Are Financial Ratios Still Relevant for Capturing Credit Risk? Evidence from the CDS Market

Transcription

1 Are Fnancal Ratos Stll Relevant for Capturng Credt Rsk? Evdence from the CDS Market by George Chalamandars 1 and Nkos E. Vlachogannaks 2 Abstract In ths artcle we explore n depth the way changes n CDS spreads relate to fnancal ratos changes, and delve nto the underlyng propertes of ths relatonshp under the structural models theory. Our results suggest that fnancal ratos can help explan part of the CDS spread varablty, whle at the same tme the CDS market s effcent n correctly antcpatng the greatest part of changes n the fnancal ratos well before these are offcally released. Usng threshold regresson s statstcal theory we confrm the nsghts of both structural model theory and practcal ntuton behnd the workngs of that partcular market. Our evdence mples that the relatonshp between CDS changes and the changes n the fnancal statements s ndeed nonlnear, and the use of Leverage and Prce to Book ratos as threshold varables can lead to pecewse lnear approxmatons n that assocaton. We also observe that systemc factors become the domnant determnants of CDS changes n perods of fnancal turmol. Fnally, we verfy the asymmetrcal mpact of fnancal ratos on the market s percepton of a company s credt rsk by employng quantle regresson. Keywords: CDS; Fnancal ratos; Structural models; Lasso estmator; Threshold regresson; Asymmetrcal mpact; Quantle regresson. JEL: C21, G14, G15, G33 1 Department of Accountng and Fnance, Athens Unversty of Economcs and Busness, 76 Patsson street, Athens , Greece. Emal: gchalamandar@aueb.gr. 2 Market and Lqudty Rsk Analyss Secton, Bank of Greece, 3 Amerks street, Athens, Ph.D canddate n the Department of Accountng and Fnance, Athens Unversty of Economcs and Busness. Emal: nvlachogannaks@bankofgreece.gr. Ths Workng Paper should not be reported as representng the vews of the Bank of Greece (BoG). The vews expressed are those of the authors and do not necessarly reflect those of the BoG. 1

2 1. Introducton - Motvaton Fnancal ratos are wdely used both by the management and by the credtors of a frm. Management manly employs fnancal ratos as a quanttatve tool n ts daly decson makng process, whle credtors utlze fnancal ratos to evaluate a frm s credt rsk. In partcular, credtors have to decde on a seres of ssues regardng the extenson or not of credt to a customer, the amount of credt to be granted as well as the proper prcng of the credt rsk undertaken. In that context, fnancal ratos have been used for 4 decades for evaluatng a company s credt standng (Altman 1968, Ohlson 1982), though, some studes reveal that structural models (Merton 1974) perform better n evaluatng a frm s credt rsk compared to models based exclusvely on fnancal statement varables. Hllegest et al. (2004) nfer that the probablty of bankruptcy estmated usng a Black-Scholes-Merton opton prcng model ncorporates much more nformaton than Z-score and O-score, whch are both calbrated on accountng data. The results of these studes ndcate that the mssng component that enables structural models to potentally outperform pure accountng-based measures of default s asset volatlty 3 (Campbell et al 2003, Vassalou and Xng, 2004). Nevertheless, some studes reveal that fnancal ratos have complementary nformatonal content when used n parallel wth structural models, especally n the presence of not fully effcent markets. Demrovc and Thomas (2007) conclude that accountng varables contan addtonal nformaton not captured by a structural model that ncludes only the dstance-to-default. Agarwal and Taffler (2008) fnd that structural models encapsulate dfferent aspects of credt rsk compared to accountng based models. Furthermore, Das et al (2009) and Ponce (2012) examne the explanatory power of market-based and accountng-based models on a company s CDS spread, just to confrm once agan that the model wth the hgher explanatory power s the one that combnes both market and accountng data. To that end, our study ncorporates fnancal ratos coupled wth the man varables prescrbed under structural models theory, namely leverage and volatlty, to examne how do fnancal ratos affect a frm s credt rsk, as t s reflected n ts CDS spread. 3 Alternatvely, equty volatlty can be used as a proxy for asset volatlty. 2

3 In partcular, the prncpal am of ths study s to explore n depth the relatonshp between changes n a frm s fnancal ratos and changes n ts CDS spread. In partcular, n pursung ths objectve, we seek to dentfy the propertes of the aforementoned relatonshp, n the context of structural models theory and market effcency. We can encapsulate the objectves of ths study n the form of four questons: ) Do changes n the fnancal ratos of a company explan part of the observed CDS spread changes?, ) Does the CDS market s effcent n antcpatng changes n fnancal ratos, before these are offcally made publc?, ) Do company-specfc characterstcs, such as low vs. hgh leverage, or the temporal phase of the economc cycle affect the relatonshp between changes n fnancal ratos and CDS spread changes?, and v) Is the mprovement or the deteroraton of a company s credt-standng drven by the same fnancal ratos and n a symmetrcal way? Wth regard to the frst queston, there s a number of relatvely recent studes that employ fnancal ratos n an attempt to explan the varaton of credt spread levels (Campbell and Taksler 2003, Das, Hanouna and Sarn 2009) or of credt spread changes (Colln-Duffresne, Goldsten and Martn (CGM ), Ercsson, Jacobs and Ovedo 2009). We examne a larger dataset on a much more generc set of changes n fnancal ratos to verfy that the CDS spreads are ndeed affected by changes n fnancal ratos. Instead of usng arbtrarly chosen models, the selecton of the fnancal ratos that are ncluded n our model specfcaton s based on the Lasso estmator, whch consders both parsmony and predcton accuracy. Our results suggest that fnancal ratos related to proftablty, valuaton and fnancal flexblty are statstcal sgnfcant n determnng a frm s CDS spread, mprovng the performance of structural models. Concernng the absorpton of the nformaton contaned n the fnancal ratos by the CDS market, we examne whether any news ncluded n the fnancal ratos of a company s ncorporated n the CDS market before these are offcally made publc. We search for potental jumps n the CDS spread around the announcement date of fnancal statements, arsng from new unprocessed nformaton contaned n a frm s fnancal ratos. Our evdence suggest that the CDS market s largely effcent n antcpatng changes n fnancal statement varables at least a week before these are announced, mplyng that the very announcement of fnancal statements doesn t offer any fresh news n the CDS market. 3

4 The motvaton of our thrd queston arses from the nonlneartes nherent n contngent clams theory (Merton 1974, Jones P., Mason S.and Rosenfeld E 1984, Eom Y., Helwege J. and Huang J.-Z 2004), accordng to whch a frm s debt can be consdered as a long poston on a rsk-free bond and a short poston on a put on the assets of the frm. In most cases, whle ths frm s vable, ths put remans deeply out-of-the-money resultng n a small Delta. Intutvely, ths mples that the default premum of a healthy company exhbts a relatvely small senstvty to the changes of the frms enterprse value that ncreases n magntude as the frm approaches the default threshold. Snce CDS spread changes are expected to have a non lnear relaton wth the frm s fundamentals, and to be also affected by varables such as volatlty, t s thus very lkely that the relatonshp between changes n fnancal ratos and changes n CDS spreads does not reman nvarant throughout the cross-secton of debt ssuers and the tme dmenson of the panel, but t depends on the captal structure, the growth prospects of the company, other frm-specfc varables as well as the regme of the economy. By employng threshold analyss we verfy the non-lnearty of the relatonshp between credt rsk and ts determnants even when we consder changes rather than levels. We fnd that Debt to Market Captalzaton (Leverage rato) and Prce to Book Rato (Valuaton rato) can be used as cross-sectonal dscrmnatory varables to defne statstcally approprate categorzaton wthn whch regresson coeffcents are sgnfcantly dfferent n the cross secton. At the same tme, our fndngs pont to the exstence of structural breaks n the relatonshp we explore, whch occur along the tme dmenson. The relatve mpact of fnancal ratos n the CDS spread changes decreases n tmes of fnancal turbulence and the systemc factors become the domnant determnants of CDS spread changes. Fnally, we explore whether the effect of changes n certan fnancal ratos on the changes of CDS spread s symmetrcal. We examne whether a partcular type of returns,.e. very postve or very negatve, are drven by the same set of fnancal ratos and n the same manner by utlzng quantle regresson 4. Our analyss ponts changes n Leverage and Prce to Book rato as the most mportant ones, n drvng the wdenng of CDS spreads. We also fnd evdence that the deteroraton n the market s percepton of a 4 For more detals on the economc applcatons of quantle regresson see Ftzenberger B., Koenker R. and Machado J.,

5 company s credt rsk s more lkely to depend on bad news, rather than ts mprovement to depend on good news contaned n a frm s fnancal ratos. The mplcatons of our fndngs can be sgnfcant to a great number of nterested partes. Our results suggest that the partcpants n the CDS market have to consder fnancal ratos as a complementary tool for determnng a frm s credt rsk, snce fnancal ratos are statstcal sgnfcant and ncrease the explanatory power of a model that consders only leverage and volatlty. The aspects of our study regardng the asymmetrcal mpact of fnancal ratos and the varablty of the regresson coeffcents both along the cross-sectonal and the tme dmenson, may help the partcpants n the CDS market gan a deeper understandng of the factors that drve CDS changes and thus better manage the rsk of ther postons as well as properly formulate ther nvestments strateges. Our results ndcate that dfferent fnancal ratos matter for low vs. hgh leveraged frms and for growth vs. value frms, hence, t becomes evdent that we have to zoom n on dfferent focal ponts when examnng frms wth dfferent characterstcs. The rest of the paper s organzed as follows. Secton 2 presents our data set and some summary statstcs. In secton 3 we ntroduce our man hypotheses. In secton 4 we present all the emprcal results of our analyss. Secton 5 provdes the necessary robustness checks that further assess our results n the context of relevant lterature and secton 6 concludes. 2. Data and Summary Statstcs 2.1 Data Descrpton Our sample conssts of 5yr CDS spreads on senor unsecured debt that s downloaded from Bloomberg. These data range from 31 Dec 2003 to 31 Dec 2008 and nclude companes all over the world for whch data s avalable n Bloomberg. The selecton of the 5yr CDS spread s prmarly due to ts hgher lqudty among all tradable tenors, and snce Bloomberg provdes most quotes for t, compared to all other tradable tenors. 5

6 In performng our analyss we employ changes of our varables (CGM 2001, Avramov, Jostova and Phlpov 2007, Ercsson, Jacobs and Ovedo 2009) nstead of levels (Campbell and Taksler 2003, Carlng, Jacobson, Lnd and Roszbach 2007, Das, Hanouna and Sarn 2009, Tang and Yan 2009, Bonfm 2009), as the mechansm that underles the market s aggregate percepton of an entty s credt rsk as an absolute level tends to shft wdely between dfferent phases of the busness cycle. Ercsson et al (2009) and Avramov et al (2007) note that ths very fact could pont to possble problems of non-statonarty of the varables used for the estmaton of the level of CDS spreads. Furthermore, the use of percentage changes of fnancal ratos rather than ther levels s more robust not only because fnancal ratos are non-statonary (Ioannds, Peel and Peel 2003), but also because the level of ratos may be structurally dfferent between dfferent sectors, thus basng regresson estmates. CDS spreads are downloaded for one week before and one week after each announcement date of the fnancal statements. We denote as: CDS(+5d) : the CDS level 5 days after the announcement date for quarter, CDS(-5d) : the CDS level 5 days before the announcement date for quarter, CDS(+5d) -1 : the CDS level 5 days after the announcement date of the prevous quarter -1 We calculate two week changes ( ) as well as quarterly CDS spread changes ( CDS Q ) as follows: CDS 2W CD CDS( 5d) CDS( 5d) S ln 2W CDS( 5d) CD S Q ln CDS( 5d) -1 For every announcement date and for each company we also estmate 90 days equty return volatlty changes, followng the same ratonal as descrbed above, both for two weeks and for the quarterly changes. For each company and for each quarter we also download from Bloomberg a broad range of fnancal ratos that cover the broad categores of Leverage & Captal structure, Cash flow protecton & Lqudty, Valuaton, Proftablty and Fnancal Flexblty, as Bloomberg descrbes them. For each of these ratos we then calculate quarterly percentage changes. Fnally, we nclude n our sample data regardng general (1) 6

7 characterstcs of a company that can be used as Dummy varables n the analyss (.e. Country, Sector, Credt Ratng, and Quarter). 2.2 Summary Statstcs After excludng all fnancal ratos wth a few observatons to maxmze our sample, we end up wth 6,244 observatons from 533 companes. There are on average 12 quarterly observatons, out of a total of 20 quarters that our dataset spans, for each company n our sample. Our fnal sample ncludes 22 fnancal ratos that cover the aforementoned broad accountng categores, and we present them n the table that follows. Leverage & Captal structure Debt to market captalzaton Common equty to assets Debt to total assets Long term debt to common equty Proftablty Valuaton Fnancal Flexblty Proft margn Operatng ncome per share Earnngs per share Revenue per share Return on common equty Return on assets Prce to sales per share Prce to book rato Prce to Earnngs rato Sales growth to Tangble Assets Asset Turnover Book Value per share Increase n Equty as a percent of Total Labltes Retaned Earnngs to Tangble Assets Cash flow protecton & Lqudty Cash and cash equvalents per share Cash from operaton to Total Debt Operatng ncome to Long term debt Operatng Income to Total Captal In table 1 we present the Country and Credt ratng profle of our sample. About two thrds of our observatons come from the US, and about two thrds have a credt ratng of A or BBB, at the date of downloadng the dataset. Furthermore, table 2 exhbts our sample by quarter dmenson. There are on average 312 observatons per quarter and the average CDS spread for all companes ncluded n our sample s 123 bps. [Insert Table 1 about here] 7 [Insert Table 2 about here]

8 3. Hypothess Development We develop our analyss n four successve stages, each of whch follows naturally from the prevous one. We frame each stage n the form of an ndvdual hypothess whch we then put to test. The successon of hypotheses that we examne n ths study can be stated as follows: Hypothess 1: Changes n the fnancal ratos of a company explan part of the observed CDS spread changes. Ths hypothess can be expressed n the form of a panel regresson: CD S Q C b X 1 Q (1) b X 2 Q (2)... b N X Q ( N ) ε In equaton (2), CDS Q s defned as n (1), X Q ( m) (2) correspond to quarterly logarthmc changes n fnancal ratos calculated from quarter -1 and, and perod fxed effects are also ncluded. The alternatve hypothessb b... b 0 mples that any varaton n the fnancal ratos s not related to the 1 2 N observed CDS changes. Hypothess 2: The CDS market s effcent n antcpatng changes n the fnancal ratos of a company, before these are offcally made publc. Ths hypothess can be structured wth the help of a panel regresson: CD S 2W C b X 1 Q (1) b X 2 Q (2)... b N X Q ( N ) ε (3) In (3), CDS 2W s defned as n (1), testng n effect whether the adjustment n the CDS spread takes place wthn the two-week perod that encloses the announcement date, X Q ( m) correspond agan to quarterly changes n publshed fnancal ratos, and perod fxed effects are agan consdered. The stated 8

9 hypothess concdes wth the restrcton b b... b 0 whch can be tested aganst the unrestrcted 1 2 N model n (3). Non-rejecton of the restrcton mples that the CDS spread adjustment precedes the announcement date. Hypothess 3: The set of fnancal ratos that explan CDS spread changes remans nvarant throughout the cross-secton of debt ssuers and the tme dmenson. We test ths hypothess usng threshold analyss. To that end, we adopt equaton (2) as the null hypothess of no threshold and as the alternatve a threshold lnear model of the form: CDS Q C b ' X 1 Q (1) b 2 ' X Q (2)... b N ' X Q ( N ) ε, x CDS Q C b '' X 1 Q (1) b 2 '' X Q (2)... b N '' X Q ( N ) ε, x (4) The regresson above s estmated usng the fxed effect estmator. The threshold varable x n equaton (4), determnng the shft from one regme to the alternatve, can be ether a cross-sectonal or a tme varable 5. In the frst case, the valdty of equaton (4) aganst the null mples the non-lnearty of the response functon that lnks the dependent wth the regressor varables. In the second case, the exstence of a tme threshold mples a structural break n that relatonshp that occurred at some pont n tme. Hypothess 4: The response of CDS changes s symmetrcal both n sgn and magntude wth respect to the changes n fnancal ratos. An OLS estmaton of equaton (2) reveals a relatonshp between the means of the dependent and the ndependent varables. We use quantle regresson to estmate (2) for a number of quantles to test, n effect, whether changes n fnancal ratos of dfferent sgn and magntude have a dstnct mpact on the CDS spreads. In partcular, we nvestgate whether there are alteratons n the estmates as well as n the sgnfcance of the coeffcents n (2) across dfferent quantles. 5 In ths case the perod fxed effects are omtted. 9

10 4. Emprcal Analyss Generally speakng, an ncrease n Proftablty, Valuaton and Lqudty are lkely to ndcate an mprovement n the fnancal health of a frm, thus tghtenng ts CDS spread. On the other hand, an ncrease n Leverage suggest the heghtened fnancal rsk of a gven company, wdenng ts CDS spread. Intally, we defne our basc model (eq. 2) to the end of testng hypothess 1. We then use ths model to examne the valdty of the next 3 hypotheses. The set of regressors conssts of fnancal ratos calculated from a frm s publshed quarterly fnancal statements, as well as perod fxed effects as n Tang (2009). In secton 5, we employ macroeconomc varables n the sprt of CGM (2001), Campbell J. et al (2003), Das S. et al (2009) and Bonfm D. et al (2009) among others, to evaluate the robustness of our results and to verfy that our man fndngs stll hold. Among the 22 fnancal ratos that we examne, we make the fnal choce of the varables to nclude n our prncpal model on the bass of ther sgnfcance when testng hypothess 1. To ths end, we mplement the Least Angle Regresson (LARS) algorthm proposed by Efron, Haste, Johnstone and Tbshran (2004) to derve the Least Absolute Shrnkage and Selecton Operator (LASSO) estmates, whch combne both model parsmony and predcton accuracy. The Lasso estmator can be consdered as an adjusted verson of the Ordnary Least Square (OLS) estmator. The OLS estmator mnmzes the sum of the squared resduals, whle the LASSO estmator apples the constrant that the L 1 norm (rectlnear dstance) of the parameters vector s not hgher than a gven value. For smaller values of the constrant the Lasso contracts the OLS regressors towards zero, enhancng the predcton accuracy (Haste et al 2001). The mplementaton of the LARS 6 algorthm enable us to derve all Lasso estmates n a more computatonally effcent way. The calculaton burden s substantally decreased under the LARS algorthm, snce t consders rather larger steps compared to the Forward Stagewse method, stll not so large as the Forward Stepwse regresson. The Least Angle Regresson (LARS) starts by selectng the explanatory varable wth the hghest absolute correlaton (X 1 ) wth the dependent varable and performs 6 For a detaled overvew of the LARS algorthm, please consult Efron, Haste, Johnstone and Tbshran (2004). 10

11 an OLS estmaton. The resduals obtaned from the prevous regresson are then consdered as the dependent varable so as for the next varable (X 2 ) wth the hghest absolute correlaton to be dentfed. As opposed to the Forward Stagewse and Forward Stepwse approaches, n whch the process carres on along (X 1 ), under the LARS algorthm both predctors are equally consdered untl a thrd regressor s dentfed to be ncluded n the set of varables that are the most correlated, and so on. The number of steps requred by the LARS algorthm equals the number of regressors whose ncluson or not n the model s examned, hence the computatons are substantally speeded up. Due to multcollnearty problems that arse when testng our ntal set of 22 varables, we remove 8 fnancal ratos 7 for whch there exst such concerns and we perform the LARS algorthm agan. Furthermore, from the varables that are fnally selected by the LARS algorthm, we remove the ones that are not statstcal sgnfcant from our fnal model specfcaton. As a result of the abovementoned process, the quarterly changes of the followng varables are ncluded n our base model (eq. 2): Debt to Market Captalzaton (DM): Ths rato s calculated as (Short Term Debt + Long Term Debt)/Market Captalzaton. Market Captalzaton s calculated as (Closng Prce as of fscal perod end date) x (Shares outstandng at that perod end date). DM Short Term Debt Long Term Debt Closng Prce Shares outstandng Prce to Book Rato (PB): Rato of the stock prce to the book value per share. PB Stock prce Book value per share Earnngs per Share (EPS): Computed as Net Income Avalable to Common Shareholders dvded by the Basc Weghted Average Shares outstandng. Net ncome ncludes the effects of all one-tme, nonrecurrng and extraordnary gans/losses. In calculatng the Basc Weghted Average Shares, the effects of convertbles are excluded. 7 The varables that are removed due to multcollnearty problems, the results of the statstcal tests for multcollnearty as well as the ntermedate regresson results are avalable by the authors upon request. 11

12 EPS Net Income Basc Weghted Average Shares outstandng Sales Growth to Tangble Assets (SGTA): Annual Sales change s calculated usng for nterm perods the comparatve perod of the precedng year. We use a full year n calculatng Sales Growth so as to exclude potental seasonal effects from the analyss. SGTA Net Sales for thecurrent perod - Net Sales Tangble Assets for the last perod Not surprsngly, the fnancal ratos that are ncluded n our model as determnants of a frm s CDS spread cover the categores of leverage, valuaton, proftablty and fnancal flexblty common to many studes from 1968 (Altman E.) up to now. The sgnfcance of Debt to Market Captalzaton (leverage) s n lne wth most artcles tryng to predct fnancal dstress (Campbell J. et al, 2003, Molna C. A. et al 2005, Campbell J. et al, 2008) or explan corporate bond credt spread changes (CGM 2001). The Prce to Book rato (valuaton rato) s found as an mportant determnant of credt rsk n Vassalou M. et al (2004) and n Avramov D. et al (2007). In partcular, Vassalou notes that frms wth hgh probabltes of default have hgh Book to Market ratos. Expressng ths fndng n terms of changes n probabltes of default, one can say that frms facng ncreases n ther probabltes of default, as t s denoted by an ncrease n ther CDS spread, also experence a decrease n ther Prce to Book rato. It s worth mentonng that the fnancal ratos regardng Debt to Market Captalzaton and Prce to Book ratos ncorporate, along wth the potentally unprocessed nformaton of fnancal statements, nformaton comng from the stock market. Therefore, the channel of the nformaton from the equty market s ncorporated n our establshed models, n lne wth the approach followed by CGM (2001) who employ DM or nterchangeably a frm s stock return to explan bond credt spreads. 12

13 Earnng per Share s a proftablty rato smlar to the ones used n Altman E. (1968) and Ohlson J. (1982). Fnally, the sgnfcance of Sales Growth to Tangble Assets rato s n lne wth metrcs that ncorporate sales expanson, common n Das S. et al (2009) and n Moody s Prvate Debt Manual 8. For all of the fnancal ratos, except for Earnngs per Share, we estmate logarthmc changes. For Earnng per Share we use the dfference from the prevous quarter, as earnngs per share can turn negatve, and hence logarthmc changes cannot be defned. Table 3 presents descrptve statstcs for the varables we use n the analyss, whether ncluded as dependent or as ndependent varables n the models we set up. [Insert Table 3 about here] 4.1. Testng Hypothess 1: Changes n the fnancal ratos of a company explan part of the observed CDS spread changes. We ntate our analyss by frst examnng the set of fnancal ratos that explan part of the CDS spread changes varablty. In our basc model (eq. 5) we ft through a panel regresson wth perod fxed effects the CDS quarterly changes ( CDS Q ) wth respect to a set of ndependent varables that conssts of a constant (C) and the changes n fnancal ratos (ΔDM, ΔPB, ΔEPS, ΔSGTA). Model 1: CD S Q C b DM 1 b PB 2 b EPS 3 b SGTA 4 ε Usng a smple F- Statstc, we test hypothess 1 by testng the restrcton of b b b b 0 aganst the unrestrcted alternatve. The valdty of ths restrcton s rejected at the 0.01 level as the F-statstc equals In estmatng Model 1 (base model) we use the Whte methodology to adjust the standard errors of the coeffcents for heteroskedastcty and autocorrelaton so as to produce robust estmates. Moreover, we (5) 8 It s publshed on Moody s KMV webste. 13

14 examne the ndependent varables for multcollnearty usng varance nflaton factors, but no such evdence s found. Our results, presented n table 4, ndcate that Model 1 (eq. 5) explans about 44% of the changes n the CDS spread. We can also notce that the coeffcents for the change of Debt to Market Captalzaton (leverage rato) and Sales Growth to Tangble Assets (fnancal flexblty rato) are postve, ndcatng a wdenng n CDS spreads when leverage or Sales ncrease, whle the coeffcents of the change of Prce to Book rato (valuaton rato) and Earnng per Share rato (proftablty rato) are negatve, mplyng a tghtenng n CDS spreads when Prce to Book rato and Earnng per Share rato ncrease. At the same tme both the sgnfcance and the magntude, n absolute terms, of the coeffcents suggest that the effect of the change n Debt to Market Captalzaton (0.12) and Prce to Book rato (0.16) domnate over the effect of Sales Growth to Tangble Assets (0.04) and Earnng per Share rato (0.0095). We next augment our Model 1 (eq. 5) by ncludng the equty return volatlty quarterly change (ΔVOL), calculated from the most recent 90 tradng days 9, both to examne whether the fnancal ratos used n our model encompass some of the nformaton content of equty return volatlty as well as to algn our analyss wth contngent clams theory that asserts the mportance of asset volatlty n determnng a frms credt rsk. Not surprsngly, the total explanatory power of our new model (Model 2 - eq. 6) ncreases by 2.3 percentage unts, as depcted n the ncreased Adjusted R-squared n Table 4. Model 2: CD S Q C b ΔDM b ΔPB 1 2 b ΔEPS 3 b ΔSGTA b ΔVOL 4 5 ε (6) 9 Of course, we should note here that the equty volatlty the structural model theory alludes to, s the marketexpected, forward-lookng volatlty of total assets. Ths s usually nferred from stock optons, whch however are not readly avalable for a sgnfcant part of our sample. For exactly ths reason, we use hstorcal volatlty nstead of the mpled one as many authors have done before us (for example Ercsson et al (2009)). The argument goes that mpled volatlty s lnked to past realzed volatlty (Chrstensen, Prahbala, 1999, Chalamandars, Rompols, 2010), thus we use ths varable as one that s smply correlated to total asset volatlty. 14

15 In settng up Model 2, we agan utlze the LASSO algorthm after ncludng volatlty changes nto our set of ndependent varables. The LASSO algorthm selects 10 the fnancal ratos employed under Model 1, as well as prce to earnngs, revenue per share and cash flow from operatons to total debt. However, these latter fnancal ratos are not statstcal sgnfcant and hence we do not nclude them n Model 2. Moreover, to further assess the robustness of our result we also nclude quarterly equty return changes nto the set of our ndependent varables. Our fndngs 11 ndcate that the fnancal ratos already dentfed reman statstcal sgnfcant, apart from earnngs per share that becomes nsgnfcant and prce to book rato that remans sgnfcant at the 10% confdence level. Therefore, the dentfed fnancal ratos do explan a part of the CDS spreads that s not captured nether by equty returns nor by equty returns volatlty. The coeffcent for the volatlty change appears to have the strongest effect n determnng CDS spread changes, as t has both the largest magntude and sgnfcance. Fnancal varables reman sgnfcant, though, both ther magntude and ther sgnfcance slghtly decrease, apart from the sgnfcance of the coeffcent for ΔDM whch remans almost at the same level. By dentfyng leverage and volatlty as the cornerstones n determnng fnancal dstress, our fndngs verfy the mportance of structural models theory. However, leverage and volatlty do not ncorporate the 100% of a frm s credtrelated nformaton snce fnancal ratos reman hghly statstcal sgnfcant. Thus, any models attemptng to capture credt rsk have to be complemented wth fnancal ratos. Furthermore, changes n equty volatlty seem to convey part of the nformaton regardng a frm s changes n proftablty, as the coeffcent for Earnngs per Share becomes margnally nsgnfcant at the 10% confdence level. The change n the coeffcents of tme dummes 12 for Model 1 (eq. 5) and Model 2 (eq. 6) suggest that equty volatlty n Model 2 captures some of the tme seres varaton attrbuted to tme dummes. The mean of the tme dummes s closer to zero for Model 2 (Model 1: , Model 2: ), as well as the standard devaton of the tme dummes n Model 2 s smaller (Model 1: , Model 2: ). These 10 All results are avalable on request. 11 All results are avalable on request. 12 The coeffcents of tme dummes are not presented here for the sake of savng space. All results are avalable on request. 15

16 fndngs are smlar to those of Campbell (2003), mplyng that equty volatlty contans systemc prema prced n the CDS market. Furthermore, our results are also n lne wth Tang (2009) who dentfes mpled volatlty as the most mportant frm specfc credt spread determnant, and Campbell (2003) who hghlghts the sgnfcance of dosyncratc equty volatlty n corporate bond yeld spreads, even n the presence of other mportant factors that drve credt rsk. Last but not least, our emprcal evdence s supported by the ntuton ganed from structural models, n the sense that the prce of an out-of-the-money put depends on the mpled volatlty of the underlyng more than on anythng else. Comparng our fndngs wth those of CGM (2001) on credt spread changes of corporate bonds, we can note that ther proposed model has an explanatory power of about 19%-25% across all maturtes and dfferent leverage group, whch s much lower than the explanatory power of Model 2 (46%). A part of ths dfference could be attrbuted to the macro varables used by CGM (2001) that take nto account only the observable systemc components, whle our model employs perod fxed effects to capture both the observed and the potental latent systemc components. However, our results 13 suggest that even f we use macroeconomc varables nstead of tme dummes, the explanatory power of our model s 33.7%, about 10% hgher than n CGM (2001). Therefore, the hgher explanatory power of our model verfes Ercsson s et al (2009) conclusons that the CDS market offers a better measure of an ssuer s credt rsk compared to the corporate bonds market. [Insert Table 4 about here] 4.2. Testng Hypothess 2: The CDS market has predctve power n antcpatng changes n the fnancal ratos of a company, before these are offcally made publc. Havng decded on the set of fnancal ratos whose changes nfluence CDS spread changes, we contnue to the next stage of our analyss n whch we nvestgate whether the CDS market has predctve power and t s able to antcpate fnancal statement alteratons well before ther announcement date. To 13 These results are avalable on request. 16

17 frame ths research queston nto a testable hypothess, we use as dependent varable the CDS 2-week change ( CDS 2W ) 14 and examne whether or not ths CDS spread change s explaned by the recorded quarterly changes n the fnancal ratos of the frm. These fnancal statement varable changes 15 become offcally known to the publc only on ther announcement date. If the CDS market s unable to predct these changes at all, then the adjustment of the CDS spreads that s caused by the release of the new fnancal report, wll take place n ts entrety wthn these 2 weeks. The volatlty change that we nclude n these regressons s calculated for the same tme perod of two weeks ( change ( ). CDS 2W VOL 2W ) as the CDS 2 weeks Model 3a: CD S 2W C b ΔDM b ΔPB b ΔEPS b ΔSGTA ε (7a) Model 4a: CD S 2W C b ΔDM b ΔPB b ΔEPS b ΔSGTA b ΔVOL W ε Our evdence from these regressons n conjuncton wth our results n the prevous secton, pont to the concluson that changes n fnancal ratos have already been ncorporated n the CDS spread a whole week before ther offcal announcement. Indeed, n table 5, panel A, we fnd that all the coeffcents of the fnancal statement varables for Models 3a and 4a are not statstcally sgnfcant. In other words, the CDS changes varablty of ths partcular 2-week perod s drven by factors that are not related to the fnancal report n queston. If the fnancal ratos hadn t already been ncorporated n the CDS spread by the begnnng of our testng perod, we would have expected a jump at the announcement date of fnancal statements, leadng to ncreased explanatory power for some of the fnancal ratos. The only varable that remans sgnfcant s the 2-week volatlty change ( VOL 2W (8a) ) n Model 4, further hghlghtng the 14 Ths s defned as the (log-) change for the perod that starts 1 week before the announcement date and ends 1 week after the announcement date t. See also equaton (1). 15 The fnancal rato changes are defned as the (log-) dfferences between the fgures recorded on the fnancal statement of 17 t mnus the fgures recorded on the prevous quarterly release on t -1.

18 mportance of volatlty as the domnant short-term determnant of credt rsk, and one that we would expect anyway gven the systemc factors that are consttuents of ths varable. Intuton suggests that there mght not be the same fnancal ratos that explan quarterly CDS changes and two week CDS changes, hence, we next proceed by explorng whether there are any fnancal ratos from the ones under our ntal set that are sgnfcant for determnng two week CDS changes. The selecton of the fnancal ratos s agan performed by the LASSO algorthm, and we present our results n Panel B of Table 5. Out fndngs mply that changes n Prce to Earnngs, Sales Growth to Tangble Assets and Cash Flow from Operatons to Total Debt are statstcal sgnfcant n determnng 2 week CDS spread changes, whle Operatng Income to Total Captal becomes sgnfcant at the 10% confdence level when we also consder 2 week volatlty n our model. It seems that the announcement of the fnancal ratos that are related somehow wth frm proftablty slghtly affect ts CDS spread, by n a sense performng some sort of mld fne-tunng between expected and publshed results. Model 3b: CD S 2W C b ΔPE b ΔSGTA b ΔCFTD ε (7b) Model 4b: CD S 2W C b ΔPE b ΔSGTA b ΔCFTD b ΔOITC b ΔVOL 4 5 2W ε (8b) However, gven the very small senstvtes of the fnancal ratos as well as the explanatory power of Models 3b and 4b that s almost dentcal to the explanatory power of Models 3a and 4a respectvely, there s no evdence of a jump n the CDS spread after the announcement date of fnancal statements. Furthermore, the explanatory power of the abovementoned models s manly attrbuted to the systemc factors, whch are captured by the fxed effects n the tme dmenson, and t approaches zero n case the fxed effects are removed. Takng all the above nto account, we can conclude that the CDS market has ncorporated almost any changes n fnancal ratos, before these have been offcally announced. To put t dfferently, the CDS spreads mply the use of unbased estmates for the mmnent fnancal ratos n the prcng of a frm s credt rsk, thus ncorporatng any necessary adjustment well before the announcement 18

19 date of the fnancal statements. Ths fndng s n agreement wth studes that examne the mpact of ratng announcements (Hull et al 2004, Norden and Weber 2004), rather than the announcement of fnancal reports, on CDS spreads. [Insert Table 5 about here] 4.3. Testng Hypothess 3: The set of fnancal ratos that explan CDS spread changes remans nvarant throughout the cross-secton of debt ssuers and the tme. In ths subsecton, we explore n greater depth the ncorporaton of fnancal ratos n CDS changes by examnng f the causal pattern that Model 2 dctates remans nvarant throughout the cross-secton of debt ssuers and the tme dmenson. Ths research queston s agan motvated by opton prcng theory. Indeed, by constructon, structural models stpulate non-lnear relatonshps between fnancal ratos and changes n credt rsk. However, n emprcal studes of the relevant lterature, a number of researchers approach the problem wth lnear approxmatons smlar to our Model 2 (e.g. Schaefer S.M. and L.A. Strebulaev 2008, Tang 2009, Ercsson et al 2009). A queston that arses naturally n those artcles refers to whether the regresson coeffcents are stable when the model s calculated for dfferent sub-samples. CGM, 2001 proceed n an ad-hoc segmentaton of the orgnal panel n terms of leverage and fnds that the coeffcents are ndeed not stable 16. In partcular, whle GCM (2001) use arbtrarly chosen leverage rato subgroups to examne the same thng, we dfferentate our work by pursung a statstcally consstent way to dentfy any clusters as well as we search for more than one cross-secton varable that could potentally dscrmnate between subgroups. 16 Our attempts to nclude n the orgnal regressons (Model 1, 2) non-lnear terms of leverage or other crosssectonal varables does not mprove the ft of the model whch s a fndng n lne wth CGM 2001 et al and Avramov et al, All the coeffcents for the nonlnear terms are nvarable statstcally non-sgnfcant. The results of these regressons are avalable from the authors upon request. 19

20 Splttng a panel dataset n subsamples, requres a choce to be made on the approprate varable that determnes the splt,.e. the threshold varable. At the same tme, the value of that varable at whch the panel s dvded nto subsamples,.e. the threshold value, must also be determned. Whle for categorcal varables such as credt ratng, country, etc., we can dentfy the threshold values relatvely easly, we need to select a statstcally plausble method to do the same for contnuous varables. To avod dong so n an arbtrary fashon, one needs the approprate asymptotc theory that provdes tests about ts exstence and grants the requred statstcal confdence concernng ts poston 17. For the purposes of our analyss we follow Hansen, E. B. (2000) approach, who has developed the requred statstcal theory for the exstence of a sngle threshold. Accordng to Hansen (2000) the threshold regresson model can be wrtten as y y x e, q 1 x e, q 2 (9) In the model of equaton (9) q s the threshold varable used to splt the sample, the value of the threshold varable, e the regresson error, 1 the coeffcent of the frst sub-sample and 2 s the coeffcent of the second subsample. We test for the exstence of the threshold based on a partcular threshold varable q wth the help of the heteroskedastcty-consstent Lagrange multpler (LM) test for a threshold of Hansen (1996). Snce the threshold s not dentfed under the null of no-threshold effect, the p-values are computed wth the help of a bootstrap procedure. If we fnd evdence that a threshold value does exst by rejectng the null of the LM test, then equaton (9) can be used as a vald representaton of the relatonshp between the dependent and the ndependent varables. In ths case, the above two equatons are merged nto one usng a dummy varable that s ndexed on the threshold varable, ntroducng the effect of the dfferent sub-samples nto our model. Least squares estmators are then derved by mnmzng the sum of squared errors functon Regresson trees have been one alternatve (for example Durlauf N. and Johnson P, 1995), whch however requres a number of rather arbtrary decsons on the parameterzaton of the tree. 18 For a detaled llustraton of the methodology employed here, see Hansen E. B

21 The estmator we use assumes heteroskedastc errors snce prelmnary tests n our sample reveal that ths s necessary. In our analyss at ths stage, we use Model 2 as the Null Hypothess of no threshold effect. We then splt our panel n 3 dfferent ways, each tme by adoptng one of the 3 canddate threshold varables as the one that determnes the statstcally approprate categorzaton. We successvely test as threshold varables () Debt to Market Captalzaton rato, () Prce to Book rato and () Tme, that s an nteger denotng the quarter of the partcular observaton. The am of ths exercse s dual. Frstly to examne the degree at whch a specfcaton of the varables n Model 2 s nherently non-lnear, and secondly to nvestgate the ntrcaces and mplcatons of ths specfcaton n the ndvdual sub-samples. We utlze the fxed effects estmator n the calculatons that follow. Startng wth our frst canddate, we use Hansen s (1996) threshold (LM) test to examne f there s some value of the Debt to Market Captalzaton rato for whch a statstcally approprate categorzaton exsts. The bootstrap p-value (<0.0001) ndcates that f we splt once the regresson n two subsamples, based on ths varable, we wll end up wth a nonlnear model n the form of equaton (10), n whch the respectve regresson coeffcents (a) b and (b) b are sgnfcantly dfferent. CDS b 4 a Q ΔSG C a a a a b b 5 a 1 ΔDM b ΔVOL ε, 2 ΔPB b DM 3 ΔEPS CDS b 4 b Q ΔSG C b b b b b b 5 b 1 ΔDM b ΔVOL ε, 2 ΔPB b DM 3 ΔEPS (10) The frst threshold value s and s presented n table 6, mplyng that frms havng a Debt to Market Captalzaton rato hgher than behave dfferently from frms havng a smaller one. After the frst sample splt, the explanatory power of our model ncreases to as ndcated by the Total jont R-squared, an ncrease of about 1.5 percentage unts compared wth our base Model 2. We replcate the procedure by testng whether the next larger subsample,.e. frms wth Debt to Market Captalzaton from 0 to , can be further splt n two. Agan, we reject the hypothess of no- 21

22 threshold effect n ths subsample based on the bootstrap p-value (0.0006). The estmated threshold level for Debt to Market Captalzaton n ths case s Havng splt our sample nto 3 sub-samples the total jont R-squared grows further to 0.484, that s, 2 percentage unts hgher compared to our base model (eq. 6 - Table 4, Model 2). Fnally, we examne f the upper part of our sample,.e. frms wth Debt to Market Captalzaton rato hgher than , can be further splt nto two sub-samples. The Bootstrap p-value (zero) provdes evdence of an addtonal splt. The threshold level for Debt to Market Captalzaton rato for ths case s equal to The estmated model of 4 regmes has a total jont R-squared of 0.492, about 3 percentage unts hgher compared to the lnear model (eq. 6 - Table 4, Model 2). We must stress here, that we use the aforementoned testng procedure only to acqure an nformatve comparson wth the nsghts of structural models. We nether seek the true number of breaks nor the true break sequence 19. [Insert Table 6 about here] In Table 7, we dsplay the coeffcents of the four regme model (eq. 10) when usng as threshold varable the Debt to Market Captalzaton rato. The volatlty changes regressor varable remans sgnfcant n all 4 regmes, exhbtng however lower sgnfcance (lower t-statstc) for frms wth very hgh leverage. Ths s consstent wth theory, snce hghly leveraged frms have a hgher delta wth respect to ther fundamentals, thus dmnshng the effect of volatlty n ths segment of the panel. Furthermore, the hgher magntude and sgnfcance of the coeffcent for the Debt to Market Captalzaton change ndcates that the more leveraged a frm s, the stronger the mpact of leverage on ts CDS spread. Indeed, the Debt to Market coeffcent becomes nsgnfcant for low leveraged frms. These results are n lne wth CGM 2001, who examned the sgnfcance of changes n a frm s leverage, for a seres of some ad hoc leverage groups. Fnally, for hghly leveraged frms the effects of prce to book rato 19 It s mportant to note ths, because the test tself assumes a unque threshold break and thus t s not clear how the theoretcal results regardng the confdence ntervals of the threshold levels extend n our applcaton. Ths s why we avod the presentaton of the standard errors of the nferred thresholds as ths could be msleadng. 22

23 changes decreases n magntude and n sgnfcance when examnng CDS spread changes, suggestng that the mpact of prce to book rato on CDS spreads s surpassed by other factors. The coeffcents for Debt to Market Captalzaton change, for Prce to Book change and for Volatlty changes n the lnear model are 0.115, and respectvely (Table 4, Model 2, eq. 6), whle for the statstcally dfferent sub-samples, they range from 0.07 to 0.19 for ΔDM, from to for ΔPB and from 0.29 to 0.39 for ΔVol, when takng nto account only the sub-samples for whch these coeffcents reman sgnfcant. These fndngs confrm the non-lneartes nherent n the structural model theory, snce the senstvtes of the fnancal ratos do not reman nvarant throughout the cross-secton of debt ssuers, across dfferent levels of leverage. [Insert Table 7 about here] In order to valdate the robustness of our observatons n terms of alternatve cross-sectonal varables, we repeat the above procedure usng ths tme as threshold varable not a leverage rato, but the next most sgnfcant fnancal rato examned,.e. the Prce to Book rato. Successve applcaton of the Hansen s heteroskedastcty consstent (LM) test provdes evdence for (at least) 4 splts based on dfferent Prce to Book rato categorzatons: (0 0.89), ( ), ( ), and (>2.28), ndcatng agan that the regresson coeffcents of Model 2 are sgnfcantly dfferent n each subsample (Table 6). The total jont R-squared of the trple-threshold model ncreases to 0.485, about 2.2 percentage unts hgher than the lnear Model 2 of the no-threshold hypothess. The coeffcents of Model 2 n each of the four sub-samples, presented n Table 8, ndcate that volatlty remans agan sgnfcant n all subsamples, valdatng once agan ts exceptonal poston n the context of structural models. The magntude and the sgnfcance of changes n Debt to Market Captalzaton are larger for low Prce to Book frms, ndcatng that leverage does matter more for value than for growth frms. What s more, earnngs per share changes are not sgnfcant n any regme for ths panel categorzaton, whle the coeffcent for Sales to Tangble Assets changes ncreases n sgnfcance for hgh Prce to Book frms. Ths latter observaton may ndcate the addtonal rsks that 23

24 sales expanson entals for growth frms and that these rsks are prced n the CDS market by the market partcpants. The coeffcents for Debt to Market captalzaton change, for Prce to Book change and for Volatlty changes for our base model are 0.115, and and respectvely (Table 4, Model 2), whle for the statstcally dfferent sub-samples, they range from 0.11 to 0.22 for Debt to Market Captalzaton change, from to for Prce to Book change and from 0.24 to 0.44 for volatlty changes, when consderng only the sub-samples for whch these coeffcents reman sgnfcant. All n all, our fndngs provde clear evdence that the mpact of fnancal statement varables on CDS spread changes depends on the partcular segment of the populaton that the frm n queston belongs. Ths s demonstrated usng ether the Leverage or the Prce to Book rato as the dscrmnatng varable. [Insert Table 8 about here] At a fnal applcaton of the threshold analyss for CDS changes, we use t to test for structural breaks along the tme dmenson n the relatonshp of Model 2 (Table 4, eq. 6). To ths end, we exclude the Tme Dummes from Model 2, and employ Quarters as the threshold varable. We then replcate the analyss we descrbed prevously. At each applcaton of the threshold test, and the subsequent estmaton of the larger model, we fnd that ts explanatory power ncreases, from 18% to 25% after the frst splt, then to 28.5% after the second and fnally to 33.8% after the thrd splt, suggestng hgh varaton across the tme dmenson. In table 9 we present the coeffcents of the resultng threshold model. We can observe that volatlty changes become the domnant drver of CDS changes after the 17th quarter (31/03/2008) due to the outbreak of the fnancal crss. After that tme, changes n volatlty and n leverage have a domnant role n determnng CDS spread changes. Ths evdence s n lne wth the common percepton that snce that tme, correlaton n CDS spreads has ncreased dramatcally. The coeffcents n the dfferent subsamples, when takng nto account only the sub-samples for whch these coeffcents reman sgnfcant, range from 0.06 to 0.15 for Debt to Market Captalzaton change, 24

25 from to for Prce to Book change and from 0.27 to 0.74 for volatlty changes. These fndngs verfy that fnancal ratos do not reman nvarant throughout the tme dmenson of our sample. [Insert Table 9 about here] 4.4. Testng Hypothess 4: The response of CDS changes s symmetrcal both n sgn and magntude wth respect to the changes n fnancal ratos. On the release of a new fnancal report concernng the quarterly results of a company, both ntuton and theory stpulate the percepton that a frm s credt rsk wll change n an asymmetrcal fashon. Notwthstandng whether a company s healthy or not, we expect negatve results to have a larger mpact n the CDS market, gven that the latter prces a pure downsde rsk whch s asymmetrc by nature. The am of ths subsecton s to verfy ths lack of symmetry by examnng whether the set of factors that drve frms to default,.e. wdenng of CDS spreads (postve changes), s dentcal to the one that drves frms to prosperty,.e. tghtenng of CDS spreads (negatve changes), and whether ther respectve coeffcents reman relatvely constant. Hull et al, (2004) and Norden and Weber (2004), among others, have shown that the mpact of negatve news regardng the ratng of a company on CDS returns s more pronounced and statstcally sgnfcant when compared to the statstcally nsgnfcant mpact of postve news. Usng the same framework from another pont of vew, we examne ths asymmetry wth the help of quantle regresson analyss. Indeed, whle the method of ordnary least squares provdes us wth estmates of the condtonal mean of the dependent varable gven certan values of the ndependent varables, quantle regresson provdes us wth estmates of ether the medan or other quantles of the response varable. Therefore, t seems to be the natural tool to help us dstngush between potentally dfferent sets of regressor varables and ther respectve coeffcent patterns, that are responsble for causng dfferent response n the CDS market. 25