NUS-RMI Credit Rating Initiative

Technical Report Version: 2011 update 1 (07-07-2011)

2011 NUS Risk Management Institute (RMI). All Rights Reserved.
The information contained in this technical report is for information purposes only and is believed to
be reliable, but NUS Risk Management Institute (RMI) does not guarantee its completeness or
accuracy. Opinions and estimates constitute our judgment and are subject to change without notice.
NUS Risk Management Institute (RMI)
Address: 21 Heng Mui Keng Terrace, I
3
Building, Level 4, Singapore 119613
Tel: (65) 6516 3380
Fax: (65) 6874 5430
Website: rmi.nus.edu.sg

About RMI
The NUS Risk Management Institute (RMI) was established in August 2006 as a research
institute of the National University of Singapore dedicated to the area of financial risk
management. The establishment of RMI was supported by the Monetary Authority of Singapore
(MAS) under its program on Risk Management and Financial Innovation. RMI seeks to
complement, support and develop Singapore's financial sector's knowledge and expertise in risk
management, and thereby helps to take on the challenges arising from globalization, structural
change and volatile financial markets. RMI has three main functions: research, education and
training.

For further information on RMI, please visit: rmi.nus.edu.sg

About the RMI Credit Rating Initiative
In 2009, RMI commenced the Credit Rating Initiative as a constructive response to the criticisms
aimed towards Credit Rating Agencies following the recent financial crisis. The primary
objective of the Credit Rating Initiative (CRI) is to advance the state of research and
development in the critical area of credit rating systems. The CRI takes a public good approach
to credit rating with the goal of keeping the credit rating system current, evolutionary and
organic, and functions like a selective Wikipedia. To achieve this goal the CRI comprises
various initiatives.

An operational probability of default system started producing daily probabilities of default for
exchange listed firms in July 2010. The implementation of this system was undertaken to
demonstrate the operational feasibility of the system and as a source of credit information for
credit professionals in financial institutions, corporate treasury departments and regulatory
agencies.

In addition to the operational PD system the CRI also publishes material targeted towards
finance professionals, policy makers and academics with an interest in credit markets. The
publications consist of an annual volume of the Global Credit Review as well as the more
frequent Quarterly Credit Reports (forthcoming).

For further information on the CRI, please visit: rmi.nus.edu.sg/cri

Call for Credit Research Proposals
We are pleased to invite applications for grants to support the research and development of
credit rating methodologies. The intent of these grants is to develop, refine and study credit
rating methodologies for exchange listed firms, in order to bring in new ideas and spur research
in this vital area of the financial world. Applications are open to researchers from around the
world.

All grants awarded will include access to RMI's database of nearly 60,000 listed firms in all
major exchanges world-wide. The data on the firms include comprehensive financial statement
data, equity data and any credit events that occur. The database can be accessed and used at
NUS, and grants may include financial support for travel to Singapore for a short or long visit.
Researchers retain the right to use and publish the methods they develop as they see fit. RMI
retains the right to use the methods in full or in part in our rating activities. This may include
the use of the methodology in the daily update of results on our web portal and/or publication
of findings and results in the Global Credit Review. Any usage will be attributed to the
contributing researchers.

For further information on our call for Credit Research proposals, please visit: rmi.nus.edu.sg/cri
GLOBAL CREDIT REVIEW 65
INTRODUCTION
T
his document describes the im-
plementation of the system which
the NUS Risk Management In-
stitutes Credit Rating Initiative uses to
produce probabilities of default (PDs). As
of this version of the Technical Report,
these PDs cover exchange listed rms
in 30 economies in Asia, Asia-Pacic,
North America and Western Europe. The
individual PDs for nearly 30,000 rms
are computed daily. 2,200 of these rms
default forecasts are freely available
to all users at www.rmi.nus.edu.sg/cri,
along with aggregate PDs at the
economy and sector level for all the
rms.
The primary goal of this initiative is
to drive research and development in
the critical area of credit rating systems.
As such, a transparent methodology is
essential to this initiative. Having the
details of the methodology available to
everybody means that there is a base from
which suggestions and improvements can
be made. The objective of this Technical
Report is to provide a full exposition of
the CRI system. Readers of this document
who have access to the necessary data and
who have a sufcient level of technical
expertise will be able to implement a
similar system on their own.
The system used by the CRI will evolve
as new innovations and enhancements
are applied. This Technical Report will be
updated to reect changes in the system.
All versions will be available via the web
portal.
The remainder of this Technical
Report is organized as follows. The next
section describes the quantitative model
that is currently used to compute PDs
from the CRI. The model was rst de-
scribed in Duan, Sun and Wang (2011).
The description includes calibration
procedures, which are performed on a
monthly basis, and individual rm PD
computations, which are performed on a
daily basis.
Section 2 will describe the input vari-
ables of the model as well as the data used
to produce the variables for input into
the model. This model uses both input
variables that are common to all rms in
an economy and input variables that are
rm-specic. Another critical component
.532-)
#REDIT2ATING)NITIATIVE
4ECHNICAL2EPORT
version: 20ll update l
RMI sta articIe
|c. .ry ,Jost|crs c.
ccnnorts cr t||s ..t|c|o,
p|o.so ccrt.ct
O||vo. |or
ochen@nus.edu.sg
66 GLOBAL CREDIT REVIEW
when calibrating a credit rating system is the default
data, and this is also described in this section.
While Section 1 provides a broader description of
the model, Section 3 will describe the implementation
details that are necessary to apply given real world
issues of, for example, bad or missing data. The
specic technical details needed to develop an
operational system are also given, including details
on the monthly calibration, daily computation of
individual rm PDs and aggregation of the individual
rm PDs. Distance-to-default (DTD) in a Merton-
type model is one of the rm-specic variables. The
calculation for DTD is not the standard one, and has
been modied to allow a meaningful computation of
the DTD for nancial rms. While most academic
studies on default prediction exclude nancial rms
from consideration, it is important to include them
given that the nancial sector is a critical component
in every economy. The calculation for DTD is detailed
in this section.
Section 4 shows an empirical analysis for those
economies that are currently covered. While the analysis
shows excellent results in several economies, there is
room for improvement in a few others. This is because,
at the CRIs current stage of development the economies
all use the variables used in the academic study of US
rms in Duan et al. (2011). Future development within
the CRI will deal with variable selection specic to
different economies, and the performance is then
expected to improve. Variable selection and other
planned developments are discussed in Section 5.
I. MODEL DESCRIPTION
The quantitative model that is currently being used by the
CRI is a forward intensity model that was introduced in
Duan et al. (2011). This model allows default forecasts
to be made at a range of horizons. In the current CRI
implementation of this model, PDs are computed from
a horizon of one month up to a horizon of two years. In
other words, for every rm, the probability of that rm
defaulting within one month, three months, six months,
one year, eighteen months and two years is given. The
ability to assess credit quality for different horizons
is a useful tool for risk management, credit portfolio
management, policy setting and regulatory purposes,
since short and long-term credit risk proles can differ
greatly depending on a rms liquidity, debt structure
and other factors.
The forward intensity model is a reduced form
model in which the probability of default is computed
as a function of different input variables. These can be
rm-specic or common to all rms within an econ-
omy. The other category of default prediction model is
the structural model, whereby the corporate structure
of a rm is modeled in order to assess the rms pro-
bability of default.
A similar reduced form model by Dufe, Saita
and Wang (2007) relied on modeling the time series
dynamics of the input variables in order to calculate
PDs for different horizons. However, there is little
consensus on assumptions for the dynamics of variables
such as accounting ratios, and the model output will be
highly dependent on these assumptions. In addition, the
time series dynamics will be of very high dimension.
For example, with the two common variables and two
rm-specic variables that Dufe et al. (2007) use, a
sample of 10,000 rms gives a dimension of the state
variables of 20,002.
Given the complexity in modeling the dynamics of
variables such as accounting ratios, this model will be
difcult to implement if different forecast horizons
are required. The key innovation of the forward inten-
sity model is that PD for different horizons can be
consistently and efciently computed based only on
the value of the input variables at the time the predic-
tion is made. Thus, the model specication becomes
far more tractable.
Fully specifying a reduced form model includes
the specication of the function that computes a PD
from the input variables. This function is parameter-
ized, and nding appropriate parameter values is called
calibrating the model. The forward intensity model
can be calibrated by maximizing a pseudo-likelihood
function. The calibration is carried out by economy
and all rms within an economy will use the same
parameter values along with each rms variables in
order to compute a rms PD.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 67
Subsection 1.1 will describe the modeling framework,
including the way PDs are computed based on a set of
parameter values for the economy and a set of input
variables for a rm. Subsection 1.2 explains how the
model can be calibrated.
1.1 Modeling Framework
While the model can be formulated in a continuous
time framework, as done in Duan et al. (2011), an
operational implementation will require discretization
in time. Since the model is more easily understood in
discrete time, the following exposition of the model
will begin in a discrete time framework.
Variables for default prediction can have vastly
different update frequencies. Financial statement data is
updated only once a quarter or even once a year, while
market data like stock prices are available at frequencies
of seconds. A way of compromising between these two
extremes is to have a fundamental time period t of
one month in the modeling framework. As will be seen
later, this does not preclude updating the PDs on a daily
basis. This is important since, for example, large daily
changes in a rms stock price can signal changes in
credit quality even when there is no change in nancial
statement data.
Thus, for the purposes of calibration and subse-
quently for computing time series of PD, the input
variables at the end of each month will be kept for
each rm. The input variables associated with the i
th

rm at the end of the n
th
month (at time t = nt) is
denoted by X
i
(n). This is a vector consisting of two
parts: X
i
(n) = (W(n), U
i
(n)). Here, W(n) is a vector of
variables at the end of month n that is common to all
rms in the economy and U
i
(n) is a vector of variables
specic to rm i.
In the forward intensity model, a rms default is
signaled by a jump in a Poisson process. The probability
of a jump in the Poisson process is determined by the
intensity of the Poisson process. The forward intensity
model draws an explicit dependence of intensities at
time periods in the future (that is, forward intensities)
to the value of input variables at the time of prediction.
With forward intensities, PDs for any horizon can be
computed knowing only the value of the input variable
at the time of prediction, without needing to simulate
future values of the input variables.
There is a direct analogy in interest rate modeling. In
spot rate models where dynamics on a short-term spot
rate are specied, bond pricing requires expectations
on realizations of the short rate. Alternatively, bond
prices can be computed directly if the forward rate
curve is known.
One issue in default prediction is that rms can
exit public exchanges for reasons other than default.
For example, in mergers and acquisitions involving
two public companies, there will be one company that
delists from its stock exchange. This is important in
predicting defaults because a default cannot happen if
a rm has been previously delisted. An exception is if
the exit is a distressed exit and is followed soon after
by a credit event. See Subsection 2.4 for details on how
this case is handled in the CRI system.
In order to take these other exits into account, de-
faults and other exits are modeled as two independent
Poisson processes, each with their own intensity. While
defaults and exits classied as non-defaults are mutual-
ly exclusive by denition, the assumption of independ-
ent Poisson processes does not pose a problem since the
probability of a simultaneous jump in the two Poisson
processes is negligible. In the discrete time framework,
the probability of simultaneous jumps in the same time
interval is non-zero. As a modeling assumption, a simul-
taneous jump in the same time interval by both the de-
fault Poisson process and the non-default type exit Pois-
son process is considered as a default. In this way, there
are three mutually exclusive possibilities during each
time interval: survival, default and non-default exit. As
with defaults, the forward intensity of the Poisson proc-
ess for other exits is a function of the input variables.
The parameters of this function can also be calibrated.
To further illustrate the discrete framework, the
three possibilities for a rm at each time point are di-
agrammed. Either the rm survives for the next time
period t, or it defaults within t, or it has a non-
default exit within t. This setup is pictured in Figure 1.
Information about rm i is known up until time t = mt
and the gure illustrates possibilities in the future
68 GLOBAL CREDIT REVIEW
II GUR 1
|o|.J|tct|o. ox|tsJ.v|v.| t.oo |c. |.n J, v|o.oJ |.cn t|no U NU
between t = (n 1) t and (n + 1) t. Here, m and n
are integers with m < n.
The probabilities of each branch are, for example:
p
i
(m, n) the conditional probability viewed from
t = mt that rm i will default before (n + 1) t, con-
ditioned on rm i surviving up until nt. Likewise,
p

i
(m, n) is the conditional probability viewed from
t = mt that rm i will have a non-default exit before
(n + 1) t, conditioned on rm i surviving up until nt.
It is the modelers objective to determine p
i
(m, n) and
p

i
(m, n), but for now it is assumed that these quantities
are known. With the conditional default and other exit
probabilities known, the corresponding conditional
survival probability of rm i is 1 p
i
(m, n) p

i
(m, n).
With this diagram in mind, the probability that a
particular path will be followed is the product of the
conditional probabilities along the path. For example,
the probability at time t = mt of rm i surviving until
(n 1)t and then defaulting between (n 1)t and
nt is:
(1)
.
Here, V
i
is the default time for rm i measured in
units of months,

V
i
is the other exit time measured in
units of months, and the product is equal to one if
there are no terms in the product. The condition V
i
<

V
i
is the requirement that the rm defaults before it has a
non-default type of exit. Note that by measuring exits in
units of months, if, for example, a default occurs at any
time in the interval ((n 1)t, nt] then V
i
= n.
Using equation (1), cumulative default probabilities
can be computed. At mt the probability of rm i
defaulting at or before nt and not having an other exit
before t = nt is obtained by taking the sum of all of
the paths that lead to default at or before nt:
(2)
While it is convenient to derive the probabilities
given in equations (1) and (2) in terms of the conditional
probabilities, expressions for these in terms of the
forward intensities need to be found, since the forward
intensities will be functions of the input variables
X
i
(m). The forward intensity for the default of rm
.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 69

i that is observed at time t = mt for the forward
time interval from t = nt to (n +1)t, is denoted by
h
i
(m, n) where m n. The corresponding forward
intensity for a non-default exit is denoted by h

i
(m, n).
Because default is signaled by a jump by a Poisson
process, its conditional probability is a simple function
of its forward intensity:
(3)
Since joint jumps in the same time interval are assigned
as defaults, the conditional other exit probability needs
to take this into account:
(4)
The conditional survival probabilities in equations (1)
and (2) are computed as the conditional probability that
the rm does not default in the period and the rm does
not have a non-default exit either:
(5)
It remains to specify the dependence of the forward
intensities on the input variables X
i
(m). The forward
intensities need to be positive so that the conditional
probabilities are non-negative. A standard way to impose
this constraint is to specify the forward intensities
as exponentials of a linear combination of the input
variables:
(6)
Here, D and D

are coefcient vectors that are func-

tions of the number of months between the observation
date and the beginning of the forward period (n m),
and Y
i
(m) is simply the vector X
i
(m) augmented by
a preceding unit element: Y
i
(m) = (1, X
i
(m)). The unit
element allows the linear combination in the argument
of the exponentials in equation (6) to have a non-zero
intercept.
.
.
.
.
In the current implementation of the forward inten-
sity model in the CRI, the maximum horizon is 24
months and there are 12 input variables plus the
intercept. So there are 24 sets of each of the coefcient
vectors denoted by D(0),, D(23) and D

(0),, D

(23)
and each of these coefcient vectors has 13 elements.
While this is a large set of parameters, as will be seen
in the next subsection, the calibration is tractable
because the parameters for each horizon can be done
independently from each other, and the default para-
meters can be calibrated separately from the other exit
parameters.
Before giving the probabilities in (1) and (2) in
terms of the forward intensities, a notation is intro-
duced for the forward intensities that makes clear
which parameters are needed for the forward intensity
in question:
(7)
This is the forward default intensity. The correspond-
ing notation for other exit forward intensities is then
just . So, the probability in (1) is
expressed in terms of the forward intensities, using
(3) for the conditional default probability and (5) for
the conditional survival probability:

(8)
This probability will be relevant in the next subsection
during the calibration. The cumulative default prob-
ability given in equation (2) in terms of the forward
intensities is then:
.
70 GLOBAL CREDIT REVIEW
(9)
This formula is used to compute the main output of
the CRI: an individual rms PD within various time
horizons. The D and D

parameters are obtained when the

rms economy is calibrated, and using those together
with the rms input variables yields the rms PD.
1.2 Model Calibration
The empirical dataset used for calibration can be des-
cribed as follows. For the economy as a whole, there are
N end of month observations, indexed as n = 1,, N.
Of course, not all rms will have observations for each
of the N months as they may start later than the start of
the economys dataset or they may exit before the end
of the economys dataset. There are a total of I rms
in the economy, and they are indexed as i = 1,, I.
As before, the input variables for the i
th
rm in the n
th

month is X
i
(n). The set of all observations for all rms
is denoted by X.
In addition, the default times V
i
and non-default exit
times V

i
for the i
th
rm are known if the default or other
exit occurs after time t = t and at or before t = Nt.
The possible values for V
i
and V

i
are integers between
2 and N, inclusive. If a rm exits before the month
end, then the exit time is recorded as the rst month
end after the exit. If the rm does not exit before t =
Nt, then the convention can be used that both of these
values are innite. If the rm has a default type of exit
within the dataset, then V

i
can be considered as innite.
If instead the rm has a non-default type of exit within
the dataset, then V
i
can be considered as innite. The
set of all default times and non-default exit times for
all rms is denoted by V and V

, respectively. The rst

month in which rm i has an observation is denoted by
t
0i
. Except for cases of missing data, these observations
continue until the end of the dataset if the rm never
exits. If the rm does exit, the last needed input variable
X
i
(n) is for n = min(V
i
, V

i
) 1.
The calibration of the D and D

parameters is
done by maximizing a pseudo-likelihood function.
The function to be maximized violates the standard
assumptions of likelihood functions, but Appendix A in
Duan et al. (2011) derives the large sample properties
of the pseudo-likelihood function.
In formulating the pseudo-likelihood function, the
assumption is made that the rms are conditionally
independent from each other. In other words, corre-
lations arise naturally from sharing common factors
(n) and any correlations there are between different
rms rm-specic variables. With this assumption,
the pseudo-likelihood function for horizon of l
months, a set of parameters D and D

and the dataset

(V, V

, X) is:
(10)
Here, P
l
( D, D

V
i
, V

i
, X
i
(m)) is a probability for rm i, with
the nature of the probability depending on what happens
to the rm during the period from month m to month
m + l. This is dened as:

(11)

.
.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 7l
In words, if rm i survives from the observation
time at month m for the full horizon l until at least
m + l, then the probability is the model-based survival
probability for this period. This is the rst term in
(11). The second term handles the cases where the
rm has a default within the horizon, in which case
the probability is the model-based probability of the
rm defaulting at the month that it ends up defaulting,
as given in equation (8). The third term handles the
cases where the rm has a non-default exit within the
horizon, in which case the probability is the model-
based probability of the rm having a non-default type
exit at the month that the exit actually does occur.
The expression for this probability uses the conditional
non-default type exit probability given in equation (4).
The nal two terms handle the cases where the rm
is not in the data set at month m either the rst
observation for the rm is after m or the rm has
already exited. A constant value is assigned in this
case so that this rm will not affect the maximization
at this time point.
The pseudo-likelihood function given in (10) can
be numerically maximized to give estimates for the
coefcients D and D

. Notice though that the sample

observations for the pseudo-likelihood function are
overlapping if the horizon is longer than one month.
For example, when l = 2, default over the next two
periods from month m is correlated to default over
the next two periods from month m + 1 due to the
common month in the two sample observations.
However, in Appendix A of Duan et al. (2011), the
maximum pseudo-likelihood estimator is shown
to be consistent, in the sense that the estimators
converge to the true parameter value in the large
sample limit.
It would not be feasible to numerically maximize
the pseudo-likelihood function using the expression
given in (11), due to the large dimension of the D and
D

parameters. Notice though that each of the terms in

(11) can be written as a product of terms containing
only D and terms containing only D

. This will allow

separate maximizations with respect to D and with
respect to D

.
The D and D

specic versions of (11) are:

(12)
Then, the D and D

specic versions of the pseudo-

likelihood function are given by:

(13)
With the denitions given in (12) and (13), it can be
seen that:
(14)
Thus, and can be separately maximized to
nd their respective parameters. A further important
separation is a separation by horizons. Notice that we
can decompose and as:
72 GLOBAL CREDIT REVIEW
(15)
where:

(16)
Thus, the D and D

specic pseudo-likelihood functions

can be decomposed as:
(17)
where,
(18)
Thus, for every horizon and
can be separately maximized.
In summary, for the current CRI implementation
where the horizons are from one month to 24 months,
and where there are 13 variables, a 2 24 13
dimensional maximization is turned into a 13 dimen-
sional maximization done 2 24 times. This makes the
calibration problem tractable. Additional implementation
details on the calibration are given in Section 3.
II. INPUT VARIABLES AND DATA
Subsection 2.1 describes the input variables used in the
quantitative model. Currently, the same description of
input variables is common to all the economies under
the CRIs coverage. Future enhancements to the CRI
system will allow different input variables for different
economies. The effect of each of the variables on the
PD output will be discussed in the empirical analysis
of Section 4.
Subsection 2.2 gives the data sources and relevant
details of the data sources. There are two categories of
data sources: current and historical. Data sources used
for current data need to be updated in a timely manner
so that daily updates of PDs are meaningful. They also
need to be comprehensive in their current coverage of
rms. Data sources that are comprehensive for current
data may not necessarily have comprehensive historical
coverage for different economies. Other data sources are
thus merged in order to obtain comprehensive coverage
for historical and current data.
Subsection 2.3 indicates the elds from the data
sources that are used to construct the input variables.
For some of the elds, proxies need to be used for a rm
if the preferred eld is not available for that rm.
Subsection 2.4 discusses the denition and sources
of defaults and of other exits used in the CRI.
2.1 Input Variables
Following the notation that was introduced in Section 2,
rm is input variables at time t = nt are represented
by the vector X
i
(n) = (W(n), U
i
(n)) consisting of a vector
W(n) that is common to all rms in the same economy,
and a rm-specic vector U
i
(n) which is observable
from the date the rms rst nancial statement is
released, until the month end before the month in which
the rm exits, if it does exit.
In Duan et al. (2011), different variables that
are commonly used in the literature were tested as
candidates for the elements of W(n) and U
i
(n). Two
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 73
common variables and ten rm-specic variables, as
described below, were selected as having the greatest
predictive power for corporate defaults in the United
States. In the current stage of development, this same
set of twelve input variables is used for all economies.
Future development will include variable selection for
rms in different economies.
Common variables
The vector W(n) contains two elements, consisting of:
1. Stock index return: the trailing one-year simple
return on a major stock index of the economy.
2. Interest rate: a representative 3-month short term
interest rate.
Firm-specic variables
The ten rm-specic input variables are transforma-
tions of measures of six different rm characteristics.
The six rm characteristics are: (i) volatility-adjusted
leverage; (ii) liquidity; (iii) protability; (iv) relative
size; (v) market mis-valuation/future growth opport-
unities; and (vi) idiosyncratic volatility.
Volatility-adjusted leverage is measured as the
distance-to-default (DTD) in a Merton-type model.
The calculation of DTD used by the CRI allows a
meaningful DTD for nancial rms, a critical sector
that must be excluded from most DTD computations.
This calculation is detailed in Section 3.
Liquidity is measured as a ratio of cash and short
term investments to total assets, protability is mea-
sured as a ratio of net income to total assets, and
relative size is measured as the logarithm of the ratio of
market capitalization to the economys median market
capitalization.
Duan et al. (2011) transformed these rst four
characteristics into level and trend versions of the
measures. For each of these, the level is computed as
the one-year average of the measure, and the trend is
computed as the current value of the measure minus the
one-year average of the measure. The level and trend
of a measure has seldom been used in the academic or
industry literature for default prediction, and Duan et al.
(2011) found that using the level and trend signicantly
improves the predictive power of the model for short-
term horizons.
To understand the intuition behind using level and
trend of a measure as opposed to using just the current
value, consider the case of two rms with the same
current value for all measures. If the level and trend
transformations were not performed, then only the
II GUR 2
.c |.ns .|t| .|| cJ..ort v.|Jos o,J.| tc o.c| ct|o., bJt || t.orJ|rg |r t|o cppcs|to J|.oct|cr
74 GLOBAL CREDIT REVIEW
current values would be used and the two rms would
have identical PD. Suppose that for the rst rm the
DTD had reached its current level from a high level,
and for the second rm the DTD had reached its current
level from a lower level (see Figure 2). The rst rms
leverage is increasing (worsening) and the second
rms leverage is decreasing (improving). If there is a
momentum effect in DTD, then rm 1 should have a
higher PD than rm 2.
Duan et al. (2011) found evidence of the momentum
effect in DTD, liquidity, protability and size. For the
other two rm characteristics, applying the level and
trend transformation did not improve the predictive
power of the model.
One of the remaining two rm characteristics is
the market mis-valuation/future growth opportunities
characteristic, which is taken as the market-to-book asset
ratio and measured as a ratio of market capitalization
and total liabilities to total assets. One can see whether
the market mis-valuation effect or the future growth
opportunities effect dominates this measure by looking
at whether the parameter for this variable is positive or
negative. This will be further discussed in the empirical
analysis of Section 4.
The nal rm characteristic is the idiosyncratic
volatility which is taken as sigma, following Shumway
(2001). Sigma is computed by regressing the monthly
returns of the rms market capitalization on the
monthly returns of the economys stock index, for the
previous 12 months. Sigma is dened to be the standard
deviation of the residuals of this regression. Shumway
(2001) reasons that sigma should be logically related to
bankruptcy since rms with more variable cash ows
and therefore more variable stock returns relative to a
market index are likely to have a higher probability of
bankruptcy.
Finally, the vector U
i
(n)

contains ten elements, con-
sisting of:
1. Level of DTD.
2. Trend of DTD.
3. Level of (Cash + Short term investments)/Total
assets, abbreviated as CASH/TA.
4. Trend of CASH/TA.
5. Level of Net income/Total Assets, abbreviated
as NI/TA.
6. Trend of NI/TA.
7. Level of log (Firm market capitalization/Econ-
omys median market capitalization), abbrevi-
ated as SIZE.
8. Trend of SIZE.
9. Current value of (Market capitalization + total
liabilities)/Total asset, abbreviated as M/B.
10. Current value of SIGMA.
The data elds that are needed to compute DTD
and Short term investments are described in Subsection
2.3. The remaining data elds required are straightforward
and standard. The computation for DTD is explained in
Section 3.
2.2 Data Sources
There are two data sources that are used for the
daily PD updates: Thomson Reuters Datastream and
the Bloomberg Data License Back Ofce Product.
Many of the common factors such as stock index
prices and short term interest rates are retrieved from
Datastream.
Firm-specic data comes from Bloombergs Back
Ofce Product which delivers daily update les by
region via FTP after the respective market closes.
All relevant data is extracted from the FTP les and
uploaded into the CRI database for storage. From
this, the necessary elds are extracted and joined with
previous months of data.
The Back Ofce Product includes daily market
capitalization data based on closing share prices and
also includes new nancial statements as companies
release them. Firms will often have multiple versions
of nancial statements within the same period, with
different accounting standards, ling statuses (most
recent, preliminary, original, reclassied or restated),
currencies or consolidated/unconsolidated indicators.
A major challenge lies in prioritizing these nancial
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 75
statements to decide which data should be used. The
priority rules are described in Section 3.
The rm coverage of the Back Ofce Product is
of sufcient quality that nearly 30,000 rms can be
updated on a daily basis in the 30 economies under
the CRIs coverage. While the current coverage is
quite comprehensive, historical data from the Back
Ofce Product can be sparse for certain eco-nomies.
For this reason, various other databases are merged in
order to ll out the historical data. The other databases
used for historical data are: a database from the Taiwan
Economics Journal (TEJ) for Taiwanese rms, a database
provided by Korea University for South Korean rms,
and data from Prowess for Indian rms.
With all of the databases merged together and for
the 30 economies under CRIs coverage, over 50,000
exchange listed rms are in the CRI database. This
includes over 20,000 delisted rms. The historical
coverage of the rm data goes back to the early
1990s.
2.3 Constructing Input Variables
The chosen stock indices and short term interest rates
for the 30 economies under the CRIs current coverage
are listed in Table A.2 and Table A.3, respectively. All
economies are listed by their three letter ISO code given
in Table A.1.
Most of the rm-specic variables can be readily
constructed from standard elds from rms nancial
statements in addition to daily market capitalization
values. The only two exceptions are the DTD and the
liquidity measure.
The calculation for DTD is explained in Section 3.
In the calculation, several variables are required. One
variable is a proxy for a one-year risk-free interest
rate, and the choices for each of the 30 economies
are listed in Table A.4. Total assets, long term bor-
rowing and total liabilities are also required, but are
standard nancial statement elds and present no
difculties.
Total current liabilities are also required, and due to
the relatively large numbers of rms that are missing
this value, proxies had to be found. The preferred
Bloomberg eld for this is BS_CUR_LIAB. If this
is missing, then the sum of BS_ST_BORROW, BS_
OTHER_ST_LIAB and BS_CUST_ACCPT_LIAB_
CUSTDY_SEC (customers acceptance and liabilities/
custody securities) is used. If one or two of these are
missing, zero is inserted for those elds, but at least
one eld is required.
The liquidity measure requires different elds
between nancial and non-nancial rms. For non-
nancial rms, the numerator (Cash+Short-term
investments) of the ratio is taken as the sum of BS_
CASH_NEAR_CASH_ITEM and BS_MKT_SEC_
OTHER_ST_INVEST (marketable securities and other
short term investments).If BS_MKT_SEC_OTHER_
ST_INVEST is missing, replace it with zero (but
BS_CASH_NEAR_CASH_ITEM is required).
It was found that this sum frequently overstated the
liquidity for nancial rms. In place of BS_MKT_
SEC_OTHER_ST_INVEST, nancial rms use the
sum of ARD_SEC_PURC_UNDER_AGR_TO_
RESELL (securities purchased under agreement to
re-sell), ARD_ST_INVEST and BS_INTERBANK_
ASSET. If one or two of these are missing, zero
is inserted for those elds, but at least one eld is
required. The ARD_ prex indicates that these are
as reported numbers directly from the nancial
statements. As such, for some rms these elds may
need to be adjusted to the same units before adding
them to other elds.
Table A.5 contains summary statistics of the rm-
specic variables: DTD, CASH/TA, NI/TA, SIZE, M/B,
and SIGMA, with the summary statistics provided for
rms grouped by economy.
2.4 Data for Defaults
The CRI database contains credit events of over 4,000
rms from 1990 to the present. The defaults events
come from numerous sources, including Bloomberg,
Compustat, CRSP, Moodys reports, TEJ, exchange
web sites and news sources.
76 GLOBAL CREDIT REVIEW
The CRI system considers two broad categories
of default: bankruptcy lings and default corporate
actions. Within bankruptcies, the sub-categories are
listed in Table A.6. Delistings that are labeled as being
due to bankruptcy are categorized as a bankruptcy ling
and the delisting date is used if the actual bankruptcy
ling event date cannot be found.
Default corporate actions can include missed or
delayed interest or principal payments by the due date,
or as debt restructuring. The more precise sub-categories
of default corporate actions used by the CRI are also
listed in Table A.6.
The exit events that are not considered as defaults
in the CRI system are listed in Table A.7. Firms that
are delisted from an exchange and then experience a
default event within 365 calendar days of the delisting
have the exit event re-classied as a credit default.
This is consistent with Shumway (2001) who re-
classies a delisting event to credit default if there
is any bankruptcy event within ve years after the
delisting event.
In addition to the aforementioned events, there are
various cases that require special attention. A non-
exhaustive overview of such events and the way they are
treated in the CRI system is listed below. The treatments
of these events are compared to the treatments by the
three major credit rating agencies (CRAs), as described
in Moodys Investor Services (2011), Standard & Poors
(2011) and Fitch Ratings (2011).
Missed or delayed payments made within the
grace period are not counted as defaults: The
major CRAs forgive such events as they focus on
the ability or willingness to pay when assessing
credit risks of an issuer. Therefore, a payment made
within the grace period is seen as the issuer being
willing to uphold the debt contract. Considering
that the typical grace period for missed interest or
principal payment is 7 to 14 days, this criterion is
within the Basel II guidelines (Basel Committee
on Banking Supervision, 2006), which allows a
90 day grace period.
Related obligor default will be assessed on a case-
by-case basis: The CRI, like major CRAs, does
not consider related party-default (e.g. subsidiary
bankruptcy) as a default event. However, the
Monetary Authority of Singapore (2004) advises
that related obligor defaults can be default events
depending on the economic dependence and
integration between the subsidiary and the parent
company. When a non-operating holding parent
company relies heavily on its subsidiary, bankruptcy
by the subsidiary will cause a considerable economic
impact on the parent company. These cases will be
carefully reviewed.
Selective default on one obligation but not on
the others is counted as a default event: The
CRI considers default on one of the obligations
but not the others as default by the overall
company. This is because, in general, a default in
one obligation is a sign of business deterioration
and often is followed by a bankruptcy ling.
S&P and Moodys consider selective default as
default. Fitch Ratings categorizes such events as
a Restricted Default.
Another important challenge in identifying default
events is linked to the fact that denitions of credit
default can vary across different jurisdictions and
between different data sources. An area of continuing
development is in normalizing to a common set of
denitions.
Table A.8 lists the total number of rms, number
of defaults and number of other exits in each of the
30 economies each year from 1992 to 2011. Note that
the total number of rms here includes all rms where
the primary listing of the shares are on an exchange
in that economy and may include rms where there
are too many missing data values for a PD estimate to
be made. However, the number of rms listed on the
CRI web portal under the tab Aggregate Forecast
includes rms that are domiciled in that economy and
excludes rms where a PD cannot be produced due
to missing data.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 77
III. IMPLEMENTATION DETAILS
Section 1 described the modeling framework under-
lying the current implementation of the CRI rating
system. It focused on theory rather than the details
encountered in an operational implementation. The
present section describes how the CRI system handles
these more specic issues.
Subsection 3.1 describes implementation details re-
lated to data, mainly dealing with data cleaning and
missing data. Subsection 3.2 describes the specic
computation of distance-to-default (DTD) used by the
CRI system that leads to meaningful DTD for nan-
cial rms. Subsection 3.3 explains how the calibration
previously described in Subsection 1.2 can be imple-
mented. Subsection 3.4 gives the implementation details
relevant to the daily output. This includes an explanation
of the various modications needed to compute daily
PD so that the daily PD is consistent with the usual
month end PD, and a description of the computation of
the aggregate PDs provided by the CRI.
3.1 Data Treatment
Fitting data to monthly frequency: Historical end
of month data for every rm in an economy is required
to calibrate the model. For daily data such as market
capitalization, interest rates and stock index values,
the last day of the month for which there is valid data
is used.
For nancial statement variables, data is used start-
ing from the period end of the statement lagged by
three months. This is to ensure (insofar as is possible)
that predictions are made based on information that
was available at the time the prediction was made. Of
course, for more recent data where the CRI database
contains the nancial statement but the period end
lagged by three months is after the current day, the
nancial statement is used in computing PD. The CRI
considers nancial statement variables to be valid for
one year without restriction after they are rst used.
Currency conversions are required if the market
capitalization or any of the nancial statement variables
are reported in a currency different than the currency
of the economy. If a currency conversion is required,
the foreign exchange rate used is that reported at the
relevant market close. For rms traded in Asia and
Asia-Pacic, the Tokyo closing rate is used; for rms
traded in Western Europe, the London closing rate is
used; and for rms traded in North America, the New
York closing rate is used. For market capitalizations,
the FX rate used is for the date that the market capitaliza-
tion is reported. For nancial statement variables, the
FX rate used is for the date of the period end of the
statement.
Priority of nancial statements: As described in
Subsection 2.2, data provided in Bloombergs Back
Ofce Product can include numerous versions of n-
ancial statements within the same period. If there are
multiple nancial statements with the same period end,
priority rules must be followed in order to determine
which to use. The formulation and implementation
of these rules is a major challenge and an area of
continuing development. The current priority rules are
as follows.
The rst rule prioritizes by consolidated/unconso-
lidated status. This status is relevant only to rms in
India, Japan, South Korea and Taiwan, so this rule is
only relevant in those economies. Most rms in these
economies issue unconsolidated nancial statements
more frequently than consolidated ones, so these are
given higher priority. This simple prioritization can,
however, lead to cases where the nancial statements
used switch from consolidated statements to unconso-
lidated statements and back again. A more complex
prioritization rule is currently under development, with
the intention of avoiding this situation.
If, after the rst prioritization rule has been applied,
there are still multiple nancial statements, the second
rule is applied. This is prioritization by scal period. In
most economies, annual statements are required to be
audited, whereas other scal periods are not necessarily
audited. The order of priority from highest to lowest is,
therefore: annual, semi-annual, quarterly, cumulative,
and nally other scal periods. The one variable that
is currently an exception to this rule is net income.
78 GLOBAL CREDIT REVIEW
Net income is a ow variable so adjustments need
to be made to annualize the net income from non-
annual nancial statements. Study needs to be done to
determine whether seasonal effects will unduly affect
PD estimates before net income from non-annual
statements can be introduced.
The third prioritization rule is based on ling
status. The Most Recent statement is used before
the Original statement, which is used before the
Preliminary statement.
The nal prioritization rule is based on the account-
ing standard. Here, nancial statements that are report-
ed using Generally Accepted Accounting Principles
(GAAP) are given higher priority than nancial state-
ments that are reported using International Financial
Reporting Standards (IFRS). If an accounting standard
is not indicated at all, the nancial statement is not
used.
Provisions for missing values and outliers: Missing
values and outliers are dealt with by a three step
procedure. In the rst step, the ten rm-specic input
variables are computed for all rms and all months. In
the second step, outliers are eliminated by winsorization.
In the nal step, missing values are replaced under
certain conditions.
The rst step is to compute the input variables and
determine which are missing. As mentioned previously,
nancial statement variables are carried forward for
one year after the date that they are rst used. This
is generally three months after the period end of the
statement. If no nancial statement is available for the
company within this year, then the nancial statement
variable will be missing. For market capitalization, if
there is no valid market capitalization value within the
calendar month, then the value is set to missing.
For illiquid stocks, if there has been no valid market
capitalization value for a rm within the last 90 cal-
endar days, then the market capitalization is deemed
to not properly reect the value of the rm. The rm
is considered to have exited with a non-default event.
Once the rm starts trading again and a new nancial
statement is released, the rm can enter back into the
calibration. With regard to historical PD, the PD can be
reported again once there are enough valid variables.
With regard to the level variables, the current month
and the last eleven months are averaged to compute
the level. There is no lower limit on the number of
valid observations. Only if all of the values are missing
is the level variable considered to be missing.
For the trend variable, the level is subtracted from
the current month. If the current month is missing, then
the trend variable is set to missing.
The value of M/B is set to missing if any of the
following values: market capitalization, total liabili-
ties or total assets of the rm, are missing. For the
computation of SIGMA, seven valid returns over the
last twelve months of possible returns are required for
the regression. If there are less than seven valid returns,
SIGMA is set to missing.
In this way, the eight trend and level variables plus
M/B and SIGMA are computed and evaluated as miss-
ing or present. Winsorization can then be performed
as a second step to eliminate outliers. The volume of
outliers is too large to be able to determine whether
each one is valid or not, so winsorization applies a
oor and a cap on each of the variables. The historical
0.1 percentile and 99.9 percentile for all rms in the
economy are recorded for each of the ten variables. Any
values that exceed these levels are set to equal these
boundary values.
With a winsorization level and 0.1 percentile and
99.9 percentile, the boundary values still may not be
reasonable. For example, NI/TA levels of nearly 25
have been observed at this stage. In these cases, a
more aggressive winsorization level is applied, until
the boundary values are reasonable. Thus, the winso-
rization level is economy and variable specic, and
will depend on the data quality for that economy and
variable. Winsorization levels different than the default
of 0.1 percentile and 99.9 percentile are indicated in
Table A.5.
A third and nal step can be taken to deal with miss-
ing values. If, during a particular month, no variables
for a rm are missing, then the PD can be computed.
If six or more of these ten variables are missing, there
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 79
are deemed to be too many missing observations and
no replacements are made.
If between one and ve variables are missing out of
the ten, the rst step is to trace back for at most twelve
months to use previous values of these variables instead.
If this does not succeed in replacing all of the variables,
a replacement by sector medians is done. The median is
for the nancial or non-nancial rms (as indicated by
their Level I Bloomberg Industry Classication System)
within the economy during that month. Replacement
by the sector median should have a neutral effect on
the PD of the rm; the rm is assessed by the other
variables that it does have values for. This sector
median is always performed in calibration. However,
when reporting historical PD, the sector replacement
is not done if it results in a relative change in PD of
10% or more where the initial PD was at or above
100bps, or an absolute change in PD of 10bps or more
where the initial PD was below 100bps.
Inclusion/exclusion of companies for calibration:
Firms are included within an economy for calibration
when the primary listing of the rm is on an exchange
in the economy. This ensures that all rms within
the economy are subject to the same disclosure and
accounting rules.
There are a relatively small number of rms that
are dual listed, in which two corporations listed in
different exchanges operate as a single entity but retain
separate legal status. In the CRI system, a combined
company will be assigned to the single economy it
is most associated with. An example is the Rio Tinto
Group. This consists of Rio Tinto plc, listed in the
UK; and Rio Tinto Limited, listed in Australia. Most
of Rio Tintos operations are in Australia rather than
the UK, so Rio Tinto is assigned to Australia.
In the US, rms traded on the OTC markets or the
Pink Sheets are not considered as exchange listed so
are not included in calibration or in the reporting of
PDs. Many of these rms are small or start-up rms. In-
cluding this large group of companies would skew the
calibration and the aggregate results. The TSX Venture
Exchange in Canada also contains only small and
start-up rms, so rms listed here are also excluded.
Other examples include Taiwans GreTai Securities
Market and Singapores Catalist. The challenge for
markets outside of the US or Canada is that the data
on whether rms are listed on the smaller markets
rather than the main board is difcult to obtain. For
all economies besides the US and Canada, there is
continuing work being done in the CRI system to
exclude rms that are not listed on major exchanges
within a country.
Firms that record an exit (other than due to no
trading for 90 calendar days) are not entered back into
the calibration even if the rm continues to trade and
issue nancial statements, as can happen after rms
declare bankruptcy. There are two exceptions to this
exclusion. The rst, determined on a case by case basis,
is if the rm should be deemed to have re-emerged
from bankruptcy. The second exception is for all rms
in China, where two situations are prevalent. The rst
situation is that the rm experiences few repercussions
from the default and continues operating normally. The
other situation is for one rm to take over a defaulted
rms listing. This happens due to the limited supply
of exchange listings. Because of these two situations,
the norm for rms based in China is to emerge from a
default, so the CRI system enters all of these companies
back into the calibration as new companies.
3.2 Distance-to-Default Computation
The distance-to-default (DTD) computation used in
the CRI system is not a standard one. Standard com-
putations exclude nancial rms, but excluding the
nancial sector means neglecting a critical part of any
economy. So the standard DTD computation must be
extended to give meaningful estimates for nancial
rms as well. The description of the specialized DTD
computation starts with a brief description of the Merton
(1974) model. Mertons model makes the simplifying
assumption that rms are nanced by equity and a single
zero-coupon bond with maturity date T and principal
L. The asset value of the rm V
t
follows a geometric
Brownian motion:
80 GLOBAL CREDIT REVIEW
(19)
Here, B
t
is standard Brownian motion, O is the drift
of the asset value in the physical measure and U is the
volatility of the asset value. Equity holders receive
the excess value of the rm above the principal of
the zero-coupon bond and have limited liability, so
the equity value at maturity is: E
T
= max (V
T
L, 0).
This is just a call option payoff on the asset value with
a strike value of L. Thus, the Black-Scholes option
pricing formula can be used for the equity value at
times t before T:
(20)
where r is the risk-free rate, N() is the standard normal
cumulative distribution function, and:
. (21)
In Mertons model, DTD is dened as volatility
scaled distance of the expected asset value under the
physical measure at maturity T from the default point L:
. (22)
The standard KMV assumptions given in Crosbie
and Bohn (2003) are to set the time to maturity T t at a
value of one year, and the principal of the zero-coupon
bond L to a value equal to the rms current liabilities
plus one half of its long-term debt. Here, the current
liabilities and long-term debt are taken from the rms
nancial statements. If the rm is missing the current
liabilities eld, then various substitutes for this eld can
be used, as described in Subsection 2.3.
This is a poor assumption of the debt level for n-
ancial rms, since they typically have large liabilities,
such as deposit accounts, that are neither classied as
current liabilities nor long-term debt. Thus, using these
standard assumptions means ignoring a large part of the
debt of nancial rms.
To properly account for the debt of nancial rms,
Duan (2010) includes a fraction F of a rms other
liabilities. The other liabilities are dened as the rms
total liabilities minus both the short and long-term
debt. The debt level L then becomes the current liabili-
ties plus half of the long-term debt plus the fraction F
multiplied by the other liabilities, so that the debt level
is a function of F. The standard KMV assumptions are
then a special case where F = 0.
The fraction F can be optimized along with O and
U in the maximum likelihood estimation method de-
veloped in Duan (1994, 2000). Following Duan et al.
(2011), the rms market value of assets is standardized
by its book value A
t
so that the scaling effect from a
major investment or nancing by the rm will not
distort the time series from which the parameter values
are estimated. Thus, the log-likelihood function is:
(23)
Here, n is the number of days with observations of
the equity value in the sample, is the implied asset
value found by solving equation (20), is computed
with equation (21) using the implied asset value, and
h
t
is the number of trading days as a fraction of the
year between observations t 1 and t. Notice that the
implied asset value and are dependent on F by virtue
of the dependence of L on F.
Implementation of DTD computation: The DTD
at the end of each month is needed for every rm
in order to calibrate the forward intensity model. A
moving window, consisting of the last one year of data
before each month end is used to compute the month
end DTD. Daily market capitalization data based on
closing prices is used for the equity value in the implied
asset value computation of equation (20). If there are
fewer than 50 days of valid observations for the market
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 8l
capitalization, then the DTD value is set to missing. An
observation is valid if there is positive trading volume
that day. If the trading volume is not available, the
observation is assumed to be valid if the value for the
market capitalization changes often enough. The pre-
cise criterion is as follows: if the market capitalization
does not change for three days or more in a row, the rst
day is taken as a valid observation and the remaining
days with the same value are set to missing.
The log-likelihood function given in (23) can be
maximized as a three dimensional maximization pro-
blem over O, U and F. After estimates for these three
variables are made, the DTD can be computed from
equation (22).
However, with quarterly nancial statements there
will never be more than three changes in the corporate
structure (dened in this model by L and A
t
) throughout
the year, leading to possibly unstable estimates of F.
This problem is mitigated by performing a two stage
optimization for O, U and F.
In the rst stage, the optimization for each rm is
performed over all three variables. For each rm, in
the rst month in which DTD can be computed the
optimization is unconstrained in O and U, while F is
constrained to being in the unit interval [0,1]. There-
after, at month n, the optimization is still unconstrain-
ed in O and U while F is constrained to the interval
[max (0,
n1
0.05), min (1,
n1
+ 0.05)], where
n1

is the estimate of F made in the previous month. In
other words, a ten percent band around the previous
estimate of F (where that band is oored with 0 and
capped with 1) is applied so that the estimates do not
uctuate too much from month to month.
It was found that this was not enough to obtain
stable estimates of F. For many rms, the estimate of
F would frequently lie on the boundary of the constrain-
ing interval. To impose greater stability, a second stage
is added. At each month end, the average estimate for F
in all nancial sector rms in the economy is used for
every nancial sector rm in the economy, meaning the
optimization is only over O and U. The same is done
for non-nancial rms. In fact, the optimization can be
reduced to only be over U by using the sample mean
of the log returns of the implied asset values in place
of O.
Since the rst stage is done to obtain a stable, sector
average estimate of F, the criteria used to include a
rm-month is more strict. In the rst stage, a two year
window is used instead of one year, and a minimum
of 250 days of valid observations of the market capi-
talization are required instead of 50. If a rm has less
than 250 days of valid observations within the last two
years of a particular month end, F will not be estimated
for that rm and that month end.
In summary, the DTD for each rm is computed using
the economy and sector (nancial or non-nancial)
average for F in that month, and the estimate of O and
U based on the last year of data for the rm.
Carrying out this two stage procedure would take
several months of computation time on a single PC,
given the millions of rm months that are required.
However, each of the stages is parallelizable. In the rst
stage the DTD can be computed independently between
rms. In the second stage, once the sector averages of
the F have been computed for each month, the DTD
can again be computed independently between rms.
In the CRI system, a grid of several hundred computers
administered by the NUS Computer Center is used.
With this, the DTD computation can be performed for
all rms over the full history of twenty years in less
than two days.
3.3 Calibration
Implementation: As shown in Section 1, the calibra-
tion of the forward intensity model involves multiple
maximum pseudo-likelihood estimations, where the
pseudo-likelihood functions are given in equation
(18). The maximizations are of the logarithm of these
expressions, and are performed independently between
the default parameters and the exit parameters, and
between parameters for different horizons. In the nota-
tion of Section 2, the vectors of parameters D(0),,
D(23) and D

(0),, D

(23) are independently estimated.

A few input variables have an unambiguous effect
on a rms probability of default. Increasing values
of both the level and trend of DTD, CASH/TA, and
82 GLOBAL CREDIT REVIEW
NI/TA all indicate that a rm is becoming more credit
worthy and should lead to a decreased PD. For large
and relatively clean datasets such as the US, an uncon-
strained optimization leads to parameter values which
largely have the expected sign. For each of DTD level
and trend, CASH/TA level and trend, and NI/TA level,
the default parameters at all horizons are negative. A
negative default parameter at a horizon means that if
the variable increases, the forward intensity will de-
crease (by equation (6)), so that the conditional default
probability at that horizon will decrease. The one
exception is the NI/TA trend variable. Since the current
implementation in the CRI system does not include net
income values from quarterly statements, the trending
effect is weak.
For some of the smaller economies and economies
with lower quality datasets, an unconstrained optimi-
zation leads to the default parameters for some of
these variables to be positive at several horizons. This
leads to counter-intuitive results. For example, if the
default parameters for CASH/TA are positive, a rm
that increases its cash reserves, all other factors being
equal, will have a PD that increases. To prevent such
situations, the CRI system performs a constrained
optimization with only non-positive values allowed for
the default parameters associated with the level and
trend of DTD, CASH/TA, and NI/TA.
For this, the Matlab function fmincon from the
Optimization Toolbox is used. The analytic gradient
and Hessian are not supplied and the algorithm used by
fmincon is the active-set optimization. If fmincon
fails to converge, fminsearch is used. This uses a
simplex search method which takes more time but is
generally more likely to converge.
Each evaluation of the pseudo-log-likelihood func-
tion can be done in a fraction of second on a standard
CPU, even for the largest economies. But since the
optimization is over 13 dimensions, thousands of eva-
luations are required. It is therefore important to make
each function evaluation as fast as possible.
Notice that at each time point and at any horizon,
there are far more surviving rms than exiting rms.
Thus, from equation (16) and (18), it can be seen
that the most time-consuming part of evaluating the
pseudo-log-likelihood function is the term for the
surviving rms. Evaluating the forward intensity
function of equation (7) can be formulated as a matrix-
vector multiplication, where the rows of the matrix
are the different surviving rms variables, and the
vector is the vector of parameters. The matrix will
typically have several hundreds of thousands of rows
and does not change during the optimization (though
it will change for different optimizations at different
horizons). This type of problem is well-suited for a
programmable graphics processing unit (GPU). The
CRI system runs the calibrations on an NVIDIA
Tesla C2050 card. For each economy, the calibrations
for the default and other exit parameters for horizons up
to 24 months typically require ten minutes or less.
Grouping for small economies: There are not enough
defaults in some small economies and calibrations of
these individual economies are not statistically mean-
ingful. In order to ensure that there are enough defaults
for calibration, the 30 economies are categorized into
groups according to similarities in their stage of deve-
lopment and their geographic locations. Within these
groups the economies are combined and calibrated
together. The calibration group for each economy is
listed in Table A.1.
Once the economies are combined into groups, the
minimum amount of alteration to the input variables
is done. Firms will still use the economy-specic
variables for the stock index return and short-term
interest rate as indicated in Tables A.2 and A.3, and
the economy-specic interest rate for the DTD computa-
tion as indicated in Table A.4. For the second stage
of the DTD computation, described in Subsection 3.2,
the sector average estimates for F are still economy-
specic rather than common to the group. For the
SIZE variable, the median market capitalization is the
economy-specic median. For replacement of missing
data, the sector medians are those for the economy. The
only exception is for winsorization. So that outliers
can be treated consistently, the 0.1 and 99.9 percentile
levels (or more aggressive levels) are determined based
on all economies within the group.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 83
If the parameters are restricted to being equal in
value between all economies in a group, the predic-
tive power of the model can be relatively poor for
individual economies. It was found that some of the
parameter estimates gave statistically signicant dif-
ferences between economies if the values were allowed
to differ. Different estimated parameter values indicate
the difference the effect of a variable has on different
economies. For example, on estimation it is found that
the short-term interest rate has little signicance in
default prediction for Indonesian rms at short time
horizons, but a signicant effect on the rest of the
Original ASEAN ex-SG group consisting of Malaysia,
Philippines and Thailand.
The reason for combining economies into groups is
so that the aggregate default experience can be used
to estimate the parameters. Different economies are
therefore allowed to have different parameter values
for a variable only on a selective basis. There are a
large number of combinations of sub-groups and vari-
ables, and extensive, on-going experiments are being
performed on different combinations. Each sub-group
must contain at least 15 default events and variables are
allowed to differ in the sub-group only if the difference
in parameter estimation is statistically signicant and
signicantly improves the outcome of the pseudo-log-
likelihood function.
Currently the sub-groups and variables that are
allowed to differ are as follows. In the Original ASEAN
ex-SG group, Indonesia has different parameters than
the other three economies for its intercept, stock index
return, short-term interest rate, CASH/TA level and
SIZE level. In the SG & HK group, Singapore and
Hong Kong have different parameters for their inter-
cept and SIGMA parameters. In the North America
group, the US and Canada have a different parameter
for their intercept only. In the Western Europe 1 group,
the Netherlands and Belgium form a sub-group. Their
parameter for their short-term interest rates is different
from that of the other ve economies in the group. In
the Western Europe 2 group, Germanys parameter for
its short-term interest rate differs from that of the other
seven economies in the group.
3.4 Daily Output
Individual rms PD: In computing the pseudo-
log-likelihood functions in equation (18), only end of
month data is needed. The data needs to be extended to
daily values in order to produce daily PDs.
For the level variables, the last twelve end of month
observations (before averaging) are combined with the
current value. The current value is scaled by a fraction
equal to the current day of the month divided by the
number of calendar days in the month. The earliest
monthly value is scaled by one minus this fraction. The
sum is then divided by the number of valid monthly
observations, with the current value and the earliest
monthly value counting as a single observation if either
or both are not missing. Not performing this scaling
can lead to an articial jump in PD at the beginning of
the month. When performing the scaling, the change in
level is more gradual throughout the month.
A similar procedure is done for SIGMA. Here the ear-
liest month is not scaled, but the return from the current
day to the previous month end is scaled by the square
root of the fraction equal to the current day of the month
divided by the number of calendar days in the month.
Computing the DTD for all rms on a daily basis
using the two stage process described in Subsection
3.3 would be time consuming, even on the grid. Since
there should be little change to O, U and F on a day
to day basis, for the daily computation of DTD these
are assumed to have the same value as in the previous
months DTD calculation. In other words, the previous
months values for U and F together with the new
days equity value are used in equation (20) to obtain
the implied asset value. This implied asset value with
the previous months values for O, U and F is used in
equation (22) to obtain the new days DTD.
Aggregating PD: The CRI provides term structures
of the probability distributions for the number of
defaults as well as the expected number of defaults for
different groups of rms. The companies are grouped
by economy (using each rms country of domicile),
by sector (using the rms Level I Bloomberg Industry
Classication) and sectors within economies. With the
84 GLOBAL CREDIT REVIEW
individual rms PD, the expected number of defaults
is trivial to compute. The algorithm used to compute
the probability distribution of the number of defaults
was originally reported in Anderson, Sidenius and
Basu (2003). It assumes conditional independence and
uses a fast recursive scheme to compute the necessary
probability distribution.
Note that while this algorithm is currently used to
produce the probability distribution of the number of
defaults within an economy or sector, it can easily be
generalized to compute loss distributions for a portfolio
manager, where the exposure of the portfolio to each
rm needs to be input.
Inclusion of rms in aggregation: As explained
in Subsection 3.1, rms are included in an economy
for calibration if the rms primary listing is on an
exchange in that economy. This is to ensure that all
rms in an economy are subject to the same disclosure
and accounting requirements. In contrast, a rm is
included in an economys aggregate results if the rm
is domiciled in that economy. This is because users
typically associate rms with their economy of domicile
rather than the economy where their primary listing is,
if they are different. For example, the Bank of China
has its primary listing in Hong Kong, but its economy
of domicile is China so the Bank of China is included
in the aggregate PDs for China, and is included under
China when searching for the individual PDs.
IV. EMPIRICAL ANALYSIS
This section presents an empirical analysis of the CRI
outputs for the thirty economies that are currently being
covered. In Subsection 4.1, an overview is given of the
default parameter estimates. Subsection 4.2 explains
the tests that are performed on the PDs, including accu-
racy ratios, the Spiegelhalter test statistic and the trafc
light test. The results of the tests are also discussed.
4.1 Parameter Estimates
With 24 months of forecast horizons, two sets of
parameters (default and other exits) and 12 different
groups of economies, tables of the parameter estimates
occupy almost 50 pages and are not included in this
Technical Report. They are available in an Annex to this
report that is available via the CRI web portal. In the
Annex, the parameter estimates are from calibrations
performed in June 2011 using data up until the end
of May 2011. As an example, plots of the default
parameters for the US are given in Figures B.1 and B.2,
along with the 90% condence level. In this subsection,
a brief overview is given of the general traits and
patterns seen in the default parameter estimations of
the economies covered by the CRI.
Recall that if a default parameter for a variable at
a particular horizon is estimated to be positive (resp.
negative) from maximizing the pseudo-likelihood
function, then an increasing value in the associated
variable will lead to an increasing (decreasing) value
of the forward intensity at that horizon, which in
turn means an increasing (decreasing) value for the
conditional default probability at that horizon.
For the stock index one-year trailing return variable,
most groups have default parameters that are slightly
negative in the shorter horizons and then become posi-
tive in the longer horizons. When the equity market
performs well, this is only a short-term positive for
rms and in the longer-term, rms are actually more
likely to default. This seemingly counter-intuitive result
could be due to correlation between the market index
and other rm-specic variables. For example, Dufe
et al. (2009) suggested that a rms distance-to-default
(DTD) can overstate its credit-worthiness after a strong
bull market. If this is the case, then the stock index
return serves as a correction to the DTD levels at these
points in time.
The default parameters for the short-term interest
rate variable are signicantly positive at one to two-
year horizons for most of the groups. This is consistent
with an increase in short-term interest rates signaling
increased funding costs for companies in the future,
increasing the probability of default. The values at
shorter horizons are varied between economies from
slightly negative to signicantly positive, possibly
indicating different lead-lag relationships between
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 85
credit conditions and the raising and cutting of short-
term interest rates.
On the other hand, a sub-group of Western Europe
1 group, consisting of France, Greece, Italy, Portugal
and Spain, has signicantly negative default parame-
ters associated with the short-term interest rate. This
is consistent with an effect that competes with the
funding costs argument: central banks typically raise
interest rates to relieve inationary pressures during
expansionary periods. Thus, a high level of the short
term interest rate is signaling an expansionary period
where rms have a lower probability of default.
DTD is a measure of the volatility-adjusted leverage
of a rm. Low or negative DTD indicates high leverage
and high DTD indicates low leverage. Therefore, PD
would be expected to increase with decreasing DTD.
This is conrmed by the estimates, where almost all
of the calibrations for the different groups lead to
negative default parameters for the DTD Level, with
only Chinas default parameter estimations hitting the
constraint at zero for longer horizons.
The ratio of the sum of cash and short term invest-
ments to total assets (CASH/TA) measures liquidity
of a rm. This indicates whether a company can meet
its short-term obligations such as interest and princi-
pal payments. As expected, for almost all economies
(Indonesia being the only exception) the default para-
meters for CASH/TA Level in shorter horizons are
signicantly negative. The magnitude of the default
parameters decreases for longer horizons, conrming
that CASH/TA Level is a better indicator of a rms
ability to make payments in the short term than the
long term.
The ratio of net income to total assets (NI/TA) mea-
sures protability of a rm. The relationship between
PD and NI/TA is as expected: the default parameters
for NI/TA Level are signicantly negative for most
economies and most horizons.
The logarithm of the market capitalization of a
rm over the median market capitalization of rms
within the economy (SIZE) does not have a consistent
effect on PD across different economies. For example,
in the US the default parameters for SIZE Level are
negative for shorter horizons and positive for longer
horizons, suggesting that the advantages enjoyed by
larger rms, such as diversied business lines and
funding sources, are a benet in the shorter term but
not in the longer term. On the other hand, in Japan
the default parameters for SIZE Level are negative
across all horizons. These differences may reect differ-
ences in the business environments in the respective
economies.
The default parameters associated with DTD Trend,
CASH/TA Trend and SIZE Trend, are negative across
almost all economies and horizons. The trend variables
reect momentum. The momentum effect is a short
term effect, and evidence of this is seen in the lower
magnitude of the default parameters at longer horizons
than at shorter horizons. The remaining trend variable
is the NI/TA Trend. The current implementation of the
CRI system retrieves net income only from annual
nancial statements. The default parameters for NI/
TA Trend are constrained to be negative, but for most
economies there is no clear relationship between
the NI/TA Trend and the horizon. Once NI/TA from
quarterly statements can be used, this will likely be
more informative.
The ratio of the sum of market capitalization and
total liabilities to total assets (M/B) can either indicate
the market mis-valuation effect or the future growth
effect. This default parameter is positive in most eco-
nomies, indicating that higher M/B implies higher
PD, and the market mis-valuation effect dominates.
The Western Europe 2 group, consisting of Austria,
Denmark, Finland, Germany, Iceland, Norway, Sweden
and Switzerland, is an exception. For these economies,
the associated default parameters are weakly negative
at short time horizons, indicating that the future growth
effect dominates.
Shumway (2001) argued that a high level of the idio-
syncratic volatility (SIGMA) indicates highly variable
stock returns relative to the market index, indicating
highly variable cash ows. Volatile cash ows suggest
a heightened PD, and this nding is consistent across
all economies and most horizons, with the exception
of India.
86 GLOBAL CREDIT REVIEW
4.2 Prediction Accuracy
In-sample and out-of-sample testing: Various tests
are carried out to test the prediction accuracy of the
CRI PDs. These tests are conducted either in-sample
or out-of-sample.
A single calibration is conducted for the in-sample
tests, using data to the end of May 2011. As an example,
one-year PDs are computed for Dec 31, 2000 by using
the data at or before Dec 31, 2000 and the parameters
from the calibration. These PDs can be compared to
actual defaults that occurred at any time in 2001.
The out-of-sample analysis is done over time. The
rst calibration is conducted using only data up to
the end of December 2000. For example, one-year
PDs can be made for Dec 31, 2000 using the data at
or before Dec 31, 2000 with the parameters from this
rst calibration. These are PDs that could have been
computed at the time, since the parameters are not
based on data available after that date. This process is
repeated every month. That is, the second calibration
is conducted using only data up to the end of January
2001, and so on.
It should be noted that for these repeated calibra-
tions based on an expanding window of data, nothing
else is changed besides the dataset. In other words, the
same choice of input variables and the same choice of
economy dummies within the groups are used through-
out all of the calibrations.
Some of the calibration groups have too few defaults
in the period before December 2000 to be able to pro-
duce stable calibration results (see Table A.8). If this
is the case, the start date is advanced. Subsequently,
if there are too few defaults after the start date to
perform meaningful tests, only in-sample tests are
performed for that calibration group. Out-of-sample
tests are performed for (starting month of calibration
in parentheses): China (12/2000), Japan (12/2003),
India (12/2001), South Korea (12/2000), ASEAN ex-
SG group (12/2000), North America group (12/2000),
and Western Europe 2 group (12/2002).
Accuracy Ratio: The accuracy ratio (AR) is one of
the most popular and meaningful tests of the discrimina-
tory power of a rating system (BCBS, 2005). The AR
and the equivalent area under the Receiver Operating
Characteristic (AUROC) are described in Duan and
Shrestha (2011). In short, if defaulting rms had
been assigned among the highest PD of all rms
before they defaulted, then the model has discrimi-
nated well between safe and distressed rms.
This leads to higher values of AR and AUROC. The
range of possible AR values is in [0, 1], where 0 is
a completely random rating system and 1 is a perfect
rating system. The range of possible AUROC values
is in [0.5, 1]. AUROC and AR values are related by:
AR = 2 AUROC 1.
Table B.1 lists AR and AUROC values for the one-
month, six-month, one-year and two-year horizons,
with standard errors in parentheses. In Table B.1, both
in-sample and out-of-sample results are available for
calibration groups where out-of-sample testing could
be performed. Other calibration groups include only
in-sample results. The in-sample AR and AUROC
are computed only from the starting date of the corre-
sponding out-of-sample tests, so that the results between
in-sample and out-of-sample are comparable. Only
economies with more than 20 defaults entering into
the AR and AUROC computation are listed. The PDs
are taken to be non-overlapping. For example, the one-
year AR is based on PDs computed on 31/12/2000,
31/12/2001, , 31/12/2010 and rms defaulting within
one year of those dates, while the two-year AR is based
on PDs computed on 31/12/2000, 31/12/2002, ,
31/12/2008 and rms defaulting within two years of
those dates.
The AUROC values have been provided only for the
purpose of comparison with rating systems that report
their results in terms of AUROC. In this report, the
discussion will focus only on AR. The model is able to
achieve strong AR results mostly greater than 0.80 at the
one and six-month horizons for developed economies.
There is a drop in AR at one and two-year horizons, but
the AR are still mostly acceptable. Australia, the UK
and Singapore have sharp drops in AR to below 0.60
at the two-year horizon. Hong Kong has worse than
expected AR over all horizons. With Hong Kong and
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 87
Singapore, the low AR may be helped if combined with
Taiwan to achieve a more stable calibration.
The AR in emerging market economies such as
China, India, Indonesia, Malaysia, Philippine and
Thailand are noticeably weaker than the results in the
developed economies. This can be due to a number of
issues. The quality of data is worse in emerging markets,
in terms of availability and data errors. This may be
due to lower reporting and auditing standards. Also,
variable selection is likely to play a more important
role in emerging markets. The variables were selected
based on the predictive power in a developed economy,
the US. Performing variable selections specic to the
calibration group are expected to improve predictive
accuracy, especially in emerging market economies.
Finally, there could be structural differences in how
defaults and bankruptcies occur in emerging market
economies. If the judicial system is weak and there are
no repercussions for default, rms may be less reluctant
to default.
At horizons of one and six-months, out-of-sample
AR are comparable to their in-sample counterparts.
At horizons of one and two-years, out-of-sample AR
can be substantially lower than the in-sample AR. This
issue is more easily examined with the tests presented
later.
Finally, the US has a sufcient number of nancial
rms and nancial defaults to produce separate AR and
AUROC. These are also listed in Table B.1 as out-of-
sample results. The nancial sector ARs are actually
stronger than the non-nancial sector AR. This is
achieved by having only minimal differences between
how nancial and non-nancial rms are treated.
The AR is a test of discriminatory power, or how
well the rating system ranks rms in terms of credit
worthiness. The following plots and tests are direct
test of the PD values computed from the rating system.
Each will be explained and a discussion of the results
will follow.
Aggregate defaults: The time series of aggregate
predicted number of defaults and actual number of
defaults in each calibration group is plotted in each of
the rst rows of graphs in Figures B.3 to B.15. The left
graphs are for a horizon of one-month and the right
graphs are for a horizon of one-year. For each month,
the predicted number of defaults computed from the
aggregate PD is plotted in the red line, and the actual
number of defaults in the following one month (or one
year) is plotted as the blue bar.
Spiegelhalter test statistic: The Brier score at time
t and horizon is a mean square error (MSE) between
the model predicted PD and the realized defaults:
.
Here, N
t
is the number of valid -horizon PDs
observed at time t, PD
it
( ) is the model generated
PD for rm i at time t for horizon . Also, y
it
( ) is
the default indictor: if the i
th
obligor defaults before
t + , then y
it
( ) = 1; otherwise y
it
( ) = 0. The MSE
statistic is small if the model generated PDs assigned
to rms that eventually default is high and the PDs
assigned to rms that do not default is low. In general,
a low MSE indicates a good rating system. However,
the best that a modeler can hope to do is to correctly
estimate probabilities of default, rather than to know
defaults with certainty, since defaulting is a random
event (Rauhmeier and Scheule, 2005).
The Spiegelhalter test has a null hypothesis that
the model has exactly predicted the actual default
probabilities, and test statistic formulated as:
.
One can compute:
,
and
.
Following Spiegelhalter (1986), under the assumption
of independence among PDs and using the central limit
88 GLOBAL CREDIT REVIEW
theorem, it can be shown that under the null hypothesis
the test statistic Z
t
( ) approximately follows a standard
normal distribution whose condence interval can be
computed routinely.
The time series of the Spiegelhalter test statistic for
each calibration group is plotted in each of the second
rows of graphs in Figures B.3 to B.15. The left graphs are
for a horizon of one-month and the right graphs are for
a horizon of one-year. The solid line is the test statistic
computed at a monthly frequency, and the dashed lines
are the 90% and 95% condence intervals.
Trafc light test: Coppens, Gonzlez and Winkler
(2007) describe a test that is of particular relevance to
regulators and parctitioners. The basic idea is to classify
rms as investment grade by their model generated
PD, observe how many of these rms end up default-
ing and assess whether the actual number of defaults
exceed certain condence levels. The condence levels
are set at 80% and 99%. If the 80% condence level is
exceeded without exceeding the 99% level, the rating
system is in the amber zone, which is allowed once
every ve years. Any exceedance of the 99% level
places the rating system in the red zone.
The trafc light test results are plotted in the third
rows of Figures B.3 to B.15. Following Coppens et al.
(2007), the threshold for one-year PD is set at 10bps.
The upper bound of the green zone is the green line,
and the lower bound of the red zone is the red line. The
condence levels are computed based on the binomial
distribution where all of the PDs are assumed to be
10bps. For some calibration groups, if there are too few
rms with a one-year PD less than 10bps, the green line
(and even the red line) will remain constant at zero. The
threshold is increased by steps of 10bps until the green
line is substantially different from zero.
Discussion of tests: Figure B.3 and B.4 are in-
sample and out-of-sample results for the North America
group. Comparing the actual versus predicted number
of defaults, two points are notable. One is that for the
early period in 2001 to mid-2002, the out-of sample
results are very poor. One possible reason is that the
datasets for these calibration dates cover a period main-
ly consisting of only an economic expansion. NBER
(2010) deems March 1991 to have been the trough of
the business cycle, while the peak was in March 2001.
Having default experience from only half a cycle would
not allow the model to calibrate for different parts of
the cycle. This would have played a part in variable
selection at that time.
The other notable point is that the out-of-sample
predicted number of defaults during the 2008-2009
nancial crisis is much higher than the corresponding
in-sample predictions. This could be due to the fact that
because of the extraordinary government intervention at
the time, the predicted defaults never occurred. As time
has passed and the data window expands to include the
period where nancial ratios are still distressed but the
default rates are relatively low, more recent calibrations
from the in-sample results have modulated the predicted
number of defaults in that period.
It is noteworthy that although out-of-sample ARs for
the North America group are acceptable, the predicted
numbers of defaults are far off in several periods. This
is due to the fact that AR primarily measures how well a
rating system ranks different rms looking at PD values
relative to each other, and ignores the magnitude of PD
values. The plots of the out-of-sample tests for the other
calibration groups show a similar difference between
the out-of-sample and in-sample results. The remainder
of this subsection will focus on the in-sample results.
The predicted numbers of defaults are slow to capture
the actual numbers of defaults in many of the economies.
A possible remedy to this is to incorporate different
common factors that are better leading indicators of
economic distress, such as realized market volatility.
Related to this lag between actual defaults and pre-
dicted defaults are the rejections of the null hypothesis
in the Spiegelhalter test. For the economies that have
weak AR, the difference between the predicted defaults
and actual defaults is apparent.
The trafc light test is passed by all calibration
groups. The red zone is never breached and there is
never more than one violation into the amber zone in
a period of ve years.
NUS-RMI CREDIT RATING INITIATIVE TECHNICAL REPORT 89
V. PLANNED DEVELOPMENTS
The credit rating system in the CRI is meant to evolve
and improve as additional features and elements are
developed. Besides modications to the current model-
ing framework of the forward intensity model, a change
in modeling platform will be undertaken if another
model proves more promising in terms of accuracy and
robustness of results. In this version of the Technical
Report, no platform changes are anticipated. This sec-
tion describes the developments that are planned within
the current modeling platform.
The existing modeling framework of the forward
intensity model has proven to be robust and exible
while providing consistent term structure information
on default probabilities that is difcult to obtain in
different models. At the same time, much development
remains to be done to improve our existing results.
Areas for future development are listed below.
Default denitions: As described in Subsection 2.4,
a major challenge in the implementation of a credit
assessment system is in dening and classifying de-
faults. Different jurisdictions treat and report defaults
differently, and additional work needs to be done on
normalizing to a common denition of default.
Besides dening default, the search for default events
is on-going. Defaults that occur on a current basis need
to be captured. As well, some economies seem to have
fewer defaults than would be expected so the historical
default event search is also an on-going effort.
Variable selection: Up to this point, the CRI system
has used the same input variables across all economies.
This has been done in order to expand the coverage
to a global scale as quickly as possible. Restricting
to a common denition of input variables removes a
bottleneck in expanding to a different geographical
region.
With only a few remaining major economies to be
covered (e.g. Brazil and Russia), the issue of variable
selection, including both rm-specic variables and
common factors, can be investigated. It is expected that
removing the restriction of having commonly dened
input variables will improve the prediction accuracy,
especially for emerging market economies.
In addition to new variables, the existing input
variables can be ne-tuned. For example, the net income
is currently taken only from annual nancial reports.
Once investigation is done on seasonal effects of using
quarterly net income, the protability measure will be
updated in a timelier manner. Also, the idiosyncratic
volatility measure uses monthly returns, and does
not react quickly to sudden changes in idiosyncratic
volatility. Experimentation is being conducted on using
daily returns to compute this measure.
Mergers and acquisitions: Whenever there is a
major merger or acquisition, the PD for the remaining
company cannot be calculated until a new nancial
statement is released, since the merger or acquisition
is reected in the market capitalization but not in the
nancial statement variables. Preliminary investigations
on simply combining balance sheets are promising.
Global coverage: The next region that coverage will
extend to is Latin America. After which, Europe will
be completed with coverage of Eastern Europe. Africa
and the Middle East will remain after that, and in a
nal step, coverage will be extended to economies that
have been missed along the way.
The RMI Credit Rating initiative is premised on the
concept of credit ratings as a public good. Being a
non-prot undertaking allows a high level of transpar-
ency and collaboration that other commercial credit
rating systems cannot replicate. The research and sup-
port infrastructure is in place and researchers from
around the world are invited to work on our database.
Any methodological improvements that researchers
develop will be incorporated into the CRI system. In
essence, the initiative operates as a selective Wikipedia
where many can contribute but implementation control
is retained.
If you have feedback on this Technical Report or
wish to work with us in this endeavor, please contact
us at rmicri@globalcreditreview.com.
90 GLOBAL CREDIT REVIEW
References
Anderson, L., J. Sidenius and S. Basu, (2003), All your
Hedges in one Basket, Risk, pp. 6772.
Basel Committee on Banking Supervision (BCBS), (2005),
Studies on the Validation of Internal Rating Systems,
Working Paper 14.
Basel Committee on Banking Supervision (BCBS), (2006),
International Convergence of Capital Measurement and
Capital Standards A Revised Framework.
Coppens, F., F. Gonzlez and G. Winkler, (2007), The
Performance of Credit Rating Systems in the Assessment
of Collateral used in Eurosystem Monetary Policy Oper-
ations, European Central Bank.
Crosbie, P. and J. Bohn, (2003), Modeling Default Risk,
Moodys KMV.
Duan, J.-C., (1994), Maximum Likelihood Estimation Using
Price Data of the Derivative Contract, Mathematical
Finance, 4, pp. 155167.
Duan, J.-C., (2000), Correction: Maximum Likelihood
Estimation Using Price Data of the Derivative Contract,
Mathematical Finance, pp. 461462.
Duan, J.-C., (2010), Clustered Defaults, Singapore: RMI,
National University of Singapore.
Duan, J.-C., J. Sun and T. Wang, (2011), Multiperiod
Corporate Default Prediction A Forward Intensity
Approach, Singapore: RMI, National University of
Singapore.
Duan, J.-C. and K. Shrestha, (2011), Statistical Credit Rating
Methods, Global Credit Review, pp. 4364.
Dufe, D., A. Eckner, G. Horel and L. Saita, (2009), Frailty
Correlated Default, Journal of Finance, pp. 20892123.
Dufe, D., L. Saita, and K. Wang, (2007), Multi-Period
Corporate Default Prediction with Stochastic Covariates,
Journal of Financial Economics, 83, pp. 635665.
Fitch Ratings, (2011), Denitions of Ratings and Other
Forms of Opinions.
Merton, R.C., (1974), On the pricing of corporate debt: The
Risk Structure of Interest Rates, Journal of Finance, 29,
pp. 449470.
Monetary Authority of Singapore, (2004), IRB Approach:
Denition of Default.
Moodys Investors Service, (2011), Corporate Default and
Recovery Rates, 19202010.
National Bureau of Economic Research, (2010), US
Business Cycle Expansions and Contractions.
Rauheimer, R. and H. Scheule, (2005), Rating properties
and their implications for Basel II capital, Asia Risk, 5,
pp. 5154.
Spiegelhalter, D., (1986), Probabilistic prediction in patient
management and clinical trials, Statistics in Medicine,
5, pp. 421433.
Standard & Poors, (2011), 2010 Annual Global Corporate
Default Study and Rating Transitions.