Anda di halaman 1dari 13

Simulating Multiple Defaults and Migration I:

Algorithms and Default Correlations


Chuang Yi∗
Market Risk, Bank of Montreal
November 16, 2009

∗ Manager of Market Risk, Bank of Montreal; 100 King St. West, 51st Floor, Toronto, Ontario, Canada, M5X 1H3;
Tel: 647-836-8124; Email: chuang.yi@live.ca. The author thanks Xiaofang Ma for helpful discussion and comments. This
article presents the personal views of the author and does not reflect any views or policies of Bank of Montreal.

1
Abstract
Zhou (2001) studied the first passage problem of multi-variate Brownian motions and provided
explicit formulas for joint default probabilities of two names. In this article, we propose an effi-
cient algorithm for simulating correlated defaults and migration based on the first passage time
framework. This algorithm guarantees both correct theoretical default probabilities and correct the-
oretical conditional marginal distribution of the distance to default. Moreover, this simulation is
able to generate consistent default correlations. For the driftless case, our numerical study shows
that the simulated default correlation closely matches theoretical default correlation, especially for
short time horizons. Finally, we illustrate an application of this simulation algorithm to value at fu-
ture of credit portfolios. Yi (2009) in a companion paper entitled “Simulating Multiple Defaults and
Migration II: Credit Value Adjustment of Credit Default Swaps”, applys the simulation algorithm to
study the credit value adjustment on credit default swaps.

Keywords: First Passage Times; Default Correlations; Multiple Defaults and Migration; Copula.

2
1 Introduction
Zhou (2001) studied the first passage problem of multi-variate Brownian motions and provided explicit
formulas for joint default probabilities of two names. For risk management purposes, it is important to
jointly simulate correlated defaults and migration. This is particularly emphasized in the most recently
proposed revisions to the Basel II market risk framework; see BIS (2008) for instance. Ideally, for a
given time horizon t, one would like to jointly simulate paths of all names and set the distance to default
to zero for those who have defaulted before time t. However, this is not practical due to computational
limitation, especially for a large pool of names. To complicate things further, the multi-name joint
distribution of the conditional marginals of distance to default is unknown in explicit form. In fact, even
for 3-dimentional standard Brownian motion, joint distribution of the first passage time is not known
explicitly.
There are currently three approaches that are practiced in the industry. The first approach is based on
copula that takes the marginal probabilities of default as inputs and outputs joint default probabilities.
However, it is difficult to incorporate joint migration to the copula model for varying time horizons. The
second approach is termed continuous cumulative model by Algorithmics, see Iscoe & Kreinin (2001),
and is also proposed by KMV, see Morokoff (2003). This approach is essentially a discrete version of the
ideal appoach and is therefore numerically intensive. The third approach is called Bucketing-Incremental
model by Algorithmics, and is also proposed by RiskMetrics. This model defines default time based on
barrier crossing of independent increments instead of the total distance to default. Effectively, the default
correlation is significantly reduced, see Iscoe (2003) for a discussion.
Given these difficulties, two important objects we are concerned with are: first, the probability of
default (PD) for a given time horizon t; second, the conditional marginal distribution of the distance
to default given survival up to time t. We aim to exactly match these two important quantities. The
normal copula of the underlying distance to default is used as a proxy to build the joint distribution
among the conditional marginals (conditional on survival). The default correlation generted by this
simulation matchs the theoretical default correlation amazingly well, at least for zero drift cases.
The rest of this document is organized as follows. Section 2 introduces some preliminaries of multi-
dimensional Brownian motions. Section 3 gives methodology of simulating multiple defaults. Section 4
discusses an alternative approach which produces much smaller default correlations. Section 5 provides
some numerical examples. In section 6, an application to credit portfolios is analysed.

2 Preliminaries
For company i, let Xt,i denote a drifted Brownian motion (DBM) with initial point x0,i > 0 given by

Xt,i = x0,i + θi t + σi Wt,i . (1)

where Wt,i denotes standard Brownian motion and dWt,i dWt,j = ρi,j dt. Define its default time as the
first passage time τi given by
τi = inf{t > 0 : Xt,i ≤ 0}.
The distance to default for company i is modeled as the stopped process:

X̂t,i = Xt∧τi ,i .

We have the following propositions.

Proposition 2.1. For a given time t > 0, the probability that τi ≤ t has the following explicit expression,
we have    
x0,i + θi t 2 x0,i − θi t
P (τi ≤ t|x0,i ) = Φ − √ + e−2x0,i θi /σi Φ − √ , (2)
σi t σi t
where Φ denotes the cumulative distribution function (cdf ) of a standard normal variable. Furthermore
the survival distribution of Xti , namely the distribution of Xti conditional on τi > t, is given by
√ 2 √
φ(x; x0,i + θi t, σi t) − e−2x0,i θi /σi φ(x; θi t − x0,i , σi t)
fXt,i |τi >t (x) =     , ∀x > 0. (3)
x +θi t 2 x −θi t
Φ 0,i σ t
√ − e−2x0,i θi /σi Φ − 0,iσ t

i i

3
where φ(x; µ, σ) denotes the probability density function (pdf ) of a normal distribution with mean µ and
volatility σ.
Equation (2) has appeared in numerous textbooks, such as Steele Steele (2004). Equation (3) has
appeared in Yi, Tchernitser & Hurd (2008), where they term it mirrored normal distribution.
Proposition 2.2. Assume θ1 = θ2 = 0, we have
2
r0 ∞
2r0 e− 4t sin( nπγ
  2  2 
α )
0
X r0 r0
P (τ1 > t, τ2 > t) = √ I vn −1 + I vn +1 ,
2πt n=1,3,...
n 2 4t 2 4t

where vn = nπ/α, Iv (x) is the modified Bessel function with order v and
  √ 
−1 1−ρ21,2
tan − , if ρ1,2 < 0


 ρ1,2
α =  √ 
1−ρ2
π + tan−1 − ρ1,21,2 , if ρ1,2 ≥ 0


  √ 
−1 y0,2 1−ρ21,2
tan , if y0,1 > ρ1,2 y0,2


 y0,1 −ρ1,2 y0,2
γ0 =  √ 2 
y0,2 1−ρ
π + tan−1 y0,1 −ρ1,2 y1,2 , if y0,1 ≤ ρ1,2 y0,2


0,2

r0 = y0,2 / sin(γ0 ),
y0,i = x0,i /σi .

The proof of Proposition 2.2 and a more general form for θi 6= 0 are both available in Zhou (2001).
For simplicity, let us denote pi := P (τi ≤ t) and pi,j := P (τi ≤ t, τj ≤ t). Then the default correlation
can be computed as:
pi,j − pi pj
ρdi,j = p . (4)
pi (1 − pi )pj (1 − pj )

3 Simulating Multiple Defaults and Migration


As mentioned before, we would like to match both default probabilites and conditional marginal distri-
bution of distance to default. In addition, we also want to use a reasonable copula for simulating joint
default and migration. This methodology is composed of three large steps to take care of each of the
concerns.

3.1 Step I: Copula


The first step is to simulate multi-variate drifted Brownian motions, for a given time horizon. This step
uses the underlying normal copula as a proxy for the real copula. More specifically, correlated xt,i are
simulated from Equation (1).
We will see from the rest of the steps that the simulated joint default probability p̂i,j can be computed
as

p̂i,j = P (Xt,i ≤ bi , Xt,j ≤ bj )


 
bi − x0,i − θi t bi − x0,i − θi t
= Φ2 √ , √ , ρi,j
σi t σi t
= Φ2 (Φ−1 (p1 ), Φ−1 (p2 ), ρi,j ).

Therefore, the simulated default correlation ρ̂di,j is given by

p̂i,j − pi pj
ρ̂di,j := p . (5)
pi (1 − pi )pj (1 − pj )

4
3.2 Step II: PD
The second step is to match the PD for each of the companys. In Step I, we have simulated the
marginals of Xt,i , which have both positive and negative support. However, in reality, only survived
firms are realized with positive support, while for defaulted firms, their distance to default at time t are
set to zero. For each company i, we determine a cut-off barrier bi to match the PD

P (Xt,i ≤ bi ) = P (τi ≤ t), (6)

where the left hand side (LHS) is given by


 
bi − x0,i − θi t
P (Xt,i ≤ bi ) = Φ √ (7)
σi t

and the right hand side (RHS) is given by Equation (2). Consequently, bi can easily be computed from

bi = x0,i + θi t + Φ−1 (pi )σi t. (8)

We then compare the simulated xt,i with the cut-off barrier bi . Let x̂t,i be one sample of our final target
simulated results.
• If xt,i ≤ bi , set x̂t,i = 0.
• If xt,i > bi , go to Step III to determine x̂t,i .

In Step II, we are therefore able to exactly match default probability pi .

3.3 Step III: Conditional Marginal


For those Xt,i which have survived from Step II, namely xt,i > bi , we determine x̂t,i as follows

P (Xt,i ≤ xt,i |Xt,i > bi ) = P (Xt,i ≤ x̂t,i |τi > t), (9)

where the LHS is given by

P (Xt,i > bi ) − P (Xt,i > xt,i )


P (Xt,i ≤ xt,i |Xt,i > bi ) =
P (Xt,i > bi )
x0,i +θi t−xt,i
Φ( σi t
√ )
= 1− x0,i +θi t−bi
(10)
Φ( σ √t )
i

and the RHS is given by


Z x̂t,i
P (Xt,i ≤ x̂t,i |τi > t) = fXt,i |τi >t (x)dx
0
   
x +θ t−x̂ 2 θ t−x −x̂t,i
Φ 0,i σ i√t t,i − e−2x0,i θi /σi Φ i σ 0,i √
i t
= 1− i    . (11)
x0,i +θi t −2x θ /σ 2 θi t−x0,i
Φ σ t √ −e 0,i i i Φ σ t √
i i

Consequently, for given xt,i > bi , x̂t,i is found from


     
x0,i + θi t − xt,i x0,i + θi t − x̂t,i −2x0,i θi /σi2 θi t − x0,i − x̂t,i
Φ √ =Φ √ −e Φ √ . (12)
σi t σi t σi t
Standard root finding procedures such as Newton-Raphson method could be used for solving the above
equation.

5
4 Alternative Simulation Scheme
An alternative approach to simulate multiple defaults is as follows:

• Simulate xt,i from Equation (1);


• If xt,i ≤ 0 set x̂t,i =0;
• If xt,i > 0, calculate conditional survival probability si ;
2x0,i xt,i

σ2 t
si := P (τi > t|xt,i ) = 1 − e i ; (13)

• Simulate independent Bi ∼ Bernoulli(si ) set


(
xt,i if Bi = 1;
x̂t,i = (14)
0 if Bi = 0.

This approach has been discussed in Yi (2008) for one dimensional case. This approach can also match
both default probability and the marginal distribution conditional on survival. The simulated joint
default probabiilty in this case becomes

p∗i,j := P (Xt,i ≤ 0, Xt,j ≤ 0) + P (Xt,i ≤ 0, Xt,j > 0)Dj


+ P (Xt,i > 0, Xt,j ≤ 0)Di + P (Xt,i > 0, Xt,j > 0)Di Dj ;

where
2 θi t−x0,i
Z ∞

2x0,i x
e−2x0,i θi /σi Φ( √ )
σi t
σ2 t
Di = e i fXt,i |Xt,i >0 (x)dx = θi t+x0,i
;
0 Φ( √ )
σi t
( √ x +θi t
φ(x; x0,i + θi t, σi t)/Φ( 0,i
σ t
√ ) if x > 0;
fXt,i |Xt,i >0 (x) = i

0 if x ≤ 0.

As a result, p∗i,j is given explicitly as

dj Φ(yj− )Φ2 (yi+ , yj+ , ρi,j )


p∗i,j = Φ2 (−yi+ , −yj+ , ρi,j ) + dj Φ(yj− ) −
Φ(yj+ )
di Φ(yi− )Φ2 (yi+ , yj+ , ρi,j ) di dj Φ(yi− )Φ(yj− )Φ2 (yi+ , yj+ , ρi,j )
+ di Φ(yi− ) − +
Φ(yi+ ) Φ(yi+ )Φ(yj+ )

where
2 x0,i + θi t −x0,i + θi t
di = e−2x0,i θi /σi ; yi+ = √ ; yi− = √ .
σi t σi t

5 Numerical Results
Figure 1 plots an example of simultaneously simulating distance to default for two firms. About 20000
scenarios are used for each of the subplots. The top two figures show the distance to default distribution
for firm 1 at times t = 1/4 year and t = 1/2 year respectively. The bottom two figures show the
distance to default distribution for firm 2 at times t = 1/4 and t = 1/2 year respectively. Please note
that at the origin of the coordinate for each of the subplot there is a positive mass. This positive mass
represents the probability of default before the given time horizon for a given firm. As time t = 1/4
moves to time t = 1/2, the simulated default probabilities for both firms become larger as we would
have expected. Meantime, the conditional survival marginal pdf is pushed to be flatter. The simulated
default probability for firm 2 is also consistently smaller than the ones for firm 1. This is because firm
2 has larger initial distance to default while other parameters are the same as firm 1.

6
Figures 2 shows the comparison of theoretical default correlation with simulated default correlations
and the alternative simulated default correlations for the zero drift case. The theoretical default correla-
tion is computed from equation 4 using proposition 2.2. There are several observations that deserve to be
mentioned. First, the simulated default correlation is very close to the theoretical one, especially for short
time horizons. Second, as the time horizon increases, the discrepancy between these two correlations
becomes larger. The increase of this discrepency is slow and tends to be stable as time increases. Third,
smaller distance to default introduces larger discrepancy between these two correlations. Fourth, the
alternative simulated default correlation is not a good proxy for the theoretical default correlation. This
is mainly due to inaccurate approximation for the joint default probability of the alternative approach.
Remark 5.1. There are ways to match the theoretical default correlation exactly from simulation. For
the first simulation approach, we can modify the distance to default simulation ρi,j slightly to match the
default correlation. Let us rewrite the notation pi,j (ρi,j ) := pi,j in order to show the dependence of the
distance to default correlation. Then we can solve ρm m m
i,j from pi,j (ρi,j ) = p̂i,j (ρi,j ). The solved ρi,j will
replace ρi,j for the simulation. The simulated default correlation from this modified approach will exactly
match the theoretical default correlation. Figure 3(a) shows the the relationship between ρm i,j and ρi,j .
Because these two correlations are too close, we plot ρm i,j − ρ i,j against ρ i,j . We notice that we need
to decrease the absolute value of ρi,j for the simulation in order to match exactly the theoretical default
correlation.
For the alternative approach, we can simulate correlated Bernoulli random variables instead of inde-
pendent ones. The correlation between two Bernoulli random variables ρB i,j can be computed so that the
simulated default correlation will match exactly the theoretical one. It is easy to show that ρB i,j is given
by
pi,j − p∗i,j
ρB
i,j = p .
Di Dj (1 − Di )(1 − Dj )Φ2 (yi+ , yj+ , ρi,j )
Figure 3(b) plots ρB
i,j as a function of time t for varying ρi,j . However, this approach is not applicable
since one has to validate the compatibility of the Bernoulli correlation matrix ρBi,j and its marginals for
every scenarios(see Chaganty & Joe (2006)).

6 Application to Credit Portfolios


6.1 A Toy Porfolio
In this section, we consider a toy credit portfolio which consists of N homogeneous defaultable zero
coupon bonds with maturity T , equal notional 1/N with zero recovery. For illustrative purposes, we
also assume zero interest rates and assume the risk-neutral default probability equals historical default
probability. The methodology described in this article can also be applied to more realistic inhomoge-
neous portfolios with non-zero recovery. For more realistic cases, historical parameters should be used
for simulation and risk-neutral parameters should be used for pricing of the bonds in order to calculate
P&L.
At time zero, the portfolio has an initial value of
N
X 1
V0 = [1 − P (τi ≤ T |x0,i )];
i=1
N

where P (τi ≤ t|x0,i ) is given by equation 2. We simulate the distance to default to a future time t for
M scenarios. Following the notation from simulation section, at future time t, the value of the portfolio
can be computed as
X 1
Vt = [1 − P (τi ≤ T − t|x̂t,i )];
N
i:survived

where P (τi ≤ t|x0,i ) is given by equation 2. The the profit and loss, P&L, is thus given by

P&L = Vt − V0 .

7
In general, the distribution of P&L has no explicit formula due to the complicated joint distribution of
x̂t,i . However, when N = 1, we are able to find its explicit distribution that is given by
(
P (τ ≤ T |x0 ) − 1, if τ ≤ t
P&L =
P (τ ≤ T |x0 ) − P (τ ≤ T − t|x̂t ), if τ > t

When x̂t = 0 we have P&L = −V0 = P (τ ≤ T |x0 ) − 1 representing largest loss. When x̂t = +∞ we
have P&L = 1 − V0 representing biggest gain.
Figures 4-6 show the simulated P&L distributions of homogeneous zero coupon bond pools for differ-
ent N = 1, 3, 100 number of bonds (from different issuers) for varying time horizons t = 1/4, 1/2, 1, 1.99
years, where T = 2 is the bond maturity. The P&L distribution lies in between its maximum loss, -0.8427
in these examples, and its the maximum gain, 0.1573 in these examples. As time horizon t increases from
a short period to the maturity of the bond, the distribution moves its masses towards multiple modes.
In Figure 5 for example, we can clearly identify four modes for t = 1 and t = 1.99. The first mode from
the left denotes the worst scenario that no bond has survived up to time t. The second mode represents
the case that only one bond has survived up to time t. The third mode represents the case that two
bonds have survived up to time t. The last model represents the case when all bonds have survived.
Figure 6 plots the P&L distribution of a homogeneous pool of 100 bonds. We can see from the figure
that the P&L distribution has been smoothed out a lot by increasing the number of bonds in the pool.
This smoothing effect comes from the small probability for a large number of bonds to default within
the same time range. However, multiple modes are also very clear when time horizon is close to the
maturity. Table 1 reports the 99.9% VaR numbers for these homogeneous zero coupon bond pools. The
statistics SVaR is defined as the quantitle of the P&L distribution given at least one company survived.
As time horizon t increases, the VaR is monotonic increasing (in size). It is interesting that the SVaR
is not monotonic in time. As number of bonds increases, the VaR decreases to account for portfolio
diversification. For N = 1, 3, there are several cases that the VaR number hits the maximum loss of
-0.8427. This is because the default probability of all bonds default within the time horizon exceeding
10 bps. If the probability of all bonds default within given time horizon is small enough, the SVaR will
be equal or very close to VaR.

t=1/4 t=1/2 t=1 t=1.99


VaR, N=1 -0.5501 -0.8427 -0.8427 -0.8427
SVaR, N=1 -0.5501 -0.7341 -0.7735 -0.1388
VaR, N=3 -0.3610 -0.5821 -0.8427 -0.8427
SVaR, N=3 -0.3610 -0.5821 -0.7463 -0.5305
VaR, N=100 -0.2525 -0.4076 -0.5689 -0.6283
SVaR, N=100 -0.2525 -0.4076 -0.5689 -0.6283

Table 1: VaR and SVaR Statistics for Homogeneous Zero Coupon Bond Pool: 99.9% confidence level
with maximum loss of -0.8427 and maximum gain of 0.1573. SVaR is defined as the quantitle of the
P&L distribution given at least one company survived.

6.2 Maturity Issues


In a real portfolio, it is common that many securities have different maturities. In this section we will
consider the effect of different tenors on the P&L of a portfolio. Consider another toy portfolio which
consists of two zero coupon bonds with maturities T1 and T2 and T1 < T2 . For simplicity, we will assume
again zero interest rates and zero recovery rates. Let their notionals be 1/2 each. At current time 0, the
initial value of the portfolio is given by
1 1
V0 = P (τ1 > T1 |x0,1 ) + P (τ2 > T2 |x0,2 ).
2 2
For any furture time t ∈ (0, T1 ], let us calculate the future value Vt of the portfolio. There are four
scenarios. First, if 0 < τ1 ≤ t and 0 < τ2 ≤ t, then Vt = 0. Second, if 0 < τ1 ≤ t and τ2 > t, then
Vt = 21 P (τ2 > T2 − t|x̂t,2 ). Third, if τ1 > t and 0 < τ2 ≤ t, then Vt = 21 P (τ1 > T1 − t|x̂t,1 ). Fourth, if

8
τ1 > t and τ2 > t, then Vt = 12 P (τ1 > T1 − t|x̂t,1 ) + 12 P (τ2 > T2 − t|x̂t,2 ). For t ∈ (0, T1 ], the simulation
is conducted the same as before.
For any future time t ∈ (T1 , T2 ], there are also four scenarios. First, if 0 < τ1 ≤ T1 and 0 < τ2 ≤ t,
then Vt = 0. Second, if 0 < τ1 ≤ T1 and τ2 > t, then Vt = 21 P (τ2 > T2 − t|x̂t,2 ). Third, if τ1 > T1
and 0 < τ2 ≤ t, then Vt = 12 . Fourth, if τ1 > T1 and τ2 > t, then Vt = 21 + 12 P (τ2 > T2 − t|x̂t,2 ).
For t ∈ (T1 , T2 ], two steps are needed for the simulation of Vt . We first jointly simulate distance to
default to time T1 . If x̂T1 ,2 ≤ 0, no further simulation is required and Vt can be computed. If x̂T1 ,2 > 0,
another simulation to time t is required for the second company conditional on x̂T1 ,2 . In the second step
simulation, the dimension has been reduced by one.
This simulation approach can be generalized to a portfolio that consists of multiple different matu-
rities. In general, when there are T1 < T2 < ... < Tn maturities in the portfolio. For t ∈ (Ti , Ti+1 ], one
requires i + 1 steps for simulating Vt . We first simulate to time T1 , and then simulate to T2 excluding
those instruments who have maturity T1 and those who have defaulted before or at T1 , conditional on
the distance to default at T1 of the survived companies. Keep doing this procedure until we arrive at
time t. There are two extreme scenarios that can be simulated in a simpler way: perfectly correlated
and the independent case.

6.2.1 Perfect Correlation Case


Consider the case that the two bonds are from the same issuer. Consequently, we have τ1 = τ2 and
Xt,1 = Xt,2 with probability one. We will hence suppress the index of τi and x̂t,i and use τ and x̂t
instead. For simulation convenience, it is clearer to write Vt in the following way:
• For 0 < t ≤ T1 : (
0, if τ ≤ t
Vt = 1 1
2 P (τ > T1 − t|x̂t ) + 2 P (τ > T2 − t|x̂t ), if τ > t

• For T1 < t ≤ T2 : 
0,
 if τ ≤ T1
Vt = 21 , if T1 < τ ≤ t
1 1

2 + 2 P (τ > T2 − t|x̂t ), if τ > t

As t increases from 0 to T1 , the probability P (τ ≤ t) increases. The masses of the distribution of Vt are
pushed to the two tails: more and more masses are added at point zero and the other masses are pushed
to the right. At time T1 , the distribution becomes discontinuous and there is no mass between (0, 0.5).
When t increases from T1 to T2 , the probability P (T1 < τ ≤ t) increases, hence putting more masses at
the point 0.5. The other masses are pushed again to the right. At time t = T2 , the distribution becomes
discrete with three possible outcomes (0, 0.5 and 1 with probability P (τ < T1 ), P (T1 < τ ≤ T2 ) and
P (τ > T2 ) respectively). This analysis can be generalized for n bonds with different maturities.

6.2.2 Independent Case


Consider another extreme case where the two issuers are independent. Then the future value of the
portfolio Vt can be simulated separately from the future value of the bonds Vti , i = 1, 2. More specifically
we have
• For 0 < t ≤ T1 :
(
0, if τ1 ≤ t
Vt1 = 1
2 P (τ1 > T1 − t|x̂t,1 ), if τ1 > t
(
0, if τ2 ≤ t
Vt2 = 1
2 P (τ2 > T2 − t|x̂t,2 ), if τ2 > t
Vt = Vt1 + Vt2 .

9
• For T1 < t ≤ T2 :
(
0, if τ1 ≤ T1
Vt1 = 1
2 , if τ1 > T1
(
0, if τ2 ≤ t
Vt2 = 1
2 P (τ2 > T2 − t|x̂t,2 ), if τ2 > t

Vt = Vt1 + Vt2 .

As t increases from 0 to T1 , the probability P (τ1 ≤ t, τ2 ≤ t) increases. The masses of the distribution
of Vt are pushed to the three points: more and more masses are added at zero point and the other
masses are pushed to the center point 0.5 and the right point 1. At time T1 , the distribution is still
continuous from 0 to 1. When t increases from T1 to T2 , the masses from (0, 0.5] are shifting to 0.5
and the masses from (0.5, 1] are shifting to point 1. Eventually, at time t = T2 , the distribution
becomes discrete with three possible outcomes 0, 0.5 and 1 with probability P (τ1 ≤ T1 )P (τ2 ≤ T2 ),
P (τ1 ≤ T1 )P (τ2 > T2 ) + P (τ1 > T1 )P (τ2 ≤ T2 ) and P (τ1 > T1 )P (τ2 > T2 ) respectively. This analysis
can be generalized for n bonds with different maturities.

Figure 1: Simulated Distance to Default: x0,1 /σ1 = 1, x0,2 /σ2 = 1.5, θ1 = θ2 = 0, ρ = 0.5.

References
BIS (2008), ‘Guidlines for computing capital for incremental risk in the trading book’. Consultative
document of Bank for International Settlements.

Chaganty, N. R. & Joe, H. (2006), ‘Range of correlation matrices for dependent bernoulli random
variables’, Biometrika (1).
Iscoe, I. (2003), Losing default correlation, step by step. Algorithmics technical paper.
Iscoe, I. & Kreinin, A. (2001), Default boundary problem. Algorithmics technical paper.

Morokoff, W. (2003), Simulation methods for risk analysis of collateralized debt obligations. KMV.
Steele, J. (2004), Stochastic Calculus and Financial Applications, Springer.

10
Figure 2: Theoretical Default Correlation vs Simulated Default Correlation: θ1 = θ2 = 0.

(a) Modified Correlation ρm


i,j (b) Bernoulli Correlations ρB
i,j

Figure 3: Modified Correlations and Bernoulli Correlations: x0,1 /σ1 = x0,2 /σ2 = 1.67, θ1 = θ2 = 0.

11
Figure 4: P&L of Homogeneous Zero Coupon Bond Pool: N = 1, M = 20K, x0 /σ = 2, θ = 0, T = 2.

Figure 5: P&L of Homogeneous Zero Coupon Bond Pool: N = 3, M = 20K, x0 /σ = 2, θ = 0, T = 2,


ρ = 0.3.

12
Figure 6: P&L of Homogeneous Zero Coupon Bond Pool: N = 100, M = 20K, x0 /σ = 2, θ = 0, T = 2,
ρ = 0.3.

Yi, C. (2008), ‘Value at future of path dependent instruments’. Technical document, Market Risk, Bank
of Montreal.
Yi, C. (2009), ‘Simulating multiple defaults and migration II: Credit value adjustment of credit default
swaps’. Working Paper, Market Risk, Bank of Montreal.

Yi, C., Tchernitser, A. & Hurd, T. (2008), ‘Randomized structual model on credit risk’. Third Interna-
tional Conference on Credit and Operational Risk, HEC Montreal, April 12-13, 2007.
Zhou, C. (2001), ‘An analysis of default correlations and multiple defaults’, The review of financial
studies (2).

13

Anda mungkin juga menyukai