Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
QUANT PERSPECTIVES The Hazard Rate Matrix Approach to Credit Rating Transitions By J. M. Pimbley 1 D ebt securities are financial instruments that obligate the issuer of the debt to make principal and interest payments at specified times to the investor. Failure of the issuer to make such payments is “default.” One may characterize all outstanding debt instruments as being “in default” or “not in default.” The great majority of debt instruments — to which we now refer as “debt” or “bonds” — will not default prior to maturity. Let’s consider the designations of “in default” or “not in default” as a primitive rating system for bond performance. Every outstanding bond falls into one of these two categories. Credit rating agencies (CRAs) are incorporated financial firms that analyze bonds and provide opinions on bond default risk in the form of “ratings” that extend the “not in default” category of the primitive rating system.2 These CRAs generally give a symbol D to the “in default” category and give symbols such as AAA, AA, A, BBB, BB, B, and CCC to seven “not in default”categories that represent increasing risk of future default. The precise symbols for these “letter grades” vary from one CRA to another. Further, each CRA has an expanded scale that boosts the number of “not in default” ratings from seven to roughly 20. In this article, we will generally use these seven “major” rating levels merely to simplify the discussion. Meaning of the CRA Credit Ratings The bond rating that carries the least risk is AAA. The “lower” rating levels of AA down to CCC denote progressively increasing default risk. The four largest CRAs have published www.garp.org “default tables” that show either targets or historical assessments of default probability over time for each rating level.3 These tables would seem to provide a definition, or at least a benchmark, for the meaning of the CRA ratings. The CRAs, however, have generally avoided making the explicit claim that their default tables define their ratings.4 The CRAs also publish “rating transition matrices” that show the frequency of rating changes from one level (such as AA) to another (such as A) over periods of one year or longer.5 These transition matrices are consistent with and complementary to the default tables since “default” is synonymous with rating transition to the default rating D. Thus, for example, the default probabilities for each rating for a tenor of one year form the (bottom) row of the one-year transition matrix for the final rating of D.6 Time-Dependence of the Credit Rating Contemplation of the CRA rating transition matrix (RTM) leads us to consider the process for time evolution of credit ratings. Previous work begins with the one-year transition matrix and makes the assumption that rating changes are Markovian.7 In this context, the description “Markovian” means simply that the likelihood of a rating change depends only on a bond’s current rating and the RTM. In practice, rating changes are not Markovian because the CRAs — inadvertently or otherwise — behave in a manner that depresses rating volatility. For example, if a currently AArated bond deserves a downgrade to BB, CRAs will typically downgrade the bond to A or BBB rather than impose the full A U G U S T 2 0 1 2 RISK PROFESSIONAL 1 not in default and that the probability of the bond not being in default at time is ( ). The initial condition is ( ) . We specify the probability of the bond (which is identical the T probability Q U A defaulting NT P ER S P Eto C I V EofSthe bond being downgraded to the “in default” rating level) during the time period ( ) as ( ). Since a rating transition to the “in default” rating is a rating transition out of the “not in default” rating, we write ( ) . (1) rating change in one rating action.8 Moody’s Investors Service Equation (1) shows familiar that is Equation (1) shows the familiar hazardthe rate processhazard that is rate well process known for 10 states: “Moody’s ratings management practice, by avoiding re- well known for bond defaults and other applications. The bond defaults and other applications.9 The solution of equation (1) for the survival versals, necessarily produces positive serial momentum in rat- solution of equation (1) for the survival time S(t) with a time) with ) isis time to( be a time-dependent rate ings changes. A downgrade is much more likely followed dependenthazard hazard rate((t) by another downgrade than it is an upgrade.”9 (2) ( ) ( )] . (2) [ ∫ Before continuing, let us emphasize this point. Existing treatments of rating transitions employ a Markov paradigm. This ( ) We specify for to indicate presence of default risk at of all detreatment is most tractable and most sensible: ratings should Weallspecify (t)>0 for allthet>0 to indicate the presence be Markovian to maintain accuracy of default risk estimation. fault risk at all times. While the probability S(t) of the bond times. While the probability ( ) of the bond being in the “not in default” rating Yet CRA rating transitions are clearly not Markovian in prac- being in the “not in default” rating level declines from 1 at level declines from 1 at timet=0 to zero to zero t goes infinity,the theprobability probabilityofof the tice. Hence, we will not try to fit empirical rating transition time as as t goes to toinfinity, data with our model results. Rather, we view the relevance of bond being in default 1–S(t)moves in the opposite manner ( ) - moves in the opposite manner from zero the bond being in default at this study to be the development of idealized (Markov) rat- from zero at t=0 to 1 as t goes to infinity. The default state is to 1 as t goes to infinity. The default state is “absorbing” in the sense that a ing transition behavior to assist banks, portfolio managers and “absorbing” in the sense that a bond remains in default once bond remains defaultit once it reaches default. other practitioners in the development of accurate defaultinratreaches default. ing scales. Consequently, we develop in this article an alternative of Marof Rating Transitions: Full Rating Scale Mathematics RatingMathematics Transitions – Full Rating Scale kovian process for rating changes that we find to be more intui- The purpose of our explaining the well-known hazard rate The purpose our ofexplaining well-known rate model tive than existing methods that begin with the one-year RTM. of model equationsthe (1) and (2) for thehazard bond default processofis to The new treatment, which ultimately merges with the convenshow motivation for a similar “hazard rate matrix” equations (1) and (2) for the bond default process is to show motivation approach for a tional method, has the advantage of both an easier derivation for credit rating transitions. Instead of beginning with a bond 2012 similar “hazard rate matrix” approach for credit rating transitions.J. M. Pimbley Instead ofthat and better examination of the bond rating behavior for large that is simply “in default” or “not in default,” we specify times. We find, for example, that a certain matrix eigenvector has a “in designated of beingwe in one of the beginning with a bondthe thatbond is simply default” probability or “not in default”, specify sum toallone and rating be written as the column vector (AAA, ⃗. This vector has components forms a “natural rating distribution” to which initial eight rating categories AA, A, BBB, BB, B, CCC, or D). that the bond has a designated probability of being in one of the eight rating states evolve. Let these eightrating probabilities one and be written with designating the eight levels. Ifsum onetowishes to specify, for as the → CCC, categories (AAA, AA,column A, BBB,vector BB, B, or D).has Letcomponents these eight probabilities The nature of this finding — that we identify an eigenvecr. This vector ri with i=1,...,8 example, current designating bond rating the is definitely BBB, thenIfone would setto specify, for tor with a credit rating distribution — implies thatthat thethe matheight rating levels. one wishes 9 See, for example, Löffler andexample, Posch (2011). ematical development and the finance implication are strongly that the current bond rating is definitely and all other to zero. More generally, let us say there are N rating categories BBB, with then linked. Much of the remaining exposition of this article is un- one would set r4=1 and all other r to zero. More generally, i 4 J. M.NPimbley category being “default”. avoidably mathematical. The following sections (up to “Practi- let us say there are N rating categories with category N being cal Result – Examples”) constitute this analysis, Analogous and the ensu“default.” Formatted: Font: 12 pt, Italic to equation (1), the rating probability vector ⃗( ) evolves in time matics of Rating Transitions Simple Case ing sections–provide the finance interpretation. Analogous to equation (1), the rating probability vector r(t) as evolves in time as To begin, let us consider the “primitive rating system” in which a bond is Mathematics of Rating Transitions: Simple Case ⃗ (3) ⃗ (3) ⃗( ) Toinbegin, let usAt consider rating “in default” or “not default”. time the “primitive , we specify that system” the bondiniswhich a bond is either “in default” or “not in default.” At time t=0, default and that the probability of the bond not being in default at time( ) where is the time-dependent matrix (HRM)hazard for rating we specify that the bond is not in default and that the probwhere Q(t)hazard is therate time-dependent ratetransitions. matrix (HRM) ). The initial condition ( ) is . We specify the probability of the bond ability of the bond not being in default atIntime S(t). The , the for rating In the limit the t>0 limitis as HRM transitions. has the meaning that as ∆t→0,isthe theHRM rating Q has initialtocondition is S(0)=1. Webond specify thedowngraded probabilitytoof the the meaning that I–Q∆t is the rating transition matrix for the ting (which is identical the probability of the being transition matrix timeperiod period∆t where where is identity the identity matrix. bond defaulting (which is identical to the probability of for the thetime I is the matrix. More More specifically, ( ) as ( ). Since a n default” rating level) during the time period bond being downgraded to the “in default” rating level) is the probability a bond specifically, the ijdurelementthe ofij elementisoftheI–Q∆t probability that a bond that in rating levelinj rating ingdefault” the timerating period S(t)dt. Since rating level j will transition to rating level i in the time period ∆t. We transition to the “in is (t,t+dt) a ratingastransition out ofa the “nottransition in will transition to rating level i in the time period . We specify a negative sign in to the “in default” rating is a rating transition out of the “not specify a negative sign in equation (3) to maintain consonance t” rating, we write equation (3) to maintain consonance with(1). equation (1). As a result, the components in default” rating, we write with equation As a result, the components Qij of Q satisfy the following: of satisfy ( ) . (1) (1) → Equation (1) shows the familiar hazard rate process that is well known for defaults and other applications.9 The solution of equation (1) for the survival 2 RISK PROFESSIONAL A U G U S T 2 0 1 2 ( ) with a time-dependent hazard rate ( ) is (4a) (4b) www.garp.org (4c) Forma where ⃗( ) (3) ⃗ ( ) is the time-dependent hazard rate matrix (HRM) for rating transitions. In the limit as , the HRM transition matrix for the time period specifically, the ij element of has the meaning that Q U A N T P EisRtheSrating PECTIVES where is the identity matrix. More is the probability that a bond in rating level j will transition to rating level i in the time period . We specify a negative sign in equation (3) to maintain consonance with equation (1). As a result, the components of satisfy ing transition dynamics have treated this one-year RTM as the J. M. Pimbley analytical starting point and then created a “generator matrix” 12 similar to the HRM Q of this article. (4b) The primary difference between these earlier ma-in less Our HRM differs from the generator matrices of generator past studies trices and our HRM is our interpretation of Q as the agent of (4c) significant terms as well. The HRM has a preceding negative algebraic 13sign for time evolution of the rating probability vector in equation (3). consistency We withconsider the scalar rate.guiding To accommodate rating probability ∑ . (4d) thedefault HRMhazard to be the concept and relegate the RTM to an ancillary column vectors in equations (3), (5),role. and (6), our hazard rate matrix indices are The special The designation in (4a) for rating N category reflects the condition that a HRM differs from the generator matrices of past studspecial designation in (4a)category for rating N reflects Our carries transposed from the convention of past studies. That is, the component the condition a bond in defaulttransition cannot experience a rating in less significant terms as well. The HRM has a preceding bond in default cannot that experience a rating to leave this default ies state. transition to leave this default state. Equation (4d) imposes the negative algebraic of sign for migration consistency with the scalar information on the probability rating from rating level j default to rating level Equation (4d) imposes the requirement that the sum of probabilities of all rating requirement that the sum of probabilities of all rating levels hazard rate. To accommodate rating probability column veci rather than from rating level i to rating level j. levels must be be constant (at the value 1). 1). Thus, equations (4a-d) show us thattors is must constant (at the value Thus, equations (4a-d) show in equations (3), (5) and (6), our hazard rate matrix indices that Q in is an N x column N matrixNiniswhich N is filled withPast are transposed fromhave the convention past studies. That is, the research projects applied a ofhazard rate matrix approach to an N x Nus matrix which filled column with zeroes, all other diagonal zeroes, all other diagonal elements are positive, all other off- component Q ij carries information on the probability of rating problems than time evolution of credit ratings. Takada and Sumita (2011) elements diagonal are positive, all other and in theother sum of thefrom elements areoff-diagonal negative andelements the sumare of negative, the elements migration rating level j to rating level i rather than from studied financial default based eachincolumn is zero. rating level i to risk rating level on j. obligor industry and macro-economic the elements each column is zero. Just as we can solve the scalar hazard rate equation (1) to get Past research projects have a hazard rateMarkov matrix apconsiderations. Singer and Spilerman applied (1976) investigated processes J. M. Pimbley Just we can time solveS(t) theofscalar hazard get the survival theassurvival equation (2),rate theequation solution(1) to to equation proach to problems other than the time evolution of credit ratrelevant to ings. sociology such as immigration and juvenile J. M. Pimbley Takadatopics and Sumita (2011) studied financial defaultdelinquency risk time ( )(3) of is equation (2), the solution to equation (3) is ( ) ⃗⃗⃗⃗ (5) ⃗( ) based on obligor industry and macro-economic considerations. recidivism. Keilson and Kester (1974) provided mathematical background and Singer and Spilerman (1976) investigated Markov processes ( ) ⃗⃗⃗⃗ (5) 5 ⃗( ) properties (5) for hazard rate matrices in Markov processes. special case in which has no time dependence. In equation (5), the relevant to sociology topics such as immigration and juvenile which has time equation (5), the everyIn delinquency recidivism. Keilson and Kester (1974) provided for the special hasIn“no time dependence. tl case of thein exponential is no an case N x independence. N which matrixQwith ” multiplying Properties background of the Hazard Matrix equation argument of the is anevery N Necessary x N ma- mathematical and Rate properties for hazard rate mahe of exponential is an(5), N the x N matrix with “ exponential multiplying ent .10 The initial rating probability vector is”⃗⃗⃗⃗. For a time-dependent 11 trix, with “–t” multiplying every component of Q. The initial trices in Markov processes. .10it The initial rating vector is ⃗⃗⃗⃗.a time-dependent For atotime-dependent ( ), is not possible toprobability determine a solution equation HRM (5) unless rating probability vector is r→ Q(t),We list here and discuss briefly a collection of properties of the HRM in o. Forsimilar addition of equations (4a) – (4d). itisisofnot possible determine a solution similar(5) to unless equation (5), to those Necessary Properties of the Hazard Rate Matrix is not possible tothe determine aa solution to multiplied equation dependence form ofto constant similar matrix by an arbitrary unless the time dependence is of the form of a constant matrix We list here and discuss briefly a collection of properties of the dence is of( the form of awith constant( matrix multiplied by an arbitraryrating ) ( scalar ), thefunction unction Thus, time-dependent a zeroHRM eigenvalue A).multiplied by an arbitrary g(t). Thus,has with Q in addition to those of equations (4a) – (4d). → ( ). ( ) ( ), nty vector ⃗(Q(t)=Ag(t), Thus, therating time-dependent rating r(t) is probability vector ) is withthe time-dependent Since column N of Q has a zero eigenvaluehas all zero entries, this matrix is singular. All ctor ⃗( ) is singular have at one zero entries, eigenvalue. Oneis singular. determines the Since column N ofleast Q has all zero this matrix (6) matrices ( )] ⃗⃗⃗⃗ (6) ⃗( ) [ ∫ All singular matrices have at least one zero eigenvalue. Oneall zero corresponding eigenvector with little effort to be the column vector with ( )] ⃗⃗⃗⃗ (6) ⃗( ) [ ∫ Unless stated otherwise for the remainder of this study, we determines the corresponding eigenvector, with little effort, to Unless stated otherwise for the remainder of this study, we take the HRM setentries, to 1. Writing ⃗ as this entries thecolumnelement take the HRM Q to be constant, so that equation (5) is thesavebeforthe vector which with allwe zero save for theand Nth elestated otherwise for the remainder of this study,expression we take thefor HRM → onstant so that equation (5) is the the rating relevant expression for therelevant rating probability vector. It would and ment, which werespectively, set to 1. Writing this eigenvalue N and N as eigenvalue eigenvector, these statements become tty so that equation (5) is the relevant expression for the rating ( ) vector. be It would be worthwhile a time-dependent in worthwhile to employtoaemploy time-dependent Q(t) in modeling and eigenvector, respectively, these statements become exercises for which one desires a rating volatility that increases ( ) ctor. It would be worthwhile to employ a time-dependent in g exercises for which one desires a rating volatility that increases or or decreases with a prescribed g(t). (7) ( ) . (7) ⃗ ⃗ ⃗ (4a) 2012 rcises which one ( ).desires a rating volatility that increases or s with afor prescribed ( ). a prescribed Standard Rating Transition Matrix (5) permits us to write the RTM for any future time t rd RatingEquation Transition Matrix as exp(–Qt). The most typical period is one year, so the one-year ating Transition Matrix quation (5)RTM permits us be to simply write exp(–Q). the RTMMany for any futurestudies time of ratas would previous on). (5) us toperiod write isthe anyone-year future time as be Thepermits most typical oneRTM year, for so the RTM would he (most). typical period is studies one year, so thetransition one-yeardynamics RTM would Many previous of rating havebetreated ). Many studies ofstarting rating transition have atreated year RTMprevious as the analytical point anddynamics then created “generator www.garp.org RTM astothe starting point 11 andThe thenprimary createddifference a “generator similar theanalytical HRM of this article. between 11 All other eigenvalues of Q are real and positive For all other eigenvalues i and eigenvectors → i with i=1,...,N-1, it would be highly desirable if the eigenvalues were real and positive. While the matrix Q is real, it is not symmet7 A U G U S T 2 0 1 2 RISK PROFESSIONAL 3 ⃗ with For all eigenvalues and eigenvectors, we can write ⃗ J. M. Pimbley . Excluding the last eigenvalue/eigenvector pair, we know 012 All other eigenvalues of are real and positive . We Q U A N T define P EtheRrow S Pvector E C⃗ Tto Ihave V rank E SN with all elements equal to 1. Pre- multiplying the HRM by ⃗ gives another rank N row vector in which all , it elements are zero (because the sum of each column of is zero). This observation would be highly desirable if the eigenvalues were real and positive. While the means that ⃗ ⃗ . The pre-multiplication of ⃗ by ⃗ simply sums the matrix is real, it is not symmetric. It is in principle possible for some eigenvalues elements of ⃗ . Since , the sum of the elements of each eigenvector ⃗ must to be complex even with the properties of equations (4a) – (4d).13 Therefore, we be zero. Note that this zero-sum condition does not apply to the last eigenvector ⃗ impose the constraint on the elements of that all eigenvalues be real. Our → ric. It is in principle possible for some eigenvalues sinceto be .elements of each eigenvector i must be zero. Note that this 14 motivation for this requirement the the behavior of solutions of equation that we zero-sum condition does not apply to the last eigenvector →N, complex even iswith properties of equations (4a) (5) – (4d). The eigenvectors of the rating transition matrix are readily Therefore, we impose the constraint on the elements Qij eigenvalues of Q sinceand N =0. discuss later. For all other eigenvalues and eigenvectors ⃗ with that all eigenvalues be real. Our motivation for thisdetermined requirement is the behavior of solutions of equation (5), which we The eigenvalues and eigenvectors of the rating transition matrix are readily Considering the sub-matrix of that excludes the row and column, discuss later. determinedearlier that the one-year RTM is ( ) and the more Wetoobserved Gerschgorin’s Circle Theorem (GCT) and equations (4b) - (4d) are sufficient Considering the sub-matrix of Q that excludes the Nth row We observed earlier that the one-year RTM is exp(–Q) and the 14 ( ). Working with the latter, this RTM is a general RTM for time is show that all eigenvalues of the sub-matrix are greater than zero (i.e., and column, Gerschgorin’s Circle Theorem (GCT) and equa-). more general RTM for time t is exp(–Qt). Working with the latwith seriesis representation (4b) - (4d)ofarethe sufficient to show all eigenvalues of the ter,infinite this RTM a matrix with the infinite series representation Note that thesetions “eigenvalues sub-matrix” arethat simply thematrix remaining the sub-matrix are greater than zero (i.e., i>0).15 Note that Equations (4b) – (4d) are sufficient eigenvalues these “eigenvalues ofofthe .sub-matrix” are simply the remain( ) (9) . (9) ..., N-1 ofdiagonally support for the assertion that the sub-matrix dominant. Since ing eigenvalues Q. Equations (4b) – (4d) i with i=1,is strictly are sufficient support for the assertion that the sub-matrix is We’ve We’ve written the scalar thescalar left oft to thethe terms with of thepowers HRM of. The writtentothe left of thepowers terms with all diagonal elements are positive, the radii of the Gerschgorin circles of the GCT 2012 M. Pimbley strictly diagonally dominant. Since all diagonal elements are the HRM Q. The notation Qk simply means that J. the matrix Q isSince notation simply means that the matrix is multiplied by itself kk times. → do not permit zero or negative eigenvalues. → positive, the radii of the Gerschgorin circles of the GCT Q Pimbley →i=ik →i, 2012do multiplied by itself k times. Since Q i = i i implies J. M. not permit zero or negative eigenvalues. ⃗ we implies ⃗ eigenvector , we see that ⃗ is alsowith an eigenvector of ⃗ see that →⃗i is also an of exp(–Qt), For convenience, we number the eigenvalues from largest to smallest. For convenience, we number the eigenvalues from largest to ( ) ( ) ⃗ ⃗ ( 2012 ) with To be published in Risk Professional J. M. Pimbley Incorporating thesmallest. possibility that two orthe more eigenvalues mayorhave same ( )⃗ ( ) ⃗ Incorporating possibility that two morethe eigenvalue, we write values may have the same value, we write . (8) The matrix [( (8) hasThe a set of N Q linearly independent matrix has a set of N linearlyeigenvectors independent eigenvectors )⃗ ( ) (() )( [ (⃗ [ ( 9 → the eigenvectors of the RTM eigenvectorThus, N Thus, the eigenvectors ) ] ⃗ ] )⃗ ⃗ )(10) ⃗ ( (10) .) . (10) ] ⃗ are simply the eigenvectors of the the of( the )RTM exp(–Qt) are simply We’ve already shown in equation (7) the ( ) are simply the eigenvectors of the eiThus, the eigenvectors of the RTM We’ve already shown in to equation (7) the eigenvector ⃗ Q corresponding ⃗ eigenvalues (10) . corresponding the eigenvalue =0. While has N–1 HRM ad-to . In terms of the eigenvalues of the HRM, the RTM are . N genvectors of the HRM Q. In terms of the eigenvalues of the t HRMas . In of the eigenvalues of the HRM, theare RTM are . ditional eigenvalues real and positive, assumed or . While has that are additional eigenvalues that areterms realThus, and HRM, the vector RTM eigenvalues e-ieigenvalues . of eigenvectors ) are the eigenvectors of the simply the eigenvectors of the The rating probability ⃗( )RTM is a linear( combination proven earlier, it is possible that Q does not have a full set of . HRM vector . In terms of of the HRM,of theeigenvectors RTM eigenvalues are ) istheaeigenvalues The rating probability ⃗( linear combination Since matrix has a set of N→ linearly independent eigenvectors ⃗ with N linearly independent eigenvectors. The the rating probability vector r(t) is a linear combination of eigenvectors As an example, imagine there are only four rating states (including the default state) so that . Let ) isNa rating If the N–1 positive eigenvalues of equation (8) are distinct, The rating, probability vector linear of independent eigenvectors wethe can matrix write the ⃗( rank probability vector ⃗( ) as a timehas set ofcombination N linearly eigenSince the matrixSince has →a set of NQlinearly independent eigenvectors ⃗ with ... then Q would have this complete set of eigenvectors. But there vectors with i=1, , N, we can write the rank N rating i dependent, weighted sum of the eigenvectors: ). Of the four eigenvalues of this matrix, two are real (0 and Since the matrix has a set of N linearly independent eigenvectors ⃗ probe HRM be ( with , weconcan write rank →r(t) N rating probability vectorweighted ⃗( ) as sum a timeis no assurance of distinct eigenvalues. As with the earlier abilitythevector as a time-dependent, of the ∑ ) ⃗ probability ⃗( ) the rank . (11) , we can write N (rating vector ⃗( ) as a timecern(0.252 for complex eigenvalues, we add a constraint on weighted the per-off-sum 097) and two are complex + i 0.048 and 0.252 – i 0.048). Relatively small changes to some eigenvectors: dependent, of the eigenvectors: dependent, weighted sum of the eigenvectors: agonal matrix elementsmitted produceelements four real eigenvalues. Qij of Q that there exists a set of N linearlyInin-equation (11), the coefficients are time-dependent while the eigenvectors ⃗ → See, for example, Weisstein (2012). In our application of this theorem, we consider the sum of column, dependent eigenvectors i with i=1,..., N. With this condition, ) ∑ ( ) ∑⃗ . ((11) aresince constant. equation (11) ) ⃗ . (11) ⃗( ) (3), ther than row, element absolute values for the Gerschgorin radius. This substitution is permissible a ⃗(Applying the matrix Q is diagonalizable. atrix and its transpose have the same eigenvalues. equation (11), the coefficients are time-dependent while the eigenvectors ⃗ In equation (11), Inthe coefficients are the time-dependent while the eigenvectors ⃗while In⃗ equation ⃗ (11), ∑ ⃗coefficients ∑ i⃗ are time-dependent ∑ ⃗ → are constant. Applying equation (3), The elements of each of the8first N-1 eigenvectors sum to zero the eigenvectors are constant.→ Applying equation (3), i are constant. Applying equation (3), → For all eigenvalues and eigenvectors, we can write Q i = i i ) ⃗ ( ) ∑ ( . (12) with i=1,..., N. Excluding the last eigenvalue/eigenvector pair, ⃗ ∑ ⃗ ∑ ⃗ ∑ ⃗ ⃗ ⃗ rankWeNuse⃗ the notation ∑ ⃗to denote the ∑ ⃗ ∑ value ⃗of . we know i0. We define the row vector →yT to have (time-independent) initial (12) with all elements equal to 1. Pre-multiplying the HRM Q by ) ⃗ the probability ( vector ) ∑ ( (11) and (12), . (12) With equations ⃗( ) becomes y→T gives another rank N row vector in which all elements are ) )o ∑ ob( use the notation . (12) We denote (time-independent) value of . )notation ∑ the( ⃗( to ⃗ . (13a) initial zero (because the sum of each column of Q is zero). This We use⃗ the i to denote the (time-independent) → →T → servation means that i y i=0. The pre-multiplication of The i by initial value of(11) vector . With equations (11)vector and (12), the. probability iand )of ) is initial With equations (12), ⃗( becomes rating probability ⃗ the ⃗( probability We use the notationinitial to denote the (time-independent) value y→T simply sums the elements of →i. Since i0, the sum of the vector →r(t) becomes the eigenvalue ⃗) ⃗ . ⃗(13b). (13a) ∑ With equations (11) and (12), ⃗(the probability vector ⃗( ) becomes 4 RISK PROFESSIONAL A U G U S T 2 0 1 2 ∑ Since the eigenvectors ⃗ are known, (13b) to determine the ) is ⃗( equation The initial rating probability vector ⃗we use ⃗( ) ∑ given the choice of ⃗ . The initial rating probability vector ⃗ ⃗ ∑ ⃗( ) is ⃗ . (13a) ⃗ . (13b) www.garp.org Since the eigenvectors ⃗ are known, we use equation (13b) to determine the given the choice of ⃗ . 11),(11), the the coefficients time-dependent while the the eigenvectors ⃗ ⃗ on coefficients are are time-dependent while eigenvectors Applying equation (3), (3), ant. Applying equation ⃗ ∑ ∑( ⃗ ( ⃗ ∑∑ ⃗ ) ⃗) ⃗ ⃗ ∑∑ ⃗ ( )( ) ⃗ QU A N⃗ T P E R S P E C T I V E S ∑ ∑ ⃗ 2012 2012 . (12) . (12) J. M. Pimbley J. M. Pimbley otation the (time-independent) initial value of vector and all other zero. From equation one with as theone ascomponent he notation to denote to denote the (time-independent) initial value of. . with component and elements all other elements zero. From equation vector the (7), this is(7), alsothis theis alsoeigenvector ⃗ . Stated⃗ symbolically, equations (11)(11) and and (12),(12), the probability vector ⃗( )⃗(becomes the eigenvector . Stated symbolically, ) becomes ith equations the probability vector ⃗( )⃗( ) ∑ ∑ ⃗ ⃗. (13a) . (13a) ting probability vector ⃗ ⃗ ⃗( )⃗(is ) is → l rating probability vector The initial rating probability vector ro = r(0)→ is (13a) ⃗( ) ( ⃗( )) ⃗ ) ( ⃗ . (14) . (14) (14) Byofinspection of(13a), equation (13a), we can improve the equaBy(13b) inspection equation we(13a), can we improve the equation (14) Bytion inspection of equation can improve thelargest equation (14) (14) asymptotic expression by retaining the next → asymptotic expression by retaining the next largest term: the eigenvectors i are known, use equation to term: asymptotic expression by retaining the next largest term: envectors ⃗Since are known, we we use use equation (13b) towe determine the (13b) eigenvectors ⃗ are known, equation (13b) → to determine the determine the io given the choice of ro. ice of ⃗of. ⃗ . ⃗( ) ⃗⃗( ) ⃗ . (15) . (15) choice (15) ⃗ ⃗ ⃗ ⃗ ∑ ∑ ⃗ ⃗. (13b) . (13b) → The probability character of the vector r (t) is preserved The ordering of eigenvalues of equation (8) motivates this expression. The → The all ordering of eigenvalues of equation (8) motivates this expression. The To interpret the vector r (t) as a “rating probability” vector, The ordering of eigenvalues of equation (8) motivates this exand we designate this as the “penultimate smallest positive eigenvalue is its elements must be non-negative and sum to one. Clearly, one pression. The smallest positive eigenvalue is and we desigand we designate thisN-1 as the “penultimate smallest positive eigenvalue is → would choose an probability vector r o that satisfies theseWe nate this N-1 as the “penultimate eigenvalue.” We consider N eigenvalue”. consider ( ) to be the “last eigenvalue” and makeand the 10initial 10 eigenvalue”. We consider ( ) to be the “last eigenvalue” make the conditions. From equation (13b), we see requiring the sum of (=0) to be the “last eigenvalue” and make the assumption that → assumption that the antepenultimate eigenvalue, , is strictly greater than the assumption the antepenultimate eigenvalue, , isgreater strictlythan greater the elements of r o to be one implies that No=1.16 With this as- thethat antepenultimate eigenvalue, N-2, is strictly the than the penultimate eigenvalue. All other terms of equation (13a) are exponentially smaller signment, equation (13a) then shows that the sum of elements penultimate eigenvalue. All other terms of equation (13a) are penultimate eigenvalue. All other terms of equation (13a) are exponentially smaller → of r (t) for all t >0 is also one (since N =0). exponentially smaller thetotwo terms of equation (15) as than the two terms of equation (15) as timethan goes infinity. → than the two terms of to equation To show that all elements of r (t) remain non-negative protime t goes infinity.(15) as time goes to infinity. → vided they are non-negative at time t, consider r (t+∆t).AFor suf- feature A striking feature (15) of equation (15)“penultimate is that the “penultimate striking equation is that the eigenvector” A strikingoffeature → → → of equation (15) is that the “penultimate eigenvector” ficiently small ∆t, r (t+∆t)[I–Q∆t]r (t). Given that Q has the eigenvector” N-1 provides a fixed distribution of rating probprovides aprovides fixed distribution of rating probabilities above theabove defaultthestate (N). state (N). → ⃗ a fixed distribution of rating probabilities default ⃗ properties (4a)-(4d) and that the elements of r (t) are non-negaabilities above the default state (N). That is, imagine that the That is, imagine that the probability of each rating level above D is specified as tive, the observation that all elements of the matrix I–Q∆t probability rating level above D is specified as Thatareis, imagine that of theeach probability of each rating level above D being is specified as → → positive implies that all elements of r (t+∆t) arebeing non-negative. proportional to each corresponding element ofThen, N-1. Then, as . as time proportional to each corresponding element of ⃗ Then, being proportional to each corresponding element of ⃗ of .each Applying these time steps ∆t repeatedly shows that all elements time increases, the relative probability “amount” rat- as time → increases, the relative probability “amount” of each rating above D remains of r (t) remain non-negative as time t increases. D remains the “amount” same evenofthough the absolute increases,ing theabove relative probability each rating above probD the remains the -N-1t It is worth noting that this property of persistent abilities of each rating are diminishing withdiminishing the e exponensame non-negaevensame though the absolute probabilities of each rating are with the even though the absolute probabilities of each rating are diminishing with the → tivity for the elements of r (t) does not arise from a non-nega- tial of equation (15). exponential of equation If the financial world deliberately added new → of (15). equation (15). If the financial added new tivity property of the elements of the eigenvectors i (i=1,..., N) exponential If the financial world deliberately addedworld newdeliberately bonds to re→ to bonds replaceto place bonds that default, then one could interpret the penultimate of the HRM Q. Other than the eigenvector Nbonds corresponding bonds that default, then one could interpret the penultireplace bonds that default, then one could interpret the penultimate → to the zero eigenvalue, all the i will have at least one negative mate eigenvector as rating providing a steady-state rating distribution eigenvector as providing a steady-state distribution above D. In reality, eigenvector as providing a steady-state rating distribution above D. of Inthere reality, there element (given that the sum of vector elements is zero and the above D. In reality, there is no replacement mechanism this →we will think of ⃗ as providing is no replacement mechanism of this type, so elements are not all zero). type, so we will think as so providing as providing is no replacement mechanism of of thisN-1 type, we will merely think ofa ⃗“natural distribution.” Regardless of the initial rating state, the merely a merely “naturala rating rating distribution”. Regardless of the initial “natural rating distribution”. Regardless of the initial rating state, the Asymptotic Behavior of Credit Ratings for Large rating probability vector will evolve over time to the shape of rating probability vector will evolve timeover to the shape of shape this “natural rating probability vector willover evolve time to the of this rating “natural rating Time this “natural rating distribution.” → Just as we know that the survival probability S(t)distribution”. for the simple Given this meaning of , we conclude that each of the N-1 distribution”. hazard rate model of equation (2) must approach zero as time first N-1 elements of this penultimate eigenvector must either , we that each that of the firstofN-1 Given thisGiven ⃗ ,sign we conclude each the elements firstthese N-1 elements meaning of ⃗conclude t goes to infinity, we have a similar expectation for the rating meaning havethis theof same algebraic or be zero. To add clarity, → of this penultimate eigenvector must have the same algebraic sign or be zero. Tozero. To probability vector r (t) as t→ . N-1 elements cannot consist of both positive and negative enof this penultimate eigenvector must have the same algebraic sign or be → Regardless of the initial rating state r o, all bonds eventually tries. This statement is not an additional requirement to place default. Hence, all probability will accumulate in the default on the HRM. Rather, the properties we developed earlier are state. The probability vector will approach the column vector sufficient to prove this12behavior. 12 → with one as the Nth component and all other elements zero. Since r (t) maintains its character as a probability vector, then → From equation (7), this is also the Nth eigenvector N. Symboli- it must do so in the asymptotic limit expressed in equation (15). → cally, this is depicted as follows: If the first N-1 elements of N-1 are non-negative, (non-posi- www.garp.org A U G U S T 2 0 1 2 RISK PROFESSIONAL 5 QUANT PERSPECTIVES tive), then the constant N-1o is positive (negative). One can also show that the Nth element of the penultimate eigenvector must have the opposite algebraic sign to that of the first N-1 elements. Equations (13a) and (15) also show why complex eigenvalues are undesirable. Earlier, we announced a restriction on the HRM to ensure that all eigenvalues are real. In the absence of this restriction, complex eigenvalues would be possible and they would occur in complex conjugate pairs. If the penultimate eigenvalue N-1 were one of a complex conjugate pair, then we would need to include its conjugate, the antepenultimate eigenvalue N-2, in the asymptotic expression of equation (15), since both N-1 and N-2 would have the same real part. This inclusion is certainly feasible mathematically, but it produces an asymptotic rating probability vector → r (t) that oscillates like a cosine function. Such oscillation is not permitted within the context of rating transition to default, so this is the basis for our forbidding complex eigenvalues. We should add that our “anti-complexity argument” strictly applies only to the smallest non-zero eigenvalues N-1 and N-2, but intuition suggests a ban on all complex eigenvalues is most reasonable. rating state.17 Hence, since it is not possible to transition out of the default state D, the last column is filled with zeroes. The bottom row consists of default hazard rates, because the values connote transition into the default state D. To create the example of Figure 1, we first conjured the diagonal entries in a manner that is loosely consistent with values of Moody’s (2011). We imposed the behavior of the diagonal value monotonically increasing from the highest (AAA) category to the lowest (CCC) category above default, because the CRAs often describe their ratings as being less volatile in the higher categories. For off-diagonal entries, we imposed plausible default hazard values in the last row, geometric decreases in absolute values moving up and down from the diagonal, and the zero-sum condition for each column (equation (4d)). To compute eigenvalues and eigenvectors of the Figure 1 and other matrices, we apply the methods and algorithms of Press, Flannery, Teukolsky and Vetterling (1992).18 For this first example, we list the eigenvalues and associated eigenvecPimbley tors 2012 in Figure 2 (below) with all values rounded toJ. M. the nearest 0.001. Practical Result - Examples As we related earlier, we do not gather empirical data from the 2012 J. M. Pimbley CRAs to test this hazard rate matrix model for rating transitions since rating agency downgrades and upgrades are markPractical Result - Examples edly non-Markovian. Rather, we create and study an idealized rating transition model. As a first example, consider the HRM As we related earlier, we do not gather empirical data from the CRAs to test ofthisFigure 1 (below) with all values rounded to the nearest hazard rate matrix model for rating transitions since rating agency downgrades 0.0001. and upgrades are markedly non-Markovian. Rather, we create and study an Figure 2: Eigenvalues and Eigenvectors for the HRM of Figure 1 idealized rating transition model. Figure 1: Example of Hazard Rate Matrix with N = 8 (0.0665) 0.2100 (0.0739) (0.0370) (0.0185) (0.0092) (0.0046) (0.0002) (0.0334) (0.0668) 0.2400 (0.0743) (0.0371) (0.0186) (0.0093) (0.0005) (0.0181) (0.0361) (0.0723) 0.2700 (0.0803) (0.0402) (0.0201) (0.0030) (0.0099) (0.0198) (0.0395) (0.0791) 0.3000 (0.0878) (0.0439) (0.0200) (0.0057) (0.0115) (0.0230) (0.0459) (0.0918) 0.3300 (0.1021) (0.0500) (0.0022) (0.0044) (0.0089) (0.0178) (0.0356) (0.0711) 0.3600 (0.2200) 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 With N = 8, the rating categories are {AAA, AA, A, BBB, BB, B, CCC, D}. Each With N = 8, the rating categories are {AAA, AA, A, BBB, off-diagonal entry is related to a probability of transition from the column rating BB, B, CCC, D}. Each off-diagonal entry is related to a probstate to the row rating state.16 Hence, the last column is filled with zeroes since it is ability of transition from the column rating state to the row not possible to transition out of the default state D. The bottom row consists of default hazard rates since the values connote transition into the default state D. To create the example of Figure 1, we first conjured the diagonal entries in 6 Figure 2: Eigenvalues andwe list Eigenvectors for the the eigenvalues and associated Vetterling (1992). For this first example, HRMeigenvectors of Figure in Figure 21with all values rounded to the nearest 0.001: 17 Eigenvalues 1 0.440 Eigenvector 1 (0.000) Formatted: Page2break(0.000) before Components 3 (0.001) 4 (0.009) 5 0.137 6 (0.341) 7 0.363 8 (0.149) 2 0.384 3 0.335 4 0.293 5 0.238 6 0.148 7 0.021 8 0.000 (0.000) (0.001) (0.026) 0.184 (0.326) 0.042 0.274 (0.147) 0.001 (0.048) 0.219 (0.256) (0.072) 0.105 0.174 (0.124) (0.057) 0.237 (0.185) (0.142) 0.025 0.110 0.128 (0.115) 0.175 (0.169) (0.169) (0.025) 0.089 0.121 0.114 (0.138) (0.210) (0.086) 0.031 0.111 0.139 0.123 0.096 (0.204) 0.092 0.095 0.092 0.080 0.064 0.046 0.031 (0.500) 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.000 As a first example, consider the HRM of Figure 1: Example of Hazard Rate Matrix with Figure 1 below with all values rounded to the nearest 0.0001: N=8 0.1500 (0.0761) (0.0381) (0.0190) (0.0095) (0.0048) (0.0024) (0.0001) To compute eigenvalues and eigenvectors of the Figure 1 and other matrices, we apply the methods and algorithms of Press, Flannery, Teukolsky, and a manner that is loosely consistent with values of Moody’s (2011). We imposed RISKthe PROFESSIONAL A Uvalue G U monotonically S T 2 0 1 increasing 2 behavior of the diagonal from the highest (AAA) category to the lowest (CCC) category above default since the CRAs often Since an eigenvector multiplied by any non-zero constant remains an eigenvector of the same eigenvalue, we aremultiplied free to choose a normalizing In Figure 2 Since an eigenvector by anycondition. non-zero constant and an othereigenvector representations of eigenvectors, our normalization is that thewe sumare of thefree to remains of the same eigenvalue, values of the elements is one and the last element of each eigenvector is chooseabsolute a normalizing condition. In Figure 2 and other reprenegative (with the exception of the last eigenvector). sentations of eigenvectors, our normalization is that the sum Consistent with equation (8), we’ve ordered the eigenvalues in Figure 2 of the absolute values of the elements is one and that the last . The from largest to smallest. The last and smallest eigenvalue element of each eigenvector is negative (with the exception of eigenvector for this last eigenvalue, ⃗ , is shown in the last column of Figure 2 to the lastconsist eigenvector). of all zeroes for the first seven elements and 1 for the last element as we Consistent with equation (8), iswe’ve ordered thepenultimate eigenvalues and the expect. The penultimate eigenvalue eigenvector lies in largest the columnto beneath this eigenvalue. of the in Figure 2 from smallest. The The lastcomponents and smallest ei→ penultimate above the D rating the eigenvalue, “natural rating , is genvalue 8=0.eigenvector The eigenvector for represent this last 8 for the HRM of Figure 1. These seven components that begin with showndistribution” in the last column of Figure 2 to consist of all zeroes for the first seven elements and 1 for the last element, as we More specifically, we reduced the HRM to upper Hessenberg form by a series of similarity 17 transformations. We then applied an iterative “QR algorithm” to determine the eigenvalues of the upper Hessenberg matrix (which are also the eigenvalues of the original HRM). With these eigenvalues, we next calculated all eigenvectors by inverse iteration. 15 www.garp.org QUANT PERSPECTIVES expect. For this second example, we list the eigenvalues and associThe penultimate eigenvalue is 7=0.021 and the penulti- ated eigenvectors in Figure 4, with all values rounded 2012 J. M. Pimbley to the mate eigenvector lies in the column beneath this eigenvalue. nearest 0.001. Standard & Poor’s (2008). The diagonal values differ markedly between the two The components of the penultimate eigenvector above the D HRM examples. rating represent the “natural rating distribution” for the HRM Figure 4:ForEigenvalues and Eigenvectors for this first example, we list the eigenvalues and associated eigenvectors in of Figure 1. These seven components, which begin with 0.092 the HRM ofallFigure Figure 2 with values rounded2 to the nearest 0.001: and 0.095 and end with 0.031, show this natural distribution Figure 4: Eigenvalues and Eigenvectors for the HRM of Figure 2 8 7 6 5 4 3 2 1 is primarily (and somewhat uniformly) in the AAA, AA and A 0.449 0.314 0.214 0.173 0.113 0.061 0.006 0.000 Eigenvalues ratings, and then trails off at lower ratings. Eigenvectors 1 (0.000) 0.001 (0.001) 0.013 0.134 (0.373) 0.185 0.000 Based on equations (13a) and (13b), a group of bonds that Components 2 (0.000) 0.004 (0.007) 0.159 (0.412) 0.086 0.110 0.000 3 (0.001) 0.020 0.148 (0.422) 0.019 0.115 0.074 0.000 all have initial AA ratings will have a rating probability vector 4 (0.004) 0.103 (0.412) 0.043 0.132 0.130 0.061 0.000 2012 J. M. Pimbley that we write as a linear combination of the eigenvectors of 5 (0.033) (0.434) 0.106 0.110 0.095 0.079 0.035 0.000 6 (0.323) 0.285 0.178 0.126 0.085 0.063 0.025 0.000 the HRM. Assuming the HRM of Figure 1, the eigenvalues of 7 0.500 0.087 0.068 0.050 0.035 0.027 0.011 0.000 0.092 and 0.095 and end with 0.031 show this natural distribution is primarily (and 8 (0.137) (0.066) (0.080) (0.078) (0.088) (0.127) (0.500) 1.000 the top row of Figure 2 give the rate of exponential decay of somewhat uniformly) in the AAA, AA, and A ratings and then trails off at lower each eigenvector component. Since the penultimate eigenvalratings. Our qualitative observations for the eigenvalues and eigenvectors of Figure ue 7=0.021, the “time constant” for decay of the penultimate 2 (the first example) also hold for this second example. The last and penultimate Basedcomponent on equations (13a) (13b), a group ofrating bonds that all have initial eigenvector (i.e.,andthe “natural distribution Our eigenvectors qualitative observations forthe the eigenvalues and eigenare of the form we expect. Here penultimate and antepenultimate AA ratings will have a rating probability thatofwe0.021). write asThus, a linear vectors eigenvalues are smaller than first those ofexample) the first examples which implies the this time second component”) is roughly 48 years (the vector inverse of Figure 2 (the also hold for constants for decay of eigenvector components to the initial rating probability of theiseigenvectors of the HRM. Assuming the HRM of Figure 1, the example. thiscombination component long-lived. The last and penultimate eigenvectors are of the vector will be longer. The natural rating distribution (proportional to penultimate eigenvalues the top (antepenultimate) row of Figure 2 give theeigenvalue rate of exponential of each form we expect. Here, the penultimate and antepenultimate The next of largest is decay 6=0.148, eigenvector) is more skewed to AAA in this second example which may be due to the “time eigenvalues eigenvector component. Since the penultimate eigenvalue which implies a much shorter time constant of just about seven than ofof the example, which the smallare AAA smaller diagonal element (0.04)those of the HRM Figure first 3. years. Hence, afterofroughly seveneigenvector years, the rating distribution timewe constants for indecay of eigenvector constant” for decay the penultimate component (i.e., the “natural implies the Finally, present a third example which the HRM, like that of Figure 1, compomotivated by therating Moody’s (2011) rating transition data. Now add longer. in the of the undefaulted bonds that(the were initially rated initial probability vector willwe be The ratingremaining, distribution component”) is roughly 48 years inverse of 0.021). Thus, nents tois the in which(proportional all letter ratings from AA to CCC have eigenAAthis willcomponent be approximately proportional to the natural rating is naturalcustomary ratingsub-categories distribution todown penultimate is long-lived. The next largest (antepenultimate) eigenvalue distinct rating levels (described by “+” and “-“ signs or by numerical distribution (components of the penultimate eigenvector of vector) three is more skewed to AAA in this second example which which implies a much shorter time constant of just about 7 years. Figure 2). may be due to the small AAA diagonal element (0.04) of the Hence, after roughly 7 years the rating distribution of the remaining, undefaulted We emphasize again that we are not attempting here to infer hazard rate matrices from rating agency data. This is inappropriate because the amount of data is insufficient to generate the precision we ascribe to As a second example, consider the alternative HRM of FigHRM of Figure 3. our matrices and because credit ratings in practice are not Markovian. Our intent in referencing past data bonds that were initially rated AA will be approximately proportional to the natural from Moody’s Investors Service and Standard & Poor’s is simply to show that the numerical values we use are similarwe to real-world credit rating ure rating 3 (below), with all values rounded toeigenvector the nearest 0.0001. Finally, present a transition thirddata.example in which the HRM, like distribution (components of the penultimate of Figure 2). 17 by the Moody’s (2011) rating that ofFormatted: Figure 1, is motivated No page break before As a second example, consider the alternative HRM of Figure 3 below with Figure 3: Example of Hazard Rate Matrix with transition data. Now we add in the customary sub-categories all values rounded to the nearest 0.0001: N= 8 in which all letter ratings from AA down to CCC have three Figure 3: Example of Hazard Rate Matrix with N = 8 distinct rating levels (described by “+” and “-” signs or by numerical modifiers or by “high” and “low” designations). 0.0400 (0.0316) (0.0212) (0.0115) (0.0087) (0.0048) (0.0040) 0.0000 (0.0203) 0.0900 (0.0425) (0.0230) (0.0175) (0.0095) (0.0079) 0.0000 Instead of N=8, this example has N=20 rating levels. Rather (0.0101) (0.0301) 0.1400 (0.0460) (0.0350) (0.0190) (0.0159) 0.0000 than showing all 400 elements of this HRM, we describe it as (0.0051) (0.0150) (0.0404) 0.1600 (0.0699) (0.0381) (0.0317) 0.0000 (0.0025) (0.0075) (0.0202) (0.0438) 0.2400 (0.0761) (0.0635) 0.0000 Figure 1, with interpolation along the bottom row (transition (0.0013) (0.0038) (0.0101) (0.0219) (0.0666) 0.2600 (0.1270) 0.0000 to default) and along the diagonals with one modification: the (0.0006) (0.0019) (0.0051) (0.0109) (0.0333) (0.0725) 0.4000 0.0000 diagonal elements need to increase by at least 50% relative (0.0001) (0.0002) (0.0005) (0.0030) (0.0090) (0.0400) (0.1500) 0.0000 to the Figure 1 diagonal values. This increase in the diagonal elements is sensible, because increasing the number of rating Whereas the HRM of Figure 1 is an approximate, idealized version of data from Whereas the HRM of Figure 1 is an approximate, ideal- levels should necessarily increase the number of rating transiMoody’s (2011), Figure 3 is an idealized and approximate rendering of data in ized version of data from Moody’s (2011), Figure 3 is an ideal- tions. If we do not increase the diagonal values in this manner, ized and approximate rendering of data in Standard & Poor’s then the HRM will have complex pairs of eigenvalues. (2008). The diagonal values differ markedly between the two This HRM for 20 rating levels has 20 eigenvalues and 20 HRM examples.19 corresponding eigenvectors. The eigenvalues are all real and 18 18 16 www.garp.org A U G U S T 2 0 1 2 RISK PROFESSIONAL 7 QUANT PERSPECTIVES distinct, and range from a high of 0.662 for 1 down to 0.013 and 0.000 (zero) for 19 and 20, respectively. Figure 5 (below) shows the largest two eigenvalues with associated eigenvectors and the smallest three eigenvalues with eigenvectors (with all values rounded to the nearest 0.001). Figure 5: Five of the Eigenvalues and Eigenvec2012 J. M. Pimbley tors for the HRM with N = 20 Figure 5: Five of the Eigenvalues and Eigenvectors for the HRM with N = 20 Eigenvalues Eigenvectors Components 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 20 0.662 19 0.632 3 0.080 2 0.013 1 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) 0.002 (0.017) 0.075 (0.196) 0.310 (0.222) 0.113 (0.065) 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 (0.001) 0.007 (0.038) 0.116 (0.191) 0.122 0.103 (0.170) 0.152 (0.101) (0.050) (0.056) (0.045) (0.032) (0.015) 0.003 0.020 0.034 0.045 0.052 0.056 0.057 0.054 0.049 0.042 0.034 0.026 0.017 0.012 (0.303) 0.035 0.044 0.045 0.045 0.044 0.042 0.039 0.036 0.032 0.028 0.025 0.021 0.017 0.014 0.011 0.009 0.006 0.004 0.003 (0.500) 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.000 The two columns furthest to the right of Figure 5 show that the last and penultimate eigenvalues and eigenvectors haveright the properties we derived and that The two columns furthest to the of Figure 5 show earlier. In particular, eigenvalues the penultimate eigenvector gives the shape have of the the the described last and penultimate and eigenvectors natural rating distribution. properties we derived and described earlier. In particular, the penultimate eigenvector gives the shape of the natural rating distribution. Parting Thoughts This article considers the well-known credit rating transition dynamic and re-formulates the problem in continuous time with a hazard rate matrix. As an extension and analogy with the hazard rate model for bond default, we create the rating 19 probability vector and show how the hazard rate matrix governs its evolution in time. We explain the mathematical properties that this matrix must satisfy (both by imposition and by proof) and also deduce analytical insights of the hazard rate matrix (HRM). The eigenvalues and eigenvectors of this HRM provide the best view of the behavior of bond credit ratings at large time 8 RISK PROFESSIONAL A U G U S T 2 0 1 2 from initial rating. In particular, the penultimate eigenvector gives the shape of the “natural rating distribution.” We believe this study is the first to identify and derive this rating distribution that prevails long after inception. Our analysis applies only to credit rating migration that is Markovian. We do not attempt to fit or explain actual credit rating agency data, since rating changes of the large public rating agencies are decidedly not Markovian. Rating accuracy, however, requires Markovian evolution. As investors, banks or asset managers create their own (internal) rating scales and procedures, it is our hope that the framework, analysis and results of this article will be helpful. Footnotes 1. The author is Principal of Maxwell Consulting (http://www.maxwell-consulting.com/) and welcomes followers on Twitter (http:// www.twitter.com/JoePimbley). 2. There exist 10 to 20 CRAs worldwide that have some significant form of governmental, regulatory or market recognition. The largest four are Moody’s Investors Service, Standard & Poor’s, Fitch Ratings and DBRS. 3. See, for example, Moody’s (2011), Standard & Poor’s (2008), Fitch (2012) and DBRS (2012). 4. The published default tables are often associated with the structured finance ratings of the CRAs, with no clear statement that these tables also apply to the ratings of non-structured finance debt such as sovereigns or banks. Further, Moody’s Investors Service describes its ratings in terms of “expected loss” rather than “default probability,” but the conversion between these two risk measures is straightforward. 5. See, for example, Moody’s (2011), Standard & Poor’s (2008), Fitch (2012) and DBRS (2012). 6. Alternatively, the CRA may define its transition matrix such that the default probabilities occupy the last column rather than the last row. 7. See, for example: Israel, Rosenthal and Wei (2001), and Gupton, Finger and Bhatia (2007). 8. See, for example, Lando and Skødeberg (2002), Altman and Kao (1992), and Carty and Fons (1993). 9. See the Moody’s Investors Service “Special Comment” of Mann and Metz (2011). 10. See, for example, Löffler and Posch (2011). 11. One interprets the exponential function of a matrix argument as the MacLaurin series expansion of the exponential with scalar product powers of the matrix argument. www.garp.org QUANT PERSPECTIVES 12. See, for example: Israel, Rosenthal and Wei (2001); Lando and Skødeberg (2002); and Jarrow, Lando and Turnbull (1997). 13. Jarrow, Lando and Turnbull (1997) also wrote a “Kolmogorov equation” similar to equation (3) to relate the “generator matrix” to the time evolution of the RTM. Our work differs from this study in our focus on the HRM and the rating probability vector. 14. As an example, imagine there are only four rating states (includ0.15 -0.17 -0.01 0 ing the default state), so that N=4. Let the HRM be -0.01 0.20 -0.02 0. ( -0.13 -0.01 0.25 -0.01 -0.02 -0.22 0 0 ) Of the four eigenvalues of this matrix, two are real (0 and 0.097) and two are complex (0.252 + i 0.048 and 0.252 – i 0.048). Relatively small changes to some off-diagonal matrix elements produce four real eigenvalues. 15. See, for example, Weisstein (2012). In our application of this theorem, we consider the sum of column (rather than row) element absolute values for the Gerschgorin radius. This substitution is permissible since a matrix and its transpose have the same eigenvalues. 16. Recall that the sum of all elements of each eigenvector → vi is zero for iN and that the sum of the elements of → vN is one from equation (7). 17. As stated previously, the HRM Q has the meaning that I – Q∆t is the rating transition matrix for the time period ∆t,where I is the identity matrix. Thus, the ij element of I – Q∆t is the probability that a bond in rating level j will transition to rating level i in the time period ∆t. 18. More specifically, we reduced the HRM to upper Hessenberg form by a series of similarity transformations. We then applied an iterative “QR algorithm” to determine the eigenvalues of the upper Hessenberg matrix (which are also the eigenvalues of the original HRM). With these eigenvalues, we next calculated all eigenvectors by inverse iteration. 19. We emphasize again that we are not attempting here to infer hazard rate matrices from rating agency data. This is inappropriate because the amount of data is insufficient to generate the precision we ascribe to our matrices and because credit ratings in practice are not Markovian. Our intent in referencing past data from Moody’s Investors Service and Standard & Poor’s is simply to show that the numerical values we use are similar to real-world credit rating transition data. Joe Pimbley (PhD) is Principal of Maxwell Consulting, a consulting firm he founded in 2010, and a member of Risk Professional’s editorial board. His recent and current engagements include financial risk management advisory, underwriting for structured and other financial instruments, and litigation testimony and consultation. In a prominent engagement from 2009 to 2010, Joe served as a lead investigator for the Examiner appointed by the Lehman bankruptcy court to resolve numerous issues pertaining to history’s largest bankruptcy. www.garp.org References E. Altman and D. Kao. “The implications of corporate bond rating drift,” Financial Analysts Journal, 64-75, 1992. B. Belkin, S. Suchower and L. Forest Jr. “A one-parameter representation of credit risk and transition matrices,” Credit Metrics Monitor, 1998. L. Carty and J. Fons. “Measuring changes in corporate credit quality,” Moody’s Special Report, 1993. J. Christensen, E. Hansen and D. Lando. “Confidence sets for continuous-time rating transition probabilities,” J. Banking & Finance 28, 2575-2602, 2004. DBRS Industry Study. “2011 DBRS Corporate Rating Transition and Default Study,” March 2012. Fitch Ratings Special Report. “Fitch Ratings Global Corporate Finance 2011 Transition and Default Study,” March 2012. G. M. Gupton, C. C. Finger and M. Bhatia. CreditMetrics – Technical Document, RiskMetrics Group, Inc., 2007. R. B. Israel, J. S. Rosenthal and J. Z. Wei. “Finding generators for Markov chains via empirical transition matrices with applications to credit ratings,” Mathematical Finance 11(2), 245-265, 2001. R. A. Jarrow, D. Lando, and S. M. Turnbull. “A Markov model for the term structure of credit risk spreads,” Review of Financial Studies 10, 481-523, 1997. J. Keilson and A. Kester. “Monotone matrices and monotone Markov processes,” Technical Report No. 79 for National Science Foundation Grant GP-32326X, Stanford University, 1974. D. Lando and T. M. Skødeberg. “Analyzing rating transitions and rating drift with continuous observations,” J. Banking & Finance 26, 423-444, 2002. G. Löffler and P. N. Posch. Credit risk modeling using Excel and VBA, Second Edition, John Wiley & Sons, Ltd., 2011. J. Mann and A. Metz. “Measuring the Performance of Credit Ratings,” Moody’s Investors Service Special Comment, November 2011. Moody’s Investors Service Special Comment. “Corporate Default and Recovery Rates, 1920-2010,” February 2011. W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling. Numerical Recipes in C: The Art of Scientific Computing, 2nd Edition, Cambridge University Press, 1992. B. Singer and S. Spilerman. “The representation of social processes by Markov models,” American Journal of Sociology 82, 1-54, 1976. Standard & Poor’s Ratings Direct. “2007 Annual Global Corporate Default Study and Rating Transitions,” February 2008. H. Takada and U. Sumita. “Credit risk model with contagious default dependencies affected by macro-economic condition,” European Journal of Operational Research, 2011. E. W. Weisstein. “Gerschgorin Circle Theorem,” from MathWorld – a Wolfram Web Resource: http://mathworld.wolfram.com/GershgorinCircleTheorem.html, 2012. A U G U S T 2 0 1 2 RISK PROFESSIONAL 9