** 309.93*** 756.25***

Coupon (%) 7.49 7.57 7.81 6.56*

Maturity (years) 28.68 31.32 19.95*** 7.94***

Age (years) 24.39 24.4 15.69* 6.47***

Moody's Issue Rating A- A- BBB- NA

S&P's Issue Rating A- A- BB+*** A-

Number of bonds 141 105 46 56

Panel B. Issuer characteristics Control Group Size group High-Yield Group S&P Group

Profitability 0.18 0.17 0.14 0.14

Tangibility 0.37 0.39 0.40 0.44

Leverage 0.32 0.29 0.45** 0.31

Tobin's Q 1.41 1.48 1.19 1.39

Interest coverage 0.10 0.11 0.041*** 0.08

Size 10.28 10.44 8.62*** 10.53

Moody's Issuer Rating BBB+ A- BBB- BBB+

S&P's Issuer Rating BBB+ A- BB+ A-

Number of issuers 132 94 37 44

108

3.4. Empirical strategy

Identifying a “pure” index effect, above and beyond its information about the bond’s or issuer’s riskiness, is empirically challenging since index addition or deletion is endogenously determined by specific characteristics, which may also be manipulated by the firm. For example, Kisgen (2006) shows how firms adjust their leverage to “match” a specific rating group; since bond index inclusion generally depends on having a specific rating, this means that index addition following a rating change is, at least partially, endogenously determined within the firm.

As discussed in Section 3.1, the 2003 change in the inclusion criteria for Lehman Brothers’ IG index is an excellent setting for testing whether there is such an effect, since these changes were not driven by firm-specific information. The ratings, as well as the outstanding bond size, were the same before and after the event, meaning that the change in yields/returns would only be driven by the “index effect.”

Before delving into the main empirical analysis, I provide some informal evidence for the index effect by looking at average cumulative returns (CR) of the bonds included in the analysis during year 2003.34

As seen in Figure 1, the bonds that were negatively affected by the change, i.e., the HY and Size groups, experience a negative return relative to the control group following the announcement date (day 0), marked with a dotted line. The SP group, conversely, exhibits a positive return after the announcement date. The changes are large relative to both the pre-event and control group values (between 20 and 40 basis points) and compare favorably to previous research on abnormal bond returns (Bessembinder et al., 2008). HY bonds seem to have a smaller relative change since their pre-event returns are already low, but, as discussed in Bessembinder et al. (2008), high-risk bonds are more sensitive to yield/price changes, and thus smaller “abnormal returns” are to be expected.

34 In my empirical analysis, I use both yield and return data in order to make my results comparable

with Chen et al. (2014). The values are cumulative and are derived from the yield-to-maturity and price variables collected on a daily basis from Capital IQ. The two variables are negatively correlated, i.e., “bad” news should lead to higher yields and lower returns.

109

Fig. 1. Average cumulative bond returns over time This figure plots the average cumulative return for the bonds affected by the changes in index inclusion criteria, relative to the index bonds which weren't affected. The HY group consists of the bonds with a speculative S&P rating and an investment-grade Moody's rating, the SIZE group includes all bonds with an outstanding amount below USD 200 million, and the SP group bonds are only rated (investment-grade) by S&P. The dotted line, ALL, represents the control bonds, i.e. all the index constituents which were not affected by the changes. On the horizontal axis, day 0 is the announcement date (June 17, 2003), while day 106 marks the effective date.

To this end, there are two important aspects for my further analysis. First, prior to the announcement date, bond prices seem to correlate well in all subsamples, confirming that the post-event changes are driven by the “index effect” and not by other omitted variables. However, there is still a possibility that unobserved factors or events happened during the same time, contaminating the results. In addition to focusing on a short event window, I will also use a matched-return approach to address this concern. Second, the effect of the announcement seems to be long-lived; the difference between the treated and the control groups persists across the entire period of 165 days after the event. While this might be partly the consequence of the fact that there are approximately 100 days between the announcement and the effective date (second dotted line), the effective date does not seem to have any visible impact on any of the groups. For this reason, I choose to focus on the announcement date.35

35 Chen et al. (2014) also don’t find any significant impact on the effective date, suggesting that

it wasn’t new information to the market.

110

Having established that the announcement has triggered a significant market response for the affected bonds, I further test whether these changes are statistically significant and not driven by other bond- or issue-specific factors in the next sections.

To do this, I follow Chen et al. (2014) by estimating cumulative yield changes over different time horizons as a function of the treatment indicator and bond- and issue-specific control variables:

∆ = + + + (1)

Where ∆ is the change in the cumulative yields between the last and the first day of the different event windows, is an indicator variable taking the value of 1 for the treated bonds36 (therefore is the coefficient of interest), and is a vector of bond- and issuer-specific control variables. As in Chen et al. (2014), the set of control variables includes bond size, offering amount, age, coupon, issuer profitability, tangibility, leverage, interest coverage, Tobin’s Q, size and an indicator variable for the broader S&P rating groups.37 All variables are defined in detail in Appendix 1. The regressions are estimated using OLS, industry fixed effects based on the 2-letter SIC and issuer-clustered errors. All variables are winsorized at 1% respectively 99% levels.

For comparability purposes, the estimation and event windows are similar to Chen et al. (2014). Since the date of implementation (106 days after the announcement day) doesn’t result in any visible market reaction (as seen in Fig.1), I only focus on the announcement day. To reduce the risk of contamination, I choose a relatively short event window of up to 60 days before and after the announcement date.

36 will therefore be different for each of the three treated groups, S&P, SIZE and HY, resulting

in three different regression outputs.

37 Several control variables used in Chen et al. (2014) were not included as they were either highly correlated with the existing ones (e.g. maturity and age and Moody’s ratings with S&P), or they significantly reduced the number of observations due to lack of data availability (e.g. R&D expenses). In Appendix 4, I include a number of alternative control variables, which don’t affect significantly the coefficients of interest.

111

3.5. Index effect results

3.5.1. The “addition” effect

Table 2 reports the regression output for the bonds that were positively affected by the changes announced on June 17, the SP group.

Table 2. Addition effect The table reports the full sample cross-section regression output for cumulative yield changes (measured as the cumulative yields for T+1 and T, respectively, for the time windows defined in the parentheses) as a function of bond- and issuer-specific variables. The variable of interest, “SP dummy,” takes value 1 for the bonds that have been included in the index as a result of the change, and therefore should capture the addition effect. The rating variables (A, BBB, BB, and below BB) are indicator variables based on the S&P overall rating (issue or issuer if issue is not available) grouped by wider rating groups (i.e. BBB group includes BBB-, BBB and BBB+ ratings), where the reference is the highest rating group (above AA+). Industry fixed effects are based on two-digit SIC values. All other variables are defined in Appendix 1. Errors are clustered at issuer level and robust to heteroskedasticity. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Control window Event window

[-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

SP dummy -0.185 -0.534*** -0.178** 0.005 -0.621*** -0.600***

Offering amount -0.000* 0.000 0.000 -0.000 0.000** 0.000

Coupon 0.001 -0.058* -0.027* -0.013* -0.078** -0.157***

Age 0.005 -0.001 0.000 0.001 -0.002 0.001

Profitability 0.738 -0.496 0.118 0.239** -0.682 -0.636

Tangibility -0.192 0.130 -0.166 -0.021 0.574 0.275

Leverage 0.575 0.121 -0.062 -0.127 0.233 -0.737

Tobin's Q -0.223* 0.202* 0.038 0.052** 0.185 0.222

Interest coverage -0.126 -0.249 -0.219 -0.244** 0.038 -0.677

Size 0.084 -0.051 -0.031 -0.025** -0.022 -0.015

A+/A/A- 0.066 -0.026 -0.013 0.010 -0.030 -0.040

BBB+/BBB/BBB- -0.086 -0.184 -0.066 -0.003 -0.138 -0.307

BB+/BB/BB- -0.355 -0.570** -0.286** -0.090 -0.541* -0.969**

< BB 0.072 -0.685*** -0.350*** -0.199*** -0.202 -0.628

Industry FE Y*** Y*** Y*** Y*** Y*** Y***

N 386 382 379 380 381 384

adj. R-sq 0.194 0.306 0.130 0.204 0.257 0.183

AIC 931.730 818.155 228.773 -316.950 876.085 1142.617

112

In this regression, (corresponding to the variable SP dummy in the table) takes the value of 1 for all the bonds that have an investment-grade issue/issuer rating from S&P, but no Moody’s rating.38 Since the S&P rating was not taken into account before June 17, these bonds were included in the index solely due to the change in the index rating rule. Since this is entirely exogenous and cannot be manipulated short-term by the firm, a decrease in yields for the affected bonds would measure, ceteris paribus, the “addition effect.”

The dependent variable for each column is the cumulative yield change for the respective period. Therefore, the first column is the pre-event window and the following columns are estimated over increasingly longer time periods centered around the announcement date, 0.

The coefficient of the SP dummy variable, , is insignificant in the control period, but statistically and economically significant in the period including the announcement day [-10,0], as well as in the following periods approaching the effective date (which would be day 106). There is an immediate decrease of 0.53% following the announcement day, and this decrease persists for up to 100 days. This result is comparable in magnitude with other events implying “good news” for the bond market, such as a decrease of 0.39% following the NRSRO designation of DBRS (Kisgen and Strahan, 2010), or a difference of up to 0.60% following Moody’s rating scale refinement, when the new rating was better than expected (Kliger and Sarig, 2000). Chen et al. (2014) has similar results over the long term, with a yield decrease of 0.73% after 240 days.

Most control variables are insignificant, except for Tobin’s Q, bond’s coupon, speculative ratings and industry fixed effects. However, since the dependent variable is the change in yields, there is no reason to expect a priori that these variables would have a significant explanatory power.

38 As mentioned in the variable description, issuer rating is taken into account only if there is no

issue rating. An example for a bond included in this group would be one that has no issue ratings from either Moody’s or S&P, and its direct issuer (not ultimate parent) has only an investment- grade S&P rating (and thus no Moody’s issuer rating).

113

3.5.2. The “deletion” effect

A second group affected by Lehman Brothers’ announcement consists of the bonds with investment-grade ratings and an offering amount between USD 150 million and 200 million, which would therefore be excluded from the IG index following the new minimum size requirement of USD 200 million. Table 3 displays the regression output for equation (1), where is the indicator variable for the Size group.

Table 3. Deletion effect The table reports the full sample cross-section regression output for cumulative yield changes (measured as the cumulative yields for T+1 and T, respectively, for the time windows defined in the parentheses) as a function of bond- and issuer-specific variables. Here, the variable of interest is "Size dummy,” which takes the value 1 for the bonds which have been excluded in the index as a result of the change in bond size requirements. The rating variables (A, BBB, BB, and below BB) are indicator variables based on the S&P overall rating (issue or issuer if issue is not available) grouped by wider rating groups (i.e. BBB group includes BBB-, BBB and BBB+ ratings), where the reference is the highest rating group (above AA+). Industry fixed effects are based on two-digit SIC values. All other variables are defined in Appendix 1. Errors are clustered at issuer level and robust to heteroskedasticity. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Control window Event window

[-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

Size dummy -0.037 0.206* 0.039 0.009 0.197* 0.302**

Offering amount -0.000 0.000* 0.000 -0.000 0.000*** 0.000

Coupon 0.013 -0.034 -0.018 -0.013 -0.049 -0.134**

Age 0.006*** 0.003** 0.001 0.001 0.002 0.005**

Profitability 0.809 -0.249 0.189 0.238** -0.400 -0.355

Tangibility -0.227 0.104 -0.178 -0.018 0.532 0.263

Leverage 0.605 0.156 -0.051 -0.129 0.281 -0.703

Tobin's Q -0.244* 0.153 0.023 0.053** 0.125 0.166

Interest coverage -0.086 -0.085 -0.171 -0.244* 0.222 -0.467

Size 0.083 -0.040 -0.028 -0.025 -0.011 0.000

A+/A/A- 0.039 -0.054 -0.025 0.012 -0.071 -0.065

BBB+/BBB/BBB- -0.114 -0.223 -0.081 -0.001 -0.189 -0.345

BB+/BB/BB- -0.373 -0.406 -0.248*** -0.084 -0.382 -0.739

< BB 0.053 -0.558** -0.321** -0.194* -0.080 -0.447

Industry FE Y*** Y*** Y*** Y*** Y*** Y***

N 386 382 379 380 381 384

adj. R-sq 0.192 0.288 0.116 0.204 0.231 0.175

AIC 932.939 828.065 235.060 -317.091 889.058 1146.521

114

The coefficient for the Size dummy variable becomes positive, and statistically and economically significant for most time periods within the event window. For example, the second column suggests that, all else equal, bonds affected by the size change experience an additional yield increase of 20 BPS in the period [-10, 0] relative to the rest of the bonds. This increase is persistent when taking into account different event windows, increasing to 30 BPS for the period [-10,100].

While economically significant, the coefficient is only half of the value of the addition effect. This asymmetry suggests that investor awareness is relatively more important than investment-related factors while, at the same time, both types of factors seem to influence the market response. Further, I test whether this result holds for the other “negative” change in my sample, resulting in a number of bonds migrating from the IG to the HY bond index.

3.5.3. The “transition” effect

Finally, there is a smaller group of bonds that were also affected negatively by the change in index rules, the HY group. These bonds have an investment-grade Moody’s rating and a speculative one from S&P. Given that, following the change, S&P ratings are also taken into account (the “index” rating would be the more conservative between S&P and Moody’s), these bonds have an overall speculative rating and were thus moved from the IG to the HY index.

Since the HY indicator variable has, by construction, a value of 1 for all bonds with an S&P speculative rating, I cannot include the indicator variables for ratings used in the other specifications due to nearly perfect multicollinearity. However, omitting these variables might raise the concern that the change in yields is driven by the ratings per se, and not by the change in index rating rules.

To address this potential omitted variable bias, I select a matched sample of otherwise similar bonds, and estimate the regression on the subsample using only the matched and treated bonds. Fortunately, the available data allows me to select a perfectly matched control group: the bonds that had the exact same split-ratings, but with Moody’s being more conservative, i.e. an IG S&P rating and a HY rating from S&P. Since the new index rating takes into account the most conservative one, these bonds were not affected by the change, making them a perfect match. Results are displayed in Table 4.

115

Table 4. Transition effect The table reports the full sample cross-section regression output for cumulative yield changes (measured as the cumulative yields for T+1 and T, respectively, for the time windows defined in the parentheses) as a function of bond- and issuer-specific variables. The variable of interest is "HY dummy,” which takes the value 1 for the bonds with an investment-grade rating from Moody's and a speculative rating from S&P, which migrated to the High-Yield Bond Index following the change in index inclusion rules. Panel A includes the full sample, while Panel B displays the output for a sample of rating-matched bonds (with investment-grade S&P rating and speculative grade Moody's rating). Industry fixed effects are based on two-digit SIC values. All other variables are defined in Appendix 1. Errors are clustered at issuer level and robust to heteroskedasticity. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Control window Event window

Panel A. Full sample [-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

HY dummy -0.157 0.280** 0.205** 0.111*** 0.232 0.305

Offering amount -0.000 0.000 0.000 0.000 0.000 -0.000

Coupon -0.568 -0.149*** -0.014 -0.013 -0.250* -0.409**

Age 0.015* 0.005*** 0.002* 0.001* 0.005* 0.010***

Profitability 2.442* 1.033 -0.712 -0.145 2.415 3.725

Tangibility 2.897 -0.306 0.390 0.311 -0.743 -1.492

Leverage 8.854 0.913 -0.274 -0.303 2.509* 1.885

Tobin's Q -0.177 0.026 0.324 0.189* -0.385 -0.491

Interest coverage -0.076 -0.366 -0.276 -0.354** 0.070 -1.030

Size 0.828 0.097 -0.013 -0.015 0.198 0.298*

Industry FE Y*** Y*** Y*** Y*** Y*** Y***

N 391 386 383 384 385 388

adj. R-sq 0.030 0.266 -0.007 0.039 0.187 0.171

AIC 2202.297 930.502 606.797 16.217 1223.385 1483.486

Control window Event window

Panel B. Matched sample [-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

HY dummy 0.048 0.284* 0.279* 0.222*** 0.428** 0.300

Offering amount -0.001 -0.000 0.000 -0.000 0.000 -0.000

Coupon -0.063 -0.102 -0.111 -0.006 -0.106 -0.078

Age 0.001 0.008 0.007 -0.001 0.010 0.009

Profitability 2.721 -0.231 0.625 -0.298** -0.612 -0.412

Tangibility -0.864 -0.837 -1.033 0.011 -1.314 -0.784

Leverage 2.189 1.990 1.310 0.265 1.907 2.478

Tobin's Q -0.291 0.578 0.424 0.040 0.784* 0.197

Interest coverage -1.570 0.505 -0.646 0.250 1.223 0.552

Size 0.146 -0.080 -0.048 0.009 -0.132 0.017

Industry FE Y*** Y*** Y*** Y*** Y*** Y***

N 67 66 66 66 66 66

adj. R-sq 0.174 0.605 -0.391 0.280 0.596 0.549

AIC 107.424 140.202 128.283 -171.147 149.800 156.273

Panel A is estimated using the full sample (for comparison purposes), and Panel B is estimated using the matched control bonds. The coefficients of the HY dummy in Panels A and B are strikingly similar (though the one in Panel B is only significant at the 0.1 level, possibly due to the small sample size), suggesting that the change in yields following Lehman Brothers’ announcement

116

is not driven by the ratings. Since credit ratings supposedly take into account a range of unobservable/qualitative variables, the coefficient in Panel B can be regarded as closer to the “true” magnitude of the market reaction.

This result is also similar to the one in Section 3.5.2. (0.28 vs. 0.20), suggesting that this asymmetry is consistent for the two types of negative events. Since I could identify a well-matched group for this particular change, this part of my analysis also serves a robustness test against other confounding effects. Furthermore, my findings are very similar (in absolute terms) with Chen et al. (2014), which look at the bonds going through the reverse transition, from HY to IG index.39

3.6. Good news versus bad news: market response (a)symmetry

As mentioned in the introduction, one of the main goals of this empirical exercise was to investigate whether there is a symmetric effect for positive versus negative changes (in my case, comparing the index addition effect with the deletion/transition to HY).

In their corresponding regressions, each of the indicators are statistically and economically significant, whereas the addition effect is twice as large as those of deletion or transition. To formally test whether the coefficients are statistically different from each other, I include the SP and the Size indicator variables in the same regression and perform a Wald test for coefficient equality.40 Table 5 displays the coefficients and the test statistics.

39 The average index effect in Chen et al. (2014) across all event windows is approximately 0.4.

40 I choose the indicator variable for the Size and not for the HY group, since the coefficients are similar and I use the main regression specification, which includes the rating variable.

117

Table 5. Good news vs. bad news The table displays the coefficients from the regression specification in (1), where Gi is replaced by the indicator variables for the SP group and the Size group for each event window. *,** and *** denote the statistical significance at the 10%, 5% and 1% levels within the regression. The last row represents the p-values of the Wald test for equal coefficients [SP dummy=Size dummy].

Control window Event window

[-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

SP dummy -0.19 -0.50* -0.17 0.00 -0.59* -0.54*

SIZE dummy -0.05 0.15* 0.02 0.01 0.13 0.24*

Wald test (prob>F) 0.32 0.04 0.25 0.93 0.03 0.02

As seen in the last row, the coefficients are statistically different from each other at the 0.05 level for most of the event windows. This suggests that investor awareness is relatively more important than other investment-related factors, as discussed in Elliott et al. (2006). While this argument might seem counterintuitive given that bond indices are not as popular as their stock counterparts, one needs to bear in mind that Lehman Brothers’ is one of the largest and most widely followed bond indices, and a change in bond index composition happens very rarely, unlike changes or recalibrations of stock indices. One interesting topic for further research could therefore be exploring how investor awareness or attention affect bond markets – especially given that most investors are institutional, informed investors.

While my results are not directly comparable with stock index effects, it is possible to gauge their economic significance by looking at the bond yield changes following an event triggered by real changes in bond riskiness, such as a rating change. For instance, the transition and deletion effects are comparable with market responses triggered by real changes in issuer creditworthiness, being equal to roughly half of the yield spread between BBB- and BB+ bonds (Chen et al., 2014).

3.7. Matching returns

While using fixed effects addresses the potential risk of omitted variables in the cross-section, there is still a possibility that my results are driven by another unidentified event which happened simultaneously with the announcement, leading to confounding effects. While another event is unlikely to affect the three groups of bonds precisely in the same direction as the announcement itself, I perform an alternative estimation to verify that this is not the case.

118

As in Bessembinder et al. (2009) and Chen et al. (2014), I construct a matched sample of bonds and investigate whether the post-event returns differ significantly from the treated groups. I follow the standard in the bond literature (Bessembinder et al., 2009) by using ratings and maturity as matching criteria.41 Since I have a large sample of control bonds, I use 1:1 exact matching by using the nearest neighbor algorithm for each of the treated groups.

Figure 2 displays graphically the average cumulative returns over the entire period for each of the three groups. The dotted lines represent the matched group, and the vertical lines the announcement and effective dates.

The graphs confirm my main results, as well as the accuracy of the matching procedure: the matched and control groups’ returns are very similar in the period before the announcement date, but start to diverge afterwards, and this divergence seems to persist over time.

To further test whether the differences between returns are statistically significant, I compute the differences for the same windows as in the main specification. Table 6 shows the values of these differences for all windows, as well as the p-values for the unequal variance two-sample t-test for equal means.42

The output confirms that all differences are statistically insignificant for the control window [-50;-10], but very significant for all event windows. In terms of economic significance, the results are similar to Chen et al. (2014). For instance, for the event window [-10,0] they find a difference of approximately 1.10 between the matched and control groups, while I have a corresponding average of 1.28 for the SP group, which was the one affected positively by the change.43

41 Since industry, size, and age are significant in some of my regressions, I construct alternative

matched samples by adding these variables, without materially affecting the outcome. The abnormal returns for these matches are reported in Appendix 6.

42 Bessembinder et al. (2009) show that, for sufficiently large samples, t-test performs equally well as non-parametric tests.

43 I compare this group to Chen et al. (2014), since their event also involved “good” news. The resulting value 1.28 is thus (0.61-(-0.66).

119

Figure 2. Matched cumulative returns The figures display the average cumulative returns (CR) over time for each of the three treated groups and their respective matches, represented by the dotted lines. Matching is done using nearest-neighbor 1:1 exact matching on maturity and ratings. The announcement day is marked with 0 (June 17, 2003). The effective date is day 106 (October 1, 2003).

120

Table 6. Matched CRs The table compares the average cumulative returns for the treated groups and their respective matches. As in Figure 2, the matching is 1:1 based on rating and maturity. The third row for each group displays the p-value for the unequal variance two-sample t-test for equal means.

Control window Event window

[-50,-10) [-10,0] [-10,+10] [-10,+30] [-10,+50] [-10,+100]

HY group 1.309 1.296 -0.963 -1.876 -4.010 -3.361

matched group 1.550 0.474 0.984 0.039 -0.036 -0.430

p-value 0.200 0.000 0.000 0.000 0.000 0.000

SP group 0.791 0.617 2.744 3.658 3.687 2.828

matched group 0.993 -0.668 -0.703 -0.793 -1.431 0.696

p-value 0.170 0.000 0.000 0.000 0.000 0.000

Size group 1.690 -0.193 -0.375 -2.342 -4.007 -3.450

matched group 1.976 0.454 1.328 1.296 0.363 0.955

p-value 0.190 0.000 0.000 0.000 0.000 0.000

3.8. Conclusion

By exploiting an exogenous change in Lehman Brothers’ bond index inclusion criteria which affected a large number of bonds both positively and negatively, the current paper provides new evidence on how indexing can affect bond yields/returns. More specifically, by using a new dataset from S&P’s Capital IQ platform including detailed bond- and issuer-specific data, I can compare the bond market response for three different types of changes following the change in size and index rating criteria announced by Lehman Brothers on June 17, 2003.

For the bonds with an investment S&P rating and with an outstanding amount of at least USD 200 million, the June 17th announcement meant that (starting on October 1st) they will be part of the Lehman Brothers’ IG bond index. I find that there is a significant positive market response for this group of bonds (S&P group). More specifically, the “addition” effect leads to a decrease in the cumulative yield of 0.53% in the period immediately following the announcement, and the decrease persists up to 100 days after the announcement.

121

A smaller group of bonds (the “Size group”), which had an outstanding amount between USD 150 million and 200 million, were excluded following the new liquidity requirement (of having a minimum of USD 200 million). For these bonds, I also find a significant, but negative, market reaction: an increase in yield of 0.20% following the announcement date.

The third change implies the transition from the IG index to the HY index for the bonds that had an S&P speculative rating (since the new rule considered the most conservative rating between Moody’s and S&P). As expected, these bonds (the “HY group”) also experienced a negative market reaction, with a 28% yield increase following the announcement.

My findings are robust to using a range of bond- and issuer- specific control variables and several matching procedures. Furthermore, the yield changes are economically significant, being comparable to the market reactions following “informative” announcements such as Moody’s rating scale refinement (which meant a reduction in information asymmetry) or to the yield change following an actual rating change (Chen et al., 2014).

Furthermore, since I have three types of changes, I can compare their relative magnitude. Since the positive change (i.e. “addition”) triggers a market reaction that is twice as large compared to the other negative changes (deletion/transition to HY), this paper suggests that investor awareness may be the primary factor underlying the addition effect, although other channels, such as benchmarking, changes in expectations, liquidity or monitoring, may also contribute to the overall index effect.

While my findings cannot single out a particular channel driving the index effect, they emphasize the importance of exploring further bond market frictions, given that these continuously growing markets function differently from stock markets, and that index investing is an increasingly popular investment strategy.

122

References

Barber, B. M., & Odean, T. (2007). All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. The Review of Financial Studies, 21(2), 785-818.

Basak, S., & Pavlova, A. (2013). Asset prices and institutional investors. American Economic Review, 103(5), 1728-58.

Barclays (2015). US Corporate Bond Index Factsheet. Barclays Bank Research.

Bessembinder, H., Kahle, K. M., Maxwell, W. F., & Xu, D. (2008). Measuring abnormal bond performance. The Review of Financial Studies, 22(10), 4219-4258.

Chang, Y. C., Hong, H., & Liskovich, I. (2014). Regression discontinuity and the price effects of stock market indexing. The Review of Financial Studies, 28(1), 212-246.

Chen, H., Noronha, G., & Singal, V. (2004). The price response to S&P 500 index additions and deletions: Evidence of asymmetry and a new explanation. The Journal of Finance, 59(4), 1901-1930.

Chen, Z., Lookman, A. A., Schürhoff, N., & Seppi, D. J. (2014). Rating-based investment practices and bond market segmentation. The Review of Asset Pricing Studies, 4(2), 162-205.

Cornaggia, J., Cornaggia, K. J., & Israelsen, R. D. (2017). Credit ratings and the cost of municipal financing. The Review of Financial Studies, 31(6), 2038-2079.

Denis, D. K., McConnell, J. J., Ovtchinnikov, A. V., & Yu, Y. (2003). S&P 500 index additions and earnings expectations. The Journal of Finance, 58(5), 1821-1840.

Elliott, W. B., Van Ness, B. F., Walker, M. D., & Warr, R. S. (2006). What drives the S&P 500 inclusion effect? An analytical survey. Financial Management, 35(4), 31-48.

Goldman Sachs, (2016). Portfolio Strategy Research. Goldman Sachs Investment Management.

Hand, J. R., Holthausen, R. W. & Leftwhich, R. W. (1992). The effect of bond rating agency announcements on bond and stock prices. The Journal of Finance, 47(2), 733-752.

Kisgen, D. J. (2006). Credit ratings and capital structure. The Journal of Finance, 61(3), 1035-1072.

Kisgen, D. J., & Strahan, P. E. (2010). Do regulations based on credit ratings affect a firm’s cost of capital? The Review of Financial Studies, 23(12), 4324-4347.

Kliger, D., & Sarig, O. (2000). The information value of bond ratings. The Journal of Finance, 55(6), 2879-2902.

Moody’s (2008). Updated Summary Guidance for Notching Bonds, Preferred Stocks and Hybrid Securities of Corporate Issuers. Moody’s Investor Service.

123

Petajisto, A. (2011). The index premium and its hidden cost for index funds. Journal of Empirical Finance, 18(2), 271-288.

Securities Industry and Financial Markets Association, (2018). SIFMA Fact Book. New York.

Shleifer, A. (1986). Do demand curves for stocks slope down? The Journal of Finance, 41(3), 579-590.

PR Newswire (2003). Lehman Brothers to make several changes to its global family of indices.

Vanderbilt Avenue Asset Management (2003). Adjustments to Lehman Brothers Global Family of Indices.

Wurgler, J., & Zhuravskaya, E. (2002). Does arbitrage flatten demand curves for stocks? The Journal of Business, 75(4), 583-608.

124

Appendices

Appendix 1. Variable definitions

Bond characteristics

Offering amount (MUSD) offering amount at the date of issuance (in million US dollars)

Maturity (years) difference between maturity date and offering date

Age (years) difference between maturity and the current year (2003); set to 0 if negative

Coupon (%) coupon at offering; since all bonds are fixed-coupon the rate is not time-varying

Yield (%)

The Yield to Maturity is the discount rate at which the present value of future payments (investment income and return of principal) equals the price of the security; yield data is readily available on a daily basis on Capital IQ platform for the more liquid bonds

Return(%) the bond return is derived from the price changes; price data is collected on a daily basis from S&P's Capital IQ platform

Issuer characteristics

Size (log) log (total assets)

Profitability (%) operating income/total assets

Leverage (%) total debt/total assets

Tangibility (%) plant, property and equipment/ total assets

interest coverage(%) ebitda/ interest expense

Tobin's q (book assets minus book common equity minus deferred taxes plus market equity) / book assets (as in e g Hadlock and Pierce,2010)

SIC2 two-digit SIC code, which is used for the fixed-effects specification

Ratings and group identifiers

SP_rating/ M_rating

Ordinal variable ranging from 1 (highest rating) to 20 (lowest rating) for either SP's or Moody's ratings. The type of rating taken into account is long-term, local currency issue-specific rating; for the case when this rating is not available, the corresponding rating for the issuer (direct issuer and not parent) is taken into account in order to maximize the number of observations.

SP/Moody's group rating Broader rating classification ranging from 1 to 5. The correspondence between the coarse rating variable and the letter ratings' scale is described in Appendix 2.

HY dummy indicator variable taking the value of 1 if SP_rating is below BB+ (speculative grade) AND M_rating is above BBB- (investment-grade).

SIZE dummy indicator variable taking the value of 1 if offering amount is below USD 200 million AND M_rating is above BBB-

SP dummy indicator variable taking the value of 1 if SP_rating above BBB- AND M_rating is not available (i.e. a bond is only rated by S&P)

125

Appendix 2. Rating scales and frequencies The table displays the rating scales for S&P and Moody's, with the corresponding number values used in the analysis. The frequencies are calculated relative to the sample size with non-missing yield values, which have a size of 449 observations for S&P ratings and 448 for Moody's. The rating variables used here are the same as the ones used in regression, i.e. issue OR issuer rating (if issue rating is not available). The number of observations used in the regressions is lower and varies depending on the availability of the variables used. The dotted line marks the threshold between investment and speculative grade.

Rating variable (coarse) Rating group variable S&P Frequency(%) Moody's Frequency(%)

1 1 AAA 3.34 AAA 3.57

2 1 AA+ 1.56 Aa1 2.90

3 1 AA 2.45 Aa2 2.23

4 1 AA- 5.57 Aa3 6.47

5 2 A+ 12.92 A1 5.36

6 2 A 6.68 A2 11.16

7 2 A- 6.90 A3 8.71

8 3 BBB+ 10.47 Baa1 10.04

9 3 BBB 11.49 Baa2 13.51

10 3 BBB- 9.58 Baa3 11.27

11 4 BB+ 6.56 Ba1 3.12

12 4 BB 5.00 Ba2 2.90

13 4 BB- 4.68 Ba3 3.79

14 5 B+ 5.12 B1 4.03

15 5 B 4.01 B2 2.68

16 5 B- 1.34 B3 2.68

17 5 CCC+ 0.89 Caa1 2.23

18 5 CCC 1.00 Caa2 1.12

19 5 CCC- 0.45 Caa3 0.67

20 5 CC 0.00 Ca 1.56

126

Appendix 3. Group composition The diagram explains how bonds were assigned to the SP, SIZE, HY and control groups (CTRL respectively CTRL* which is a more narrow control for the HY group, with the same split rating), based on their size (amount outstanding) and rating. INV is the abbreviation for investment-grade rating, SPEC means speculative rating; NR stands for not rated, while NI means not included in the sample. The first level (size) classifies the sample by the amount outstanding in million US dollars (MUSD). Bonds with an amount outstanding lower than MUSD 150 could not be part of the index at any point in time and thus were not included in the sample. The second level represents Moody’s rating, while the third level refers to S&P rating. The last column displays the final group distribution.

>150

<200

INV

SPEC HY

INV SIZE

NA SIZE

SPEC

SPEC NI

INV CTRL*

NA NI

NA

SPEC NI

INV NI

NA NI

>200

INV

SPEC HY

INV CTRL

NA CTRL

SPEC

SPEC NI

INV CTRL*

NA NI

NA

SPEC NI

INV SP

NA NI

Size(MUSD) Moody’s rating S&P’s rating Group

127

Appendix 4. Index effects – robustness to additional control variables The table reports an alternative specification for the full sample cross-section regression output for cumulative yield changes for the estimation window [-50,-10] and for the immediate event window [-10,0] by including additional variables to the controls described in Section 3.3. The first variable, “Group dummy,” represents the indicator variable for each of the three groups (“SP” for the addition regressions, “Size” for the deletion model, and “HY” for the transition effect). In addition to the existing control variables (offering amount, coupon, profitability, tangibility, Q, interest coverage, size and S&P broad rating groups), I add cash, intangibles, capital expenditures and dividends (all relative to total assets) in order to control for investment opportunities and financial constraints. I also control for Moody’s rating by including two dummy variables for (1) when Moody’s rating is higher than S&P; (2) when the two ratings are equal. The third case (3) when Moody’s rating is smaller than S&P is here the reference category. Lastly, as an alternative to bond age I use maturity, and as an alternative to total leverage I use long-term leverage only. Industry fixed effects are based on two-digit SIC values. Errors are clustered at issuer level and robust to heteroskedasticity. *,** and *** denote statistical significance at the 10%, 5% and 1% levels.

Addition (SP) Deletion (SIZE) Transition (HY)

[-50,-10) [-10,0] [-50,-10) [-10,0] [-50,-10) [-10,0]

Group dummy -0.160 -0.450*** -0.037 0.206* -0.157 0.280**

Offering amount -0.000* 0.000 -0.000* 0.000 -0.000* 0.000

Maturity 0.002 -0.003 0.006*** 0.003** 0.015* 0.005***

Coupon 0.020 -0.087* 0.028 -0.058 0.056* -0.060*

Profitability 0.607 -0.530 0.736 -0.305 2.564* 1.714

Long-term leverage 1.065* 0.244 1.077* 0.016 1.254** 0.259

Tobin's Q -0.089 0.020 -0.053 -0.051 -0.011 0.088

Interest coverage -0.004 0.003 0.002 0.006 -0.001 0.004

Size 0.084 -0.051 0.083 -0.040 0.828 0.097

CAPEX -1.825 -1.381 -1.460 -2.618 2.045 -2.988

Cash 1.217 0.796 0.986 -0.777 1.250 -0.878

Intangibles 0.668* 0.909* 0.692* 0.695 0.292 0.620

Dividends 0.751 -0.369 -0.891 -0.663 2.774 -0.839

A+/A/A- -0.077 0.032 -0.083 0.031

BBB+/BBB/BBB- -0.297 -0.063 -0.307 -0.060

BB+/BB/BB- 0.146 -0.286 0.207 0.058

< BB- 0.353 -0.450 0.415 -0.183

Moody’s>S&P -0.265** -0.061 -0.231* 0.091 -0.240* 0.092

Moody’s=S&P -0.199 -0.128 -0.199 -0.127 -0.060 -0.108

Industry FE Y*** Y*** Y*** Y*** Y*** Y***

N 269 267 269 267 269 267

adj. R-sq 0.484 0.234 0.477 0.170 0.446 0.191

AIC 423.602 470.173 427.981 492.318 440.353 482.303

128

Appendix 6. Matched CRs – extended matching The table compares the outcome of different matching procedures with the treated group CARs. The first two rows for each group are as in Table 6. “Extended match” is based on nearest-neighbor matching on the following variables: age, broad rating group, two-digit SIC and size. The number of matches varies for each group. The third row for each group displays the p-value for the unequal variance two-sample t-test for equal means between the treated and the extended match group.

[-50,-10) [-10,0]

HY group 1.309 1.296

1:1 matched group 1.550 0.474

extended match 1.415 0.380

p-value 0.520 0.000

SP group 0.791 0.617

1:1 matched group 0.993 -0.668

extended match 0.680 -0.430

p-value 0.370 0.000

Size group 1.690 -0.193

1:1 matched group 1.976 0.454

extended match 1.814 0.580

p-value 0.370 0.000

129

Chapter 4. Credit rating agencies’ conservatism revisited: the sovereign market

4.1. Introduction

Sovereign credit ratings have been the target of much criticism in recent years. After the Asian crisis, Credit Rating Agencies (hereafter referred to as CRAs) were criticized for failing to predict it (Mora, 2006). During the Eurozone sovereign debt crisis, however, CRAs were instead criticized for being too aggressive. For example, the President of the European Commission mentioned that “ratings appear to be too cyclical, too reliant on the general market mood rather than on fundamentals – regardless of whether market mood is too optimistic or pessimistic” (Barroso, 2010).

While, to the best of my knowledge, there is no research focusing on sovereign rating standards, empirical evidence on US corporate ratings seems to support Barroso’s concern. Two relatively recent papers, Alp (2013) and Baghai et al. (2014), find that corporate credit ratings have become increasingly conservative over time. They also show that the changes in ratings seem to be unrelated to the variation in firm-specific factors. If sovereign credit ratings exhibit a similar pattern, this would not only undermine the usefulness of ratings, but may also distort capital markets through the impact on corporations via the “ceiling” rule44 and on financial institutions via the rating-based capital requirements.45

44 This refers to the fact that the large majority of firms do not receive a rating above their

sovereign.

45 Starting with Basel II, credit ratings can be used for risk-weighting assets (BCBS, 2004).

130

In this paper, I investigate whether sovereign rating standards have changed over the last two decades. More specifically, I ask whether a sovereign with the same riskiness would receive the same rating in 2013 as in 1993, ceteris paribus. A lower rating in 2013 would imply that CRAs have become more conservative over time, while a higher rating would imply more lenient rating standards.

I find that there is an increase in conservatism of up to 2.2 notches from 1993 to 2013. This implies that a sovereign that received an AA rating in 1993 would only get A+ twenty years later, in spite of maintaining the same fundamentals. This finding is robust to including additional factors and using alternative estimation methods and different subsamples. Furthermore, my result is similar in terms of magnitude to both Baghai et al. (2014) and Alp (2013), who find an increase in conservatism of 2.9 and 1.6 notches, respectively, over earlier time periods of similar length.

Consistent with research on corporate rating standards, I also find that this conservatism is not fully justified. First, sovereign bond markets take this conservatism into account when pricing debt by “undoing” about 20% of its effect.46 Second, by using Bank of Canada’s detailed sovereign default dataset and S&P’s rating-specific predicted default frequencies, I show that default rates have actually been decreasing over time. Third, I find that other “global” factors such as credit scarcity, business cycles or global economic growth only explain up to 15.5% of the year indicators, i.e. my measure of conservatism. Lastly, I show that the downward trend in rating standards is consistent across various subsamples: CRAs are equally conservative with developed and developing economies, and the notorious investment-speculative threshold doesn’t affect rating standards. Therefore, factors shown previously to distort the credit rating process, such as information asymmetry or rating-based regulations, do not seem to influence sovereign rating standards.

The rest of the paper is organized as follows. Section 4.2. refers to related theoretical and empirical research. The empirical analysis starts with a description of the data and variables in Section 4.3, followed by the main results in Section 4.4 and several robustness tests in Section 4.5. The following part of the paper investigates potential explanations for my main finding – the increase in conservatism: Section 4.6 presents the results for different sample splits,

46 When modelling yield spreads as a function of bond ratings and a measure of conservatism, both factors are significant, and the measure of conservatism is of the opposite sign and about 20% of the ratings’ coefficient magnitude.

131

while in Section 4.7, I investigate whether the increased conservatism is justified. Section 4.8 concludes.

4.2. Rating standards and the sovereign market

Over the last two decades, there has been a steady decrease in the average sovereign rating,47 from A+ to BBB-. While this decline seems consistent with the increasing number of emerging markets being rated, the extent to which it fully reflects increasing default probabilities over time is still an open question.

Theoretical and empirical literature on corporate ratings provides a less intuitive explanation for this decline: a change in rating standards over time. This means that the same rating would imply a different creditworthiness today as opposed to, say, 20 years ago. How is this possible?

There is a wealth of research showing how the market structure and business model of CRAs’ can alter rating standards, both over time and across different markets or asset types. Theoretically, the issuer-pays model triggers an incentive for rating inflation (Bolton et al., 2012), while reputation concerns can motivate more conservative ratings (Morris, 2001). Regulation, on the other hand, is a double-edged sword: while investment restrictions based on rating levels can lead to rating inflation (Opp et al., 2013), an asymmetric liability for overly optimistic ratings can trigger more conservatism (Goel and Thakor, 2011).

Empirical literature is largely consistent with these theoretical arguments, providing mixed evidence on the variation in rating standards. This paper relates most to the narrow, but growing body of research focused on the time variation of (corporate) rating standards, started by Blume and McKinlay (1998). Using a sample of US investment-grade firms, they are the first to address the question of whether a firm with the same risk factors would receive different ratings in different time periods, and they find that this is indeed the case. More recently, Alp (2013) and Baghai et al. (2014) confirm this finding for larger samples of US firms. However, Alp (2013) shows that, surprisingly, ratings have become more lenient over time for speculative graded companies, while Baghai et al. (2014) find the opposite after including firm-fixed effects.

47 Sovereign credit ratings represent “an opinion on the government’s capacity and willingness to

meet its financial obligations as they come due” (Standard and Poor’s, 2008).

132

There is also empirical evidence for cross-sectional variation in rating standards, i.e., across different groups of issuers or debt instruments. One theoretical argument for this variation is provided in Skreta and Veldkamp (2009), who show that incentives for rating inflation increase with the complexity of the rated product. Further, in an extensive comparison of different asset classes, Cornaggia et al. (2014) show that the rating process and accuracy is significantly different for firms, structured finance products, municipalities and sovereigns, making any comparisons inconsistent. They also discuss how these differences can lead to variations in rating standards. Indeed, when taking into account different asset types, empirical evidence is mixed. While the overall findings for corporate ratings suggest more conservative rating standards relative to other asset classes, research on structured finance products (see e.g., He et al., 2012) concludes that ratings for these types of assets have been significantly inflated. So what about sovereign ratings?

Sovereigns are a relatively new market for the CRAs, and a market that is substantially different. First, unlike in national lending relationships, the default risk premium is determined by the borrower’s willingness, rather than ability, to pay (Dimitrijevic et al., 2011). Governments rarely default, since they can either modify their fiscal policies in order to facilitate debt repayment, or receive support and debt relief from higher-authority institutions (Moody’s, 2013). Second, CRAs’ revenue from sovereign ratings represents a small fraction of the total revenue (e.g., in 2013, sovereigns accounted for 10% of S&P’s revenue). Third, the sovereign rating process is less elaborate, given the small team of analysts and little interaction with the issuer relative to the corporate market (Gaillard, 2011). Fourth, CRAs rely mostly on publicly available data, and therefore there is little proprietary information benefit. Most importantly, the impact of a sovereign downgrade can be more severe, given that a sovereign’s economy, companies and financial institutions are affected. Gande and Parsley (2005) show that there are significant spillover effects to bond/CDS markets, while Ferreira et al. (2007) document the same effect for international stock markets. The question is, then, whether there is a change in rating standards for sovereigns?

To the best of my knowledge, the sovereign ratings literature has not addressed this question yet. Most of the research focuses on the determinants of sovereign ratings or on the impact of rating events on capital markets (e.g., inter alia, Cantor and Packer, 1996; Brooks et al., 2004; Hill et al., 2010; Ferreira et al., 2007). The only papers partially related to sovereign rating standards are Ferri

133

et al. (1999) and Mora (2006), who investigate CRAs’ rating behavior surrounding the Asian crisis. However, my paper extends their sample and time horizon, instead investigating a systematic change in rating standards, above and beyond a purely cyclical reaction.

4.3. Data

4.3.1. Sovereign Ratings

Following related literature, I use long-term foreign currency sovereign issuer ratings. I focus on S&P, since there is a relatively larger variation in their ratings, and they tend to lead other agencies in re-rating (Hill et al., 2010). Furthermore, since the rate of disagreement among CRAs is rather small (Brooks et al., 2004), it seems implausible that using Moody’s or Fitch ratings would significantly affect my results.

The rating data are collected on a yearly basis from S&P’s Capital IQ, for all the 120 countries rated for at least one year between 1993 and 2013. As in Hill et al. (2010), the rating scale is converted into numbers, with a maximum of 21 corresponding to AAA, and 1 corresponding to default rating notches (D/SD). Appendix 1 displays the full rating scale with its assigned values, as well as a broader classification (defined by merging 3-notch groups: e.g. AA+, AA and AA- are coded with 7). AAA rating dominates (15.5% of the country-year observations) and the second largest frequency (8.63%) corresponds to the lower speculative grade rating B+. Except for the near-default ratings (below B-), all other rating groups are well represented (with frequencies of 4-6%), reflecting a quite even distribution of various rating groups.

With 1935 country-year rating observations and a timespan of 20 years, the sample compares very favorably with related literature.48 In contrast with corporate ratings, however, the sample is small, since sovereign ratings are a relatively recent phenomenon. It wasn’t until the beginning of the 1990s that a significant number of less developed countries received a rating; it thus makes little sense to investigate any time trends prior to this decade.

48 This number represents all country-year observations with available rating data. For a more

detailed comparison on sample size, see Hill et al. (2010).

134

Figure 1 illustrates the time distribution of broad rating groups starting from the first years of sovereign rating activity. It is clearly visible how the rating scale distribution shifts over time towards the lower end: while there were no ratings below AAA at the beginning of the period, the B group (B-/B/B+) represents the largest fraction in 2013. The highest rating, AAA, remains quite stable over time, consistent with the fact that high-quality sovereigns have little variation in their ratings.

This time pattern has two important implications for the study. Firstly, since there is little variation in the beginning of the period, there will be an unavoidable trade-off between cross-sectional variation and longer time series. Since one priority of this empirical exercise is to investigate the time trend, I narrow down the time frame to the first year when there is sufficient variation in the levels of speculative ratings (i.e., a regression using only the speculatively graded sovereigns can be estimated), which is 1993.

Figure 1. Time distribution of broad rating groups

135

4.3.2. Independent variables

The choice of variables for the baseline specification relies to a large extent on the widely used model proposed by Cantor and Packer (1996):49 (1) log of GDP per capita; (2) real GDP growth; (3) log of consumer price index (inflation); (4) government debt divided by GDP; (5) government balance, measured as the difference between costs and revenues, divided by GDP; (6) external balance, measured as the difference between exports and imports divided by GDP and (7) default history measured as the inverse of the number of years from last default. The detailed description of all variables can be found in Appendix 2.

Macroeconomic indicators are lagged by one year. As in Cantor and Packer (1996) and Hill et al. (2010), I use three-year averages for GDP per capita, inflation, government and external balances. This is consistent with CRAs’ stated attempt to rule out cyclical influences by using averages over several years (Dimitrijevic et al., 2011). Macroeconomic data is collected primari ly from World Bank’s Development Indicators. I complement missing data with Oxford Economics, Moody’s public database, and the debt statistics used in Jaimovic and Panizza (2015). Rating data are collected from S&P’s Capital IQ and Fitch’s website.

Although the initial number of rated sovereigns in the sample is 120, the number drops to 111 countries for which I have at least three years of non-missing data for all variables. The resulting unbalanced panel consists of 1557 country-year observations for the main specification.

Table 1 summarizes the median values for the main explanatory variables across broad rating groups. The values are comparable with Cantor and Packer (1996) (e.g., the median of GDP growth for the AA group is 2.1 for my sample, whereas CP have a corresponding value of 2.47), in spite of the fact that they only use the cross-section of 49 rated countries for 1996.

49 For a detailed explanation of the expected impact and economic rationale for different variables,

see Cantor and Packer (1996).

136

Table 1. Median values across broad rating groups The table displays the median values of the main explanatory variables. Medians are computed using raw values (without any transformation, such as logarithm /average/lag)

GDP per Capita

GDP growth

Inflation Government debt

Fiscal Balance

External Balance

Default history

AAA 39.724 1.811 1.820 51.778 -0.950 2.940 0.000

AA 30.077 2.144 2.235 53.452 -1.705 1.305 0.000

A 14.640 3.277 3.535 35.753 -1.625 -1.455 0.000

BBB 5.885 3.754 5.090 40.845 -2.321 -2.485 0.000

BB 3.130 3.048 6.180 40.273 -2.460 -2.085 0.059

B 1.552 2.336 7.285 44.092 -3.240 -5.010 0.067

CCC-CC 2.829 -0.180 10.590 53.958 -1.360 -1.190 0.200

C/D/SD 4.572 2.023 7.930 92.663 -1.160 -0.365 1.000

Total 12.801 2.277 5.583 51.602 -1.853 -1.043 0.166

Most indicators have a non-linear distribution across the rating scale, as highlighted as well in Hill et al. (2010). While seemingly counterintuitive, this distribution makes economic sense: for instance, quite a few developed economies have a very high debt-to-GDP ratio (e.g. in 2013, Japan’s total debt was 243% of its GDP while maintaining a rating of AA-), while several emerging economies have relatively higher GDP growth rates compared to many of the developed countries. Hence, I will include quadratic terms for these variables as a robustness test.

The time patterns of the control variables, shown in Appendix 3, are not entirely consistent with the decrease in ratings, suggesting that increased “conservatism” might not be justified by a decrease in issuers’ quality. More specifically, over the entire period there is a significant decrease in government debt and inflation (13% respectively 25%), while default history is slightly decreasing by the end of the period. On the other hand, several factors which should be positively correlated with ratings, such as GDP per capita and external balance, are decreasing over time – which means that they could be at least partially driving the overall decrease in ratings. Therefore, from a purely univariate perspective, it seems unclear whether the downward time trend of ratings is driven by country-specific variables. In the following section, I will investigate various alternative explanations in a multivariate setting.

137

4.4. Empirical results

I analyze the time variation in rating standards by modelling ratings as a function of macroeconomic variables and year indicators (time fixed effects). The variables of interest are thus the year indicators which – if the model is stable over time – would capture the change in rating standards relative to the base year, 1993, after controlling for other factors.

The general form of the main specification is: = + + , + + (1)The dependent variable represents the sovereign rating coded from 21 (highest rating) to 1 (lowest rating), Xit,k is a vector of K control variables described in the previous section, Si are country fixed effects and Yt represent the year indicators (thus a are the coefficients of interest).

Since it is important to include cross-section fixed effects,50 the preferred estimation method is OLS. This is because, as discussed in e g Lancaster (2000), using within-estimator for ordered probit may produce biased estimates due to the incidental parameter problem. Furthermore, the coefficients’ magnitudes and predicted values would be more difficult to interpret using ordered choice estimators. However, for comparison purposes I also report OLS without fixed effects, as well as ordered probit estimation.

Table 2 presents the estimation output for the baseline specifications.

50 Country fixed effects address time-invariant unobservable variables, such as qualitative data

CRAs claim to use in their methodologies; Baghai et al. (2014), for instance, find that the conservatism patterns for speculative graded firms in Alp (2013) are reversed once including fixed effects.

138

Table 2. Main estimation output This table displays baseline results for the entire sample, from 1993 to 2013. Year indicator coefficients are relative to the base year 1993. The sample includes all sovereigns rated by S&P with non-missing macroeconomic data for at least three consecutive years. The dependent variable is S&P's long-term foreign currency sovereign rating converted into numerical identifiers (with a minimum of 1 corresponding to ratings below CCC and a maximum of 21 for AAA). The independent variables are defined in App. 2. Standard errors are corrected for heteroscedasticity, autocorrelation and are clustered at country level. Models (1) and (2) are estimated using OLS. (3) is estimated using ordered probit. Country fixed effects are used in model (1). *, **, and *** correspond to 10%, 5%, and 1% significance levels.

(1) (2) (3)

OLS OLS Ord.Probit

1994 0.251** 0.261** -0.014

1995 0.231* 0.292** -0.045

1996 -0.002 0.070 -0.255*

1997 -0.529 -0.410 -0.466***

1998 -0.978** -0.820** -0.591***

1999 -1.014*** -0.810** -0.597***

2000 -1.103*** -0.908*** -0.865***

2001 -1.396*** -1.151*** -0.980***

2002 -1.315*** -1.046*** -1.006***

2003 -1.181*** -0.897*** -1.013***

2004 -1.034*** -0.697** -0.805***

2005 -1.067*** -0.682** -0.808***

2006 -1.333*** -0.893*** -0.845***

2007 -1.549*** -1.054*** -0.933***

2008 -2.054*** -1.489*** -1.079***

2009 -2.196*** -1.609*** -1.235***

2010 -1.604*** -1.039*** -1.013***

2011 -1.726*** -1.120*** -1.151***

2012 -1.975*** -1.312*** -1.155***

2013 -2.207*** -1.543*** -1.329***

GDP per capita 4.437*** 2.881*** 0.943***

Real GDP growth 0.070*** 0.087*** 0.065***

Inflation -0.187* -0.230** -0.464***

Government debt -0.030*** -0.028*** -0.005

Fiscal balance 0.072*** 0.070*** 0.014

External balance 0.066*** 0.062*** 0.039***

Default history -3.032*** -3.434*** -3.649***

Country fixed effects Yes No No

N 1557 1557 1557

RMSE 1.064 1.145 -

adj./pseudo R-sq 0.955 0.757 0.261

Model (1) is the base-case specification, which will be employed in later stages of the analysis. The year coefficients express the change relative to 1993 and are all significant except for the years 1996 and 1997. Almost all years have negative coefficients, consistent with more conservative ratings. For instance, in 2013, a sovereign would receive 2.2 notches less (i.e. from AA- to BBB+)

139

relative to 1993, ceteris paribus. To put this into perspective, the average rating has decreased by 4.15 notches from 1993 to 2013 (Appendix 4). Thus, the results in (1) suggest that as much as 53% of this drop is the result of increased conservatism. This result is similar in terms of sign and magnitude to the findings in Baghai et al. (2014), who find a decrease of up to 2.9 notches for US firms over a period of 24 years, using fixed effects OLS estimation. In terms of control variables, most coefficients are highly significant and with magnitudes comparable with prior literature. For instance, GDP per capita has, as expected, an economically and statistically significant positive effect, being by far the largest coefficient, both in absolute and standardized terms.51 This is consistent with CRAs’ methodologies, where GDP per capita is the single most important ratio when assessing sovereigns (e.g. Dimitrijevic et al., 2011). Furthermore, in Cantor and Packer’s model, the coefficient for GDP is of comparable magnitude. In contrast to previous studies (e.g., Cantor and Packer, 1996; Mora, 2006; Hill et al., 2010), all coefficients have expected signs: GDP variables and balances are positively correlated with ratings, while inflation, debt and default history decrease creditworthiness.

Model (2) replicates the estimation in (1), but without country fixed effects. There is a noticeable difference in the levels of significance and magnitudes for both year indicators and control variables. For example, the coefficient for the last year, 2013, is only -1.5 (compared to 2.2), meaning that the model without fixed effects would underestimate the conservatism. Furthermore, if we were to interpret the fixed effects as proxies for the qualitative country-specific information CRAs use for their assessments,52 this information seems to be of a “negative” nature. This could be explained by an increased level of information asymmetry and corruption in developing countries, which have become more frequently rated over recent years. I will investigate this alternative explanation in Section 4.5.

Finally, model (3) is estimated using ordered probit. While the coefficients are not directly comparable to OLS estimates, they are consistent with the results found in other ordered probit estimations of conservatism for US firms: I find a decrease of 1.3 notches over my sample period of 20 years, while Blume and McKinlay (1998) document a drop of 0.7 notches from 1978 to 1995 for

51 The standardized coefficients are used in Figure 3.

52 CRAs stress in their methodologies the importance of qualitative information in their sovereign assessments as well (Standard and Poor’s, 2008).

140

investment-grade companies and Alp (2013) finds a decrease of 1.6 notches from 1986 to 2007.

4.5. Robustness tests

4.5.1. Additional variables

I re-estimate the base-case specification using a variety of additional controls, without materially affecting the economic and statistical significance of the year indicators. In this section, I describe some of these alternative specifications.53 The results are reported in Table 3 and represented graphically in Appendix 5; additional robustness tests are reported in Appendix 6. For comparison purposes, I include the base specification as model (1) in Table 3.

First, as discussed in the descriptive section, there might be a “newly rated” bias: CRAs might be relatively more conservative when assigning the first rating due to greater information asymmetry, especially in the case of emerging markets. On the other hand, since the largest part of the rating fee occurs when first rated (S&P, 2015), this might lead to “inflated” ratings. In model (2), I test this effect by restricting the sample to the countries rated prior to 1993, resulting in 731 country-year observations.54 The resulting year coefficients are lower in magnitude (e.g. the year coefficient is -1.9 for 2013, while the corresponding coefficient in the base-case specification is -2.2), but the downward trend of all year coefficients is very similar to (1). In model (3), I add a country-specific dummy variable taking the value of one for the first year with an available rating for that country. However, the inclusion of these variables doesn’t change the time fixed effect coefficients.

53 I perform a range of additional tests reported in Appendix 6, using alternative measures of

debt, newly rated issuers, and adding additional variables reported in Moody’s rating methodology (Moody’s, 2013).

54 The composition of this sample is also different, since countries with long rating histories have relatively higher ratings and less variation across time, as visible in Fig.1.

141

Table 3. Robustness tests: additional variables The table reports several robustness tests. All models are estimated using OLS, the control variables in Table 2 (not reported) and country fixed effects. Model (1) is the base-case specification from Table II. Model (2) reproduces model (1) in Table II for the subsample of sovereigns rated before 1993; Models (3), (4), (5), and (7) include additional variables described in Appendix II. Model (6) includes the square terms for GDP growth, debt, fiscal and external balances, which, for brevity, are not reported. Standard errors are corrected for heteroskedasticity and clustered at country level. *. **. and *** correspond to 10%. 5% and 1% significance levels.

(1) (2) (3) (4) (5) (6) (7)

1994 0.251** 0.341** 0.253** 0.247** 0.279** 0.257**

1995 0.231* 0.215 0.213* 0.227* 0.266** 0.238*

1996 -0.002 -0.025 -0.030 0.000 -0.012 0.030 0.007

1997 -0.529 -0.744* -0.534 -0.505* -0.542 -0.509 -0.520

1998 -0.978** -0.988** -0.981** -0.992*** -0.996** -0.968** -0.964**

1999 -1.014*** -0.936** -1.039*** -1.016*** -1.041*** -1.010*** -0.990***

2000 -1.103*** -0.965** -1.089*** -1.083*** -1.136*** -1.123*** -1.074***

2001 -1.396*** -1.361*** -1.407*** -1.383*** -1.407*** -1.399*** -1.366***

2002 -1.315*** -1.291*** -1.323*** -1.345*** -1.341*** -1.304*** -1.285***

2003 -1.181*** -1.021** -1.188*** -1.217*** -1.212*** -1.167*** -1.151***

2004 -1.034*** -0.743* -1.030*** -1.076*** -1.080*** -1.018*** -1.004**

2005 -1.067*** -0.927** -1.052*** -1.033*** -1.122*** -1.057*** -1.037**

2006 -1.333*** -1.317*** -1.316*** -1.303*** -1.398*** -1.298*** -1.302***

2007 -1.549*** -1.612*** -1.537*** -1.525*** -1.627*** -1.464*** -1.516***

2008 -2.054*** -1.947*** -2.049*** -2.058*** -2.144*** -1.920*** -2.019***

2009 -2.196*** -1.853*** -2.163*** -2.192*** -2.304*** -2.058*** -2.162***

2010 -1.604*** -1.104** -1.591*** -1.630*** -1.732*** -1.516*** -1.564***

2011 -1.726*** -1.389*** -1.702*** -1.749*** -1.871*** -1.648*** -1.685***

2012 -1.975*** -1.603*** -1.937*** -2.004*** -2.139*** -1.910*** -1.934***

2013 -2.207*** -1.922*** -2.163*** -2.197*** -2.377*** -2.161*** -2.166***

GDP per capita 4.437*** 4.101*** 4.430*** 4.457*** 4.488*** 4.384*** 4.455***

Real GDP growth 0.070*** 0.183*** 0.066** 0.052** 0.063** 0.097** 0.070***

Inflation -0.187* -0.060 -0.185* -0.115 -0.178* -0.198* -0.189*

Government debt -0.030*** -0.031*** -0.031*** -0.033*** -0.031*** -0.025** -0.030***

Fiscal balance 0.072*** 0.062** 0.073*** 0.064*** 0.070*** 0.067*** 0.072***

External balance 0.066*** 0.063* 0.065*** 0.070*** 0.064*** 0.067*** 0.066***

Default history -3.032*** -2.714 -3.110*** -2.941*** -3.027*** -2.969*** -3.035***

Newly rated 0.654***

Gov. effectiveness 0.045***

Bond issues 0.001***

Quadratic terms Yes

Market share -0.093

Country fixed effects Yes Yes Yes Yes Yes Yes Yes

N 1557 731 1557 1479 1557 1557 1557

adj. R-sq 0.423 0.451 0.427 0.454 0.432 0.430 0.422

AIC 4636.182 2245.279 4626.065 4348.357 4611.680 4620.058 4638.131

142

Second, given the importance of sovereign governance stressed in CRAs’ rating methodologies (e.g., Moody’s, 2013), I collect the World Governance Indicators published since 1996 by the World Bank, which aggregate 32 survey-based and proprietary data sources on six different dimensions: voice and accountability, political stability, government effectiveness, regulatory quality, rule of law and control of corruption.55 Consistent with CRAs’ rating methodologies, government effectiveness has the highest explanatory power of the governance indicators. As visible in model (4), the coefficient is highly significant and increases the R squared by 3.1% compared to the baseline. However, it has no significant impact on the year indicators, and neither do the other governance indicators (see Appendix 6).

Third, I use issue size as a proxy for the revenue-based incentive. He et al. (2012) show that issue size, as measured by the amount of securities issued by each company, leads to more lenient ratings for the structured finance market. As in He et al. (2012), I estimate issue size by collecting all sovereign bond issues for each year and country from capital IQ, and dividing them by the total amount of issues. As shown in (5), the resulting coefficient is statistically significant and of the expected sign but has a small magnitude and little influence on the other coefficients.

Fourth, given the non-linear relation between several factors and ratings, I follow Hill et al. (2010) and add quadratic terms for GDP growth, government debt and external balance. While the quadratic factors are significant, the cumulative magnitude and the explanatory power are similar to the linear specification. For brevity, the quadratic coefficients are not reported in model (6).

Last, I test whether competition has any effect on rating standards in model (7). Becker and Millbourn (2011) show that CRAs assign inflated ratings in industries where Fitch rates more companies. However, their measure does not capture the smaller CRAs, which, in the case of the sovereign market, have a significant combined market share. Moreover, their measure would not be varying in the cross-section for my sample, since Fitch’s market share will be the same for all countries. I instead use data on nine CRAs that are rating sovereigns, which I collect from the online appendix in Fuchs and Gehring (2015). The authors collected the first year of rating for the following CRAs:

55 These indicators range from 0% (worst) to 100% (best). For more details, see Kauffman et al.

(2011).

143

S&P, CI, Dagong, DBRS, Feri, Fitch, JCR, Moody’s and R&I. Using their data, I construct a country-specific yearly indicator by dividing the number of CRAs who rated the country in that year by the total possible number, 9. Since SEC’s certification as NSRSO is important for regulatory purposes, I also compute an alternative variable taking into account the ratings assigned by CRAs after they received SEC certification.56 However, none of the market share variables are significant. This could be driven by the fact that the fee structure is rather fixed in comparison with other rated products, minimizing the possibility for rating shopping.

4.5.2. Model stability

Equation (1) relies on the strong assumption of coefficient stability over time, which may not be realistic. For instance, CRAs may have become more conservative as a result of being criticized for failing to predict sovereign crises, or as a result of regulations penalizing them for being too lenient, such as the Dodd-Franck Act (as discussed in more detail in Dimitrov et al., 2015). A further concern is that the rating methodology may have changed over time, i.e. certain factors may have become relatively more/less important.

To test this, I estimate yearly cross-sectional regressions for each year t from 1993 to 2013 using the same variables as in model (1):

Figure 2 shows the standardized coefficients for the explanatory variables for all regressions, along with the yearly adjusted R2 values. No factor varies substantially over time, with only inflation having relatively lower coefficients in the beginning of the period (-0.54 in 1993 compared to around 0.2 from 1998 onwards), and several of the factors having a “peak” around 2001-2002. R2 is remarkably stable across the 20 years (the standard deviation is 0.05). Hence, it is unlikely that the increase in conservatism is a consequence of changes in rating methodology.

56 To exemplify, Albania has been rated by three CRAs in 2010 and hence has a ratio of 3/9;

however, only two CRAs were certified by SEC at that time, so the NSRSO market share would be 3/7, since only 7 CRAs received SEC certification prior to 2013. In Appendix 6, I use NSRSO market share instead of all CRAs without any significant change in the output.

= + + ′ (2)

144

Figure 2. Yearly regressions The figure displays the coefficients for the variables included in the base-case model (1) for each yearly (cross-sectional) regression. The dotted line represents the adjusted R2 for the corresponding regression. To further test the stability of the year indicators in (1), I estimate an alternative measure of conservatism as the difference between actual and predicted rating, similar to Alp (2013). To estimate the predicted rating, I use two alternative methods.

To further test the stability of the year indicators in (2), I estimate an alternative measure of stringency as the difference between actual and predicted rating, similar to Alp (2013). To estimate the predicted rating, I use two alternative methods.

First, I follow Fama and French (1973) by using the averages of the coefficients and estimated in the yearly regressions in (2) as an input for calculating

the predicted rating :

Second, I use the coefficients from the first regression in (2) estimated only for year 1993 to compute an alternative fitted value for ratings, 93 :

I then estimate the difference between actual and predicted ratings for the entire period using the two alternative estimations of the predicted rating:

= − (3c)

93 = − 93 (3d)

= ̅ + ̅ + ′′ (3a)

93 = + + ′′′ (3b)

145

As shown in Figure 3, the trends of the base specification estimated using (1) and the two predicted ratings are quite similar, meaning that the main specification and the downward trend are indeed stable over time. Figure 3. Year coefficient stability The figure compares the year coefficients (year_coeffs) in the baseline specification (1) with the differences between the actual and predicted ratings, where the predicted ratings were computed using the average yearly coefficients (RDFF) and the coefficients estimated using the 1993 values (RD93).

4.5.3. Impact of other time-varying factors

Another potential limitation of the main specification (1) is that year coefficients could capture other time-varying “global” factors, which cannot be included in the model since they do not vary across countries. To the extent that these factors affect local economies, it is important to control for them. To do this, I first use an alternative specification as in Baghai et al. (2014), by replacing year indicators with a time trend (1 for 1993, 2 for 1994, etc.):

= ′ + ′ + , + + ′ (4)

where b’ captures the trend in rating conservatism.57 Table 4 displays the results for the trend regressions.

57 Since this specification imposes a linear trend on rating standards, it serves solely as a

robustness test for whether the trend is significant and decreasing.

146

Table 4. Trend regressions The table presents the regression output after replacing year indicators with a trend variable. Year trend takes values of 1 for 1993 to 20 for 2013. NBER indicator takes value 1 for the periods considered to be US recession by NBER. Yield represents the lagged difference between US 10-year treasury bonds rate and 3-month LIBOR. All models are estimated using OLS and country fixed effects. Model (2) excludes US, while models (4) and (5) exclude both UK and US.

(1) (2) (3) (4) (5)

Year trend -0.097*** -0.099*** -0.100*** -0.120*** -0.120***

GDP per capita 4.278*** 4.469*** 4.351*** 4.701*** 4.742***

Real GDP growth 0.051** 0.061** 0.052** 0.065** 0.062**

Inflation -0.146 -0.144 -0.168* -0.133 -0.155*

Government debt -0.030*** -0.031*** -0.031*** -0.033*** -0.034***

Fiscal balance 0.050** 0.064*** 0.049** 0.063*** 0.060***

External balance 0.067*** 0.065*** 0.064*** 0.071*** 0.068***

Default history -3.160*** -3.090*** -3.121*** -3.026*** -3.002***

NBER indicator -0.492***

World GDP growth 0.028*** -0.110***

Yield -0.135*** 0.023***

N 1557 1545 1557 1501 1501

adj. R-sq 0.395 0.412 0.406 0.423 0.429

AIC 4689.598 4622.122 4663.018 4492.552 4478.052

Model (1) replicates the base specification by replacing the year indicators with the year trend. In (2), I follow related literature and add the NBER recession indicator (see, e.g., Hill et al., 2010), which takes the value 1 for the months classified as recession months by NBER. Since this indicator is a proxy for US recession, I use in (3) an alternative variable proposed by IMF as a measure of global economic cycles: global GDP growth (IMF, 1998). As an indicator for credit availability, I use the lagged yield spread as in, e g, Cantor and Packer (1996), Mora (2006) and Hill et al. (2010), measured as the difference between the US 10-year treasury bond yield and the 3-month LIBOR (model (4)). Finally, model (5) includes both the yield and world GDP growth. To alleviate endogeneity concerns, I exclude the UK and the US from models that include the yield spread.

In all specifications, the year trend coefficient changes minimally (-0.097 to -0.12) and is approximately equivalent to the year indicator regressions: a 0.1

147

decrease in ratings for each year is equivalent to a difference of 2 notches over the 20-year period. This value is very close to the 2013 year coefficient of the baseline specification (1), -2.2.

An important limitation of this approach is the assumption of linearity for the year trend, which might obscure the correlation with the other factors. I address this concern by regressing the year indicators from equation (1) on the NBER recession indicator, global GDP growth, and the lagged yield spread.

= + ′′ + ′′ (5)

where at are the year indicators estimated in (1), Bt are the time-varying global factors, and rt are the residuals. Table 5 displays the regression output. I use the same control variables as in the trend regressions.

As opposed to specification (4), where the linear trend eliminated year-to-year variation, equation (5) allows me to directly test whether these global factors drive my fixed effects. However, since the regression is based on 20 observations, the results should be interpreted with caution.

The adjusted R2s are rather low, with the highest being 15.5% for model (4),58 where I include both GDP growth and the credit spread. I also report the residuals, which, given that there are no omitted variables, would be a more accurate proxy for rating standards. While the magnitudes are lower relative to the original year coefficients, there is still a downward trend. For example, in model (2), which includes GDP growth, the residual for 1993 is 1.167, but it decreases to -1.041 for the last year, 2013. Therefore, the absolute change for the entire period is similar, corresponding to about two notches.

58 Compared to the first-stage regressions where it was as high as 95%.

148

Table 5. Second stage regression output The table presents the regression output of the second stage model (5), where the dependent variable is the year coefficient. NBER indicator takes value 1 for the periods considered to be US recession by NBER. Yield represents the lagged difference between US 10-year treasury bonds rate and 3-month LIBOR. World GDP growth is calculated in real terms. All models are estimated using OLS with robust standard error. The year coefficients in this output are the residuals of the regression rt.

(1) (2) (3) (4)

NBER indicator -0.874***

World GDP growth 0.012 -0.001

Yield 0.228** 0.228**

Residuals

1993 1.008 1.167 1.289 1.286

1994 1.259 1.362 1.123 1.125

1995 1.239 1.294 1.192 1.198

1996 1.005 1.166 0.875 0.871

1997 0.479 0.677 0.170 0.162

1998 0.030 0.231 -0.137 -0.145

1999 -0.006 0.139 -0.419 -0.422

2000 -0.095 0.061 -0.019 -0.022

2001 0.486 -0.188 0.096 0.089

2002 -0.307 -0.163 0.199 0.198

2003 -0.173 -0.133 0.263 0.272

2004 -0.026 0.011 -0.123 -0.115

2005 -0.059 0.032 -0.449 -0.447

2006 -0.325 -0.236 -0.759 -0.757

2007 -0.541 -0.505 -0.556 -0.548

2008 -0.172 -0.972 -0.612 -0.607

2009 -0.314 -0.931 -0.693 -0.705

2010 -0.596 -0.518 -0.197 -0.192

2011 -0.718 -0.657 -0.580 -0.574

2012 -0.967 -0.796 -0.663 -0.667

2013 -1.199 -1.041 . .

N 21 21 20 20

adj. R-sq 0.128 -0.045 0.155 0.105

AIC 46.985 50.779 42.986 44.984

4.6. Sample splits

As discussed in the introduction, there is both empirical and theoretical evidence for cross sectional variation in corporate rating standards. In this section, I test whether there is such variation for different groups of sovereigns, by taking into account two of the main “culprits” found throughout the corporate ratings literature: regulations and information asymmetry.

149

One implication of the theoretical model in Opp et al. (2013) is that rating-based regulations can lead to changes in rating standards. To test this, I use the risk-weighting scheme in the Basel accords as a proxy for regulatory restrictions based on ratings. More specifically, the scheme recommends certain risk weights as a function of the asset type and a risk criterion; these weights are then used as an input for the calculation of minimum capital requirements. For sovereign debt instruments, Basel I assigned a 0% risk weight to issuers that were OECD members, and 100% to non-OECD members (BCBS, 1988). In Basel II, there are instead five risk weights corresponding to specific rating groups: 0% for ratings above AA-; 20% for the A group; 50% for BBB group; 100% for BB, and 150% for below BB- and unrated sovereigns.59 Hence the first sample split follows the risk-weighting criteria across time (i.e., the shift from OECD membership to ratings in 2007 when Basel II was implemented). Since there are only two possible risk weights prior to Basel II, I create two groups: countries with a risk weight below 100% (i.e., OECD members prior to 2007 and sovereigns with investment ratings from 2007 onwards) and above 100% (non-OECD members and sovereigns with speculative ratings). The risk weights with their respective coding are displayed in Appendix 7. Since Alp (2013) and Baghai et al. (2014) find differences in rating standards between investment and speculatively graded firms, I also split my sample of countries into these two rating groups, prior to 2007.

Secondly, following the arguments developed in e.g. Skreta and Veldkamp (2009), I test whether information asymmetry leads to differences in sovereign rating standards. In the context of this paper, information asymmetry can be associated with less developed countries, since they tend to be more opaque, corrupt, and have less reliable macroeconomic data compared to advanced economies. Additionally, these countries have a relatively short rating history, meaning that CRAs need to put substantially more effort into assessing their creditworthiness. While Skreta and Veldkamp (2009) argue that “asset complexity” leads to inflated ratings for the MBS market, the outcome might be different for sovereign ratings. For example, CRAs may choose to err on the conservative side instead of investing additional resources, especially given the fact that the revenue generated by sovereign ratings is very low and that

59 As described in more detail in Appendix 7, these risk weights are different for corporate debt,

thus reinforcing the argument in Cornaggia et al. (2014) that ratings cannot be compared across different asset classes.

150

there is a relatively higher reputation cost for assigning lenient ratings.60 To test this hypothesis, I split the sample into developed and developing economies by using two widely used classifications: OECD membership and IMF’s development indicator.

Table 6 displays the year coefficients for the base-case specification for the various groups. The odd numbers represent the “high-quality” groups: (1) with a risk weight of below 100%, (3) having an investment rating, (5) being OECD members and (7) being classified by IMF as a developed economy. The even numbers are assigned to the “low-quality” groups. For a meaningful coefficient comparison, I also include the standard errors below the coefficients, and the P-values for the differences in the means of the two groups are reported in the P (χ2) column.

At first glance, most year coefficients seem relatively higher for the “low quality” groups. More specifically, the two-group difference in the year coefficients for 2013 ranges from 33% for the risk weight classification to 13% for the IMF classification. However, in order to assess whether these differences are significant, I re-estimate the sample splits using seemingly unrelated regression (SUR) estimation, which has the advantage of using the same covariance matrix when estimating the coefficients for both groups, allowing for a consistent comparison of the mean values. As displayed in the P (χ2) column, the differences are not statistically significant,61 except for a few years.

This finding is different from Alp (2013) and Baghai et al. (2014), and provides evidence that rating standards are consistent for different sovereign groups.62 Hence, in this case, neither rating-based regulation nor information asymmetry seem to affect rating standards.

60 This argument is derived from regulations that penalize CRAs for being too lenient, but not

for being too conservative.

61 An important aspect to note here is that, even if coefficients have different magnitudes, they also have proportionally different standard errors, resulting in low Chi-test values.

62 They do not report standard errors or p-values for differences in mean coefficients.

151

Tab

le 6

. S

amp

le s

pli

ts

The

tab

le d

ispl

ays

the

year

coe

ffici

ents

for

the

bas

e-ca

se m

odel

re-

estim

ated

for

var

ious

sub

sam

ples

. T

he e

stim

atio

n an

d th

e un

repo

rted

var

iabl

es a

re t

he s

ame

as in

m

odel

(1)

, T

able

2.

Gro

up(1

) in

clud

es s

over

eign

s w

ith a

ris

k w

eig

ht lo

wer

tha

n 10

0% a

ccor

ding

to

Bas

el I

and

II;

grou

p(2)

incl

udes

sov

erei

gns

with

ris

k w

eigh

ts h

ighe

r th

an 1

00%

. G

roup

s (3

) an

d (4

) co

nsis

t of

cou

ntrie

s w

ith a

n in

vest

men

t, re

spec

tivel

y sp

ecul

ativ

e,

ratin

g in

tha

t ye

ar.

Gro

ups

(5)

and

(6)

incl

ude

sove

reig

ns w

ith O

EC

D

mem

bers

hip,

res

pect

ivel

y no

n-O

EC

D m

embe

rs.

Gro

ups

(7)

and

(8)

repr

esen

t de

velo

ped,

res

pect

ivel

y de

velo

ping

/em

ergi

ng,

mar

kets

as

clas

sifie

d by

IM

F.

Sta

ndar

d er

rors

are

incl

uded

in s

quar

e br

acke

ts. *

, **,

and

***

cor

resp

ond

to 1

0%, 5

% ,

and

1% s

igni

fican

ce le

vels

. P-v

alue

s fo

r pa

irwis

e m

ean

diffe

renc

es a

re r

epor

ted

in th

e P

rob

(χ2)

colu

mn.

(1

) (2

)

(3)

(4)

(5

) (6

)

(7)

(8)

Lo

w r

isk

Hig

h ris

k P

(χ2)

Inv.

S

pec.

P

(χ2)

OE

CD

N

o-O

EC

D

P(χ

2)

Inds

N

on-in

ds

P(χ

2)

1994

0.

324*

**

-0.1

43

0.04

0.

244*

* -0

.118

***

0.00

0.

452*

**

-0.2

24

0.00

0.

404*

**

-0.2

24

0.01

[0

.119

] [0

.196

]

[0.1

18]

[0.0

30]

[0

.163

] [0

.184

]

[0.1

45]

[0.1

81]

1995

0.

212

-0.0

88

0.47

0.

225*

-0

.380

0.

03

0.19

8 -0

.236

0.

28

0.21

3 -0

.251

0.

29

[0

.170

] [0

.370

]

[0.1

31]

[0.2

46]

[0

.201

] [0

.344

]

[0.2

31]

[0.3

47]

1996

0.

287

-0.6

31

0.06

0.

065

-0.8

89**

0.

02

0.14

3 -0

.801

* 0.

06

0.10

0 -0

.930

**

0.08

[0

.218

] [0

.465

]

[0.1

89]

[0.3

75]

[0

.215

] [0

.462

]

[0.3

19]

[0.4

63]

1997

-0

.451

-0

.701

0.

72

-0.3

46

-0.9

93**

0.

15

-0.6

94

-0.9

49*

0.72

-0

.501

-1

.287

**

0.35

[0

.403

] [0

.554

]

[0.2

57]

[0.4

28]

[0

.431

] [0

.556

]

[0.5

62]

[0.5

82]

1998

-0

.616

* -1

.379

*

0.35

-0

.762

**

-1.3

50**

* 0.

33

-0.9

02**

-1

.645

**

0.38

-0

.636

-2

.173

***

0.11

[0

.341

] [0

.750

]

[0.3

38]

[0.4

91]

[0

.374

] [0

.756

]

[0.5

61]

[0.7

50]

1999

-0

.606

* -1

.617

**

0.19

-0

.884

**

-1.5

18**

* 0.

20

-0.8

70**

-1

.842

***

0.21

-0

.593

-2

.233

***

0.06

[0

.322

] [0

.701

]

[0.3

36]

[0.3

70]

[0

.347

] [0

.695

]

[0.5

38]

[0.6

74]

2000

-0

.839

**

-1.7

76**

0.

25

-0.9

98**

* -1

.676

***

0.15

-1

.106

**

-2.0

16**

* 0.

27

-0.8

89

-2.2

41**

* 0.

16

[0

.412

] [0

.702

]

[0.3

46]

[0.3

30]

[0

.430

] [0

.699

]

[0.6

27]

[0.6

80]

2001

-1

.032

**

-2.1

69**

* 0.

18

-1.2

74**

* -2

.227

***

0.08

-1

.338

***

-2.4

13**

* 0.

19

-1.1

68*

-2.5

18**

* 0.

17

[0

.425

] [0

.729

]

[0.3

48]

[0.4

34]

[0

.442

] [0

.703

]

[0.6

60]

[0.6

87]

2002

-1

.021

**

-2.0

68**

* 0.

21

-1.1

77**

* -2

.222

***

0.04

-1

.251

**

-2.2

85**

* 0.

21

-1.1

13

-2.3

49**

* 0.

20

[0

.439

] [0

.712

]

[0.3

46]

[0.3

82]

[0

.473

] [0

.685

]

[0.6

64]

[0.6

56]

2003

-0

.972

**

-1.9

61**

* 0.

25

-1.0

62**

* -2

.138

***

0.05

-1

.153

**

-2.1

58**

* 0.

25

-1.0

92

-2.1

83**

* 0.

28

[0

.484

] [0

.719

]

[0.3

62]

[0.4

16]

[0

.538

] [0

.688

]

[0.7

24]

[0.6

72]

2004

-0

.885

* -1

.768

**

0.34

-0

.867

**

-2.0

54**

* 0.

03

-0.9

90

-1.9

88**

* 0.

28

-0.9

16

-2.0

15**

* 0.

30

[0

.507

] [0

.761

]

[0.3

52]

[0.4

52]

[0

.592

] [0

.720

]

[0.7

71]

[0.6

94]

2005

-0

.995

* -1

.719

**

0.44

-0

.939

**

-2.0

27**

* 0.

05

-1.2

00*

-1.9

57**

* 0.

42

-1.1

20

-1.9

34**

* 0.

46

152

Tab

le 6

. S

amp

le s

pli

ts (

con

tin

ued

)

[0

.536

] [0

.778

]

[0.3

66]

[0.4

36]

[0

.615

] [0

.723

]

[0.8

23]

[0.6

97]

2006

-1

.294

**

-1.8

94**

0.

54

-1.1

68**

* -2

.114

***

0.11

-1

.538

**

-2.1

67**

* 0.

52

-1.4

05

-2.1

20**

* 0.

53

[0

.581

] [0

.798

]

[0.3

94]

[0.4

53]

[0

.659

] [0

.728

]

[0.8

81]

[0.6

97]

2007

-1

.411

**

-2.2

03**

* 0.

44

-1.4

43**

* -2

.032

***

0.34

-1

.756

**

-2.3

68**

* 0.

55

-1.7

35*

-2.2

53**

* 0.

67

[0

.613

] [0

.826

]

[0.4

24]

[0.4

58]

[0

.707

] [0

.744

]

[0.9

34]

[0.7

14]

2008

-1

.860

***

-2.7

32**

* 0.

43

-1.8

34**

* -2

.432

***

0.36

-2

.153

***

-2.8

85**

* 0.

49

-2.1

44**

-2

.762

***

0.62

[0

.664

] [0

.860

]

[0.4

48]

[0.4

89]

[0

.763

] [0

.752

]

[0.9

89]

[0.7

27]

2009

-1

.999

***

-2.8

88**

* 0.

42

-1.9

19**

* -2

.495

***

0.39

-2

.006

**

-3.0

93**

* 0.

31

-2.0

93**

-3

.042

***

0.45

[0

.678

] [0

.865

]

[0.4

55]

[0.5

15]

[0

.767

] [0

.761

]

[1.0

07]

[0.7

32]

2010

-1

.419

**

-2.3

34**

* 0.

40

-1.3

12**

* -1

.955

***

0.32

-1

.300

-2

.559

***

0.25

-1

.411

-2

.514

***

0.37

[0

.669

] [0

.861

]

[0.4

23]

[0.5

08]

[0

.799

] [0

.757

]

[0.9

82]

[0.7

21]

2011

-1

.552

**

-2.5

28**

* 0.

40

-1.3

52**

* -2

.116

***

0.28

-1

.670

* -2

.603

***

0.43

-1

.855

-2

.547

***

0.62

[0

.766

] [0

.892

]

[0.4

61]

[0.5

43]

[0

.884

] [0

.788

]

[1.1

52]

[0.7

38]

2012

-1

.776

**

-2.7

39**

* 0.

46

-1.5

05**

* -2

.186

***

0.36

-2

.075

**

-2.8

05**

* 0.

57

-2.4

14*

-2.6

49**

* 0.

87

[0

.850

] [1

.015

]

[0.5

19]

[0.5

59]

[0

.970

] [0

.857

]

[1.2

41]

[0.7

60]

2013

-2

.003

**

-3.0

10**

* 0.

43

-1.7

49**

* -2

.455

***

0.35

-2

.425

**

-3.0

35**

* 0.

62

-2.5

46**

-2

.905

***

0.80

[0

.864

] [0

.953

]

[0.5

45]

[0.5

35]

[0

.904

] [0

.841

]

[1.1

91]

[0.7

71]

N

750

807

92

6 63

1

579

978

56

0 99

7

adj.R

2 0.

334

0.39

8

0.

410

0.36

6

0.43

5 0.

446

0.

442

0.45

3

AIC

19

90

2329

23

50

1777

1774

27

88

17

59

2751

RM

SE

0.

897

1.00

8

0.

849

0.97

3

1.09

6 0.

994

1.

138

0.94

9

153

4.7. Is conservatism justified?

4.7.1. Bond market implications

So far, I have shown that ratings have become more conservative over the last 20 years and that this conservatism seems to be unrelated either to observable sovereign-specific characteristics or to global economic patterns. If this is indeed the case, bond markets would take this conservatism into account when pricing debt (Baghai et al., 2014), and hence conservatism would also explain credit spreads, above and beyond the explanatory power of rating levels. Furthermore, to the extent that the bond market perceives this conservatism as not justified, it would “undo” its effects on the spread, i.e., strictly rated countries would have lower credit spreads relative to the loosely rated ones, given the same rating.

Similarly to Baghai et al. (2014), I use one of the conservatism measures calculated in (4), RD93it, as explanatory variable, in addition to the current rating level, in the following equation:

= ′′′ + 93 + + ′′′ + ′′′ (6)

where Spreadit is the difference between US treasury 10-year bond and the sovereign bond yield,63 and RD93it is the difference between actual and predicted ratings as described in (4), Rit is the current rating level, and Si are sovereign-specific fixed effects. For robustness, I also estimate (6) using RDFFit instead of RD93it. The results are displayed in Table 7.

63 I collect the data for plain, fixed-coupon bonds with a 5-year maturity from Bloomberg, on a

yearly basis from 1993 to 2013, resulting in 742 country-year observations. Since all bonds are of the same type, I don’t use additional control variables; any other sovereign-specific variables are controlled for using fixed effects.

154

Table 7. Effect of conservatism on bond spreads The table displays the regression output for the credit spread models. The dependent variable is the difference between US T-bill and the 5-year sovereign bond spread. Rating represents the current sovereign rating level, while “Conservatism” is the difference between the actual and the predicted rating. In (1), conservatism is RD93 (predicted rating is calculated using the coefficients from the 1993 year regression), while in (2) it is RDFF (predicted rating is computed using the average of all yearly regressions' coefficients). Country and period fixed effects are included. Errors are clustered at country level and corrected for heteroskedasticity.

(1) (2)

Conservatism 0.255* 0.191**

Rating -1.198** -1.080*

N 742 742

adj. R-sq 0.186 0.198

AIC 3625.253 3613.5

As expected, ratings are significant and negatively correlated with credit spreads. More importantly, bond markets seem to regard parts of the conservatism as unjustified, since they “undo” about ¼ of its effect (i.e., the coefficient for conservatism is roughly 20% of the ratings coefficient, and of the opposite sign). This effect is quite similar to the one found in Baghai et al. (2014) and Alp (2013), who find that roughly one-third of the conservatism is reversed when pricing corporate debt spreads.

4.7.2 Time variation of default rates

Could it be that sovereign risk increased over time, due to various factors such as technological development, globalization, increasingly competitive business environment, etc.? In this case, the trend may simply reflect an increase in riskiness over time for the very same ratios: it is possible that a 60% debt-to-GDP ratio is riskier today, as opposed to 20 years ago. If this is the case, an increase in conservatism would be justified. I test this by investigating the time variation of sovereign default rates.

The definition of sovereign default is less straightforward than for firms, being at times context- and instrument-specific. The definition varies among CRAs and there is a distinction between foreign and domestic debt. For my analysis, I use S&P’s broader definition of default for foreign debt,64 where a sovereign default occurs when the central government fails to pay the scheduled debt service on the due date or tenders an exchange offer of new debt with less-

64 Since I look at foreign sovereign ratings, only foreign debt-related defaults are of interest.

155

favorable terms than the original issue. Furthermore, S&P only takes into account debt held outside the public sector (e.g., failure to pay back an IMF loan is not considered a default), while the amount owed per se is irrelevant.65

Following this definition, I collect data on sovereign defaults from Bank of Canada’s database described in Beers and Nadeau (2014). The dataset contains detailed data on central government debt defaults for 143 countries over 30 years and is much more comprehensive than IMF and WDI debt data. Furthermore, the dataset includes a breakdown of defaulted debt by type of creditor and currency: (a) Paris Club; (b) Other official creditors;(c) IMF; (d) FC bank loans; (e) FC bonds; (f) other private debt and (g) LC debt. Therefore, I can easily disentangle the type of debt relevant for S&P’s definition of default: foreign currency private debt (categories d, e and f).

Figure 4 displays historical data on sovereign defaults, both in terms of number of sovereigns (Panel A) and in terms of value of defaulted debt relative to world GDP (Panel B).

Both the number of defaulted countries and the amount of defaulted foreign debt exhibit downward trends over time, indicating that my results are not driven by an increase in sovereign debt riskiness.

65 I got this information from an S&P consultant.

156

Figure 4. Time variation of total sovereign defaults Panel A shows the number of sovereign issuers who defaulted on either foreign currency loans or bonds as reported in Standard & Poor’s (2006). Panel B displays the ratios of total defaulted public debt, respectively foreign currency private debt, relative to world GDP; the data is calculated from Bank of Canada’s database described in Beers and Nadeau (2015).

157

I further investigate whether this trend holds in a multivariate setting. As in Baghai et al. (2014), I test the relation between the year trend (as defined in 4) and the default frequencies for various rating groups, after controlling for business cycles. I estimate rating-specific default frequencies using S&P’s “static pools.”66 These “static pools” represent a mapping of predicted default probabilities for each broad rating group with non-zero default frequencies. Each year, these default probabilities are estimated for up to 15 years ahead. As an example, for a sovereign rated BBB in 2002, the default frequency is 0% for one year ahead (2003), 6.3% for two years ahead, etc. Since these matrices are triangular (i.e., for ratings assigned in recent years, the time horizon is shorter), I average the default frequencies across all time horizons in order to obtain enough observations for my entire sampling period.67 Appendix 9 exemplifies the static pool for the BBB group, and the average default probabilities I calculated for each year. I regress these default frequencies for each of the lower rating groups (A, BBB, BB, B and below CCC) against the year trend, while controlling for other global economic conditions by using inflation-adjusted world GDP growth.68 The output is reported in Table 8.

Table 8. Default rates over time The dependent variable is the average default probability over all time horizons for each broad rating group. The default probabilities corresponding to ratings for CCC group and below are grouped together due to data limitations. Global GDP growth represents the inflation-adjusted annual GDP growth rate aggregated for all countries. Trend takes value 1 for first year of the sample, 2 for the second, etc. Errors are corrected for heteroskedasticity.

Variable A BBB BB B <CCC

Global GDP growth -0.155 0.243 0.418 -0.805 7.816

Year trend 0.076 -0.473*** -0.729*** -1.520*** 1.472

N 23 23 23 20 13

R-sq 0.158 0.410 0.544 0.744 0.142

66 These data are available on Ratings Direct, in “Sovereign Rating and Country T&C

Assessment Histories” reports.

67 S&P’s time horizons in these matrices range from 0 to 15 years over a time span from 1992 to 2014, meaning that until year 2000 I would have the average of 15 values (2014-15), while in 2014 I have only one value.

68 I also test adding more control variables measuring the global macro environment in Appendix 8, with no material changes in the trend coefficients.

158

For BBB, BB and B rating groups the trend is significant and negatively correlated with default rates. While the sample size is too small to lead to any robust statistical inferences, the results suggest that the increase in conservatism is not consistent with increases in actual sovereign defaults on foreign debt.

4.8. Conclusion

By using several measures of conservatism, I find that sovereign rating standards have become increasingly stricter over time, displaying a similar trend to corporate rating standards. More specifically, I find that a country with the same credit risk would receive in 2013 a rating that is up to 2.2 notches lower (i.e., AA+ to AA-) relative to 1993.

This finding is robust to including a range of control variables measuring both country-specific and global factors that could impact credit risk, as well as factors that may be relevant for the accuracy of the rating process, such as information asymmetry, competition or regulatory reliance.

Having established that my measure of conservatism is not biased by various omitted variables, I also show that the bond market, as well as the actual default rates, are consistent with the concept of increasing conservatism over time. First, the bond market seems to “undo” about 20% of the effect of conservatism when pricing sovereign debt, which would not make sense if the ratings accurately reflected their credit risk. Second, by using a new, detailed dataset compiled by Bank of Canada, as well S&P’s rating-specific predicted default frequencies, I show that default rates have been decreasing over time. This decrease is persistent even after controlling for changes in the global economic environment.

Lastly, in contrast to previous studies on corporate ratings, I find that rating standards do not vary cross-sectionally. There are no significant differences between developed and developing countries, nor between the speculative and the investment rated sovereigns. I conclude that factors found to influence the rating process, such as information asymmetry and regulatory reliance, do not have a distorting effect for sovereign ratings.

The Eurozone crisis, as well as the reactions following more recent downgrades of several developed countries, are clear proof of the far-reaching

159

effect of sovereign rating changes. While a corporate rating change will affect its investors, customers and suppliers, a sovereign rating change will affect the entire economic and business environment (given that the sovereign “ceiling rule” is still de-facto implemented). Bearing this in mind, my findings are important and economically relevant. However, since my measure of conservatism (which is estimated by using the same method as in prior research on credit ratings) fundamentally depends on the model specification, more research is needed to confirm whether CRAs are indeed systematically biased in their sovereign rating assessments.

160

References

Alp, A. (2013). Structural shifts in credit rating standards. The Journal of Finance, 68(6), 2435-2470.

Baghai, R. P., Servaes, H., & Tamayo, A. (2014). Have rating agencies become more conservative? Implications for capital structure and debt pricing. The Journal of Finance, 69(5), 1961-2005.

Barroso, J.M.D. (2010). Statement to the European Parliament prior to the meeting of the Heads of State and Government of the Euro Area. European Parliament Plenary, Brussels.

Basel Committee on Banking Supervision (1988). International Convergence of Capital Measurement and Capital Standards. Bank for International Settlements, Basel.

Basel Committee on Banking Supervision (2004). International Convergence of Capital Measurement and Capital Standards: A Revised Framework. Bank for International Settlements, Basel.

Becker, B., & Milbourn, T. (2011). How did increased competition affect credit ratings? Journal of Financial Economics, 101(3), 493-514.

Beers, D., & Nadeau, J. S. (2014). Introducing a new database of sovereign defaults. Available at SSRN 2609162 (2014).

Blume, M. E., Lim, F., & MacKinlay, A. C. (1998). The declining credit quality of US corporate debt: Myth or reality? The Journal of Finance, 53(4), 1389-1413.

Bolton, P., Freixas, X., & Shapiro, J. (2012). The credit ratings game. The Journal of Finance, 67(1), 85-111.

Brooks, R., Faff, R. W., Hillier, D., & Hillier, J. (2004). The national market impact of sovereign rating changes. Journal of Banking & Finance, 28(1), 233-250.

Cantor, R., & Packer, F. (1996). Determinants and impact of sovereign credit ratings. The Journal of Fixed Income 6.3: 76-91.

Cornaggia, J., Cornaggia, K. R., & Hund, J. (2014). Credit ratings across asset classes: A long-term perspective. Available at SSRN 1909091.

Dimitrijevic, A., Beers, D., Chambers, J., & Kraemer, M. (2011). Sovereign government rating methodology and assumptions. RatingsDirect.

Dimitrov, V., Palia, D., & Tang, L. (2015). Impact of the Dodd-Frank act on credit ratings. Journal of Financial Economics, 115(3), 505-520.

Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal of Political Economy, 81(3), 607-636.

Ferreira, M. A., & Gama, P. M. (2007). Does sovereign debt ratings news spill over to international stock markets? Journal of Banking & Finance, 31(10), 3162-3182.

161

Ferri, G., Liu, L. G., & Stiglitz, J. E. (1999). The procyclical role of rating agencies: Evidence from the East Asian crisis. Economic Notes, 28(3), 335-355.

Fuchs, A., & Kai Gehring (2015). The home bias in sovereign ratings. Courant Research Centre ‘Poverty, Equity and Growth in Developing and Transition Countries’ Discussion Paper 179.

Gaillard, N. (2011). A Century of Sovereign Ratings. Springer Science & Business Media.

Gande, A., & Parsley, D. C. (2005). News spillovers in the sovereign debt market. Journal of Financial Economics, 75(3), 691-734.

Goel, A. M., & Thakor, A. V. (2011). Credit ratings and litigation risk. Unpublished working paper.

He, J., Qian, J., & Strahan, P. E. (2012). Are all ratings created equal? The impact of issuer size on the pricing of mortgage‐backed securities. The Journal of Finance, 67(6), 2097-2137.

Hill, P., Brooks, R., & Faff, R. (2010). Variations in sovereign credit quality assessments across rating agencies. Journal of Banking & Finance, 34(6), 1327-1343.

Jaimovich, D., & Panizza, U. (2010). Public debt around the world: a new data set of central government debt. Applied Economics Letters, 17(1), 19-24.

Kaufmann, D., Kraay, A., & Mastruzzi, M. (2011). The worldwide governance indicators: methodology and analytical issues. Hague Journal on the Rule of Law, 3(2), 220-246.

Lancaster, T. (2000). The incidental parameter problem since 1948. Journal of Econometrics, 95(2), 391-413.

Mora, N. (2006). Sovereign credit ratings: Guilty beyond reasonable doubt? Journal of Banking & Finance, 30(7), 2041-2062.

Morris, S. (2001). Political correctness. Journal of Political Economy, 109(2), 231-265.

Moody’s (2013). Sovereign Rating Methodology. Moody’s Investor Service.

Opp, C. C., Opp, M. M., & Harris, M. (2013). Rating agencies in the face of regulation. Journal of Financial Economics, 108(1), 46-61.

Skreta, V., & Veldkamp, L. (2009). Ratings shopping and asset complexity: A theory of ratings inflation. Journal of Monetary Economics, 56(5), 678-695.

Standard and Poor’s (2008), Sovereign Ratings’ Methodology. RatingsDirect.

Standard & Poor’s Ratings Services (2015). U.S. Ratings Fees Disclosure. RatingsDirect.

162

Appendices

Appendix 1. Sovereign rating scale The appendix displays the coding of the dependent variable and the frequencies for each value. The dashed line separates the investment- from the speculative-graded rating groups.

Rating number (fine)

Rating number (broad)

S&P foreign currency rating scale

Country-year observations

Frequency

21 8 AAA 300 15.50

20 7 AA+ 96 4.96

19 7 AA 87 4.50

18 7 AA- 80 4.13

17 6 A+ 62 3.20

16 6 A 124 6.41

15 6 A- 103 5.32

14 5 BBB+ 81 4.19

13 5 BBB 91 4.70

12 5 BBB- 125 6.46

11 4 BB+ 115 5.94

10 4 BB 123 6.36

9 4 BB- 135 6.98

8 3 B+ 167 8.63

7 3 B 130 6.72

6 3 B- 80 4.13

5 2 CCC+ 16 0.83

4 2 CCC 4 0.21

3 2 CCC- 2 0.10

2 2 CC 3 0.16

2 2 C 0 0.00

1 1 SD/D 11 0.57

Total 1935 100

163

Ap

pen

dix

2. V

aria

ble

des

crip

tio

n

Var

iabl

e D

escr

iptio

n un

it tr

ansf

orm

atio

n la

g ex

p.si

gn

sour

ce

Mai

n m

odel

Yea

r ef

fect

s (1

993.

1994

…20

13)

D

umm

y va

riabl

e w

ith f

irst

year

as

a ba

se

bina

ry

none

N

-

own

calc

ulat

ions

GD

P p

er c

apita

G

DP

/ c

apita

M

US

D

log

Y

+

Wor

ld B

ank

GD

P g

row

th

Yea

rly

real

GD

P g

row

th

%

3-ye

ar m

ovin

g a

vera

ge

Y

+

Wor

ld B

ank

Infla

tion

CP

I gr

owth

rat

e at

cur

rent

pric

es

%

3-ye

ar m

ovin

g a

vera

ge

Y

- O

xfor

d E

cono

mic

s

Cur

rent

acc

ount

ba

lanc

e

Sum

of t

he v

alue

of i

mpo

rts

of g

oods

and

ser

vice

s pl

us n

et r

etur

ns o

n in

vest

men

ts a

broa

d, m

inus

the

va

lue

of e

xpor

ts o

f goo

ds a

nd s

ervi

ces

rela

tive

to

GD

P

%

3-ye

ar m

ovin

g a

vera

ge

Y

+

Oxf

ord

Eco

nom

ics

Bud

get

bala

nce

Diff

eren

ce b

etw

een

gove

rnm

ent

reve

nues

and

ex

pend

iture

s re

lativ

e to

GD

P

%

3-ye

ar m

ovin

g a

vera

ge

Y

+

Wor

ld B

ank

Cen

tral

go

vern

men

t de

bt

Cen

tral

gov

ernm

ent

debt

rel

ativ

e to

GD

P

%

none

Y

-

Wor

ld B

ank

Def

ault

hist

ory

Num

ber

of y

ears

sin

ce la

st d

efau

lt [1

;0]

1/n,

whe

re N

is t

he n

umbe

r of

ye

ars

from

last

def

ault;

0 f

or n

o pa

st d

efau

lts

N

- S

&P

's G

loba

l C

redi

t Por

tal

Rob

ustn

ess

test

s

New

ly r

ated

1

for

the

first

yea

r w

ith a

vaila

ble

ratin

g bi

nary

no

ne

N

- ow

n ca

lcul

atio

ns

New

rat

ings

/yea

r 1

for

whe

ther

a c

ount

ry is

firs

t ra

ted

in a

spe

cific

yea

r bi

nary

no

ne

N

- ow

n ca

lcul

atio

ns

164

Gov

ernm

ent

effe

ctiv

enes

s

Per

cent

ile r

anki

ng o

f cou

ntrie

s ba

sed

on s

ever

al

qual

itativ

e (s

urve

y-ba

sed)

gov

erna

nce

fact

ors

[0;

10

0]

none

Y

+

W

orld

Ban

k

Bon

d is

sues

N

umbe

r of

issu

ed s

over

eign

bon

ds a

nd C

DS

rel

ativ

e to

the

tota

l num

ber

for

the

entir

e sa

mpl

e %

no

ne

Y

+

S&

P's

Cap

itaI

IQ.

own

calc

ulat

ions

Mar

ket s

hare

N

umbe

r of

CR

As

ratin

g th

e so

vere

ign

in a

giv

en

year

/tot

al n

o. o

f C

RA

s th

at r

ate

sove

reig

ns

%

none

N

+

F

uchs

and

G

ehrin

g(20

15)

Yea

r_co

effs

year

(tim

e fix

ed e

ffect

s) c

oeffi

cien

ts e

stim

ated

fro

m

the

base

-cas

e m

odel

(1)

cont

inuo

us

none

N

-

own

calc

ulat

ions

RD

FF

alte

rnat

ive

cons

erva

tism

mea

sure

com

pute

d as

the

diffe

renc

e be

twee

n ac

tual

and

pre

dict

ed r

atin

g,

whe

re t

he p

redi

cted

rat

ing

is e

stim

ated

usi

ng th

e av

erag

e co

effic

ient

s es

timat

ed fr

om th

e 20

cro

ss-

sect

iona

l yea

rly r

egre

ssio

ns u

sing

the

bas

e ca

se

spec

ifica

tion

cont

inuo

us

none

N

-

own

calc

ulat

ions

RD

93

alte

rnat

ive

cons

erva

tism

mea

sure

com

pute

d as

the

diffe

renc

e be

twee

n ac

tual

and

pre

dict

ed r

atin

g,

whe

re t

he p

redi

cted

rat

ing

is c

ompu

ted

usin

g th

e co

effic

ient

s of

the

base

-cas

e m

odel

est

imat

ed fo

r th

e fir

st y

ear,

199

3

cont

inuo

us

none

N

-

own

calc

ulat

ions

Yea

r tr

end

Ord

inal

yea

r ra

nkin

g {1

.2...

20}

none

N

-

own

calc

ulat

ions

NB

ER

indi

cato

r 1

for

"cris

is"

year

s as

cla

ssifi

ed b

y N

BE

R o

r w

ith

mor

e t

han

5 s

yste

mic

cris

es

bina

ry

none

Y

-

NB

ER

. IM

F

WP

Glo

bal G

DP

gr

owth

ye

ar-o

n-ye

ar r

eal w

orld

GD

P g

row

th

%

none

N

+

W

orld

Ban

k

Yie

ld

Spr

ead

betw

een

10 y

ears

T-b

ill a

nd 3

mon

ths

LIB

OR

%

no

ne

Y

- D

atas

trea

m

A;B

BB

;BB

;B;<

CC

C

aver

age

pred

icte

d de

faul

t pr

obab

ilitie

s ov

er a

15-

year

hor

izon

for

the

broa

d ra

ting

grou

ps:

A (

A+

/A/A

-);

BB

B (

BB

B+

/BB

B/B

BB

-);

BB

(B

B+

/BB

/BB

-);

B

(B+

/B/B

-);

CC

C (

CC

C+

/CC

C/C

CC

-/C

C/C

/SD

/D)

%

none

N

S&

P.

Rat

ings

' D

irect

165

Ap

pen

dix

3.

Tim

e av

erag

es o

f m

ain

exp

lan

ato

ry v

aria

ble

s

166

Ap

pen

dix

4.

Ave

rag

e ra

tin

g o

ver

tim

e A

vera

ge r

atin

g ov

er t

ime

for

the

full

sam

ple

(unb

alan

ced)

and

for

the

sam

ple

of c

ount

ries

rate

d pr

ior

to 1

993

(bal

ance

d).

167

Ap

pen

dix

5. Y

ear

coef

fici

ents

all

reg

ress

ion

s

The

app

endi

x di

spla

ys t

he y

ear

coef

ficie

nts

for

all t

he r

egre

ssio

n m

odel

s de

scrib

ed in

Tab

les

2 an

d 3.

“B

ase”

is t

he b

asel

ine

mod

el (

1),

“No

FE

” is

th

e ba

selin

e m

odel

est

imat

ed w

ithou

t co

untr

y fix

ed e

ffect

s (m

odel

(2)

),

“OP

” is

the

sam

e m

odel

est

imat

ed u

sing

ord

ered

pro

bit

in (

3),

Tab

le 2

. T

he

rest

of

the

mod

els

are

estim

ated

in

Tab

le 3

: “B

efor

e 19

93”

corr

espo

nds

to (

2) a

nd i

s es

timat

ed u

sing

onl

y co

untr

ies

rate

d be

fore

199

3; “

New

” co

rres

pond

s to

(3)

and

inc

lude

s th

e ne

wly

rat

ed d

umm

y; “

Gov

” is

est

imat

ed i

n (4

) an

d in

clud

es t

he w

orld

gov

erna

nce

inde

x va

riabl

e; “

Siz

e”

corr

espo

nds

to (

5) a

nd in

clud

es t

he b

ond

issu

es v

aria

ble;

"Q

V"

is m

odel

(6)

and

incl

udes

qua

drat

ic t

erm

s fo

r G

DP

gro

wth

, de

bt,

fisca

l and

ext

erna

l ba

lanc

es;

and

“Sha

re”

corr

espo

nds

to (

7) a

nd in

clud

es th

e m

arke

t sha

re v

aria

ble.

168

Appendix 6. Robustness tests: additional variables The table reports additional robustness tests. All models are estimated using OLS and country fixed effects, and include the control variables for the baseline specification (for comparison, year coefficients for this specification are reported in the first column, which is the same as in Table 2. In (2), I use the market share of all SEC-approved rating agencies (NSNROs). (3) reports the year coefficients of a specification where I add indicator variables for the specific year when the sovereign was first rated (for brevity, the coefficients for each of these variables are not reported). In the fourth column, I add several variables mentioned in the rating methodology or used in prior research: external debt relative to GDP, an indicator variable taking the value 1 if the country’s currency is traded, and a measure of foreign reserves relative to exports. Lastly, I include all the World Governance Indicators (survey-based measures for various government governance aspects), described in detail in e g Kaufmann et al. (2011). Standard Errors are corrected for heteroskedasticity and clustered at country level. *, **, and *** correspond to 10%, 5% and 1% significance levels.

(1) (2) (3) (4) (5)

1994 0.251** 0.184 0.258** 0.296

1995 0.231* 0.152 0.231* 0.412

1996 -0.002 -0.086 0.057 0.163 0.000

1997 -0.529 -0.617 -0.576 -0.501 -0.474*

1998 -0.978** -1.092*** -0.976** -1.087* -0.860***

1999 -1.014*** -1.146*** -1.026*** -1.127** -0.909***

2000 -1.103*** -1.243*** -1.082*** -1.140** -0.914***

2001 -1.396*** -1.546*** -1.420*** -1.423*** -1.179***

2002 -1.315*** -1.474*** -1.305*** -1.296*** -1.178***

2003 -1.181*** -1.335*** -1.249*** -1.044** -0.958***

2004 -1.034*** -1.197*** -1.020*** -0.834* -0.795**

2005 -1.067*** -1.232*** -1.021*** -1.016** -0.742**

2006 -1.333*** -1.509*** -1.284*** -1.243** -1.011***

2007 -1.549*** -1.817*** -1.511*** -1.480*** -1.221***

2008 -2.054*** -2.330*** -2.003*** -2.024*** -1.741***

2009 -2.196*** -2.469*** -2.146*** -2.158*** -1.828***

2010 -1.604*** -1.878*** -1.550*** -1.496*** -1.287***

2011 -1.726*** -2.001*** -1.670*** -1.686*** -1.387***

2012 -1.975*** -2.261*** -1.916*** -1.950*** -1.640***

2013 -2.207*** -2.494*** -2.144*** -2.103*** -1.824***

Share SEC 0.696

Rating_1994 – Rating_2013 YES

External debt/GDP -0.003

Traded currency -0.303

Foreign reserves/exports -0.119

Accountability index

Corruption index 0.006

Government effectiveness index -0.007

Rule of law index 0.024**

Regulatory strength index 0.023

Government stability index 0.025**

N 1557 1557 1557 1238 1478

adj. R-sq 0.423 0.424 0.430 0.451 0.473

AIC 4636.182 4633.501 4624.215 3778.365 4298.344

169

Appendix 7. Basel risk-weighting scheme The appendix compares the risk-weighting scheme for claims on sovereigns and corporates for Basel I and II. The dates in brackets represent the implementation period. The last column describes the group coding according to the risk weighting.

Risk weight Sovereigns Corporates Variable code

Basel I (1988)

0% OECD member - 0

100% non-OECD all 1

Basel II (2007/2008)

0% AAA to AA- - 0

20% A+ to A- AAA to AA- 0

50% BBB+ to BBB- A+ to A- 0

100% BB+ to B-; not rated BBB+ to BBB-; not rated 1

150% below B- below BB- 1

Appendix 8. Default rates over time – other global factors The appendix reports the output from the specification in Table 8, after adding several macroeconomic aggregates for the world economic/credit climate collected from the World Bank: GDP growth, net flow of stocks relative to GDP, total unemployment relative to available labor force, trade flow relative to GDP, and inflation rate. The variable of interest, trend, takes value 1 for first year of the sample, 2 for the second, etc. The dependent variable is the average default probability over all time horizons for each broad rating group. The default probabilities corresponding to ratings for CCC group and below are grouped together, due to data limitations. Errors are corrected for heteroskedasticity. *, **, and *** correspond to 10%, 5% and 1% significance levels.

Variable A BBB BB B <CCC

Trend 0.205 -0.579* -1.055* -1.789** -15.364*

GDP growth 0.022 -0.199 -0.141 -0.525 -0.111

Stocks -0.036*** -0.041 -0.070* -0.070* 0.113

Unemployment 1.771 -3.651 8.424 0.758 113.044

Trade -0.150 0.009 0.602 0.690 21.431*

Inflation 0.285 -0.230 -0.242 0.460 -30.935**

N 21 21 21 19 12

adj. R-sq 0.512 0.543 0.628 0.642 0.041

170

Ap

pen

dix

9.

Fo

reig

n-c

urr

enc

y 'B

BB

' sta

tic

po

ols

an

d d

efau

lt r

ates

(19

92–2

014)

T

he ta

ble

disp

lays

, exc

ept f

or th

e la

st c

olum

n, S

&P

's p

redi

cted

def

ault

freq

uenc

ies

for

the

BB

B r

atin

g gr

oup.

The

tabl

e is

take

n di

rect

ly fr

om S

&P

's la

test

rep

ort “

Sov

erei

gn

Rat

ing

And

Cou

ntry

T&

C A

sses

smen

t H

isto

ries,

” av

aila

ble

on R

atin

gs’ D

irect

; th

e re

port

incl

udes

sim

ilar

tabl

es f

or a

ll ra

ting

gro

ups

with

a n

on-z

ero

defa

ult

freq

uenc

y.

The

dat

a st

arts

fro

m 1

992

(and

late

r fo

r lo

wer

rat

ing

grou

ps)

sinc

e ea

rlier

the

re w

as n

o av

aila

ble

data

on

low

er-r

ated

sov

erei

gns.

The

firs

t co

lum

n re

pres

ents

the

act

ual

date

whe

n de

faul

t fr

eque

ncie

s ar

e co

mpu

ted,

whi

le t

he c

olum

n he

adin

gs r

epre

sent

the

tim

e ho

rizon

; e

g, f

or a

sov

erei

gn r

ated

BB

B in

201

0, t

he 3

-yea

r ah

ead

defa

ult

freq

uenc

y is

5.6

. I c

ompu

te th

e av

erag

es o

ver

the

15-y

ear

horiz

on in

ord

er t

o m

axim

ize

the

num

ber

of o

bser

vatio

ns; t

his

aver

age

is d

ispl

ayed

in t

he la

st c

olum

n.

Yea

r

No.

of

issu

ers

1 2

3 4

5 6

7 8

9 10

11

12

13

14

15

A

vera

ge

1992

2

0 0

0 0

0 0

0 0

0 0

0 0

0 0

0 0.

0

1993

6

0 0

0 0

0 0

16.7

16

.7

16.7

16

.7

16.7

16

.7

16.7

16

.7

16.7

10

.0

1994

8

0 0

0 0

0 12

.5

12.5

12

.5

12.5

12

.5

12.5

12

.5

12.5

12

.5

12.5

8.

3

1995

7

0 0

0 0

14.3

14

.3

14.3

14

.3

14.3

14

.3

14.3

14

.3

14.3

14

.3

14.3

10

.5

1996

4

0 0

0 25

25

25

25

25

25

25

25

25

25

25

25

20

.0

1997

9

0 0

11.1

11

.1

11.1

11

.1

11.1

11

.1

11.1

11

.1

11.1

11

.1

11.1

11

.1

11.1

9.

6

1998

16

0

0 0

0 0

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

12.5

4.

6

1999

16

0

0 0

0 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 12

.5

12.5

5.

4

2000

16

0

0 0

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

5.0

2001

16

0

0 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3 6.

3

5.4

2002

16

0

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

6.3

5.8

2003

13

0

0 0

0 0

0 0

0 0

0 0

0

0.0

2004

12

0

0 0

0 0

0 0

0 0

0 0

0.0

2005

13

0

0 0

0 0

0 0

0 0

0

0.0

2006

13

0

0 0

0 0

0 0

0 0

0.0

2007

13

0

0 0

0 0

0 0

0

0.0

2008

14

0

0 0

0 0

0 0

0.0

2009

17

0

0 0

0 0

0

0.0

2010

18

0

0 5.

6 5.

6 5.

6

3.

4

2011

19

0

0 0

0

0.0

2012

24

0

4.2

4.2

2.8

2013

23

0

0

0.0

2014

26

0

0.0

171

Chapter 5. Tightening credit standards: the importance of common trends

5.1. Introduction

Between 1980 and 2009, the average corporate credit rating for US non-financial companies has been steadily declining (see, e g, Alp, 2013). While this trend has been initially attributed to a deterioration in the rated issuers’ credit quality (e.g., Lucas and Lonski, 1992), more recent research suggests that this may be the outcome of increasing conservatism over time (Blume et al., 1998; Alp, 2013; Baghai et al., 2014). The increase in conservatism implies that CRAs assign lower ratings over time, for reasons unrelated to companies’ credit risk. Accordingly, a company with the same level of creditworthiness would receive a relatively lower rating today, relative to, say, 20 years ago.

For measuring conservatism, the aforementioned papers model ratings as a function of several company-specific variables, and interpret the unexplained time variation (captured by year fixed effects) as indicative of the level of conservatism in each period. While this variation is monotonically decreasing over time, interpreting this trend as conservatism seems far-fetched, from both an economic and an econometric perspective.

Firstly, none of the papers on conservatism suggests a plausible explanation for why conservatism increases over time. While CRAs might be more scrutinized today relative to 20 years ago, existing research on CRA incentives does not offer any arguments supporting a monotonic increase over time in conservatism. For example, Opp et al. (2013) show that the increasing prevalence of rating-contingent regulation over time, combined with increasing asset complexity, leads to an incentive to assign more lenient ratings. Sangiorgi and Spatt (2009) suggest that increased competition may lead to rating inflation, since the companies paying for their rating will choose

172

the best offer (the so-called “rating shopping”). Bar-Isaac and Shapiro (2013) show instead that rating quality is countercyclical, i.e., CRAs are more likely to issue less accurate ratings in good times, when fee-income is high and default probabilities are low.

Secondly, interpreting the unexplained time trend as a proxy for conservatism relies on the strong assumption that there is no confounding time variation from either omitted or existing variables. As long as the time trend is included in the same model with other variables with a similar time pattern, it will capture the time variation in the rating, which may be due to trend in both observed and omitted time-varying variables, making it impossible to disentangle conservatism from other time-varying factors.

In order to investigate whether the time pattern in ratings depends on pure conservatism or whether it mainly reflects the reaction of the CRAs to the existing trend in variables that affect firms’ credit risk, I revisit the most recent findings in Baghai et al. (2014), by using a two-step approach instead of the standard one-step fixed effects panel estimation. In the first step, I estimate a panel regression of rating on all the firm-specific and macroeconomic variables used in Baghai et al. (2014), but without including year effects. In the second step, I analyze the time pattern existing in the residual obtained in the first step. By orthogonalizing ratings to all the explanatory variables, I can separate the trend component related to creditworthiness from the one related to other unobservable factors, such as conservatism. Since the unexplained time trend in the credit rating may also depend on omitted variables, I also investigate whether adding other variables that have been changing over time, such as various measures of accounting quality (Givoly et al., 2017; Jorion et al., 2009) or macroeconomics (Baghai et al., 2014), can explain the remaining time trend.

I find that several of the existing variables, namely interest coverage, profitability, size, convertible debt, rent expense, PPE (plant, property and equipment) and CAPEX (capital expenditures) explain most of the variation in the time trend. More specifically, the remaining (unexplained) trend coefficient after controlling for the firm-specific variables is only 0.02, which is less than 20% of the original coefficient in, e g, Baghai et al. (2014), 0.12. Furthermore, I find that the included macroeconomic variables have a statistically significant, but modest, impact on the time trend (which decreases further to 0.016), while the proxies for accounting quality do not have any significant impact. Lastly, I propose an alternative approach to assess conservatism by testing whether the weights of the relevant accounting variables have been

173

changing over time by running yearly cross-sectional regressions (similarly to the first step in Fama and McBeth, 1973), and I find that the coefficients for most of the variables used in Baghai et al. (2014) do not exhibit any time trend. This finding doesn’t support the conservatism argument, since more conservative ratings would imply assigning a lower weight to “positive” factors (e.g., investment, size), and a relatively larger weight for the “negative” ones (e.g., leverage, risk).

While the remaining variation in the time trend may be indeed caused by increased conservatism, my findings suggest that other methods might be more suitable for assessing its true magnitude. Given that Blume et al. (1998), Alp (2013) and Baghai et al. (2014) have been cited very frequently and that their research was also used by regulators and policy-makers (Jorion et al., 2009), this can be an important contribution to the literature.

The remainder of this paper proceeds as follows. In Section 5.2, I present a more detailed literature review. In Section 5.3, I describe the dataset and variables. The empirical method and model are presented in Section 5.4, and the results are presented and discussed in 5.5. Finally, Section 5.6 provides a concluding discussion.

5.2. Literature review and conceptual framework

This paper builds mainly on the narrow research strand focused on the changes in rating standards over time, while also relating to the literature about the effect of the accounting and macroeconomic factors on credit risk and credit ratings.

By using a sample of investment-rated companies, Blume et al. (1998) are the first to suggest that the decrease in the average corporate ratings is not a mere reflection of an overall increase in companies’ credit risk, but is – at least partially – the result of a systematic change in rating standards. More specifically, by interpreting the year fixed effects in a panel ordered probit estimation as a proxy for conservatism, they find that a company with the same creditworthiness would be rated approximately a notch less (e.g., from AA+ to AA) in 1995 relative to 1978, ceteris paribus. The intuition behind their analysis is that, as long as all the relevant company and market-specific control variables are included and as long as their relative impact is stable over time,

174

the year indicators can be interpreted as a proxy for rating standards. Several years later, Alp (2013) replicates the model in Blume et al. (1998). She also finds that ratings have become more conservative for investment-grade rated companies. However, she finds the opposite pattern for speculative issuers, which is more difficult to explain – why would CRAs be more lenient with higher-risk companies? Lastly, Baghai et al. (2014) re-specifies the model in Alp (2013) by adding company fixed effects and using OLS instead of ordered probit.69 In contrast to Alp (2013), they find that both investment- and speculative-rated companies seem to receive more conservative ratings over time. Furthermore, they show that this increase in conservatism has a statistically and economically significant impact on capital structure, suggesting that CRA behavior has a real effect on companies’ long-term performance. While these three papers have a consistent overall result, which they interpret as an increase in conservatism, the validity of their conservatism measure relies on a strong and unrealistic assumption: that there are no omitted variables “hidden” in the year indicators (which thus correlate with both ratings and other company-specific variables). I explore, and test, this assumption, by extending the model in Baghai et al. (2014) with additional variables which, according to prior research, exhibit a secular time trend.

The first group of such variables are company-specific ratios (measuring either financial or market data). A well-known such variable is cash holdings. In an influential paper, Bates et al. (2009) attempt to explain the dramatic increase over time in cash-to-assets ratio for US industrial firms by showing that cash flows have become riskier, while companies have become increasingly R&D intensive over time. Another relevant empirical finding is that idiosyncratic risk has increased over time. Campbell et al. (2001), among others, attempt to identify the cause of this increase by using volatility decomposition models, and find that firm-level volatility accounts for the greatest proportion of the variation, while market volatility exhibits the strongest time trend. They show that the volatility measures are countercyclical and tend to lead GDP growth, but they emphasize that their results are descriptive rather than causal. By using a similar method, Ferreira and Gama (2005) instead show that industry risk has changed over time, with more assets needed to achieve the same level of

69 Using fixed effects with ordered probit is problematic due to the incidental parameter problem

(Lancaster, 2000). Since the interpretation of the year coefficients depends crucially on controlling for all relevant rating determinants, including company fixed effects is very important in this context (Baghai et al., 2014).

175

diversification. While it is difficult to disentangle empirically the exact causes for structural changes in accounting or market variables, previous research seems to agree on the fact that there is an overall increase in risk over time. In an influential keynote speech, Rajan (2006) suggests that technological advances in finance and increasing integration and sophistication of financial markets encourage risk-taking behavior, suggesting that, under certain conditions, companies and economics are more exposed to financial turmoil today than they were in the past. If this is true, then the decrease over time in average ratings may be a mere reflection of this increase in riskiness, which cannot be captured solely by a handful of financial ratios.

Even if the rating model did include all relevant variables (and hence capture their time variation), a further concern is whether the information content of those variables changed over time. Jorion et al. (2009) show that changes in accounting quality (measured by discretionary accruals) can explain much of the variation in the year indicators in Blume et al.’s (1998) rating model for investment-grade rated companies, but not for the speculative rated ones. Franzen et al. (2007) show that the increase in R&D intensity over time has altered the information content of several accounting variables that are commonly used in default prediction models. Givoly et al. (2017) focus on accounting measures relevant for debtholders, and show that they have actually become more informative over time. On the other hand, Beaver et al. (2005) show that changes in accounting standards, discretionary reporting, and changes in unrecognized/intangible assets over time affected only to a minor extent the predictive power of the bankruptcy model they used.

An overall change in riskiness could also be correlated with secular changes in macroeconomic variables. Givoly et al. (2017) show that reporting quality becomes more important in times when GDP growth is low. An extensive body of research has shown that credit risk is cyclical (e.g., Koopman et al., 2005; Allen and Saunders, 2002), which makes sense economically, since credit supply and demand rise in good times and shrink in recessions. At the same time, there are several macroeconomic variables that have been steadily increasing or decreasing over time. Inflation, for instance, has been decreasing since 1980. While at first glance this may sound like a positive development, lower inflation results in a lower cost of debt, which may trigger an increase in leverage. Furthermore, a similar type of argument can be made for interest rates, which have been decreasing since the 1980s. Aggregate measures such as money supply and loans outstanding, which exhibit strong secular trends,

176

have been included as growth rates in credit risk models (e.g., Koopman and Lucas, 2005). While this has been done in order to isolate secular from cyclical patterns, the trend in the year indicators may reflect the secular trends in these variables instead. CRAs claim to rate “through the cycle,” which implies that ratings should not be affected by short-term fluctuations in the economic environment since that would compromise rating stability (Standard and Poor’s, 2008). This means that ratings should be relatively insensitive to cyclical developments. At the same time, secular changes in ratings may co-vary with more long-term macroeconomic trends. Research on whether or not ratings are influenced by cyclical movements yielded mixed results. Ferri et al. (1999) show that sovereign ratings were excessively pro-cyclical during the Asian crisis, while Mora (2006) shows that this result disappears once a longer time period spanning the crisis is included in the analysis. When it comes to corporate ratings, most of the related research focuses on transition matrices. Nickel et al. (2000) and Bangia et al. (2002) show that rating transitions depend on the business cycle. In contrast, Amato and Furfine (2004), which is the only paper to include business cycle indicators in a rating model, find no evidence of excess pro-cyclicality.

5.3. Data

Data calculations and variable definitions follow Baghai et al. (2014), Jorion et al. (2009) and Givoly et al. (2017). The resulting sample is an unbalanced panel ranging from 1985 to 2016, and consisting of 2425 companies. Accounting variables are collected from Compustat’s annual database, while the market data are collected from CRSP. Macroeconomic data is collected from FRED, CBOE, the U.S. Bureau of Labor Statistics and Robert Shiller’s public database.70

5.3.1. Ratings

Rating data is collected from the monthly Compustat Ratings database. Since the analysis is performed on an annual basis, the end-of-year rating is

70 http://www.econ.yale.edu/~shiller/data.htm

177

selected.71 As in prior literature (e.g. Blume et al., 1998; Amato and Furfine, 2004), S&P domestic long-term issuer credit rating is chosen, and financial (SIC 6000-6999), utility (SIC 4900-4999) and governmental (SIC>9000) companies are excluded (Baghai et al., 2014).

To estimate the regression models, I translate the ratings into a numerical scale, starting with 1 for the highest rating (AAA) to 21 for C, meaning that the rating variable is positively correlated with an increase in default risk.72 Table 1 contains the average rating for each year, as well as the corresponding number of observations for the full sample, which sums up to 41,581 company-year rating observations.

Consistent with previous research, there is a decrease in the average ratings over time. The total decrease from 1985 until 2016 corresponds to about 2 notches, from BBB to BB+. However, during more recent years, which have not been included in previous research, there is virtually no change in the average rating.73 Since prior related research focused on the period before the 2009 financial crisis, it is interesting to investigate an extended timeframe.

A decrease in the average rating can be a consequence of many factors, such as a change in the average creditworthiness of rated companies, changes in the sample (in 1985, there are only 830 rated companies in the sample, whereas in 2016 the sample size increases to 1293), or, simply, an increasingly risky US market. If the downward trend in ratings is indeed caused by conservatism, one would expect an increasing number of downgrades over time and a decrease in the number of upgrades. Figure 1 displays the average frequency of downgrades (panel A) and upgrades (Panel B). In order to control for any sampling differences or rating conservatism for newly rated issuers, I estimate the frequencies for both the entire (unbalanced) sample, and for a balanced sample of 150 companies that have been rated between 1985 and 2016.

71 To avoid errors caused by differences between fiscal and calendar years, all the variables are

selected according to their calendar point in time.

72 The full rating scale with its corresponding numeric scale is displayed in Appendix 1.

73 Baghai et al. (2014) has the longest sample, from 1985 to 2009.

178

Table 1. Average and number of ratings by year This table contains the average and number of ratings for each year. The second column represents the rating group corresponding to the rounded numeric average in the first column. The last column refers to the selected sample of companies with available rating data.

Year Numeric average Rating group Rated companies

1985 9.000 BBB 830

1986 9.871 BBB- 981

1987 10.181 BBB- 1055

1988 10.282 BBB- 1039

1989 10.143 BBB- 998

1990 9.991 BBB- 935

1991 9.683 BBB- 888

1992 9.639 BBB- 920

1993 9.761 BBB- 1025

1994 9.904 BBB- 1072

1995 9.930 BBB- 1150

1996 10.080 BBB- 1271

1997 10.330 BBB- 1418

1998 10.545 BB+ 1604

1999 10.676 BB+ 1630

2000 10.847 BB+ 1638

2001 10.901 BB+ 1598

2002 11.077 BB+ 1565

2003 11.143 BB+ 1576

2004 11.207 BB+ 1587

2005 11.240 BB+ 1587

2006 11.328 BB+ 1544

2007 11.460 BB+ 1486

2008 11.467 BB+ 1378

2009 11.453 BB+ 1315

2010 11.357 BB+ 1355

2011 11.343 BB+ 1360

2012 11.348 BB+ 1355

2013 11.312 BB+ 1369

2014 11.314 BB+ 1395

2015 11.340 BB+ 1364

2016 11.311 BB+ 1293

Total 10.774 BBB- 41581

179

Figure 1. Average frequencies of rating changes over time Panel A shows the average frequencies of downgrades (relative to the total number of rated companies) for each year, while Panel B shows the corresponding frequency for upgrades. The continuous line takes into account all rated companies, while the dotted line represents the rating change frequencies for 150 companies with a rating for the entire period.

180

For both samples and rating actions, the time pattern is cyclical rather than secular, i.e., there is no monotonic increase or decrease over time. Instead, the frequency of downgrades peaks during the 2009 financial crisis (16%), while the number of upgrades is the lowest during the 2001 bubble (2%). Therefore, the upgrade-to-downgrade ratio has no discernible trend, and therefore does not support the idea of the increased conservatism.

5.3.2. Financial variables

As in Baghai et al. (2014), in my main specification, I use the following firm-specific variables: (1) EBITDA to interest expense (INT_COV);74 (2) profitability, measured as EBITDA/sales (PROF); (3) volatility of profitability, measured as the standard deviation of (2) over a 4-year period (VOL); (4) leverage measured as total debt to assets (LEV); (5) total debt relative to EBITDA (D_EBITDA); (6) dummy taking the value of 1 if EBITDA is negative and 0 otherwise (N_EBITDA); (7) convertible debt divided by total assets (CD); (8) log of real assets (SIZE); (9) cash divided by total assets (CASH); (10) Net Plant, Property and Equipment divided by assets (PPE); (11) rent expense relative to assets (rent); (12) Capital expenditures relative to assets (CAPEX); (13) beta, computed by using a market-model regression estimated using CRSP value-weighted index (BETA) and (14) idiosyncratic risk, estimated as the root mean squared error of the market-model regressions (ID_RISK). All variables are winsorized at 1% and 99%.

The first seven ratios measure three factors central to CRAs’ rating methodology: interest coverage, profitability and leverage (e.g., Standard and Poor’s, 2008). Furthermore, since the debt-to-EBITDA ratio has a non-monotonic relationship with default risk,75 I follow Baghai et al. (2014) by using a dummy variable for negative EBITDA (6), which should be positively correlated with credit risk. Size, on the other hand, has been shown to be the most significant explanatory variable for rating models, which is not surprising

74 To reduce the influence of outliers, Blume et al. (1998) and Alp (2012) create a piecewise

linear function from the raw interest coverage ratio. However, I follow Baghai et al. (2014) and instead use the winsorized version of the raw variable. In the first column of Appendix 3, I instead use the variable calculated as a piecewise linear function, but the results remain qualitatively similar.

75 When debt/EBITDA is positive, it also has a positive correlation with default risk; however, negative debt/EBITDA is associated with a higher default risk.

181

given that big companies are more stable and diversified (Shivdasani and Zenner, 2005). Cash reserves can indicate a stable company, as common intuition suggests, but they can also exist for precautionary reasons (Bates et al., 2009), and hence the relation between cash and ratings is not as straightforward. Variables 10, 11 and 12 measure asset tangibility, which indicates generally safer investments. Lastly, several papers (e.g., Hoberg and Prahbala, 2009) show that equity risk has changed over time. Since equity risk can be either systematic (market-specific) or idiosyncratic, both firm-specific betas and standard errors from market model regressions are included as control variables (13, 14).

Table 2 displays the yearly averages for these variables for the full sample of rated companies. Furthermore, in order to quantify the extent to which these variables exhibit a secular pattern, in the last row I compute the correlation between each specific variable and a linear trend (taking the value of 1 for 1985 up to 32 for year 2016).

The averages are very similar to Baghai et al. (2014), who briefly mention that several variables exhibit a visible trend over time. The last row suggests though that many of the variables are strongly correlated with the time trend: interest coverage, profitability, size, convertible debt, rental expense, and PPE all have correlation coefficients of over 80% (in absolute value). This suggests that there is a secular trend common to many of the variables, which might be captured by a time variable, as I will discuss in more detail in Section 5.5.1. Some of these trends have been well-documented in previous research, such as the increase in cash reserves (Bates et al., 2009), idiosyncratic risk (Campbell et al., 2001) and the decrease in tangible assets (Givoly et al., 2017).

182

Tab

le 2

. Y

ear

aver

ages

: B

agh

ai e

t al

. (2

014)

var

iab

les

The

tabl

e be

low

dis

play

s th

e ye

arly

ave

rage

s fo

r the

var

iabl

es u

sed

for

the

mai

n sp

ecifi

catio

n, a

s w

ell a

s th

eir c

orre

latio

n w

ith a

line

ar ti

me

tren

d (b

otto

m r

ow).

All

varia

bles

ar

e w

inso

rized

at

1% a

nd 9

9%, r

espe

ctiv

ely.

Var

iabl

e de

finiti

ons

can

be f

ound

in A

ppen

dix

2.

Yea

r IN

T_C

OV

P

RO

F

LEV

S

IZE

D

/EB

ITD

A

N.E

BIT

DA

VO

L C

AS

H

CD

R

EN

T

PP

E

CA

PE

X

BE

TA

ID

_RIS

K

1985

6.

381

0.13

2 0.

311

6.69

6 2.

883

0.04

1 0.

002

0.08

8 0.

050

0.02

4 0.

406

0.08

5 1.

380

0.68

9

1986

5.

763

0.12

9 0.

332

6.61

4 2.

381

0.07

4 0.

029

0.09

5 0.

057

0.02

4 0.

391

0.08

1 1.

507

0.69

3

1987

6.

220

0.14

8 0.

350

6.67

7 2.

823

0.04

7 0.

033

0.10

0 0.

068

0.02

4 0.

377

0.07

4 1.

382

0.70

6

1988

5.

884

0.15

4 0.

350

6.93

4 2.

810

0.04

0 0.

033

0.08

7 0.

059

0.02

4 0.

379

0.07

8 1.

623

0.62

8

1989

5.

354

0.15

1 0.

357

7.12

0 2.

719

0.04

3 0.

034

0.08

2 0.

053

0.02

3 0.

385

0.07

8 1.

622

0.59

4

1990

5.

543

0.15

4 0.

347

7.36

3 2.

952

0.02

9 0.

032

0.07

6 0.

045

0.02

4 0.

400

0.07

9 1.

468

0.57

6

1991

5.

213

0.14

6 0.

343

7.42

9 3.

041

0.03

5 0.

030

0.07

8 0.

043

0.02

4 0.

402

0.07

1 1.

402

0.55

2

1992

6.

009

0.14

4 0.

338

7.42

1 3.

096

0.03

3 0.

028

0.07

5 0.

045

0.02

7 0.

403

0.06

8 1.

345

0.49

3

1993

6.

746

0.15

1 0.

342

7.34

2 2.

962

0.03

1 0.

026

0.07

7 0.

045

0.02

6 0.

397

0.07

2 1.

291

0.52

4

1994

7.

491

0.15

6 0.

340

7.40

6 2.

719

0.04

4 0.

027

0.07

3 0.

042

0.02

4 0.

396

0.07

6 1.

295

0.51

8

1995

7.

348

0.16

3 0.

349

7.42

1 2.

769

0.03

7 0.

029

0.07

3 0.

034

0.02

2 0.

390

0.08

0 1.

041

0.51

7

1996

7.

503

0.16

5 0.

355

7.42

8 2.

691

0.04

2 0.

031

0.07

3 0.

031

0.02

3 0.

385

0.08

0 1.

058

0.54

1

1997

7.

965

0.17

2 0.

371

7.42

7 2.

928

0.04

3 0.

033

0.07

8 0.

033

0.02

3 0.

381

0.08

4 1.

255

0.59

1

1998

7.

056

0.15

2 0.

401

7.50

7 2.

924

0.05

7 0.

039

0.07

2 0.

030

0.02

3 0.

374

0.08

4 1.

161

0.62

6

1999

6.

477

0.16

2 0.

395

7.62

6 2.

792

0.06

4 0.

039

0.07

6 0.

028

0.02

2 0.

362

0.07

0 1.

117

0.63

8

2000

6.

849

0.16

2 0.

378

7.75

1 2.

290

0.07

6 0.

040

0.08

2 0.

032

0.02

2 0.

346

0.06

9 0.

844

0.67

0

2001

6.

571

0.16

2 0.

368

7.86

2 2.

926

0.05

7 0.

041

0.08

4 0.

031

0.02

2 0.

343

0.06

2 1.

006

0.61

7

2002

7.

267

0.16

3 0.

358

7.89

2 3.

004

0.04

9 0.

040

0.09

2 0.

034

0.02

4 0.

342

0.05

1 1.

268

0.67

1

183

Tab

le 2

. Y

ear

aver

ages

: B

agh

ai e

t al

. (2

014)

var

iab

les

(co

nti

nu

ed)

2003

8.

702

0.16

6 0.

339

7.98

9 2.

963

0.04

0 0.

037

0.10

3 0.

035

0.02

3 0.

330

0.04

8 1.

214

0.61

5

2004

11

.414

0.

179

0.31

8 8.

072

2.67

0 0.

019

0.03

6 0.

110

0.03

5 0.

022

0.31

8 0.

051

1.03

0 0.

625

2005

12

.684

0.

185

0.30

5 8.

149

2.64

9 0.

014

0.03

2 0.

107

0.02

7 0.

021

0.30

9 0.

056

1.09

9 0.

640

2006

12

.800

0.

191

0.31

1 8.

217

2.55

8 0.

010

0.03

0 0.

097

0.02

6 0.

021

0.31

0 0.

061

1.13

8 0.

669

2007

11

.166

0.

188

0.32

5 8.

353

2.59

7 0.

028

0.02

8 0.

093

0.02

3 0.

019

0.30

9 0.

062

1.16

2 0.

654

2008

10

.909

0.

175

0.35

1 8.

375

2.53

5 0.

046

0.03

5 0.

092

0.02

3 0.

021

0.32

1 0.

065

1.25

3 0.

748

2009

9.

235

0.17

1 0.

331

8.42

7 2.

997

0.04

7 0.

040

0.11

4 0.

021

0.02

1 0.

325

0.04

9 1.

201

0.67

8

2010

11

.694

0.

201

0.32

4 8.

487

2.80

0 0.

016

0.03

7 0.

117

0.01

8 0.

020

0.31

6 0.

048

1.14

4 0.

651

2011

12

.726

0.

204

0.32

7 8.

567

2.67

0 0.

019

0.03

6 0.

109

0.01

5 0.

019

0.31

1 0.

055

1.11

1 0.

669

2012

11

.420

0.

198

0.33

9 8.

611

2.95

6 0.

018

0.03

6 0.

106

0.01

2 0.

018

0.31

7 0.

062

1.15

9 0.

658

2013

10

.887

0.

201

0.34

5 8.

624

2.94

1 0.

023

0.03

7 0.

109

0.01

0 0.

017

0.32

3 0.

060

1.09

6 0.

626

2014

11

.236

0.

200

0.36

4 8.

620

3.35

9 0.

021

0.03

4 0.

103

0.00

9 0.

018

0.31

7 0.

060

1.01

1 0.

675

2015

9.

459

0.16

9 0.

394

8.63

5 3.

226

0.06

4 0.

048

0.10

0 0.

008

0.01

6 0.

318

0.05

5 1.

163

0.71

5

2016

8.

613

0.19

0 0.

401

8.68

3 3.

512

0.03

8 0.

048

0.10

2 0.

007

0.01

6 0.

316

0.04

8 1.

136

0.73

3

corr

. 0.

805

0.88

2 0.

121

0.96

2 0.

296

-0.3

31

0.59

5 0.

659

-0.9

50

-0.8

96

-0.9

05

-0.7

91

-0.6

50

0.45

9

184

5.3.3. Accounting quality

I follow two papers focusing on the time variation in accounting quality, Jorion et al. (2009) and Givoly et al. (2017), and collect several measures related to earnings management, accounting conservatism, uncertainty and fair value adjustments.

Jorion et al. (2009) find that discretionary accruals (DAC) have been increasing over time, particularly for investment-grade rated companies, and that this trend dominates the year coefficients in a replication of the rating model of Blume et al. (1998) for the time period 1985-2002.

First, I follow Jorion et al. (2009)’s method of calculating DAC as a modified version of the Jones (1991) model. First, total accruals (TAC) for each company i in year t is regressed against inverse assets, the change in revenues and the level of PPE of the other industry peers. More specifically, the model is run as a cross-sectional regression on a subsample of same two-digit SIC companies:

= 1 + ∆ + + (1a)

Where is total accruals (USD), defined as income before extraordinary items minus operating cash flow, ∆ represents the changes in net revenues, is net Property, Plant and Equipment, and is total assets in previous year.

The estimated regression coefficients, , and , are then used to calculate nondiscretionary accruals (NDAC) for each company i:

= 1 + ∆ −∆ + (1b)

Where ∆ represents the change in accounts receivables. Finally, discretionary accruals (DAC), are computed as the difference between total and non-discretionary accruals:

= − (1c)

185

DAC may be a noisy proxy for earnings management, since it cannot be observed or measured. Fields et al. (2001) write that “the only convincing conclusion appears to be that relying on existing accruals models to solve the problem of multiple method choices may result in serious inference problems.” Therefore, I collect additional variables capturing various aspects of accounting quality. To do this, I follow Givoly et al. (2017), who investigate whether changes in accounting quality lead to more informative financials from a debtholder perspective. The authors find frequency of losses, fair value measurement, intangible intensity and accounting conservatism to be trending over the last 20 years, and that this structural change improved the information value of bond-prediction models.

Second, frequency of losses can reflect increased riskiness, often being correlated with adverse changes in the business environment. Baghai et al. (2014) include a dummy for negative EBITDA, but Givoly et al. (2017) suggest that using operating cash flow is better, since it is not affected by accounting conservatism. Indeed, since the correlation between the two ratios is only 40%, they seem to convey different information, and therefore I include an indicator variable taking the value of 1 if the operating cash flow is negative:

= 1, ocf < 00, ocf ≥ 0 (2)

Third, in order to assess the impact of fair value accounting, I use the measure suggested by Demerijan (2011), which captures the volatility of book value of equity, relative to core net income.76 This is based on the fact that fair-value adjustments tend to appear in the equity section, rather than as part of income. The volatility ratio, as Demerijan (2011) defines it, can be written as:

= (∆ )( ) (3)

Where SD represents the 5-year standard deviation, ∆ is the change in retained earnings, DIV are total dividends and ANI is net income minus special and non-operating items. As Givoly et al. (2017) also argue, it is unclear from a debtor’s perspective if fair value accounting is a positive factor. On one hand, it may increase the information content of several financial variables, which

76 One of the company-specific variables in Baghai et al. (2014) is volatility of profitability.

However, this variable simply captures uncertainty in profitability, as opposed to fair value adjustments, as Demerijan’s (2011) ratio. The correlation between the two variables is only 7.4%.

186

should be valuable from a rating perspective. On the other hand, fair value adjustments result in unrealized gains or losses in shareholders’ equity, which are difficult to verify. Therefore, because of these offsetting effects, there is no specific directional hypothesis on the correlation between VR and ratings.

Fourth, I use intensity of investment in intangibles as a proxy for increased uncertainty. Since R&D is considered to be a higher risk investment, and since former research has shown that R&D expenses have been continuously increasing over time for US companies (e.g., Franzen et al., 2007), I expect this factor to have significant explanatory power in a rating model. Intensity of investment in intangibles (INI) is calculated as follows:

= ∗ . (4)

Where is intangible assets, and is the latest 5-year mean for R&D expenses deflated by total assets. It is multiplied by 2.5, since Lev and Sougiannis (2003) have documented that the average duration of the benefit period of R&D expenses is 2.5 years.

As one last measure, I estimate accounting conservatism (CONS) as the relative persistence of profits versus losses, following Ball and Shivakumar (2005). The coefficient of interest ∝ is estimated from the following piecewise linear regression: ∆ , =∝ +∝ ∆ , +∝ ∆ , +∝ ∆ , ∗ ∆ , + , (5)

Where ∆ is the change in net income from year t-1 to t, scaled by beginning of the year total assets, and ∆ represents an indicator variable set to one if ∆ in the previous year is negative and zero otherwise. The idea behind this model is that increased accounting conservatism involves recognizing economic losses in a timelier fashion relative to gains, meaning that ∝ <0. The regression is estimated on a yearly basis.

Table 3 displays the yearly averages for the accounting-quality variables, as well as their time trend correlation.

Although the overall averages are comparable with Jorion et al. (2009) and Givoly et al. (2017), unlike them I find that most of the variables do not exhibit a significant time trend.77 The only exceptions are intangible intensity and

77 Jorion et al. (2009) study a different time period, 1990-2002, while Givoly et al. (2017) use a

proprietary dataset of smaller companies that issue bonds (but are not necessarily public).

187

volatility ratio, which have a correlation of 98% and 79% with a linear time trend.

Table 3. Year averages: accounting quality The table below displays the yearly averages for the variables measuring various aspects of accounting quality, as well as their correlation with a linear time trend (bottom row). All variables are winsorized at 1% and 99%, respectively. Variable definitions can be found in Appendix 2.

Year DAC CONS VR NOCF INI

1985 0.012 0.203 2.112 0.003 0.100

1986 0.014 0.220 2.056 0.005 0.111

1987 0.013 0.190 1.872 0.009 0.116

1988 0.009 0.180 1.974 0.063 0.124

1989 0.008 0.510 1.966 0.067 0.138

1990 0.006 0.130 2.088 0.044 0.139

1991 0.012 0.140 2.027 0.046 0.145

1992 0.000 0.320 2.045 0.053 0.159

1993 0.008 0.002 1.944 0.077 0.165

1994 0.008 0.068 1.854 0.077 0.170

1995 0.000 0.056 1.926 0.068 0.170

1996 0.013 0.140 1.885 0.079 0.181

1997 0.015 0.028 1.994 0.079 0.185

1998 0.009 0.230 2.126 0.077 0.205

1999 0.011 0.350 2.144 0.074 0.215

2000 0.028 0.400 2.259 0.088 0.236

2001 0.021 0.270 2.198 0.065 0.248

2002 0.016 0.335 2.288 0.050 0.257

2003 0.014 0.250 2.239 0.053 0.261

2004 0.011 0.005 2.108 0.044 0.269

2005 0.015 0.100 1.958 0.035 0.273

2006 0.022 0.020 2.115 0.045 0.294

2007 0.010 0.430 2.356 0.024 0.307

2008 0.013 0.100 2.876 0.042 0.318

2009 0.015 0.140 2.978 0.022 0.318

2010 -0.009 0.620 2.826 0.027 0.326

2011 -0.005 0.170 2.852 0.032 0.333

2012 -0.014 0.410 2.806 0.036 0.347

2013 0.014 0.330 2.522 0.036 0.340

2014 0.015 0.070 2.522 0.042 0.355

2015 0.001 0.110 2.666 0.028 0.372

2016 0.009 0.220 2.623 0.031 0.376

corr. -0.134 0.311 0.790 -0.158 0.986

188

These trends are consistent with former research, documenting an increase in R&D expenses. This argument can also explain the increasing trend in the volatility ratio, which decreases the information content of certain financial ratios.

5.3.4. Macroeconomic variables

Macroeconomic factors may also explain much of the time trend in the year indicators. Baghai et al. (2014) are the only ones to acknowledge this and they perform a robustness test where they replace year indicators with a time trend, while adding several macro variables: (1) GDP growth, (2) inflation, (3) VIX, (4) aggregate PE, (5) term slope and (6) TED spread.

For comparability, I use the same macroeconomic variables as in Baghai et al. (2014). The real GDP growth (%), which is commonly used as a proxy for economic cycles (e.g., Carling et al., 2007; Feng et al., 2008), has been collected from the U.S. Bureau of Economic Analysis. Inflation rate is taken from the U.S. Bureau of Labor Statistics and is calculated from the All Urban Consumers Consumer Price Index; Gilchrist and Zakrajšek (2012), for example, show that inflation affects not only treasuries, but also corporate bond spreads. Further, I use two commonly used measures for macro-level credit risk: the TED spread and the term slope (e.g., Cornett et al., 2011), collected from the FED. The TED spread measures the difference between yields on corporate bonds and treasuries for the same maturity, while the term slope (or yield curve) represents the difference between long- and short-term treasury bond yields. The last two variables represent aggregate measures for US capital markets. The CBOE Volatility Index (VIX) is a proxy for near-term volatility, conveyed by S&P 500 stock index option prices. VIX has been shown to influence bond yields (Bao et al., 2011). The aggregate P/E ratio, collected from Robert Schiller’s public database, is another measure of market expectations/volatility, commonly used in the asset pricing literature (Bollerslev et al., 2011).78 Similarly with the previous sections, Table 4 displays the yearly averages for these variables and the time trend correlations. Here, the correlations are relatively low, with inflation having the highest coefficient of -0.56.

78 http://www.econ.yale.edu/~shiller/data.htm

189

Table 4. Year averages: Macroeconomics The table below displays the yearly averages for the macroeconomic variables, as well as their correlation with a linear time trend (bottom row). All variables are winsorized at 1% and 99%, respectively. Variable definitions can be found in Appendix 2.

Year GDP SPREAD SLOPE PE VIX INFL

1985 4.239 0.800 2.860 10.313 17.780 3.200

1986 3.512 0.900 0.610 14.250 18.710 2.018

1987 3.462 0.760 2.950 18.267 39.450 2.551

1988 4.204 1.320 3.040 14.313 18.530 3.501

1989 3.681 1.170 0.230 12.017 17.390 3.888

1990 1.919 0.800 0.910 14.865 25.360 3.699

1991 -0.074 1.100 1.640 15.253 20.910 3.329

1992 3.555 0.330 3.210 26.054 17.400 2.280

1993 2.746 0.360 2.870 22.799 12.420 2.379

1994 4.038 0.290 4.100 21.608 10.630 2.128

1995 2.719 0.810 0.870 15.204 11.960 2.086

1996 3.796 0.660 1.440 18.092 12.530 1.826

1997 4.487 0.510 1.320 19.784 19.470 1.712

1998 4.450 0.690 0.220 24.254 21.470 1.085

1999 4.685 0.690 1.300 33.115 26.250 1.530

2000 4.092 0.750 0.710 29.595 24.950 2.276

2001 0.976 0.680 -0.130 26.713 22.020 2.279

2002 1.786 0.150 2.960 46.181 21.090 1.535

2003 2.807 0.180 2.850 32.470 31.170 1.994

2004 3.786 0.240 3.390 23.236 16.630 2.750

2005 3.345 0.300 1.960 20.178 12.820 3.218

2006 2.667 0.450 0.550 18.286 12.950 3.072

2007 1.779 0.420 -0.350 17.472 10.420 2.661

2008 -0.292 1.500 0.920 20.833 26.200 1.962

2009 -2.776 1.330 3.130 58.171 44.840 0.759

2010 2.532 0.170 3.150 22.044 24.620 1.221

2011 1.601 0.160 2.640 16.582 19.530 2.065

2012 2.224 0.560 1.770 14.958 19.440 1.842

2013 1.677 0.220 2.280 21.900 14.280 1.615

2014 2.370 0.170 2.500 24.860 18.410 1.790

2015 2.596 0.240 2.110 26.490 20.970 1.076

2016 1.616 0.390 1.580 24.210 20.200 1.315

correlation -0.432 -0.435 0.025 0.356 0.017 -0.567

190

5.4. Empirical framework

The focus of my analysis is to investigate the extent to which the time pattern in ratings depends on pure conservatism. I suggest a two-step approach; the first step estimates a regression of rating on all the firm-specific and macroeconomic variables and the second step analyzes the time pattern existing in the residual obtained from the first step. The purpose is to filter out the trend component related to firms’ creditworthiness to uncover the “conservatism” component (if any).

Since conservatism may also mean that the weights of the factors used by CRAs have been changing over time, I propose an alternative method for gauging conservatism, by estimating 32 yearly cross-sectional regressions as in Fama and McBeth (1973). A consistent decrease over time in the coefficients for the favorable factors (e.g., size, profitability) and an increase for the risk factors (e.g., leverage, beta) would then support the notion of conservatism.

5.4.1. Two-step model

I follow Baghai et al. (2014) in using OLS rather than ordered probit (OP) estimation. While OP may be more appropriate given that ratings are ordinal, there are several advantages to using OLS. Firstly, using fixed effects can lead to biased coefficients in ordered probit (OP) estimation due to the incidental parameter problem (Lancaster, 2000). Since year fixed effects are the coefficients of interest, OLS is the preferred choice, allowing at the same time to control for unobservable company-specific variables (since CRAs emphasize on the importance of qualitative factors in rating assessment). Secondly, since I need to estimate the residuals from the first regression for the later part of my empirical analysis, the maximum-likelihood estimator (used for OP models) is neither straightforward nor directly interpretable (e.g., Greene and Hensher, 2010).

The starting point in my modelling is Baghai et al.’s (2014) model:

191

= + + , + + (6)

where is the rating, represent the year indicators, are n firm-specific rating determinants, and represent company fixed effects. The coefficient of interest, , indicates the difference between the intercept for year and the first year 1985. Since previous research has found that is monotonically decreasing over time, and firm-specific rating determinants seem to exhibit a similar trend, may simply be a parsimonious representation of this common trend. To investigate whether this is the case, I propose a testing strategy involving a two-step estimation. First, I re-estimate model (6) without the year indicators. Having included ,I control for the time pattern in them, while any remaining (unexplained) variation is captured by the residuals ′ : = ′ + ′ , + ′ + ′ (7)

In the second step, I model the residuals ′ as a function of the year indicators :

′ = + ′ + (8)

Where ′ captures any remaining trend in the year indicators.

If the year fixed effects in (6), , were a consequence of conservatism, and not just an artifact of the rating determinants exhibiting a common trend, then = ′ . The more of that is due to a common trend in the rating determinants, the smaller ′ would be relatively to . Therefore, the main goal of this empirical exercise is to test whether = ′ . Throughout the rest of the paper, I will further refer to as the “unadjusted” coefficient, and to ′ as the “adjusted” one. As I will show in Section 5.5.1. (Figure 2), the unadjusted year coefficients are monotonically increasing over time and are highly correlated with a linear time trend ((ρ =0.97). Hence, for the rest of my empirical analysis, I replace the year dummies with a trend variable T (taking the value of 1 for 1985 up to 32 for 2016). I do this firstly because having one single coefficient makes it easier to highlight the differences between (6) and (8), and secondly because it allows me to add company-invariant variables such as macroeconomics. While year dummies are

192

generally preferred since they allow a more flexible modelling of time effects, in this particular case the results would be very similar, since the year effects exhibit a linear trend.

5.4.2. Yearly cross-sectional regressions

This part of my empirical analysis is inspired by the cross-sectional two-step method suggested originally in Fama and MacBeth (1973), which has been extensively used in the analysis of the cross-section of stock returns. As stated by Fama and French (2002), the purpose of Fama MacBeth (FM) estimation is to correctly estimate standard errors in panels where autocorrelation (both spatial and time-specific) is present.

The traditional FM approach consists of two steps. In the first step, for each time period, coefficients of the parameters of interest are estimated by using cross-sectional regressions. In the second step, the time series of these coefficients are used to obtain final estimates, and standard errors so that correct t-statistics can be computed. As shown in Cochrane (2001), this approach is equivalent to OLS under the assumption that explanatory variables are stable through time.

I only use the first step of the FM approach, and my purpose is opposite to its “traditional” use: if the yearly estimates of rating determinants exhibit a significant time trend, this would be consistent with a change in rating standards.

More specifically, I propose an alternative way to define conservatism as a systematic change in the weighting of variables used for credit rating assessment. This implies that rating determinants that are positively correlated with credit risk (e.g., leverage, beta) would receive an increasingly higher weight over time, while the reverse would be true for factors negatively correlated with credit risk (e.g., investment, assets).

I thus estimate the following cross-sectional regression for each year: = + , + (9)Where , are the ratings for year t, and ∑ , represent the vector of coefficients for the rating determinants in Baghai et al. (2014) for each year.

193

The main goal of this empirical exercise is therefore to investigate whether the coefficients of each k variables , , … . exhibit a significant time trend, in a direction consistent with increasing conservatism. I first estimate the cross-sectional regressions using the full sample, and then I repeat the same estimations for investment and speculative subsamples.

5.5. Results

5.5.1. Baseline specification

As a baseline model, I replicate the main regression in Baghai et al. (2014) without year indicators, which I then use as an input for the second-stage estimation. For comparability, the first column in Table 5 displays the coefficients and significance levels taken directly from Baghai et al. (2014),79 while the following columns correspond to specifications (6), (7) and (8) described in the previous section.

The second column (model (6)) displays the coefficients of re-estimating the Baghai et al. (2014) model by using my sample. The year coefficients follow a similar time trend, and the significances, magnitudes for most control variables and the adjusted R2 (which is 0.90 in both estimations) are very similar to the original estimates in Baghai et al. (2014).

Table 5. Baseline model The columns correspond to the specifications (6), (7) and (8) described in Section 2.2. Therefore, the dependent variables are ratings for (6) and (7), and the residuals in (8). Variable descriptions can be found in Appendix 2. All models are estimated using OLS and robust standard errors (clustered at firm-level). All continuous variables are winsorized at 1% and 99%, respectively. *,**, and *** correspond to coefficients being significant at 10%, 5% and 1%, respectively. Models (6) and (7) include firm-fixed effects.

Year indicators Baghai et al.(2014) (6) (7) (8)

1986 0.228*** -0.427** -0.036

1987 0.437*** -0.303* -0.062

1988 0.474*** -0.081 0.135**

1989 0.497*** -0.025 0.072

1990 0.619*** 0.209 0.169**

1991 0.638*** 0.273* 0.165***

1992 0.656*** 0.462*** 0.329***

1993 0.730*** 0.490*** 0.246***

1994 0.876*** 0.662*** 0.347***

79 Baghai et al. (2014), model (6), Table III

194

Table 5. Baseline model (continued)

Variables Baghai et al.(2014) (6) (7) (8)

1996 1.084*** 0.890*** 0.345***

1997 1.159*** 0.823*** 0.207***

1998 1.209*** 0.879*** 0.111**

1999 1.347*** 1.036*** 0.067

2000 1.687*** 1.364*** 0.111***

2001 1.875*** 1.669*** 0.342***

2002 2.060*** 1.728*** 0.338***

2003 2.233*** 2.042*** 0.544***

2004 2.449*** 2.315*** 0.600***

2005 2.609*** 2.451*** 0.645***

2006 2.798*** 2.609*** 0.707***

2007 2.841*** 2.702*** 0.715***

2008 3.002*** 2.638*** 0.606***

2009 2.869*** 2.842*** 0.770***

2010 2.911*** 0.744***

2012 2.994*** 0.735***

2013 3.012*** 0.705***

2014 2.905*** 0.528***

2015 2.753*** 0.411***

2016 2.853*** 0.458***

INT_COV -0.017*** -0.016*** -0.011***

PROF -0.672** -0.326 -0.291

LEV 2.685*** 2.612*** 2.645***

SIZE -0.976*** -1.069*** -0.248***

D_EBITDA 0.040*** 0.037*** 0.029***

N_EBITDA 0.469*** 0.750*** 0.621***

VOL 0.635 -0.577 -0.138

CASH -0.028 -0.368 1.360***

CD 0.505 0.620 0.556

RENT 1.832 0.493 4.850

PPE -0.944*** -1.820*** -2.938***

CAPEX -3.910*** -3.981*** -5.133***

BETA 0.079** 0.138*** -0.100**

ID_RISK 1.165*** 1.899*** 2.489***

N 22705 20849 20849 20849

adj. R-sq 0.903 0.905 0.898 0.039

195

The third column (model (7)) includes the same variables, but without the year indicators. Here, the adjusted R2 is almost the same (0.89), suggesting that the year indicators do not add much explanatory power to the existing company-specific variables. However, most of the coefficients for the existing variables are now different (in particular for those exhibiting a strong time trend as shown in Table 1, such as size, cash, PPE, CAPEX and beta), suggesting that there is time heterogeneity which is not controlled for, since the year indicators are omitted. The question is if this heterogeneity comes from the existing variables or from conservatism.

I attempt to answer this question in model (8), where I use the saved residuals from (7) as a dependent variable in a univariate regression against the year indicators. As visible in the last column of Table 5, the year coefficients are much smaller – yet they are still slightly decreasing over time. For example, in the last year, 2016, the original year coefficient in (6) is 2.85, while the “adjusted” coefficient for the same year (in 8) is only 0.45. Assuming that the adjusted coefficient is not the result of other (trending) omitted variables,80 but rather the outcome of increasing conservatism, this would imply that the same company is rated half a notch lower in 2016 relative to 1985. By contrast, the unadjusted year coefficient for 2016, which is very similar to the one in Baghai et al. (2014), would suggest that a company would receive almost three notches less in 2016 compared to 1985.

Given that the year indicators in (6) and (8) are monotonically decreasing (as shown in Figure 2), I re-estimate the models by replacing the year indicators with a linear time trend (taking the value of 1 for the base year 1985, up to 33 in 2016). I do this for two reasons. First, one single coefficient is easier to interpret and the differences between the adjusted and the unadjusted trend will be easier to visualize and compare. Second, Baghai et al. (2014), Amato and Furfine (2004) and Jorion et al. (2009) use a linear trend variable in models including factors that only vary across time (such as macroeconomics), since having year fixed effects together with such variables would result in perfect multicollinearity.81 However, these papers could replace year indicators with a time trend precisely because the year indicators exhibit a secular pattern.

80 I will investigate this in the following parts of my analysis.

81 Year indicators can also be included with e.g. macroeconomic variables if two years are being dropped, but in this case the interpretation of the remaining year coefficients is less straightforward, while the coefficients might still be biased.

196

Figure 2. Year fixed effects vs. time trend The figure shows the year coefficients estimated in Table 5 relative to the trend line calculated by multiplying the linear trend value (marked with a dotted line) with its coefficient from Table 6 (e.g., for year 1988, the adjusted trend is 3*0.022). Trend (6) and FE (6) refer to the original values (first column of Table 6, and respectively 5), while Trend (8) and FE (8) refer to the adjusted values obtained from the second-step estimation in (8).

Table 6. Baseline model - linear trend The table re-estimates the regressions in Table 5 by replacing the year indicators with a linear time trend (ranging from 1 for 1985 to 32 for 2016). All models are estimated using OLS and robust standard errors (clustered at firm-level in (6) and (7)). All continuous variables are winsorized at 1%, respectively 99%. *,**, and *** correspond to coefficients being significant at 10%, 5% and 1%, respectively. Models (6) and (7) include firm-fixed effects.

Baghai et al. (2014) (6) (7) (8)

trend 0.129*** 0.115*** 0.022***

INT_COV -0.016*** -0.015*** -0.011***

PROF -0.776** -0.270 -0.364

LEV 2.505*** 2.165*** 2.685***

SIZE -0.949 -1.022*** -0.245***

D/EBITDA 0.039*** 0.035*** 0.024***

N.EBITDA 0.450*** 0.701*** 0.521**

VOL 0.569* -0.653 -0.114

CASH 0.242 -0.106 1.360***

CD 0.650*** 1.094** 0.579

RENT 1.596* 0.921 4.786

PPE -1.037*** -2.107*** -2.933***

CAPEX -3.960 -4.377*** -5.154***

BETA 0.084*** 0.107*** -0.100**

ID_RISK 1.199*** 2.035*** 2.498***

N 22228 20849 20849 20849

adj. R-sq 0.903 0.901 0.885 0.026

197

Table 6 displays the results for the same regressions as in Table 5, but with a time trend variable instead of year dummies. As in Table 5, the first column reproduces the coefficients from Baghai et al. (2014). My estimation again yields very similar results to their model. Concerning the rest of the columns (6), (7) and (8), the coefficients and R2 are very similar to the ones in Table 5, as expected. Most importantly, the adjusted time trend coefficient (from (8)) is approximately five times lower than the unadjusted one in (6).82 At the same time, it is statistically and economically significant, meaning that there are either (a) other omitted variables trending over time or (b) an increase in CRA conservatism.83

In the following parts of my analysis, I will attempt to address (a) by adding several variables measuring accounting quality or changes in the US macro environment, and (b) by proposing an alternative method to gauge conservatism.

5.5.2. Accounting quality

Since there is empirical evidence regarding secular changes in various aspects of accounting quality, such as earnings’ management, conservatism, fair value accounting or R&D intensity, I investigate whether these factors can explain the remaining time trend.

In Table 7, I replicate the models from Table 6 by successively adding different measures for accounting quality as described in 4.3. To conserve space, the variables in Baghai et al. (2014) are not reported in the regression output.

Even if all variables except DAC are significant, they do not influence the time trend coefficients. The adjusted trend coefficients in (8) range from 0.023 to 0.020, while the unadjusted ones in (6) have values between 0.116 and 0.103. As seen in Table 7, the adjusted trend is 0.022, whereas the unadjusted one is 0.129; the relative difference is therefore similar.

82 This is a similar relative change as for the year indicator coefficients

83 31*0.022= 0.68, which means more than half a notch.

198

Tab

le 7

. A

cco

un

tin

g q

ual

ity

T

he t

able

dis

play

s th

e re

gres

sion

out

put

for

estim

atio

ns (

6),

(7)

and

(8),

afte

r ad

ding

to

the

varia

bles

in

Bag

hai

et a

l.’s

(201

4) f

ive

diffe

rent

mea

sure

s fo

r ac

coun

ting

qual

ity. T

he c

oeffi

cien

ts in

Bag

hai e

t al.

(201

4) a

re s

imila

r to

the

one

s in

Tab

le 6

and

thus

not

rep

orte

d to

con

serv

e sp

ace.

The

last

thre

e es

timat

ions

, "al

l,” in

clud

e al

l the

fiv

e ad

ditio

nal v

aria

bles

in o

ne m

odel

. All

mod

els

are

estim

ated

usi

ng O

LS a

nd r

obus

t sta

ndar

d er

rors

(cl

uste

red

at fi

rm-le

vel i

n (6

) an

d (7

)).

All

cont

inuo

us v

aria

bles

are

w

inso

rized

at

1% a

nd 9

9%,

resp

ectiv

ely.

*,*

*, a

nd *

** c

orre

spon

d to

coe

ffici

ents

bei

ng s

igni

fican

t at

10%

, 5%

and

1%

, re

spec

tivel

y. M

odel

s (6

) an

d (7

) in

clud

e fir

m-f

ixed

ef

fect

s.

d

iscr

etio

nary

acc

rual

s

nega

tive

oper

atin

g ca

sh fl

ow

vola

tility

rat

io

(6

) (7

) (8

) (6

) (7

) (8

) (6

) (7

) (8

)

TR

EN

D

0.11

2***

0.02

3***

0.

114*

**

0.02

2***

0.

116*

**

0.

023*

**

DA

C

-0.2

92

-0.3

02

NO

CF

0.27

8***

0.

309*

**

VR

-0.0

02

0.01

9**

N

2005

2 20

052

2005

2

2084

9 20

849

2084

9

1972

9 19

729

1972

9

adj.

R-s

q 0.

904

0.89

0 0.

027

0.

901

0.88

6 0.

026

0.

901

0.88

5 0.

027

in

tang

ible

inte

nsity

ac

coun

ting

cons

erva

tism

a

ll

(6

) (7

) (8

) (6

) (7

) (8

) (6

) (7

) (8

)

TR

EN

D

0.10

3***

0.02

0***

0.

113*

**

0.02

2***

0.

103*

**

0.

021*

**

DA

C

-0

.403

-0

.292

NO

CF

0.24

5**

0.25

0**

VR

-0.0

03

0.01

5

INI

1.53

1***

2.

554*

**

1.

518*

**

2.58

3***

CO

NS

0.

233*

**

0.40

2**

*

0.

128*

**

0.24

0***

N

1145

5 11

455

1145

5

2084

6 20

846

2084

6

1061

3 10

613

1061

3

adj.

R-s

q 0.

915

0.90

4 0.

022

0.90

1 0.

886

0.02

5

0.91

9 0.

908

0.02

4

199

Even if the five variables measure accounting quality, they are very weakly correlated, and therefore I can include all of them in the same model, which is reported in the last three columns (all) of Table 7. Their joint influence on the adjusted trend coefficient is qualitatively similar to previous estimations.

This result might seem to be at odds with previous empirical findings. Jorion et al. (2009) find that discretionary accruals explain the yearly trend in ratings,

while Givoly et al. (2017) show that all their variables except negative OCF change monotonically over time.84 However, Jorion et al. (2009) extend their analysis only until 2002 and use a different rating model,85 while Givoly et al. (2017) use a proprietary sample of smaller companies, most of which are not rated.

In Appendix 3, I also use the variables used in Beaver et al. (2005) and adjust the existing variables for changes in R&D reporting as suggested in Franzen et al. (2007). However, the trend coefficients are very similar irrespective of the choice of variables or adjustments.

Therefore, time variation in accounting quality doesn’t explain the remaining time trend. In 5.3, I test another group of variables that can have an impact on the adjusted trend coefficient, measuring what CRAs call “country risk factors.”

5.5.3. Macroeconomic variables

Given that the empirical evidence on the impact of macroeconomic changes on credit ratings is mixed, I also investigate whether macroeconomic variables can explain the remaining time trend in ratings. Similar to the previous section, I first test the relative impact of each variable: GDP growth, inflation, TED spread, Term slope, VIX and PE ratio, and finally I include all of them. As in 5.2., I only report the output for specifications (6) and (8), and the coefficients from the variables in Baghai et al. (2014) are included in the estimation, but not reported. The output is displayed in Table 8.

84 Their result is mostly driven by investment-grade rated companies

85 Since they use ordered probit as in Blume et al. (1998), they can’t include company-fixed effects, and the year coefficients might be biased due to incidental parameter problems (Lancaster, 2000)

200

Tab

le 8

. Mac

roec

on

om

ics

The

tab

le d

ispl

ays

the

regr

essi

on o

utpu

t fo

r es

timat

ions

(6)

, (7

) an

d (8

), a

fter

addi

ng t

o th

e va

riabl

es in

Bag

hai e

t al

.’s (

2014

) si

x di

ffere

nt m

acro

econ

omic

fac

tors

. T

he

coef

ficie

nts

in B

agha

i et a

l. (2

014)

are

sim

ilar

to th

e on

es in

Tab

le 6

and

thus

not

rep

orte

d to

con

serv

e sp

ace.

The

last

thre

e es

timat

ions

, "al

l,” in

clud

e al

l the

six

add

ition

al

varia

bles

in o

ne m

odel

. A

ll m

odel

s ar

e es

timat

ed u

sing

OLS

and

rob

ust

stan

dard

err

ors

(clu

ster

ed a

t fir

m-le

vel i

n (6

) an

d (7

)).

All

cont

inuo

us v

aria

bles

are

win

soriz

ed a

t 1%

and

99%

, re

spec

tivel

y. *

,**,

and

***

cor

resp

ond

to c

oeff

icie

nts

bein

g si

gnifi

cant

at

10%

, 5%

and

1%

, re

spec

tivel

y. M

odel

s (6

) an

d (7

) in

clud

e fir

m-f

ixed

effe

cts.

G

DP

gro

wth

infla

tion

vol

atili

ty in

dex

P

/E r

atio

(6

) (7

) (8

)

(6)

(7)

(8)

(6)

(7)

(8)

(6

) (7

) (8

)

TR

EN

D

0.11

1***

0.02

1***

0.12

4***

0.

021*

**

0.11

4***

0.

022*

**

0.

115*

**

0.02

2***

GD

P

-0.0

46**

* -0

.102

***

INF

L

0.

171*

**

-0.0

72**

*

VIX

-0.0

02

-0.0

03**

PE

-0

.000

0.

003*

*

N

2084

9 20

849

2084

9

2084

9 20

849

2084

9

2084

6 20

846

2084

6

2084

9 20

849

2084

9

adj.

R-s

q 0.

901

0.88

7 0.

023

0.

901

0.88

6 0.

024

0.

901

0.88

5 0.

026

0.

901

0.88

5 0.

026

t

erm

slo

pe

T

ED

spr

ead

all

(6

) (7

) (8

)

(6)

(7)

(8)

(6)

(7)

(8)

TR

EN

D

0.11

4***

0.02

2***

0.11

5***

0.

021*

**

0.12

0***

0.

016*

**

GD

P

-0

.059

***

-0.2

19**

*

IN

FL

0.

223*

**

-0.0

97**

*

V

IX

-0

.001

-0

.006

***

PE

0.00

4 -0

.010

***

SLO

PE

0.

013*

0.

052*

**

0.

028*

* 0.

016

SP

RE

AD

0.

025

-0.2

33**

* -0

.072

**

-0.6

08**

*

N

20

849

2084

9 20

849

20

846

2084

6 20

846

20

846

2084

6 20

846

adj.

R-s

q 0.

901

0.88

6 0.

025

0.

901

0.88

6 0.

024

0.

902

0.89

1 0.

014

201

Individually, none of the macro factors has any significant impact on the trend, above and beyond the Baghai et al. (2014) variables. However, when all of them are included, the “adjusted” trend decreases further to 0.016, though still remains statistically significant. This value represents 13% (0.016/0.12) of the “unadjusted” trend coefficient, which represents a further decrease from the results in Table 6 and 7, where the smallest ratio of adjusted-to-unadjusted trend coefficients was 17% (0.022/0.13).

Most of the variables are significant in all models, although the ones having a trend component, primarily inflation and GDP growth, change their magnitude or sign. While it is impossible to attribute causality to a certain factor in this context, some of the variables are more likely to actually influence ratings. For example, GDP growth, which is the most popular proxy for business cycles in the credit risk literature (e.g., Amato and Furfine, 2004; Nickell et al., 2000; Bangia et al., 2002), has the same (expected) positive sign in the regression with and without trend, although the magnitudes of the coefficients are significantly different: GDP growth leads to an increase in ratings, ceteris paribus. This finding is consistent with a number of studies showing that there is a positive correlation between ratings and GDP growth (Bangia et al., 2002). Inflation has been trending downwards over the sample period, while ratings have become worse, and it therefore has a negative coefficient in the model without trend – suggesting that higher inflation leads to better ratings – which seems difficult to explain. However, once the trend is included in the model, the sign reverses, leading to a more plausible interpretation. At the same time, since there is a common trend in both inflation and ratings, it seems far-fetched to draw any causal conclusion, since inflation itself is the result of a variety of factors that are not controlled for in the model (e.g. fiscal policy, credit supply, etc.).

Lastly, I also test other macroeconomic variables following Koopman et al. (2009) and Figlevski et al. (2012): money supply growth rate, output gap, unemployment, consumer sentiment index, and S&P 500 return with similar results. The results are reported in Appendix 3.

5.5.4. Investment vs. speculative-grade rated companies

One of the most important divides in capital markets is the distinction drawn between investment- and speculative-grade ratings (e.g., Chernenko and Sunderam, 2012). A large number of regulations and investment restrictions

202

refer to this distinction; e g, many financial institutions cannot invest in speculative-grade rated bonds. Several papers suggest that this divide has real effects on capital structure (e.g., Kisgen and Strahan, 2010). Furthermore, certain CRAs have separate teams for investment- and speculative-grade issuers, such as Moody’s Leveraged Finance Group, which could create conflicts of interest when one of the teams loses its revenue from the issuer “crossing” the investment-speculative threshold.

Does this segmentation matter for the time variation in credit ratings? Empirical evidence provides mixed results. Alp (2013) finds that there is a significant difference between the two groups: there is a decreasing trend for investment-grade rated companies (which she interprets as conservatism), while the year trend for the speculative group has an opposite sign (which, according to Alp, suggests more lenient ratings). Baghai et al. (2014) instead show that, once you include company fixed effects this difference disappears, suggesting that all ratings have become more conservative. On the other hand, Jorion et al. (2009) find that earnings management (measured by DAC) has been increasing over time for investment-grade rated companies, but not for speculative-grade rated ones.

Given the inconclusive empirical evidence, I investigate whether there are significant differences in the adjusted time trend for investment- vs. speculative-grade rated firms. Table 9 displays the estimation output for the two groups for each set of variables: baseline, accounting and macroeconomic. To conserve space, I only report the results from the original specification (6) and from the second-stage trend regression (8). The coefficients of interest are the trend in (6) and the adjusted trend in (8) (adj_trend).

203

Table 9. Investment vs. Speculative The table displays the regression output for estimations (6) and (8) for the baseline model in Baghai et al. (2014), and the last models from Table 6 and 7 ("All"), for the subsamples of investment and speculative rated companies, respectively. All models are estimated using OLS and robust standard errors. All continuous variables are winsorized at 1% and 99%, respectively. *,**, and *** correspond to coefficients being significant at 10%, 5% and 1%, respectively. Estimation (6) includes firm-fixed effects. The last row, "Trend ratio,” displays the difference between the original trend in (6) and the adjusted one in (8)

Baseline (6) Accounting (6) Macros (6)

inv spec inv spec inv spec

Trend 0.093*** 0.045*** 0.082*** 0.047*** 0.095*** 0.044***

INT_COV -0.009*** -0.009*** -0.005** -0.008*** -0.009*** -0.010***

PROF -1.995*** 0.535* -4.400*** -0.784 -1.957*** 0.422

LEV 2.273*** 1.237*** 2.239*** 1.622*** 2.456*** 1.372***

SIZE -0.668*** -0.652*** -0.803*** -0.603*** -0.678*** -0.651***

D_EBITDA 0.022 0.030*** 0.003 0.023*** 0.016 0.025***

N_EBITDA -0.818* 0.973*** -0.932* 0.654** -0.832** 0.890***

VOL 1.179 -0.423 6.728*** 1.727 1.106 -0.357

CASH -0.552 -0.013 0.314 -0.087 -0.666 -0.087

CD 0.717 0.258 0.092 0.139 0.421 0.181

RENT 2.384 3.171** 2.029 5.346** 2.118 2.844*

PPE -2.830*** -0.078 -2.535*** -0.452 -2.808*** -0.049

CAPEX -5.037*** -2.732*** -5.825*** -3.384*** -4.636*** -2.657***

BETA 0.013 0.014 0.049 0.097** 0.010 -0.003

ID_RISK 1.572*** 1.585*** 1.793*** 1.577*** 1.620*** 1.554***

DAC -0.437* -0.188

CONS 0.134*** 0.030

VR -0.001 0.005

NOCF 0.203 0.110

INI 2.136*** 0.102

GDP -0.037*** -0.069***

INFL 0.141*** 0.116***

VIX -0.004* -0.001

PE 0.006** -0.004

SLOPE 0.020 0.047***

SPREAD -0.072** -0.017

N 10546 10303 6159 4454 10545 10301

adj. R2 0.831 0.743 0.844 0.748 0.833 0.746

Baseline (8) Accounting (8) Macros (8)

adj_trend 0.016*** 0.006*** 0.014*** 0.008*** 0.012*** 0.004***

N 10546 10303 6159 4454 10545 10301

adj. R-sq 0.021 0.004 0.015 0.006 0.012 0.002

Trend ratio 0.17 0.13 0.17 0.17 0.13 0.09

204

When comparing the unadjusted trend coefficients, as in Baghai et al. (2014), I find a positive value for both subsamples, suggesting that the there is a decreasing time trend for both investment- and speculative-grade ratings. My result is consistent with Alp (2013), since the coefficient for the speculative-grade subsample is roughly half that of the investment-grade one, suggesting that the trend (“conservatism”) is not as prominent for speculative issuers. Interestingly, this result remains very similar when including accounting quality or macroeconomic variables.

Turning instead to the adjusted trend (as a percentage of the unadjusted one, shown by the last row in Table 9), we see no significant difference between the investment and speculative groups. At most, the adjusted coefficient is 17% of the original trend coefficient for both the original variables in Baghai et al. (2014) and for the estimation including the five accounting quality measures. Although statistically significant, the economic magnitude of the adjusted trend is negligible. More specifically, if one would compare the rating differences between 1985 and 2016, an adjusted trend coefficient of 0.016 would translate to roughly half a notch (32*0.016=0.51).

Consistent with Section 5.5.3, the macroeconomic variables explain more of the original trend. Furthermore, the difference between investment- and speculative-grade rating is highest in this case: the adjusted trend coefficient is 13% of the original trend coefficient for the investment-grade subsample, while only 9% for the speculative-grade one. While this difference is consistent with the fact that distressed companies are more susceptible to changes in the macroeconomic environment (e.g., Standard and Poor’s, 2008), the magnitude is economically negligible.

To conclude, this part of my empirical analysis provides further evidence regarding why using a time trend/year indicators in a regression with trending variables can be misleading. In model (6), it seems like ratings have become more “conservative” for investment-grade than for speculative-grade rated firms, as suggested by Alp (2013). Given that CRAs have been criticized for being too lenient with risky assets (e.g., in the aftermath of the financial crisis for assigning high ratings to mortgage-backed securities), and while there is greater information asymmetry and risk of moral hazard in speculative-grade rated companies (Ashbaugh-Skaife et al., 2006), it is difficult to justify why CRAs would have become relatively more lenient towards them over the last four decades. My results instead suggest that the weaker trend for the speculative-grade subsample is a consequence of the model specification. For

205

example, the included variables may exhibit a weaker secular trend, which would be reflected in a lower coefficient for the time trend. On the other hand, given that the R2 is always lower for the speculative-grade subsample and that several explanatory variables are insignificant, other unobservable factors may be more relevant for speculative-grade issuers.

5.5.5. Yearly cross-sectional regressions

Using one model for the entire time period implies the rather strong assumption that the weights for the rating determinants do not change over time. However, if ratings have indeed become more conservative, this assumption seems implausible. More specifically, an increasingly conservative rating assessment would mean that factors that negatively affect creditworthiness, such as leverage or equity risk, are more important today relative to 20 years ago. Conversely, a company would need more cash, assets, or increased profitability to maintain the same rating in recent years, relative to the 1980s (and therefore the weights for these factors would be decreasing over time).

I test whether this is the case by running cross-sectional regressions for each year by using the original variables in Baghai et al. (2014), as in specification (9). Conservatism would then be consistent with an increase over time in the coefficients for leverage, debt/EBITDA, negative EBITDA, convertible debt, rent expenses, volatility, beta and idiosyncratic risk; and with a decreasing trend in the coefficients for interest coverage, profitability, size, cash, tangibility and capex.

Furthermore, I investigate whether there are differences in the coefficient time variation for investment-, relative to speculative-grade rated companies. If there is an increased conservatism for investment-grade rated companies, then the time trend of the individual coefficients would be more pronounced for this group.

Figure 3 displays the yearly coefficients for each of the rating determinants over time.

206

Figure 3. Time variation of yearly regressions’ coefficients The figure displays the coefficients over time for the yearly cross-sectional regressions, for each of the variables from the baseline model in Baghai et al.(2014). The coefficients for the entire sample are plotted in orange triangles, the ones for the investment group are marked by green squares and the speculative group by red circles.

207

Figure 3. Time variation of yearly regressions’ coefficients (continued) The figure displays the coefficients over time for the yearly cross-sectional regressions, for each of the variables from the baseline model in Baghai et al(2014). The coefficients for the entire sample are plotted in orange triangles, the ones for the investment group are marked by green squares and the speculative group by red circles.

208

The triangle-shaped orange line displays the coefficients for the regressions using the entire sample, while the square-shaped green line marks the coefficients for the investment subsample and the circle-shaped red line for the speculative one. The only coefficient with a clear trend for the three samples is leverage. However, since the coefficient is decreasing over time, it rather suggests more lenient ratings.

For ease of interpretation, in Table 10 I summarize the expected, as well as the actual trends displayed graphically in Figure 3. The letters in bold emphasize the coefficients which move in a direction consistent with increased conservatism.

Table 10. Summary of coefficients’ trends The table summarizes the overall trend patterns for the coefficients of the yearly regressions. “S” means stable, i.e. no visible monotonous trend, “N” stands for negative trend, and “P” for positive. The letters marked in bold font are the ones consistent with the notion of conservatism. ALL represents the entire sample, INV stands for investment and SPEC for speculative. EXP shows the expected direction of the trend, consistent with increasing conservatism.

ALL INV SPEC EXP

INT_COV S S S N

PROF S P S N

LEV N N N P

SIZE P S S N

D_EBITDA S S S P

N_EBITDA P S P P

VOL S N S P

CASH S S S N

CD P S S P

RENT S S P P

PPE S N S N

CAPEX P P S N

BETA P S S P

ID_RISK S S S P

For the full sample, only three variables, convertible debt, negative EBITDA and beta have a trend that is also consistent with conservatism. Leverage, size, PPE and CAPEX have trends suggesting more lenient rating standards, while the rest of the variables have no discernible time trend.

Concerning the coefficients for investment vs. speculative-grade rated companies, most of the variables are highly correlated and therefore moving in the same direction. None of the coefficients for the investment subsample

209

move in a direction consistent with conservatism, while for the speculative group, only negative EBITDA and rent have a positive trend over time. If ratings were indeed more conservative for the investment-grade rated issuers as suggested by Alp (2013) and by the unadjusted trend coefficient in the previous section, one would expect different time patterns than the ones shown in the third row of Table 10.

One concern regarding this approach is whether the variables are equally relevant throughout time. However, as seen in the last row, R2 is slightly increasing over time, in particular for the speculative group. This result is consistent with the overall finding in Givoly et al. (2017), who suggest that accounting ratios have become more informative to bondholders over time.

Furthermore, these results might be driven by changes in industry composition, or by different rating standards for newly rated companies. To address this, I select a subsample of 120 companies that had a rating between 1985 and 2016. While these averages might be a better reflection of potential conservatism (if any), they are also more sensitive to outliers due to the small sample size. Since the time pattern is very similar to the full sample, I choose not to include the coefficients for this subsample in Figure 3.

This approach is inspired by Fama and McBeth (1973) and used commonly in the asset pricing literature. In fact, Blume et al. (1998) and Alp (2013) use yearly regressions to show that the coefficients do not change significantly over time, and my findings are consistent with that. However, these papers do not attempt to interpret the differences in the year coefficients, but rather see whether their R2 and the statistical significance of the individual coefficients remain stable through time. By contrast, I instead focus on the time pattern of the coefficient values, as I consider it to be a plausible alternative interpretation of conservatism.

5.6. Conclusion

By using time fixed effects as a proxy for CRA rating standards, Blume et al. (1998), Alp (2013) and Baghai et al. (2014) show that ratings have become more conservative over time. According to their definition, a company with the same creditworthiness would be granted a lower rating today compared to 20 years ago. However, existing rating theories and models do not offer a

210

plausible explanation for a systematic downward bias in the rating process; if anything, rating conservatism should instead be affected by business cycles (e.g., Bar-Isaac and Shapiro, 2013).

Given that the year indicators in Blume et al. (1998), Alp (2013) and Baghai et al. (2014) exhibit a monotonic downward trend, I propose a two-step regression approach for testing whether the year trend can be attributed to conservatism, or whether it is simply capturing a common, secular trend in the existing or omitted variables.

To do this, I first replicate the baseline model in Baghai et al. (2014), and I get similar results over the period 1986-2016, suggesting that a company would receive a 3-notch lower rating (i.e. from AAA to AA-) relative to 30 years ago, ceteris paribus. Second, I re-estimate the same model but without the year trend, and I save the residuals. Third, I model the residuals as a function of the time trend. The resulting (“adjusted”) trend coefficient therefore captures the remaining variation after “weeding out” the impact of the rating variables used in Baghai et al. (2014). I find that, as expected, the “adjusted trend” is only 20% of the original one. From a rating perspective, this means less than half a notch lower – as opposed to the three notches found in Baghai et al. (2014).

Since the “adjusted” trend could still reflect the time variation in certain omitted variables, I test this assumption by adding a range of control variables that prior research (1) found relevant for explaining credit risk and (2) found to exhibit a secular time trend. First, I turn to the accounting literature and include proxies for accounting quality, fair disclosure, R&D intensity and conservatism, and I find that they do not have a significant impact on the remaining (adjusted) year trend. Second, I test a range of macroeconomic variables measuring volatility, business and credit cycles, stock market performance, and interest rates, and I find that these have a statistically significant, but economically modest impact on the adjusted trend (i.e. the adjusted trend reduces further to approximately 13% of the original one).

In contrast to previous studies, I propose a new method for gauging conservatism by testing whether the weights for the rating variables change over time. More specifically, if CRAs had become more conservative, “risk” factors such as leverage or interest coverage would become increasingly important over time, while “good” factors such as assets, sales, profitability would be assigned a lower weight in the more recent time period. I find that most rating variables do not exhibit any visible time trend in their assigned

211

weights (i.e. the coefficients from the yearly cross-sectional regressions), and therefore, from this perspective, there is no evidence of increased conservatism.

Finally, while the current paper cannot disprove the existence of rating conservatism, it suggests that interpreting a time trend as such is problematic since it relies on very strong assumptions. Given the increasing criticism and regulatory pressure faced by CRAs, a more accurate interpretation/estimation has to be done before taking the increased conservatism as given and as proof of “systematic misconduct” (Jorion et al., 2009).

212

References

Ashbaugh-Skaife, H., Collins, D. W., & LaFond, R. (2006). The effects of corporate governance on firms’ credit ratings. Journal of Accounting and Economics, 42(1-2), 203-243.

Allen, L., & Saunders, A. (2002). A survey of cyclical effects in credit risk measurement models. Working Paper.

Alp, A. (2013). Structural shifts in credit rating standards. The Journal of Finance, 68(6), 2435-2470.

Amato, J. D., & Furfine, C. H. (2004). Are credit ratings procyclical? Journal of Banking & Finance, 28(11), 2641-2677.

Baghai, R. P., Servaes, H., & Tamayo, A. (2014). Have rating agencies become more conservative? Implications for capital structure and debt pricing. The Journal of Finance, 69(5), 1961-2005.

Ball, R., & Shivakumar, L. (2006). The role of accruals in asymmetrically timely gain and loss recognition. Journal of Accounting Research, 44(2), 207-242.

Bangia, A., Diebold, F. X., Kronimus, A., Schagen, C., & Schuermann, T. (2002). Ratings migration and the business cycle, with application to credit portfolio stress testing. Journal of Banking & Finance, 26(2), 445-474.

Bao, J., Pan, J., & Wang, J. (2011). The illiquidity of corporate bonds. The Journal of Finance, 66(3), 911-946.

Bar-Isaac, H., & Shapiro, J. (2013). Ratings quality over the business cycle. Journal of Financial Economics, 108(1), 62-78.

Bates, T. W., Kahle, K. M., & Stulz, R. M. (2009). Why do US firms hold so much more cash than they used to? The Journal of Finance, 64(5), 1985-2021.

Beaver, W. H., McNichols, M. F., & Rhie, J. W. (2005). Have financial statements become less informative? Evidence from the ability of financial ratios to predict bankruptcy. Review of Accounting Studies, 10(1), 93-122.

Blume, M. E., Lim, F., & MacKinlay, A. C. (1998). The declining credit quality of US corporate debt: Myth or reality? The Journal of Finance, 53(4), 1389-1413.

Bongaerts, D., Cremers, K. J., & Goetzmann, W. N. (2012). Tiebreaker: Certification and multiple credit ratings. The Journal of Finance, 67(1), 113-152.

Bollerslev, T., Gibson, M., & Zhou, H. (2011). Dynamic estimation of volatility risk premia and investor risk aversion from option-implied and realized volatilities. Journal of Econometrics, 160(1), 235-245.

Campbell, J. Y., Lettau, M., Malkiel, B. G., & Xu, Y. (2001). Have individual stocks become more volatile? An empirical exploration of idiosyncratic risk. The Journal of Finance, 56(1), 1-43.

213

Carling, K., Jacobson, T., Lindé, J., & Roszbach, K. (2007). Corporate credit risk modeling and the macroeconomy. Journal of Banking & Finance, 31(3), 845-868.

Chernenko, S., & Sunderam, A. (2011). The real consequences of market segmentation. The Review of Financial Studies, 25(7), 2041-2069.

Cochrane, J. H. (2001). A rehabilitation of stochastic discount factor methodology (No. w8533). National Bureau of Economic Research.

Cornett, M. M., McNutt, J. J., Strahan, P. E., & Tehranian, H. (2011). Liquidity risk management and credit supply in the financial crisis. Journal of Financial Economics, 101(2), 297-312.

Dechow, P. M., Sloan, R. G., & Sweeney, A. P. (1995). Detecting earnings management. Accounting Review, 193-225.

Demerjian, P. R. (2011). Accounting standards and debt covenants: Has the “balance sheet approach” led to a decline in the use of balance sheet covenants? Journal of Accounting and Economics, 52(2-3), 178-202.

Duffie, D., Saita, L., & Wang, K. (2007). Multi-period corporate default prediction with stochastic covariates. Journal of Financial Economics, 83(3), 635-665.

Fama, E. F., & French, K. R. (2002). The equity premium. The Journal of Finance, 57(2), 637-659.

Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal of Political Economy, 81(3), 607-636.

Ferri, G., Liu, L. G., & Stiglitz, J. E. (1999). The procyclical role of rating agencies: Evidence from the East Asian crisis. Economic Notes, 28(3), 335-355.

Ferreira, M. A., & Gama, P. M. (2005). Have world, country, and industry risks changed over time? An investigation of the volatility of developed stock markets. Journal of Financial and Quantitative Analysis, 40(1), 195-222.

Feng, D., Gourieroux, C., & Jasiak, J. (2008). The ordered qualitative model for credit rating transitions. Journal of Empirical Finance, 15(1), 111-130.

Fields, T. D., Lys, T. Z., & Vincent, L. (2001). Empirical research on accounting choice. Journal of Accounting and Economics, 31(1-3), 255-307.

Figlewski, S., Frydman, H., & Liang, W. (2012). Modeling the effect of macroeconomic factors on corporate default and credit rating transitions. International Review of Economics & Finance, 21(1), 87-105.

Franzen, L. A., Rodgers, K. J., & Simin, T. T. (2007). Measuring distress risk: The effect of R&D intensity. The Journal of Finance, 62(6), 2931-2967.

Greene, W. H., & Hensher, D. A. (2010). Modeling ordered choices: A primer. Cambridge University Press.

Gilchrist, S., & Zakrajšek, E. (2012). Credit spreads and business cycle fluctuations. The American Economic Review, 102(4), 1692-1720.

214

Givoly, D., Hayn, C., & Katz, S. (2017). The changing relevance of accounting information to debt holders over time. Review of Accounting Studies, 22(1), 64-108.

Hoberg, G., & Prabhala, N. R. (2008). Disappearing dividends, catering, and risk. The Review of Financial Studies, 22(1), 79-116.

Jones, J. J. (1991). Earnings management during import relief investigations. Journal of Accounting Research, 193-228.

Jorion, P., Shi, C., & Zhang, S. (2009). Tightening credit standards: the role of accounting quality. Review of Accounting Studies, 14(1), 123-160.

Kasznik, R. (1999). On the association between voluntary disclosure and earnings management. Journal of Accounting Research, 37(1), 57-81.

Kisgen, D. J., & Strahan, P. E. (2010). Do regulations based on credit ratings affect a firm's cost of capital? The Review of Financial Studies, 23(12), 4324-4347.

Koopman, S. J., Kräussl, R., Lucas, A., & Monteiro, A. B. (2009). Credit cycles and macro fundamentals. Journal of Empirical Finance, 16(1), 42-54.

Koopman, S. J., & Lucas, A. (2005). Business and default cycles for credit risk. Journal of Applied Econometrics, 20(2), 311-323.

Lancaster, T. (2000). The incidental parameter problem since 1948. Journal of Econometrics, 95(2), 391-413.

Lev, B., & Sougiannis, T. (2003). The capitalization, amortization, and value-relevance of R&D. Intangibles: Management, Measurement, and Reporting, 123-152.

Lucas, D. J., & Lonski, J. G. (1992). Changes in corporate credit quality 1970–1990. The Journal of Fixed Income, 1(4), 7-14.

Mora, N. (2006). Sovereign credit ratings: Guilty beyond reasonable doubt? Journal of Banking & Finance, 30(7), 2041-2062.

Nelson, C. R., & Plosser, C. R. (1982). Trends and random walks in macroeconmic time series: some evidence and implications. Journal of Monetary Economics, 10(2), 139-162.

Nickell, P., Perraudin, W., & Varotto, S. (2000). Stability of rating transitions. Journal of Banking & Finance, 24(1-2), 203-227.

Opp, C. C., Opp, M. M., & Harris, M. (2013). Rating agencies in the face of regulation. Journal of Financial Economics, 108(1), 46-61.

Partnoy, Frank (2009). Overdependence on credit ratings was a primary cause of the crisis. Working Paper.

Rajan, R. G. (2006). Has finance made the world riskier? European Financial Management, 12(4), 499-533.

215

Sangiorgi, F., Sokobin, J., & Spatt, C. (2009). Credit-rating shopping, selection and the equilibrium structure of ratings. Working Paper, Stockholm School of Economics and Carnegie Mellon University.

Shivdasani, A.& Zenner M., (2002). Best Practice in Corporate governance: What Two Decades of Research Reveals, New York: Salomon Smith Barney.

Skreta, V., & Veldkamp, L. (2009). Ratings shopping and asset complexity: A theory of ratings inflation. Journal of Monetary Economics, 56(5), 678-695.

Standard and Poor's, (2008). Corporate Ratings Criteria. RatingsDirect.

216

Appendices

Appendix 1. Rating scale The appendix displays the distribution and coding of the rating variable. The first row corresponds to the original values of the long-term issuer rating variable from Compustat. Code represents the corresponding numeric value (as in Baghai et al., 2014). The last column displays the total firm-year observations for each rating.

Rating Code Count

AAA 1 594

AA+ 2 217

AA 3 868

AA- 4 954

A+ 5 1651

A 6 2533

A- 7 2174

BBB+ 8 2933

BBB 9 3758

BBB- 10 3020

BB+ 11 2393

BB 12 3381

BB- 13 4515

B+ 14 5793

B 15 3616

B- 16 1759

CCC+ 17 754

CCC 18 376

CCC- 19 158

CC 20 133

C 21 6

Total 10.77418 41586

217

Appendix 2. Variable description

Baghai et al. (2014)

Rating S&P's Issuer long-term rating (splticrm) converted into numbers as in Baghai et al. (2014): 1 for AAA, to 21 for C and below

Interest coverage (INT_COV)

Operating income after depreciation (oiadp) plus interest expense (xint) divided by interest expense (xint). In the Alp(2013) regression model, I use the modified piecewise linear function on the 3-year averages, covering the incremental value of interest coverage in the intervals of (0–5), (5–10), (10–20), and (20-100)

Profitability (PROF) Operating income before depreciation (oibdp) to sales (sale)

Leverage (LEV) Long-term debt (dltt) plus short-term debt (dlc), divided by assets (at).

Idiosyncratic risk (ID_RISK)

The root mean squared error from a regression of the daily stock returns on the CRSP value-weighted index return; in the regression, it is standardized by dividing it by the annual cross-sectional mean.

Beta (BETA) Market model beta estimated from the same regression used to define idiosyncratic risk; similarly to the idiosyncratic risk, it is standardized by dividing it by the annual cross-sectional mean.

Capital expenditures (CAPEX)

Capital expenditures (capx) to assets (at).

Plant, property and equipment (PPE)

Property, plant, and equipment - total (ppent) to assets (at).

Cash (CASH) Cash and short-term investments (che) to assets (at).

Debt/EBITDA (D_EBITDA)

Long-term debt (dltt) plus short-term debt (dlc), divided by earnings before taxes, depreciation and amortization(ebitda).

Negative EBITDA (N_EBITDA)

Dummy taking value of 1 if ebitda is negative and 0 otherwise

Volatility (VOL) Volatility of profitability, computed over the last 5 years

convertible debt (CD) convertible debt (dcvt) to assets(at)

rent expense (RENT) rental expense (xrent) to assets(at)

Size (SIZE) Log of real total assets (at)

Accounting quality

discretionary accruals (DAC)

Proxy for earnings management, calculating using a modified version of the Jones (1991) model. DAC is calculated cross-sectionally for each year and industry, and is a function of "abnormal" reported changes in sales and PPE

frequency of losses (NOCF)

Indicator variable taking the value of 1 if operating cash flow (oaindp) is 1

fair value accounting (VR) Volatility of book value of equity relative to core net income (codes used: reta, dvc, ni)

intangible intensity (INI) Intangibles relative to unrecognized R&D expense

accounting conservatism (CONS)

Persistence of profits relative to losses

Macroeconomic

GDP growth (GDP) Real GDP growth (% per annum, Bureau of Economic Analysis)

VIX (VIX) Annual average of the market volatility index (CBOE)

TED spread (SPREAD) 3-month LIBOR (Bank of England) minus the 3-month T-bill (%)

term slope (SLOPE) 10-year T- bond minus the 3-month T-bill (%, FRED)

P/E ratio (PE) Aggregate price-to-earnings ratio based on previous years' earnings (Robert Shiller’s Online Database)

Inflation (INFL) Inflation rate (%, Bureau of Labor Statistics CPI - All urban consumers)

218

Appendix 3. Robustness: additional variables and ordered probit The table compares the original trend (in (6)) with the adjusted one (adj_trend) estimated in the second step regression(8), after controlling for a range of additional variables or adjusting the existing ones for accounting biases. In (1), instead of the interest coverage variable used in the main specification, I instead use the piecewise linear variable as in Alp (2013), since this would control for the impact of outliers and for the non-monotonous relationship with credit risk. Furthermore, I replace the accounting-trending variable with alternative aggregate/ frequency measures as in Beaver et al. (2012): losses, R&D expenses higher than 1% of sales, mis-valuation (market-to-book ratios deviating significantly from one) and abnormally high accruals (higher than 1% of sales). In (2), I adjust leverage, size, debt/ebitda, cash, tangibility and capex for R&D intensity following Franzen et al. (2007). As in Table 9, the last four rows refer to the second-stage estimation, with “adj_trend” being the estimated second-stage trend coefficient. Lastly, trend ratio is the ratio between the adjusted and the original trend. All models are estimated using OLS with firm-fixed effects and robust standard errors. All continuous variables are winsorized at 1% and 99%,. *,**, and *** correspond to coefficients being significant at 10%, 5% and 1%, respectively.

(1) (2) (3) Trend 0.109*** 0.106*** 0.128*** INT_COV -0.009*** -0.016*** INT_COV1 0.122***

INT_COV2 -0.104***

INT_COV3 -0.016

INT_COV4 0.001

PROF -3.299*** -3.151*** -0.397 LEV 2.832*** 3.017*** 2.423*** SIZE -0.998*** -1.037*** -1.042*** D_EBITDA 0.024*** 0.018*** 0.029*** N_EBITDA 0.023 0.319 0.613*** VOL 6.884*** -0.616 CASH -0.230 2.521 -0.265 CD 1.113* 0.978* 0.915** RENT -0.032 -0.420 0.369 PPE -2.984*** -3.025*** -1.994*** CAPEX -5.233*** -6.884*** -4.226*** BETA 0.187*** 0.194*** 0.102** ID_RISK 2.095*** 2.101*** 1.968*** LOSS_FR 0.138*** RD_FR -0.533*** MTB_FR 0.108* DAC_FR 0.009 GDP GAP 0.177*** S&P RETURN 0.404*** M2 GROWTH -0.619 UNEMPLOYMENT 0.127*** CONSUMER SENTIMENT -0.025*** N 11720 11117 20849 adj. R2 0.913 0.413 0.903

adj_trend 0.019*** 0.022*** 0.017***

N 11720 11117 20849

adj. R-sq 0.021 0.025 0.015

Trend ratio 0.17 0.20 0.13

AN

AM

AR

IA C

OC

IOR

VA

Essays on C

redit Ratings

Department of Business AdministrationLund University, School of Economics,

Business and Management Lund Studies in Economics and Management 144

ISBN 978-91-7753-956-8ISSN 1652-8220

Essays on Credit RatingsANAMARIA COCIORVA

DEPARTMENT OF BUSINESS ADMINISTRATION| LUND UNIVERSITY

Essays on Credit RatingsThis book investigates the role, as well as the accuracy of credit ratings, both for companies and countries. In the first article, I find that credit ratings capture financial constraints more effectively relative to other popular proxies. The second paper focuses on a change in the rules of a widely-followed bond index, and finds that there is a significant market response for the bonds affected by this change. This suggests that credit ratings also have a significant role for bond market segmentation. In the third paper, I find that sovereign ratings have become increasingly conservative over time. However, in my final paper, I focus on the method used for estimating rating ”conservatism”, and , by using a new, two-step model, I find that previous research has overestimated its magnitude.

978

9177

5395

68

144