Professional Documents
Culture Documents
1
t
prmleg(0) := S
0
DV01(0)
DV01(0) :=
b
X
iD1
i
D(0. T
i
)(1
C
T
i
)
where D(s. t ) is the discount factor (often assumed to be deterministic) between
times s and t and
i
= T
i
T
i1
is the year fraction. In the second equation, DV01(0)
denotes the sumover all premiumpayment dates of the outstanding notional, weighted
by the relevant year fraction, and discounted back at time 0. The risk-neutral expec-
tation of such a quantity will provide the sensitivity of the premium leg to a unit
movement in the spread S
0
, all other things being equal (hence the name DV01,
although usually the spread is moved by 1 basis point (bp)). In the same equation,
the actual outstanding notional in each period would be an average over T
i1
. T
i
|,
but we have replaced it with 1
C
T
i
(the value of the outstanding notional at T
i
) for
simplicity, which is commonly done.
Note that, in contrast with the tranches (see Section 2.2), here the recovery is not
considered when computing the outstanding notional because it is only the number
of defaults that matters.
The market quotes the values of S
0
that, for different maturities, balance the two
legs. If we have a model for the loss and the number of defaults, we may require that
the loss and the number of defaults in the model, when plugged into the two legs, lead
to the same risk-neutral expectation (and hence price):
S
0
=
E
0
R
T
0
D(0. t ) d
1
t
|
E
0
P
b
iD1
i
D(0. T
i
)(1
C
T
i
)|
(1)
where E denotes the expectation under the pricing risk-neutral measure.
Assuming deterministic default-free interest rates, we can rewrite:
S
0
=
R
T
0
D(0. t ) dE
0
1
t
|
P
b
iD1
i
D(0. T
i
)(1 E
0
C
T
i
|)
(2)
The assumption regarding deterministic default-free interest rates is a practical one
and allows us to obtain analytical or semi-analytical pricing formulas, as is usually
Research Paper www.thejournalofcreditrisk.com
46 D. Brigo et al
done in market practice. It can be relaxed to an assumption of independence between
defaults and default-free interest rates, with the consequence that the D(t. T ) terms of
the deterministic case become 1(t. T ) = E
t
D(t. T )|, the zero-coupon bond prices.
2.2 Collateralized debt obligation tranches
Synthetic CDOs with maturity T are contracts involving a protection buyer, a protec-
tion seller and an underlying pool of names. They are obtained by putting together a
collection of CDSs with the same maturity on different names, 1. 2. . . . . M, typically
M = 125, each with notional 1,M, then tranching the loss of the resulting pool
between points and T, with 0 6 < T 6 1:
1
t
:=
1
T
(
1
t
)1
fA<
N
L
t
6Bg
(T )1
f
N
L
t
>Bg
|
A useful alternative expression is:
1
t
:=
1
T
T
1
0;B
t
1
0;A
t
| (3)
Once enough names have defaulted and the loss has reached , the count starts.
Each time the loss increases, the corresponding loss change rescaled by the tranche
thickness T is paid to the protection buyer until maturity arrives or until the total
pool loss exceeds T, in which case the payments stop.
The discounted default leg payout can then be written as:
deeg
A;B
(0) :=
Z
T
0
D(0. t ) d
1
t
As usual, in exchange for the protection payments, a premium rate S
A;B
0
, xed at
time T
0
= 0, is paid periodically at times T
1
. T
2
. . . . . T
b
= T . Part of the premium
can be paid at time T
0
= 0 as an upfront U
A;B
0
. The rate is paid on the surviving
average tranche notional. If we further assume that payments are made on the notional
remaining at each payment date T
i
, rather than on the average in T
i1
. T
i
|, the
premium leg can be written as:
prmleg
A;B
(0) := U
A;B
0
S
A;B
0
DV01
A;B
(0)
DV01
A;B
(0) :=
b
X
iD1
i
D(0. T
i
)(1
1
T
i
)
When pricing CDO tranches, we are interested in the premium rate S
A;B
0
that sets
to zero the risk-neutral price of the tranche. The tranche value is computed taking the
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 47
TABLE 1 Standardized attachment and detachment for the DJiTraxx Europe main and
CDX.NA.IG tranches.
DJiTraxx
Europe CDX.NA.IG
main (%) (%)
03 03
36 37
69 710
912 1015
1222 1530
22100 30100
(risk-neutral) expectation (in t = 0) of the discounted payout, which is the difference
between the default and premium legs above. We obtain:
S
A;B
0
=
E
0
R
T
0
D(0. t ) d
1
t
| U
A;B
0
E
0
P
b
iD1
i
D(0. T
i
)(1
1
T
i
)|
(4)
Assuming deterministic default-free interest rates, we can rewrite:
S
A;B
0
=
R
T
0
D(0. t ) dE
0
1
t
| U
A;B
0
P
b
iD1
i
D(0. T
i
)(1 E
0
1
T
i
|)
(5)
The above expression can be easily recast in terms of the upfront premium U
A;B
0
for
tranches that are quoted in terms of upfront fees.
The tranches that are quoted on the market refer to standardized pools, standardized
attachmentdetachment points and T and standardized maturities.
The standardized attachment and detachment differ slightly for the CDX.NA.IG
and the DJiTraxx Europe main tranches.
2
For the DJiTraxx and CDX pools, the equity tranche ( = 0. T = 3%) is quoted
by means of the fair U
A;B
0
, while assuming S
A;B
0
= 500bps. The reason that the equity
tranche is quoted upfront is to reduce the counterparty credit risk that the protection
seller is facing. All other tranches are generally quoted by means of the fair running
spread S
A;B
0
, assuming no upfront fee (U
A;B
0
= 0). Following recent market turmoil,
the 36% and the 37% tranches have been quoted in terms of an upfront amount and
a running S
A;B
0
= 500bps given the exceptional risk premium priced by the market
for this tranche.
2
The attachment points of the CDX tranches are slightly higher, reecting the average higher
perceived riskiness (as measured, for example, by the CDS spread or by balance-sheet ratios) of the
liquid investment grade North American names.
Research Paper www.thejournalofcreditrisk.com
48 D. Brigo et al
3 A FULLY CONSISTENT DYNAMICAL MODEL:
THE GENERALIZED-POISSON LOSS MODEL
The GPLmodel can be formulated as follows. Consider a probability space supporting
a number n of independent Poisson processes 7
1
. . . . . 7
n
with time-varying and
possibly stochastic intensities z
1
. . . . . z
n
under the risk-neutral measure Q. The risk-
neutral expectation conditional on the market information up to time t , including the
pool loss evolution up to t , is denoted by E
t
. Intensities, if stochastic, are assumed to
be adapted to such information.
Dene the stochastic process:
7
t
:=
n
X
jD1
j
7
j
(t ) (6)
for positive integers
1
. . . . .
n
. In the following, we refer to the 7
t
process simply as
the GPL process. We will use this process as a driving process for the cumulated port-
folio loss
1
t
, which is the relevant quantity for our payouts. In Brigo et al (2006a,b),
the possible use of the GPL process as a driving tool for the default counting process
C
t
is illustrated instead.
The characteristic function of the 7
t
process is:
Z
t
(u) = E
0
e
iuZ
t
| = E
0
E
0
e
iuZ
t
[
1
(t ). . . . .
n
(t )||
where:
j
(t ) :=
Z
t
0
z
j
(s) ds. i = 1. . . . . n
are the cumulated intensities of each Poisson process. Now, we substitute 7
t
,
obtaining:
Z
t
(u) = E
0
n
Y
jD1
E
0
e
iu
j
Z
j
.t/
[
1
(t ). . . . .
n
(t )|
= E
0
n
Y
jD1
Z
j
.t/j
j
.t/
(u
j
)
Z
t
(u) = E
0
exp
n
X
jD1
j
(t )(e
iu
j
1)
(7)
The marginal distribution
Z
t
of the process 7
t
can be directly computed at any
time via an inverse Fourier transformation of the characteristic function of the process.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 49
The characteristic function
Z
t
(u) can be explicitly calculated for some relevant
choices of Poisson cumulated-intensity distributions (see, for example, Brigo et al
(2006a,b)).
3.1 Loss dynamics
The underlying GPL process 7
t
is non-decreasing and, given sufciently large times,
it takes arbitrarily large values. The portfolio cumulated loss and the rescaled number
of default processes is non-decreasing, but limited to the interval 0. 1|. Thus, we con-
sider the deterministic non-decreasing function v : N L {0] 0. 1| and we dene
the cumulated portfolio loss process
1
t
as:
1
t
:= v
L
(7
t
) := min(7
t
. M
0
) and
1
t
:=
1
t
M
0
(8)
where 1,M
0
, M
0
> M > 0, is the minimum jump for the loss process. M
0
is clearly
related to the loss granularity.
Remark 3.1 Note that the loss is bounded within the interval 0. 1| by construction,
but that the possibility remains that the loss jumps more than M times, where M is
the number of names in the portfolio. If this is the case, we may check a posteriori that
the probability of such events is negligible. This is the case for all of our examples.
The marginal distribution of the cumulated portfolio loss process 1
t
can be easily
calculated. We obtain:
1
t
= min(7
t
. M
0
) = 7
t
1
fZ
t
<M
0
g
M
0
1
fZ
t
>M
0
g
Since 7
t
has a known distribution, the distribution of 1
t
can be easily derived as a
by-product. The related density (dened on integer values since the law is discrete)
is:
L
t
(.) =
Z
t
(.)1
fx<M
0
g
Q{7
t
> M
0
]1
fxDM
0
g
The density of
1
t
follows directly. Also, the intensity of 1
t
, ie, the density of the
absolutely continuous compensator of 1
t
(see, for example, Giesecke et al (2009)),
can be computed directly and is given by:
h
L
(t ) =
n
X
jD1
min(
j
. (M
0
7
t
)
C
)z
j
(t ) (9)
(see Brigoet al (2006a,b) for details). The intensityhgoes tozerowhen7exceeds M
0
,
which corresponds to total loss, as expected. Furthermore, if all the possible integer
jump sizes between 1 and M
0
are allowed, ie, if
j
= and n = M
0
, the intensity
h
L
jumps whenever the cumulated portfolio loss process 1 jumps. The intensity
Research Paper www.thejournalofcreditrisk.com
50 D. Brigo et al
jumps downward, and this would seem to go in the opposite direction with respect
to self-excitedness, which is considered a desirable feature of loss models in general.
However, self-exciting features are embedded in our model, and embedded in the
possibility of having several defaults in small intervals, contrary to most approaches
to loss modeling. Consider, for example, just two names: instead of having the loss
of one name increase the likelihood of default (intensity) of a second name, we have
both names defaulting together immediately. This embeds self-excitedness, although
in an extreme way. Finally, to view the effect of single names on each other, we need
a more sophisticated formulation based on the common Poisson shocks framework,
leading to the GPCL model, which is analyzed in Brigo et al (2007) and is presented
below with a few updates.
Example 3.2 It is worth presenting an example to further clarify the GPL for-
mulation. Let us take an example considering a GPL model where there are three
independent Poisson processes 7
1
, 7
2
and 7
3
correspondingly multiplied by sizes
1, 10 and 40, leading to:
7
t
= 17
1
(t ) 107
2
(t ) 407
3
(t )
with constant deterministic intensities z
1
= 4%, z
2
= 1.5% and z
3
= 0.75%.
Assume a 40%deterministic recovery, and take one scenario where the three processes
have eachjumpedonce before the current time. The loss of the standardizedDJiTraxx
pool of 125 names associated to this scenario with one jump of 7
1
, 7
2
and 7
3
,
respectively, would be 0.48%, 4.8% and 19.2%, respectively. The expected loss in
one year for the uncapped loss process is:
0.2352% = (0.0048 0.04) (0.048 0.015) (0.192 0.0075)
In Figure 1 on the facing page we plot the resulting default counting distribution on
four increasing maturities: 3, 5, 7 and 10 years. We note that the modes on the right
tail of the distribution become more evident as the maturity increases. These bumps
are the corresponding probabilities of a jump of a higher amplitude occurring.
From the distribution we also note that the probability is not simply assigned to
the higher-amplitude jumps (10 and 40 defaults in our example) but also to a number
of defaults immediately above: this would be the probability of having jumps of both
a high amplitude (either 10 or 40) and of a lower amplitude (1 in our case), where
jumps with lower amplitudes have higher probability.
Furthermore, we note that some second-order modes start to appear as maturity
increases. These are the bumps corresponding to multiple jumps in the higher-size
(10 and 40), smaller probability and GPL components.
Let us now assume we are to price two tranches: 24.8% and 1019.2%. The
expected tranche loss of these tranches, and, ultimately, their fair spread, will depend
primarily on the probability mass lying above the tranche detachment.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 51
FIGURE 1 Default counting M
C
T
probability distribution associated to a GPL model
featuring three independent Poisson processes with jump amplitudes
1
= 1,
2
= 10 and
3
= 40, with constant deterministic intensities z
1
= 4%, z
2
= 1.5% and z
3
= 0.75%.
10
8
6
4
2
0
0 10 20 30 40 50 60
3 year
%
(a)
10
8
6
4
2
0
0 10 20 30 40 50 60
%
5 year
(b)
10
8
6
4
2
0
0 10 20 30 40 50 60
%
7 year
(c)
Research Paper www.thejournalofcreditrisk.com
52 D. Brigo et al
FIGURE 1 Continued.
10
8
6
4
2
0
0 10 20 30 40 50 60
%
10 year
(d)
3.2 Model limits
The GPL model that we have introduced can currently be viewed as a particularly
simple parametrization of the market-implied loss-distribution dynamics. A positive
feature is that the loss changes only by positive jumps, which should be the case
in any sensible loss model. Furthermore, this choice allows us to achieve a good
calibration to market data, as we will see in the following section. However, we
are not making explicit assumptions on two important issues: rstly, we have not
addressed possible ways of making our model consistent with single-name dynamics,
and secondly, we have not explained how to choose a full-featured pool spread and
recovery dynamics. One possibility would be to make the intensities in the Poisson
processes driving 7 stochastic and to consider more general transformations of 7 to
obtain the loss process.
Since we are focusing only on the calibration of CDO tranches, which depend
only on the loss marginal distribution, we will avoid discussing such problems here
(we direct the interested reader to Brigo et al (2006a,b) for an extensive analysis
of candidate spread and recovery dynamics, and to Brigo et al (2007) for further
discussion, including consistency with single-name data). We point out that signicant
progress and testing in loss modeling will only be possible when more liquid market
quotes for tranche options and forward-start tranches are available.
3.3 Model calibration
We work with the basic GPL model specication as given by the driving GPL process
7 in (6), which we use to model the pool loss through (8). In this basic formulation,
each Poisson mode 7
j
has a deterministic piecewise-constant intensity z
j
(t ).
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 53
Given that we have modeled the pool loss
1
t
directly, we do not completely charac-
terize the rescaled default counting process
C
t
, but instead give only its expectations.
Indeed, suchexpectations are the onlyinformationondefault countingthat are implicit
in index market quotes (Equation (1)), whereas tranche quotes (Equation (4)) depend
only on the loss and not on default counting explicitly. We therefore assume:
E
0
C
t
| :=
1
1 R
E
0
1
t
|. 0 6 R < 1 E
0
1
T
b
| (10)
where the range of denition of the constant Ris taken in order to ensure that, at each
time t , the expected value of the rescaled number of defaults is greater than or equal
to the cumulated portfolio loss, and that both are smaller than or equal to one. Note
that we avoid introducing an explicit dynamics for the recovery rate (see Brigo et al
(2006a,b, 2007) for an initial discussion on recovery dynamics). Our R here can be
interpreted as a kind of average recovery rate.
3.4 Detailed calibration procedure
The model parameters found using the calibration procedure are the amplitudes
j
{m N: m 6 M
0
], = 1. . . . . n, and the cumulated intensities
j
(T ), which are
real non-decreasing piecewise linear functions in the tranche maturity.
The optimal values for the amplitudes are selected in the following way.
1) Fix the minimum jump size to 1,M
0
by choosing the integer M
0
> M > 0.
2) Find the integer value for
1
by calibrating the cumulated intensity
1
for each
value of
1
in the range 1. M
0
|, all other modes being set to zero. Calibration is
obtainedbyminimizingthe objective function(11) below. Finally, choose the
1
for which the calibration error is at its minimum when using the corresponding
1
. We call this chosen
1
the best integer value for
1
.
3) Add the amplitude
2
and nd its best integer value by calibrating the cumulated
intensities
1
and
2
, starting from the previous value for
1
as a guess, for
each value of
2
in the range 1. M
0
|.
4) Repeat the previous step for
i
with i = 3 and so on by calibrating the cumu-
lated intensities
1
. . . . .
i
, starting from the previously found
1
. . . . .
i1
as an initial guess, until the calibration error is under a given threshold or until
the intensity
i
can be considered negligible.
5) Check a posteriori that the probability of having more than M jumps is negli-
gible and that the value of R is within the arbitrage-free range given in (10).
Research Paper www.thejournalofcreditrisk.com
54 D. Brigo et al
TABLE 2 DJiTraxx index and tranche quotes in basis points on May 13, 2005, with
bidask spreads in parentheses.
Maturities
Attachment 3 5 7 10
detachment years years years years
Index 38 54 65 77
(4) (1) (3) (2)
Tranche
03 2,060 4,262 5,421 6,489
(100) (118) (384) (124)
36 72 173 398 590
(10) (68) (40) (20)
69 28 57 141 188
(6) (6) (17) (15)
912 13 31 72 87
(2) (5) (20) (15)
1222 3 21 42 60
(1) (3) (13) (10)
Index and tranches are quoted through the periodic premium, while the equity tranche is quoted as an upfront
premium (see Section 2).
The objective function to be minimized in the calibration is the squared sum of
the errors shown by the model to recover the tranche and the index market quotes
weighted by market bidask spreads:
(. ) =
X
i
c
2
i
. c
i
=
.
i
(. ) .
mid
i
.
bid
i
.
ask
i
(11)
where the .
i
(with i running over the market quote set) are the index values S
0
for DJiTraxx index quotes and either the periodic premiums S
A;B
0
or the upfront
premium rates U
A;B
for the DJiTraxx tranche quotes.
3.5 Calibration results
The GPL model is calibrated to the market quotes observed weekly from May 6,
2005 to October 18, 2005. Following Albanese et al (2006), we take R = 30% as
our reference value for the recovery rate in the DJiTraxx Europe market for spot
and forward contracts. The quality of our calibration below is not altered if we select
a value R = 40% resembling the recovery typically used in simplied quoting
mechanisms in the market (see Brigo et al (2006a,b, 2007) for examples of this). We
start with M
0
= 200, corresponding to a minimum loss jump size of 50bps.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 55
TABLE 3 (a) Calibration error c
i
in (11) (calculated with respect to the bidask spread)
for tranches quoted on May 13, 2005; (b) cumulated intensities (integrated up to tranche
maturities) of the GPL model with M
0
= 200.
(a)
Maturities
Attachment 3 5 7 10
detachment years years years years
Index 0.0 0.1 0.3 0.0
Tranche
03 0.0 0.1 0.2 0.2
36 0.0 0.0 0.2 0.0
69 0.0 0.0 0.3 0.1
912 0.1 0.1 0.1 0.4
1222 0.0 0.0 0.2 0.3
(b)
.T /
3 5 7 10
years years years years
1 1.955 3.726 4.464 7.694
3 0.000 0.062 0.305 0.305
8 0.016 0.033 0.109 0.109
12 0.004 0.013 0.026 0.026
19 0.006 0.006 0.017 0.017
72 0.000 0.009 0.026 0.049
185 0.000 0.002 0.002 0.008
Each row corresponds to a different Poisson component with jump amplitude . Recovery rate is 30%.
As a rst example, consider the calibration date May 13, 2005 (see Table 2 on the
facing page). In Table 3 we list the calibration result and the values of the calibrated
parameters. The calibration errors c are very low for all maturities. Note that a cal-
ibration error smaller than one means that the difference between the market quote
and the model price is smaller than the bidask spread.
Consider, as a second example, the calibration date October 11, 2005 (see Table 4
on the next page). InTable 5 on page 57, we list the calibration results and the values of
the calibrated parameters. The calibration errors show that the 10-year equity tranche
Research Paper www.thejournalofcreditrisk.com
56 D. Brigo et al
TABLE 4 DJiTraxx index and tranche quotes in basis points on October 11, 2005, with
bidask spreads in parentheses.
Maturities
Attachment 3 5 7 10
detachment years years years years
Index 23 38 47 58
(2) (1) (1) (1)
Tranche
03 762 3,137 4,862 5,862
(26) (26) (76) (74)
36 20 95 200 515
(10) (1) (3) (10)
69 7 28 43 100
(6) (1) (2) (4)
912 12 27 54
(2) (4) (5)
1222 7 13 23
(1) (2) (3)
is not correctly priced. We nd such mispricing in many calibration examples, in
particular after October 2005.
For the minimum loss jump size 1,M
0
, besides 50bps, we try the values 2bps and
10bps, corresponding, respectively, to M
0
equal to 5,000 and 1,000. As can be seen
from Table 6 on the facing page, the 10-year maturity tranches (in our experience the
most difcult to calibrate) are stable through the three different loss sizes, suggesting
that going below 50bps does not add much exibility to the model. This is conrmed
by further tests. In particular, the difference between the M
0
= 1.000 calibration
and the M
0
= 5.000 calibration is always small. Furthermore, the behavior of the
mean calibration error, ie, of the mean of the absolute values of the c
i
across time and
quotes, for the three different choices of M
0
is quite similar and within one bidask
spread.
We also note that as the minimum jump size decreases (granularity increases),
the loss distribution becomes noisier, due to the presence of small amplitudes. Fur-
thermore, very small modes, appearing when the minimum jump size is as small as
a few basis points, may violate the requirement that the loss process jumps fewer
than M times (see Remark 3.1). We also try calibrations with M
0
less than 200, ie,
with a minimum loss jump greater than 50bps. In this case the calibration error grows
quickly. Indeed, the minimumjump size, in this case, becomes greater than the typical
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 57
TABLE 5 (a) Calibration error c
i
in (11) (calculated with respect to the bidask spread) for
tranches quoted on October 11, 2005; (b) cumulated intensities (integrated up to tranche
maturities) of the GPL model with M
0
= 200.
(a)
Maturities
Attachment 3 5 7 10
detachment years years years years
Index 0.0 0.0 0.1 0.1
Tranche
03 0.1 0.1 1.2 2.1
36 0.1 0.1 0.3 1.0
69 0.0 0.1 0.3 0.9
912 0.4 0.8 0.8
1222 0.0 0.0 0.0
(b)
.T /
3 5 7 10
years years years years
1 0.441 2.498 4.466 7.555
2 0.435 0.435 0.435 0.671
11 0.004 0.023 0.023 0.023
22 0.000 0.001 0.006 0.030
29 0.000 0.000 0.001 0.001
32 0.001 0.004 0.004 0.004
192 0.000 0.001 0.005 0.011
The three-year maturity quotes lack two tranches.
TABLE 6 GPL calibration error for different minimum loss sizes 1,M
0
with respect to the
bidask spread for 10-year tranches on October 11, 2005.
Attachment
detachment 50bps 10bps 2bps
03 2.1 1.8 1.8
36 1.0 1.0 1.0
69 0.9 0.9 0.9
912 0.8 0.9 0.8
1222 0.0 0.2 0.0
Recovery rate is 30%.
Research Paper www.thejournalofcreditrisk.com
58 D. Brigo et al
TABLE 7 Values of the Poisson amplitudes ,M
0
for different values of the minimum loss
jump 1,M
0
.
(a) 50bps
Poisson amplitudes
1 2 3 4 5 6 7
Date (%) (%) (%) (%) (%) (%) (%)
May 6, 2005 0.50 1.50 4.00 6.00 9.50 39.50 92.50
September 2, 2005 0.50 1.00 4.00 5.50 12.50 39.00 100.00
October 11, 2005 0.50 1.00 5.50 11.00 14.50 16.00 96.00
(b) 10bps
Poisson amplitudes
1 2 3 4 5 6 7
Date (%) (%) (%) (%) (%) (%) (%)
May 6, 2005 0.10 1.50 4.60 5.90 9.60 39.60 53.00
August 5, 2005 0.20 1.10 1.40 8.10 11.30 49.00 62.40
October 11, 2005 0.10 0.70 1.00 6.30 11.50 14.50 93.70
(c) 2bps
Poisson amplitudes
1 2 3 4 5 6 7
Date (%) (%) (%) (%) (%) (%) (%)
May 6, 2005 0.02 1.50 5.26 9.64 17.58 39.64 99.78
August 12, 2005 0.38 1.06 1.14 7.38 12.24 41.34 99.80
October 3, 2005 0.02 0.98 1.16 7.52 9.74 43.34 65.16
October 11, 2005 0.16 0.68 1.00 6.30 10.98 14.46 94.90
Only the calibration dates between May 6, 2005 and October 18, 2005 where the =M
0
values change are listed.
portfolio loss given when one name defaults. 50bps seems, then, to be a reasonable
reference value.
Also, the values of the Poisson amplitudes are quite stable across the calibration
dates. Indeed, in six months we observe at most four changes in their values, as shown
in Table 7.
The loss distribution implied by the GPL model is multimodal and the probabil-
ity mass moves toward larger loss values as the maturity increases. These features
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 59
FIGURE 2 Loss-distribution evolution of the GPL model with a minimum jump size of
50bps at all the quoted maturities up to 10 years, drawn as a continuous line.
0 0.02 0.04 0.06 0.08 0.10
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
Loss
0.10 0.15 0.20 0.25 0.30
0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
x 10
3
Loss
(a)
3 years
5 years
7 years
10 years
(b)
3 years
5 years
7 years
10 years
are shared by different approaches. For instance, static models, such as the implied
default-rate distribution in Torresetti et al (2006c), suggest multimodal loss distribu-
tions, as mentioned in the introduction regarding the implied copula. The evolution
of the implied loss distribution is shown in Figure 2.
The dynamic credit correlation model of Albanese et al (2006) shows implied
loss distributions whose modes tend to group as the maturity increases, leading to a
Research Paper www.thejournalofcreditrisk.com
60 D. Brigo et al
distribution approaching normality. The GPL model reproduces this behavior (see,
for example, Brigo et al (2006a,b)).
4 APPLICATIONSTO MORE RECENT DATA ANDTHE CRISIS
In this section we check whether the critical features that we discussed regarding
implied correlation and the subsequent elements coming frommore advanced models
are still present in-crisis after mid 2007. We will observe that the features are still
present and are often amplied in the market after the beginning of the crisis.
4.1 Customizing the generalized Poisson loss model
to deal with sectors
We now consider the GPL model to assess how well it performs in-crisis. We change
the model formulation slightly in order to have a model that is more in line with the
current market, while maintaining all the essential features of the modeling approach.
Compared with the model described in Section 3, we introduce the following modi-
cations. We x the jump amplitudes a priori rather than calibrating themthrough the
detailed calibration procedure seen in Section 3.4, and we associate different recov-
eries to different . Fixing the before calibration will make the model less exible
but quicker to calibrate and possibly more stable. To x the a priori, we reason as
follows. Fix the independent Poisson jump amplitudes to the levels just above each
tranche detachment, when considering a 40% recovery.
For the DJiTraxx index, for example, this would be realized through jump ampli-
tudes a
i
=
i
,125, where:
5
= roundup
125 0.03
1 1
.
6
= roundup
125 0.06
1 1
7
= roundup
125 0.09
1 1
.
8
= roundup
125 0.12
1 1
9
= roundup
125 0.22
1 1
.
10
= 125
and, in order to have more granularity, we add the sizes 1, 2, 3 and 4:
1
= 1.
2
= 2.
3
= 3.
4
= 4
In total we have n = 10 jump amplitudes. We then modify the obtained sizes slightly
in order to account for CDX attachments that are slightly different. Eventually, we
obtain the set of amplitudes:
125 a
i
i
{1. 2. 3. 4. 7. 13. 19. 25. 46. 125]
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 61
We want to associate a recovery of 40% to all of these amplitudes , except
125
, to
which we want to associate a recovery of 0% in order to size the premium quoted
by the market for super-senior tranches. We start by considering the default counting
process which, by introducing the non-Armageddon indicator 1
t
, can be cast in the
following form:
C
t
= 1 1
t
(1 c
t
) (12)
with:
c
t
:= min
n1
X
iD1
a
i
7
i
(t ). 1
. 1
t
:= 1
fZ
n
.t/D0g
We refer to the component associated with
n
=
10
= 125 as the Armageddon
mode, since whenever 7
n
= 7
10
jumps, the whole pool defaults. We refer to the
jump 7
n
= 7
10
as an Armageddon event. Indeed, whenever theArmageddon com-
ponent 7
n
jumps for the rst time, the default counting process
C
t
jumps to the entire
pool size and every name in the pool has defaulted, with no more defaults being
possible.
We then introduce the stopping time t as the minimumtime between the Armaged-
don jump event and the time when the reduced pool without the Armageddon compo-
nent has completely defaulted. This is also the time when the full pool has defaulted:
t = inf
t :
n
X
iD1
a
i
7
i
(t ) > 1
We note that 1 is 0 after the Armageddon event, and otherwise it is equal to 1, while
c is equal to 1 after the reduced pool is over, so, after some algebra, we obtain:
C
t
= 1
fO 6tg
c
t
1
fO >tg
(13)
Note that, as expected,
C depends on c only at terminal time t .
To derive the loss process with a zero recovery for the Armageddon event, we rst
calculate the evolution of the counting process:
d
C
t
= 1
t
d c
t
d1
t
(1 c
t
)
Here, notation such as d
C
t
denotes the left increment d
C
t
=
C
t
C
t
, since we are
considering right-continuous jump processes. Intuitively, this quantity is zero if
C
does not jump at t , whereas it is one if
C jumps at t .
We then apply the recovery only to the terms coming from jumps that are not due
to the Armageddon event:
d
1
t
= (1 1)1
t
d c
t
d1
t
(1 c
t
)
Research Paper www.thejournalofcreditrisk.com
62 D. Brigo et al
We cannowintegrate the previous equationtoobtainthe loss process. Tomanipulate
the above equation we observe three things. Firstly, 1 is zero after the Armageddon
event, otherwise it is equal to one. Secondly, d1 is non-zero only at the Armageddon
event. Finally, c is equal to one and d c is zero after the reduced pool is over:
1
t
= (1 1 c
O
)1
fO 6tg
(1 1) c
t
1
fO >tg
(14)
Note that
1 now depends on c both at full-pool exhaustion time t and at terminal
time t .
Remark 4.1 (Interpretation of the model) Whenever the Armageddon component
7
n
jumps the rst time, we will assume that the recovery rate associated to the
remaining names defaulting in that instant will be zero. The pool loss, however, will
not always jump to one, since there is the possibility that one or more names already
defaulted before the Armageddon component 7
n
jumped, and that they defaulted
with recovery rate 1. If, at a given instant t , the whole pool defaults, ie,
C
t
= 1, this
may happen in one of two ways.
1) 7
n
jumped by t . In this case, the portfolio has been wiped out with the help of
an Armageddon event. Note that, in this case, d
C
t
= d
1
t
if 7
n
jumps at t . In
fact, the recovery associated to the pool fraction defaulting in that instant will
be equal to zero.
2) 7
n
has not jumped by t . In this case the portfolio has been wiped out, not with
the help of anArmageddon event, but because of defaults of more small or large
sectors that do not comprise the whole pool. Note that, in this case, the loss is
less than the whole notional, as all these defaults had recovery 1 > 0.
In this way, whenever 7
n
jumps at a time when the pool has not yet been wiped
out, we can rest assured that the pool loss will be above 1 1. We do this because
the market in 2008 quoted CDOs with prices assuming that the super-senior tranche
would be impacted to a level that was impossible to reach with recoveries xed at
40%. For example, there was a market for the DJiTraxx ve-year 60100% tranche
on March 25, 2008 quoting a running (bid) spread of 24bps bid.
In the dynamic loss model, recovery can be made a function of the default rate
C
(or other solutions are possible; see Brigo et al (2006b) for more discussion). Here
we use the above simple methodology to allow losses of the pool to penetrate beyond
1 1 and hence severely affect even the most senior tranches, in line with market
quotations.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 63
4.2 Numerical calculation of loss distribution
We know how to calculate the distribution of both
C
t
and
1
t
. The distribution of the
counting process may be directly calculated from Equation (12) as the distribution of
a reduced GPL, ie, a GPL where the jump 7
n
is excluded, whose counting process
will be c
t
.
The distribution of the loss process can be calculated starting from Equation (14).
A simple approach involves explicitly calculating the forward Kolmogorov equation
satised by the probability distribution:
N
L
t
(.) = Q{
1
t
= .].
d
dt
N
L
t
(.) =
X
y
t
(.. ,)
N
L
t
(,) (15)
where . and , are generic states of the loss process and where the generator matrix
t
is given by:
t
(.. ,) := lim
t!0
Q{
1
tCt
= . [
1
t
= ,] 1
fxDyg
^t
This procedure is allowed since, at each point in time, starting only from the value
of the loss process, it is possible to say whether the Armageddon event happened, so
we are able to reduce the loss process, which depends on the stopping time t, to a
Markov process. In order to do so, we consider the following states for the pool loss
process divided by M:
1
t
AL B
where Ais the set of states where the Armageddon event has not happened, and B is
the set of states where the Armageddon event has happened. The two sets are dened
as follows:
A :=
0. (1 1)
1
M
. (1 1)
2
M
. . . . . (1 1)
M 1
M
. 1 1
B :=
(1 1)
M 1
M
1
M
. . . . . (1 1)
1
M
M 1
M
. 1
Note that the two sets are disjoint if 1 > 0, so if we know the value of
1 only at time
t , we can state whether the Armageddon event has happened.
Example 4.2 Consider a pool of six names with the following GPL modes with
constant deterministic default intensities:
7
t
= 7
1
(t ) 37
2
(t ) 67
(t )
z
1
= 0.020. z
3
= 0.010. z
= 0.003
Research Paper www.thejournalofcreditrisk.com
64 D. Brigo et al
TABLE 8 Generator matrix for a simple GPL process of three amplitudes for a pool of six
names with 40% recovery for each mode, and the Armageddon mode, whose recovery is
zero.
y
x 0 10 20 30 40 50 60 100
0 3.30 0.00 0.00 0.00 0.00 0.00 0.00 0.00
10 1.00 3.30 0.00 0.00 0.00 0.00 0.00 0.00
20 0.00 1.00 3.30 0.00 0.00 0.00 0.00 0.00
30 2.00 0.00 1.00 3.30 0.00 0.00 0.00 0.00
40 0.00 2.00 0.00 1.00 3.30 0.00 0.00 0.00
50 0.00 0.00 2.00 0.00 1.00 3.30 0.00 0.00
60 0.00 0.00 0.00 2.00 2.00 3.00 0.00 0.00
67 0.00 0.00 0.00 0.00 0.00 0.30 0.00 0.00
73 0.00 0.00 0.00 0.00 0.30 0.00 0.00 0.00
80 0.00 0.00 0.00 0.30 0.00 0.00 0.00 0.00
87 0.00 0.00 0.30 0.00 0.00 0.00 0.00 0.00
93 0.00 0.30 0.00 0.00 0.00 0.00 0.00 0.00
100 0.30 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Starting states (y) on columns, arrival states (x) on rows (see Equation (15)). See text for intensity and amplitude
data. All values in the table are in percent.
where we have called 7
N
L
t
(.) = exp
Z
t
0
X
y
u
(.. ,) du
N
L
0
(,) = exp
t
X
y
(.. ,)
N
L
0
(,)
where the last step is due to the fact that, in this case, the generator matrix does
not depend on time. Matrix exponentiation can be quickly computed with the Pad
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 65
FIGURE 3 Loss distribution evolution up to 10 years predicted by the GPL model for a
pool of six names.
0
20
40
60
80
100
2
4
6
8
10
0
Years to maturity
Attachment points (%)
L
o
s
s
(
%
)
10
5
0
40% recovery for each mode except the Armageddon mode, whose recovery is zero. See text for intensity and
amplitude data.
approximation (Golub andVan Loan (1983)), leading to a closed-formsolution for the
probability distribution
N
L
t
(.). This distribution can then be used in the calibration
procedure. In Table 9 on the next page and in Figure 3, we show the loss distribution
for our example.
4.3 Calibration results
Bearing this interpretation of the modes in mind, we decided to select the GPL ampli-
tudes by choosing the independent Poisson jump amplitudes to be at the level just
above each tranche detachment when using a 40% recovery. This led to the specic
values of
5
. . . . .
10
that we introduced earlier in the paper. With an eye on the vari-
ety of shapes of the implied copula historical calibrated distributions, as summarized
in Brigo et al (2010) and in Torresetti et al (2006c), we added four amplitudes (from
one to four), corresponding to a small number of defaults.
Research Paper www.thejournalofcreditrisk.com
66 D. Brigo et al
TABLE 9 Loss distribution numerical values for 5-year and 10-year maturity dates, pre-
dicted by the GPL model for a pool of six names.
0 5 10
years years years
(%) (%) (%)
0% 100.00 84.79 71.89
10% 0.00 4.24 7.19
20% 0.00 0.11 0.36
30% 0.00 8.48 14.39
40% 0.00 0.42 1.44
50% 0.00 0.01 0.07
60% 0.00 0.46 1.72
67% 0.00 0.00 0.00
73% 0.00 0.00 0.02
80% 0.00 0.07 0.25
87% 0.00 0.00 0.00
93% 0.00 0.04 0.12
100% 0.00 1.38 2.55
40% recovery for each mode, except the Armageddon mode whose recovery is zero.
We now present the goodness of t of this GPL through history. We measure the
goodness of t by calculating, for each date, the relative mispricing:
MisprRel
A;B
=
8
<
:
S
A;B;theor
0
S
A;B;bid
0
S
A;B;mid
0
if S
A;B;theor
0
< S
A;B;bid
0
S
A;B;theor
0
S
A;B;ask
0
S
A;B;mid
0
if S
A;B;theor
0
> S
A;B;ask
0
0 otherwise
where S
A;B;theor
0
is the tranche theoretical spread as in Equation (4), where, for the
calculation of the expectations of both numerator and denominator, we take the loss
distribution resulting from the calibrated GPL.
In Figure 4 on the facing page and Figure 5 on page 68 we present the relative
mispricing for all tranches, for all maturities and for both indexes throughout the
sample (March 2005 to June 2009).
We note that, while the 5-year tranches could be repriced fairly well in the cur-
rent credit crisis for both DJiTraxx and CDX, the 10-year tranche calibrations for
both indexes have sensibly been made less precise following the collapse of Lehman
Brothers. Recently, with the stabilization of credit markets we again see fairly precise
calibration results (ie, a relative mispricing in the 24% range) for all tranches and
for both indexes and both maturities.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 67
FIGURE 4 Relative mispricing resulting from the GPL calibration: DJiTraxx.
20
15
10
5
0
5
10
15
20
03/2005 01/2006 11/2006 09/2007 08/2008 06/2009
03%
36%
69%
912%
1222%
22100%
Index
%
(a)
20
15
10
5
0
5
10
15
20
03/2005 01/2006 11/2006 09/2007 08/2008 06/2009
%
03%
36%
69%
912%
1222%
22100%
Index
(b)
To highlight where the problems in calibration come from, we have grouped mis-
pricings across three different categories.
Instrument: we have grouped all tranches, independent of maturity and seniority, into
one group and we have put the remaining calibrated instruments, ie, the 5-year and
10-year indexes, in the other group.
Maturity: we have grouped all tranches and indexes in two groups according to their
maturity (5 and 10 years).
Research Paper www.thejournalofcreditrisk.com
68 D. Brigo et al
FIGURE 5 Relative mispricing resulting from the GPL calibration: CDX.
20
15
10
5
0
5
10
15
20
%
03/2005 01/2006 11/2006 09/2007 08/2008 06/2009
03%
36%
69%
912%
1222%
22100%
Index
(a)
20
15
10
0
5
10
15
20
%
03%
36%
69%
912%
1222%
22100%
Index
5
01/2006 11/2006 09/2007 08/2008 06/2009 03/2005
(b)
Seniority: we have grouped all tranches (leaving out the indexes) into three categories
depending on the seniority of the tranche in the capital structure.
1) Equity: equity tranche for both the DJiTraxx and CDX indexes.
2) Mezzanine: comprising the two most junior tranches after the equity tranche.
For the DJiTraxx index this means the 36% and 69% tranches, whereas
for the CDX this means the 37% and 710% tranches.
3) Senior: comprising the remaining most senior tranches. For the DJiTraxx
this means the 912%, 1222%and 22100%tranches, whereas for the CDX
this means the 1015%, 1530% and 30100% tranches.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 69
FromFigure 6 on the next page and Figure 7 on page 71 we note that the GPLmodel
produced calibrations that resulted in a relative mispricing that was larger (though of
relatively small magnitude) in the period from June 2005 to October 2006. For both
the DJiTraxx and the CDX pools the mispricing could be ascribed to the 10-year
tranches, in particular to the equity and mezzanine tranches.
We also note that from October 2008 to June 2009 (ie, after the default of Lehman
Brothers), the GPL calibration again resulted in a non-zero but still fairly contained
relative mispricing that, in this case, could be ascribed to both the 5-year and 10-year
tranches independent of their positions in the capital structure (equity, mezzanine or
senior).
5 ADDING SINGLE-NAME AND CLUSTER DYNAMICS:
THE GENERALIZED-POISSON CLUSTER LOSS MODEL
The GPL model introduces a simple mechanism to avoid repeated defaults by cap-
ping the default counting process to the number of names in the pool. However, this
approach prevents us from extending the model to incorporate single names, since
we are limiting the number of defaults irrespective of which name is defaulting.
In Brigo et al (2007), to overcome this issue we introduced the GPCL model,
starting from the standard CPS framework with repeated defaults and using it as an
engine to build a new model for (correlated) single-name defaults, clusters defaults
as well as for the default counting process and the portfolio loss process. Two possible
strategies (aside from the GPL capping function) were introduced to deal with single
names.
We now review the GPCL strategy based on an adjustment at the single-name level
to avoid repeated defaults, leading to (correlated) single-name default processes. This
approach leads to a less clear cluster dynamics in terms of the original cluster repeated
default processes, but it shows interesting properties at the single-name level. We refer
the reader to Brigo et al (2007) for the strategy based directly on cluster dynamics.
5.1 Common Poisson shock basic framework
In the CPS framework we consider the observation that a single names default can
be originated by different events or factors. Occurrence of the event/factor number
e is modeled as a jump of independent Poisson processes N
e
, e = 1. . . . . m. Each
event can be triggered many times r = 1. 2. . . . as jumps go on. The rth jump of N
e
triggers a default event for name k with probability
e
r;k
. Note that a defaulted name
k may default again. We address this limitation later.
The CPS framework may accommodate systemic and idiosyncratic factors in a
natural way. Consider, for instance, the following two cases:
Research Paper www.thejournalofcreditrisk.com
70 D. Brigo et al
FIGURE 6 DJiTraxx relative mispricing resulting from the GPL calibration grouped by
(a) maturity (5 and 10 years), (b) instrument type (index and tranches) and (c) seniority
(equity, mezzanine and senior).
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
10 year all
5 year all
(a)
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
Tranche
Index
%
%
(b)
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
Mezzanine
Equity
%
14
(c)
Senior
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 71
FIGURE 7 CDX relative mispricing resulting from the GPL calibration grouped by (a)
maturity (5 and 10 years), (b) instrument type (index and tranches) and (c) seniority (equity,
mezzanine and senior)
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
10 year all
5 year all
(a)
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
Tranche
Index
%
%
(b)
12
10
8
6
4
2
0
03/2005 10/2005 06/2006 02/2007 10/2007 06/2008 02/2009
Mezzanine
Equity
%
14
(c)
Senior
Research Paper www.thejournalofcreditrisk.com
72 D. Brigo et al
1) e is a totally systemic factor if
e
r;1
=
e
r;2
= =
e
r;n
= 1;
2) e is a totally idiosyncratic factor if there exists a k such that, for all = k, we
obtain
e
r;k
= 1 and
e
r;j
= 0.
We dene the dynamics for the single-name default process N
k
, jumping each time
name k defaults, as:
N
k
(t ) :=
m
X
eD1
N
e
.t/
X
rD1
1
e
r;k
where 1
e
r;k
is a Bernoulli variable with probability Q{1
e
r;k
= 1] =
e
r;k
. Note that the
process N
k
itself turns out to be Poisson. Furthermore, by differentiating the above
relationship, we obtain:
dN
k
(t ) =
m
X
eD1
1
e
1CN
e
.t/;k
dN
e
(t )
so that, if we consider two names k and h, we have that the respective default counting
processes N
k
and N
h
are not independent, since their dynamics is explained by the
same driving events N
e
(t ).
The core of the CPSframeworkinvolves mappingthe single-name default dynamics
(which consist of the dependent Poisson processes N
k
) into a multiname dynamics
explained in terms of independent Poisson processes
N
s
, where s is a subset (or
cluster) of names of the pool, dened as follows:
N
s
(t ) =
m
X
eD1
N
e
.t/
X
rD1
X
s
0
s
(1)
js
0
jjsj
Y
k
0
2s
0
1
e
r;k
0
where [s[ is the number of names in the cluster s. In a summation, s k means
that we are adding up across all clusters s containing k, k s means that we are
adding across all elements k of cluster s, [s[ = means that we are adding across all
clusters of size , and, nally, s
0
_ s means that we are adding up across all clusters
s
0
containing cluster s as a subset.
The proof of the independence of
N
s
for different subsets s canbe foundinLindskog
and McNeil (2003). Note that a jump in an
N
s
process means that all the names in
the subset s, and only those names, have defaulted at the jump time. We denote by
z
s
the intensity of the Poisson process
N
s
(t ), and we assume it to be deterministic for
the time being. We refer to Brigo et al (2007) for model extensions.
5.2 Cluster processes as common Poisson shock building blocks
We do not need to remember the above construction. All that matters for the following
developments are the independent clusters default Poisson processes
N
s
(t ). These
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 73
can be taken as fundamental variables from which (correlated) single-name defaults
and default counting processes follow. The single-name dynamics can be derived
based on these independent
N
s
processes in the so-called fatal shock representation
of the CPS framework:
N
k
(t ) =
X
s3k
N
s
(t ) or dN
k
(t ) =
X
s3k
d
N
s
(t ) (16)
where the second equation is the same as the rst one but in instantaneous jump
form. We now introduce the process 7
j
(t ), which describes the occurrence of the
simultaneous default of any names whenever it jumps (with jump size 1):
7
j
(t ) :=
X
jsjDj
N
s
(t ) (17)
Note that each 7
j
(t ), being the sum of independent Poisson processes, is itself Pois-
son. Furthermore, since the clusters corresponding to the different 7
1
. 7
2
. . . . . 7
M
do not intersect, the 7
j
(t ) are independent Poisson processes.
The multiname dynamics, that is, the default counting process 7
t
for the whole
pool, can be easily derived by carefully adding up all the single-name contributions:
7
t
:=
M
X
kD1
N
k
(t ) =
M
X
kD1
X
s3k
N
s
(t ) =
M
X
kD1
M
X
jD1
X
s3k;jsjDj
N
s
(t ) =
M
X
jD1
X
jsjDj
N
s
(t )
leadingtothe relationshipthat links the set of dependent single-name default processes
N
k
with the set of independent and Poisson distributed counting processes 7
j
:
M
X
kD1
N
k
(t ) =
M
X
jD1
7
j
(t ) =: 7
t
(18)
Hence, the CPS framework offers us a way of consistently modeling the single-
name processes along with the pool-counting process, taking into account the cor-
relation structure of the pool, which remains specied within the denition of each
cluster process
N
s
. Note, however, that the 7
t
,M process is not, properly speaking,
the rescaled number of defaults
C
t
, since the former can increase without limit, while
the latter is bounded in the 0. 1| interval. We address this issue in the next section,
along with the issue of avoiding repeated single-name defaults.
5.3 The single-name adjusted approach
In order to avoid repeated defaults in single-name dynamics, we can introduce con-
straints on the single-name dynamics ensuring that each single name makes only one
default. Such constraints can be implemented by modifying Equation (16) in order to
Research Paper www.thejournalofcreditrisk.com
74 D. Brigo et al
allow for only one default. Given the same repeated cluster processes
N
s
as before,
we dene the new single-name default processes N
1
k
replacing N
k
as solutions of the
following modication of Equation (16) for the original N
k
:
dN
1
k
(t ) := (1 N
1
k
(t
))
X
s3k
d
N
s
(t )
=
X
s3k
d
N
s
(t )
Y
s3k
1
f
Q
N
s
.t
/D0g
(19)
Remark 5.1 (Interpretation) This equation amounts to saying that name k jumps
at a given time if some cluster s containing k jumps (ie,
N
s
jumps) and if no cluster
containing name k has ever jumped in the past.
We can compute the new cluster defaults
N
1
s
consistent with the single names
N
1
k
as:
d
N
1
s
(t ) =
Y
j2s
dN
1
j
(t )
Y
j2s
c
(1 dN
1
j
(t )) (20)
where s
c
is the set of all names that do not belong in s.
Now we can use Equation (18), with the N
1
k
replacing the N
k
, to calculate how the
new counting processes 7
1
j
are to be dened in terms of the new single-name default
dynamics:
M
X
kD1
dN
1
k
(t ) =
M
X
kD1
(1 N
1
k
(t
))
X
s3k
d
N
s
(t )
=
M
X
kD1
(1 N
1
k
(t
))
M
X
jD1
X
s3k;jsjDj
d
N
s
(t )
=
M
X
jD1
X
jsjDj
d
N
s
(t )
X
k2s
(1 N
1
k
(t
))
=
M
X
jD1
X
jsjDj
d
N
s
(t )
X
k2s
Y
s
0
3k
1
f
Q
N
s
0 .t
/D0g
This expression should match:
d7
1
(t ) :=
X
j
d7
1
j
(t )
so that the counting processes are to be dened as:
d7
1
j
(t ) :=
1
X
jsjDj
d
N
s
(t )
X
k2s
Y
s
0
3k
1
f
Q
N
s
0 .t
/D0g
(21)
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 75
The intensities of the above processes can be directly calculated in terms of the
density of the process compensator. By direct calculation, we obtain:
h
N
1
k
(t ) =
Y
s3k
1
f
Q
N
s
.t
/D0g
X
s3k
z
s
(t )
h
Z
1
j
(t ) =
1
X
jsjDj
z
s
(t )
X
k2s
Y
s
0
3k
1
f
Q
N
s
0 .t
/D0g
9
>
>
>
>
=
>
>
>
>
;
(22)
where, in general, we denote by h
X
(t ) the compensator density of process X at time t ,
referred to as intensity of X, and where
z
s
is the intensity of the Poisson process
N
s
.
If we consider the repeated Poisson cluster default building blocks
N
s
to be
exogenously given, the model N
1
k
.
N
1
s
. 7
1
j
is a consistent way of simulating the
single-name processes, the cluster processes and the pool-counting process from the
point of view of avoiding repeated defaults. In particular, we obtain:
C
t
:=
1
M
X
k
N
1
k
(t ) =
1
M
7
1
t
6 1
Note, however, that the denition of N
1
k
in (19), even if it avoids repeated defaults of
single names, is not consistent with the spirit of the original repeated cluster dynamics.
Consider the following example.
Example 5.2 (Single names versus clusters) Consider two clusters s = {1. 2. 3]
and : = {3. 4. 5. 6]. Assume that no names defaulted up to time t except for cluster :,
in that in a single past instant preceding t names 3, 4, 5 and 6 (and only these names)
defaulted together (ie,
N
z
jumped at some past instant). Now suppose that, at time t ,
cluster s jumps, ie, names 1, 2 and 3 (and only these names) default, so
N
s
jumps for
the rst time. Does name 2 default at t ?
According to our denition of N
1
2
, the answer is yes, since no cluster containing
name 2 has ever defaulted in the past. However, we have to be careful in interpreting
what is happening at cluster level. Indeed, clusters : and s cannot both default since,
in this way, name 3 (in both clusters) would default twice. So we see that the actual
clusters default dynamics of this approach, implicit in Equation (20), does not have
a clear intuitive link with processes
N
s
. This is why, in Brigo et al (2007), we present
a second strategy to avoid repeated defaults that is also consistent at the cluster level.
5.4 Homogeneous pool limit
To simplify the parameters and combinatorial complexity, we may assume that the
cluster intensities
z
s
depend only on the cluster size [s[ = . Then it is possible to
Research Paper www.thejournalofcreditrisk.com
76 D. Brigo et al
directly calculate the intensity of the pool-counting process C = 7
1
as:
h
Z
1(t ) =
1
7
1
t
X
j
z
j
where
z
j
is the common intensity of clusters of size .
We see that the pool-counting process intensity h
Z
1 is a linear function of the
counting process C = 7
1
itself, as would be expected from general arguments for
a pool of independent names (again with homogeneous intensities). In such a pool,
a default of one name does not affect the intensity of default of other names, and
the pool intensity is the common homogeneous intensity multiplied by the number
of outstanding names. Each new default simply diminishes the pool intensity of one
common intensity value and the pool intensity is always proportional to the number
(fraction) of outstanding names (1
C).
5.5 Credit default swap calibration
In this section we tackle the single-name CDS calibration using the above strategy for
avoiding repeated defaults. In such a strategy, the compensator for the single default
event of name k is given by Equation (22). This leads to the following expression for
the default leg of a CDS on name k with deterministic recovery 1 paying protection
up to maturity T , under deterministic risk-free interest rates, and assuming that the
cluster intensities
z
s
are deterministic:
E
0
deeg(0)| = LGD
Z
T
0
1(0. u)E
0
dN
1
k
(u)|
= LGD
Z
T
0
1(0. u)E
0
h
N
1
k
(u)| du
= LGD
Z
T
0
1(0. u)
Y
s3k
E
0
1
f
Q
N
s
.u
/D0g
|
X
s3k
z
s
(u) du
= LGD
Z
T
0
1(0. u)
Y
s3k
e
s
.u/
X
s3k
z
s
(u) du
= LGD
Z
T
0
1(0. u) exp
X
s3k
s
(u)
X
s3k
z
s
(u) du
which is the same as the default leg in a standard deterministic intensity single-name
model for the CDS when the intensity of name k at time t is given by:
h
CDS
k
(t ) :=
X
s3k
z
s
(t )
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 77
In particular, if cluster intensities
z
s
are also constant in time, besides being deter-
ministic, and if we consider CDSs with continuous payments in the default leg, we
can see that single-name consistency with CDSs on name k having fair running spread
S
k
CDS
(0. T ) for maturity T is achieved with:
X
s3k
z
s
=
S
k
CDS
(0. T )
LGD
Imposing this (linear) constraint on the
z
0
s
for all k values in the CDO calibration
ensures single-name consistency with CDSs. Of course, there is no need to assume
constant cluster intensities, so the method also works when one tries to t the term
structure of both CDSs and CDOs. It is sufcient to go through the obvious general-
ization for piecewise constant cluster intensities.
This highlights a great advantage of this capping strategy for avoiding repeated
defaults: it makes single-name CDS calibration very easy to achieve.
5.6 Customizing the generalized Poisson cluster loss model to
deal with sectors
As was previously done for the GPL model, we can split the default counting process
and the loss process to factor out the Armageddon event in order to assign a zero
recovery to such a mode to allow for a better calibration of super-senior tranches.
We consider the rescaled pool-counting process
C
t
and the corresponding reduced
process c
1
t
, which can be dened as:
C
t
:=
1
M
M
X
jD1
7
1
j
(t )
c
1
t
:=
1
M 1
M1
X
jD1
d7
1
j
(t )
Starting from the counting process, we obtain:
d
C
t
=
1
M
M
X
jD1
d7
1
j
(t )
=
1
M
M
X
jD1
X
jsjDj
d
N
s
(t )
X
k2s
Y
s
0
3k
1
f
Q
N
s
0 .t
/D0g
Research Paper www.thejournalofcreditrisk.com
78 D. Brigo et al
= 1
f
Q
N
.t
/D0g
1
M
M1
X
jD1
X
jsjDj
d
N
s
(t )
X
k2s
Y
s
0
3k;s
0
1
f
Q
N
s
0 .t
/D0g
1
f
Q
N
.t
/D0g
d
N
(t )
1
M
X
k2
Y
s
0
3k;s
0
1
f
Q
N
s
0 .t
/D0g
= 1
f
Q
N
.t
/D0g
d c
1
t
1
f
Q
N
.t
/D0g
d
N
(t )(1 c
1
t
)
and, by introducing the non-Armageddon indicator 1
1
t
, we obtain:
d
C
t
= 1
1
t
d c
1
t
(1 c
1
t
)1
1
t
d1
1
t
. 1
1
t
:= 1
fZ
1
M
.t/D0g
= 1
f
Q
N
.t/D0g
where is the set of all possible names.
For the GPL model we then introduce the stopping time t as the minimum time
between the Armageddon event time and the time when the reduced pool is over:
namely, the time when the full pool is over. Nowwe can integrate in order to obtain the
counting process by observing the following: rstly, 1
1
is zero after the Armageddon
event, otherwise it is equal to one; secondly, d1
1
is non-zero only at the Armageddon
event; and nally, c is equal to one and d c is zero after the reduced pool is over. After
some algebra, we obtain:
C
t
= 1
fO 6tg
c
1
t
1
fO >tg
Note that
C
t
depends on the value of c
1
only at time t and that the equation relating
the two counting processes has the same form as the corresponding Equation (13) of
the GPL model.
The same line of reasoning used for the GPL model may be repeated for the GPCL
loss process as well by considering a zero recovery for the Armageddon event. We
obtain:
d
1
t
= (1 1)1
1
t
d c
1
t
(1 c
1
t
)1
1
t
d1
1
t
and, by integrating:
1
t
= (1 1 c
1
O
)1
fO 6tg
(1 1) c
1
t
1
fO >tg
we obtain an equation relating the loss of the pool to the counting process of a reduced
GPCLmodel, which has the same formas the corresponding Equation (14) of the GPL
model, allowing for the same numerical techniques to calculate the loss distribution.
6 CONCLUSIONS AND FUTURE DEVELOPMENTS
In this paper we have considered CDOs. We analyzed their valuation (both pre-crisis
and in-crisis) using our pre-crisis GPL model, an arbitrage-free dynamic loss model
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 79
capable of calibrating all the tranches for all the maturities simultaneously. We have
also conrmed the CDO model-implied loss-distribution features that we had already
highlighted in Brigo et al (2006a): we found a multimodal loss probability distribution
that can also be obtained independently within the implied copula framework. This
behavior can be associated to the default of clusters of names within the CDO pool,
which, in turn, may be interpreted as default of sectors of the economy. We repeat
the analysis in-crisis, updating our model to make it more functional. Our discussion
is abundantly supported by market examples through history, both pre-crisis and in-
crisis. Our results give an example of how, even before the crisis, attempts had been
made to go beyond the static and inconsistent notions of copulas and base correlation,
targeting dynamic dependence models for CDOs and credit derivatives (for a more
extensive overview and references we refer the reader to Brigo et al (2010)).
REFERENCES
Albanese, C., Chen, O., Dalessandro, A., and Vidler, A. (2006). Dynamic credit correlation
modeling. URL: www.defaultrisk.com/pp_corr_75.htm.
Balakrishna, B. S. (2006). A semi-analytical parametric model for dependent defaults. URL:
www.defaultrisk.com/pp_crdrv128.htm.
Balakrishna, B. S. (2007). Delayed default dependency and default contagion. URL: http://
papers.ssrn.com/sol3/papers.cfm?abstract_id=980684.
Baxter, M. (2007). Lvy simple structural models. In Credit Correlation: Life After Copulas,
Lipton, A., and Rennie, A. (eds). World Scientic.
Bennani, N. (2005). The forward loss model: a dynamic term structure approach for the
pricing of portfolio credit derivatives. URL: www.defaultrisk.com/pp_crdrv_95.htm.
Berd, A. M., Engle, R. F., and Voronov, A. B. (2007). The underlying dynamics of credit
correlations. URL: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1295847.
Bielecki, T., and Rutkowski, M. (2001). Credit Risk: Modeling, Valuation and Hedging.
Springer.
Bielecki, T., Vidozzi, A., andVidozzi, L. (2006). Pricing and hedging of basket default swaps
and related derivatives. Preprint.
Bielecki, T., Crepey, S., and Jeanblanc, M. (2008). Up and down credit risk. Working Paper.
URL: www.defaultrisk.com/pp_model190.htm.
Brigo, D., Pallavicini, A., and Torresetti, R. (2006a). The dynamical generalized-Poisson
loss model. I. Introduction and CDO calibration. Risk 20(5), 7075.
Brigo, D., Pallavicini, A., andTorresetti, R. (2006b).The dynamical generalized-Poisson loss
model. II. Calibration stability and spread dynamics extensions. URL: http://ssrn.com/
abstract=900549.
Brigo, D., Pallavicini, A., and Torresetti, R. (2007). Cluster-based extension of the general-
ized Poisson loss dynamics and consistency with single names. International Journal of
Theoretical and Applied Finance 10(4), 607631.
Brigo, D., Pallavicini, A., and Torresetti, R. (2009). Credit models and the crisis, or:
how I learned to stop worrying and love the CDOs. URL: http://papers.ssrn.com/sol3/
papers.cfm?abstract_id=1529498.
Research Paper www.thejournalofcreditrisk.com
80 D. Brigo et al
Brigo, D., Pallavicini, A., and Torresetti, R. (2010). Credit Models and the Crisis: A Journey
into CDOs, Copulas, Correlations and Dynamic Models. Wiley, Chichester.
Chapovsky, A., Rennie, A., and Tavares, P. A. C. (2007). Stochastic intensity modelling for
structured credit exotics. In Credit Correlation: Life After Copulas, Lipton, A., and Rennie,
A. (eds). World Scientic.
Cont, R., and Minca, A. (2008). Recovering portfolio default intensities implied by CDO
quotes. URL: http://ssrn.com/abstract=1104855.
Cont, R., and Savescu, I. (2008). Forward equations for portfolio credit derivatives. URL:
http://ssrn.com/abstract=1124954.
Di Graziano, G., and Rogers, C. (2005). A new approach to the modeling and pricing of
correlation credit derivatives. URL: www.statslab.cam.ac.uk/
chris/papers/cdo21.pdf.
Duan, J. C. (2009). Clustered defaults. URL: http://ssrn.com/abstract=1511397.
Elouerkhaoui, Y. (2006). Pricing and hedging in a dynamic credit model. Working Paper,
Citigroup. Presented at the Credit Correlation: Life After Copulas conference, London,
September 29, 2006.
Errais, E., Giesecke, K., and Goldberg, L. (2009). Afne point processes and portfolio credit
risk. URL: www.defaultrisk.com/pp_cdo_16.htm.
Fermanian, J. D. (2009). Atop-down approach for MBS, ABSand CDOof ABS: a consistent
way to manage prepayment, default and interest rate risks. URL: http://papers.ssrn.com/
sol3/papers.cfm?abstract_id=1233962.
Giesecke, K., Goldberg, L., and Ding, X. (2009). A top down approach to multi-name credit.
URL: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1142152.
Golub, H., and Van Loan, C. (1983). Matrix Computation. Johns Hopkins University Press,
Baltimore, MD.
Halperin, I., and Tomecek, P. (2008). Climbing down from the top: single name dynamics
in credit top down models. Research Paper, JP Morgan. URL: http://arxiv.org/PS_cache/
arxiv/pdf/0901/0901.3404v1.pdf.
Hull, J., and White, A. (2006). The perfect copula. Working Paper, Rothman School of
Management.
Inglis, S., Lipton, A., Savescu, I., and Sepp, A. (2008). Dynamic credit models. Statistics
and Its Interface 1, 211227.
Jaeckel, P. (2008). The gamma loss and prepayment model. Risk 9(9), 134139.
Joshi, M. S., and Stacey, A. M. (2006). Intensity gamma: a newapproach to pricing portfolio
credit derivatives. URL: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=907384.
Li, D. X. (2000). On default correlation: a copula function approach. Journal of Fixed Income
9(4), 4354.
Li, D. X., and Hong Liang, M. (2005). CDO
2
pricing using Gaussian mixture model with
transformation of loss distribution. URL: www.defaultrisk.com/pp_crdrv105.htm.
Lindskog, F., and McNeil, A. (2003). Common Poisson shock models: applications to insur-
ance and credit risk modeling. ASTIN Bulletin 33, 209238.
Longstaff, F., and Rajan, A. (2008). An empirical analysis of the pricing of collateralized
debt obligations. Journal of Finance 63(2), 529563.
Lopatin, A. V. (2008). A simple dynamic model for pricing and hedging heterogeneous
CDOs. URL: www.defaultrisk.com/pp_cdo_68.htm.
Lopatin, A. V., and Misirpashaev, T. (2007). Two-dimensional Markovian model for dynamics
of aggregate credit loss. URL: www.defaultrisk.com/pp_model159.htm.
The Journal of Credit Risk Volume 6/Number 4, Winter 2010/11
Credit models and the crisis: default cluster dynamics and the GPL model 81
Meng, C., and Sengupta, A. N. (2008). CDO tranche sensitivities in the Gaussian copula
model. URL: www.defaultrisk.com/pp_cdo_67.htm.
Papadopoulos, P., and Tan, C. (2007). Credit risk analysis of cashow CDO structures.
URL: www.defaultrisk.com/pp_cdo_38.htm.
Rosen, D., and Saunders, D. (2009). Analytical methods for hedging systematic credit risk
with linear factor portfolios. Journal of Economic Dynamics and Control 33(1), 3752.
Schnbucher, P. (2005). Portfolio losses and the term structure of loss transition rates: a
newmethodology for the pricing of portfolio credit derivatives. URL: www.defaultrisk.com/
pp_model_74.htm.
Sidenius, J., Piterbarg, V., and Andersen, L. (2008). A new framework for dynamic credit
portfolio loss modeling. International Journal of Theoretical and Applied Finance 11(2),
163197.
Standard & Poors (2009). Request For Comment: Update to Global Methodologies and
Assumptions for Corporate Cash Flow CDO and Synthetic CDO Ratings, March 18.
Standard & Poors Financial Services LLC, NewYork.
Torresetti, R., and Pallavicini, A. (2007). Stressing rating criteria allowing for default clus-
tering: the CPDO case. Working Paper. URL: http://papers.ssrn.com/sol3/papers.cfm?
abstract_id=1077762.
Torresetti, R., Brigo, D., and Pallavicini, A. (2006a). Implied expected tranched loss surface
from CDO data. URL: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=933291.
Torresetti, R., Brigo, D., and Pallavicini, A. (2006b). Implied correlation in CDO tranches:
a paradigm to be handled with care. URL: http://papers.ssrn.com/sol3/papers.cfm?
abstract_id=946755.
Torresetti, R., Brigo, D., and Pallavicini, A. (2006c). Risk-neutral versus objective loss dis-
tribution and CDO tranches valuation. Working Paper. URL: http://ssrn.com/abstract=
900784.
Torresetti, R., Brigo, D., and Pallavicini, A. (2009). Risk-neutral versus objective loss distri-
bution and CDOtranches valuation. Journal of Risk Management in Financial Institutions
2(2), 175192.
Research Paper www.thejournalofcreditrisk.com