EU Banks Rating Assignments: Is There Heterogeneity Between New And Old Member Countries?

Transcription

1 Department of Economcs and Fnance Workng Paper No Economcs and Fnance Workng Paper Seres Guglelmo Mara Caporale, Roman Matousek and Chrs Stewart EU Banks Ratng Assgnments: Is There Heterogenety Between New And Old Member Countres? October

2 EU Banks Ratng Assgnments: Is There Heterogenety Between New And Old Member Countres? Guglelmo Mara Caporale 1 Roman Matousek 2 Chrs Stewart 3 October 2009 Abstract We model EU countres bank ratngs usng fnancal varables and allowng for ntercept and slope heterogenety. We fnd that country-specfc factors (n the form of heterogeneous ntercepts) are a crucal determnant of ratngs. Whlst new EU countres typcally have lower ratngs than old EU countres, after controllng for fnancal varables, all countres are found to have sgnfcantly dfferent ntercepts, whch confrms our hypothess. Ths ntercept heterogenety may reflect dfferences n country rsk and the legal and regulatory framework that banks face (such as foreclosure laws). In addton, ratngs may respond dfferently to the lqudty and operatng expenses to operatng ncome varables across countres: typcally ratngs are more responsve to the former and less senstve to the latter for new EU countres compared wth old EU countres. Keywords: EU countres, banks, ratngs, ordered choce models, ndex of ndcator varables JEL Classfcaton: C25, C51, C52, G21. We wsh to acknowledge helpful comments and suggestons from partcpants n the CICM Conference on 20 Years of Transton n Central and Eastern Europe: Money, Bankng and Fnancal Markets, London Metropoltan Busness School, London, September Centre for Emprcal Fnance, Brunel Unversty, West London, UB8 3PH, UK. Emal: Guglelmo- Mara.Caporale@brunel.ac.uk 2 Centre for Internatonal Captal Markets, London Metropoltan Busness School, London Metropoltan Unversty, 84 Moorgate, London, EC2M 6SQ. Tel: E-mal: r.matousek@londonmet.ac.uk. 3 London Metropoltan Busness School, London Metropoltan Unversty, 84 Moorgate, London, EC2M 6SQ. Tel: E-mal: c.stewart@londonmet.ac.uk.

3 1. Introducton Ratngs of banks and companes conducted by External Credt Assessment Insttutons (ECAIs) may be seen as nstruments that provde nvestors wth prma face nformaton about the fnancal poston of the subject n queston and on the prce of credt rsk. Ratngs are ordnal measures that should not only reflect the current fnancal poston of soveregn natons, frms, banks, etc. but also provde nformaton about ther future fnancal postons. The objectve of our paper s to analyse the determnants of ndvdual bank ratngs conducted by Ftch Ratngs (FR) and to nvestgate whether the country of orgn matters for ndvdual ratngs. For ths purpose, we frst consder whether (and whch of) the key fnancal ratos of banks reflect ndvdual ratngs (that s, accordng to FR, a key component for long- and short-term ratng). Second, we examne whether bank ratngs are systematcally determned by the country orgn of commercal banks. One hypothess s that FR mght assgn hgher ratngs to commercal banks from old EU countres that have the same fnancal poston as those from new EU countres. Ths could reflect dfferences n country rsk (gven that bank ratngs cannot exceed soveregn ratngs) or dfferences n legal and regulatory factors (ncludng ther enforcement), such as foreclosure laws. Another hypothess s that FR mght set ratngs dfferently for old and new EU countres n terms of response to fnancal factors. That s, the coeffcents on fnancal varables n a regresson explanng ratngs may be dfferent for old and new EU countres. In other words, we test f commercal banks from new EU countres are assgned ratngs on the bass of ther fnancal ratos n the same way as old EU countres or f other factors are consdered. To ths end, we ncorporate new EU and country-specfc ndcator varables to capture heterogeneous varatons n ratngs under that ratonale that a bank s ratng s related to the country n whch t s based. As country-specfc ndcators we construct ndex-of-ndcator varables that are n the sprt of the method appled n Hendry (2001) and Hendry and Santos (2005), although we extend t to allow heterogeneous slopes. Ths methodologcal approach has recently been proposed by Caporale et al. (2009) and represents a novel contrbuton n the context of modellng bank ratngs. We also assess the predctve power of our model to classfy the ndvdual ratngs of the commercal banks n queston. The ablty to predct the fnancal soundness of banks, corporatons and soveregn countres has been of central mportance for analysts, regulators and polcy makers. A large number of studes have employed fnancal ratos to predct falures of ndvdual frms 1

4 (banks), for example, Altman et al. (1977) and Ohlson (1980). Models that predct bank falures usng so-called Early Warnng Systems (EWS) have appeared n a number of studes, ncludng Mayer and Pfer (1970), and Kolar et al. (2002). Wthn ths context, the fnancal varables of commercal banks have been utlsed n several ways. Yet the ablty of ECAIs to assgn ratngs correctly has been extensvely questoned (Altman and Saunders, 1998, Levch et al., 2002, Altman and Rjken, 2004, Amato and Furfne, 2004, Portes, 2008). One of the most frequent arguments about the predcton abltes of ratng agences (RAs) s that they could provde msleadng nformaton snce the analyss s backward- rather than forward-lookng. In addton, the low transparency of ratngs assgnments contrbutes to the concern over the accuracy of ratngs. Further, ECAIs do not have, and cannot have, superor nformaton to market partcpants about uncertanty and the degree of nsolvency (llqudty) of companes. By modellng ratngs we seek to dentfy ther determnants and, usng measures of ft, gauge how transparent ratngs assgnments are. There are numerous studes that predct bond ratngs such as Kamstra et al. (2001), who utlse ordered-logt regresson. Other evdence from recent studes (Km, 2005; Huang et al., 2004 and Lee, 2007) show that artfcal ntellgence methods do not provde superor predctons of bond ratngs compared wth standard ordered-choce methods. Hence, usng ordered logt/probt regressons s a vald way of addressng the man challenge n modellng ratngs, whch s to ncrease the probablty of correct classfcatons. However, we are not aware of any prevous studes that seek to model and predct ndvdual bank ratngs allowng for heterogeneous country effects, whch s the am of ths paper. The organzaton of the paper s as follows. Secton 2 descrbes the data and the methods appled, whle Secton 3 dscusses the prncpal emprcal fndngs. The last secton concludes. 2. Data and Methodology We model the ndvdual ratngs of EU banks as produced by Ftch Ratngs (FR). These ratngs are dvded nto sx man categores (A, B, C, D, E, F) whch, wth ntermedate subdvsons (A/B, B/C, C/D, D/E), gve ten categores of bank performance. We use data on 1168 European banks ratngs, denoted Y, between 1996 and Y s ordnal and has ten categores that are assgned nteger values, 0 to 9: lower values ndcate a lower ratng. The 2

5 ten ratng categores are: F (0), E (1), D/E (2), D (3), C/D (4), C (5), B/C (6), B (7), A/B (8), A (9). We apply ordered-choce estmaton technques to model ths ordnal dependent varable because, as s well known, they are the approprate method to use n ths case. The ordered dependent varable model assumes the followng latent varable form (see Greene, 2008): K * Y = β k X k + u k =1 (1) where X k s the k th explanatory varable for the th bank, u s a stochastc error term, and s the unobserved dependent varable that s related to the observed dependent varable, Y, (assumng ten categores) as follows: * Y Y Y Y = 1 = j = 10 * f Y λ1 f λ < Y f j-1 λ < Y 9 * * λ, j j = 2, 3,..., 9 (2) where λ 1, λ 2,, λ 9 are unknown parameters (lmt ponts) to be estmated wth the coeffcents (the β k s). We are prmarly nterested n the general drecton of correlaton between the dependent and ndependent varables. Therefore, we use the sgn of β k to provde gudance on whether the estmated sgns of the coeffcents are consstent wth our a pror expectatons. Ths s nstead of lookng at the margnal effects whch ndcate the drecton of change of the dependent varable (for each value of the dependent varable) n 3

6 response to a change n nterpret. X k. For ordered-choce models these margnal effects are dffcult to The probt form of ths model assumes that the cumulatve dstrbuton functon employed s based upon the standard normal, whle the logt form assumes a logstc dstrbuton. Greene (2008) suggests that probt and logt models yeld results that are very smlar n practce and so we focus on those from the probt form. The frst explanatory varable that we consder s for the year n whch the ratng was made [ Date ]. Ths s 3 n 1996, 4 n 1997, 5 n 1998 and so on. 4 The second set of covarates consdered s the frst lagged values of the followng seven fnancal varables: the rato of equty to total assets [denoted Equty ], the rato of lqud assets to total assets [ Lqudty ], the natural logarthm of total assets [ ln ( Assets) ], the net nterest margn [ NIM ], the rato of operatng expenses to total operatng ncome [ OE _ OI ], other operatng ncome to total OOIA and the return on assets [ ROA ]. 5 Current values of fnancal varables are not assets, [ ] used as they may contan nformaton not known when the ratng was made. 6 The choce of varables s guded by the past lterature. A thrd set of varables employed are country ndcator (or dummy) varables. Two broad types of ndcators are consdered. Frst, we construct a shft dummy varable, New D, that s defned to take the value of unty for new EU countres and s zero for the 15 old EU countres. 7 Ths dummy varable, multpled by a fnancal varable, Z, yelds the shft n that varable s slope coeffcent for a new EU country, Z = Z D. Second, we New k k New develop ndex-of-ndcator varables that allow each country to have dfferent ntercept and slope coeffcents. However, an ordered-choce model ncorporatng 27 dummy varables for each covarate cannot be estmated; hence, we employ a method that s n the sprt of Hendry (2001) and Hendry and Santos (2005) to construct ndces-of-ndcator varables for each covarate. 4 Orgnally we had data from 1994 where 1994 took the value of 1. However, data pror to 1996 was lost due to mssng observatons on some varables. 5 Some other varables were consdered but were omtted from the analyss due to multcollnearty. 6 For example, f a bank s ratng was decded n January 2007 then the value of any explanatory factor measured over the whole of 2007 would be unknown when the ratng was made. 7 The twelve new EU countres n our sample are: Bulgara, Cyprus, Czech Republc, Estona, Hungary, Latva, Lthuana, Malta, Poland, Romana, Slovaka and Slovena. The ffteen old EU countres are Austra, Belgum, Denmark, Fnland, France, Germany, Greece, Ireland, Italy, Luxembourg, Netherlands, Portugal, Span, Sweden and the UK. 4

7 To construct a country ndex for the ntercept we estmate two probt models, one for new EU countres and one for old EU countres. That s, one probt regresson of ratngs on the 12 new EU countres (ntercept) dummy varables, D, m = 1,2,...,12, s estmated, thus: m 12 ˆ * Y = ˆ δ D m= 1 m m (3) where, δˆ m denotes the respectve estmated coeffcents. The ntal ndex for new EU countres s constructed as the sum of the products of the coeffcents for the sgnfcant varables and ther correspondng dummy varables, thus: 12 N I = ˆ δ D m= 1 1 m (4) Smlarly, the followng ordered-choce model s ftted to the 15 old EU country dummy varables, D, m = 13,14,..., 27 : m 27 ˆ * Y = ˆ δ D m= 13 m m (5) The ntal ndex for old EU countres s correspondngly constructed as: 27 O I = ˆ δ D 1 m= 13 m (6) To obtan a prelmnary ndex for all countres, ratngs are then regressed on these two ndces, thus: ˆ * Y = ˆ γ I N N + ˆ γ I O O (7) The ntal country ndex s constructed as: 5

8 I C = ˆ γ I + ˆ γ I N N O O (8) Ths ndex was checked for approprateness by runnng a sngle regresson that ncluded the ntal country ndex plus one ndvdual country s dummy, that s: ˆ * Y = ˆ λi C + ˆ α D m m (9) If the latter ndvdual dummy varable was sgnfcant the value of ts coeffcent, αˆ m, was ncorporated nto the country ndex. Ths was repeated for all 27 countres, that s, 27 regressons contanng only two varables (the country ndex and a partcular country s dummy) were estmated. After all the coeffcents of the ndvdual country dummes that were sgnfcant n these 27 regressons had been ncorporated nto the ndex ths step was repeated untl no ndvdual country dummes were sgnfcant at the 5% level (when ncluded n a regresson wth the country ndex). The result s the ntercept country ndex reported n Table 4. A modfed procedure was employed to construct ndces for the non-ntercept covarates. For each covarate (except for Date ) a slope nteracton varable, Z C km, was constructed as: Z = Z D C km k m (10) For the k th covarate one regresson s estmated for the new EU countres as ratngs on the fnancal varables, date and the 12 new EU countres slope nteracton term for the k th varable, thus: Yˆ K 12 * = ˆ + ˆ φ k Z k θ m k = 1 m= 1 Z C km (11) A correspondng regresson for the k th fnancal varable s estmated for the group of 15 old EU countres, as: 6

9 Yˆ K 27 * = ˆ + ˆ φ k Z k θ m k = 1 m= 13 Z C km (12) Intal ndces for the k th covarate for new and old EU countres are constructed usng only the statstcally sgnfcant nteracton terms, as: 12 N I = ˆ θ Z k m= 1 m C km (13) 27 O I = ˆ θ Z k m m= 13 C km (14) To obtan a prelmnary ndex of the k th covarate for all countres we regress ratngs on these two ndces, thus: ˆ * Y = ˆ ω I N N k + ˆ ω I O O k (15) The ntal country slope ndex for the k th fnancal varable s constructed as: I C k = ˆ ω I + ˆ ω I N N k O O k (16) Ths ndex was refned by the followng teratve process. A sngle regresson that ncluded the date, the fnancal varables, the ntal country ndex plus one ndvdual country s nteracton term was estmated as follows: Yˆ * K = ˆ φk Z k = 1 k + ˆ ρi C k + ˆ μ Z m C km (17) If the latter ndvdual nteracton term was sgnfcant the value of ts coeffcent, was ncorporated nto the country ndex. Ths was repeated for all 27 countres. After all the coeffcents of the ndvdual country nteracton terms that were sgnfcant n these 27 regressons had been ncorporated nto the ndex ths teraton was complete. Further teratons were repeated untl there was convergence gvng the fnal country slope ndex, 7 μˆ m,

10 CF I k. Complete convergence would be acheved when no Z km term was sgnfcant at the 5% level for any country n (17) n a full teraton. Convergence may also be acheved even f nteracton varables can be added wth sgnfcance between teratons f the change n the ndex s small between teratons (to some tolerance level). We found that 999 teratons was suffcent for all but the lqudty ndex to acheve complete convergence or make the changes between the values n the ndces suffcently small to conclude that they had converged. For the lqudty ndex there s non-convergence such that the ndex s not the same between adjacent teratons but s exactly the same for every other teraton. In ths case we tred both possble ndces for lqudty n our regressons. 8 Plots of the 998 th and 999 th teratons of the ndex for each of the fnancal varables are gven n Fgure 1 to Fgure Emprcal Results The frst set of ordered probt regresson results for the determnants of bank ratngs are presented n Table 1. We report a general model and one favoured parsmonous specfcaton obtaned usng a cross-sectonal varant of the general-to-specfc methodology. 10 When there was ambguty over whch model to favour we selected the model wth the lowest SBC. In all cases the favoured parsmonous models only nclude varables that are ndvdually sgnfcant accordng to z-statstcs and jontly sgnfcant accordng to a lkelhood rato test, denoted LR statstc. The restrctons placed on the general model to obtan the parsmonous model cannot be rejected accordng to a lkelhood rato test [LR(general favoured)]. The favoured parsmonous models wll yeld more effcent nference relatve to the general model and so they are used for nference. The model n the column headed No shft n Table 1 contans no coeffcents that shft for new EU countres (all the coeffcents are the same for all countres). In the favoured 8 Ths happened for the lqudty ndex where for one country, Luxembourg, the value n the ndex could take on one of two values: or We used the ndex that produced the best ft n our experments, beng the value ( 1.046) correspondng to the 998 th teraton. See Fgure 2 for a plot of the 998 th and 999 th teraton of the ndex for ths varable s ndex. 9 The ndces for assets, operatng expenses to operatng ncome and other operatng ncome to assets converge completely by the 999 th teraton. The ndces for equty, net nterest margn and return on assets almost completely converge by the 999 th teraton. 10 In ths method we frst delete all varables wth z-statstcs below one (or, exceptonally, 0.5 f the z-statstcs are very small for a large number of varables) and apply a Lkelhood Rato (LR) test relatve to the general model. If the restrctons cannot be rejected, we delete all varables wth z-statstcs below 1.5 and then all explanatory factors wth z-statstcs below 1.96 (applyng all LR tests relatve to the general model). If any LR test for jont restrctons s rejected, we experment to fnd the varable(s) that cause ths rejecton and retan t (them) n the model. 8

11 model all the sgnfcant coeffcents have plausble sgns. That s, lqudty has a postve effect on ratngs: banks wth greater lqudty have a hgher ratng; the natural log of assets has a postve effect on ratngs: banks wth a larger sze of assets have a hgher ratng; the net nterest margn ( NIM ) has a postve correlaton wth ratngs: a bank wth a hgher margn has a hgher ratng. 11 Further, operatng expenses to operatng ncome ( OE _ OI ) has a negatve correlaton wth a bank s ratng: a bank wth a greater rato of operatng expenses to operatng ncome has a lower ratng. Ths benchmark model s percentage of correct predctons s 33.6% whch exceeds the predctve accuracy of 10% (gven 10 ratng categores) expected f the ratngs were assgned randomly. Hence, the model adds predctve performance that s 22.6 percentage ponts greater than that obtaned by chance. The favoured model n the column headed Intercept shft n Table 1 contans the ntercept dummy varable that shfts for new EU countres, New D, but no slope coeffcent shft varables. The same fnancal varables as for the No shft model are sgnfcant and have the same plausble coeffcent sgns, whle the shft n the ntercept s sgnfcant and negatvely sgned. The latter mples that, gven the fnancal varables, new EU countres receve a systematcally lower ratng than old EU countres. Ths may reflect, for example, hgher country rsk and/or regulatory and legal defcences n new EU countres and confrms our hypothess that the country of orgn s an mportant determnant of a bank s ratng. Ths model s percentage of correct predctons of s 37.4%, thus allowng the ntercept to shft notably ncreases the model s predctve performance. 12 The favoured model n the column headed All shft contans varables that allow both the ntercept and slope coeffcents to shft dependng upon whether the naton s an old EU or new EU country. Sx non-shft varables are sgnfcant (equty, lqudty, ln(assets), NIM, OE_OI and ROA) and ther coeffcents represent these varable s correlatons wth ratngs for old EU countres. Seven of the shft varables are sgnfcant (ntercept, equty, lqudty, ln(assets), NIM, OOIA and ROA) whch ndcates that the nfluence of these varables on ratngs s dfferent for new EU countres and old EU countres. 13 The model s percentage of correct predctons s 39.6% and demonstrates that allowng slopes to 11 A hgh NIM contrbutes to a bank s proftablty and enables them to buld up suffcent reserves/provsons for potental losses The other reported measures of ft, pseudo R and SBC, confrm ths ncrease n ft and, beng broader measures of ft, guard aganst the result arsng because the former measure focuses only on whether a model predcts wth complete accuracy or not. 13 The lkelhood rato statstcs ndcate that these shft varables are jontly sgnfcant, confrmng that the coeffcents for old and new EU countres are dfferent for all of these varables. 9

12 shft as well as the ntercept further ncreases the model s predctve performance. 14 The negatve coeffcent on the ntercept shft term suggests that, as for the prevous model, new EU countres have systematcally lower ratngs than old EU countres after the effects of fnancal varables have been taken nto account. Further, the sgnfcance of the slope shft varables coeffcents demonstrates that bank ratngs responses to fnancal varables are dfferent for old and new EU countres. Table 2 reports the slope coeffcents and t-ratos for old and new EU countres mpled by the models reported n Table 1. From the results correspondng to the favoured specfcaton 5 of the 6 sgnfcant coeffcents have the expected sgns for the old EU countres. An ncrease n lqudty, assets, net nterest margn and return on assets wll have a postve mpact on ratngs whereas an ncrease n operatng expenses relatve to operatng ncome has a negatve effect on ratngs. All of these relatons are plausbly sgned. However, the negatve correlaton of equty and ratngs s unexpected. One possble ratonalsaton s that banks use equty to create a buffer aganst possble loss or non-performng assets. 15 Thus, a hgher equty to assets rato may ndcate potental problems wth asset qualty, whch s reflected n a lower ratng. 16 For new EU countres 3 of the 4 sgnfcant coeffcents of the favoured model reported n Table 2 have the expected sgns. Increases n assets and operatng ncome to assets have a postve mpact on ratngs whlst an ncrease n operatng expenses relatve to operatng ncome has a negatve effect on ratngs. In contrast, the negatve correlaton of return on assets wth a bank s ratng s not expected. 17 However, the coeffcent s only just sgnfcant and may be due to a Type-I error (of whch there s a 5% chance gven our chosen sgnfcance level). Indeed, ths fndng of a postve coeffcent on return on assets s not repeated n any other regressons and may, therefore, be regarded as a fragle result. The results of the favoured model reported n Table 2 provde clear evdence that ratngs are determned dfferently for old and new EU countres. The coeffcent for new EU countres s sgnfcantly larger than for old EU countres for equty, assets and operatng ncome. Conversely, the coeffcent for new EU countres s sgnfcantly smaller than for 14 2 The other reported measures of ft, pseudo R and SBC, confrm ths ncrease n ft. 15 Untl recently (before the crss) equty (or captalsaton) was not a problem n bankng. 16 In transton economes t has been essental that banks buld up hgh equty because of hgher rsk, although we do not fnd a negatve correlaton between ratngs and equty for new EU countres. 17 Return on assets s an ndcator of proftablty. In ths specfc case hgh proftablty can be consdered as a weakness that s assocated wth mprudent lendng polces. In other words, a hgh proft may result from reckless lendng. Ths would be especally relevant for new EU countres. 10

13 old EU countres for lqudty, net nterest margn and return on assets. Only for operatng expenses to operatng ncome are the coeffcents the same for old and new EU countres. Table 3 reports results where a heterogeneous ntercept and slopes (for the fnancal covarates) are allowed for all countres and not just for the new and old EU country groupngs. The models reported n the column headed Intercept heterogenety contan the ntercept country ndex but no country ndces for the covarates slopes. From the favoured model we see that all sgnfcant coeffcents have expected sgns except equty. Date, lqudty, assets, net nterest margn and operatng ncome have plausble postve effects on ratngs whle operatng expenses has a plausble negatve correlaton wth a bank s ratng. As before, equty has an unexpected negatve mpact on ratngs suggestng that ths may not be a fragle result. 18 It s partcularly noteworthy that the ntercept country ndex s hghly sgnfcant and ts ncluson n the model rases the model s percentage of correct predctons substantally compared wth prevous models to 48.0%. 19 Ths suggests that country-specfc factors, beyond those captured by fnancal covarates, are very mportant determnants of ratngs. The models reported n the column headed All heterogenety of Table 3 contan both heterogeneous ntercept and slope ndces. The same non-ndex covarates as reported n the favoured model under the Intercept heterogenety column are sgnfcant, except for Date, and have the same coeffcent sgns. The ndex varables that are sgnfcant are for the ntercept, lqudty and operatng expenses: these are the only varables that exhbt coeffcent heterogenety. The percentage of correct predctons s 50.5%, whch suggests that addng covarate ndces (gvng slope heterogenety) rases the predctve performance by 2.5 percentage ponts relatve to the model only allowng ntercept heterogenety. The values of the ntercept coeffcents from the ntercept country ndex are gven n Table 4. All of the countres have dfferent ntercepts, ndcatng that all countres ratngs contan a country-specfc element. All of the old EU countres have larger ntercepts than the new EU countres, ndcatng that country-specfc factors lower new EU countres ratngs relatve to old EU natons, whch confrms our ntal hypothess. However, t s worth emphassng that wthn old and new EU country groupngs there s ntercept heterogenety. Hence, factors such as soveregn rsk and country dfferences n the legal and 18 A hgher equty to assets rato may be an ndcaton of potental problems wth asset qualty whch s reflected n a lower ratng. 19 Ths ntercept ndex varable substantally mproves predctve performance relatve to a model wth no heterogenety or shfts by 14.4 percentage ponts. The model headed Intercept heterogenety n Table 3 s predctve performance s 48.0% compared wth the model headed No shft n Table 1 of 33.6%. 11

14 regulatory frameworks n whch banks specfcally operate affect the ratngs at the ndvdual country level. Whlst we confrm that new EU countres have lower ratngs than old EU countres (after controllng for fnancal varables) our results emphasse that ratngs do not smply dffer by old and new EU country cohorts. The country-specfc coeffcents for the lqudty and operatng expenses to operatng ncome varables are reported n Table 5. All of the countres coeffcents have the expected sgns, except for Romana s lqudty coeffcent whch s relatvely small n magntude, beng vrtually zero. Wth the excepton of Romana (and Span) new EU countres tend to have larger coeffcents for both varables compared wth old EU countres. Further, ratngs tend to be more senstve to lqudty for new EU countres relatve to old EU countres, whle ratngs tend to be less responsve to operatng expenses to operatng ncome for new EU countres compared wth old EU countres. Whlst there s some heterogenety for both varables, many coeffcents are the same. That s, for 16 out of 27 countres the coeffcents are the same for lqudty and for 13 out of 27 countres they are the same for operatng expenses. We note that only two fnancal varables show coeffcent heterogenety and wthn these varables many of the dfferent countres are the same, whch contrasts wth the ntercept ndex whch ndcates a dfferent ndex for all countres. It therefore appears that the man country heterogenety comes from the ntercept varable and only a small part comes from the dfferent country responses of ratngs to fnancal varables. Further, recall that the predctve performance of the benchmark model contanng no heterogeneous (or shftng) coeffcents s 33.6%. Thus, the ncorporaton of a heterogeneous ntercept ncreases ths performance by 14.4 percentage ponts to 48.0%. Addng ndces for both heterogeneous slopes and a heterogeneous ntercept rases the model s predctve accuracy to 50.5%, whch s a relatvely modest ncrease of 2.5 percentage ponts (compared wth the model contanng a heterogeneous ntercept). Ths suggests that most of the mprovement n ft comes from addng a heterogeneous ntercept and only a small percentage from the addton of heterogeneous slopes. Thus, the heterogeneous ntercept appears to be a crucal determnant of ratngs. 4. Conclusons Our models of EU country ratngs show that ratngs are determned by fnancal varables and that these covarates have the expected coeffcent sgns except for equty. We suggest that the explanaton for ths latter result may be that a hgher equty to assets rato can be an 12

15 ndcaton of potental problems wth asset qualty whch s reflected n a lower ratng. Country-specfc factors (n the form of heterogeneous ntercepts) are a crucal determnant of ratngs. Whlst new EU countres typcally have lower ratngs than old EU countres, after controllng for fnancal varables, t should be emphassed that all countres have sgnfcantly dfferent ntercepts ths confrms our ntal hypothess. Ths ntercept heterogenety may reflect dfferences n country rsk and the legal and regulatory framework that banks face (such as foreclosure laws). There may be some dfferences across countres n the assgnment of ratngs due to the lqudty and operatng expenses to operatng ncome varables. There s some evdence that ratngs are typcally more responsve to lqudty and less senstve to operatng expenses for new EU countres compared wth old EU countres. However, t s clear that the prmary country heterogenety n ratngs arses from the ntercept rather than from the slopes. Constructon of slope heterogenety ndces s a novel development n the methodology of constructng ndex-of-ndcator varables. 13

16 References Altman, E. The Z-Score Bankruptcy Model: Past, Present, and Future, Wley, New York, Altman, E. I., and Saunders, A. (1998). Credt rsk measurement: Developments over the last 20 years. Journal of Bankng and Fnance, 21, Altman, E., and Rjken, H., A. (2004). How Ratng Agences Acheve Ratng Stablty, Journal of Bankng & Fnance, 28, Amato, J. D. and Furfne, C.H. (2004). Are Credt Ratngs Procyclcal?, Journal of Bankng and Fnance 29, Caporale, G.M., Matousek, R. and Stewart, C. (2009). Ratng Assgnments: Lessons from Internatonal Banks, DIW Berln Dscusson Paper No Avalable at SSRN: Greene W. H., (2008). Econometrc Analyss, Pearson, Prentce Hall, 6 th edton. Hendry D. F. and Santos C. (2005). Regresson models wth data-based ndcator varables Oxford Bulletn of Economcs and statstcs, 67, 5, Hendry D. F., (2001). Modellng UK Inflaton, Journal of Appled Econometrcs, 16, Huang, Zan, Hsnchun Chen, Cha-Jung Hsu, Wun-Hwa Chen, Soushan Wu. (2004). Credt ratng analyss wth support vector machnes and neural networks: a market comparatve study, Decson Support Systems, 37, pp , I Kamstra, M., Kennedy, P., Suan, T.K. (2001). Combnng bond ratng forecasts usng logt. The Fnancal Revew, 37, pp Km, S. K. (2005). Predctng bond ratngs usng publcly avalable nformaton, Expert Systems wth Applcatons, 29, pp Kolar, J. D. Glennon, H. Shn and M. Caputo, (2002). Predctng large US commercal bank falures, Journal of Economcs and Busness, 54, pp Kolar, J., Caputo, M., Wagner, D. (1996). Trat recognton: An alternatve approach to early warnng systems n commercal bankng. Journal of Busness Fnance and Accountng, 23, Lee, Y.C. (2007). Applcaton of support vector machnes to corporate credt ratng predcton, Expert Systems wth Applcatons 33, pp Levch, R., Majnon, G., Rnhart, C. (2002). Ratngs, Ratng and Agences and the Global Fnancal System, Kluwer Publshng. Meyer, P., & Pfer, H. (1970). Predcton of bank falures. Journal of Fnance, 25, Ohlson, J. (1980). Fnancal ratos and the probablstc predcton of bankruptcy. Journal of Accountng Research, 18, Portes, R. (2008). Ratngs agency reform, Vox, 22 January,2008, < org/ndex.php?q=node/887> 14

17 Table 1: Bank ratngs probt regressons wth new EU coeffcent shft Varables (expected sgn) Date ( 0.229) Equty (+) ( 0.631) Lqudty (+) (8.049) ln( Assets ) t 1 (+) (14.430) NIM ( /+) (1.560) OE _ OI ( ) ( ) OOIA t 1 (+) ( 1.693) ROA (+) (1.110) No shft New EU ntercept shft New EU ntercept and slope shft Gen Fav Gen Fav Gen Fav (8.354) (15.683) (2.115) ( ) (1.276) ( 1.023) (7.358) (8.332) (4.493) ( 6.874) (1.271) (1.355) Intercept _ New Equty _ New ( ) (7.714) (9.030) (4.780) ( 8.748) ( ) (1.509) ( 2.277) (5.735) (6.944) (4.032) ( 5.615) ( 0.476) (4.000) ( 1.609) (2.989) Lqudty _ New ( )_ New ( 2.870) ln Assets (2.801) NIM _ New ( 2.778) OE _ OI _ New ( 1.126) OOIA _ New (2.469) ROA _ New ( 4.291) ( 2.252) (5.946) (6.929) (3.953) ( 6.172) (3.976) ( 2.674) (2.902) ( 3.273) (2.790) ( 3.231) (3.546) ( 4.735) Ft Measures % correct Pseudo R SBC LR statstc [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] LR(general favoured) [0.382] [0.096] [0.276] LR(slope shft) [0.000] [0.000] LR(slope/ntercept shft) [0.000] [0.000] Observatons Table 1 notes. The dependent varable s a bank s ratng whch has ten categores that correspond to the nteger values n the range of 1 to 10 and yelds nne lmt ponts, λ, = 1, 2,..., 9 (the ntercept s not separately dentfed from the lmt ponts). Z-statstcs (n parentheses) are based upon Huber-Whte standard errors and the percentage of correct predctons (% correct) use the category wth the 2 hghest probablty to gve the predcted ratng. Also reported are the Pseudo R and Schwartz s nformaton crteron, SBC. Lkelhood rato tests for the model s explanatory power, LR Statstc, the deleton of varables from the general model to obtan the parsmonous model, LR(general favoured) the deleton of slope shft varables, LR(slope shft), and the deleton of slope and ntercept shft varables, LR(slope/ntercept shft) from a model are addtonally reported. Probablty values are gven n square parentheses. All regressons were estmated usng E-Vews

18 Table 2: Impled slope coeffcents and t-ratos of EU shft models General Favoured Varables (expected sgn) Old EU New EU Old EU New EU Date (1.509) Equty (+) ( 2.277) * (1.917) ( 2.252) * (1.829) Lqudty (+) (5.735) * ( 0.542) (5.946) * ( 0.818) ln( Assets ) t 1 (+) (6.944) * (7.773) * (6.929) * (8.067) * NIM ( /+) (4.032) * ( 0.378) (3.953) * ( 1.011) OE _ OI ( ) ( 5.615) * (3.208) * ( 6.172) * ( 6.172) * OOIA (+) ( 0.476) (3.684) * (3.546) * ROA (+) (4.000) * ( 1.890) (3.976) * ( 1.991) * Table 2 notes. The (mpled) coeffcents and t-ratos are reported for new EU and old EU countres based upon the general and favoured regressons reported n Table 1 under the column headed New EU ntercept and slope shft. The coeffcents and t-ratos for the old EU countres are exactly the same as those reported n Table 1. The coeffcents for new EU countres are the sum of the coeffcents on the varable of nterest and ts correspondng shft term. The t-ratos for new EU countres are calculated based upon the varance of the sum of a partcular varable s coeffcent (a) and ts correspondng shft varable s coeffcent (b), that s, Var(a + b) = Var(a) + Var(b) + 2Cov(ab). An asterx ndcates that a varable s sgnfcant at the 5% level (usng a crtcal value of 1.96 n absolute value). 16

19 Table 3: Bank ratngs probt regressons wth country heterogenety Intercept heterogenety Intercept and slope heterogenety Varables (expected sgn) Gen Fav Gen Fav Date (2.489) (2.448) (1.714) Equty t 1 (+) ( 3.704) ( 3.537) ( 2.770) ( 2.723) Lqudty (+) (3.212) (3.424) (1.903) (2.370) ln( Assets ) t 1 (+) (13.367) (13.461) (9.256) (9.248) NIM ( /+) (4.402) (4.987) (2.968) (3.176) OE _ OI ( ) ( 8.795) ( ) ( 5.884) ( 7.434) OOIA t 1 (+) (2.329) (2.486) (2.022) (2.551) ROA (+) (1.101) (0.162) Intercept _ Country (24.159) (24.159) (19.883) (22.507) Equty _ Country t (1.570) Lqudty _ Country t (1.161) (3.332) ln( Assets )_ Country t (1.294) NIM _ Country t ( 1.088) OE _ OI _ Country t (1.964) (2.475) OOIA _ Country t ( 0.201) ROA _ Country t ( 0.768) Ft Measures % correct Pseudo R SBC LR statstc [0.000] [0.000] [0.000] [0.000] LR(general favoured) [0.271] [0.318] LR(slope heterogenety) [0.000] [0.000] LR(slope/ntercept heterogenety) [0.000] [0.000] Observatons Table 3 notes. The dependent varable s a bank s ratng whch has ten categores that correspond to the nteger values n the range of 1 to 10 and yelds nne lmt ponts, λ, = 1, 2,..., 9 (the ntercept s not separately dentfed from the lmt ponts). Z-statstcs (n parentheses) are based upon Huber-Whte standard errors and the percentage of correct predctons (% correct) use the category wth the 2 hghest probablty to gve the predcted ratng. Also reported are the Pseudo R and Schwartz s nformaton crteron, SBC. Lkelhood rato tests for the model s explanatory power, LR Statstc, the deleton of varables from the general model to obtan the parsmonous model, LR(general *) the deleton of slope shft country varables, LR(slope heterogenety), and the deleton of slope and ntercept country varables, LR(slope/ntercept heterogenety) from a model are addtonally reported. Probablty values are gven n square parentheses. The varables correspondng to the country shft are all determned after 999 teratons except the one for lqudty, whch alternated between two dfferent forms, we used the form correspondng to the 998 th teraton. All regressons were estmated usng E-Vews

20 Table 4: Heterogeneous ntercept (country weghts) Country Weght Country Weght Old EU New EU Luxembourg Estona Netherlands Slovaka UK Malta Denmark Hungary Span Cyprus Sweden Slovena Ireland Czech R Portugal Poland Fnland Bulgara Belgum Romana Austra Lthuana Italy Latva France Germany Greece Table 4 notes. The coeffcent of the ndvdual countres emboded n the ndex of ndcators varable, Intercept _ Country, are gven. The coeffcents are ranked from hghest to lowest value. 18

21 Table 5: Heterogeneous slopes Lqudty Oe_o Malta Sweden Lthuana Denmark Latva Fnland Bulgara Romana Slovena Germany Span Austra Austra France Belgum Italy Cyprus Belgum Czech Republc Cyprus Estona Czech Republc Fnland Estona France Greece Greece Ireland Hungary Luxembourg Ireland Netherlands Italy Poland Netherlands Portugal Poland Slovaka Portugal Span Slovaka UK UK Slovena Sweden Bulgara Denmark Lthuana Germany Malta Luxembourg Hungary Romana Latva Table 5 notes. The coeffcents for each ndvdual country mpled by the fnancal varables parameters and the ndex of ndcator varables, Lqudty _ Country and OE _ OI _ Country, are gven. These are constructed as the coeffcent on the k th varable, βˆ k, and the product of the k th CF CF varable s ndex, I k, and ts assocated coeffcent, βˆ k, that CF s, as, ˆ CF β k + ˆ β k I k. The coeffcents are ranked from the hghest to lowest value for lqudty and lowest to hghest for operatng expenses to operatng ncome. 19

22 Fgure 1: Equty Index Iteratons 6,000 4,000 2, ,000-4,000-6, SLOPE_DUM_EQUITY_998 SLOPE_DUM_EQUITY_999 Fgure 1 notes: slope_dum_equty_998 and slope_dum_equty_999 are the 998 th and 999 th teratons of the equty ndex. Fgure 2: Equty Index Iteratons SLOPE_DUM_LIQ_998 SLOPE_DUM_LIQ_999 Fgure 2 notes: slope_dum_lq_998 and slope_dum_lq_999 are the 998 th and 999 th teratons of the lqudty ndex. 20

23 Fgure 3: Assets Index Iteratons SLOPE_DUM_LNASSETS_998 SLOPE_DUM_LNASSETS_999 Fgure 3 notes: slope_dum_lnassets_998 and slope_dum_lnassets_999 are the 998 th and 999 th teratons of the assets ndex. Fgure 4: NIM Index Iteratons 8,000 6,000 4,000 2, ,000-4,000-6,000-8, SLOPE_DUM_NIM_998 SLOPE_DUM_NIM_999 Fgure 4 notes: slope_dum_nm_998 and slope_dum_nm_999 are the 998 th teratons of the NIM ndex. and 999 th 21

24 Fgure 5: OE_OI Index Iteratons SLOPE_DUM_OE_OI_998 SLOPE_DUM_OE_OI_999 Fgure 5 notes: slope_dum_oe_o_998 and slope_dum_oe_o_999 are the 998 th and 999 th teratons of the operatng expenses to operatng ncome ndex. Fgure 6: OOIA Index Iteratons SLOPE_DUM_OOIA_998 SLOPE_DUM_OOIA_999 Fgure 6 notes: slope_dum_ooa_998 and slope_dum_ooa_999 are the 998 th teratons of the other operatng ncome to assets ndex. and 999 th 22

25 Fgure 7: ROA Index Iteratons 80,000 60,000 40,000 20, ,000-40,000-60, SLOPE_DUM_ROA_998 SLOPE_DUM_ROA_999 Fgure 7 notes: slope_dum_roa_998 and slope_dum_roa_999 are the 998 th teratons of the return on assets ndex. and 999 th 23