Sie sind auf Seite 1von 20

The Valuation of Callable Bonds with

Floored CMS-spread Coupons


David Skovmand
University of Aarhus, Department of Management, University Park 322,
DK-8000 Aarhus C, Denmark, e-mail: dskovmand@econ.dk
Peter Løchte Jørgensen
Aarhus School of Business, Department of Business Studies,
Fuglesangs All 4, DK-8210 Aarhus V, Denmark, e-mail: plj@asb.dk

the LIBOR Market Model as well as a Gaussian two-factor short rate model. We show
Abstract
how to adapt the Least-Squares Monte Carlo procedure to handle the callability of the
A new type of structured bond has recently been introduced with enormous
product in a numerically efficient manner. We also calculate lower bounds for the
success–primarily among private investors–in many countries in Europe. The bonds
product as well as delta and vega ratios.
are medium term and with fixed and very high initial coupons. The remaining
coupons are determined as a constant multiplier times the spread between a long
and a short swap interest rate. These coupons are floored at or near zero, and the JEL Classificatio n Co des: G12, G13.
bond investment can thus be seen as a bet on the steepening of future term structure
curves. However, if the term structure becomes too steep, the bonds may be called by Keywords
the issuer. The paper studies the pricing and the optimal call strategy of these highly Structured products, callable bonds, optimal call strategies, Least-squares Monte
exotic bonds in a stochastic interest rate framework. We implement two versions of Carlo, G2++ and LIBOR market models.

1 Introduction From the investor’s point of view this type of product is essentially a
bet on the steepening of the term structure. Other than in the unlikely
The market for complex interest rate dependent products has been growing event of issuer default, the worst case scenario for the investor is being
rapidly in the past two years. Fueled by the low interest rate environment locked into a below market coupon until maturity of the bond. The high
investors have been seeking higher returns on their investments which has initial coupons are meant to compensate for this risk. The issuer has a
caused banks to come up with more and more complicated products to sell. limited downside since he can call the bond at par if the CMS spread
This paper investigates one of these products called the callable coupons he has to pay become too high.
Constant Maturity Swap (CMS)-steepener. This type of product is typically CMS spread related products (callable and non-callable) were extremely
sold as a medium term bond. The bond provides the investor with a high popular with investors in 2005 with an estimated $50 billion sold on a
fixed coupon (typically 6–10%) for an initial period but after this initial Worldwide basis (see Jeffery (2006)). However, the recent flattening of the
period the bond pays a floating coupon determined as some multiplum term structure of interest rates has caused a significant drop in the
of the spread between a long and short constant maturity swap rate secondary market value of the products. This has led to a wide criticism
determined in arrears. The coupons are floored at or near zero to avoid of issuers from the investor community where many feel they have
negative payments. The bond can also be called at par on coupon dates purchased something they did not understand. Indeed the largest pool
after the initial period making it a Bermudan type product. of buyers have been unsophisticated retail investors.

106 WILMOTT magazine


TECHNICAL ARTICLE 3

Adding to the controversy banks have also been reluctant to disclose The paper is arranged as follows. First, we fix notation and describe
the profits made from CMS Steepeners. According to Sawyer (2005) there the CMS Steepener in greater detail. Secondly, we describe the different
have even been suggestions that the banks themselves have had difficulties models of the term structure. Third, we outline the LSMC procedure as
with both pricing and managing risk. well as a theoretical derivation of the duality gap of a generic suboptimal
In the Nordic region sales of CMS Steepeners have been particularly strategy. Fourth, we show how hedge ratios are calculated and outline
high. The market was primarily triggered by the huge success of the numerical procedure which ends the theoretical part of our paper.
Forstædernes Bank (Forbank) who–in cooperation with Dexia Bank–issued Next we deal with the calibration and numerical implementation of the
a DKK 2.4 Billion ($400 Million) callable CMS Steepener in the beginning of different models of the term structure. Finally, we present the numerical
2005. We use this particular issue as a case of study in our investigation of results and discuss the implications. We end with a conclusion.
the products.
The academic literature has dealt with these products only on a very
abstract level where the most notable is Piterbarg (2004b). In a sense any 2 Basic Setup
callable bond can be viewed as a non-callable bond subtracted the value In this section we introduce some basic notation and briefly recapitulate
of the issuer’s embedded option to call the bond at par. Since the bond some central valuation expressions regarding simple interest rate derivatives.
can be called only at discrete times this embedded option is a Bermudan. These are used later in the paper mainly in relation to market calibration of
Therefore Piterbarg (2004b) uses the well-known method of Least-Squares our various dynamic models. The present section also describes the CMS
Monte Carlo (LSMC) (see Longstaff and Schwartz (2001), Carrière (1996), Steepener and its valuation on an abstract level.
and Tsitsiklis and Roy (1999)) to price the option part independent from We assume a standard probability space (, F , P ) with sigma-algebra
t=0 and real world probability measure P . We also make the
filtration {Ft }∞
the product itself. In this paper we propose a simpler and more efficient
way of pricing these bonds that is based on pricing the entire bond instead standard assumptions of no arbitrage and frictionless markets.
of splitting it into a non-callable and an option part. We present the tech- For a tenor structure T0 < . . . . . . < TM and day-count fractions
niques in a framework that is general enough to be used for any issuer τk = Tk − Tk−1 we define the discretely compounded time t forward LIBOR
callable bond. rates which apply between times Tk−1 and Tk as
A well-known issue in pricing American or Bermudan options using
the LSMC procedure is finding the proper exercise strategies. This has P(t, Tk−1 ) − P(t, Tk )
Fk (t) = F(t, Tk−1 , Tk ) = , (1)
been widely studied in the case of simple fixed for floating Bermuda τk P(t, Tk )
swaptions (Andersen (2000) and Svenstrup (2005)). In the case of CMS
spread callables this problem is more complex, and we therefore perform where P(t, Tk ) denotes the time t price of a zero coupon bond which
a numerical investigation to evaluate the quality of different exercise matures at Tk .
strategies. The technology to perform this quality check is developed in A call option on the forward rate giving the payoff τk (Fk (Tk−1 ) − K̄)+ at
the papers of Andersen and Broadie (2004) and Haugh and Kogan (2004). Tk is referred to as a caplet. The time t value of the caplet can be expressed
The intuition behind these papers is basic: The price of any option is as
equal to the value of the hedge portfolio, but to hedge a Bermudan option
we would need an optimal stopping rule. The LSMC procedure only Cpl(t) = P(t, Tk )τk EFt k [(Fk (Tk−1 ) − K̄)+ ], (2)
provides us with an estimated stopping rule. The present value loss or
duality gap from following a suboptimal stopping rule determines the
where Fk refers to the forward measure with corresponding numeraire
quality of the stopping rule approximation. We adapt the procedure in
P(t, Tk ) (see e.g. Björk (2004)). As is well-known, the forward rate Fk (t) must
Andersen and Broadie (2004) and show both theoretically and numerically
be a martingale under this measure.
how the duality gap can be calculated in the case of issuer callable bonds.
A Tα × (Tβ − Tα ) payer swap is a contract that entitles the holder to a
We would not expect the pricing of these products to be independent
series of floating LIBOR payments in exchange for paying a series of fixed
of the underlying assumption of the distribution of interest rates. We
payments at times Tα+1 < . . . . . . < Tβ .Tα is the starting point of the swap
therefore implement three different models of the term structure: The
and (Tβ − Tα ) is the tenor. The forward swap rate is defined as the level of
simple Gaussian short rate G2++ model of Brigo and Mercurio (2006) as
the fixed rate that makes the present value of the swap equal to zero. The
well as the log-normal (see Miltersen et al. (1997), Brace et al. (1997), and
forward swap rate at time t ≤ Tα is traditionally written as
Jamshidian (1997)) and the constant elasticity of variance (CEV) extended
version of the LIBOR market model (see Andersen and Andreasen (2000)
P(t, Tα ) − P(t, Tβ )
and Hull and White (2001)). We discuss the different properties of these Sα,β (t) = β . (3)
models as well as the specifics of their calibration. i=α+1 τi P(t, Ti )

Perhaps even more important than pricing are the sensitivities of the
product–the deltas and vegas. Calculating these in the LIBOR market model An equivalent representation of the forward swap rate (see e.g. Rebonato
is a non-trivial exercise. We show how the pathwise approach to sensitivities (2002)) is
in Glasserman and Zhao (1999), Piterbarg (2004a), and Piterbarg (2004b) can 
β
^
be adapted to our specific case and our results are consistent with intuition. Sα,β (t) = wi (t)Fi (t), (4)
i=α+1

WILMOTT magazine 107


with The entire bond value can therefore be written as
τi P(t, Ti )  
wi (t) = β . (5) 
N
CFi
j=α+1 τj P(t, Tj ) Vt = Nt ENt Nt−1 + − Ct , (11)
N
Nti
i=1
This expression shows that the forward swap rate can be seen as a weighted
average of forward LIBOR rates. Note, however, that the weights will be where the first term on the right-hand side of course corresponds to the
changing in a stochastic manner over time. value of an otherwise identical non-callable CMS Steepener bond. The
It is often more convenient to express the spot swap rate in terms of LSMC algorithm can be implemented on both (9) and (11). We shall
the length of the swap tenor, l: argue below that it will be computationally more efficient to work with
the single expression (9) than to work with the decomposition in (11) as
Sl (Ti ) ≡ Si,β (Ti+l) (Ti ), (6) is often done in similar contexts (see e.g. Piterbarg (2004b)).

where β(t) = i if ti−2 < t ≤ ti−1 .


A Tα × (Tβ − Tα ) payer swaption of strike K̄ is a contract that gives the 3 Model Choice
right but not the obligation to enter into a Tα × (Tβ − Tα ) payer swap at
In order to evaluate the model risk in pricing the CMS Steepener we will
time Tα and with fixed rate K̄. It can be shown (see e.g. Björk (2004), p. 381)
value this structured bond using three distinct models for the evolution
that the time t value of the payer swaption can be represented as
of the term structure of interest rates. We start out with the 2-factor
classical Gaussian model termed the G2++ model and proposed by
PS(t) = Cα,β (t)Et [(Sα,β (Tα ) − K̄)+ ],
α,β
(7) Brigo and Mercurio (2006). This model’s main advantage is tractability.
α,β Since the model has Gaussian state variables a closed form expression
where Et [·] denotes expectation formed under the swaption measure
 for their transition densities can be derived. This makes the evaluation
C α,β . The associated numeraire is defined as Cα,β (t) = βi=α+1 τi P(t, ti ).
of complex floating payoffs particularly easy. But the tractability comes at
We define a callable CMS Steepener as a bond with a principal of 1
a price and the model is not able to simultaneously price a large segment
and payments CFi at t1 < . . . ≤ ti ≤ . . . < tN and maturity at tN . We set the
of the vanilla instruments used for calibration.
day-count fraction δi = ti − ti−1 . The bond can be called at par by the
An entirely different approach is the so-called market models pioneered
issuer on all payment dates after the lockout period ends at time tc ≥ t1 .
by Miltersen et al. (1997), Brace et al. (1997), and Jamshidian (1997). These
In practice a pre-advise of a few business days is typically required, but this
models differ from the classical interest rate models in that they model
will be ignored here. In general the call date is unknown and therefore a
directly the evolution of discretely compounded market rates such as the
stochastic variable. We denote the time index of the call date by η where
LIBOR or the swap rate. As we will see, the main advantage of these
N ≥ η ≥ c. The payments are defined as
models is the ease at which they can be calibrated to price a large segment
 of vanilla products such as caps and swaptions. Many different market
δi R if i ≤ c
CFi = , (8) models exist and in this paper will use the standard log-normal forward
δi max[m[Sl2 (ti−1 ) − Sl1 (ti−1 )], f ] if i > c
LIBOR market model as well as the CEV-extension proposed by Andersen
and Andreasen (2000).
where R is the fixed initial coupon, and Sl2 and Sl1 are the swap rates set in
arrears as defined in (6) with l2 > l1 . These rates are referred to as
Constant Maturity Swap (CMS) rates because as time passes the rates have 3.1 The G2++ Model
a fixed time to maturity.1 m is a constant multiplier (e.g. 3) and f is a small
positive constant (e.g. 5 bps) which effectively floors the coupon payments In the G2++ model the instantaneous short rate under the risk-neutral
just above zero (for tax reasons). measure, Q c , is given by
Let us now consider the general valuation at time t, t < t1 , of the CMS
Steepener bond.2 From the First Fundamental Theorem of Finance (see r(t) = x(t) + y(t) + ϕ(t), r(0) = r0 , (12)
Björk (2004)) we know that this value can be represented as
  with stochastic factor processes, x(·) and y(·), described by the Ornstein-

CFi Uhlenbeck system
−1
N
Vt = Nt Et Nt n + , (9)
Nti
i=1
dx(t) = −ax(t)dt + σ dW 1 (t), x(0) = 0, (13)
where N is a generic martingale measure with corresponding numeraire
Nt . Now, if the issuer calls the bond before expiry he avoids paying
coupons but forfeits the interest earned on the principal. Assuming that dy(t) = −bx(t)dt + ηdW 2 (t), y(0) = 0, (14)
this interest rate is the LIBOR rate we can value the issuers option as
with
⎡ ⎤

N
CFi − δi Fi (ti − 1)
Ct = Nt ENt ⎣ ⎦. (10) dW 1 (t)dW 2 (t) = ρdt. (15)
Nti
i=η+1

108 WILMOTT magazine


TECHNICAL ARTICLE 3

In this dynamic system W1 (t) and W2 (t) are correlated Brownian motions, realistically replicates the so-called volatility skew observed in market
a, σ, b and η are positive constants, and −1 ≤ ρ ≤ 1. In (12), ϕ(·) is a deter- prices of interest rate options. The volatility skew refers to the decreasing
ministic function used to match the initial observed term structure of curve of Black-76 volatilities as a function of strike observed in market prices.
interest rates. The model is affine and the zero coupon bond price can The simplest extension is the CEV model investigated in Andersen and
thus be written as Andreasen (2000) and Hull and White (2001). This model uses a different dis-
tribution for the forward rate which allows for replication of simple versions
P(t, T) = A(t, T) exp{−B(a, t, T)x(t) − B(b, t, T)y(t)}, (16) of the skew. In reality the skew is more like a hockey-stick shape and if one
wants to replicate this pattern exactly one would need stochastic volatility or
where jumps in the LIBOR rate process. This causes problems with calibration as
 well as inherent difficulties with hedging as these type of models describe
P M (0, T) 1
A(t, T) = exp [V(t, T) − V(0, T) + V(0, t)] , (17) incomplete markets. We therefore stick to the CEV-extension of Andersen
P M (0, t) 2
and Andreasen (2000) as this retains completeness.
The general version of the LIBOR market model we work with is the
1 − e−z(T−t) following:
B(z, t, T) = , (18)
z
dFk (t) = σk (t)φ (Fk (t))dZk (t), t ≤ Tk−1 , ∀ k = 1, . . . , M, (20)
2

σ 2 −a(T−t) 1 −2a(T−t) 3
V(t, T) = T−t+ e − e − where Zk (t) is a Brownian motion under the forward measure with
a2 a 2a 2a
numeraire P(t, Tk ). Note that forward rates are martingales under
η2 2 1 −2b(T−t) 3
+ T − t + e−b(T−t) − e − their respective measures as they should be. We also assume constant
b2 b 2b 2b
correlations, i.e.
ση e−a(T−t) − 1 e−b(T−t) − 1
+ 2ρ T−t+ +
ab a b dZi (t)dZj (t) = ρi,j dt, ∀ i, j = 1, . . . , M. (21)

e−(a+b)(T−t) − 1
− ,
a+b (19) The choice of φ -function determines the distribution of the forward
rates. We look at two different specifications,
where P M (0, ·) denotes the market observed discount function. The forward
φ (x) = x, (22)
LIBOR and swap rates can now be generated in closed form through
formula (1) and (3)-(5). A pricing formula for swaptions is described in the
φ (x) = xp , 0 < p < 1. (23)
Appendix. For further details we refer to Brigo and Mercurio (2006).

The first of these yields a log-normal distribution of the forward rates


3.2 LIBOR Market Models and the second implies a non-central χ 2 -distribution (see Andersen and
The log-normal LIBOR market model (LMM) was derived with the intent Andreasen (2000)).
of having a model that was consistent with the standard market practice The two most important products we need to price are caplets and
of using the Black-76 formula to value caplets. Consistence with the swaptions. If we were to price a caplet the current formulation of the
Black-76 formula requires the discretely compounded or simple forward model would suffice since a caplet depends on one forward rate only.
rate to be a log-normal martingale under the forward measure, and short Recall from (2) that the time t price of a caplet on Fk (·) and expiring at
rate models are not able to achieve this. The HJM-model by Heath et al. time Tk is given as
(1992) is similar to LMM in spirit but models the instantaneous forward
rate instead of the simple forward rate. HJM was also the starting point Cpl(t) = P(t, Tk )τk EFt k [(Fk (Tk−1 ) − K̄)+ ]. (24)
for one derivation of the market model of Brace et al. (1997), and indeed
Hunt and Kennedy (2004) show that the market models can be seen as a With φ = x this immediately leads to the well-known Black-76 pricing
subset of the HJM model class. A PDE approach to market models can be formula, cf. Appendix B.2. When φ (x) = xp the distribution of the forward
found in Miltersen et al. (1997) but perhaps a more natural derivation rate is skewed to the left and the Black-76 formula no longer applies. This
which is also the one used by most textbooks is the change-of-numeraire skewing of the distribution causes in-the-money caplets to be priced higher
technique advocated by Jamshidian (1997). We refer to these papers or and out-of-the-money caplets to be priced lower implying a volatility skew.
the textbooks by Brigo and Mercurio (2006) and Rebonato (2002) for an A closed-form formula can also be derived and is shown in Appendix B.2.
extensive treatment. For swaptions no exact pricing formula exists in our reference LIBOR
The log-normal assumption of the forward rate distribution was known market models and one must resort to approximations or simulation to
to be at odds with reality long before LIBOR Market Models were invented. calculate prices. In such a simulation we would need realizations from the
Therefore several extensions have been proposed some of which still nest joint distribution of the forward rates comprising the underlying swap,
^
the log-normal version. The goal of the extensions was to have a model that and we therefore need the dynamics of all relevant forward rates under a

WILMOTT magazine 109


single measure. As it turns out a forward rate Fk (t) is only driftless under so they have to be estimated. Consider, for example, the option embedded
its “own” measure with numeraire P(t, Tk ). The no arbitrage conditions in the callable CMS Steepener whose value was defined in (10). Once this
determine the drift under other numeraires. The usual risk neutral option is exercised the holder “receives” a series of payments determined
measure (Q c cf. earlier) takes the continuously accrued bank account as by CFi − δi Fi (ti−1 ). The value of these payments cannot be calculated in
the numeraire, but since we are working with discrete rates it is more closed form for the LIBOR Market Models or the G2++ model. We must
convenient to work under the discrete money market measure which we therefore either approximate or simulate the value. Piterbarg (2004b)
denote by Q d and which has the discrete bank account as numeraire. The shows in a quite general setting how this value can be approximated by
discrete bank account is defined as follows: regression.
However, if instead of focusing on the option to call the bond we look
B0 = 1, (25) at the bond in its entirety, then the exercise value for the issuer is simply
the principal and therefore readily observable. Although this simple in-
BTk = BTk− 1 [1 + Tk Fk (Tk−1 )], 1 ≤ k < M, (26) sight avoids an entire regression step in the LSMC procedure it is rarely
Bt = P(t, Tk )BTk , t ∈ [Tk−1 , Tk ]. (27) used in the literature. Notable exceptions are Joshi (2006) and Amin
(2003) which both use this “trick” in different contexts. To the authors’
The dynamics of forward rates under this measure can be established knowledge it has never been applied in the case of issuer callable bonds,
using Itô’s Lemma and Girsanovs Theorem (see Andersen and Andreasen and we will therefore more carefully describe the procedure in a setting
(2000)). The result is appropriate for these types of bonds.
We assume an economy with a time t vector of state variables Xt . The
dFk (t) = σk (t)φ (Fk (t))(µ k dt + dZk (t)), objective of our analysis is to price a bond with maturity tN and a principal
of 1. The bond is issued at time t0 = 0, has payments CFi on payment dates

k
ρk,j τj σj (t)φ (Fj (t)) T = {t1 , . . . , ti , . . . , tN }, and is callable at par on all payment dates. The val-
µk = 1{t<Tj } , (28) uation of this bond at issuance is an optimal stopping problem of type
1 + τj Fj (t)
j=1

η

where we have redefined dZk (t) as a Brownian motion under Q . We can d V0P /N0 = inf EN0 [Nt−1
η
+ Nt−1
i
CFi ]. (29)
η
i=1
now simulate from the SDE in (28) and use formula (3) to get realizations
of the swap rate and value the expectation of B−1T α Cα,β (Tα )(Sα,β (Tα ) − K̄)
+
The superscript on V indicates that this is the solution to what we shall
under the discrete money market measure through simulation to get the later refer to as the Primal problem. As before N is a martingale measure
swaption price. with corresponding numeraire Nti at time ti . Note also that the index of
However, simulation is not feasible for calibration purposes. We exercise η is unknown and therefore a stopping time in a mathematical
therefore use an approximative formula for swaption prices given in sense. To price the product we need an estimate of the stopping time, or a
Appendix B.2. The approximation is mainly founded on assuming that stopping rule. To obtain such an exercise strategy we proceed as follows:
the swap rate follows the same distribution as the forward rates. This is of At each exercise time the issuer can either pay back the principal or wait
course not true because the swap rate will have a non-zero drift under until the next exercise time. If we assume that the coupons are reinvested
any measure in the LMM. However, the discrepancy between the distribu- in numeraire bonds the (accumulated) value at an exercise time is therefore
tions have been shown to be very small by Brace et al. (2001) and Hull and
 
White (2001) for the CEV-case, making this procedure accurate enough 
i
 −1 
for calibration purposes. Vti = min 1 + Nti Nt−1
k
CF k , N E N
ti ti N V
ti+ 1 ti+ 1 . (30)
k=0

4 Pricing Methodology A more convenient way of writing this is to define


Standard Monte Carlo simulation techniques cannot be used for pricing 
  i
callable products like the CMS Steepener since these methods do not Qti = Nti ENti Nt−1
i+ 1
Vti+ 1 − Nti Nt−1
k
CFk
identify the stopping rule (call strategy) necessary for characterizing the k=0 (31)
 
value of the security. One way to estimate such a stopping rule goes via N
= Nt i Et i Nt−1
i+ 1
{CFi+1 + min(1, Qti+1 )} ,
the Least-Squares Monte Carlo procedure normally attributed to Tsitsiklis
and Roy (1999), Carrière (1996), and Longstaff and Schwartz (2001). This and then rewrite (30) in terms of Qti as follows,
procedure has countless extensions and applications where Piterbarg
(2004b) is the most notable in our context since he focuses exclusively on 
i
Vti = Nti Nt−1 CFk + min(1, Qti ). (32)
the LIBOR Market Model. The idea of the LSMC procedure as used in k
k=0
Piterbarg (2004b) is described in the following.
To price any American or Bermudan product we essentially need two Vti cannot be directly evaluated since it depends on the exercise behavior
things everywhere in the state space: The exercise value and the continu- of the issuer at ti+1 . The LSMC procedure is essentially an algorithm to
ation value of the product. In general neither of these are readily available approximate Vti by simulation. But instead of approximating Vti directly

110 WILMOTT magazine


TECHNICAL ARTICLE 3

we approximate the Qti ’s instead. The two approaches are essentially The stopping time index for each path is
equivalent as one can always calculate one from the other using the above
formula. But approximating Qti means that we avoid approximating η(ω j ) = min(i ∈ [1, . . . , N] : Ii (ω j ) = 1). (37)
coupons that have already been determined. for the approximation of Qi
Finally, the MC price estimate of the bond price can the be calculated as
we choose for each time a polynomial function f (·, ·) : Rr×o+1 × Rr+ → R
follows
where o is the order of the polynomial ⎛ ⎞
1  
K η(ω j )
(0) (1) (2) (o) V̂0 = N0 ⎝Ntη (ω ) (ω j ) +
−1
Nti (ω j )CFi (ω j )⎠ .
−1
(38)
Qti ≈ f (βi ; R(Xti ) ) = βi + βi R1 (Xti ) + βi R1 (Xti )2 + . . . βi R1 (Xti )o (33) K j
j=1 i=1
..
. (34)
(r×o−o+1) (r×o−o+2) (r×o)
4.1 Lower bounds
+βi Rr (Xti ) + βi Rr (Xti )2 · · · + βi Rr (Xti )o
(35) The method described in the previous section does not guarantee that
the chosen fi -functions are optimal. This means that exercise might occur
Xti is a M dimensional vector of state variables and R(X) = (R1 (X), . . . , IRr (X)) at sub-optimal points in time. If this is the case the estimate in (38) will
where the Ri ’s are vector functions RM not converge to the true value of the bond as K is increased. Exercise is
+ → R+ that each map the state
variables into a key rate determined a priori to be important for the done by the issuer who seeks to minimize the value of the bond.
continuation value of the bond. Therefore if exercise is done sub-optimally the value of the bond will be
The important thing here is that the rates used must be measurable higher than in the optimal case. One can therefore interpret the price in
with respect to the filtration Fti , as the issuers decision to call can only (38) as an upper bound on the true bond price. The method does not give
be based on information known at the call time. Choosing these rates is any measure on how far one is from the true price and the estimate (38)
not an exact science, and it is very specific the particular problem at could thus in principle be severely upward biased. In this section we
hand. In general one should include rates that determine the level and therefore show how to calculate a corresponding lower bound estimate
the slope of the term structure at a given point in time. Piterbarg (2004b) to complement the upper bound. The difference between the lower and
also advocates the use of a so-called core swap rate. That is the swap rate upper bound is called the duality gap and can be seen as a measure of the
that sets on the day of exercise and expires on the day the entire bond quality of the estimated exercise strategy.
matures. For the case of CMS Steepeners we have found that using the short The problem of calculating a lower bound is similar to calculating
forward rate, the core swap rate, and the two swap rates that determine the an upper bound for American options. In the case of straight American
coupons give excellent results. options it is the holder of the option that exercises, and the challenge
As the functional form for f (·; R(Xti )) we use polynomial functions which in that case is therefore to establish an upper bound since a suboptimal
seems to be the industry standard. Choosing f to be linear in β is a smart strategy provides a downward biased estimate. A method of finding
choice from a computational perspective since it allows us to estimate β proper upper bounds for American options using simulation has been
with standard ordinary least squares (OLS). This choice does not appear to be recently developed in three seminal papers by Rogers (2002), Haugh and
overly restrictive (see Piterbarg (2004b)) and we will therefore use the linear Kogan (2004) and Andersen and Broadie (2004) (for other methods see
model throughout the paper. The LSMC procedure goes as follows: Glasserman (2004)). A less mathematical approach is that of Joshi (2006)
that gives a very good economic interpretation of the upper bound. In
1. Simulate K paths of the state variables {Xti (ω j )}Ni=1 and calculate all the cited cases the tool to calculate the upper bound is the concept
the numeraires {Nti (ω j )}Ni=1 and coupons {CFi (ω j )}Ni=1 for each path of duality which is well-known in, for example, the operations research
j = 0, . . . , K, ω j ∈ . literature. For more details on this subject we refer to the above cited
2. At the terminal date tN set QtN (ω j ) = 0 ∀ j = 0, . . . , K, ω j ∈ . papers and the references therein.
Below we show how a lower bound in relation to issuer callable bonds
3. At the last exercise time tN − 1 set QtN − 1 (ω j ) = NtN − 1 (ω j )NtN (ω j )−1 (1+CFN (ω j )) can be developed using the ideas from the American options literature.
K

R N − 1 (X N − 1 (ω j ))Q tN − 1 (ω j ) The main points can easily be lost in the mathematics and we therefore
and calculate the OLS estimator N−1 β̂ = K
j=1
.
j= 1
R N −1 (X N − 1 (ω j ))
R N − 1 (X N − 1 (ω j )) end with an economic argument based on the intuition provided in Joshi
(2006). Recall that the problem of pricing an issuer callable bond is that
4. For the next exercise time tN−2 set QtN − 2 (ω j ) = NtN − 2 (ω j )Nt−1
N−1
(ω j )
of computing
{CFN−1 (ω j ) + min(1, f (β̂N−1 ; RN−1 (XtN − 1 (ω j ))}, and calculate the OLS
estimator β̂N−2 in the same manner as in step 3. 
η

5. Repeat previous step until the first the exercise time t1 . V0P /N0 = inf EN0 [Nt−1
η
+ Nt−1
i
CFi )], (39)
η∈i
i=0
We now have an exercise strategy through the estimated parameters of
the fi -functions. We can define the indicator function for exercise as for a series of exercise times T = {t1 , . . . , tN }, with indices i = {1, . . . , N}.
We have earlier termed this the primal problem. Now, for an arbitrary

1 if 1 ≤ f (β̂i ; Ri (Xti (ω j )) or ti = tN finite submartingale Mtn – abbreviated Mn – we can define a so-called dual
Ii (ω j ) =
^
. (36)
0 if 1 > f (β̂i ; Ri (Xti (ω j ))) function F(n, Mn ) as follows,

WILMOTT magazine 111


     

η
Ṽtn+ 1 Ṽtn+ 1
F(n, M) = Mn + Etn N
min Nt−1 + Nt−1 CFi − Mη ) . (40) Mn + ENtn − ENtn = Mn . (46)
η∈[n,N]∩i η i
Ntn+ 1 Ntn+ 1
i=1

The dual problem consists of finding the maximum submartingale, This means that Mn is a martingale in the continuation region as well. If
we assume the strategy says exercise at tn so In = 1, then we have
V0D /N0 = sup F(0, M) =   

M
   Ṽtn+ 1 
n
−1 −1
η (41) N
Etn [Mn+1 ] = Mn + Etn N
− Ntn + Nti CFi (47)
sup M0 + EN0 min Nt−1 + Nt−1 CFi − Mη . Ntn+ 1
η i
i=1
M η∈i
i=1
  
Ṽtn+ 1 
n
−1 −1
Recall from the previous section that we defined Vtn as the value of the
N
−Etn − Ntn + Nti CFi = Mn . (48)
Ntn+ 1
i=1
bond newly issued at time tn plus the coupons up until that time invested
in numeraire bonds:
So Mt is also a martingale in the exercise region. Martingales are also
 

n
  submartingales so we can insert this in equation (43) to obtain a lower
Vtn = min 1 + Ntn Nt−1
i
CFi , Ntn ENtn Nt−1
n+ 1
Vtn+ 1 . (42) bound. Hence, a lower bound for the bond can be written at time 0:
i=0
 

η

We can now show the following theorem: V0L = Ṽ0 + E0 N


min (Nt−1
η
+ Nt−1
i
CFi − Mη ) . (49)
η∈i
i=1
Theorem 1: Given the primal and dual problems in (39) and (41), the
following statements are true
The first term is the upper bound calculated from the LSMC algorithm
1. The duality relation holds i.e V0P = V0D . described in the previous section. At the first exercise time (or at maturity

2. An optimal solution Mn for the dual problem is the discounted bond if the bond is never called) the exercise value is always equal to M. This
price process Vtn /Ntn . means that the second term will always be less than or equal to zero.
V0L can be interpreted as the price the buyer of the bond will be willing
Proof: See Appendix A.  to pay, whereas the price in (38) can be seen as the sellers price. The argument
The dual problem can now be used to construct a lower bound for goes as follows. The seller of the bond decides when he wants to call the
the bond price. We know the optimal submartingale, but solving this bond. This may or may not be at an optimal time. The buyer wants to be
process is equivalent to solving the primal problem. But if we where to hedged even if the bond is called optimally, but without knowledge of the
use an arbitrary submartingale instead we get the inequality, optimal strategy, exact replication is not possible. But the buyer can form
a subreplicating strategy based on selling the bond himself and following
  

a suboptimal exercise strategy. The initial value of the hedge will be V0L .
V0 /N0 = M0∗ + EN0 min Nt−1η
+ N −1
ti CF i − M ∗
t When following this strategy 4 situations can occur at a given point in
η∈i
i=1
   (43) time:

−1 −1
N
≥ M0 + E0 min Ntη + Nti CFi − Mt . 1. The buyer and seller exercise. In this case their prices will agree and
η∈i
i=1 the hedge will be perfect.
2. The buyer and seller do not exercise. Again the prices will agree.
Hence any submartingale will give a lower bound. Suppose we use the
3. The seller exercises and the buyer does not. If the time is optimal the
suboptimal strategy estimated in the previous section and let Ṽti be
buyer will take a loss on his hedging position which is financed by
defined as Vt in (42) when following this estimated suboptimal strategy
Ṽ selling numeraire bonds.
instead. Set M0 = Ṽ0 /N0 and M1 = Ntt1 and
1 4. The buyer exercises and the seller does not. The buyer can resell the
product with 1 less exercise date to continue the hedge. If the time
  
Ṽt Ṽt Ṽtn+ 1 
n was not optimal then losses from reselling are financed by selling
−1 −1
Mn+1 = Mn + n+ 1 − n + In ENtn − Ntn + Nti CFi . (44) numeraire bonds.
Ntn+1 Ntn Ntn+ 1
i=1
In all 4 cases the buyer’s price will be equal to or below that of the seller’s.
All the small losses are discounted back to present time. The losses are
We first notice that this is actually a martingale. Assume that the suboptimal
equal to the second term in (49) which is called the duality gap. We define
strategy says “continue” at time tn , that is In = 0. In this case we have:
this term as follows:
   
Ṽtn+ 1 Ṽt η
−1 −1
N
Etn [Mn+1 ] = Mn + EtnN
− n = (45) L N
0 = V0 − Ṽ0 = E0 min (Ntη + Nti CFi − Mη ) . (50)
Ntn+ 1 Ntn η∈i
i=1

112 WILMOTT magazine


TECHNICAL ARTICLE 3

This term is calculated through simulation. The algorithm is as follows the amount of payments on the relevant path will have changed.
Another reason to avoid using (53) is the computational cost since a new
1. Estimate an exercise strategy using the technique from the previous
price would have to be calculated for each perturbation.
section.
A superior method that both alleviates the instability problem as well
2. Simulate K paths of the state variables {Xti (ω j )}Ni=1 and calculate the the computational cost is the pathwise approach described in Glasserman
numeraires {Nti (ω j )}Ni=1 and coupons {CFi (ω j )}Ni=1 for each path and Zhao (1999) and refined in Piterbarg (2004a) and Piterbarg (2004b).
j = 0, . . . , K, ω j ∈ . The idea is as follows. Recall the forward rate dynamics

3. For the first path, ω1 , do the following: Start from t1 and calculate an
dFk (t) = σk (t)φ (Fk (t))(µ k dt + dZk (t)), t ≤ Tk−1 , ∀ k = 1, . . . , M, (54)
estimate, M̂1 , using a sub-simulation with K1 subpaths. The subpaths
are evolved from time t1 in path ω1 . Repeat this for all possible exercise
times creating a series of estimates M̂1 , . . . , M̂N−1 . 
k
ρk,j τj σj (t)φ (Fj (t))
 µk = 1{t<Tj } . (55)
Now calculate Ntn (ω1 )−1 + ni=1 Nti (ω1 )−1 CFi (ω1 ) − M̂n ∀n = 1, . . . , N 1 + τj Fj (t)
j=1
and find the minimum.
4. Repeat the previous step for all paths ω j and average the result to get We can differentiate through this system and calculate dynamics for the
ˆ 0.
an estimate of duality gap  delta ratios with respect to rates,

With N exercise times, K paths, and K1 subpaths the computational time d(m Fk (t)) = σk (t)φ
(Fk (t)) m Fk (t)(µ k (t, Fk (t))dt + dZk (t))
for calculating the lower bound is approximately of order K × N × K1 and ⎛ ⎞
M
∂µk (t, F̄(t))
therefore extremely slow. + σk (t)φ (Fk (t)) ⎝ m Fj (t)⎠ (56)
∂xj
j=0

5 Hedging and Sensitivities t ≤ Tk−1 , ∀ k = 1, . . . , M,

The hedging of derivatives is in general at least as important as pricing, where


and we therefore dedicate this section to some considerations on the
determination of the two most important hedge ratios–the delta and 
the vega. This analysis focuses mainly on the LIBOR market models. ∂µk  (φ
(Fj (t))(1 + τj Fj (t)) − φ (Fj (t))τj )τj ρk,j σj (t)
 = , (57)
∂xj xj =Fj (t) (1 + τj Fj (t))2
5.1 Deltas
φ (x) = xp , (58)
Delta denotes the sensitivity of the price with respect to shifts in the
initial vector of forward rates
φ
(x) = pxp−1 . (59)
F̄(0) = (F0 (0), F1 (0), . . . .., FM (0)). (51)
The starting value of this SDE is
The mth individual delta ratio is defined as follows 
1 for m = k
m Fk (0) = (60)
∂V0 0 for m = k
m = . (52)
∂Fm (0)
A nice trick that saves computational time is fixing the drifts at the time
One can easily calculate a first order approximation for this number by zero value of the forward rates. This avoids having to recalculate the drift
shifting the initial term structure and calculating the corresponding shift for each time step. This results in the following SDE,
in the price using the methodology provided in the previous sections. If we
denote the shifted price V̄0 and rate F̄m (0) then d(m Fk (t)) = σk (t)φ
(Fk (t)) m Fk (t)(µ k (t, Fk (0))dt + dZk (t))
⎛ ⎞
∂V0 V̄0 − V0 M
∂µk (t, F̄(0))
≈ . (53) + σk (t)φ (Fk (t)) ⎝ m Fj (t)⎠ , (61)
∂Fm (0) F̄m (0) − Fm (0) j=0
∂xj

In simple cases the payoff will be a smooth function of the initial curve t ≤ Tk−1 , ∀ k = 1, . . . , M.
making this “bump-and-revalue” approach stable enough for practical
purposes. However, in our case the price of a callable CMS Steepener is Glasserman and Zhao (1999) show that the errors resulting from this
not a smooth function of the initial term structure. The reason is the approximation are insignificant.
callable nature of the product. The decision to call the bond at a point in We can now calculate the delta of the product. Given an estimate of
time is binary and a small shift in the initial rates could therefore change the optimal exercise time η we have seen that the initial bond price can
^
this decision. The corresponding price change could be quite large since be represented as

WILMOTT magazine 113


 

η This term is very small especially for large l and can in practice be ignored but
Ṽ0 = E0 N
B−1
tη + B−1
ti CF i . (62) we include it since it comes at little computational cost as it can be calculated
i=1
before any simulation is performed. We can now evaluate all the relevant
The deltas can now be calculated as follows terms and calculate the deltas in equation (61) by Monte Carlo.
  Note that the approach described above is not applicable to G2++ model

−1 −1
N
m Ṽ0 = E0 m (Btη + Bti CFi ) since this is not a model of forward rates. In this case we therefore use (53).
i=1 The standard error of the delta estimate depends on the size of the rate shift
    (63)

η

η or perturbation F̄m (0) − Fm (0). Because of the discontinuity this might not
= EN0 [m B−1 N
t η ] + E0 m (B−1 N
ti )CF i ) + E0 B−1
ti m (CF i ) . be at the smallest possible level. We have experimented with several shifts
i=1 i=1
and found that a 1 basis point perturbation gives decent results.
Notice that when differentiating through we have ignored that the optimal
time η also depends on the initial forward rates. Piterbarg (2004a) gives a 5.2 Vegas
formal argument why this is reasonable based on the smooth pasting
conditions of the optimization problem. The first two terms are calculated Vega denotes the price sensitivity with respect to the volatility in the
as follows with t ∈ [tn , tn+1 ], market. In practice traders not only perform delta hedging but also vega
hedging in order to be hedged against moves in volatility. One can form a

n
τi
m (B−1 −1 vega hedging portfolio in several ways by using volatility dependent
t ) = −Bt m Fi (ti−1 ) (64)
1 + τi Fi (ti−1 ) products such as cap, floors and swaptions. We have initially chosen to
i=1
calibrate our models to a large segment of swaptions. However, finding
t − tn
−B−1 m Fn+1 (tn ). (65) the hedge portfolio of swaptions is difficult since swaption volatilies are
t
1 + (tn+1 − t)Fn+1 (tn )(1 + τn Fn+1 (tn )) not direct inputs to the versions of the LIBOR market model used in this
paper. One could proceed with the naive bump-and-revalue approach by
For details see Piterbarg (2004a). The last term in (63) is more difficult.
shifting a market price and then recalibrating the model. In our case this
For l > s we have
does not give meaningful results since the perturbation will be obscured

τi max((Sl (ti−1 ) − Ss (ti−1 )) ∗ m, f ) for η ≥ i > Tlockout by calibration error. One way to get meaningful swaption vegas would
CFi = (66) therefore be to start over and apply a local calibration procedure (see for
τi R for 0 ≤ i ≤ Tlockout
example Brigo and Mercurio (2006) or Pietersz and Pelsser (2006)) instead
where R is the fixed rate in the lockout period. of our more global approach. This procedure fits the model to a small
A standard way to approximate the swap rates is the following formula subset of swaptions without error. This subset could for example be the
due to Rebonato (2002), most liquid or the core swaptions. The downside of the local approach is
of course that you take the chance of fitting the model perfectly to prices
β (ti− 1 +l) β (ti− 1 +l)
  that might reflect noise such as illiquidity and non-synchronous trading.3
Sl (ti ) = wj (ti )Fj (ti ) ≈ wi (0)Fj (ti ) (67)
This avenue is beyond the scope of our paper and we therefore proceed
j=i−1 j=i−1
with calculating the sensitivities of the volatilities of the individual
β (ti− 1 +s) β (ti− 1 +s) rates which amounts to finding a vega hedge portfolio of caplets. This
 
Ss (ti ) = wj (ti )Fj (ti ) ≈ wi (0)Fj (ti ) (68) approach also has the advantage of having a more direct interpretation
j=i−1 j=i−1 with respect to forward rates than the swaption approach.
We proceed in a manner similar to the previous section. We calculate
β(t) = m if tm−2 < t ≤ tm−1 . (69) the price sensitivity of a unit shock to the volatility curve for each rate
and we denote each vega: ∂∂ σVm0 . This partial derivative can be calculated
This approximation can be justified by the fact that the variability in the using the pathwise approach in a similar manner as we did for the deltas
weights is very small compared to the variation in the forward rates. With in the previous section. Differentiating through and fixing the drifts we
this approximation we can calculate the delta for the long swap rate: now get the following SDE:
β (ti− 1 +l)
  
m (Sl (ti−1 )) = (m wj (0)Fj (ti ) + wj (0)m Fj (ti )). (70) ∂ ∂
d Fk (t) = φ (Fk (t))1{m=k} + σk (t)φ
(Fk (t)) Fk (t)
j=i−1 ∂σm ∂σm
From simple calculations we get that × (µk (t, Fk (0))dt + dZk (t)) + σk (t)φ (Fk (t))
⎛ ⎞ (72)
1 
M
m wj (0) = −τm wj (0) ∂µk (t, F̄(0)) ∂
1 − τm Fm (0) ×⎝ Fj (t)⎠ ,
⎛ ⎞ ∂xj ∂σm
(71) j=0
1
×⎝ β (ti− 1 +l) k ⎠.
1 + k=m j=i−1
1
1+τj Fj (0)
t ≤ Tk−1 , ∀ k = 1, . . . , M.

114 WILMOTT magazine


TECHNICAL ARTICLE 3


Note that ∂
Fk (0) = 0 ∀ m and δi 9.55% if i ≤ 2
∂ σm CFi = (76)
δi max[3[S20 (ti−1 ) − S2 (ti−1 )], 0.05%] if 23 ≥ i > 2,

∂  (φ
(Fj (t))(1 + τj Fj (t)) − φ (Fj (t))τj )τj ρk,j σj (t)
 =
∂xj xj =σj (1 + τj Fj (t))2 where δi is determined from a 30/360 day-count fraction. The payments are
(73) made on the dates listed in Table 1. The bond can be called by the issuer at
φ (Fj (t))τj ρk,j time t2 and onwards. If the bond is called the coupon on the call date is
+ 1{j=k} .
(1 + τj Fj (t)) paid as well as the principal which we for simplicity assume is 1.

Besides the extra indicator function term this is the same as in the previous
section. We can now calculate the vegas of the CMS Steepener bond as follows: 7 Model Calibration
  The concept of calibration is to fit the parameters in the model so they are
∂ ∂  η

Ṽ0 = EN0 (B−1


tη + B−1
ti CF i )
able to reproduce the relevant prices observed in the market with a certain
∂σm ∂σm amount of accuracy. There is naturally a tradeoff between the number of
i=1

   η  calibration instruments used and the fitting quality. We therefore make a


∂ −1 
η
∂ 
−1 ∂ selection of calibration instruments that we deem most relevant for the
= EN0 B + EN0 (B−1 )CF i + E N
B CF i .
∂σm tη ∂σm ti 0 ti
∂σm the callable CMS Steepener described in the previous section.
i=1 i=1
From Jyske Bank we have obtained market quotes on Euro denominated
(74)
EURIBOR caps and swaptions. All market quotes are in Black-76 implied
volatility. The swaptions in the European market are settled annually and
The three terms are calculated with Monte Carlo as in the previous section
the caps are settled either quarterly or semiannually. Both the LMMs and
but with m replaced by ∂ σ∂ m .
the G2++ model are marked to market with the zero curve available from
Since prices are quoted in terms of Black-76 implied volatility, it is
Datastream. All rates and quotes are from January 25, 2005–the day the
necessary to express the vegas in this metric. By using the chain rule of
bond was issued.
partial differentiation we have

∂ Ṽ ∂σBS76 ∂ Ṽ 7.1 G2++


=
∂σm ∂σm ∂σBS76 The G2++ model has a limited number of parameters which again limits
(75) the amount of calibration instruments. Since the bond we are aiming to
∂ Ṽ
∂ Ṽ ∂ σm
⇒ = ∂ σ BS 76
price depends mainly on swap rates we choose to calibrate to swaptions
∂σBS76 ∂ σm only. A subsection of the swaption matrix of implied volatilities can be
seen in Table 2. The bond price depends among other things on the twenty
The numerator is calculated in 74. The denominator cannot be calculated year swap rate that sets in 11 years. This means that swaptions with a
in closed form for the CEV model but it can easily be found by numerical timespan that stretches further than 31 years are of little relevance to the
differentiation. For the standard LMM case we of course have σm = σBS76 bond price. Excluding this portion of the swaption matrix leaves us with
so the vegas are given in the proper metric from 74 directly 106 datapoints to fit.
In the G2++ model there is, in our setup, no proper way to calculate The G2++ model has a closed form solution for the swaption price.
vegas. A pathwise approach as well as a bump-and-revalue approach will This formula is shown in Appendix B.1. Following the market convention
be too noisy since all volatility parameters in the G2++ model have a
global influence.
Table 1: Payment dates
6 Case Study: The Dexia Dannevirke t1 20-06-2005 t13 20-06-2011
2005/2016 CMS Steepener t2 20-12-2005 t14 20-12-2011
t3 20-06-2006 t15 20-06-2012
In this section we further investigate the particular issue of “Dexia t4 20-12-2006 t16 20-12-2012
Dannevirke”. This was the largest structured bond issue ever seen in t5 20-06-2007 t17 20-06-2013
Denmark with DKK 2.4 billion (about EUR 320 million or USD 425 million) t6 20-12-2007 t18 20-12-2013
sold. The success of this particular issue incited the larger share of all t7 20-06-2008 t19 20-06-2014
Nordic banks to issue a string of similar products. The bond was issued on t8 20-12-2008 t20 20-12-2014
January 25, 2005 which we denote time zero. The initial fixed coupon was t9 20-06-2009 t21 20-06-2015
set at 9.55% and the floating coupons were determined as 3 times the t10 20-12-2009 t22 20-12-2015
floored (at 5bps) difference between the 20 and the 2 year EURIBOR swap t11 20-06-2010 t23 25-01-2016
rate determined in arrears at the previous coupon time. The following t12 20-12-2010
^
payments CFi were specified,

WILMOTT magazine 115


Table 2: Swaption Matrix (of implied volatilities in percent)
Maturity tiM /Tenor tjT 1 2 3 4 5 6 7
1 22.3 21.8 21.1 20.5 19.7 18.6 17.8
2 22.2 21.3 20.2 19.1 18.1 17.4 16.7
3 20.6 20.0 19.1 17.9 17.1 16.4 16.0
4 19.5 18.9 17.9 16.8 16.2 15.7 15.4
5 18.3 17.7 16.8 15.9 15.4 15.0 14.8
10 15.1 14.5 13.9 13.3 13.0 12.9 12.8
15 13.0 12.5 12.0 12.0 11.9 11.8 11.8
20 11.9 11.6 11.6 11.6 11.6 11.6 11.7
25 11.4 11.4 11.4 11.5 11.5 11.5 —
30 11.0 — — — — — —
Maturity tiM /Tenor tjT 8 9 10 15 20 25 30
1 17.1 16.5 16.1 14.5 13.3 12.9 12.6
2 16.2 15.8 15.4 14.2 13.1 12.7 —
3 15.6 15.3 15.1 13.9 13.0 12.6 —
4 15.1 14.8 14.6 13.8 12.9 12.5 —
5 14.6 14.4 14.3 13.4 12.7 12.4 —
10 12.7 12.6 12.5 13.0 11.4 — —
15 11.8 11.8 11.8 11.3 — — —
20 11.7 11.7 11.7 — — — —
25 — — — — — — —
30 — — — — — — —

we translate the model prices into volatilities by inverting the Black-76 7.2 LIBOR Market Models
formula. We then minimize the distance between model volatilities and
˜ = (ã, b̃, σ̃ , η̃, ρ̃),
market volatilities to get the optimal set of parameters,  The LIBOR Market Model in its currently stated form does not readily
using the following Root-Mean-Square distance metric,4 lend itself to calibration. First recall the general market model under the
forward numeraire. With tenor structure T1 < · · · < TM we have

  2 dFk (t) = σk (t)φ (Fk (t))dZk (t), t ≤ Tk−1 , ∀ k = 1, . . . , M, (78)
 1  i,j i,j
σmarket − σmodel (a, b, σ, η, ρ)
˜ = arg min 
 
(a,b,σ,η,ρ ) 106 i,j
1≤i≤10,1≤j≤10,tiM +tjT ≤31
σmarket where

(77) dZi (t)dZj (t) = ρi,j dt. (79)

i,j
Here σmarket is the implied volatility of a payer swaption with maturity date Since we have semiannual payments in the bond we are currently analyzing,
i,j
tiM and swap length (tenor) equal to tjT , and σmodel (·) is the corresponding we choose forward rates with equidistant maturities of 6 months, meaning
model implied volatility from inverting formula (92) in the Appendix. Tk − Tk−1 = 0.5. The furthest maturity we need to deal with in pricing is
Minimization is done in the Ox programming language using the 31 years. This means that TM = 31 and the number of rates is 62. In the
Simulated Annealing optimization procedure to get starting values and log-normal case, ie. when φ (x) = x, we have only the volatilities σk (t) and
the standard BFGS optimization routine to refine the solution. As Table 3 the correlations ρi,j to determine. As recommended by Rebonato (2002)
shows the errors are tolerable for smaller tenors but increase up to 13% for and Brigo and Mercurio (2006) we use the following parametric form for
larger tenors and swaption maturities. This can be expected from a model the volatilities,
with only five parameters and 106 datapoints to fit. Note from the table  
that we estimate a correlation between the factors equal to –0.891 similar σk (t) = k [a(Tk−1 − t) + d]e−b(Tk− 1 −t) + c . (80)
to the case investigated in Brigo and Mercurio (2006). Since the factors are
theoretical constructs this number does not have a direct interpretation. This specification is very tractable. It allows the volatilities to exhibit the
But with a value different from 1 or −1 it does mean that the model is typical humped shape observed in the market. It is also time-inhomogeneous
able to pick up some of the information on the correlation between rates since the term k is forward rate dependent which allows a better initial fit
available in the swaption matrix. to the market. We do however set the ’s pairwise equal:

116 WILMOTT magazine


TECHNICAL ARTICLE 3

Table 3: Calibration errors (in percent) in G2++


Maturity tiM /Tenor tjT 1 2 3 4 5 6 7
1 –0.41 –0.37 –2.28 –1.92 –1.79 –3.16 –3.35
2 –0.64 –1.54 –3.47 –4.81 –5.69 –5.32 –5.49
3 –2.72 –1.72 –2.29 –4.43 –4.85 –5.22 –4.25
4 –1.10 –0.01 –1.42 –3.92 –3.96 –3.89 –2.85
5 0.11 0.74 –1.08 –3.37 –3.62 –3.51 –2.19
10 2.25 0.36 –1.79 –4.01 –3.87 –2.30 –0.78
15 0.42 –1.40 –3.78 –2.04 –1.09 –0.26 1.35
20 –1.66 –2.75 –1.40 –0.24 1.15 2.66 4.77
25 –2.76 –1.95 –0.65 1.63 3.11 4.64 —
30 –3.91 — — — — — —
Maturity tiM /Tenor tjT 8 9 10 15 20 25 30
1 –3.46 –3.52 –2.73 0.36 1.84 5.77 8.46
2 –4.99 –4.26 –3.81 0.92 2.25 5.63 —
3 –3.67 –2.68 –1.24 2.39 4.32 7.17 —
4 –2.02 –1.38 –0.21 5.13 6.32 8.76 —
5 –1.00 0.07 1.77 5.48 7.38 10.20 —
10 0.54 1.69 2.74 13.26 6.82 — —
15 2.86 4.18 5.49 7.39 — — —
20 6.10 7.40 8.66 — — — —
25 — — — — — — —
30 — — — — — — —
Parameter values are a = 3.47, b = 0.0461, σ = 0.0537, η = 0.00842 ρ = 0.891. With data calibrated to the Swaption implied volatilities on January 28, 2005.

k = k+1 , ∀ k odd. (81) comparison with the G2++ model we choose to calibrate to the same
swaption data as in the previous section.5
This eases calibration as it reduces the number of parameters to estimate. In A practical issue often ignored is the fact that the LMM models
unconstrained calibration a common problem is that the parameters oscillate semiannual rates whereas the EURIBOR swaptions we calibrate to are
between forward rates yielding an unrealistic evolution of the volatility (see settled annually. Throughout the paper we therefore use the following
Brigo and Mercurio (2006)). To avoid this we impose a regularity constraint, adjusted swap rate for a Tα × (Tβ − Tα ) -swap.

|k+1 − k | < 0.1, ∀ k. (82) P(t, Tα ) − P(t, Tβ ) β

Sα,β (t) = (α −β )/2 = wi (ti )Fi (t), (84)


−b(T k− 1 −t) i=1 P(t, Tα+2i ) i=α+1
The second term, [a(Tk−1 − t) + d]e + c, is homogeneous through
time and depends only on time to maturity of the forward rate. This term
determines the shape of the volatility structure. P(t, Ti )
wi (t) = (β −α)/2 . (85)
The correlations are parametrized through the well-known approach i=1 2τi P(t, Tα+2i )
justified in Schoenmakers and Coffey (2000),
This swap rate is described in Schoenmakers (2002). As before we minimize

|j − i| the distance between the market quoted volatilities and the model implied
ρi,j = exp − − (ln(ρ∞ )
M−1 volatilities using a root mean square distance metric, yielding the optimal
i2 + j2 + ij − 3Mi − 3Mj + 3i + 3j + 2M2 − M − 4 parameters:
+ η1
(M − 2)(M − 3)
 ˜ = (ã, b̃, c̃, d̃, {˜ k }k≥1 , η˜1 , η˜2 , ρ˜∞ ) =

i + j + ij − Mi − Mj − 3i − 3j + 3M + 2
2 2
−η2
(M − 2)(M − 3) (83) 
 
 1
 106
 1≤i≤10,1≤j≤10,tiM +tjT ≤31
This particular approach yields a correlation matrix that is flexible arg min   2 . (86)

enough to have the particular characteristics observed in several empirical  i,j i,j
σ market −σ model (a,b,c,d,{ k },η 1 ,η 2 ,ρ ∞ )
×
^
i,j
studies, see e.g. the survey in Rebonato (2002)). In order to facilitate direct σ market

WILMOTT magazine 117


Table 4: Calibration errors (in percent) in the log-normal LMM
Maturity tiM /Tenor tjT 1 2 3 4 5 6 7
1 –0.78 –0.58 0.64 2.07 2.26 0.73 0.65
2 –0.27 1.06 0.56 –0.54 –1.46 –0.83 –0.69
3 –0.64 0.98 0.88 –1.00 –0.81 –0.60 –0.59
4 –0.60 0.93 0.37 –0.87 0.09 –0.74 –0.45
5 –1.01 0.87 1.16 0.50 –0.70 –1.30 –0.35
10 0.42 0.36 –0.39 –0.99 –2.50 –1.88 –0.84
15 0.27 0.17 –0.29 –0.66 –0.72 –0.87 0.12
20 0.17 0.05 –0.14 –0.37 –1.44 –0.69 0.57
25 0.12 –0.09 –0.24 –0.29 0.08 0.80 —
30 –0.10 — — — — — —
Maturity tiM /Tenor tjT 8 9 10 15 20 25 30
1 0.70 –0.26 –0.29 –0.26 –2.06 –1.20 –1.10
2 –1.13 –1.22 –1.35 0.33 –1.57 –1.31 —
3 –0.80 –0.35 0.68 1.40 –0.12 –0.34 —
4 –0.06 0.26 1.37 3.18 1.08 0.40 —
5 0.59 1.70 2.37 2.82 1.24 1.13 —
10 0.39 0.35 0.55 7.20 –3.80 — —
15 0.50 0.72 0.49 –2.52 — — —
20 0.82 0.69 0.96 — — — —
25 — — — — — — —
30 — — — — — — —
Average Error = 1.74 % Max error = 7.19 %

i,j
Here σmarket is the Black-76 implied volatility of a payer swaption with
i,j
maturity date tiM and swap length (tenor) equal to tjT , and σmodel (·) is the Table 5: Parameters in the log-normal LMM
corresponding model Black-76 volatility from formula (108) in the a 0.033807 η1 0.881421
Appendix. We use the constrained optimization algorithm of Lawrence b 0.904988 η2 0
and Tits (2001) to obtain the minimum distance parameters. In Table 4 c 0.141077 ρ∞ 0.414192
we report the corresponding pricing errors and in Table 5 the estimated d 0.000169
parameters. We obtain an average error of 1.7% which might seem high
i= i i= i
when using 39 parameters to fit 106 datapoints. But 32 of the parameters
1, 2 1.374 33, 34 0.849
(the ’s) are heavily constrained and have only a local influence.
3, 4 1.473 35, 35 0.795
If we look at the correlation parameters we see that the second parameter
5, 6 1.473 37, 38 0.884
η2 has reverted to zero as in the calibration experiments performed in
7, 8 1.388 39, 40 0.840
Schoenmakers and Coffey (2000). We also observe that ρ∞ = 0.41. This param-
9, 10 1.326 41, 42 0.833
eter is interpreted as the limiting correlation between rates when distance
11, 12 1.260 43, 44 0.798
between maturity is increased. Looking at the 3d-plot of the correlation matrix
13, 14 1.177 45, 46 0.830
in Figure 1 we can observe two important features. The first is decorrelation,
15, 16 1.077 47, 48 0.838
i.e. the fast reduction in dependency when the distance between maturity is
17, 18 1.007 49, 50 0.883
increased. The second is the increase in interdependency between equidistant
19, 20 1.084 51, 52 0.799
rates as maturity is increased. Graphically this means that the steepness of the
21, 22 1.045 53, 54 0.807
surface when moving away from the diagonal is decreasing.
23, 24 0.985 55, 56 0.812
25, 26 0.937 57, 58 0.843
7.3 The CEV LMM 27, 28 0.864 59, 60 0.801
29, 30 0.957 61, 62 0.774
We also perform a market calibration of the CEV LIBOR market model which
31, 32 0.906
has φ = xp . As in the previous section we assume the same parametric form

118 WILMOTT magazine


TECHNICAL ARTICLE 3

estimated in step 1 and recalibrate (a, b, c, d, {k }k≥1,η 1 ,η 2 ,ρ ∞ ) to swaption


implied volatilities by using formula (117) using the values from step 1 as
1.0 starting values in the optimization procedure.7 Again optimization is done
using the standard root mean square distance metric. The errors we get in
0.8 Table 6 are tolerable and comparable to the LMM case.
ρi,j

0.6
8 Numerical Results
30 8.1 Simulation Details
20 25 30
10 15 20 We use an Euler discretization of the forward LIBOR dynamics under
Ti 5 10
Tj the discrete money market measure given in the SDE in (28). We choose
stepsizes so each coupon payment date/exercise time t1 , . . . , t23 are hit.
Figure 1: Implied Correlation We use 4 Euler steps between each coupon date totaling 92 steps per path.
In the the pre-simulation step we use 25000 antithetic paths (total 50000)
for volatilities and correlation. The difference is that now we have an extra to estimate the parameters of the exercise strategy. In the pricing step
parameter p to determine. This parameter is highly dependent on the skew we use another 25000 (50000) paths. The sub-simulations involved in
of the implied volatilies of interest rate options. Ideally we would like to calculating the lower bound are extremely slow and we therefore estimate
determine p from OTM-Swaptions, since swap rates are our main concern. this with only 25 (50) antithetic paths for 200 (400) sub-simulations. To
But data for these products are not readily available as the market has further increase speed we have also employed a so-called Runge-Kutta drift
very low liquidity. We therefore follow Hull and White (2001) and use approximation in the sub-simulations only. This significantly increases
OTM-Caplets instead.6 We have access to prices quoted in Black-76 volatilities the speed as we avoid having to continuously calculate the state dependent
for 13 different strikes ranging from 1.5% to 10%. The procedure is in two drift for each Euler step. The errors from this approximation are in general
steps. First we calibrate p and (a, b, c, d, {k }k≥1,η 1 ,η 2 ,ρ ∞ ) from OTM-Caplet extremely small (see for example Rebonato (2002) and Glasserman and
prices using formula (111) in the Appendix. Secondly we fix p at the value Zhao (1999)). Zero coupon bond prices setting on non-tenor dates are

Table 6: Calibration errors (in percent) in the Extended/CEV LMM


Maturity tiM /Tenor tjT 1 2 3 4 5 6 7
1 –0.88 –0.45 0.85 2.19 2.27 0.62 0.56
2 –0.38 1.15 0.66 –0.52 –1.57 –0.90 –0.77
3 –0.69 1.04 0.93 –1.08 –0.82 –0.63 –0.62
4 –0.64 0.97 0.33 –0.80 0.13 –0.69 –0.46
5 –1.02 0.82 1.33 0.64 –0.56 –1.23 –0.34
10 0.47 0.37 –0.36 –0.98 –2.51 –1.93 –0.92
15 0.32 0.19 –0.28 –0.62 –0.70 –0.90 0.07
20 0.20 0.08 –0.14 –0.38 –1.40 –0.70 0.53
25 0.11 –0.07 –0.24 –0.26 0.12 0.79 —
30 –0.08 — — — — — —
Maturity tiM /Tenor tjT 8 9 10 15 20 25 30
1 0.60 –0.34 –0.39 –0.30 –2.03 –1.11 –0.97
2 –1.20 –1.31 –1.45 0.30 –1.53 –1.22 —
3 –0.86 –0.43 0.61 1.38 –0.08 –0.25 —
4 –0.11 0.22 1.33 3.18 1.13 0.49 —
5 0.58 1.69 2.36 2.85 1.32 1.23 —
10 0.30 0.28 0.48 7.17 –3.80 — —
15 0.43 0.65 0.45 –2.56 — — —
20 0.77 0.65 0.93 — — — —
25 — — — — — — —
30 — — — — — — —
Average Error = 1.76 % Max error = 7.20 %
^

WILMOTT magazine 119


Table 7: Parameters in the extended/CVV LMM 8.2 Prices
In Table 8 we have calculated the theoretical price of the Dexia Dannevirke
a 0.047104 η1 0.922098
bond with a 10000 bp principal. The value of the call option was calculated
b 1.410679 η2 0
by finding the difference in values to the price of a similar non-callable
c 0.092664 ρ∞ 0.395819
bond valued using the same random numbers which can be done at almost
d –0.01779 p 0.665
zero computational expense. We used a second order polynomial of the
i= i i= i core swap rate, the 6m LIBOR rate and the 20 and 2 year swap rate with all
1, 2 0.610 33, 34 0.466 rates setting on the respective exercise dates to calculate the exercise strategy.
3, 4 0.691 35, 35 0.452 We have experimented with many other rates such as medium term swap
5, 6 0.722 37, 38 0.469 rates and other LIBOR rates as well as higher order polynomial functions.
7, 8 0.698 39, 40 0.468 In all of these cases the duality gap was larger, in some cases higher than
9, 10 0.683 41, 42 0.459 100 bp. The results of one of these experiments are in Table 9 in the
11, 12 0.663 43, 44 0.435 Appendix where we have used a 4th order polynomial function as well as
13, 14 0.628 45, 46 0.450 included more rates. From this table we can see that the duality gaps
15, 16 0.580 47, 48 0.451 have increased with an order of almost 8, and the option prices have been
17, 18 0.570 49, 50 0.479 approximately halved. These results corroborate the existing notion in the
19, 20 0.557 51, 52 0.432 literature (see for example Piterbarg (2004b)) that basis functions should be
21, 22 0.570 53, 54 0.430 kept simple.
23, 24 0.534 55, 56 0.435 We have also supplied a 95% confidence interval which is calculated
25, 26 0.522 57, 58 0.431 as follows:
27, 28 0.469 59, 60 0.446 ⎡  ⎤
29, 30 0.520 61, 62 0.408 s2V s2 sV
31, 32 0.498 ⎣Ṽ0 − ˜ 0 − z1−α/2 + ; Ṽ0 + z1−α/2 √ ⎦ , (87)
K K1 K

calculated using the constant interpolation technique described in where sV and s are the standard deviations for the upper bound and
Piterbarg (2004a). the duality gap respectively. In Table 8 we see that the prices for the
The short rate dynamics of the G2++ model is simulated in all cases different models are statistically different from each other. The fact
without discretization error under the terminal forward measure with that the log-normal model gives higher prices than the CEV model is
numeraire P(t, t23 ). We use the same amount of paths as in the LIBOR expected since the non-central chi-square distribution of the CEV
Market Models. For details on this we refer to Brigo and Mercurio (2006). model gives more weight to the lower end of the distribution of forward

Table 8: Prices
Model Price Call option Duality gap 95% CL
Log-normal LMM 9315.45(2.0) 91.02(1.6) 12.74(1.6) [9297.68 ; 9319.37]
CEV-LMM 9292.54(1.77) 112.29(1.8) 13.79(1.8) [9273.80 ; 9296.01]
G2 + + 9091.80(0.8) 127.42(1.3) 17.09(2.1) [9070.30 ; 9093.36]
All prices in basis points. Standard errors in parentheses.
All prices are calculated with the same basis function which is a second order polynomial of four rates. The 6m forward LIBOR, core swap rate and 2 and 20 year swap rate.
All rates setting on each exercise date.

Table 9: Prices with an alternative stopping strategy


Model Price Call option Duality gap 95% CL
Log-normal LMM 9382.49(2.2) 24.02(1.1) 96.88(8.5) [9268.40 ; 9386.81]
CEV-LMM 9344.51(1.9) 46.83(1.4) 91.05(8.1) [9237.15 ; 9348.23]
G2++ 9158.98(0.9) 60.24(1.3) 106.09(6.1) [9040.80 ; 9160.74]
All prices are calculated using a basis function consisting of a 4th order polynomial of the 6m forward LIBOR, the core swap rate, the 2 and 20 year swap rates, and the
20 year forward LIBOR. All rates setting on exercise dates.

120 WILMOTT magazine


TECHNICAL ARTICLE 3

Table 10: Call time distribution


Times LMM CEV-LMM G2++ Times LMM CEV-LMM G2++
t2 0.68% 0.64% 0.54% t13 0.75% 0.84% 0.88%
t3 1.13% 1.10% 1.26% t14 0.67% 0.81% 0.99%
t4 1.14% 1.12% 1.30% t15 0.84% 0.88% 1.02%
t5 1.24% 1.25% 1.32% t16 1.01% 0.99% 1.00%
t6 1.08% 1.17% 1.14% t17 1.13% 1.14% 1.04%
t7 1.06% 1.14% 1.19% t18 1.38% 1.32% 1.16%
t8 0.75% 0.83% 1.09% t19 1.16% 1.38% 1.16%
t9 0.74% 0.88% 0.96% t20 1.43% 2.02% 1.46%
t10 0.71% 0.79% 1.06% t21 1.88% 2.30% 2.56%
t11 0.74% 0.80% 0.98% t22 0.65% 1.00% 5.09%
t12 0.67% 0.76% 0.87% t23 79.16% 76.87% 71.91%

rates. The normal distribution of the G2++ model implies even more the fact that higher initial interest rates make the short CMS rate higher
weight to the lower end than the CEV model and therefore yields the causing smaller coupons. We can see that the effect begins to shift around
smallest price of the three. rates setting in 13 years or more which is natural since these rates only affect
In Table 10 we have tabulated the probability of exercise at each exercise the long CMS rate and not the short one. Overall the delta ratios are very
time point. The overall distribution is very similar for the 3 models having close to each other for the different models. This would suggest that all
at most 28.09% percent chance of early exercise. This is consistent with models are appropriate for delta hedging.
papers such as Andersen and Andreasen (2001) and Svenstrup (2005) which
investigate the exercise strategies from lower factor models in Bermudan 8.4 Vegas
swaptions. The only thing that stands out is the large discrepancy at the
final exercise time t22 , where the G2++ model has more than 5 times the The vegas in Figure 3 are calculated with the pathwise approach for
higher probability of exercise. One reason for this discrepancy is that there is the LMM and CEV-LMM for the volatility of all forward rates. We notice
less than one month between the last exercise time t22 and maturity at t23 . a somewhat complex vega profile with large differences in the two
The small day-count fraction makes the coupon payment at t23 very small models. The specific complex pattern is mainly due to the negative
and therefore the exercise decision is more idiosyncratic and prone to error. vega component in the product. As the volatility of the 2 year CMS rate
Perhaps one of the most interesting things to observe from this section increases the price of the coupons decreases. However, the volatility of
is the implied estimate of the profit the issuing bank must have made the 2 year CMS rate cannot increase without also affecting the 20 year
from selling these bonds. The bonds where sold at par (10,000bps) and swap, and hence the effects will offset each other at some points.
DKK 2.4 billion worth of bonds where sold. Looking at the lower end of the Looking closer at Figure 3 we see that rates of low maturity have close
confidence interval for the G2++ model and the upper end for the log to zero vega but as maturity increases the vegas begin to increase as
normal LMM in Table 8 suggests that the profit margin for Forbank ranges the rates affect more and more coupon payments. This effect is offset
between 6.8% to 9.07%. In monetary terms this translates to a profit ranging by the negative vega component in coupon payments which causes the
from DKK 163.2 million to DKK 217.6 million. Considering that this profit is dip around rates of 7-9 years maturity. The vegas increase again around
riskless if a proper hedge portfolio is formed, the amounts are considerable. 11 years which is the maturity date of the bond, but drop as maturity
If we make the somewhat unfair and unscientific extrapolation of this increases further due to the decreasing effect they have on coupons.
result to a USD 50 billion market of CMS spread related products (see
Jeffery (2006)) then these type of products have made considerable profits
to the issuers.
9 Conclusions
This paper has presented a methodology to price and risk manage the
8.3 Deltas so-called callable CMS Steepener structured bond. We have used the
simple G2++ model for the short rate as well as two different LIBOR
In figure 2 we show the delta profile for the product for each forward rate market models. We have adapted the Least-Squares Monte Carlo procedure
in the term structure. All deltas are in basis points. The LIBOR market and shown how looking at the bond in its entirety can avoid a regression
model deltas are calculated using the pathwise approach and the G2++ step. We have also adapted the procedure of Andersen and Broadie (2004)
model deltas are calculated by the standard bump-and-revalue approach to give a downward biased price estimate of issuer callable bonds. This is an
with a 1bp perturbation. The effect of shifting the initial term structure important supplement to the upward biased estimate of the LSMC procedure
has two opposite effects. The downwards effect is due to two reasons. One is as the gap between the two determines the quality of the exercise strategy. We
^
simple and due to higher discounting of the coupons; the other is due to then proceeded to show how hedge ratios can be calculated.

WILMOTT magazine 121


1

0.5

5
5
5
5
5
5
5
5
5
5
.5

12 5
13 5
14 5
15 5
16 5
17 5
18 5
19 5
20 5
21 5
22 5
23 5
24 5
.5

26 5
27 5
28 5
29 5
30 5
31 5
.5
0.
1.
2.
3.
4.
5.
6.
7.
8.
9.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
10
11

25
-0.5

-1

-1.5

-2
G2
CEV-LMM
LMM

Figure 2: Deltas

2.00%

1.50%

1.00%

0.50%

0.00%
5
5
5
5
5
5
5
5
5
5
.5

12 5
13 5
14 5
15 5
16 5
17 5
18 5
19 5
20 5
21 5
22 5
23 5
24 5
25 5
26 5
27 5
28 5
29 5
30 5
31 5
.5
0.
1.
2.
3.
4.
5.
6.
7.
8.
9.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10
11

-0.50%

-1.00%
CEV-LMM
LMM

Figure 3: Vegas

We have applied the theoretical insights on a particular callable shown that our specific exercise strategy gives duality gaps of a tolerable
CMS spread bond. We have found that the different models of the size. An interesting bi-product of our analysis is an estimate of the
term structure yield significantly different prices when calibrating to sizeable profits the banks must have made from selling this particular
the same data and using the same exercise strategy. We have also bond.

122 WILMOTT magazine


TECHNICAL ARTICLE 3

To the authors’ knowledge this paper is the first about these particular
products and we have therefore opted to use relatively simple models of
B Calibration formulas
the term structure of interest. However, the use of stochastic volatility
and/or jumps in the forward rates are becoming more and more normal B.1 G2 ++ Model
in both academia and practice (see Hagan et al. (1997) and Glasserman
and Kou (2003)). Our results on pricing could easily be extended to these The formula for a payer swaption is rather involved and it is stated here
cases. However, in the area of hedging there are still many unanswered for completeness. For the proof we refer to Brigo and Mercurio (2006).
questions in the more advanced model setups. Consider an interest rate swap with nominal value of 1 and strike K̄
Future work on CMS spread callables could be to evaluate the hedging and payments at t1 < . . . .. < tn and denote τi = Ti+1 − Ti . Set ci = K̄τi for
performance of different models over time. At lot of work has been done i = 1 . . . .. < n − 1 and cn = τn K̄. The price of a payer swaption at t = 0
recently on simpler products such as European and Bermudan swaptions which gives the right to enter the swap at t0 < t1 is
(see Pietersz and Pelsser (2006) and the references therein) and the conclusions
 ∞ x− µx 2
have mostly been that properly calibrated simple models perform very well e−1/2( σ x )
PS(0, K̄) = √ [(−h1 (x))
against more complicated alternatives. Our results for delta hedging also point −∞ σx 2π
 (92)
in that direction but further analysis are still needed. n
− κ i (x)
λi (x)e (−h2 (x)) dx,
i=0
A Proof of Theorem 1 where
For an arbitrary finite submartingale we have
ȳ − µy 2
ρxy (x − µx )
h1 (x) = − ! (93)

η
σy 1 − ρxy
2 σx 1 − ρxy
V0P /N0 = inf E0 [Nt−1
N
η
+ Nt−1
i
CFi )]
η∈i
i=0


η
h2 (x) = h1 (x) + B(b, t0 , ti )σy 1 − ρxy
2 (94)
= inf EN0 [Nt−1
n
+ Nt−1
i
CFi ) − Mη + Mη ]
η∈i
i=0
(88)

η
λi (X) = ci A(t0 , ti )e−B(a,t0 ,ti )x (95)
≥ inf EN0 [Nt−1
η
+ Nt−1
i
CFi ) − Mη ] + M0
η∈i
i=0
   "

η
N
≥ E0 min Nt−1 + Nt−1 CFi − Mη + M0 . κi (x) = −B(b, t0 , ti ) µy − 1/2(1 − ρxy
2
)σy2 B(b, t0 , ti )
η∈i η i
(96)
i=0 x − µx
+ρxy σy ,
σx
The first inequality follows from from the optional sampling theorem and
for submartingales (see for example Hoffmann-J% orgensen (1994)).
Let us now set Mn = Vtn /Ntn where M0 = V0P /N0 . We first observe Mn is a  
σ2 ση σ2
submartingale as µx = − 2
+ ρ [1 − e−at 0 ] + 2 [1 − e−2at 0 ]
a ab 2a
(97)
  ρσ η −t0 (a+b)
Vtn 
n + [1 − e ]
Mn = = min Nt−1 + Nt−1 CFi , ENti [Nt−1 Vti+ 1 ] (89) b(a + b)
Ntn n i i
i=0
 
η2 ση η2 ρσ η
≤ Eti [Nt−1
N
V ] N
= Etn [Mn+1 ]. (90) µy = − +ρ [1 − e−bt 0 ] + [1 − e−2bt 0 ] + [1 − e−t0 (a+b) ]
n +1 t n+ 1 b2 ab 2b2 a(a + b)
(98)

Plugging this into the dual problem gives us #


  1 − e−2at 0

η
Vtη σx = σ (99)
V0D ≥ V0P + N0 EN0 min (Nη−1 + Nt−1 CFi − ) . (91) 2a
η∈i i
Ntη
i=1 
1 − e−2bt 0
The second term in the above equation is positive since the exercise value σy = η (100)
of the bond will always be larger than or equal to the continuation value 2b
when following an optimal strategy. We therefore have V0D ≥ V0P . Taking
the infimum of (88) we get V0D ≤ V0P which proves the duality relation ρσ η
[1 − e−(a+b)t ].
0
^
ρxy = (101)
V0D = V0P attained at the optimal submartingale Mn∗ = Vtn /Ntn . (a + b)σx σy

WILMOTT magazine 123


B.2 LIBOR Market Models European market that are annually settled. The weights in the above formula
is therefore determined as follows
Caplet prices in the log-normal LIBOR market model with tenor structure
T = T1 < · · · < TM coincide with the Black (76) formula. The price of a P(t, T̃i )
caplet paying at time Ti the difference between the strike K̄ and the forward wi (t) = (β −α)/2 . (110)
2τi P(t, T̃α+2i )
rate setting at Ti−1 is i=1

For further details we refer to Schoenmakers (2002).


Cpl(0, Ti−1 , Ti , K̄, v) = P(0, Ti )τi Ei (Fi (Ti−1 − K̄)+ ) (102) In the extended LMM cap prices can be derived in closed form. With a
CEV parameter 0 < p < 1 we have

= P(0, Ti )τi [Fi (0)(d1 ) − K̄(d2)] (103)


Cpl(0, Ti−1 , Ti , K̄) = P(0, Ti )τi Ei ((Fi (Ti−1 ) − K̄)+ ) (111)
where
= P(0, Ti )τi [Fi (0)(1 − χ 2 (a, b + 2, c)) − K̄χ 2 (c, b, a)],
log(Fi (0)/K̄) + v2 /2
d1 = (104) (112)
v

log(Fi (0)/K̄) − v2 /2 where


d2 = (105)
v
K̄ 2(1−p)
a= (113)
 T i− 1 (1 − p)2 v
v= σi2 (t)dt. (106)
0
1
b= (114)
Swaption prices in the log-normal LMM cannot be derived in closed form. 1−p
However, a well established approximation due to Rebonato (2002) can be
used. Consider again an interest rate swap with nominal value of 1 and Fi (0)2(1−p)
strike K̄ and payments at the tenor structure T̃ = T̃α+1 < . . . . . . < T̃β and c= (115)
(1 − p)2 v
denote τ̃i = T̃i+1 − T̃i . The price of a payer swaption at time t = 0 which
gives the right to enter the swap at T̃α is  T i− 1
v= σi (t)2 dt, (116)
0
PS(T, K)(0) = Cα,β (t)[Sα,β (0)(d1 ) − K̄(d2 )], (107)
where χ 2 (·, x, y) is the distribution function for a non-central χ 2 -distributed
where random variable with non centrality parameter value x and y degrees of
freedom. For standard approximations of this function we refer to classical
log(Sα,β (0)/K̄) + vα,β
2
/2 textbook of Johnson and Kotz (1973). Again we use an approximation due to
d1 = Andersen and Andreasen (2000) to find the swaption value:
vα,β
log(Sα,β (0)/K̄) − vα,β
2
/2
d2 = (108) PS(T, K) = Cα,β (0)[Sα,β (0)(1 − χ 2 (a, b + 2, c)) − K̄χ 2 (c, b, a)], (117)
vα,β

β 
wi (0)wj (0)Fi (0)Fj (0)ρi,j T̃α where
2
vα,β = σi (t)σj (t)dt,
Sα,β (0) 0
i,j=1 K̄ 2(1−p)
a= (118)
(1 − p)2 v

where Cα,β (0) = i=α+1 τi P(0, T̃i ) is the present value of a series of basis
points also referred to as the swaption numeraire. The weights wi follow 1
b= (119)
from the fact that the swap rate can be written as a weighted sum of 1−p
forward rates,
Sα,β (0)2(1−p)

β c= (120)
(1 − p)2 v
Sα,β (t) = wi (t)Fi (t). (109)
i=α+1


β p 
In this paper the tenor structure of the LIBOR Market Model is semiannual wi (0)wj (0)Fi (0)p Fj (0)ρi,j T̃ α
2
vα,β = 2p
σi (t)σj (t)dt. (121)
and this does not coincide with the tenor structure of the swaptions in the i,j=1
Sα,β 0

124 WILMOTT magazine


TECHNICAL ARTICLE 3

Acknowledgments Brigo, D. and F. Mercurio (2006). Interest Rate Models, Theory and Practice, 2nd Edition.
Springer Finance.
Financial support from the Danish Mathematical Finance Network is gratefully ac- Carriére, P. (1996). Valuation of early-exercise price of options using simulations and
knowledged. The authors are grateful for helpful comments from Mark Joshi, Svend nonparametric regression. Insurance: Mathematics and Economics 19 (1), 19–30.
Jakobsen, B.J. Christensen, and Malene Shin Jensen as well as from Glasserman, P. (2004). Monte Carlo Methods in Financial Engineering,. Springer Verlag.
participants at the March 2007 World Congress on Computational Finance in London, Glasserman, P. and S. Kou (2003). The term structure of simple forward rates with jump
the conference on Quantitative Methods in Finance in Sydney, December 2005, and risk. Mathematical Finance 13 (3), 383–410.
at a seminar at Cass Business School in London. Any remaining errors are the Glasserman, P. and X. Zhao (1999). Fast greeks by simulation in forward libor models.
authors’ responsibility. Journal of Computational Finance 3 (1), 5–39.
Hagan, P. S., D. Kumar, A. S. Lesniewski, and D. E. Woodward (1997, September).
Managing smile risk. Wilmott Magazine, 84–108.
Haugh, M. B. and L. Kogan (2004). Pricing american options: A duality approach.
FOOTNOTES & REFERENCES
Operations Research 52 (2), 258–270.
Heath, D., R. Jarrow, and A. Morton (1992). Bond pricing and the term structure of interest
1 A CMS rate is not the fixed rate that sets the price of a Constant Maturity Swap equal to rates: A new methodology for contingent claims valuation. Econometrica 60 (1), 77–105.
zero. A Constant Maturity Swap is more general expression than a standard swap in that Hoffmann-Jørgensen, J. (1994). Probability with a View toward Statistics. Chapman & Hall.
the floating LIBOR rates are normally replaced with a longer term fixed maturity rate for Hull, J. and A. White (2001). Forward rate volatilities, swap rate volatilities, and
example the 20 year swap rate. Pricing a Constant Maturity swap is a non-trivial exercise implementation of the libor market model. Journal of Fixed Income 10 (2), 46–63.
(see e.g. Hunt and Kennedy (2004)). Hunt, P. J. and J. E. Kennedy (2004). Financial Derivatives in Theory and Practice. Wiley.
2 The choice of valuation date before the first coupon date is made purely for notational Jamshidian, F. (1997). Libor and swap market models and measures. Finance & Stochastics
simplicity. Valuation at later dates entails no further complications. 1 (4), 261–291.
3 An alternative to this is the indirect approach described in Piterbarg (2004b). We have Jeffery, C. (2006, February). Getting flattened. Risk Magazine 19 (2).
tried this approach but where not able to get meaningful results. Johnson, N. L. and S. Kotz (1973). Continuous Univariate Distributions, Volume 2.
4 We also tried minimizing prices instead and achieved almost identical results. Houghton-Mifflin Company Boston.
5 The LMM with the current parametrization can also be calibrated to both swaptions and Joshi, M. (2006). Monte carlo bounds for callable products with non-analytic break costs.
caps but as described in Rebonato (2002) there are inherent differences between the Working Paper available www.markjoshi.com.
swaption and the cap market that complicate simultaneous calibration to both markets. Lawrence, C. and A. Tits (2001). A computationally efficient feasible sequential quadratic
6 Caps are settled quarterly for maturities below 2 years and semiannually thereafter. We programming algorithm. SIAM Journal of Optimization 11(4), 1092–1118.
therefore translate quarterly volatilities into semiannual volatilies using the procedure in Longstaff, F. and E. Schwartz (2001). Valuing american options by simulation: A simple
Brigo and Mercurio (2006). Note also that caplet volatilites are stripped from cap least-square approach. Review of Financial Studies 14 (1), 113–147.
volatilities using the standard “bootstrap” procedure also described in Brigo and Miltersen, K. R., K. Sandmann, and D. Sondermann (1997). Closed form solutions for term
Mercurio (2006). structure derivatives with log-normal interest rates. Journal of Finance 52 (2), 409–430.
7 Note that we do not have a closed formula for the Black-76 volatility in the CEV case Pietersz, R. and A. Pelsser (2006). A comparison of single factor markov-functional and
but only a pricing formula. We therefore plug in the price from formula (117) and invert multi factor market models. Unpublished Working Paper.
the Black-76 formula. Piterbarg, V. (2004a). Computing deltas of callable libor exotics in forward libor models.
Journal of Computational Finance 7 (2), 107144.
Piterbarg, V. (2004b). Pricing and hedging callable libor exotics in forward libor models.
Amin, A. (2003). Multi-factor cross currency libor market models: Implementation, calibration Journal of Computational Finance 8 (2), 65–119.
and examples. preprint, available from http://www.geocities.com/anan2999/. Rebonato, R. (2002). Modern Pricing of Interest Rate Derivatives, The Libor Market Model
Andersen, L. (2000, Winter). A simple approach to the pricing of bermudan swaptions in and Beyond. Princeton University press.
the multifactor libor market model. Journal of Computational Finance 3 (2), 5–32. Rogers, L. (2002). Monte carlo valuation of american options. Mathematical Finance 12 (1),
Andersen, L. and J. Andreasen (2000). Volatility skews and extensions of the libor market 271–286.
model. Applied Mathematical Finance 7, 1–32. Sawyer, N. (2005, October). A difference of opinion. Risk Magazine 18 (10), 20–22.
Andersen, L. and J. Andreasen (2001). Factor dependence of bermudan swaptions: Fact or Schoenmakers, J. (2002). Calibration of libor models to caps and swaptions: A way round
fiction. Journal of Financial Economics 62 (1), 3–37. intrinsic instabilities via parsimonious structures and a collateral market criterion. Working
Andersen, L. and M. Broadie (2004). Primal-dual simulation algorithm for pricing Paper, Weierstrass Institute.
multidimensional american options. Management Science 50 (9), 1222–1234. Schoenmakers, J. and B. Coffey (2000). Stable implied calibration of a multi-factor libor
Björk, T. (2004). Arbitrage Theory in Continuous Time, Second Edition. Oxford University model via a semi-parametric correlation structure. Working Paper, Weierstrass Institute.
Press. Svenstrup, M. (2005). On the suboptimality of single-factor exercise strategies for bermudan
Brace, A., T. Dun, and G. Barton (2001). Towards a Central Interest Rate Model. Cambridge swaptions. Journal of Financial Economics 78 (3), 651–684.
University Press. Tsitsiklis, J. and B. V. Roy (1999). Optimal stopping of markov processes: Hilbert space
Brace, A., D. Gatarek, and M. Musiela (1997). The market model of interest rate dynamics. theory, approximation algorithms, and an application to pricing high dimensional financial
Mathematical Finance 7 (2), 127–154. derivatives. IEEE Transactions on Automatic Control 44, 1840–1851.

WILMOTT magazine 125

Das könnte Ihnen auch gefallen