Are Outlooks and Rating Reviews capable to bridge the gap between the agencies through-thecycle and short-term point-in-time perspectives?

Transcription

1 COVER PAGE WITH AUTHOR NAMES Are Outlooks and Rating Reviews capable to bridge the gap between the agencies through-thecycle and short-term point-in-time perspectives? Edward I. Altman 1 and Herbert A. Rijken 2 October 2005 JEL categoryification: G20, G33 1 NYU Salomon Center, Leonard N. Stern School of Business, New York University, 44 West 4 th 2 Street, New York, NY 10012, USA. ealtman@stern.nyu.edu Free University, De Boelelaan 1105, 1081 HV Amsterdam, The Netherlands. hrijken@feweb.vu.nl We thank Moody s for funding this study and for providing an extended version of data on outlooks and rating reviews for Moody s corporate issuer credit ratings

2 COVER PAGE FOR BLIND REFEREE Are Outlooks and Rating Reviews capable to bridge the gap between the agencies through-thecycle and short-term point-in-time perspectives? JEL categoryification: G20, G33 2

3 Abstract Besides corporate issuer credit ratings, outlooks and rating reviews on a Watchlist are important additional sources of credit risk information to investors. The aim of this study is to provide a guideline how interpret outlook ratings outlooks and rating reviews on a Watchlist - and how to interpret a combination of issuer ratings and outlook ratings in terms of through-the-cycle and point-in-time credit risk perspectives. In contrast to issuer ratings outlook ratings are sensitive to credit risk cycles, even more than from a one-year point-in-time perspective. Although outlook ratings have above all a timing objective to signal upcoming issuer migration events, outlook ratings contain valuable point-in-time credit risk information. We benchmark properties of actual outlook ratings by simulated outlook ratings which have no timing objective and only follow credit scores of an outlook prediction model. Actual outlook ratings appear to be less informative on short term credit risk than simulated outlook ratings. We show that the timing objective does not severely overrides the credit risk information in actual outlook ratings. Instead we suggest that credit risk information signaled by outlook ratings is affected by the absence of standardization of credit risk information in the outlook rating scale. As outlook ratings are not intended to quantify credit risk information, agencies do not standardize the credit risk signaled by outlook ratings. The absence of credit risk standardization in the outlook rating scale and the slight adverse credit cycle sensitivity of issuer ratings prevent actual outlook ratings to bridge the gap between the agencies through-the-cycle perspective and the investor s point-in-time perspective. Instead, issuer ratings adjusted by their outlook ratings reflect a more moderate version of a through-thecycle perspective. Adjusted issuer ratings will switch to a long term point-in-time perspective after standardizing the credit risk signaled by outlook ratings. 3

4 I Introduction In addition to their corporate issuer credit ratings, agencies provide outlooks and rating reviews on a Watchlist. Outlooks signal the likely direction of an issuer migration in 1-2 years time. In response to an event or a abrupt break in a trend a corporate-issuer credit rating is placed on a Watchlist by Moody s, on CreditWatch by Standard&Poor s or on a Rating Watch by Fitch. In these cases issuer ratings are said to be under review and the outcome is disclosed typically within 90 days. In this paper outlook ratings refer to both outlooks and rating reviews. Issuer ratings refer to corporate-issuer credit ratings. Issuer rating migration and outlook rating migration are shortened to issuer migration and outlook migration. Outlook ratings have become an important source of credit risk information to investors in addition to issuer ratings. Especially the timeliness of this credit risk information is valued by investors. Issuer ratings are the outcome of a through-the-cycle methodology which makes them relative stable, insensitive to credit cycles and long-term oriented. A drawback of this methodology is a lower timeliness compared to a one year point-in-time perspective which most investors have. According to a survey conducted by Moody s in 2002 investors agree with the goal of more timely rating actions including shorter review periods. However they use and appreciate the rating review and outlook signaling process; they derive substantial information from them... In their response to this survey Moody s intended to improve rating timeliness by shorter (internal) rating reviews and retaining the provision of outlook ratings. Without having to change their though-the-cycle methodology agencies can fulfill the desire of more timely credit risk information by issuing outlook ratings in addition to issuer ratings. Strictly according to their definition outlook ratings are indications of the likely direction of an issuer migration in the short or medium term. The official outlook rating information does not pretend to signal any information on the size of a potential issuer migration. However from historical data one can compute the average issuer migration at the ending of outlook ratings. The resolution of a Moody s down review is on average an issuer migration of -1.0 notch steps. Moody s negative, stable, positive outlooks and up reviews result on average in an issuer migration of -0.4, -0.1, +0.2 and +1.0 notch steps. Although outlook ratings are not meant to be a correction for issuer ratings in the first place one can use outlook ratings as a secondary credit risk rating scale on top of the issuer rating scale and if desired adjust issuer ratings by their outlook ratings. A first study of Hamilton and Cantor (2004) reveals that adjusted issuer ratings improve default prediction performance significantly. The accuracy ratio improves by 4 % for a one year prediction horizon. Another interesting aspect of adjusted ratings is the absence of serial correlation which is profoundly present in issuer ratings (see Altman an Kao, 1992). To predict future issuer migrations the monitoring of outlook ratings can replace the monitoring of recent trends in issuer migrations. 4

5 This paper focuses on the properties of outlook ratings and adjusted issuer ratings by their outlook ratings. The study is carried out with Moody s Outlook and Watchlist data for their issuer ratings. We are not aware of a reason why empirical results and conclusions presented here should not apply for outlook ratings of Standard&Poor s and Fitch as these agencies have disclosed a similar policy on outlook and rating reviews. Discussions and conclusions are therefore generalized to all outlook ratings for issuer ratings of Moody s, Standard and Poor's and Fitch, unless explicitly stated otherwise. In our study we first model the issuer rating scale and the secondary outlook rating scale with credit scoring models using the (ordered) logit regression methodology. In the same setup we model default probabilities for various prediction horizons. In terms of sensitivity to the temporary credit risk component, i.e. credit cycle sensitivity, issuer ratings and outlook ratings show up at both extremes of the spectrum. Issuer ratings are even adversely sensitive to credit cycles, while outlook ratings follow closely the temporary credit risk component, especially at the downside. Outlook ratings, especially rating reviews, are driven by events and breaks in trends. Second, we investigate the outlook migration policy conditional to an issuer migration event. The outlook migration policy is driven by the prime objective to signal upcoming issuer migration effects. Insight in outlook migration policy is obtained by benchmarking the dynamics of actual outlook ratings by the dynamics of simulated outlook ratings. Simulated outlook ratings are based on credit scores of outlook prediction models. Comparing the dynamics between actual outlook ratings and simulated outlook ratings shows that actual outlook migrations are heavily concentrated before and at the issuer migration event, more than one would expect from a pure credit risk perspective. All detected deviations can be ascribed to the outlook migration policy. Third, we adjust issuer ratings by their outlook ratings and their simulated outlook ratings. For that purpose the secondary outlook rating scale is linked to the primary issuer rating scale by the credit scores of outlook prediction models. After adjusting issuer ratings by their actual outlook ratings migration probabilities increase, rating drift disappears and default prediction performance improves significantly, i.e. the accuracy ratio increases by 3.7% for a one year prediction horizon. Adjusting issuer ratings by their simulated outlook ratings shows a larger potential to improve default prediction performance. We show that the lower level of credit risk information in actual outlook ratings is not caused by the outlook migration policy. The outlook migration policy does not severely override credit risk information in outlook ratings. Perhaps the assignment of outlook ratings is not guided by well defined credit risk standards as a safeguard for comparability. In contrast to the primary issuer rating scale standardization of credit risk information in the secondary outlook rating scale has not been an issue so far. In their definitions of outlook ratings agencies only mention the objective to signal upcoming issuer migration events. As a result actual outlook rating scale is more diffuse in terms of short term credit risk than the simulated outlook rating scale, which is standardized by the ranking of credit scores. 5

6 Once we have parameterized the credit risk perspective and migration policy of outlook ratings we are in a position to judge whether outlook ratings have the potential to bridge the gap between the through-the-cycle perspective and the short-term point-in-time perspective which most banks - and perhaps most investors have. Our main conclusion is that the absence of credit risk standardization in the outlook rating scale and the adverse credit cycle sensitivity of issuer ratings prevent actual outlook ratings to bridge the gap. Instead, issuer ratings adjusted by their outlook ratings reflect a more moderate version of a through-the-cycle perspective. Adjusted issuer ratings will switch to a long term point-in-time perspective after standardizing the credit risk signaled by outlook ratings. This paper is organized as follows. Chapter 2 elaborates on outlook rating definitions provided by agencies, the agencies through-the-cycle rating methodology and literature on outlook ratings. Chapter 3 examines the credit risk perspective of outlook ratings. Chapter 4 explores the influence of the outlook migration policy on outlook rating dynamics. Chapter 5 reports the default prediction performance and dynamic properties of adjusted issuer ratings. Chapter 6 draws conclusions. 6

7 II Definitions and literature on outlook ratings and through-the-cycle methodology 2.1 Agencies definition of outlook ratings In their guide to ratings, rating process and rating practices, Moody s describes the meaning of outlooks and rating reviews as follows: A Moody s rating outlook is an opinion regarding the likely direction of a rating over de medium term, typically 18 to 36 months... A RUR (rating(s) under review) designation indicates that the issuer has one or more ratings under review for a possible change and thus overrides the outlook designation... Moody s uses the Watchlist to indicate that a rating is under review for possible change in the short term, usually within 90 days... A credit is removed from Watchlist when the rating is upgraded, downgraded or confirmed... (Moody s, 2004). In a similar way S&P explains the meaning of their outlooks and rating reviews on CreditWatch (2005): A Standard & Poor s rating outlook assesses the potential direction of a long term credit rating over the intermediate term (typically six months tot two years). In determination a rating outlook consideration is given to any changes in economic and/or fundamental business conditions... Credit Watch highlights the potential direction of a short- or long term rating. It focuses on identifiable events and short term trends that cause ratings to be placed under special surveillance by the Standard &Poor s analytical staff. These may include mergers, recapitalizations, voter referendums, regulatory action or anticipated operating developments. Ratings appear on CreditWatch when an event or a deviation from an expected trend occurs and additional information is necessary to evaluate the current rating...such rating reviews are normally completed within 90 days, unless the outcome of a specific event is pending... (Standard & Poor s, 2005). Fitch outlooks indicate the direction a rating is likely to move over a one to two-year period. Fitch places their long-term credit ratings on a Rating Watch to notify investors that there is a reasonable probability of a rating change and the likely direction of such change. These are designated as Rating Watch is typically resolved over a relatively short period. (Fitch, 2005). According to these definitions outlooks indicate likely issuer migrations in the medium term. In their definitions agencies appear to have slight different notions what they mean by medium term. For example Moody s notion of medium term is one year longer than the notion of S&P. However these horizons are only indications. A fine average is one to two years. Outlooks and rating reviews indicate different time horizons at which an issuer migration might occur. Outlooks signal likely issuer migrations in the medium term. In their definitions agencies appear to have slight different notions what they mean by medium term. For example Moody s notion of medium term is one year longer than the notion of S&P. However these horizons are only indications. A fine average is one to two years. For rating reviews Standard & Poor s and 7

8 Moody s are more explicit on the time horizon: 90 days. Fitch speak about a relative short period. Although not mentioned explicitly in the definitions, outlooks and rating reviews differ in the likelihood an issuer migration might occur. Descriptions to define issuer migration probabilities signaled by rating reviews are slightly stronger than those for outlooks. For example Fitch uses likely for outlooks and reasonable probability for rating reviews. Moody s and Standard&Poor s implicitly enhance the possibility of an issuer migration in their definitions by emphasizing the urgency of the situation when an issuer rating is placed under review. Historical data on Moody s outlook ratings shows that 2/3 of the rating reviews will be followed by an issuer migration with indicated sign, while 1/3 of the outlooks will ultimately result in an issuer migration with indicated sign. In terms of expected issuer migration probability and urgency rating reviews can be interpreted as stronger versions of outlooks. Only Standard&Poor s (2005) provides some insight in the trigger setting issuer ratings on an outlook or a rating review. According to Standard&Poor s rating reviews are triggered by events or sudden changes in expected trends, which requires a formal review procedure in the short term. Standard&Poor s defines outlooks as a response to changes in economic and fundamental business conditions. Our interpretation of this definition is that agencies are aware of developing changes in the medium term, but judge them not to be severe enough yet to consider an issuer migration. 2.2 Trough-the-cycle rating methodology A well accepted explanation for the inadequate timeliness of ratings is the through-the-cycle methodology, which agencies apply in their rating assessment. This methodology has two aspects: first, a focus on the permanent credit risk component of default risk and ignorance of the temporary credit risk component and, second, a prudent migration policy. The first aspect of the through-the-cycle rating methodology is the ignorance of short-term fluctuations in default risk. By filtering out the temporary credit risk component, issuer ratings measure exclusively the permanent, long-term and structural credit risk component. According to Cantor and Mann (2003) the through-the-cycle methodology aims to avoid excessive rating reversals, while holding the timeliness of issuer ratings at an acceptable level: "If over time new information reveals a potential change in an issuer's relative creditworthiness, Moody's considers whether or not to adjust the rating. It manages the tension between its dual objectives - accuracy and stability by changing ratings only when it believes an issuer has experienced what is likely to be an enduring change in fundamental creditworthiness. For this reason, ratings are said to look through-the-cycle'. ". Standard and Poor s (2003) is convinced that " the value of it s rating products is greatest when it s ratings focus on the long term and do not fluctuate with near term performance.". 8

9 The second aspect of the through-the-cycle methodology is the enhancement of rating stability by a prudent migration policy. Only substantial changes in the credit risk permanent component result in issuer migrations and, if triggered, ratings are partially adjusted to the actual level in the credit risk permanent component. Although not officially disclosed by agencies, practical evidence of such a prudent migration policy exists. In their announcement to reconsider their migration policy, in January 2002, Moody's provides some insight in their migration policy: "Under consideration are more aggressive ratings changes - such as downgrading a rating by several notches immediately in reaction to adverse news rather than slowly reducing the rating over a period of time - as well as shortening the rating review cycle to a period of weeks from the current period of months". 1 How rating agencies put their through-the-cycle methodology exactly into practice is not clear. Treacy and Carey (2000) describe the through-the-cycle rating methodology as a rating assessment in a worst case scenario, in the bottom of a presumed credit risk cycle. Löffler (2004) explores the through-the-cycle effects on rating stability and default-prediction performance in a quantitative manner by modeling the separation of permanent and temporary components of default risk in a Kalman filter approach. We have taken a different approach to get insight in the agencies through-the-cycle methodology by estimating credit-scoring models with varying sensitivities to the permanent and temporary credit risk components. From a previous study (Altman and Rijken, 2004) we conissuer the exclusive focus of issuer ratings on the permanent credit risk component and the disregard of the temporary credit risk component. We have modeled the migration policy as follows. An issuer migration is triggered if the actual credit risk, as indicated by the permanent credit risk component, exceeds a threshold of 1.8 notch steps relative to the average credit risk level in a rating category. If triggered, ratings are partially adjusted to the actual credit risk level, 60% at the upside and 70% at the downside. In contrast to the through-the-cycle methodology bankers have a point-in-time perspective on corporate credit risk with a time horizon in between one and seven years (see Basle Committee, 2000). Most bankers have a one-year point-in-time perspective on credit risk. It is reasonable to assume that this perspective applies to most other investors as well. The point-in-time perspective looks at the actual credit risk without an attempt to suppress the temporary credit risk component. It weights both the temporary and permanent credit risk component. The relative weight of these two components depends on the horizon in the point-in-time perspective. For a one year horizon the temporary component has a larger relative weight than for a longer horizon. 9

10 2.3 Literature on outlook ratings Two studies by Keeman et al. (1998) and Hamilton and Cantor (2004) describe extensively the properties of outlook ratings in combination with issuer ratings. Hamilton and Cantor (2004) find that outlook ratings contain apart from timing information extra credit risk information. After adjusting issuer ratings by their outlook ratings, default prediction performance improves significantly. The adjustment of issuer ratings is done by adding or subtracting one or two notch steps depending on the type and sign of outlook ratings. In this paper we further explore in more detail the adjustment of issuer ratings by their outlook ratings. So far most studies on outlook ratings are event studies, testing whether outlook rating migrations signal new information to the market. These studies are of interest to find out whether credit risk information provided by agencies by their rating actions is timely. Recent surveys reveal that investors are not satisfied with the timeliness of ratings (Association for Financial Professionals, 2003, Ellis (1998) and Baker and Mansi (2001)). The outcome of these surveys might suggest that announcements on issuer ratings do not disclose new information to the market. However, Holthausen and Leftwich (1986) and Hand, Holthausen and Leftwich (1992) report significant abnormal stock returns and abnormal bond returns in response to issuer downgrade announcements. On the upside their results are less reliable. A similar asymmetric announcement effect is reported by Steiner and Heinke (2001) for bond returns and by Hull, Predescu and White (2004) for quotes in the credit default swap market. According to these studies, rating reviews for a downgrade negatively impact bond returns, stock returns and quotes in credit default swaps. Results are mixed for rating reviews for an upgrade. Hand, Holthausen and Leftwich (1992) find significant excess bond returns at both the downside and the upside, while Steiner and Heinke (2001) find more significant response at the downside. 10

11 III Credit risk information signaled by outlook ratings 3.1 Data and outlook rating statistics Data on Moody's issuer ratings is obtained from the July 2005 version of the Moody's DRS database, which includes all corporate credit rating revisions and default events in the period January May An extended version of the outlook dataset has been made available to us by Moody s. This database includes all outlook ratings provided by Moody s for their issuer ratings in the period September February In 1991 Moody s started to provide information on rating reviews (DOWN, UP) followed in 1995 by outlook information (positive POS, stable STA and negative NEG) as well. All five outlook rating categories are available since Therefore our analysis covers the January 1995 December 2004 period. In addition NOA outlook ratings (No Outlook Available) are included in the database if the provision of outlook information is (mostly temporarily) suspended. Not all issuers rated by Moody s are included in our analysis. For benchmarking purposes issuer ratings are linked with accounting and market data from COMPUSTAT. In order to ensure consistency in accountancy information only non-financial US issuers are selected and issuers with sufficient accounting and market data available in COMPUSTAT. This selection of issuers reduces the number of issuer-monthly observations from to 71962, including the NOA outlook ratings. When excluding the NOA outlook rating observations are left. The factor 7 in data reduction does alter the outlook rating distribution substantially. Along the different selection steps, as outlined in Table I, the fraction of DOWN and POS increase by almost a factor 2 and the fraction of NOA is reduced by a factor 1.5. As these factors do not represent a fundamental change in outlook rating distribution, we believe the conclusions of this study will largely hold for non-us, private and financial issuers as well. In the first years after the introduction of outlooks by Moody s the percentage of negative, stable and positive outlook ratings is steadily increasing (see Table I). After 1998 the outlook rating distribution becomes relative stable up to In 2004 almost all NOA (no outlook available) ratings seem to be converted to stable outlooks. Outlook rating distributions varies among issuer rating categories. Most issuers in the Caa category (including Ca and C ratings) have a negative rating outlook. Aaa rated issuers have obviously no POS- and UP ratings. Notable is the relative small fraction of DOWN and NEG ratings in the Aaa rating category. DOWN ratings are more likely to appear for investment grade categories while POS and UP ratings are more present in speculative grade categories. 3.2 Specification of credit scoring models 11

12 In order to reveal the character of credit risk information signaled by outlook ratings we benchmark the characteristics of outlook prediction models with issuer rating prediction models and default prediction models. All default prediction models are estimated by the following logit regression model in a panel data setting CS WK = α + β1 TA + β Size 5 + β SD(AR) 6 RE + β 2 ln(1 TA 7 + β AR EBIT ) + β 3 (1 TA + ε ME ) + β 4 (1 + ln BL ) (3.1) E( p i, t 1 ) = 1 + exp( CS ) (3.2) i, t CS is the credit score of issuer i at time t, E(p ) is the expected probability of default of issuer i at time t, WK is net working capital, RE is retained earnings, TA is total assets, EBIT is earnings before interest and taxes, ME is the market value of equity, and BL is the book value of total liabilities. Size is the log-transformation of total liabilities normalized by the total value of the US equity market Mkt: ln(bl/mkt). The abnormal stock return AR t is stock return relative to equal weighted market return in the twelve months preceding t. SD(AR) t is the standard deviation in monthly abnormal returns in the twelve months preceding t. The parameters of the logit regression model α and β are estimated by a standard maximum likelihood procedure. This estimation procedure seeks for an optimal match between the actual outcome p and the expected outcome of the model E(p ). p = 0 when issuer i defaults before t + T and p = 1 when issuer i survives beyond t + T. Default prediction models are estimated for various time horizons T. An SDP model is estimated for T = 6 months, an DP1 model for T = 12 months, an DP2 model for T = 24 months and an TTC model for T = 6 years. Credit scores of these models are point-in-time measures of credit risk, giving weight to both the permanent and temporary credit risk component. In addition, a through the cycle default prediction model is estimated. These models focus exclusively on default probability in a specific future period, i.e. the permanent credit risk component, and the binary variable p is set to 0 only for issuers defaulting in this future period (t + T 1, t + T 1 + T). Default events in the near future (t, t + T 1 ) are ignored by leaving out the observations of issuers defaulting in this period. An alternative, including these observations in the analysis by setting p = 1 for these observations, does not change the estimate significantly, since the number of defaults is relative small. Also a variation of T 1 beyond three years and a variation in T conditional to T 1 beyond three years do not alter the estimates considerably. A TTC model is estimated for T 1 = 5 years and T = 1 year. Credit scores of the TTC model are 12

13 through the cycle measures of credit risk, aimed only to be sensitive to the permanent credit risk component. The issuer rating prediction model (IRP model) models the discrete issuer rating scale N with an ordered logit regression model in a panel data setting. In this model, the credit score IRP is an unobservable variable IRP WK = α + β1 TA + β Size 5 + β SD(AR) 6 RE + β 2 ln(1 TA 7 + β AR EBIT ) + β 3 (1 TA + ε ME ) + β 4 (1 + ln BL ) (3.3) The IRP score is related to the issuer rating N as follows y = R if B < IR B (3.4) N 1 N where k is one of the issuer rating categories, y is the actual issuer rating, B N is the upper boundary for the IRP score in rating category N, B 0 = - and B 18 =. In the estimate we the following 16 issuer rating categories N: Aaa, Aa1, Aa2, Aa3, A1, A2, A3, Baa1, Baa2, Baa3, Ba1, Ba2, Ba3, B1, B2, B3, Caa and Ca. In order to have a reasonable number of observations in each rating category, the issuer rating categories Caa3, Caa2 and Caa1 are combined into a single rating category Caa and the issuer rating categories C and Ca are combined to Ca. In the ordered logit model, the probability that y equals k is specified by P(y = N) = F(B IRP ) F(B IRP ) (3.5) N N 1 i, t where F is the cumulative logistic function. Parameters α, β, and B k are estimated with a maximum likelihood procedure. This estimation procedure seeks for an optimal match between the actual rating y and the expected outcome of the model P(y = N). The outlook prediction model (OP model) is estimated following the same ordered logit regression methodology. Instead of 18 issuer rating categories N, five outlook rating categories O are modeled: watch down DOWN, negative outlook NEG, stable outlook STA, positive outlook POS and watch up UP. Outlook rating scale is a secondary rating scale on top of the issuer rating scale. Outlook ratings are relative credit risk measures within a rating category N. Therefore the seven model variables X in equation 3.3 are replaced by their differentials X relative to their mean values X N,t of all issuers in a rating category N at time t X X N,t X N, = with i N (3.6) SD(X) N,t 13

14 SD(X) N,t is the cross sectional standard deviation of model variable X for all issuers in rating category N at time t. Normalization of X by SD(X) makes X better comparable among different issuer rating categories N. 3.3 Parameters estimates of default prediction models and issuer prediction models At the end of each month corporate ratings are linked to stock price data and accounting data. Accounting data is assumed to be widely publicly available three months after the end of the fiscal year. The resulting panel dataset covers the period April December 2004 and includes issuer-month observations from the time series of 2239 issuers with durations in between 1 and on 273 months and an average duration of 85 months. In April 1982 Moody s started to refine their issuer ratings by adding 1, 2 and 3 modifiers. Table II reports the estimated parameters α and β i of five default prediction models (SDP, DP1, DP2, LDP, TTC) and the issuer rating prediction model (IRP). The estimation period of each model is April 1982 until December 2004 minus the time horizon. All credit scoring models employ the same model variables. This allows a fair comparison of the relative weights RW k of the model variables k RW k = 6 β σ j= 1 j j k β σ j (3.7) β k is the parameter estimate for model variable k, and σ k is the standard deviation in the pooled sample distribution of model variable k in period Table II shows the RW values for the estimated credit scoring models. Market equity information has a large share in the explanation of short term defaults. The ME/BL variable dominates the SDP model with a RW value of 37.2 %. Short term dynamics of stock returns measured by AR and SD(AR) have a relative weight of respectively 16.5% and 10.6%. Although equity information is most important, accounting information and Size add substantially to the explanation of the default incidence. Prediction horizon has a significant impact on the relative weight of the model variables. Especially for the RE/TA, ME/BL, and Size variable, a clear shift in relative weight is observed in the SDP, DP1, LDP, TTC model, in that order of sequence. Not surprisingly, the short-term oriented SDP model depends heavily on variables which follow most closely the credit / business cycle, like ME/BL and the trend variable AR, while the TTC model place relatively more weight 14

15 on the RE/TA and Size variables. As expected the TTC model is insensitive to credit cycle indicators EBIT/TA, SD(AR) and AR. The relative weight of the model variables in the IRP model matches most closely the TTC model. Both models represent a through-the-cycle perspective, aiming to be insensitive to the temporary credit risk component. However the TTC and IRP model do not fully match. The agencies through-the-cycle perspective puts less weight on WK/TA and ME/LIB but more on the uncertainty level, as proxied by SD(AR), and Size. The negative parameter for AR suggests that issuer ratings are slightly countercyclical. It appears that agencies apply a slightly more prudent through-the-cycle perspective than indicated by the TTC model. The RW values of the TTC model are more close to the LDP model, the long term point-in-time perspective than the IRP model, the agencies through-the-cycle perspective. Issuer ratings appear to switch to a point-intime LDP perspective when issuers are highly distressed. The IRP model is re-estimated for issuers rated as Ca, Caa or B3 or issuers approaching a default event (see Table II). For highly distressed issuers, RW values of the IRP model become largely comparable to the DP1 and LDP models. Default prediction models estimated with the Moody s default dataset are fairly similar to the models estimated with the S&P defaults listed in CREDITPRO (see Altman and Rijken, 2004). Because the Moody s issuers ratings differ only by 1 2 notch steps at most from S&P rating equivalents the IRP model estimated with Moody s issuer ratings is almost an exact replication of the IRP model estimated with S&P ratings. Although the Moody s default definition differs from the Standard&Poor s default definition, the DP models estimated with Moody s default events are as good as equal to DP models estimated with Standard&Poor s default events 2. In contrast to the S&P default definition Moody s counts delayed payments made within a grace period and explicitly counts issuer files for bankruptcy (Chapter 11 and Chapter 7) and legal receivership. Extensive robustness checks have been carried out to test for sector influence, time period and default definition. Parameter estimates do not change substantial after splitting the dataset in two periods and (see also Altman and Rijken, 2004). Default prediction models are only a little sensitive the definition of default (bankruptcies, Moody s default definition and Standard&Poor s definition). The parameter estimates of the TTC model do not change substantially when varying T 1 in between 5 and 8 years and allowing T only to vary in between one and two years. A specific test for the IRP model shows the robustness of the estimated parameters to a split of observations into non-investment graded (Ba1 issuers down to B2 issuers) issuers and investment graded (Baa3 and above) issuers. 3.4 Parameters estimates of the outlook prediction model Table III reports the parameter estimates of various outlook prediction models. A basic OP model is estimated with all issuer-monthly observations in the period. At first sight 15

16 the RW values of the OP outlook prediction model is close to the short term default prediction SDP model, suggesting that on average the outlook rating scale is a measure of short term pointin-time credit risk. To explore whether the entire outlook rating scale has a short term point-in-time perspective, outlook prediction models are estimated for four outlook rating subscales, each of them only including DOWN, NEG, POS or UP outlook ratings and all of them including stable outlook ratings. Parameter estimates vary considerably between these four subscales, so the outlook rating scale appears not to have a uniform credit risk perspective. As expected, rating reviews are more sensitive to trends on the stock market than outlooks. Rating reviews are responses to sudden changing circumstances (events) while outlooks signal the likely direction of any medium term rating actions. However, the asymmetry between the downside (DOWN and NEG) and the upside (POS and UP) is less straightforward. At the downside outlook ratings depend more on SD(AR) and ME/LIB, while at the upside outlook ratings are more related to variables Size and RE/TA (with a negative sign) which are less sensitive to credit cycles. Apparently, the downside of the outlook rating scale is more driven by credit risk cycles. Other tests examines whether OP model parameters are sensitive to time period, issuer rating category and occurrence of an issuer migration event in the near past or future. Results are reported in Table III. The OP model is robust to issuer rating category. Even in the extreme case of CCC issuer rating category RW values are comparable to investment graded and other speculative graded issuers. Just before and just after an issuer migration event the outlook rating scale is more sensitive to short term trends in stock prices at a cost of the ME/LIB variable. The same exchange is observed between time periods and More recently the outlook rating scale depends less on equity trends and more on the level of market leverage. In the OP model estimation rating reviews have a relative low weight because of their low share in the outlook rating distribution (see Table I). To check for the influence of the outlook rating distribution a weighted ordered regression model OPW is estimated which gives an equal weight to all five outlook categories in the model estimation. Between the OP and OPW model RW values differ by 7.5% at most. In benchmark study we use various outlook rating prediction models - OP, OP(-), OP(+) and OPW model - to account for the multidimensional character of the outlook rating scale. The OP model represents the basic model, the OPW model adjusts for the unequal outlook rating distribution - giving relative more weight to the upside - and the OP(-) and OP(+) models represent the different character of the outlook rating scale between the downside and upside. 3.5 Weighting the permanent and temporary credit risk components in credit scoring models In a first stage credit scoring models are estimated using seven model variables. In a second stage credit scoring models are re-estimated using only two model variables X P and X T, proxies for the 16

17 permanent and temporary credit risk component. The permanent credit risk component X P is proxied by TTC model scores. The TTC model explains exclusively long term defaults and is insensitive to short term credit cycle indicators (see section 3.1 and 3.2). The temporary credit risk component X T is proxied by the difference in SDP model scores and TTC model scores. Because of its short six month horizon the point-in-time SDP model is most sensitive to credit risk cycles. As SDP model scores are strongly correlated with TTC scores (correlation coefficient of 0.49) SDP model scores follows both the permanent and temporary credit risk component. To obtain a pure proxy for the temporary credit risk component TTC scores are subtracted from SDP scores. Credit scoring models are re-estimated using X P and X T. Outlook prediction scoring models are re-estimated after converting X P and X T to their differentials following equation 3.6. Results are presented in Table IV. The weight of the permanent credit risk component in the TTC model is obviously 100%. In the SDP model 48.6% of the variations in credit risk are explained by the permanent credit risk component. Increasing the horizon in point-in-time credit risk perspectives from six months to six years reduces the weight to the temporary component substantially from 51.4% to 20.9%. Issuer ratings follow almost exclusively the permanent credit risk component, consistent with the objective of the through-the-cycle methodology to filter out the temporary credit risk component. For the entire issuer rating scale only a small countercyclical effect is observed by the negative parameter for X T. As noticed earlier in section 3.2 the credit risk perspective at the low end of the issuer rating scale is different. The temporary component has a 42.4% weight in the issuer rating scale for B3 rated issuers and below, comparable to a one-year point-in-time perspective. The outlook rating scale is on average more sensitive to the temporary component than the SDP model. However a more diverse picture shows up, when estimating outlook prediction model for four subscales, each of them only including DOWN, NEG, POS or UP outlook ratings and all of them including stable outlook ratings. At the downside outlook ratings are sensitive to the temporary credit risk component, excessively compared to the SDP model. The DOWN review, driven by events - according to the Standard&Poor s definition is most responsive to credit risk cycles. The prediction model for positive outlooks has a short term credit risk perspective comparable to the SDP model. UP outlook ratings have a complete different perspective. They are only sensitive to the permanent credit risk component. Comparing pseudo R 2 values between first stage and second stage credit scoring models gives insight to what extent information is lost by replacing the seven model variables by X P and X T. Pseudo R 2 indicates the goodness of fit by the (ordered) logit regression model. As expected the pseudo R 2 of the default prediction models is hardly affected by replacing the seven model variables by X P and X T. In case of the IRP model the pseudo R 2 is reduced by 20%. For the DOWN outlook prediction model pseudo R 2 is only lowered by 5%. For the NEG, POS and UP models the reduction in pseudo R 2 is larger respectively 25%, 40% and 60%. Interestingly, the 17

18 pseudo R 2 of the OP model and OW model which model the common factors for all five outlook ratings - is only reduced by 10% and 15%. This implies that the common factors in the outlook rating scale are largely captured by X P and X T. The larger reduction in pseudo R 2 for the POS model and especially the UP model compared to NEG and DOWN models, suggests that outlook ratings at the upside are also driven by factors other than corporate credit risk. 3.6 Linking the outlook rating scale to a notch rating scale Credit risk variations within an issuer rating category N can be measured by variations in credit scores CS CS CS N,t CS N, = (3.8) γ N,t CS N,t is the average credit score for all issuers in rating category N at time t. CS is converted to notch rating scale by γ N,t. This scaling factor γ N,t is obtained as follows 1. For default prediction model scores and IRP model scores the scaling factor γ N,t reflects the slope between the numerical issuer rating scale N and the average credit model scores CS in rating categories N. The numerical issuer rating scale runs from Ca/C = 0, Caa = 1, B3 = 2, B2 = 3, up to Aaa = 18. This numerical rating scale is an arbitrary, but quite intuitive, choice that is commonly found in the mapping of bank internal-rating models to agency ratings. Roughly three groups of rating categories can be distinguished with a close to linear relationship between CS and N: N [1..4], N [5..10] and N [ ]. For each of these groups and for each month γ N,t are derived. 2. For OP, OWP, OP(+) and OP(-) model scores the scaling factor γ N,t is obtained by the average shift in these scores when the differentials X are calculated relative to their mean values X in a rating category one notch step lower or one notch step higher (see equation 3.6). So only X N,t is replaced by X N+1,t or X N-1,t and credit scores are recalculated keeping the α and β parameters of the outlook prediction models fixed. Note that the scaling factor γ N,t does not convert credit model scores to an issuer rating scale in a sense that it removes the temporary credit risk component. The scaling factor converts credit model scores to a notch rating scale. The only purpose to scale by γ N,t is to standardize various credit model scoring scales. In addition to OP scores and OW scores a combined score OC is defined. The OC score is a combination of the two outlook prediction model scores OP(-) and OP(+). It aims to combine the different character of the outlook rating scale at the downside and the upside in a single score. The OC score is defined by the following algorithm. If both OP(-) and OP(+) scores are negative than both models classify an observation at the downside and OC follows 18

19 the downside model OP(-) score. Following the same reasoning OC equals OP(+) if OP(- ) and OP(+) scores are positive. If the OP(-) score is negative and the OP(+) score is positive than the OC score equals the score with the largest absolute value. In the other case both scores have the opposite sign and OP(-) and OP(+) are actually on the wrong side the smallest absolute value is decisive. Table V reports the average CS values for five outlook rating categories. By definition the average CS values for all outlook ratings is zero. Based on OP model scores, credit risk of issuers with DOWN ratings is centered 2.7 notch steps lower than indicated by their issuer rating class N. NEG, STA, POS and UP rated issuers are centered at respectively -1.5, 0.3, 1.5 and 1.7. The average OPW and OC model scores for the five outlook ratings are comparable for all outlook ratings except for UP ratings. OPW and OC model scores make a more clear distinction between POS and UP ratings. The different character of UP ratings is allowed to show up more pronouncedly by combining two scales with different character (OC model) and by heavily weighting the upside in the outlook prediction model estimation (OPW model). As indicated by OC model scores, outlook ratings diverge 1.5 notch steps from the centre and rating reviews diverge 3 notch steps from the centre. These numbers are consistent with the adjustments Hamilton and Cantor (2004) apply in their search for an optimal adjustment of issuer ratings by their outlook ratings (see also chapter 5). Conditional to an issuer migration event the outlook rating scale seems to be stretched both at the downside and the upside (see Figure I). This is due to the increased cross sectional variance in credit risk conditional to an issuer migration event. Conditioning to an issuer migration event selects issuers with a permanent credit risk component deviating from the mean in a particular issuer rating category, either to the downside or the upside. Additionally, this dispersion in permanent credit risk component is further enhanced by a positive correlation between the permanent credit risk component and the temporary credit component. Compared to through-the cycle credit risk measures, point-in-time credit risk measures exhibit a larger cross sectional variation conditional to an issuer migration event. Because of their largest sensitivity to the temporary credit risk component this cross sectional variation is highest for OP and OC model scores. Of all default prediction models a best match in linking outlook ratings to a notch scale is found for the SDP model. This empirical finding emphasizes once more the short term point-in-time character of outlook ratings. Point-in-time measures with longer time horizons (DP1, DP2, and LDP scores) and through-the-cycle measures of risk (IRP scores) detect less credit risk dispersion in the outlook rating scale. Just like the OP model the SDP model is not capable to capture the different character of UP ratings properly. However from default prediction performance analysis (see chapter 5) we conclude that UP ratings do not contain more credit risk information than indicated by OP model scores. So, in terms of measuring corporate credit risk the OPW and OC 19

20 models, which give more weight to UP ratings, are of no added value compared the OP model. For this reason we use the OP model to benchmark outlook rating dynamics in next chapter 4. 20

21 IV Agencies migration policy for outlook ratings 4.1 Outlook rating simulation The prime objective of outlook ratings is to signal upcoming issuer migration events. Timing information is included in outlook ratings by the timing of outlook migrations, following a particular outlook migration policy. Insight in this policy is obtained by benchmarking the dynamics of actual outlook ratings by the dynamics of simulated outlook ratings. Simulated outlook ratings are based on the ranking of OP model scores and a simple migration policy with one threshold level. Simulated outlook ratings have no timing objective and their dynamics are only driven by a variation in corporate credit risk as measured by OP model scores within a rating category N. Outlook migration policies are revealed from differences between simulated and actual outlook rating dynamics. Outlook OP scores are converted to OP outlook ratings, equivalent to actual outlook ratings by the following three step procedure. Step 1: Modification of OP scores For each issuer the monthly time series of OP scores is is replaced by a time series of modified OP M scores. At the beginning of the time series at t = 0 OP M 0 is set equal to OP 0. As long as the OP score stays within the threshold interval (OP M 0 - γ TH, OP M 0 + γ TH) OP M remains constant. As soon as the OP score jumps out of the interval at t = T the OP M T is set equal to OP T and the threshold interval is adjusted to (OP M T - γ TH, OP M T + γ TH). This repeats until the end of the time series of a particular issuer in the dataset. The threshold parameter TH is expressed in notch steps, the scaling factor γ N,t converts OP model scores to a notch rating scale. Step 2: Conversion of OP M scores to equivalent outlook ratings At the end of each month all issuers in a specific rating category N which have a DOWN, NEG, STA, POS or UP rating are ranked by their OP M model score. On the basis of this ranking five equivalent outlook ratings OP(TH) are assigned to the individual issuers. As a result at the end of each month and for each rating category N the distribution of simulated outlook ratings OP(TH) matches exactly the distribution of actual outlook ratings. Step 3: Correction for non-intended outlook rating transitions The time-series of OP M scores is an irregular pattern of upward and downward jumps. Only jumps in OP M scores may trigger an outlook migration. This constraint is not guaranteed by step 2. An outlook migration may happen even if no jump occurs in the OP M score. These nonintended migrations are removed by replacing OP(TH) rating by their one-month lags if OP M t score equals the one-month lagged OP M t-1 score. As a consequence, the distribution of the OP(TH) ratings is slightly altered. The number of observations in each outlook rating category 21