See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/215991426

Dependent Credit Migrations

Article  in  Journal of Credit Risk · January 2004

DOI: 10.21314/JCR.2006.043

CITATIONS READS

36 123

2 authors, including:

Alexander John McNeil

The University of York
80 PUBLICATIONS   9,652 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Operational Risk View project

All content following this page was uploaded by Alexander John McNeil on 20 May 2014.

The user has requested enhancement of the downloaded file.

 Dependent Credit Migrations

JONATHAN WENDIN∗ ALEXANDER J. McNEIL

Department of Mathematics, ETH Zürich

First version: July 2004; Current version: January 2006

Abstract

This paper examines latent risk factors in models for migration risk. We employ the
standard statistical framework for ordered categorical variables and induce dependence
between migrations by means of latent risk factors. By assuming a Markov process for
the dynamics of the latent factors, the model can be interpreted as a state space model.
The paper contains an empirical study on quarterly migration data from Standard &
Poor’s for the years 1981–2000, in which the ordered logit model with serially correlated
latent factors is fitted by computational Bayesian techniques (Gibbs sampling). Apart
from highlighting the usefulness of the Gibbs sampler for statistical inference in models of
this kind, the survey in particular investigates the issues of rating-specific factor loadings
and heterogeneity among industry sectors, with emphasis on their implications in terms
of implied asset correlations.

J.E.L. Subject Classification: C15, C23, C35

Keywords: Credit risk, State space models, Multivariate random effects, Gibbs sam-
pling

1 Introduction

An ordered categorical variable expressing the ability of a company to fulfill its financial
obligations is known as a credit rating (or simply rating). Such classifications are provided
by agencies such as Moody’s and Standard and Poor’s, but could also be according to an
internal system of a bank. With the recent Basel II accord the importance of ratings as a tool
for risk management has increased (BCBS (2005)). This has lead to interest in statistical
models for the dynamics of ratings; the change in rating of an obligor is referred to as a
transition or a migration.
Firstly, it is obvious that the present rating of an obligor is a strong predictor for its
rating in the nearest future. A cardinal feature of any migration model is hence past and
present ratings influencing the evolution. The Markov chain is a stochastic process of this
∗
e-mail: [email protected]. Financial support from NCCR Financial Valuation and Risk Management
(a research program supported by the Swiss National Science Foundation) is gratefully acknowledged.

1
kind for which the migration probabilities, given all past ratings, depend only on the present
state. This is the simplest form of a Markov chain, and it allows all migration probabilities
for a specific time-horizon to be collected in a so-called migration (or transition) matrix.
If no obligor-specific properties other than rating are considered, the Markov assumption
is very convenient since, without loss of generality, all migrations taking place in a period
can be summarized in cross-sectional migration counts; see Section 2.2.1. The empirical
study of this paper is based on data of this kind. The intuitive and comprehensible form of
the migration matrix has made it a cornerstone of many current risk management systems;
see Jafry and Schuermann (2004) for a review. In most applications, migration matrices are
estimated in discrete time with a monthly, quarterly, or yearly time horizon. As obligors
seldom change rating (and when they do it is often to a neighbouring state), migration
matrices typically have their mass concentrated along the main diagonal. In fact, the low
occurrence of certain transitions may be a problem when estimating migration probabilities
naively by empirical proportions. One way of mitigating this is a continuous-time Markov
approach as in Lando and Skødeberg (2002).
The assumption of a firm’s rating path following a homogeneous Markov chain has been
increasingly questioned and even statistically rejected in recent contributions. An empirically
verified characteristic of agency rating data is rating momentum: for instance, the tendency
of recently downgraded obligors to be more at risk than other obligors. Such observations
contradict the Markov property; see Lando and Skødeberg (2002) and the references therein.
Statistical approaches to model non-Markovian rating paths can be found in Christensen,
Hansen, and Lando (2004) and Frydman and Schuermann (2005). The question of rating
momentum will not be treated in this paper, since our analysis is based on repeated cross-
sectional migration counts. Time inhomogeneity in migration probabilities is, however, an
important topic which we treat next.
Several empirical surveys have found evidence of time-variation in migration intensities
and confirmed that this time variation may to some extent be explained by observed macro-
economic variables, see Nickell, Perraudin, and Varotto (2000); Bangia et al. (2002); and
Hu, Kiesel, and Perraudin (2002). A second requirement of the migration model is therefore
that of dependence among transitions taking place within a time period. The underlying
economic conditions will be referred to as the systematic risk of the portfolio. Unfortunately,
observed variables as proxies for the systematic risk are seldom completely satisfactory. The
first important issue is the identification of appropriate proxies. Moreover, there may also
be a lag between the cycle of a proxy variable and that of the migration activity, and this
lag may vary stochastically over time. The above shortcomings of observed risk factors have
serious implications for regulation; see the discussion in Koopman, Lucas, and Klaassen
(2005).
The approach of this paper will be to complement observed risk factors with latent ones
that capture the residual systematic risk once any observed parts have been accounted for.

2
The unobserved systematic risk gives rise to correlated migrations in each time period; we
refer to this as cross-sectional dependence. We also expect the cyclical behaviour of economic
factors to create serial dependence among migration events in different time periods; see
McNeil and Wendin (2005) for a similar discussion in the context of default risk.
Relatively little work has been carried out on latent systematic risk in migration models.
An early contribution is Kijima, Komoribayashi, and Suzuki (2002), where a (non-standard)
statistical framework for correlated migrations is suggested. Gagliardini and Gouriéroux
(2005b) present a more general framework for rating dynamics, based on stochastic migration
matrices. In particular, they consider serially correlated migration matrices, and perform an
empirical analysis on French corporate data using a linearization of the likelihood function.
The modelling framework of this paper is conceptually close to that of Gagliardini and
Gouriéroux (2005b). We consider a general statistical model for discrete-time migration
counts, employing standard techniques for ordered, categorical responses. The ratings are
subject to both observed and unobserved systematic risk, where the unobserved risk factors
are serially correlated. The latent risk factors and their serial dependence are able to capture
the effects of cross-sectional dependence. If the latent factors are assumed to follow a Markov
process, the models formally belong to the class of state space models (Durbin and Koopman,
2001).
Serially correlated, latent risk factors yield joint migration distributions in terms of high-
dimensional integrals, which are indeed awkward for standard maximum likelihood (ML)
techniques. Koopman, Lucas, and Klaassen (2005) and Gagliardini and Gouriéroux (2005b)
consider models with continuous latent factors, although they model the ratio of defaulted
obligors instead of the actual default counts. This simplification may show undesirable
features, in particular when either the numerator or the denominator are small (Kurbat and
Korablev, 2002). An alternative approach is that of simulated ML; a recent application in the
context of default risk is given in Koopman, Lucas, and Daniels (2005). Koopman, Lucas,
and Monteiro (2005) use simulated ML to fit a duration model with stochastic intensity to
credit transition data.
Instead of following a frequentist approach, we follow McNeil and Wendin (2005) and
use computational Bayesian techniques (Gibbs sampling) for model inference. The algo-
rithms are able to handle varyingly complex specifications of the latent systematic risk, such
as serially correlated factors as well as general multivariate Gaussian risk factors that in-
duce heterogeneity across industry sectors. A brief review of Bayesian statistics and Gibbs
sampling is given in Section 3. The rest of the paper is organized as follows. Section 2 intro-
duces the basic framework for models of cross-sectionally and serially dependent migration
counts. The empirical analysis on quarterly migration data from Standard & Poor’s follows
in Section 4. Section 5 concludes.

3
2 State Space Model of Migration Counts

2.1 Notation

Consider a set K = {1, . . . , K} of rating classes of increasing creditworthiness. The state

default, which we assume is absorbing, is denoted by 0 and is included in K0 := K ∪ {0}.
P
Denote by mtk the number of firms in group k ∈ K, so that mt = k∈K mtk is the total
number of obligors in the portfolio at the beginning of period t. For each obligor i = 1, . . . , mt
in period t, let κ(t, i) denote its initial rating and Rti the rating at the end of the period t.
Notice that the former is necessarily in K, whereas the latter is in K0 .
Our aim is to derive a statistical model for Mt:k,` , the number of obligors with ratings
k ∈ K and ` ∈ K0 at the onset and at the end of period t, respectively. Although mt1 , . . . , mtK
are assumed to be known at the beginning of period t, the migration count variable Mt:k,`
is known only at the end of it. For notational convenience we introduce the migration count
vector M tk := (Mt:k,` )`∈K0 , which summarizes the migrations of the k-rated obligors during
period t. Analogously, M t := (M tk )k∈K .
The framework of this paper presupposes data of repeated cross-sectional type so that
Rti and Rsi , the ratings of obligor i in two distinct periods s 6= t, do not necessarily refer to
the same obligors. Cross-sectional migration counts are easily constructed given a dataset of
rating paths (i.e. data of panel-type, where the migration dates of each obligor are collected),
but the reverse construction is in general not possible.

2.2 Cross-sectionally Dependent Migrations

Assumption 1. Conditional on a latent factor bt (following a non-degenerate distribution

F to be specified), the ratings Rt1 , . . . , Rtmt are conditionally independent and satisfy

P (Rti ≤ ` | bt ) = g(µκ(t,i),` − x0ti β − z 0ti bt ), ` ∈ K0 , (1)

for some strictly increasing function g : R → (0, 1), and a non-decreasing sequence of inter-
cepts (µκ(t,i),` )` ..

xti and z ti of Assumption 1 denote the observed design vectors of obligor i, hold-
ing the corresponding covariates; µκ(t,i),` and β are unknown intercepts and regression
coefficients to be estimated; and g is the response function. Common choices of g are
√ Rx
Φ(x) = (1/ 2π) −∞ exp{−u2 /2} du and 1/(1 + exp{−x}), which lead to the probit and
logit models, respectively.
The formulation in (1) is a generalized linear mixed model (GLMM) for ordered, polyto-
mous responses, where the rating at the onset of period t has been included as a covariate.
If no latent factor bt is present, the model is known as a generalized linear model (GLM),
the theory of which is treated in the monograph by McCullagh and Nelder (1989). For more
information about GLMMs, see Fahrmeir and Tutz (1994) and Skrondal and Rabe-Hesketh
(2004).

4
 The design vectors xti and z ti together with the latent factor bt constitute the systematic
risk x0ti β + z 0ti bt of obligor i in period t. To simplify notation we will often refer to the
systematic risk as γti := x0ti β + z 0ti bt . A fundamental role of xti and z ti is to add observed
elements to the systematic risk: they may hold quantitative variables as well as dummy
variables indicating group membership. By including macro-economic variables, or other
observed risk factors, we may capture time inhomogeneity in transition rates. Likewise,
obligor-specific variables, such as balance sheet data, can be used to capture heterogeneity
among obligors. In fact, current rating is an obligor-specific covariate, which for sake of
clarity has been kept separate from β. Observed explanatory variables are known as fixed
effects in the GLMM-framework.
The latent factors bt (or random effects as they are known as in the literature on GLMMs)
account for unobserved systematic risk, and hereby introduce heterogeneity beyond that
which can be captured with observed covariates. In particular, they induce dependence
among the responses. The latent factors within a time period may be univariate or multivari-
ate; the latter are useful when treating migrations according to industry sector. Parameters
of F are referred to as hyperparameters and will be denoted by θ.
It immediately follows from (1) that

P (Rti = ` | bt ) = g(µκ(t,i),` − x0ti β − z 0ti bt ) − g(µκ(t,i),`−1 − x0ti β − z 0ti bt ), (2)

and that the intercepts (µk,` )k∈K,`∈K0 , which are also known as threshold values or cut-off
levels, for all k ∈ K must satisfy

−∞ = µk,−1 ≤ µk,0 ≤ µk,1 ≤ · · · ≤ µk,K−1 ≤ µk,K = ∞,

in order for the probabilities in (2) to be non-negative. Note that a model for K + 1 ordered
choices requires K thresholds. We refer to the thresholds associated with rating class k as
µk := {µk,` : ` = 0, . . . , K − 1}; the full set of thresholds is denoted by µ := {µk : k ∈ K}.
Thus, the overall number of unknown thresholds amounts to K 2 . For the state default (state
0), which we assume is absorbing, we have µ0,` = ∞ for all ` ∈ K0 .
Unconditionally, Rt1 , . . . , Rtmt of Assumption 1 are not independent, which is most easily
seen by interpreting (2) as a model of threshold-type: Let εt1 , . . . , εtmt be iid rvs with df g,
and independent of bt . Set Vti := εti + x0ti β + z 0ti bt for i = 1, . . . , mt , and notice that Rti is
generated according to

Rti = ` ⇐⇒ Vti ∈ µκ(t,i),`−1 , µκ(t,i),` .

εti is referred to as the idiosyncratic risk of obligor i in period t. It is easy to see that
Vt1 , . . . , Vtmt , despite being conditionally independent given bt , are dependent random vari-
ables. Moreover, their joint distribution (together with the threshold values µ) fully deter-
mines that of (Rt1 , . . . , Rtmt )0 . Vti is often interpreted as the asset value of obligor i, and µi
as critical liability levels as in the seminal work on structural models of credit risk in Merton

5
(1974). This representation is particularly useful for the probit case with a Gaussian random
effect bt (each εti is standard Gaussian under the probit response); the joint distribution
of Vt1 , . . . , Vtmt is then Gaussian, and fully determined by its correlation matrix as in the
well-known industry model CreditMetrics.
The representation with Vti allows us to quantify the cross-sectional migration depen-
dence in period t in terms of the so-called implied asset correlation, corr(Vti , Vtj ), of two
obligors i, j. Given relevant covariates, we trivially have
 q q
corr(Vti , Vtj ) = cov(z 0ti bt , z 0tj bt ) var(z 0ti bt ) + w2 var(z 0tj bt ) + w2 ,


(3)

where w2 := var(εti ). In the probit case, we have w2 = 1, while in the logit case w2 = π 2 /3.

2.2.1 Homogeneous Groups of Obligors

When no obligor-specific covariates are considered, it is often reasonable to the divide the
portfolio into homogeneous subgroups (or buckets) for which all ingoing obligors are assumed
to share transition probabilities. In the framework of Assumption 1, it is clear that the
transition probabilities have two components: the initial rating κ(t, i), and the systematic
risk x0ti β + z 0ti bt . As a consequence, all obligors in a time period that share both (i) rating
class, and (ii) design vectors xti and z ti define a homogeneous bucket. The ingoing members
of a bucket are modelled exchangeably. For notational convenience, we denote by H =
{1, . . . , H} an index set for the H homogeneous buckets.
A commonly seen assumption in practice is homogeneity within rating classes, which is
accomplished by setting xti := x̃tk and z ti := z̃ tk for all i with κ(t, i) = k. This corresponds
to the case H = K. Under Assumption 1 and given bt , we then have the vectors of migration
counts M t1 , . . . , M tK (see Section 2.1 for their definition) conditionally independent with

M tk | bt ∼ Multinomial mtk , pk (x̃0tk β + z̃ 0tk bt )


for all k ∈ K, (4)

where
pk (·) := {pk,` (·)}`∈K0 and pk,` (x) := g(µk,` − x) − g(µk,`−1 − x).

Thus, for each x = (x0 , x1 , . . . , xK ) ∈ {0, 1, . . . , mtk }K+1 satisfying x0 + x1 + · · · + xK = mtk

we have
K
mtk ! Y
P (M tk = x | bt ) = pk,` (x̃0tk β + z̃ 0tk bt )x` .
x0 ! x1 ! · · · xK !
`=0
The unconditional distribution of M tk is evidently not multinomial, since the effect of the
latent factors bt must be integrated out. Furthermore M t1 , . . . , M tK are not independent.
The systematic risk of group k in period t will be referred to as γtk := x̃0tk β + z̃ 0tk bt .
For applications, it might be relevant to perform the grouping according to further
attributes than merely rating class; natural candidates are industry sector and country
membership. In the case of industry sector effects, we have H = K × S, where S is an index
set of industry sectors. A convenient way of introducing heterogeneity between subgroups

6
is by means of multivariate latent factors. The transition properties of the portfolio are
determined by the joint distribution of (bt1 , . . . , btH )0 , where bth denotes the latent systematic
risk of group h ∈ H at time t. This fits naturally into the GLMM-framework by setting
bt = (bt1 , . . . , btH )0 and z̃ th = eh , where eh is the hth unit vector in RH . Applications of
this technique are given in Section 4.2.

2.2.2 Migration Correlations

A consequence of the latent factor bt is dependence between rating migrations occurring in a

time period. It is often convenient to quantify this in terms of the migration correlation, an
entity that relates to two particular obligors and a corresponding (bivariate) rating transition
from (k1 , k2 ) ∈ K2 to (`1 , `2 ) ∈ K02 . It is well-known that migration correlations alone do not
determine the full distribution of the migrations. Nevertheless, they present useful model
summaries when comparing different models for transition risk. An extensive introduction
to migration correlations is given in Gagliardini and Gouriéroux (2005a).
In the present context, migration correlations are derived with the conditional indepen-
dence property and (2). Assuming that κ(t, i) = ki for two obligors i = 1, 2, we have:


cov I{Rt1 =`1 } , I{Rt2 =`2 }
  Y2  
= E E[I{Rt1 =`1 } I{Rt2 =`2 } | bt ] − E E[I{Rti =`i } | bt ]
i=1
hY2
0 0
i Y 2 (5)
E pki ,`i (x0ti β + z 0ti bt )
 
= E pki ,`i (xti β + z ti bt ) −
Z Y i=1 i=1
Y2 Z
2 0 0
= pki ,`i (xti β + z ti bt ) dF (bt ) − pki ,`i (x0ti β + z 0ti bt ) dF (bt ).
i=1 i=1

Since the number of combinations (k1 , k2 ) and (`1 , `2 ) is large, we restrict our attention to
migrations in the same direction, as well as default correlations. In this spirit, we define the
upgrade and downgrade correlations as



corr I{Rt1 >k1 } , I{Rt2 >k2 } and corr I{Rt1 <k1 } , I{Rt2 <k2 } , (6)


respectively. Naturally, the default correlation follows as corr I{Rt1 =0} , I{Rt2 =0} . Given µ,
β, and θ, these expressions are easily calculated numerically.

2.3 Serially Dependent Migrations

The cyclic behaviour of economic factors leads us to expect dependence between migration
events taking place in different time periods. Given the interpretation of bt as a general
state-of-the-economy in period t, it seems reasonable to let its value at time t + 1 depend on
bt , and hence to impose a Markovian structure on the sequence (bt ).

Assumption 2. (i) The sequence (bt ) is a Markov chain; and (ii) conditionally on (bt ), the
M t ’s are independent, and M t depends on bt only.

7
 Assumption 2 defines a state space model (or hidden Markov model, HMM) for the count
sequence (M t ), see Künsch (2001). (Recall that mtk is treated as a known variable, cf.
Chapter 2 of MacDonald and Zucchini (1997) on binomial HMMs.) It is straight-forward to
see that M 1 , . . . , M T in general are not independent in the HMM framework. Assumption 2
thus provides a meaningful serial dependence for the migration counts (M t ), where the
components of M t already exhibit cross-sectional dependence by Assumption 1.
The first-order autoregressive, AR(1), time series (bt )
p
bt = αbt−1 + φt , t ≥ 1, b0 = φ0 / 1 − α2 , (7)

where 0 , 1 , . . . are iid N (0, 1), is a real-valued Markov process in discrete time. For |α| < 1
it has a Gaussian stationary distribution with mean 0 and variance σ 2 := φ2 /(1 − α2 ). Its
d-dimensional counterpart takes the form
p
bt = αbt−1 + A(t,1 , . . . , t,d )0 , t ≥ 1, b0 = A(0,1 , . . . , 0,d )0 / 1 − α2 (8)

with α ∈ (−1, 1), A ∈ Rd×d and (t,i ) iid N (0, 1). The stationary solution is Gaussian with
mean zero and covariance matrix Σ := Φ/(1 − α2 ), where Φ := AA0 . The parameterization
in (7) and (8) turns out to be useful for the applications of Section 4.
Before turning to model calibration, we make a remark on the implications of Assump-
tions 1 and 2 on the rating path of a single obligor. If the rating paths of the obligors follow
first-order Markov chains, then the migration counts M 1 , . . . , M T are easily seen to be inde-
pendent. An important consequence of this observation is that serially correlated migration
counts, as in the state space setting, necessarily imply non-Markovian rating paths. This
circumstance, however, should not be confounded with rating momentum: the tendency of a
recent downgrade to increase the probability of further downgrades, or default, of the same
obligor (Lando and Skødeberg, 2002). Models incorporating obligor-level rating momentum
presuppose a dataset of panel-type, and explicitly model the rating path. Non-Markovian
rating paths in the setting of Assumption 2, on the other hand, should be interpreted as
momentum effects due to past economic conditions.

3 Calibration with Markov Chain Monte Carlo

The unconditional distribution of the migration count vectors (M t ), as defined in Section 2,

requires the effect of the latent factors (bt ) to be integrated out, which greatly complicates
the use of standard maximum likelihood (ML) techniques. This difficulty applies even more
if the latent factors are serially correlated. Unless the likelihood function is approximated
numerically, some kind of simulation-based approach to inference seems inevitable; see for
instance the discussion of simulated ML in Chapter 3 of Gouriéroux and Monfort (1996). An
overview of general approaches to frequentist and Bayesian inference in a very similar context
is given in McNeil and Wendin (2005). The latter highlights the usefulness of computational

8
Bayesian procedures for inference in models of portfolio credit risk, an approach we follow
in this paper as well.

3.1 Bayesian Statistics and Markov chain Monte Carlo

In Bayesian statistics, all unobserved elements ϑ of the statistical model are treated as
random variables on a suitable space Θ. These are assigned a joint prior distribution p(ϑ),
expressing any information about ϑ we might have before observing the data D. Inference is
based on p(ϑ|D), the posterior distribution of ϑ, which is obtained by applying Bayes’ rule.
Point estimates of model parameters can be obtained by calculating the posterior mean. In
the context of Section 2, ϑ holds the parameters µ and β, the hyperparameters θ as well
as the latent risk factors b1 , . . . , bT . The migration count vectors M 1 , . . . , M T and design
vectors constitute the observed data D.
Analytical characterization of the joint posterior distribution is, in general, difficult.
Fortunately, we may simulate from p(ϑ | D) with Markov chain Monte Carlo (MCMC)
techniques. The objective of these is to generate a realization ϑ(1) , ϑ(2) , . . . of an ergodic
Markov chain, whose stationary distribution is the desired posterior. The B iterations prior
to convergence are known as burn-in. A prominent example is the Gibbs sampler which
proceeds by simulating from the so-called full conditional distributions

p(ϑi | ϑ1 , . . . , ϑi−1 , ϑi+1 , . . . , ϑr , D)

in a pre-specified order, where ϑ ≡ (ϑ1 , . . . , ϑr ) is a decomposition of ϑ with r ≥ 2 (Gilks,

1996). By exploiting conjugacy, we obtain full conditionals which are easy to simulate from
(the concept of conjugacy is explained in Casella and Berger (2002)). In many cases of
practical interest, however, non-standard full conditionals necessarily emerge. Note that ϑi ,
i = 1, . . . , r, can be univariate, which greatly facilitates simulation since the ARS (Adaptive
Rejection Sampling) and ARMS algorithms can often be employed (Gilks, 1992). While the
former is intended only for log-concave densities, the latter is capable of handling arbitrary
densities and plays an important role in Section 4. The Gibbs sampler along with other
approaches to non-standard full conditionals are treated in detail in Chapters 6 and 7 of
Robert and Casella (1999).

3.2 Inference with MCMC

(B+1) (B+N )
Inference about a single parameter ϑi ∈ R of the model is based on ϑi , . . . , ϑi ,
which is a sample of size N from p(ϑi|D), the marginal posterior distribution of ϑi . Point es-
timates and standard errors are given by forming the sample mean (or median) and standard
deviation, respectively. Other features of p(ϑi | D) are estimated accordingly; the posterior
(B+t)
probability of ϑi lying in Θ0 ⊂ R is estimated as #{t : ϑi ∈ Θ0 }/N . This offers an intu-
itive way of conducting tests of hypotheses (Casella and Berger, 2002, Ch. 8), a technique
that we will make use of when discussing the significance of regression coefficients β. As

9
ϑ also includes the latent risk factors, the output of the MCMC algorithm can be used to
investigate the posterior path of (bt ). In particular, one can visually compare the path of
(bt ) with those of observed economic variables.
A further strength of the MCMC approach is that the above techniques can be applied
to model quantities derived from several primary parameters as well, such as migration
probabilities or correlations. Assuming that these can be written in the form f (ϑ) for some
f : Θ → R, a sample from their respective posterior distributions is instantly given by
f (ϑ(B+1) ), . . . , f (ϑ(B+N ) ). This means that point estimates as well as standard errors of
derived quantities are obtained at virtually no additional computational cost.

3.3 Bayesian Model Validation

Bayesian model validation involves two steps: robustness (i.e. quantification of the impact
of the prior distribution on the posterior point estimates) and assessment of model fit,
where the latter step is closely related to model selection. The first issue can be addressed
by considering several prior specifications and comparing the resulting posterior estimates.
By restricting ourselves to non-informative (or vague) priors, the differences in posterior
estimates are generally kept small.
For the second step, we consider a modern approach to Bayesian model comparison based
on cross-validation predictive densities and regression diagnostics derived from these. Due
to the multivariate nature of M tk as basic observation, a marginal likelihood is a suitable
diagnostic to work with for the purposes of Section 4. We consider the conditional predictive
ordinate (CPO), defined as

CPOt := p(M t1,obs , . . . , M tK,obs | {M s1,obs , . . . , M sK,obs : s = 1, . . . , T, s 6= t}),

where M tk,obs denotes the observed realization of M tk . The CPO is attractive in that it
suggests how likely the joint observation M t1,obs , . . . , M tK,obs is, when the model is fitted to
all observations except M t1,obs , . . . , M tK,obs . It can be implemented promptly by re-using
the output of the Gibbs sampler. By comparing, for instance, T −1 t CPOt and the plot
P

{(t, CPOt ) : t = 1, . . . , T } for competing models, we obtain an assessment of the model fits.
We refer to Gelfand (1996) and Carlin and Louis (2000) for an in-depth survey of the above
issues and more.

4 Empirical Analysis

In this section, we apply the ideas of Section 2 and suggest a series of migration models of
increasing complexity. These are fitted to data from S&P by Gibbs sampling.

10
4.1 Description of the Data

The dataset has been extracted from Standard & Poor’s CreditProTM 6.6 database and
consists of 5,651 US and Canadian firms from 12 S&P industry sectors (see Table 1). The
migration counts M tk have been collected for three-month periods, ranging from January
1981 to December 2000 (T = 80 quarters). Obligors whose rating is withdrawn have been
excluded from consideration in the time period of the withdrawal. Obvious duplicates in the
database (such as holding companies) have been removed as well. The rating classes under
consideration are
K = {CCC, B, BB, BBB, A, AA, AAA},

where for the case of simplicity qualifiers have been suppressed (K = 7). This means that
k ∈ K corresponds to one of the actual ratings k + , k, or k − . As is customary, we merge
CCC, CC, and C into a single rating class: CCC. All other notation follows Section 2.

4.2 Models

Throughout this section, we employ the logit response g(x) = 1/(1 + exp{−x}) in the frame-
work of Sections 2.2 and 2.3. Apart from being the canonical response function for many
GLMs (McCullagh and Nelder, 1989), the logit response is also straight-forward to evalu-
ate (the probit response g(x) = Φ(x) requires numerical evaluation). We do not consider
obligor-specific covariates, so firms are arranged into homogeneous groups as discussed in
Section 2.2.1. The empirical study follows the common Bayesian approach of using non-
informative priors; detailed information about prior distributions is given in Section 4.3.1.

Model Class (P) — Preliminary Analysis

Our preliminary analysis, also known as model class (P), explores the validity of the Chicago
Fed National Activity Index (CFNAI) (xt ) as a proxy for the business cycle. The CFNAI
is published on a monthly basis and can be downloaded from the internet. It was found
significant for describing default activity in McNeil and Wendin (2005).
We assume that all obligors in rating class k in period t are exposed to the same sys-
tematic risk γtk , k ∈ K. Given γt1 , . . . , γtK , the migration count vectors M t1 , . . . , M tK are
assumed to be conditionally independent with

M tk | γtk ∼ Multinomial {mtk , pk (γtk )} ,

as in (4). We give three different specifications of γtk . First we consider two simple GLMs
which are free from latent factors:
(P1) γtk = xt β;
(P2) γtk = xt βk ,
where xt is the CFNAI for the first calendar month of time period t. Parameters to be
estimated are the threshold values (µk,` ), and the regression coefficients β and β1 , . . . , βK ,

11
respectively. Note that (P2) fits into the GLM-framework with β = (β1 , . . . , βK )0 and
x̃tk = ek , where ek is the kth unit vector in RK .
We then investigate a GLMM which in addition to (xt ) features a sequence of serially
dependent latent factors (bt ), in order to capture any cross-sectional or serial dependence
among the migrations:

(P3) γtk = xt β + bt ,

where (bt ) is the univariate AR(1) sequence with variance σ 2 = φ2 /(1 − α2 ) introduced in
(7). Given xt , the implied asset correlation is zero under the first two models, and equals
σ 2 /(σ 2 + π 2 /3) under (P3). Model (P3) requires estimation of the hyperparameters φ and
α in addition to the unknown parameters of model (P1).

Results

The posterior mean and standard deviation of all parameters of model class (P) are contained
in Tables 2, 3 and 4, from which we draw the following conclusions. Four of the threshold
parameters apparently exhibit large standard errors, as the corresponding transitions never
take place in the dataset (see rows AAA and B). The large standard errors merely indicate
the profound uncertainty in assigning these events a probability. It is also worth noting
that the thresholds of rating class CCC exhibit higher standard errors than those of other
subinvestment-grade ratings. One reason for this is the small size of the CCC-cohort (rating
classes B and BB are approximately 10 times larger than CCC).
Secondly, even though model (P1) suggests high explanatory power of the CFNAI, the
results of model (P2) are slightly contradictory: the coefficients βk remain significant (al-
though the BB and CCC rating categories are borderline cases), but βAAA carries a different
sign than expected. We interpret this as empirical evidence for differences in exposure to
systematic risk across rating classes.
Model (P3) is summarized in Table 4. The point estimates of φ and α clearly suggest the
presence of latent systematic risk in the migrations; the variance of bt , φ2 /(1 − α2 ), suggests
an implied asset correlation of 3.5 %, and the time series parameter α points to profound
serial dependence. Figure 1 presents the posterior mean of bt in all time periods. For sake of
reference, we include the CFNAI in the graph for visual comparison. There is some degree
of co-movement between the two time series (xt β) and (bt ), but it is clear that the observed
covariate does not capture the full variability in migration rates. Finally, the presence of
(bt ) has reduced the explanatory power of CFNAI to the extent that it is no longer formally
significant. We therefore leave out (xt ) in the remainder of the empirical study.
The parameters of model (P3) can be used to calculate various model summaries: Ta-
ble 11 holds the migration matrix, and Table 12 the corresponding up- and downgrade
correlations, as defined in Section 2.2. Although migration correlations are small numbers
(joint downgrade correlations usually higher than upgrade correlations), the values presented

12
here are slightly higher than those of Gagliardini and Gouriéroux (2005b), based on yearly
migration counts on French corporate data. (This holds also when we consider only the
one-step up-up and down-down correlations.)

Model Class (K) — Rating-specific Factor Loadings

The Basel II capital adequacy framework allows risk factor loadings to depend on the credit
quality of the obligor. The issue of rating-specific factor loadings has therefore been subject
to recent interest; see BCBS (2002) and Lopez (2004). This matter can be gone into within
the framework of homogeneous buckets defined by rating category: that is, H = K.
The findings of model (P2) suggest that the rating classes may be subject to different
systematic risk. Secondly, the results of model (P3) motivate us to leave out the CFNAI
when latent risk factors are included, hence all models from now onwards have xti = 0 for
all obligors. Let bt = (bt1 , . . . , btK )0 , where btk denotes the latent risk of rating class k.
The joint distribution of M t1 , . . . , M tK is fully determined by the threshold values (µk,` )
along with the distribution of bt . In this section, we propose three multivariate Gaussian
specifications of the latter, ranging from perfectly dependent components in (K1) to an
arbitrary covariance structure in (K3).
As in model class (P), we assume that M t1 , . . . , M tK are conditionally independent,
given γt1 , . . . , γtK , with

M tk | γtk ∼ Multinomial {mtk , pk (γtk )} .

Model (K1)

Let (bt ) be the univariate AR(1) sequence (7) with φ = 1, and define

(K1) γtk = btk = φk bt .

Note that we impose var(bt ) = 1/(1 − α2 ) for identifiability. The components of bt =

bt (φ1 , . . . , φK )0 are perfectly dependent (model (K1) is thus a single-factor model). The
experience with model (P2) shows that there is no reason to exclude negative φk ’s. The
implied asset correlation for two obligors in the kth rating class is σk2 /(σk2 + π 2 /3), where
σk2 = φ2k /(1 − α2 ). Model (K1) is a standard specification of rating-specific factor loadings
and has been used in Gordy and Heitfield (2002) and Rösch (2005).

Model (K2)

Let (bt ) be the univariate AR(1) sequence in (7) with variance σ 2 = φ2 /(1 − α2 ), and define

(K2) γtk = btk = bt + ξtk ,

where (ξtk ) are iid, random level shifts that are independent from (bt ), and follow the

13
N (0, ω 2 )-distribution. The vector bt is multivariate Gaussian with mean zero and

σ 2 + ω 2 if k = l,

cov(btk , btl ) = (9)
σ 2

else,

hence, its components are exchangeable (the covariance matrix exhibits compound symme-
try). Although model (K2) does not display rating-specific factor loadings, it allows us to
obtain a feeling for the covariance structure of a multivariate Gaussian distribution on bt .
In particular, the magnitude of ω 2 reveals the degree of variability in systematic risk across
different rating classes. For two obligors of rating classes k and l, model (K2) suggests an
implied asset correlation of (σ 2 +ω 2 )/(σ 2 +ω 2 +π 2 /3) if k = l, and σ 2 /(σ 2 +π 2 /3) otherwise.

Model (K3)

Assume that (bt ) follows the K-dimensional AR(1) sequence (8) and set

(K3) γtk = btk .

The covariance matrix of bt is the K × K-matrix (Σkl ) = Σ = Φ/(1 − α2 ). The matrix Φ

has the properties of a covariance matrix, and is calculated as Φ = AA0 ; the roles of A and
α follow from (8). Σ has K(K + 1)/2 free parameters, compared to only two in model (K2).
The implied asset correlation of two firms with ratings k and l equals Σkl /(Σkl + π 2/3).

Results

Point estimates of the parameters of models (K1), (K2) and (K3) are given in Tables 5, 6
and 7, respectively. The threshold parameters are similar to those of model class (P) and
will not be commented.
The variance of the latent factor in model (P3) suggests an average implied asset cor-
relation of 3.5 %, whereas those of (K1) range between a fraction of a percent (class AAA)
and 8.0 % (class B). For model (K2), obligors belonging to the same rating class have an
implied asset correlation of 6.5 %, whereas that of obligors in different rating classes is only
2.3 %. This corresponds to a point estimate of corr(btk , btl ) of 34 %. The results of model
(K3) are similar, although there is large uncertainty about many of the elements of Σ.

Model Class (S) — Heterogeneity across Industry Sectors

This section addresses the issue of heterogeneity among industry sectors by means of multi-
variate latent risk factors. Let S = {1, . . . , S} be an index set of industry sectors for which
migration counts M tsk are collected for each rating category k ∈ K, sector s ∈ S, and time
period t. We define
M tsk := (Mts:k,0 , . . . , Mts:k,K ) ,

so that Mts:k,` equals the number of firms in sector s making a transition from state k ∈ K
to ` ∈ K0 during period t. Table 1 introduces the S = 10 sectors of the study. Notice

14
that componentwise summation of M tsk over s yields M tk as previously defined. Obligors
sharing industry sector and rating class in a time period are assumed to define a homogeneous
bucket; we denote by γtsk the systematic risk associated with M tsk . The models will be
referred to as model class (S), and can be accommodated in the framework of Section 2.2.1
with H = K × S.
It seems plausible to find non-perfect dependence between the systematic risk of differ-
ent industry sectors. We therefore concentrate on Gaussian specifications of bt : both the
compound symmetry model and the fully general covariance structure are treated. We also
include the case of rating-specific factor loadings after industry effects have been accounted
for; see model (S3).
Given the systematic risks {γtsk : s ∈ S, k ∈ K} we assume that the migration count
vectors {M tsk : s ∈ S, k ∈ K} are conditionally independent with

M tsk | γtsk ∼ Multinomial {mtsk , pk (γtsk )} .

Model (S1)

Let (bt ) be the Gaussian AR(1) sequence (7), and assume that (ξts ) are iid N (0, ω 2 ) and
independent from (bt ). We define

(S1) γtsk = bts = bt + ξts .

Under (S1), we have bt Gaussian with mean zero and covariance matrix as in (9). The
implied asset correlation for two firms sharing sector equals (σ 2 + ω 2 )/(σ 2 + ω 2 + π 2 /3),
whereas that of two firms in different sectors is merely σ 2 /(σ 2 + π 2 /3).

Model (S2)

A general covariance structure is obtained by letting (bt ) follow the S-dimensional AR(1)-
process (8) with covariance matrix Σ = (Σkl )k,l∈S :

(S2) γtsk = bts .

Σ is calculated as Φ/(1−α2 ), where α is the AR(1)-parameter, cf. model (K3). Consequently,

the implied asset correlation between two firms in sectors k and l equals Σkl /(Σkl + π 2/3).
The covariance matrix of bt consists of S(S + 1)/2 parameters.

Model (S3)

Our final specification of γtsk merges models (K1) and (S1) by combining rating-specific
factor loadings with sector effects. We set

(S3) γtsk = φk bts = φk (bt + ξts ),

where (bt ) is the AR(1)-sequence (7) with var(bt ) constrained to 1/(1 − α2 ); and (ξts ) are

15
iid N (0, ω 2 ), and independent of (bt ). This specification helps to avoid confounding het-
erogeneity across rating classes with events that can be traced back to an industry sector.
The implied asset correlations under model (S3) will depend on both rating and sector
membership.

Results

Point estimates of the parameters of models (S1), (S2) and (S3) are given in Tables 8, 9
and 10, respectively. The point estimate of ω in model (S1) is 51 % larger than σ, and thus
reveals material evidence of sector-specific variability (the correlation between bts of two
different sectors is roughly 30 %). Consequently, the overall implied asset correlation 3.5 %
of model (P3) is now 2.6 or 8.7 %, depending on whether the two obligors belong to different
sectors or not. Table 9, which allows for an arbitrary covariance matrix Σ for bt , yields
a similar picture, albeit with large standard errors on many of the covariance parameters.
Note that the correlation between certain sectors is high, whereas other appear to be rather
uncorrelated.
Model (S3), which combines rating-specific factor loadings with industry sector effects,
is summarized in Table 10. The factor loading of every rating class is significant; moreover,
the magnitude of the φk ’s shows less variability than in model (K1). A point estimate of
the correlation between the bts ’s of two different sectors is 37 %.

4.3 Implementation Details

The models of the previous section are fitted by Gibbs sampling with self-customized code in
C. A derivation of the full conditional distributions is sketched in Appendix A. The running
time of a 10,000-iteration simulation of the algorithms ranges from a few minutes up to an
hour, depending on the complexity of the model.

4.3.1 Choice of Priors

This study follows the common practice of using non-informative priors for the threshold
values, regression coefficients, and hyperparameters. The intercept vectors µ1 , . . . , µK are
assigned iid zero-mean, ordered Gaussian prior distributions with variance τ 2 IK×K , where
τ is large (τ = 100) and IK×K is the identity matrix. Likewise, β and β1 , . . . , βK of models
(P1) and (P2) are given iid N (0, τ 2 )-priors. The autoregressive parameter α is assumed to
be uniform on (−1, 1) a priori.
The variance of the innovations φ2 is assigned an improper prior decaying as 1/x; this
is a limiting case of the inverse-gamma distribution (1/φ2 is assumed to follow the Γ(η, ν)-
distribution with (η, ν) = (0, 0)). This is a standard vague prior for a scaling parameter; see
McNeil and Wendin (2005). We employ the same prior for ω 2 of models (K2) and (S1) as
well. In models (K3) and (S2), Φ is assigned an inverse-Wishart distribution with parameters

16
ν and Λ0 ; see Appendix A. This is a standard prior for a covariance matrix. For small values
of ν the inverse-Wishart distribution is vague: we use ν = 0.001 with Λ0 set to the identity
matrix. References on informative priors for a covariance matrix are given in Boscardin and
Weiss (2004). The scaling parameters φ1 , . . . , φK in models (K1) and (S3) occur in the form
of regression coefficients rather than hyperparameters, which means that the inverse-gamma
prior no longer leads to an inverse-gamma full conditional. As there is no prior reason for
all these parameters to share sign, they are assigned independent N (0, τ 2 )-priors.

4.3.2 Model Comparisons

We apply the cross-validation density-approach described in Section 3.3 to compare the

models of Section 4.2. Table 13 contains the relevant summaries of the CPOs: we consider
the average log(CPO)-value (to the left) and the number of time periods, out of T = 80, in
which the more advanced model(s) perform(s) better (to the right).
It is evident from Table 13 that the nine models under study can be divided into four
groups of comparable model performance, which we now list in order of increasing model
fit. The GLMs (P1) and (P2) undoubtedly exhibit the poorest fit. The two single-factor
models (P3) and (K1) come second, and are followed by models (K2) and (K3) with Gaussian
distributions for (bt1 , . . . , btK )0 . Finally, model class (S), featuring sector-specific risk factors,
displays the superior model fit. The differences in model fit within these four groups are
barely worth mentioning.

4.4 Discussion

One of the main conclusions of model classes (P) and (K) is that the exposure to systematic
risk may vary across rating classes. This is particularly evident for rating class AAA, whose
coefficients βAAA and φAAA in models (P2) and (K1) either have a different sign than
expected, or equal zero. Moreover, the systematic risk of class AAA in model (K3) shows
very little correlation with the other rating classes. It is of course important to bear in mind
that assessment of the systematic risk is difficult for rating classes where obligors are scarce.
This applies especially to rating class CCC, whose parameter uncertainty is constantly large,
but also to class AAA to some extent.
This being said, the results of model class (K) still point towards heterogeneity among
rating classes: in model (K2), the implied asset correlations instantly fall by two-thirds if
the obligors belong to different rating classes. Clearly, there is no economic justification for
massively reduced implied asset correlation just because two obligors belong to neighbouring
rating classes. It therefore lies close at hand to suspect that this outcome is a substitute
for effects that have not been accounted for; in particular, violations of the assumption of
homogeneity within each rating class. Before proceeding, we note that models (K2) and
(K3) exhibit considerably larger estimates of var(btk ) than models (P3) and (K1). Imposing

17
perfect dependence on (bt1 , . . . , btK )0 , as in (K1), seems to force the variances to be modest.
In the next set of models, homogeneous groups are composed with the two attributes
rating class and industry sector. The correlation between the sector-specific risk factors
is 30 % in the exchangeable model (S1), whereas the correlation between industry sectors
varies considerably in the general model (S2), despite the large parameter uncertainty. As
industry sectors may be subject to different business conditions, a non-perfect dependence
structure on (bt1 , . . . , btS )0 makes good economic sense. Recall also that models featuring
sector effects are clearly favoured in the model comparison, cf. Table 13.
As to rating-specific factor loadings, we find that joint consideration of sector effects and
rating-specific loadings renders less variable factor loadings, cf. model (S3). This seems to
suggest that the need for rating-specific loadings decreases in the presence of sector effects.
As this specification is low in parameters and displays a good model fit, it will be our
preferred model.
Finally, our findings do not necessarily support the view that factor loadings drop with
increasing probability of default, at least not for the investment-grade obligors (ratings BBB
and above). We believe that properly addressing the issue of heterogeneity among industry
sectors is more imperative. It should, however, be pointed out that inference about the
investment grade is based mainly on migration events other than default. Thus, inferring
factor loadings from default data only might lead to other conclusions (although estimates
for the investment grade are usually highly uncertain due to the rare occurrence of defaults).

5 Conclusions

This paper presents a statistical framework for dependent rating migrations driven by latent,
serially correlated systematic risk factors. Unobserved risk factors have several advantages
over observed ones; see the discussion in the Introduction. The paper also highlights the use
of computational Bayesian inference (MCMC) for migration risk models that feature latent
effects. MCMC techniques are capable of handling a variety of specifications of the system-
atic risk, many of which are highly relevant for practice. In particular, the methodology
can deal with multivariate Gaussian risk factors in order to capture heterogeneity among
industry sectors.
The general conclusions of the empirical study, which is performed on three-month mi-
gration data from S&P, parallel those of McNeil and Wendin (2005): we find evidence of
substantial cross-sectional and serial dependence in transition activity, and reestablish the
fact of heterogeneity between industry sectors. For practical purposes, the effect of the la-
tent systematic risk is most easily interpreted in terms of implied asset correlations, which
are defined in Section 2.2. By model (S1), the implied asset correlation of obligors in dif-
ferent sectors is less than one-third of the implied asset correlation of firms in the same
industry sector. Model (S2), featuring a general covariance structure, yields a similar pic-

18
ture, albeit with large uncertainty about many of the non-diagonal elements of Σ. Models
featuring sector-specific effects also score the best in the model comparison. Nevertheless, a
general covariance structure on bt , as in model (S2), appears to be slightly over-ambitious,
considering the limited amount of historical data available.
We also find that single-factor models often suggest smaller implied asset correlations
than multivariate specifications of the latent risk. The implied asset correlations of firms
sharing industry sector are not far from the values prescribed by regulators BCBS (2002),
whereas the corresponding values for firms in different sectors in general are substantially
lower. Some of the implied asset correlations of our analysis (especially for the investment-
grade ratings) might be received as worryingly low if used as inputs in a risk management
context, in particular when treating only defaults. As defaults of investment-grade obligors
are very rare, the implied asset correlations of these rating classes will mainly be influenced
by migration events other than default, but they nevertheless reflect the correlation structure
of the implied asset values (Vt1 , . . . , Vtmt )0 , as defined in Section 2.2.
At this stage, it is worth pointing out that the systematic risks governing defaults and
transitions do not necessarily have to coincide—in fact, there is empirical evidence of up-
grades being less correlated with the general credit cycle (Koopman, Lucas, and Monteiro,
2005). It should also be stressed that correlations alone do generally not determine the joint
distribution of the threshold variables Vt1 , . . . , Vtmt . This is particularly material for events
far out in the tails of the threshold variables, such as defaults of investment-grade obligors,
events for which correlations alone constitute a very blunt specification of the dependence.
Frey, McNeil, and Nyfeler (2001) illustrate the substantial model risk in this context. The
previous two remarks illustrate that it is crucial to know the origins of the implied asset
correlations, as the source determines their scope and validity for applications.

Acknowledgements

We thank Standard & Poor’s for providing the dataset used in Section 4. The first au-
thor gratefully acknowledges financial support from NCCR Financial Valuation and Risk
Management (a research program supported by the Swiss National Science Foundation).

A Deriving the Full Conditional Distributions

This section employs a notation that is common in literature on Gibbs sampling: [ X ]

denotes the (unconditional) density (or mass function) of the random quantity X, [X | Y ]
the conditional counterpart given Y , and [X | ·] is the full conditional distribution of X. X
denotes the full sequence (X1 , . . . , XT ). The key to all full conditional distributions is the

19
joint distribution function of data and parameters, which reads

[M , b, µ, x, β, α, φ, ω] ∝ [M | µ, x, β, b][b | α, φ, ω][µ, β, α, φ, ω]
T Y
!
Y
= [M ti | µi , xt , β, bt ] [b | α, φ, ω][µ][β ][α][φ][ω]. (10)
t=1 i∈K

The last line follows by conditional independence arguments and a priori independence
of the parameters. For more information about the Gibbs sampler and full conditional
distributions, see Gilks (1996).
We exemplify the derivation of a full conditional distribution in the context of model
(P3). Before proceeding, we observe that b := (b1 , . . . , bT ) defined as in (7) is multivariate
Gaussian with covariance matrix Σb :

cov(bs , bt ) = φ2 α|s−t| /(1 − α2 ), s, t ∈ {1, . . . , T }.

Its inverse Σ−1 −2 2 2

b is tridiagonal with diagonal elements φ (1, 1+α , . . . , 1+α , 1), off-diagonal
elements −φ−2 α and determinant φ−2T (1 − α2 ).
The full conditional of α is calculated by applying the formula for conditional probabilities
and using (10). In each step of the derivation we drop factors without explicit dependence
on α, which explains the use of the ∝-sign:

[M , b, µ, x, β, φ, α]
[α | ·] = ∝ [M , b, µ, x, β, φ, α] ∝ [b | φ, α][α].
[M , b, µ, x, β, φ]

The final step involves properties of [b | φ, α]

q
[α|·] ∝ [b | φ, α][α] ∝ det(Σ−1
 1 −1 0
b ) exp − 2 bΣb b [α]
p
1 − α2 exp − 12 φ−2 (C1 (b)α2 − C2 (b)α) [α],

∝ (11)

where C1 ( · ) and C2 ( · ) are identified by performing the vector-matrix multiplications. If

α is assigned a uniform prior, (11) is log-concave in α and can be simulated from with
the ARS algorithm. The other full conditionals are derived analogously, see McNeil and
Wendin (2005) for a similar application. Except for φ2 , whose full conditional remains
inverse-gamma, all other full conditionals of our study rely on the ARMS algorithm for their
simulation (Gilks, 1992).
In case of multivariate bt = (bt1 , . . . , btH )0 , note that b := (b01 , . . . , b0T )0 is multivariate
Gaussian with covariance matrix

α|t−s| s, t ∈ {1, . . . , T },
cov(bsi , btj ) = Φij ,
1 − α2 i, j ∈ {1, . . . , H}.

As Φ has the properties of a covariance matrix, we assign it a inverse-Wishart prior distri-

bution. This means that its inverse Φ−1 := Λ follows the Wishart distribution

[Φ−1 ] = [Λ] ∼ Wishart (ν, Λ0 ) ,

20
where ν and Λ0 are fixed real numbers and symmetric, non-singular H × H-matrices, re-
spectively. Its prior density is thus proportional to

|Λ|(ν−H−1)/2 exp − 12 tr(Λ0 Λ) ,



see Johnson and Kotz (1972) or Muirhead (1982). The inverse-Wishart distribution is a
conjugate prior of a covariance matrix:

[Λ | · ] ∼ Wishart (ν + T, Λ0 + C3 (b)) ,

where C3 (b) is the following H × H-matrix:

T
X T
X −1 T
X
bt b0t 2
bt b0t bt b0t−1 + bt−1 b0t ,


C3 (b) = +α −α
t=1 t=2 t=2

see Appendix A.2 of Bernardo and Smith (1994) or McCulloch and Rossi (1994). An algo-
rithm for generating variates from the Wishart distribution is given in Johnson (1987). In
each iteration of the Gibbs sampler the current value of Φ is obtained by inverting Λ.
Due to the form of [ M | µ, x, β, b ], the full conditional distribution of the latent risk
factor bt does not follow a standard statistical distribution. We therefore update each bth
separately by means of the ARMS algorithm, an exercise which is greatly simplified by
first deriving the conditional distribution of bth given all other elements of b. Let b−t :=
(b01 , . . . , b0t−1 , b0t+1 , . . . , b0T )0 and observe that under the multivariate AR(1) process in (8) we
have [bt | bt −t , α, Φ] multivariate Gaussian:




N αb2 , Φ if t = 1,


[bt | b−t , α, Φ] ∼ N αbT −1 , Φ (12)


 if t = T,


α 1


N 1+α 2 (bt−1 + bt+1 ), 1+α2 Φ else.


Secondly, it is easily verified that if Z = (Z1 , . . . , ZH )0 is Gaussian with mean (µ1 , . . . , µH )0 ,

then for each h = 1, . . . , H
 
H
1 X 1 
Zh | {Zi = zi }i=1,...,H; i6=h ∼ N µh + λih (µi − zi ), , (13)

λhh i=1 λhh
i6=h

where (λij ) denotes the inverse of the covariance matrix of Z (the so-called precision matrix).
Combining (12) and (13) the above observations yields the conditional distribution of bth
given the remainder of b and the hyperparameters. All other full conditionals are simulated
as in the case of univariate latent factors.

References

Bangia, A., F. X. Diebold, A. Kronimus, C. Schagen, and T. Schuermann (2002). Rat-

ings migration and the business cycle, with application to credit portfolio stress testing.
J. Banking Finance 26, 445–474.

21
Basel Committee on Banking Supervision (BCBS) (2002, October). Quantitative impact,
Study 3, Technical guidance. Bank of International Settlements.

Basel Committee on Banking Supervision (BCBS) (2005, May). Working paper No. 14,
Studies on validation of internal rating systems. Bank of International Settlements.

Bernardo, J. M. and A. F. M. Smith (1994). Bayesian Theory. John Wiley & Sons, Chich-
ester.

Boscardin, W. J. and R. E. Weiss (2004). Fitting unstructured covariance matrices to

longitudinal data. Working paper, UCLA School of Public Health.

Carlin, B. P. and T. A. Louis (2000). Bayes and Empirical Bayes Methods for Data Analysis
(2nd ed.). Chapman & Hall/CRC.

Casella, G. and R. L. Berger (2002). Statistical Inference (2nd ed.). Duxbury, Pacific Grove.

Christensen, J. H. E., E. Hansen, and D. Lando (2004). Confidence sets for continuous-time
rating transition probabilities. J. Banking Finance 28, 2575–2602.

Durbin, J. and S. J. Koopman (2001). Time Series Analysis by State Space Methods. Oxford
University Press, Oxford.

Fahrmeir, L. and G. Tutz (1994). Multivariate Statistical Modelling Based On Generalized

Linear Models. Springer-Verlag, New York.

Frey, R., A. J. McNeil, and M. Nyfeler (2001). Copulas and credit models. Risk 14 (10),
111–114.

Frydman, H. and T. Schuermann (2005, June). Credit rating dynamics and Markov mixture
models. Preprint, Wharton Financial Institutions Center.

Gagliardini, P. and C. Gouriéroux (2005a). Migration correlation: Definition and efficient

estimation. J. Banking Finance 29, 865–894.

Gagliardini, P. and C. Gouriéroux (2005b). Stochastic migration models with application to

corporate risk. J. Financial Econometrics 3 (2), 188–226.

Gelfand, A. E. (1996). Model determination using sampling-based methods. In W. R. Gilks,

S. Richardson, and D. J. Spiegelhalter (Eds.), Markov Chain Monte Carlo in Practice,
pp. 145–161. Chapman & Hall, London.

Gilks, W. R. (1992). Derivative-free adaptive rejection sampling for Gibbs sampling. In J. M.

Bernardo, J. O. Berger, A. P. Dawid, and A. F. M. Smith (Eds.), Bayesian Statistics 4,
pp. 641–649. Oxford University Press, Oxford.

22
Gilks, W. R. (1996). Full conditional distributions. In W. R. Gilks, S. Richardson, and
D. J. Spiegelhalter (Eds.), Markov Chain Monte Carlo in Practice, pp. 75–88. Chapman
& Hall, London.

Gordy, M. B. and E. Heitfield (2002). Estimating default correlations from short panels of
credit rating performance data. Technical report, Federal Reserve Board.

Gouriéroux, C. and A. Monfort (1996). Simulation-Based Econometric Methods. CORE

Lecture series. Oxford University Press, New York.

Hu, Y.-T., R. Kiesel, and W. Perraudin (2002). The estimation of transition matrices for
sovereign credit ratings. J. Banking Finance 26, 1383–1406.

Jafry, Y. and T. Schuermann (2004). Measurement, estimation and comparison of credit

migration matrices. J. Banking Finance 28, 2603–2639.

Johnson, M. E. (1987). Multivariate Statistical Simulation. Wiley.

Johnson, N. L. and S. Kotz (1972). Distributions in Statistics: Continuous Multivariate

Distributions. John Wiley & Sons, New York.

Kijima, M. K., K. Komoribayashi, and E. Suzuki (2002). A multivariate Markov model for
simulating correlated defaults. J. Risk 4 (4), 1–32.

Koopman, S. J., A. Lucas, and R. J. Daniels (2005). A non-Gaussian panel time series model
for estimating and decomposing default risk. Working paper, Tinbergen Institute.

Koopman, S. J., A. Lucas, and P. Klaassen (2005). Empirical credit cycles and capital buffer
formation. J. Banking Finance 29, 3159–3179.

Koopman, S. J., A. Lucas, and A. Monteiro (2005). The multi-state latent factor intensity
model for credit rating transitions. Working paper, Tinbergen Institute.

Künsch, H. R. (2001). State space and hidden Markov models. In O. E. Barndorff-Nielsen,

D. R. Cox, and C. Klüppelberg (Eds.), Complex Stochastic Systems, pp. 109–173. Chap-
man & Hall/CRC, Boca Raton.

Kurbat, M. and I. Korablev (2002). Methodology for testing the level of the EDFTM credit
measure. Technical Report #020729, Moody’s KMV.

Lando, D. and T. Skødeberg (2002). Analyzing rating transitions and rating drift with
continuous observations. J. Banking Finance 26, 423–444.

Lopez, J. A. (2004). The empirical relationship between average asset correlation, firm
probability of default, and asset size. J. Finan. Intermediation 13, 265–283.

23
MacDonald, I. L. and W. Zucchini (1997). Hidden Markov Models and Other Models for
Discrete-valued Time Series. Chapman & Hall, London.

McCullagh, P. and J. A. Nelder (1989). Generalized Linear Models (2nd ed.). Chapman &
Hall, London.

McCulloch, C. E. and P. E. Rossi (1994). An exact likelihood analysis of the multinomial

probit model. J. Econometrics 64, 207–240.

McNeil, A. J. and J. Wendin (2005). Bayesian inference for generalized linear mixed models
of portfolio credit risk. Preprint, ETH Zürich.

Merton, R. C. (1974). On the pricing of corporate debt: The risk structure of interest rates.
J. Finance 29, 449–470.

Muirhead, R. J. (1982). Aspects of Multivariate Statistical Theory. John Wiley & Sons, New
York.

Nickell, P., W. Perraudin, and S. Varotto (2000). Stability of rating transitions. J. Banking
Finance 24, 203–227.

Robert, C. P. and G. Casella (1999). Monte Carlo Statistical Methods. Springer-Verlag, New
York.

Rösch, D. (2005). An empirical comparison of default risk forecasts from alternative credit
rating philosophies. Int. J. Forecasting 21, 37–51.

Skrondal, A. and S. Rabe-Hesketh (2004). Generalized Latent Variable Modeling. Chapman

& Hall/CRC, Boca Raton.

24
 Sector Name # Obligors CAN

1 ”Aerospace, automotive, capital goods, metal” 765 5.9

2 ”Consumer, service sector” 920 3.2
3 ”Leisure time, media” 571 5.3
4 ”Utility” 486 5.4
5 ”Health care, chemicals” 406 2.8
6 ”High tech, computers, office equipment” + ”Telecom” 531 6.1
7 ”Financial institutions” 717 4.2
8 ”Insurance” 444 4.0
9 ”Energy and natural resources” 359 9.1
10 ”Forest and building products, homebuilders” + ”Real estate” 452 10.9

Table 1: The table displays the industry sectors used for the analysis of Section 4. Sectors 6 and 10
are mergers of two regular S&P sectors. The rightmost column shows the proportion (%) of Canadian
firms in the sector.

25
 0.5
0.0
−0.5
−1.0

01/01/1981 01/01/1987 01/01/1993 01/01/1999

0.5
0.0
−0.5
−1.0

01/01/1981 01/01/1987 01/01/1993 01/01/1999

Figure 1: The upper plot shows the posterior mean of the latent process (bt ) under model (P3)
including its 95 % confidence bounds. Observe that the width of the confidence interval decreases
with time, as the number of obligors in the sample grows. The lower plot shows the posterior mean
(full line) together with the evolution of the CFNAI {(t, xt β) : t = 1, . . . , T } for reference. The point
estimate of β in the lower plot is taken from model (P1).

26
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.67 (4.93) −22.48 (5.73) −16.16 (4.95) −8.65 (.82) −7.62 (.54) −6.16 (.27) −3.81 (.08) β: 0.116 (.020)
AA −11.34 (1.38) −9.81 (.80) −8.11 (.40) −7.67 (.32) −6.77 (.20) −3.97 (.05) 6.37 (.17)
A −10.11 (.70) −9.77 (.62) −7.60 (.22) −6.47 (.13) −4.19 (.04) 5.25 (.07) 8.37 (.33)
BBB −8.15 (.33) −7.73 (.26) −6.18 (.13) −4.22 (.05) 4.27 (.05) 7.25 (.22) 9.93 (.80)

27
BB −6.58 (.18) −5.86 (.13) −3.66 (.04) 4.04 (.05) 6.54 (.18) 8.35 (.43) 9.79 (.80)
B −4.65 (.06) −3.68 (.04) 4.20 (.05) 6.29 (.14) 6.95 (.20) 8.67 (.47) 20.37 (5.57)
CCC −2.01 (.07) 3.25 (.12) 4.92 (.26) 5.72 (.38) 6.95 (.65) 8.84 (1.34) 9.84 (1.67)

Table 2: Posterior mean and standard deviation (in brackets) of model (P1) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
The posterior probability of {β < 0} is 0.000.
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA βk p(βk < 0 | D)

AAA −28.62 (4.93) −22.47 (5.65) −16.32 (4.95) −8.69 (.83) −7.65 (.54) −6.19 (.26) −3.85 (.09) −0.315 (.107) 0.999
AA −11.30 (1.35) −9.81 (.79) −8.11 (.37) −7.68 (.30) −6.79 (.20) −3.98 (.05) 6.39 (.16) 0.216 (.058) 0.000
A −10.10 (.73) −9.77 (.64) −7.60 (.22) −6.47 (.13) −4.19 (.04) 5.25 (.07) 8.37 (.33) 0.164 (.043) 0.000

28
BBB −8.14 (.33) −7.72 (.27) −6.18 (.13) −4.22 (.05) 4.27 (.05) 7.26 (.22) 9.94 (.80) 0.113 (.045) 0.008
BB −6.58 (.18) −5.86 (.12) −3.66 (.04) 4.04 (.05) 6.54 (.18) 8.34 (.43) 9.78 (.81) 0.067 (.045) 0.069
B −4.65 (.07) −3.68 (.04) 4.20 (.05) 6.30 (.15) 6.96 (.21) 8.69 (.47) 20.37 (5.53) 0.136 (.044) 0.001
CCC −2.01 (.07) 3.25 (.11) 4.90 (.26) 5.70 (.37) 6.93 (.64) 8.78 (1.32) 9.79 (1.66) 0.124 (.077) 0.054

Table 3: Posterior mean and standard deviation (in brackets) of model (P2) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.94 (4.78) −22.61 (5.70) −16.32 (5.04) −8.61 (.84) −7.59 (.55) −6.14 (.28) −3.82 (.13) β: 0.060 (.051)
AA −11.28 (1.38) −9.79 (.80) −8.08 (.38) −7.66 (.32) −6.78 (.22) −3.98 (.11) 6.46 (.19) φ: 0.256 (.030)
A −10.08 (.68) −9.75 (.60) −7.62 (.24) −6.48 (.15) −4.19 (.10) 5.33 (.12) 8.44 (.34) α: 0.672 (.113)
BBB −8.15 (.34) −7.73 (.28) −6.18 (.16) −4.22 (.11) 4.36 (.11) 7.34 (.24) 10.03 (.80)

29
BB −6.60 (.20) −5.88 (.15) −3.67 (.10) 4.13 (.11) 6.63 (.20) 8.42 (.42) 9.84 (.81)
B −4.69 (.12) −3.72 (.11) 4.25 (.11) 6.35 (.17) 7.02 (.22) 8.74 (.48) 20.53 (5.48)
CCC −2.04 (.12) 3.29 (.15) 4.93 (.27) 5.71 (.39) 6.91 (.65) 8.72 (1.33) 9.72 (1.68)

Table 4: Posterior mean and standard deviation (in brackets) of model (P3) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
The posterior probability of {β < 0} is 0.119.
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA φk p(φk < 0 | D)

AAA −28.69 (4.91) −22.46 (5.66) −16.18 (4.86) −8.67 (.83) −7.62 (.53) −6.16 (.27) −3.83 (.10) −0.020 (.110) 0.554
AA −11.25 (1.43) −9.75 (.82) −8.05 (.40) −7.62 (.34) −6.74 (.23) −3.94 (.11) 6.47 (.19) 0.235 (.050) 0.000
A −10.03 (.73) −9.70 (.64) −7.58 (.22) −6.44 (.14) −4.16 (.08) 5.32 (.10) 8.43 (.33) 0.191 (.038) 0.000
BBB −8.13 (.34) −7.71 (.28) −6.16 (.15) −4.19 (.10) 4.35 (.10) 7.34 (.24) 10.03 (.80) 0.222 (.041) 0.000

30
BB −6.58 (.23) −5.87 (.19) −3.67 (.15) 4.21 (.16) 6.71 (.23) 8.53 (.45) 9.99 (.82) 0.379 (.051) 0.000
B −4.68 (.16) −3.71 (.16) 4.33 (.16) 6.43 (.21) 7.09 (.25) 8.81 (.50) 20.20 (5.60) 0.387 (.051) 0.000
CCC −2.01 (.12) 3.31 (.16) 4.96 (.29) 5.74 (.42) 6.95 (.69) 8.76 (1.34) 9.73 (1.66) 0.250 (.066) 0.000

α: 0.689 (.115)

Table 5: Posterior mean and standard deviation (in brackets) of model (K1) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.82 (4.89) −22.48 (5.66) −16.26 (5.04) −8.72 (.81) −7.67 (.54) −6.20 (.28) −3.85 (.13) φ: 0.167 (.034)
AA −11.31 (1.38) −9.84 (.83) −8.15 (.39) −7.72 (.33) −6.82 (.22) −4.01 (.11) 6.54 (.19) ω: 0.387 (.026)
A −10.08 (.71) −9.74 (.63) −7.61 (.24) −6.48 (.15) −4.20 (.10) 5.40 (.12) 8.50 (.33) α: 0.801 (.090)

31
BBB −8.16 (.35) −7.74 (.29) −6.18 (.15) −4.22 (.10) 4.42 (.11) 7.40 (.23) 10.11 (.83)
BB −6.64 (.20) −5.93 (.16) −3.71 (.11) 4.20 (.11) 6.70 (.20) 8.47 (.43) 9.90 (.83)
B −4.76 (.12) −3.79 (.10) 4.31 (.11) 6.41 (.17) 7.08 (.22) 8.82 (.49) 20.55 (5.64)
CCC −2.12 (.12) 3.34 (.15) 5.01 (.27) 5.81 (.38) 7.03 (.64) 8.86 (1.30) 9.85 (1.61)

Table 6: Posterior mean and standard deviation (in brackets) of model (K2) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.88 (4.75) −22.73 (5.49) −16.65 (4.86) −9.08 (.85) −8.03 (.57) −6.55 (.35) −4.19 (.25) α: 0.470 (.076)
AA −11.42 (1.37) −9.91 (.83) −8.19 (.39) −7.77 (.33) −6.88 (.23) −4.08 (.12) 6.51 (.20)
A −10.19 (.71) −9.85 (.63) −7.67 (.24) −6.53 (.15) −4.25 (.09) 5.33 (.11) 8.44 (.34)
BBB −8.21 (.34) −7.79 (.28) −6.24 (.15) −4.28 (.09) 4.34 (.10) 7.33 (.24) 10.03 (.82)
BB −6.69 (.21) −5.98 (.17) −3.76 (.12) 4.17 (.12) 6.67 (.21) 8.46 (.45) 9.92 (.84)
B −4.86 (.12) −3.88 (.11) 4.24 (.11) 6.34 (.18) 7.01 (.23) 8.74 (.48) 20.49 (5.55)
CCC −2.32 (.16) 3.29 (.18) 4.95 (.28) 5.74 (.38) 6.95 (.63) 8.73 (1.30) 9.72 (1.65)

Covariance matrix Σ

32
Posterior mean Posterior standard deviation

AAA AA A BBB BB B CCC AAA AA A BBB BB B CCC

AAA 1.174 −0.065 0.008 0.025 −0.105 0.109 −0.011 (.467) (.203) (.187) (.193) (.202) (.195) (.221)
AA −0.065 0.322 0.072 0.061 0.091 0.065 0.018 (.203) (.099) (.060) (.060) (.066) (.064) (.076)
A 0.008 0.072 0.207 0.060 0.032 0.046 0.031 (.187) (.060) (.060) (.044) (.046) (.047) (.055)
BBB 0.025 0.061 0.060 0.210 0.083 0.071 0.020 (.193) (.060) (.044) (.062) (.048) (.048) (.058)
BB −0.105 0.091 0.032 0.083 0.321 0.031 −0.041 (.202) (.066) (.046) (.048) (.085) (.065) (.075)
B 0.109 0.065 0.046 0.071 0.031 0.293 0.134 (.195) (.064) (.047) (.048) (.065) (.080) (.071)
CCC −0.011 0.018 0.031 0.020 −0.041 0.134 0.466 (.221) (.076) (.055) (.058) (.075) (.071) (.152)

Table 7: Posterior mean and standard deviation (in brackets) of model (K3) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.95 (4.86) −22.77 (5.62) −16.50 (4.97) −8.63 (.83) −7.58 (.54) −6.13 (.28) −3.80 (.13) φ: 0.191 (.037)
AA −11.35 (1.43) −9.84 (.87) −8.09 (.41) −7.67 (.34) −6.78 (.24) −3.97 (.12) 6.67 (.20) ω: 0.471 (.024)
A −10.24 (.75) −9.89 (.67) −7.65 (.25) −6.50 (.17) −4.21 (.11) 5.53 (.12) 8.64 (.35) α: 0.792 (.096)

33
BBB −8.19 (.34) −7.77 (.28) −6.22 (.16) −4.26 (.12) 4.51 (.12) 7.49 (.25) 10.21 (.81)
BB −6.68 (.21) −5.96 (.16) −3.76 (.11) 4.21 (.12) 6.71 (.20) 8.53 (.44) 9.97 (.82)
B −4.78 (.12) −3.81 (.11) 4.31 (.12) 6.41 (.18) 7.07 (.22) 8.78 (.47) 20.53 (5.50)
CCC −2.10 (.13) 3.40 (.16) 5.07 (.27) 5.87 (.38) 7.07 (.65) 8.86 (1.32) 9.84 (1.65)

Table 8: Posterior mean and standard deviation (in brackets) of model (S1) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA

AAA −28.94 (4.72) −22.83 (5.60) −16.50 (5.02) −8.72 (.82) −7.68 (.53) −6.22 (.27) −3.87 (.11) α: 0.573 (.055)
AA −11.38 (1.37) −9.90 (.83) −8.18 (.39) −7.76 (.32) −6.87 (.22) −4.06 (.09) 6.64 (.18)
A −10.21 (.72) −9.88 (.63) −7.73 (.23) −6.60 (.15) −4.31 (.08) 5.48 (.10) 8.60 (.34)
BBB −8.32 (.35) −7.90 (.29) −6.35 (.15) −4.38 (.09) 4.44 (.09) 7.43 (.23) 10.09 (.81)
BB −6.83 (.19) −6.11 (.15) −3.91 (.09) 4.10 (.09) 6.60 (.19) 8.41 (.42) 9.86 (.83)
B −4.94 (.10) −3.96 (.08) 4.18 (.09) 6.29 (.17) 6.96 (.22) 8.67 (.47) 20.44 (5.58)
CCC −2.25 (.10) 3.29 (.14) 4.94 (.26) 5.73 (.37) 6.95 (.64) 8.79 (1.30) 9.65 (1.70)

Covariance matrix Σ

Posterior mean Posterior standard deviation

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

34
1 .363 .063 .095 .018 .096 .083 .050 .060 .247 .091 (.11) (.07) (.09) (.09) (.09) (.08) (.11) (.10) (.12) (.11)
2 .063 .269 .089 .001 .123 .057 .022 −.022 .065 .099 (.07) (.08) (.08) (.08) (.08) (.07) (.10) (.09) (.11) (.10)
3 .095 .089 .365 −.020 .062 .059 .090 .035 .072 .169 (.09) (.08) (.12) (.09) (.09) (.08) (.11) (.10) (.12) (.11)
4 .018 .001 −.020 .444 .029 .071 .024 .013 −.052 .056 (.09) (.08) (.09) (.14) (.09) (.09) (.11) (.10) (.12) (.11)
5 .096 .123 .062 .029 .406 .076 .100 .020 .101 .051 (.09) (.08) (.09) (.09) (.13) (.09) (.11) (.10) (.12) (.11)
6 .083 .057 .059 .071 .076 .339 .097 .048 .139 .092 (.08) (.07) (.08) (.09) (.09) (.11) (.11) (.10) (.11) (.11)
7 .050 .022 .090 .024 .100 .097 .582 .149 .111 .203 (.11) (.10) (.11) (.11) (.11) (.11) (.15) (.10) (.12) (.11)
8 .060 −.022 .035 .013 .020 .048 .149 .442 .076 .063 (.10) (.09) (.10) (.10) (.10) (.10) (.10) (.14) (.12) (.11)
9 .247 .065 .072 −.052 .101 .139 .111 .076 .611 .053 (.12) (.11) (.12) (.12) (.12) (.11) (.12) (.12) (.16) (.11)
10 .091 .099 .169 .056 .051 .092 .203 .063 .053 .536 (.11) (.10) (.11) (.11) (.11) (.11) (.11) (.11) (.11) (.15)

Table 9: Posterior mean and standard deviation (in brackets) of model (S2) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 Threshold parameters µk,` Remaining parameters

k µk,D µk,CCC µk,B µk,BB µk,BBB µk,A µk,AA φk p(φk < 0 | D)

AAA −28.86 (4.78) −22.46 (5.68) −16.08 (5.06) −8.67 (.78) −7.65 (.53) −6.18 (.29) −3.83 (.15) 0.202 (.068) 0.003
AA −11.30 (1.37) −9.80 (.81) −8.11 (.41) −7.68 (.34) −6.79 (.24) −3.99 (.13) 6.59 (.20) 0.215 (.038) 0.000
A −10.16 (.72) −9.82 (.63) −7.64 (.24) −6.51 (.16) −4.22 (.12) 5.45 (.14) 8.56 (.35) 0.214 (.032) 0.000
BBB −8.20 (.37) −7.78 (.31) −6.23 (.17) −4.26 (.12) 4.44 (.12) 7.42 (.24) 10.08 (.80) 0.210 (.033) 0.000

35
BB −6.83 (.25) −6.12 (.21) −3.90 (.18) 4.29 (.19) 6.80 (.26) 8.63 (.47) 10.09 (.84) 0.334 (.046) 0.000
B −4.84 (.16) −3.87 (.15) 4.31 (.16) 6.42 (.21) 7.08 (.26) 8.80 (.50) 20.62 (5.50) 0.277 (.040) 0.000
CCC −2.061 (.11) 3.28 (.14) 4.94 (.27) 5.73 (.39) 6.95 (.67) 8.82 (1.33) 9.82 (1.65) 0.148 (.041) 0.000

ω: 1.936 (.269)
α: 0.741 (.111)

Table 10: Posterior mean and standard deviation (in brackets) of model (S3) with logit response function. Gibbs sampling with 2,000 iterations (burn-in 5,000).
 AAA AA A BBB BB B CCC D

AAA 97.73 2.04 0.17 0.03 0.02 0.00 0.00 0.00

AA 0.17 97.89 1.82 0.07 0.02 0.03 0.00 0.00
A 0.02 0.49 97.91 1.42 0.11 0.05 0.00 0.00
BBB 0.00 0.06 1.27 97.13 1.31 0.17 0.02 0.03
BB 0.01 0.02 0.12 1.54 95.70 2.33 0.15 0.14
B 0.00 0.02 0.08 0.09 1.30 96.01 1.54 0.96
CCC 0.01 0.01 0.09 0.24 0.41 3.03 84.24 11.97
D 0 0 0 0 0 0 0 100.00

Table 11: Unconditional quarterly migration matrix (%) of model (P3) with xt = 0, cf. Table 4.

Upgrade correlations

AA A BBB BB B CCC

AA 0.02 0.04 0.06 0.07 0.06 0.10

A 0.06 0.10 0.12 0.11 0.17
BBB 0.17 0.19 0.18 0.27
BB 0.21 0.20 0.31
B 0.18 0.29
CCC 0.45

Downgrade correlations

AAA AA A BBB BB B CCC

AAA 0.28 0.26 0.23 0.23 0.30 0.29 0.59

AA 0.24 0.22 0.21 0.28 0.27 0.55
A 0.20 0.19 0.25 0.24 0.50
BBB 0.19 0.25 0.24 0.49
BB 0.32 0.31 0.64
B 0.31 0.62
CCC 1.27

Table 12: Upgrade and downgrade correlation matrices (%) of model (P3). The values are
based on the point estimates of Table 4 with xt = 0; cf. (6).

36
 1
PT (i)
i T t=1 log(CPOt ) i\j (P2) (P3) (K1) (K2) (K3) (S1) (S2) (S3)

(P1) −124.56 (P1) 34 60 60 76 74 79 78 78

(P2) −124.46 (P2) 60 59 76 74 78 75 76
(P3) −121.98 (P3) 41 72 62 78 75 78
(K1) −121.73 (K1) 73 64 74 73 75
(K2) −118.92 (K2) 20 57 54 55
(K3) −119.26 (K3) 60 57 59
(S1) −116.62 (S1) 40 36
(S2) −116.74 (S2) 39
(S3) −116.60

Table 13: The leftmost part of the table shows the mean CPO-value, see Section 3.3; the rightmost
(j) (i) (i)
part holds #{t ∈ {1, . . . , T } : CPOt > CPOt }, where CPOt is the CPO under model i.

37

View publication stats