asiarisk.com.

hk

Cutting edge | Credit derivatives

A market model on the iTraxx

A market model for the dynamics of credit-risky baskets and indexes such as the iTraxx has long been sought, but because of difficulties with the natural numraire has remained elusive. Here, Philippe Carpentier proposes using hedging arguments to develop such a model, identifying the resultant dynamics, deriving pricing results and examining the implied correlation
n formulating a market model for credit-risky baskets analogous to the Libor market model (LMM), we must first choose an appropriate numraire under which the spread is a martingale, then examine what drift is picked up in the pricing measure, and determine a consistent loss process. Inspired by the original work of Brace, Gatarek & Musiela (1997) on the LMM, we can extend to index spreads a proposal by Schnbucher (2004) and Brigo (2005), by taking risky zero-coupon (RZC) bonds as the natural numraire of the forward spread. However, we are prevented from a naive translation of the Libor methods by the fact that these instruments can take a zero value, and so Girsanovs theorem is not applicable. Sidenius, Piterbarg & Andersen (2005) consider the problem by directly modelling the cumulative loss density pt(x, T), that is, the probability at time t of the realised loss at T > t being less than x. They provide a general framework for the dynamics of pt(x, T), which is driftless by construction. A similar framework is presented by Schnbucher (2005), although he chooses to model the transition rates of the loss process. However, a major problem with these approaches is determining non-trivial dynamics of pt(x, T) that fulfil the no-arbitrage condition that pt be an increasing function of x. Similar no-arbitrage conditions arise in market models of vanilla options for equities and are notoriously difficult to handle. Here we focus on the case of a credit index such as the iTraxx. Our aim is to specify dynamics for its spread, derive pricing equations based on hedging arguments and understand the default correlation. We show how to negotiate the problem in applying the Girsanov theorem by the use of hedging arguments. In the case that the corresponding spread is given by a continuous diffusion process, we show its drift is zero under the RZC measure by hedging a simple caplet option. Relaxing the assumption of the dynamics of the spread by allowing it to jump at default, we determine the drift of the process between jumps under the bank account numraire, by hedging more general options. In the first part of the article, we make the link between forward spreads and RZC prices, showing how to strip the spread curve. Then we introduce an alternative to the standard index credit swaption and derive a Black-like formula for the price of such a product. Next,

we use one period collateralised debt obligations (CDOs) to hedge against defaults, and derive the spread drift in the more general case under the bank account measure. Finally, we give option pricing examples and the implied volatility and correlations.

Interest rates are assumed to be deterministic, the recovery rate R is fixed, and default is only possible at a set schedule of dates {Ti}. A one-period credit default swap (CDS) is defined to be a CDS whose spread is fixed at some date Ti and which expires at Ti+1. The value of the CDS is then zero on the interval [Ti, Ti+1). We now give the definitions of forward spreads and RZCs, suited to the case of an index. We consider a basket with M underlyings (for example, M = 125 for the iTraxx index, M = 50 for the HighVol index, M = 30 for the CrossOver index). Let hTi denote the number of survivors at date Ti, a decreasing integer process starting from M. Let ni+1 denote the number of defaults between Ti and Ti+1, an integer random variable observed at date Ti+1 with values between zero and hTi. At time t, the number of survivors is hTp(t)1, where p(t) is a discontinuous index denoting the time interval to which t belongs: t [Tp(t)1, Tp(t)). In the following, we drop the index t. We note that:

Stripping the spread curve

Ti +1 = Ti i+1
and that the description of the process ht is equivalent to the description of the loss process. The RZC for the index is an asset that pays hTi/M at date Ti. l Forward CDS contract. Following Schnbucher (2004) and Brigo (2005), we set a forward CDS fixing at Ti1 to expire worthlessly if all the obligors of the basket default prior to the fixing date, and otherwise to turn into a standard one-period CDS contract of maturity Ti. At date Ti, the protection buyer of a one-period index CDS pays the spread si on the surviving notional hTi/M and receives ni(1 R)/M at maturity. The default leg (DL) of a forward CDS contract is called a forward DL, paying ni(1 R) at Ti. The forward spread si is the spread that makes the value of this contract vanish at inception. This definition of a forward CDS is consistent with the usual stripping of the index spread curve.

67
067-072_AR_1209_Cutting_edge.ind67 67 11/12/09 15:54:22

Cutting edge | Credit derivatives

The benefit of the definition of a forward CDS is that we can relate the value of RZCs to forward spreads, and thereby strip the spread curve as in LMM. A simple no-arbitrage argument detailed in the online appendix gives the price Dti of an RZC of maturity Ti, given the forward spreads sk and the number hTp1 of surviving names t in the index, as: Taking advantage of assumption (3), this equation reads:

VT + = d VT
p p

N DT p

T p , sT
p

Dti = Pti

T p 1 M

ik = p

(1 R ) (1 R ) + stk

(1)

N = DT + ,
p

T p 1

(T , s ) ) D D T , s ) (T , s ) (
p

N DT + ,

N T p

T p 1

N T p+ ,T p 1 d

+ p

T p+

T p

+ p

T p+

where d is the day count fraction and Pti is the price at date t of a standard zero-coupon bond of maturity Ti. No assumption on the spread dynamics or on the distribution of defaults is needed since it is derived by statically replicating the RZC with a forward CDS. Thanks to equation (1), one can strip the spread curve out from suitable RZCs.

Since the spreads do not jump upon default:

sT _ = sT + and so VT + = d VT _ = 0
p p p p

Options on pure diffusion spreads

We can now examine a simple contract within this framework. A CDS caplet without loss payment of maturity TN is a contract with payout:

The terminal RZC, which is the natural numeraire for sN, is a perfect hedge against default. We do not need other hedging instruments. l Hedging spread variations. The value of the portfolio at time t + dt reads:

payout =

N DT N 1

N sT N 1

(2)

Vt + dt = Ot + dt + N DLN+ dt t

N N DLt

DtN

Ot

)D

N t + dt

Here we show that under simple assumptions its price in the appropriate numraire is given by a Black-like formula even though multiple defaults can occur. We assume in this section that upon default spreads do not jump, so that the dynamics are given by a pure diffusion. We start with the assumption that the caplet price O is proportional to its associated numeraire DN: t

DN = Ot + dt Ot t +Ndt + N DtN dt stN dt stN + + Dt

% % = Ot + dt Ot + N stN dt stN +

O ( t, t , s ) = DtN ( t, s )

(3)

with the constant of proportionality being a continuous function of t and s, as well as possibly non-dynamical parameters such as the strike, but does not depend on the number of survivors ht. We use as hedging instruments a forward DL and the terminal RZC DN. The forward DL hedges spread variations, while the RZC t hedges against defaults. The hedging strategy is to hedge spread variations with forward CDSs, which incurs no financing cost, and to invest the option premium in the final RZC. Let Vt be the value of a hedged short position on O. At t this is zero:

where in the last line a tilde denotes the asset value in RZC units. As the funding cost for the profit and loss of the portfolio is zero (recall Vt = 0), we have that in the terminal RZC measure, the spread sN is t driftless:

dstN = tN stN dwtN

Choosing the delta to be DN = Ot/sN, the pricing equation reads:
~

% % Ot 1 2O dt + d sN = 0 t 2 sN 2

( ) (

The option price is then given by:

N Ot = DtN E tRZC sT N 1 K

Vt = Ot + N DLN t

N DLN t

l Hedging defaults. Assume d names default at t = Tp. Then denoting times slightly after and before Tp with +/ subscripts respectively, we have:

DtN

Ot

)D

(4)

N t

=0

which clearly satisfies assumption (3). In conclusion, if spreads have lognormal dynamics and do not move upon default, the caplet price is given by a simple Black-like formula in the DN numraire. This is because no loss is paid to the protection buyer before TN1.

VT + = d VT
p

= O

( O (T , , s ) +
p T p 1 N DT p
p

T p+ , T p 1

d, s +

Options on spreads with jumps at default

N N N sT p DT + ,
p

N N N sT p DT
p

)D

T p 1

N T p+ ,T p 1 d

=O

T p , T p 1 , s

N DT + ,

N DT p

T p 1

O T p+ , T p 1 d, s

We now expand the spreads dynamics to general Lvy-It processes, allowing spreads to jump at default. To find the noarbitrage dynamical constraints we will consider general options whose payout depends, perhaps non-linearly, on the indexs loss process. i+1 At date Ti, knowing hTi and sTi , the distribution of the numi+1 ber n of defaults at time Ti+1 is constrained by the following relationship, which expresses the equality of the default and premium leg values:

68
067-072_AR_1209_Cutting_edge.ind68 68

December 2009/January 2010

11/12/09 15:54:30

asiarisk.com.hk

Cutting edge | Credit derivatives

i+1 PTi+1 d Ti 0 P i+1 = d Ti , sTi d = i PTi+1 d Ti 0 P = i

(1 R )

l upon default by an amount:

=
We also have:

( (

i+1

=d

i+1 Ti , sTi

M
Ti

) s

(5)

i+1 Ti

k k sT + sT
p p

l defaults

which we are free to specify in the model. l if no default occurs, by an amount given by: (6)
k k sT + sT
p p

i+1 d Ti 0 P i+1 = d Ti , sTi = 1 =

no default

There are only two equations and hTi + 1 unknowns (the probabilities of the hTi + 1 values that variable ni+1 can have). The choice for the distribution of n is a critical ingredient of the model as it determines default correlations. While in the single obligor case default can be hedged with a standard CDS, hedging defaults in the index case requires multiple instruments, as the number of defaults is random. A standard CDS on the index pays out an amount proportional to the number of defaults incurred, but the value of the option we are pricing will generally have non-linear dependence on the number of surviving names, so we are forced to posit another hedging instrument. n One-period CDO. We thus introduce a set of theoretical hedging instruments, one-period CDO contracts indexed by d. Such a CDO pays one at Tp if the number of defaults occurring at Tp is exactly d, otherwise the protection buyer simply pays the CDO premium: sdd where sd is the CDO spread, set so that the value of the CDO is zero at inception. This spread, to be distinguished from the forward CDS spread s, is obviously related to the distribution of defaults at Tp:

= d =1

P p = d k k s + sT p P p = 0 Tp

( (

) )

k DT p+ , d k d defaults D + T ,
p

The option price also jumps:

O T p+ , , sT + = O T p , , sT
d =1

P p = d O T p+ , d, sT + O T p , , sT p p P p = 0

( (

) (
) )

) ) ( )

l Resulting dynamics. The option price O(t, h) on the interval [Tp1, Tp) is then the solution of the following partial differential equation:

O 1 N t2 O dt + k , l = p+1 k l d s k , s l t 2 s s
N k = p+1

1 P =d sd = 1 P p = d

O s
k

lN p+1 =

ln Dtk stl

)d

stk , stl = rOdt

(7)

s d P =d = 1 + s d

which yields the dynamics of forward spreads in the bank account (BA) measure:

l Model ingredients. The ingredients of the models are the diffusion term of the spreads dynamics in the absence of default (volatility si and correlations rij = Cov[si, sj]/sisj) and the distribution of the jumps of the spread at default. In contrast, as we will see below, the drift terms and the spreads jumps in the absence of defaults are dictated by no-arbitrage conditions. We are free to set the distribution of the random variables ni insofar as it satisfies conditions (5) and (6). To specify what happens to forward spreads once default occurs, we pick a distribution for the jumps. For instance, in copula-based models the behaviour of the spread conditioned on default is dictated by the copula itself (see, for instance, Schmidt & Ward, 2002, for a detailed discussion in the context of the Gaussian copula). At default dates Tp, the forward CDS spreads sk jump:

ds i = ik = p+1

(1 R ) + s k

ik i k s k

s i dt + i s i dwti

We prove this via the hedging argument in the next two sections. Let O(t, h) be the value of the option and Dk the value of the RZC. The t,h hedged portfolio Vt is: l Short the option l Long Dk forward CDSs fixing at Tk1 and maturing at Tk l Long Nd one-period CDOs of maturity Tp l Financed with the bank account BAt. The Dk are set to hedge against forward spread variations, while the CDO notionals Nd are set so that the profit and loss vanishes, irrespective of the default scenario. We have:
N Vt = O ( t, t , st ) + k = p+1 k CDS k
t + d =1 N d CDOd +

O ( t, t , st ) BAt

BAt

A. Parameters and sets

Set 1 Set 2 0.05 0.25 0.003 Set 3 0 0.25 100

a b g

0.1 0.5 0

By definition of the forward spreads and of the CDO premiums, the value of the CDS and of the CDO is zero at t. l Hedging defaults. First consider what happens at Tp when defaults may occur. Let 0 < d hTp1 be the number of defaults occurring at Tp. Using +/ superscripts as before to indicate the right and left limits:

69
067-072_AR_1209_Cutting_edge.ind69 69 11/12/09 15:54:36

Cutting edge | Credit derivatives

V T p+ , d V T p ,

= O T p , , sT O T p+ , d, sT + p p

) (

) (

N k k k + N d l=1, l d N l s l + k = p+1 k sT + sT DT + , d
p p p

(8)

k sT + p

k sT p

no default

l=1

k k k sT + sT DT + ,l p p s l l defaults p T (1 + s l ) 1 l=1p 1 s l DTkp+ , (13) (1+s )

where we have explicitly expressed the fact that forward CDSs jump upon default. At T + we receive payment from the CDO that p was triggered by the d defaults and pay the premiums of the other CDOs. Equation (8) is general. Once we have specified in the model the variation of forward spreads:
k k sT + sT
p p

We can make this expression more transparent by rewriting it using the probability of defaults P(np = d), instead of the CDO spreads sl:
k k sT + sT
p p

no default

d =1

P p = d k k s + sT p P p = 0 Tp

( (

) )

k DT p+ , d k d defaults D + T ,
p

(14)

as a function of d and the spreads themselves, solving the system of hTp linear equations (8) yields the CDO notionals Nd so that V(T + , h p d) V(T , h) = 0 for any value of d hTp. p The notionals Nd are given analytically by:

The corresponding equation for the option price jump in the absence of default is given by (see online appendix):

1 Nd = O T + , d, sT + O T p , , sT p p (1 + s d ) p

N k k k + S k = p+1 k sT + , d sT DT + , d
p p p

) (

O T p+ , , sT + = O T p , , sT
d =1

(9)

P p = d O T p+ , d, sT + O T p , , sT p p P p = 0

( (

) (
) )

) ) ( )
(15)

where:

the total spread to be paid if no default occurs at Tp, is given by:

S = l =1p 1 N l s l

S=

s l l=1 (1+s ) l

O T p+ , l, sT + O T p , , sT p p
N k = p+1 k

) (

k sT p

)
(10)

k sT + , l p

k DT + ,l p

l Remark 1. If defaults could occur continuously, instead of discretely at dates Tp, the deterministic jump of spreads sk in the nodefault case would be replaced by a continuous drift. l Remark 2. Equation (15) is a discrete expression of a discounting term together with a source term dependent on the options value with fewer surviving names. l Hedging spread variations. On the time interval [Tp1, Tp), the only variations stem from the forward spread moves since we assume that the sp and hence the CDO spread stay constant:

When no default occurs at Tp, the portfolio variation reads:

O O 1 N 2O N dVt = t dt + k = p+1 kt ds k + k , l = p+1 k t l d s k , s l 2 s s s t

N N + k = p+1 k Dtk dstk + k = p+1 k lN p+1 =

T p+ ,

)V(

T p ,

= O T p , , sT O T p+ , , sT + p p

N k k k l=1 N l s l + k = p+1 k sT + sT DT + ,
p p p

In this case, the deterministic jump of the options value, O, across Tp is given by:

) (

Dtk stl

d stk , stl

(11)

+ rO ( t, t ) dt

The notionals Dk of forward CDSs are set so as to cancel the options sensitivity to forward spreads:

O T p , , sT O T p+ , , sT +
p

) (

l=1 N l s l

N k = p+1 k

)
p

k =

k sT + p

k k sT DT + ,
p

(12)

We then get the pricing equation for O on interval [Tp1, Tp):

Dtk s k

Once we have set Nd for all d > 0 in order to hedge the option variation on defaults, we have no more instruments at hand to hedge in the absence of default. No-arbitrage dictates that this variation vanishes, so that the right-hand side of equation (12) is zero. This jump depends on the spread variation in the absence of default, which is not a model ingredient but results from the no-arbitrage condition that the expected profit and loss of a long position on a forward CDS between T and T + is zero (see online appendix for p p more details). Therefore:

O 1N t2 O dt + d s k , sl t 2 k , l = p+1 s k s l
N k = p+1

O s
k

lN p+1 =

ln Dtk stl

)d

stk , stl = rOdt

which yields the dynamics of forward spreads in the BA measure between jumps:

ds i = ik = p+1

(1 R ) + s k

ik i k s k

s i dt + i s i dwtt

70
067-072_AR_1209_Cutting_edge.ind70 70

December 2009/January 2010

11/12/09 15:54:46

asiarisk.com.hk

Cutting edge | Credit derivatives

2. Credit swaption
Set 1 Set 2 Set 3 Price (bp) 180 160 140 120 100 80 60 40 20 0 Set 1 Set 2 Set 3

1. Default law
100 90 80 Probability (%) 70 60 50 40 30 20 10 0 0 1 2 3 4 5 6 7 8 9 10 11 12 1314 15 1617 18 19 20 Number of defaults

3. Constant maturity credit swap

20 18 16 14 Price (bp) 10 8 6 4 2 0 Set 1 Set 2 Set 3 12

4. Credit swaption without loss

60 50 Price (bp) 40 30 20 10 0 Set 1 Set 2 Set 3

with the drift term resulting from differentiation of equation (1).

Pricing examples
Here, we give pricing results for three typical index options: a CDS swaption or caplet, a constant maturity CDS and a swaption without realised loss. We assume for simplicity that forward spreads do not jump upon default. The payouts are as follows: n CDS swaption:

loss, in contrast, only depends on the dynamics of spreads, and is immune to default correlation. Constant maturity CDSs have a small sensitivity to default correlation as they vanish if all obligors default. l A choice for P(n). Being able to freely specify the default distribution P(n) is an important feature of our model. In the pricing examples below, we use the following form:

i, i+ n payout = loss + sTi K T

i

) )
+

k =i+1

k k DTi

P ( = d , s ) = P ( = 1 , s ) f , , ( , d ) f , , ( , d ) = 1{d =1} + 1{d = } + 1{1<d < } exp ( d )

The probability for one default, P(n = 1|h, s), is set by the consistency condition (5). Suppose that the initial three-month spread on the iTraxx is 171 basis points and that none of the 125 obligors has defaulted, as was the case on February 10, 2009. We have three sets of parameters, a, b and g, to choose. Three sets of those parameters are given in table A. In our first set, the default correlation is low to disallow sudden default of all obligors and b is rather large. The second set is more typical, while the third set is extreme in that all obligors default simultaneously. Figure 1 shows n laws consistent with the spread value for the different sets of parameters. Notice that for the third set, n

n Constant maturity CDS swaption:

payout = payout = sTi

i

i, i+ n sTi

n CDS swaption without loss payment:

i, i+ n

k n =i+1 k DTi k

i,i+n sT is the CDS spread at Ti of a spot starting CDS of maturity Ti+n.

Standard swaptions on the iTraxx involve the realised loss between inception and maturity. Because the swaptions payout then depends on the distribution of the realised loss, its value depends on default correlation. A credit swaption without realised

71
067-072_AR_1209_Cutting_edge.ind71 71 11/12/09 15:54:52

Cutting edge | Credit derivatives

5. Implied volatility
60 58 Implied volatility (%) 56 Correlation (%) 54 52 50 48 46 44 42 40 50 100 150 Strike 200 250

6. Correlation smile
100 90 80 70 60 50 40 30 20 10 0 0 5 10 15 Strike (%) 20 25 Set 1 Set 2 Set 3

can only take two values, zero and 125, and so the probability of zero default is very high. Figures 2, 3 and 4 show prices for the three types of products, using three different sets of parameters, calculated with the maturity on June 21, 2009 and a strike equal to 150bp. The underlying maturity is December 20, 2013 and its spread is 147bp. We have used one-factor lognormal dynamics for forward spreads: their volatilities are 50% and they are perfectly correlated with each other. In figure 5, we show the implied volatility for the same credit swaption without realised loss but with different strikes. The price of such an option is given in the appropriate numraire by a Blacklike formula involving the volatility of the forward spread. The swaptions implied volatility value is close to 50%, which we expect as spreads are perfectly correlated. Just like in a one-factor LMM, the slight non-lognormality of the swap rate generates a small skew that is here of the order of numerical error. Figure 6 shows the implied correlation smile of a standard CDO with maturity on December 20, 2013 for the three different sets. One may wish to calibrate the model on the correlation smile of five-year CDOs. This smile depends chiefly on the choice for P, even though higher spread volatilities also generate higher default correlation. Reasonable flexibility is achieved. l Remark 3. To calculate the base correlation, we assume that the spreads of all the names are equal. We had to consider the term structure of the spreads to imply the correlation. Indeed, if we do not take into account the decreasing term structure of spread, there was no solution for some strikes with the parameters of the first and the third set.

to the volatility and correlation parameters, as well as the jump measure at default times. Moreover, we are free to choose how the occurrence of defaults affects the index spreads. On the other hand, the correlation smile has no proper dynamics. While the equations governing the dynamics of the spreads in our model are similar to those for forward rates in LMMs, techniques for generating smiles and calibrating to the market cannot be applied directly as: most credit payouts specifically involve default events; standard issues in fixed income, such as relating caplet and swaption smiles, have limited relevance in the credit world; and issues such as the correlation smile are specific to credit. Further study is required in this area. l Philippe Carpentier is a member of the quantitative research team in the global markets division at Socit Gnrale. He would like to thank Lorenzo Bergomi, Pierre Henry-Labordre and two anonymous referees for helpful comments and suggestions. Email: phil. carpentier@sgcib.com

References
Bennani N and D Dahan, 2004 An extended market model for credit derivatives Stochastic Finance Autumn School & International Conference, available at http://nordine.bennani.free.fr/private/ ecmm.pdf Brace A, D Gatarek and M Musiela, 1997 The market model of interest rate dynamics Mathematical Finance 7(2), pages 127155 Brigo D, 2005 Market models for CDS options and callable floaters Risk January, pages 8994 Schmidt W and I Ward, 2002 Pricing default baskets Risk January, pages 111116 Sidenius J, V Piterbarg and L Andersen, 2005 A new framework for dynamic credit portfolio loss modeling Working paper, available at http:// jakobsidenius.com/papers/spa_vp_ 1.pdf Schnbucher P, 2004 A measure of survival Risk August, pages 7985 Schnbucher P, 2005 Portfolio losses and the term structure of loss transition rates: a new methodology for the pricing of portfolio credit derivatives Working paper, available at www. schonbucher.de/papers/cdo_loss_ transition_rates.pdf

Conclusion
We have presented a dynamic model for credit indexes such as the Itraxx that allows for a very flexible specification of the dynamics of forward spreads and of the correlation smile. This model produces hedge ratios involving index credit instruments only, with no reference to the single names comprising the index basket, namely forward CDS contracts and one-period CDOs. Default hedging relies on theoretical assets and a practical hedge would rather cancel the sensitivity to the model parameters. The dynamics of the spreads is very rich, with choice afforded

72
067-072_AR_1209_Cutting_edge.ind72 72

December 2009/January 2010

11/12/09 15:54:53