Pricing Convertible Bonds with

Interest Rate, Equity, Credit, FX and Volatility Risk

Ali Bora Yigitbasioglu

31
st
August 2002

First Draft: 30
th
November 2001




Abstract

Pricing convertible bonds poses numerical challenges that are not easily overcome. We present a
quasi five-factor model with interest rate, equity, credit, currency, and volatility risk. This is
implemented using unconditionally stable methods. We extend a method to address credit risk, and
propose a means to deal with cross-currency convertibles. A procedure for extracting the price of
vanilla options struck on foreign stock in domestic currency is employed to obtain local volatility.
This is useful for pricing primary issue and secondary market convertibles whose maturities may be
spanned by the currency and equity options markets. We facilitate numerical convergence with
Wolfes (1959) quadratic minimization algorithm, which assists in smoothening local volatility.
Coupons, dividends, reset clauses, calls and Bermudan style puts are easily accommodated. We
implement a functional relationship between stock price and credit spread, to capture the negative
correlation between spreads and equity.


Keywords: cross-currency convertibles, credit spread, interest rate risk, American feature, local
volatility, Crank-Nicholson, quadratic programming, Wolfes algorithm.


JEL Classification: C63, C13, C15









The author would like to thank Carol Alexander, Oliver Brockhaus, Peter Carr, Ryan Davies, Gordon Gemmill,
Giovanni Guidi, Shmuel Kandel, Olaf Korn, Beni Lauterback, Dmitri Lvov, Dilip Madan, Fabio Mercurio, Salih
Neftci, Luke Olsen, Jacques Pzier, William Press, Riccardo Rebonato, Bruce Resnick, Chris Rogers, Steven Satchell,
Steve Shreve, Mihai Sirbu, Saul Teukolsky, Steven Todd, and discussants and participants of the 2nd Bachelier World
Conference in Crete, the EFMA 2002 conference in London, the EFMA 2002 doctoral session in London, the FMA
2002 Annual Meetings in San Antonio, Texas, and Morgan Stanley (London) Ltd. for providing the convertibles,
corporate bonds, fixed income, and options data.

ISMA Centre, University of Reading, Box 242, Reading RG6 6BA, UK.


1
The standard technique in the contingent claims approach to pricing financial instruments is
to stipulate one or more stochastic processes for the risk factors that are thought to drive the
price of the instrument. Arbitrage free pricing theory then leads to a unique price for the
instrument when markets are complete. Passage to this unique price relies on the fact that
when an analytical price is not available, accurate and tractable numerical techniques exist
that can be applied to obtain the price. Quite frequently, instruments possess features that
pose additional complications, and often these cannot be resolved by introducing extra risk
factors into the simulation. The more complex instruments, such as those possessing whole or
partial American type early exercise free boundaries, cannot be priced without recourse to a
multifactor partial differential equation (PDE) approach. However, the difficulty of the
numerical solution to multifactor PDEs increases dramatically with the number of stochastic
risk factors included.

The instrument price may be exposed to multiple sources of randomness, such as to stock
prices, interest rates, credit quality, currency, stock volatility, and interest rates. As more
sources of risk are considered, the paramount issue is to find a balance between a theoretically
sound and yet numerically feasible model. Convertible bonds are one type of instrument
whose pricing poses a substantial set of complications, both with regard to the theoretical
model, but also with respect to procuring a numerical solution which is stable, flexible, fast,
and as a result commercially usable. We introduce such a model in this paper and discuss its
implementation in detail. Our approach has the advantage of nesting a wide array of dressed
down pricing models which may be substituted to price convertible bonds sensitive to a
subset of the risk factors above. Hence the model encompasses pricing of many types of
convertible bonds such as straight domestics, cross-currency bonds, or cross-currency
exchangeables, all of these with or without volatility risk or credit risk.

Convertible bonds are hybrid securities that confer upon the owner the right to receive a fixed
income stream during the life of the convertible with the embedded right to forego the fixed
income stream and irrevocably convert at the holders option into a prescribed amount of the
issuers equity any time during the life of the instrument
1
. Hence, the presence of convertible
bonds in the firms capital structure may lead to the issuance of new shares and concomitant
equity dilution. Because convertibles will be exercised when the issuer does well, convertible
debt acts to lower the issuers cost of debt financing through the implicit call option on the

1
In some cases the convertible may only be converted during a prescribed time period during its life, or into shares
of a different corporation.


2
issuers stock that is implicit in the convertible. In essence, the convertible bond is a mixture
of a corporate bond and a warrant on the underlying equity. The decision to exercise the
convertibility feature terminates the bond component of the instrument
2
. Its corporate debt
and equity nature renders it sensitive to the stock price, interest rates, and the issuers credit
quality. Furthermore, the special case of currency risk arises for those convertibles that pay
coupons and face value in some domestic currency but convert into some foreign equity
3
.

The global market for convertible bonds has gained in popularity in recent years. The
worldwide convertibles market by issue size exceeds US$ 400 billion and highly developed
markets for convertibles exist in the United States, Japan, England, France, Canada, Australia,
Sweden, and Switzerland. The majority of issues in these markets are domestic convertibles
denominated in a local currency and convertible into domestic equity. Details of outstanding
issue size, maturity, and initial premium
4
for the US market case are provided in Appendix D.
In addition to domestic convertible markets, there is a sizeable Eurobond convertible market
in which currency sensitive convertibles are actively issued and traded.
5
This market as a
whole grew from US$ 8.1 billion in 1990 to US$ 14 billion in 1995 (Calamos (1998)). The
Eurobond convertible market, as a subset, contains convertible issues that convert into the
issuers equity but are denominated (i.e. pay coupons and face value) in some other currency.
These instruments are naturally sensitive to currency risk
6
. One of the largest convertible
bond markets by issue size is Japan, which as of the first quarter of 2002 had a 17% share of
the world market in convertibles. It has produced a large share of cross-currency convertible
bonds, amongst them a US$ 2bn issue by the Mitsubishi Bank. A rapidly growing continental
Eurobond convertible market is the Swiss market, which is unique in that the majority of
companies that issue convertible debt in Swiss francs are not domestic firms. Examples of
cross-currency denominated outstanding convertible issues on the Swiss markets are
numerous, including the USD denominated Hutchison Whampoa International 2% coupon,
convertible into sterling denominated shares in Jan. 2004. The preceding examples are
augmented by numerous other highly liquid, large cross-currency convertible issues. Some of
these are listed in Appendix D. However, although the size and importance of this subset of

2
Hence the exercise price of the warrant component of the convertible is stochastic, and depends on interest rates
and credit quality of the issuer prevailing at the exercise time.
3
The third largest convertibles market is the Eurobond convertible bond market. Many of the constituent bonds in
this market have currency risk.
4
Defined as the percentage difference between the conversion price of equity and the issue price.
5
Eurobond convertible refers to the broad market category encompassing all convertibles issued outside the
issuers domestic market. The denomination currency can be issuers domestic currency (no currency risk) or some
foreign currency (leading to currency risk).


3
the convertible bonds market is unquestionable, none of the models suggested earlier can be
applied to price these.

Convertible bonds have a number of contractual features that from the outset complicate the
pricing. The American early exercise free boundary, together with (constant, resetting, or
trigger-dependent) call by the issuer and (usually Bermudan type) put by the investor,
constant or resetting trigger dependent conversion ratios, soft call protection features, anti-
dilution, anti-takeover (such as the poison put) clauses are amongst some of these. Credit risk
modelling is also more intricate due to the hybrid credit nature of convertibles. The equity
component of the convertible bond poses no credit risk to the investor. This is because the
issuer can always deliver its own equity
7
while the fixed income characteristics of the bond
such as coupons, the face value repayment, and the cash payment when the bond is put back
is subject to the credit risk of the issuer, who will need cash to fulfil its obligations
8
. We
develop a quasi five factor parabolic coupled PDE model with change of numeraire that can
be used to price international convertibles with credit and volatility risk. The credit risk of
convertibles is dealt with by introducing an auxiliary asset called the cash only part of the
convertible (hereafter abbreviated to COCB), which pays conversion contingent fixed income
cash flows only. This extends an approach introduced by Tsiveriotis and Fernandes (1998).
This security also satisfies a similar pricing equation with its boundary and initial conditions
linked to the convertible security lifecycle. The two pricing equations have to be solved
simultaneously due to the coupling. We assume the capital structure of the issuer consists of
senior corporate debt which trades at a spread over the treasury yield, and the junior
convertible issue. Extension of the credit risk framework is achieved by representing the price
of the COCB in a multifactor framework, thereby allowing the price of the cash part of the
convertible to depend on the interest rate in addition to the equity price. We then solve the
coupled two factor system with the Crank-Nicholson (semi-implicit) scheme, preferring this
over trinomial and binomial lattice and explicit finite difference schemes used earlier which
are prone to numerical instability. This enriches the credit risk modelling feature of the

6
The International Securities Markets Association (ISMA) Cupid database on convertible bonds reveals that 78%
of Japanese convertibles are cross-currency denominated.
7
Unless the convertible is exchangeable into the equity of a firm other than the issuer, in which case the issuing
firm has to purchase the equity with cash.
8
It is possible to view the credit spread associated with the same issuers nonconvertible debt as an upper bound
for the credit risk associated with the convertible bond, a fact that is often used by practitioners for valuing
convertibles with a rule of thumb approach. Using this view, a trader might use say x% of the credit spread for
nonconvertible debt as the appropriate discount rate for the convertible bond in the pricing equation, and apply this
spread throughout. Although the drawback of using such ad-hoc approaches is self-evident, it is the authors
experience that it is regularly applied in industry.



4
model, ensuring stability of results
9
, but at the cost of increased computational intricacy, as
the computational domain is doubled in size and feedback between the equity and bond
component PDEs amplifies scope for computational bottlenecks. We employ an iterative
solver with adaptive learning to successfully speed up the computation.

Major banks routinely mark their exotics to market using local volatilities obtained from
implied surfaces for equity and currency options. The underlying equity for convertible bonds
usually has a liquid options market, with banks warehousing implied volatility surfaces of up
to two years maturity (or further) for equity options, and four to five years for currency
options. By sidelining the issue of volatility sensitivity for convertible bonds, earlier studies
have implicitly endorsed a view that the (relatively) longer maturity of convertible bonds
obviates the need to account for price sensitivity to the local volatility. Convertible bond
prices will display the same sensitivity to short and medium term volatility as other exotic
derivatives such as barrier options, because they have a shorter expected life span than their
time to maturity when they are callable, putable, or both. A seasoned trader will probably deal
with this by forming an estimate of the expected life of the convertible bond, and in his
pricing use the implied volatility corresponding to this maturity and strike corresponding to
the conversion premium. Alternatively, he may use the implied volatility from options
corresponding to the maturity of the convertible. The choice of this constant parameter is
invariably subjective, and in both cases, these approaches are somewhat ad-hoc, as model
prices depend closely on traders estimates of instrument lifetime. When instead we use a
local volatility surface, the volatility estimation for the convertible bond pricing becomes an
endogenous aspect of the model. Using a local volatility surface in the pricing grid will result
in only that part of the smile and term structure of volatility being used in the grid for which
call by issuer and put by investor do not apply. This removes the dependency of the model on
the trader providing a correct implied volatility each time the instrument is re-priced, which
will have implications on convertible arbitrage and indeed on any trading situation where
many convertible bonds are priced, traded, re-priced, and hedged. As such, an automated and
more accurate method for providing the volatility estimate, via the local volatility surface, can
be very useful. Instruments whose pricing is most enhanced with local volatility are
convertible bonds trading in the secondary market
10
. It is also very useful to use this approach

9
In a practical setting, stability (i.e. no oscillations , pricing convergence) of the pricing model is crucial. Lattice
methods for pricing convertible bonds cannot guarantee this. Consequently, a trader is at the mercy of a model
that can produce large pricing errors (due to oscillation) and using it to price a portfolio of convertible bonds will
often lead to unpredictable behaviour.
10
Many of these will be actively traded by (amongst others) convertible arbitrage hedge funds, who will sell
overvalued convertible bonds and buy undervalued ones, and hedge in the equity markets.


5
to price primary issue convertible bonds as their prices are then consistent with the
information content of the liquid vanilla options market. In any case, when a primary issue is
planned, the lead manager will frequently have a good indication of long term equity and
currency volatilities from different brokers, and these can be readily incorporated in the
pricing of the primary issue.

Even though we do not explicitly treat volatility as stochastic, we allow for time and strike
dependent local volatility on the finite difference grid, after the approaches pioneered by
Rubinstein (1994), Dupire (1994), and Derman and Kani (1994)
11
. To our knowledge, this
approach has not been used in the convertible pricing literature to date
12
. To obtain local
volatility from the implied volatility surface, we propose two complementary techniques. The
first involves minimizing Arrow-Debreu market price deviations from their theoretical
minimum and gives one the choice to constrain local volatilities to stay in a bound range (say
10 to 60%) as part of the mathematical formulation of the problem. In addition to being
theoretically appealing, this method allows us to restrict the local volatility on the PDE grid.
A bound might be imposed (for example, by equating local volatility to flat at the money
volatility at that point) if the implied surface is highly spiked in some portion of the grid in
relation to its neighbouring points. This can be of immense benefit as unduly spiky local
volatilities can contaminate the solution to the CB pricing problem. The approach can be
implemented using Wolfes (1959) three-phase simplex method for quadratic programming,
or by using a commercial C++ library such as that provided by the Numerical Algorithms
Group (NAG). We provide a description of Wolfes method when applied to our problem in
Appendix C. A quicker alternative method is suggested which involves second order accurate
finite difference approximations to the partial differentials in a functional relationship
between local volatility and the (observed) implied volatility surface, obtained via the Fokker-
Plank backward equation.

Our method relies on a change of probability measure to keep the problem dimension lower
for cross-currency convertible bonds. Although the change of numeraire technique has been
around in the literature for some time (Geman et al. (1995)) and in fact underpins the risk
neutral pricing theory via Girsanovs theorem and the Black and Scholes (1973) analysis, it

11
This allows us to keep the tools of arbitrage free pricing. Moving to a model where volatility is stochastic places
us in an incomplete market world.
12
Possible use of local volatility is mentioned although the idea is not implemented in a recent working paper by
Hoogland et al. (2001).



6
has not been used to simplify the pricing problem for convertible bonds when currency
sensitivity is present. Our use of numeraire change avoids increasing the dimension of the
problem for the currency risk in cross-currency convertible bonds. The model for
international convertibles also nests domestic convertibles as a subset. Inclusion of currency
risk via change of measures poses little additional difficulty if a flat volatility is assumed.
However, unifying the change of numeraire method with the local volatility framework is not
straightforward. The change of numeraire relies on the stochastic process followed by the
foreign stock price in domestic currency, and because there is no market-given volatility
smile for options struck on this synthetic process, to get its local volatility surface we have to
develop a numerical procedure to back out the implied volatility of options struck on it, using
the implied volatility for vanilla FX and foreign stock (in foreign currency) options. This is
accomplished numerically by evaluating an expectation over the joint density of the FX rate
and the foreign equity processes, over all possible (combinations of) strikes for FX and
foreign equity in the double integral. The prices obtained are used to back out the implied
volatility for the foreign equity in domestic currency. Even though the integration is done
over a numerically obtained
13
implied volatility surface for FX and equity, and scope for
amplifying the numerical errors cannot be ruled out, we find in fact that the method
contributes a degree of smoothness to the implied surface for the stock price in domestic
currency. These implied volatilities are used to obtain local volatility.

The collapsing bond floor of convertible bonds as stock prices fall sharply is a fact well
known to traders and practitioners. Previous studies have not sought to capture this
phenomenon, which can impact model prices significantly. Credit spreads are sensitive to
firm value, and will tend to increase when equity prices decrease, and vice versa. Earlier
works have used constant credit spreads for pricing, obtained from corporate (nonconvertible)
debt of similar maturity, which is either applied in a so-called full discounting, or in a
blended discount model. In these, the spread is invariant to changes in the credit quality of
the issuer as the issuers stock price rises or declines. Our model can replicate the collapsing
bond floor as stock prices go down. On the other hand credit spread for high values of stock
price will decrease due to the improving credit quality of the issuer. This too should be
captured in a pricing model, which is applied here.
A Crank-Nicholson scheme, with adaptive grid to match time steps to jump events, such as
coupon payments or discrete dividends, is employed. The American early exercise feature is

13
via bi-cubic spline interpolation



7
posed as a linear complementarity problem and is tackled using the projective successive
over-relaxation (PSOR) method. This method has some shortcomings in that it cannot give
the exact location of the free boundary at which it is optimal to convert the convertible bonds.
By employing smaller equity steps in our grid the adverse effect of the PSOR method is
reduced: the PSOR scheme requires a diagonally dominant space-discretization matrix for
convergence, which is enforced using small time steps, and the high number of equity steps
reduces the inaccuracy due to projection.

We provide details on the pricing PDE and its discretization in various regions of the grid in
relation to the boundary and initial conditions. This serves two purposes. It provides a full
specification for the stability of the discretization schemes and also sheds light on the
implementation details. The paper is organized as follows. Section I provides a review of the
existing models for pricing convertible bonds. Section II details the basic model for currency,
equity, and interest rate; it provides details of the change of numeraire technique, local
volatility, and the credit risk approach in addition to the pricing equations and boundary and
initial conditions on the grid. Section III provides the numerical details of the model. We
provide some numerical results from the model in Section IV. Section V concludes. A
detailed survey of the numerical techniques, the discretization, and accuracy of the schemes
are provided in the appendix.


I. A survey of the literature
The literature on using the contingent claim approach for pricing convertible bonds in the
Black and Scholes (1973) framework originates in the firm value models of Brennan and
Schwarz (1977, 1980) and Ingersoll (1977). In both, the firm acts to maximize common
shareholder value, and as a corollary, minimizes the convertible bond value. A consequence
of this is that a firm behaving optimally should call the convertible bond as soon as the
conversion price is equal to the call price. Many researchers have investigated firms call
policies in the empirical literature, and the consensus is that firms very often wait until the
conversion price is significantly higher than the call price before issuing the call. Possible
causes for this have been proposed. Amongst these feature the call notice period effect and
the safety net (Asquith (1995) and Asquith and Mullins (1991)), the preferential tax
treatment of coupons over dividends as an incentive to keep the convertible bonds alive
(Asquith (1995), Asquith and Mullins (1991), and Constantinides and Grundy (1984)), and


8
signalling effects of issuing convertible bonds calls which may convey to shareholders that
management expects share prices to fall (Harris and Raviv (1985) and Mikkelson (1985)).
Another possible cause for delayed call is advanced in Dunn and Eades (1984), tested in an
empirical study by Constantinides and Grundy (1984), who say that it is not in managements
interest to awaken sleeping investors who are not optimally converting their bonds by
issuing a call. On an interesting historical note, Brennan and Schwarz (1977) are also the first
to introduce the method of finite differences into the finance literature. They use firm value
as a stochastic variable to value domestic convertible bonds which are priced using an explicit
finite difference scheme. Brennan and Schwarz (1980) subsequently extend their earlier
model to include interest rate risk. Ingersoll (1977) provides a pricing formula for one factor
callable convertible bonds using risk-neutrality, and expresses the convertible bond price as
an expectation over the terminal payoff of the convertible, which is solved using numerical
integration. Credit risk is not treated explicitly as a risk factor in the PDE in these papers
14
. In
Brennan and Schwarz (1977) the only point where the bond should be converted is possibly
immediately prior to a dividend payment date. In Ingersoll (1977), the bond should not be
converted except possibly at maturity. Hence, as early conversion in these papers is not
optimal, it is not necessary in their approach to address the American free boundary.

After a lengthy hiatus in the pricing literature, Wilmott, Howison, and Dewynne (1993)
proposed that the American feature of convertible bonds can be interpreted as a linear
complementarity problem, which after discretization can be solved using a so-called
projection version of the successive over-relaxation technique from numerical analysis. The
PSOR technique avoids solving for the early exercise boundary. Wu and Kwok (1997) use a
transformation method
15
to solve explicitly for the early exercise boundary for an American
put with continuous dividends using a one factor PDE. Consequently Zhu and Sun (1999)
extend this technique to two factors, and study convertible bonds. While a promising
approach, their method is not very practical as it does not incorporate the call and put features
of convertibles, each of which constitutes a further coupled non-linear PDE, inclusion of
which will complicate the numerical implementation enormously.

Tsiveriotis and Fernandes (1998) proposed an approach to tackle credit risk whereby the
convertible bond price is driven by its bond component and its equity component, in a
coupled Black-Scholes-Merton PDE framework. In their model, the convertible bond price is

14
Credit risk becomes an endogenous component of the model, along the lines of Mertons (1974) model.
15
The so-called front-fixing technique.


9
driven by an It process and related to the stock price and time through a unique and determinable
PDE. They write the PDE for the convertible bond, which depends on the solution to another
simultaneous PDE (hence the coupling) for its bond component. This method is inherently
superior to the blended discount models of Goldman Sachs (1994) where a probability of default
weighted discount rate is used
16
, and Ho and Pfeffer (1994) who use a single discount factor plus
spread. The latter model unnecessarily penalizes the convertible bond price by using a constant
credit spread, when the equity price is high and the risk of default is correspondingly low. Lischka
and Davis (1999) propose a two and a half factor lattice model with stochastic interest rate,
equity, and hazard rate calibrating to the initial term and also to the term structure of credit
spreads via a Hull and White (1990) interest rate framework. This framework does not
address currency risk or volatility risk, but permits some recovery in the event of default.

Recently Hoogland et al. (2001) present a one factor (stock price) model to price Japanese
resetting convertibles, and propose a method for accounting for the call notice period and the
conversion rate reset feature without increasing the dimension of the problem. Grimwood and
Hodges (2001) investigate the extent of alternative convertible bond pricing model mis-
specifications in a subset of the models in the literature. As such it is the first work to our
knowledge to compare these competing models. The paper also provides a catalogue of the
different features of convertible bond contracts for both the Japanese and USA markets in the
sizeable International Securities Markets Association (ISMA) database. Takahashi et al.
(2001) develop a lattice pricing model using equity as the only risk factor and employ a
Duffie-Singleton (1999) type model to characterize credit risk of convertibles, where the
default event is defined exogenously by a jump process and it is possible to obtain the
arbitrage free price of securities subject to default risk. There are several shortcomings of
their approach. They only consider equity risk in the PDE, and moreover implement it using
lattices which are subject to numerical instabilities. To obtain prices, they also have to make
the assumption that fractional loss of market value for convertible bonds is the same as the
fractional loss of market value for a nonconvertible bond of the same company and
comparable maturity.

Srbu, Pikovsky, and Shreve (2002) remove the time parameter from the valuation problem by
assuming the bond is perpetual. As a result the free boundary associated with optimal call and
conversion becomes a free point problem. They derive an arbitrage free price for the

16
Therefore in effect the Goldman Sachs (1994) model is a one factor approximation to the Tsiveriotis and
Fernandes (1998) model, and its implementation on a binomial tree rather than explicit finite difference makes it


10
convertible bond as the solution of a nonlinear ordinary differential equation. Barone-Adesi et
al. (2002) suggest a finite element framework for pricing the PDE for convertible bonds in the
presence of two stochastic risk factors. Their model is capable of calibrating to the initial term
structure via a Hull and White (1990) model, but does not treat credit, FX, or volatility risk.
They propose an iterative algorithm based on a sequence of variational inequalities to keep
track of the American free boundary. Olsen (2002) provides a detailed technical survey of the
existing models in the literature. This is also the first paper in the literature to suggest a credit
spread versus stock price functional relationship in the pricing routine for convertible bonds.
This is implemented in a one factor lattice model, and numerical results from empirical
investigations of the spread/stock price sensitivity parameter in their model are reported for a
range of market issues. The model presented captures many stylized features observed in the
markets such as the collapse of the bond floor value of the convertible as the issuers stock
price declines.


more prone to error.


11
II. The Model
A. Definitions and Notation
The market consists of a foreign asset paying continuous dividend yield , and the foreign
exchange rate (denoting number of foreign currency units per domestic currency unit). We
assume the foreign stock price (in foreign currency units) has the following standard form in
the objective domestic probability measure:

1 S
t)dW (S, S )dt S( dS + = (1)

Here,
1
dW is a univariate Brownian motion under the domestic objective measure, and
) t , S (
S
is some local volatility function depending on the volatility smile for foreign equity
(vanilla) options. We choose the following one factor Cox, Ingersoll, and Ross (1985)
(hereafter CIR) model to characterise domestic interest rate dynamics, although the pricing
framework is general and can accommodate other interest rate models, including those which
calibrate to the initial term structure, such as Hull and White (1990).

r
dW (t) r w (t)dt ar ( (t) dr
d d d
+ = (2)

It is proposed that the exchange rate X, the number of foreign units of currency per one unit
of domestic currency, follows the geometric Brownian motion

2 X X
t)dW (X, X dt X dX + = (3)

where
2
dW is a univariate Brownian motion under the domestic objective measure and
) t , X (
X
is some local volatility function depending on the volatility smile for FX vanilla
options. Volatility is functionally dependent on the underlying and time and hence, contingent
upon the realized value of the underlying in the future, is fully determined. For notational
convenience, we write
X
and
S
as shorthand for the local volatilities from here on,
suppressing dependence on underlying and time. We assume that the univariate Brownian
motions for the stock price and foreign exchange rate are correlated as follows:

dt dW dW
X S, 2 1
=


12
with correlation coefficient
X S,


As the market with local volatility is complete, we can use arbitrage arguments to derive the
processes for the FX rate, and the foreign equity in domestic currency, under the domestic
risk neutral measure. It is easy to show that the foreign exchange rate follows the geometric
Brownian motion with local rate of return
f d
r r in the domestic risk neutral measure:

2 X f d
W d X )dt r X(r dX + =

where
2
W d is a univariate Brownian motion in the domestic risk neutral measure.

B. Measure change
In the convertible bond pricing model, the foreign stock price in domestic currency units is
employed as a state variable. To derive the domestic price of the foreign stock, denoted
) t ( S ) t ( X S
~
= , a simple application of Its lemma yields:

1 S 2 X S X X S, f X
1 S 2 X S X X S, X
S X X S, 1 S 2 X X
2
dW S
~
dW S
~
)dt ( S
~
dW XS dW XS )dt XS(
)dt XS( ) dW S )dt X(S( ) dW X dt S(X
dXdS
S X
S
~
dS
S
S
~
dX
X
S
~
S
~
d d(XS)
+ + + + =
+ + + + =
+ + + + =

= =


This can be rewritten by combining the correlated Brownian motions W
1
and W
2
into another
Brownian motion W
17
, also in the domestic objective probability measure:

dW 2 S
~
)dt ( S
~
S
~
d
S X X S,
2
X
2
S S X X S, f X
+ + + + + =

For S
~
to have the correct risk neutral drift of r
d
, introduce the measure transformation:


17
Here W has the same drift (zero) and variance as
S
dW
1
+
x
dW
2

(4)


13
W d 2 S
~
)dt (r S
~
S
~
d
)dt ( r dW 2 W d 2
S X X S,
2
X
2
S d
S X X S d S X X S,
2
X
2
S S X X S,
2
X
2
S
+ + + =
+ + + + + = + +


Here W d is the univariate Brownian motion associated with the domestic risk neutral
probability measure. We also assume that the interest rate SDE and the SDE for S
~
are
correlated through

dt r )w ( S
~
dr S
~
d
r , S
~
X S
+ = .

C. Risk neutrality and the basic pricing equation
Its lemma applied to the convertible bond ) t , r , S
~
( u which depends on the foreign equity in
domestic currency
18
, the domestic interest rate r, and time yields the SDE followed by the
differential process du:

t) r, , S
~
du( dW 2 S
~
S
~
u
dW r w
r
u
dt r w 2 S
~
r S
~
u
r w
r
u
S
~
u
S
~
) 2 ( S
~
) (r
S
~
u
ar) (
r
u
t
u
S X
2
S
2
X
r
r , S
~
S X
2
S
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1
= + +

+
|
|
.
|
+ +

\
|

+ + +

Risk-neutrality arguments then lead to the pricing equation for the convertible bond which
pays domestic interest but converts into foreign equity

0 ru r w 2 S
~
r S
~
u
r w
r
u
S
~
u
S
~
) 2 ( S
~
) (r
S
~
u
ar) (
r
u
t
u
r , S
~
S X
2
S
2
X f
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1
= + +

+ + +

18
Earlier work by Brennan and Schwarz (1977, 1980), Ingersoll, and others have focused on firm value
as one of the stochastic risk factors driving the convertible bond price. Although this is theoretically
appealing, it is not observable in the market and computing firm value is extremely data intensive,
compared to using the stock price of the firm, which is directly observable. Furthermore, using firm
value as a stochastic variable makes local volatility modelling impractical as options are written on the
stock, not the firm value.

(5)


14
This equation is valid only up to the early exercise boundary, at which S
~
. CR ) t , r , S
~
( u =
(provided the conversion price is below the call price plus accrued interest). Here, CR refers
to the conversion ratio.
It is assumed that the convertible bondholder (the domestic investor) converts only when
S
~
. CR is high enough compared to V
L
(t) = max(PP(t)+AcInt(t), Bond_Floor_Value(t)),
where PP(t) is the prevailing put price at time t (a step function of time), AcInt(t) is the
accrued interest at time t since the last coupon payment, and Bond_ Floor_Value(t) is the time
t present value of the fixed income flows from the convertible bond, taking into account the
credit quality of the issue. Thus V
L
(t) constitutes a lower value for the convertible.

The call value of the convertible is assumed to be resetting (and is cast as a step function),
possibly with hard call protection
19
, and can be cast as an upper value for the convertible:
V
U
(t) = ) t ( AcInt ) t ), t ( r , S
~
( V
call
+ . Equivalently this is an upper value for S
~
, thereby
constituting an upper price bound for the instrument on the finite difference grid.

D. Local Volatility and Calibration
Smile and skew risk for cross-currency and domestic convertible bonds will be most relevant
in the shorter maturity secondary markets. For these shorter maturities liquid options market
data exists. For example, in the foreign exchange markets, FX options data is easily obtained
from a front office vendors system (such as MUREX) up to the first two years. Up to 5-7
years can typically be obtained from brokers. In the equity markets, the maturity of the data
available will be somewhat lower. By sidelining the issue of volatility sensitivity for
convertible bonds, earlier studies have implicitly endorsed a view that the (relatively) longer
maturity of convertible bonds obviates the need to account for price sensitivity to the local
volatility. Convertible bond prices will display the same sensitivity to short and medium term
volatility as other exotic derivatives such as barrier options, because they have a shorter
expected life span than their time to maturity when they are callable, putable, or both. A
seasoned trader will probably deal with this by forming an estimate of the expected life of the
convertible bond, and in his pricing use the implied volatility corresponding to this maturity
and strike corresponding to the conversion premium. Alternatively, he may use the implied
volatility from options corresponding to the maturity of the convertible. The choice of this
constant parameter is invariably subjective, and in both cases, these approaches are somewhat


15
ad-hoc, as model prices depend closely on traders estimates of instrument lifetime. When
instead we use a local volatility surface, the volatility estimation for the convertible bond
pricing becomes an endogenous aspect of the model. Using a local volatility surface in the
pricing grid will result in only that part of the smile and term structure of volatility being used
in the grid for which call by issuer and put by investor do not apply. We document that using
constant volatility (implied at maturity of convertible and strike at conversion premium)
creates considerable price discrepancies, compared to using the local volatility surface.
Assumptions about the long term behaviour of implied volatility have to be made when
pricing longer maturity primary issues, as the maturity may exceed the tenors of available
options data. In this case the result will be dependent on the extrapolation of implied
volatilities, although call and put features will now act to dampen the pricing impact of the
choice of extrapolation, by reducing the expected lifetime of the instrument. The situation is
simpler for secondary market convertibles, where maturities are lower, and our approach can
readily be used for assessing fair value and for arbitraging.

It is assumed that prices for OTC call/put options struck on S and X are available up to the
maturity of the convertible bond, or have been extrapolated to span the finite difference mesh
for the convertible bonds when necessary. The extrapolation can be done in many ways,
contingent on the view held of the distribution of future stock price and foreign exchange.
Some of these techniques are Shimko (1993), who appends log-normal tails to the Arrow-
Debreu securities in the fringes of the mesh, Rubinstein (1994), who minimizes the deviation
of the system of Arrow-Debreu prices from a lognormal distribution, and Jackwerth and
Rubinstein (1996). Given the implied volatilities for S and X, the local volatility
S
~ for
vanilla options struck on S
~
can be found numerically from the implied volatility for S
~
. The
local volatility replaces a constant one in the convertible bond PDE. The difficulty is that S
~

is a non-traded security, and there will be no options prices struck on S
~
, but these are needed
to obtain local volatilities for cross-currency convertibles. To solve for ) t , S
~
(
imp
for the
synthetic call option struck on S
~
, recall that the price of a call option struck on the foreign
equity in domestic currency units can be expressed as an expectation over the risk-neutral
domestic measure of the (call) payout with respect to the density of S
~
. This is written
equivalently as the joint bivariate lognormal density of X and S, expressed as the product of

19
A convertible that is not callable for a certain number of months or years is said to have hard call protection.
This is in contrast to soft call protection, when the convertible becomes callable after the underlying foreign stock


16
the conditional bivariate lognormal density of S given all X strikes that combine to the strikes
(and maturities) on the convertible pricing mesh. The risk neutral domestic price of the call
option struck on the scaled asset process, and the joint and marginal densities in the scaled
variable are given as follows.
| |

+
|
|
.
|

\
|
=
=
K
K K
T r
T r imp
X(0))dX | (X(T) S(0))dS X(0), X(T); | (S(T) K) (S(T)X(T) e
(0) | K,0) (S(T)X(T) E e ) T, K, , S
~
C(
d
d


where S(0)) X(0), X(T); | (S(T) is the bivariate conditional lognormal probability density
function for S given X at maturity, given by:

X(0)) | (X(T)
(0)) S
~
| (T) S
~
(
X(0)) | (X(T)
S(0)) X(0), | (S(T)X(T)
S(0)) X(0), X(T); | (S(T) = =

and S(0)) X(0), | (S(T)X(T) is the joint density for S and X, and can be obtained as follows:

( )
(
(

+
=
2
S X,
2
X X S f X,
2
S
2
S X, S X
- 1 2
Y Y Y 2 Y
exp
- 1 SX T 2
1
S(0)) X(0), | (S(T)X(T)

with
S
Y and Y
X
defined as

( )
( )
T
T r r
X(T)
X(0)
ln
Y
T
T r
S(T)
S(0)
ln
Y
X
2
X 2
1
f d
X
S
2
S 2
1
d
S
+
|
|
.
|

\
|
=
+
|
|
.
|

\
|
=


the marginal univariate lognormal density for X is as follows:


trades at a predetermined level above initial price.
(7)
(8)


17
( ) | |
(

=
2
X
2 2
X 2
1
f d
X
T 2
T - r r 0)) ln(X(T)/X(
exp
X T 2
1
X(0)) | (X(T)


The double integral in (7) depends on the smile for foreign equity in foreign currency and on
the exchange rate, and the numerical solution to it gives the implied volatility for the asset S
~

for one particular strike and maturity. Notice that to advance over one particular path in the
numerical integration we need to possess values of bicubic spline interpolated implied
volatility of both X and S for every dS or dX in the path. Choosing small dS and dX to
accurately evaluate the integration results in a large computational effort. This burden can be
eased by implementing routines that search for optimal combinations of X(T) and S(T) in (7)
that combine to a particular (scaled) strike, with optimal upper integration limits for S given a
particular X such that the density weighted payoff is less than some notional small amount
such as 1.0e-40. This and other numerical refinements are implemented to make the
computation more efficient.

Denote by ) t , S
~
(
imp
the implied volatility for calls struck on S
~
, which is used subsequently
to obtain the local volatility functional in the diffusion equations. There are many techniques
in the literature for solving the ill-posed function approximation
S
~
imp
) t , S
~
( , once
) t , S
~
(
imp
has been found. These include a PDE approach (Andreasen and Brotherton-
Ratcliffe (1998)) and spline functional methods (Coleman et al. (1999)). We modify an
approach pioneered by the former to obtain the local volatility. They show that, provided
observable market prices for call and put options for all strikes and maturities exist or can be
interpolated/extrapolated, and provided the underlying process (foreign stock, FX rate)
follows geometric Brownian motion, the instantaneous local volatility functional satisfies the
following PDE, written in terms of the market implied volatility smile ) T K, t; , S
~
( .

(
(

|
|
.
|

\
|

+ |
.
|

\
|

=
+ +
2 2
2
2
2
2
K

d
t T K
1

1
K

t T d
K

K
K

(T)] 2K[r(T)
t T

2
T) (K,


They employ a one-factor Crank-Nicholson technique to solve for the local volatility. The
Arrow-Debreu prices implied by call option prices are solved for and used to back out the
(10)
(9)


18
local volatility using Lemkes quadratic optimization method. Their method reports very high
accuracy for calibration to the vanilla options smile. We alter their approach via a
modification to the basic discretization scheme that they use for the method, where a
discretization term is omitted in the pricing equation. We implement Wolfes three-phase
algorithm for the quadratic optimization, a quicker and more flexible alternative to Lemkes
method. Details of implementation are provided in Appendix C.
Another technique, quicker although not as flexible as the previous one, is to approximate
(10) numerically with finite differences. This approach was suggested in Rebonato(1999). We
choose discretization schemes that are second order accurate everywhere on the grid, and
truncate local volatility above upper thresholds or below lower thresholds to a constant level.
This presupposes that implied volatilities exist for all possible points in the (K,T) space. We
first fill the space with bicubic spline interpolated (Press et al. (2002)) implied local
volatilities and then perform the following second order finite difference approximations for
(10):

(
(

|
|
.
|

\
|

+ |
.
|

\
|

=
+ +
2 2
2
2
2
2
K

d
t T K
1

1
K

t T d
K

K
K

(T)] 2K[r(T)
t T

2
T) (K,


applying:

2 2
2
K
) T , K K ( ) T , K ( 2 ) T , K K (
K
+ +
=

K 2
) T , K K ( ) T , K K (
K
+
=

T 2
) T T , K ( ) T T , K (
T
+
=

we obtain local volatilities T) (K,
2
for all K, T that lie on the original PDE grid for the
convertible bond. The discretization scheme for the partial derivatives is chosen carefully to
be second order accurate in all inner and boundary regions of the grid.

E. Credit Risk


19
Credit risk is included by introducing the cash only part of the convertible bond (COCB), a
synthetic security discussed first in Tsiveriotis and Fernandes (1998). The payoff of the
COCB depends on the underlying convertible security, and the holder of the COCB is paid all
the fixed cash flows provided the convertible remains unconverted, and any put cash flows if
the holder of the convertible decides to put the bond back to the issuer.
As no call notice periods is assumed once the convertible price reaches the call price plus
accrued interest, we impose that at call the COCB is killed (as the convertible is converted
into foreign equity). Because the COCB fixed cash flows are tied in with the issuers ability to
procure the necessary funds to honour its coupon/face repayment and put payment, the
appropriate discount rate for the COCB in the pricing equation is the risk free rate plus a
credit spread that the market attributes to the issuers nonconvertible corporate bonds.

Let ) t , r , S
~
( v denote the international COCB and let ) t , r , S
~
( u denote the international
convertible bond price. It has the pricing equation:

(t) S
~
(t) S
~
for 0 LHS with
0 )v r (r r w 2 S
~
r S
~
v
r w
r
v
S
~
v
S
~
) 2 ( S
~
) (r
S
~
v
ar) (
r
v
t
v
boundary free
c
r , S
~
S X X , S
2
S
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1

>
= + + +

+ + +

where r
c
is some spread above treasury rates that is possibly stock price dependent, and is
related but not necessarily the same as the spread on nonconvertible corporate debt by the
same issuer. This add-on (to the risk-free rate) that applies to the COCB can be above the
issuers senior corporate bond credit spread of comparable maturity.

The residual security uv is the equity component of the convertible bond, i.e. that part of the
convertible whose payoff is contingent on the original convertible being converted. This
should be discounted at the risk-free rate as the issuer of the international convertible bond
can always deliver its own equity even when it is worthless. The pricing PDE satisfied by uv
is given by:

(11)


20
S
~
(t) S
~
for 0 LHS with
0 v) r(u r w 2 S
~
r S
~
v) (u
r w
r
v) (u
S
~
v) (u
S
~
) 2 ( S
~
) (r
S
~
v) (u
ar) (
r
v) (u
t
v) (u
boundary free
r , S
~
S X
2
S
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1

>
= + +


+

+ + +

Rearranging the above leads to:

r , S
~
2
S S X X S,
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1
r , S
~
S X
2
S
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1
r w 2 S
~
r S
~
v
r w
r
v
S
~
v
S
~
) 2 ( S
~
) (r
S
~
v
ar) (
r
v
t
v
rv
ru r w 2 S
~
r S
~
u
r w
r
u
S
~
u
S
~
) 2 ( S
~
) (r
S
~
u
ar) (
r
u
t
u
+ +

+ + +

+ =
+ +

+ + +

The left hand side of (13a) is equal to r
c
v. Therefore, the international convertible bond
pricing PDE can be rearranged to yield the pricing PDE for the international convertible bond
with local volatility:

0 v r ru r w 2 S
~
r S
~
u
r w
r
u
S
~
u
S
~
) 2 ( S
~
) (r
S
~
u
ar) (
r
u
t
u
c
r , S
~
S X X , S
2
S
2
X
2
2
2
2
2
1
2
2
2 2
S S X X S,
2
X 2
1
= + +

+ + +

F. Boundary and Initial conditions: COCB case
The coupling in the equations (11) & (13b) is apparent through the term r
c
v, which must be
available at the time of solving for the new time step u
n
given the value at u
n+1
. Crank-
Nicholson with PSOR is used to solve for u
n
above, but using the previous value for v, i.e.
v
n+1
. The algorithm for v is a separate finite difference with v
n
obtained from v
n+1
using the
Crank-Nicholson scheme with the boundary/initial conditions obtained from the solution for
the convertible bond, u. That is, the boundary/initial conditions for v are given by:

Initial conditions:
(12)
(13a)
(13b)


21

+ + =
+ + =
coupon) FV coupon, max(PP(T)
AcInt(T) Value Face AcInt(T), max(PP(T)
maturity at or prior to converted was bond the if 0
T) r(T), (T), S
~
v(


FV) max(PP, S
~
n
f
f
Coupon)I FV] (max[PP, T) r(T), (T), S
~
v( i.e.
<
+ =

Boundary conditions:
0 ) S
~
( v
max
n
l , m
= where n refers to the time step, m refers to the asset step, and l to the interest
rate step. The (coupled) boundary conditions that are used on the grid are:

=
+
+ +
=
s) m ( n nS u if 0
AcInt(t) CP(t) u if 0
AcInt(t) PP(t) u if AcInt(t) PP(t)
v
n
l m,
n
l m,
n
l m,
n
l m,


l m, u v
n
l m,
n
l m,
(the contemporaneous value of u is an upper bound for v).

These are combined with the convertible bond initial/boundary conditions for u below.

G. Boundary and initial conditions for the convertible bond
Initial condition:
Coupon) CP(T) (T), S
~
min(CR. Coupon(T), PP(T) Coupon(T), max(FV T) r(T), (T), S
~
u(
f f
+ + + =

Boundary condition:

S
~
S
~
on everywhere 0
S
u

callable u for 0
callable - non u for n
s
u

callable u for 0
callable - non u for S
~
. CR
) t , r , S
~
u( S
~
S
~
max
2
2
max
S
~
max
max max f
= =

= =


We also impose the continuous upper and lower constraints:

(14)
(15)


22
AcInt(t)]] S
~
CR. , AcInt(t) Min[CP(t) , AcInt(t) max[PP(t) u
n
l m,
+ + + =

where n refers to the time step, m refers to the asset step, and l to the interest rate step. The
solution proceeds as follows: first u
T-1
is solved for using Crank-Nicholson with u
T
, and v
T
,
known from initial conditions. Then solve for v
T-1
using the boundary conditions for v
T-1
,
which depends on u
T-1
, from above. In this way, iterate between u, and v, and solve until u
0
is
obtained.

III. Numerical Methods
A. PDE formalism and general issues
For parabolic PDEs, the time and space dynamics do not evolve jointly. There are no mixed
time and space (equity and interest rate) partial derivatives in the PDE. Therefore we can
solve the convertible bond problem in two distinct steps: first construct a grid corresponding
to the space discretization; then proceed to the second step via a backward procedure where
early call, Bermudan put, discrete dividends, and any other time-dependent jump conditions
are applied in the successive over-relaxation routine. This starts with the last time-step value
of the convertible bond as an initial iterate, and continues the iteration towards a new solution.
In this way the price one time step before is obtained. Proceed in this way until the time zero
value has been obtained. Formally, we have the relationship

r D D with v r u D v r ru Du
t
u
U c U c
+ = + = + + =

where:

r , S
~
2
S S X X S,
2
X
2
2
2
2
2
2
2 2
f S X X S,
2
X
r w 2 S
~
r S
~
r w
r 2
1
S
~
S
~
) 2 (
2
1
S
~
) (r
S
~
ar) (
r
D
+ +

+ + +

=


And

) r r ( D D with v D v ) r r ( Dv
t
v
c V V c
+ + = = + + =

(16)


23
The time, equity, and interest rate grid points are indexed respectively as m, n, and l, with
corresponding step sizes M / S
~
S
~
, L / r r , N / T t
max
max
= = = , for N, M, and L taking
values in the set of positive integers. The notation
n
l , m
u then corresponds to the value of the
convertible at the grid points. In the following it is emphasized that the time index n refers to
calendar time, not inverse time
20
.

The Crank-Nicholson two-step procedure is then cast as a set of two ordinary differential
equations:

1 n
j , i c
j i,
n
j i,
n
j i, l, m, U, 2
1
j i,
1 n
j i,
1 n
j i, l, m, U, 2
1
n
l m,
1 n
l m, l m,
1 n
c
n n
U 2
1
1 n 1 n
U 2
1
c U 2
1
U 2
1
v r u D

u D

t
u u
dt
du
where
v r u D

u D

v r t) u(t D

u(t) D

dt
du(t)
+ + +
+
+ + +
+ + =

=
+ + = + + + =



And for the COCB:


+ =

=
+ = + + =
+ +
+
+ +
j , i
n
j , i
n
j , i , l , m , V 2
1
j , i
1 n
j , i
1 n
j , i , l , m , V 2
1
n
l , m
1 n
l , m l , m
n n
V 2
1
1 n 1 n
V 2
1
V 2
1
V 2
1
v D

v D

t
v v
dt
dv
where
v D

v D

) t t ( v D

) t ( v D

dt
) t ( dv


That is,

1 n
j , i c
j , i
1 n
j , i
1 n
j , i , l , m , U 2
1 1 n
j , i
n
j , i
n
j , i , l , m , U 2
1 n
1 n
c
1 n 1 n
U 2
1 n n
U 2
1
v r u D

u
t
1
u D

u
t
1
v r u D

1
t
1
u D

1
t
1
+ + + +
+ + +

|
|
.
|

\
|
=
|
|
.
|

\
|
+
|
.
|

\
|
= |
.
|

\
|
+



And

|
|
.
|

\
|
=
|
|
.
|

\
|
+
|
.
|

\
|
= |
.
|

\
|
+

+ + +
+ +
j , i
1 n
j , i
1 n
j , i , l , m , V 2
1
1 n
j , i
n
j , i
n
j , i , l , m , V 2
1
n
1 n 1 n
V 2
1 n n
V 2
1
v D

v
t
1
v D

v
t
1
v D

1
t
1
v D

1
t
1


20
By inverse time we refer to time evolving backwards away from the maturity of the convertible.
(17)
(18)
(19)
(20)


24

Here
j , i , l , m ; U
D

and
j , i , l , m ; V
D

refer to the space discretization matrices (where each element

itself consist of another matrix) for u and v, representing the (nearly everywhere second order
accurate) r and S
~
discretization schemes applied for the partial differentials in the pricing
equations (11) and (13b). The indices i and j refer to the lag and lead terms in the space (r and
S
~
) discretization, and are typically in the range 3 l and , 3 m , depending on whether
centred (in the centre of the grid) or non-centred (in the boundaries) discretization schemes
are employed. As the final payoff for the convertible and the COCB is known, the problem
consists of solving the equation above backward in time to obtain todays values, subject to
boundary conditions for the coupled PDEs. Truncation points for the grid are chosen with the
intent to delimit the grid to those plausible regions. Sensible choices are
max
S
~
at 3 standard
deviations from its expected future value (typically in the magnitude of 1200% of initial
value for a five year convertible bond), 0 S
~
min
= ,
max
r =20-25%, and
min
r =0.

We can impose that the credit spread is some function ( )
n
l , m c
S
~
r , dependent on the location in
the grid. The effect of negative correlation between credit spreads and equity can be
incorporated in this deterministic fashion. An intuitive and simple functional relationship was
suggested by Olsen (2002):

0 0
) S S )( h h ( h ) S ( h


+ =

here h(S) is the spread (c.f. r
c
in equations (19)-(22)) as a function of S,

h is the spread as
the stock price tends gets very large (which reflects the lowest credit spread for debt of
comparable maturity, ranking, and subordination as the convertible being values), and h
0
is
the spread associated with todays stock price S
0
. The decay parameter is a measure of
sensitivity of the credit spread to the stock price and can be estimated using a KMV type
model or empirically using panel time series data of yields/stock prices for nonconvertible
bonds of the same issuer with similar subordination, ranking, and duration. Olsen (2002)
reports this parameter to lie in the range 0.5 to 2.0 for most issues.

B. Computational Method


25
The solution to the matrix equations and the underlying PDE depends closely on the choice of
the space discretization matrix D

above. High order accuracy for the finite difference

approximations depends on the choice of scheme at the boundaries and the interior of the
grid. In the appendix we consider the discretized PDE in various regions of the grid and report
on the accuracy of the schemes. The space discretization (propagator) matrices
j , i , l , m ; U
D

and
j , i , l , m ; V
D

are composed of M+1 block matrices of dimension L+1 each, and look as
follows:


MLMat MLMat MLMat
MLMat MLMat MLMat
MLMat MLMat MLMat
MLMat MLMat MLMat


MLMat MLMat MLMat
MLMat MLMat MLMat
MLMat MLMat MLMat


Here each object of type MLMat is an (L+1) by (L+1) matrix encapsulating the spatial
discretization in the interest rate axes of the grid. There are (M+1) rows of these block
matrices (for the S
~
axis discretization). Note that the propagator matrix is sparse because we
use neighboring points at most 2dr 2dS, away from a reference point
n
l , m
u and
n
l , m
v in
calculating the partial derivative terms in the PDE
21
. This can easily be relaxed to make the
system even more accurate at the cost of more processing power. Due to the nature of the
problem, we can design tailored matrix storage classes in C++, with operations such as the
iterative solver method described in (25) that take advantage of the sparseness of the system
to optimize the algorithm and its computations. The specialized sparse matrix above
represented in the computational model as an (M+1).(L+1) by 3.(L+1) array object, and
matrix multiplication, addition, equality, and subscript (to access elements of the matrix)
operators are overloaded to provide quick access for operations. Each MLMat looks as
follows:


21
See the Appendix on discretization schemes for details.
M+1
M+1


26
1) 1)x(L (L
1 L m, L m, 1 L m,
1 L m, L m, 1 L m,
L m, 1 L m, 2 L m,
m,0 m,0 m,0
m,2 m,1 m,0
m,2 m,1 m,0
) (u f

) (u f

) (u f

0 0 0 0 0
) (u f
~
) (u f
~
) (u f
~
0 0 0 0 0
0 ) (u f
~
) (u f
~
) (u f
~
0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 ) (u f
~
) (u f
~
) (u f
~
0
0 0 0 0 0 ) (u f
~
) (u f
~
) (u f
~
0 0 0 0 0 ) f(u ) f(u ) f(u
MLMat
+ +
+
+

(
(
(
(
(
(
(
(
(
(
(

=
M M M
M M M

Note the tilda and stet formatting characters on the discretization function f above for the
interest rate partial derivatives with stock price level kept constant. This is to emphasize the
different discretization schemes that are effective in the interior and the boundaries of the
solution domain.

Recall, from section 3, that the value of the convertible and the COCB are constrained
between upper and lower bounds as follows:

n
U
n
l , m
n
L
U u U and
n
l , m
n
l , m
u v 0 , where it will be recalled that:
AcInt(t) CP(t) U
n
U
+ = and ) S
~
CR. , AcInt(t) PP(t) ( max U
f
n
L
+ =

We solve systems (19) and (20) as follows. First, from the initial conditions
mat
l , m
u is
computed. From these values, we can compute the maturity value of the COCB applying the
above constraints. From there, we next solve the system (19) one time step before maturity
using the Gauss-Seidel routine with successive overrelaxation as follows:

1 n
j i, c
1 n
j i,
j i,
1 n
j i, l, m, U, 2
1
j i,
n
j i, l, m, U, 2
1
v r u D

1
t
1
b
D

1
t
1
A
+ + +
+
|
|
.
|

\
|
=
|
|
.
|

\
|
+ =

Then b A =
n
j i,
u . In (21) A is a matrix of dimension (M+1)*(L+1) by (M+1)*(L+1), and b is
a vector of size (M+1)*(L+1) by 1.

(21)


27
Given the kth step of the Gauss-Seidel iteration
k n,
j i,
u the k+1th iteration can be found
(applying the projection constraints) in the following way.

|
|
|
|
|
.
|

\
|
|
|
|
|
|
.
|

\
|
|
|
|
|
|
.
|

\
|
+ =

<
=
<
<
=
>
+ +
l j
m, i
or
m i
l j
m, i
or
m i
k n,
j i, j i, l; m,
1 k n,
j i, j i, l; m, l m,
l m, l, m,
k n,
j i,
n
l m, L,
n
l m, U,
1 k n,
j i,
u A u A b
A
1
)u (1 , U max , U min u


An identical expression as above applies to the COCB but with A and b defined as below:

b Av
v D

1
t
1
b
D

1
t
1
A
n
j i,
1 n
j i,
j i,
1 n
j i, l, m, U, 2
1
j i,
n
j i, l, m, V, 2
1
=
|
.
|

\
|
=
|
.
|

\
|
+ =
+ +

The successive overrelaxation parameter in (22) is constrained to lie between zero and two.
We take an adaptive approach to search for the optimal for a particular time step, so as to
minimize the spectral radius of the iteration matrix A. This is due to the fact that iterative
schemes will in general only converge for a spectral radius less than one (Tavella and Randall
(2000)). It is known that the spectral radius is bound above by

l. m, given A max
l) (m, j) (i,
j i,
j i, l; m,
|
|
.
|

\
|

So it is easily seen that this method will converge for diagonally dominant matrices. It is
worth observing from (21) and (23) that taking small time steps dt will improve convergence
of the iterative technique.

IV. Numerical Results
We proceed first with a simplified example of a 5 year domestic convertible bond with a
coupon of 5%, flat volatility at 22%, and the credit spread associated with nonconvertible
(22)
(23)


28
debt of the same company is 4% over treasury rates. This example is an illustration of the
smooth price profiles that can be obtained from applying the model. Notice that the upper
price level is pruned off beyond the call price of US$ 175, effective as soon as hard call
protection expires in 3 years from issue date, after which the issue is callable anytime until
maturity. The truncation above call price is applied in the projective solver in (22), but can
readily be changed, at the input of the user, to some threshold k% above the call price
prevailing at the time, to accommodate empirical stylized features such as the late call safety
net. We have suppressed credit spread dependency on stock price for this example. As a
result, we observe that as stock prices go to zero the convertible price declines to a non-zero
bond floor. Currency risk is abstracted (Thus the bond is domestic with underlying and
interest payment in USD)
22
. Due to the fact that most of the more complicated features of the
model are not in use, the example serves as a benchmark for the examples that follow. The
term sheet for the convertible bond is presented in Table I.

Table I
Domestic Convertible Bond with Flat Volatility and no put features
Convertible Bond Term Sheet

Maturity 5.0 Years ConvRatio 1.5 Put No
Coupon 5% Volatility 22% S-r correl -0.5
Frequency Semi-ann Call Protection >3yrs @175 S(0) US$ 48.0
Spread 4% Model Price US$ 106.872 r(0) 0.03
ConvPremium 48.4% Conversion Price US$ 72.00 S_Max US$ 1200


Note that with a model price of US$106.872, this represents a conversion premium for the
issue of 48.4%. We present the price evolution of the convertible with calendar time in Figure
1.

22
We test in effect the simplest case here, with the COCB sensitive to interest rate risk as well as stock price,
although all other features of the model are switched off.


29
5
4.4
3.8
3.2
2.6
2.005
1.5
0.9
0.3
0
2
4
4
8
7
2
9
6
1
2
0
1
4
4
1
6
8
1
9
2
0
50
100
150
200
250
300
Convertible Price
Calendar years
Stock Price


Figure 1. (S,t) evolution of the convertible price, with time to maturity in calendar years depicted on
the x-axis. Notice the stepping behaviour of the price about the coupon dates, when the CB price drops
by an amount equal to the coupon payout, which in the example above is 5% (of $100 face value) paid
semi-annually.

The price profile displays the expected behaviour. We report a model price of $106.872.
Notice the cut-off after the hard call protection feature ceases. In addition to this, as the bond
has no put features, there is no sudden jump in price across discrete (put) dates at low stock
price levels. As stock price declines to zero, the convertible bond acts like a fixed income
instrument but the bond floor itself does not decay to zero. Observe also the discrete jumps in
price as a coupon is paid and the convertible value drops by the coupon amount. Next, we
present the interest rate versus stock price sensitivity of the issue from Table I.





30
0
0
.
0
3
0
.
0
6
0
.
0
9
0
.
1
2
0
.
1
5
0
.
1
8
0
.
2
1
0
.
2
4
0 2
4 4
8 7
2 9
6
1
2
0
1
4
4
1
6
8
1
9
2
0
50
100
150
200
250
300
Convertible Price
Interest rate
Stock Price

Figure 2. Convertible bond price at time zero as a function of stock price and interest rates.

Figure 2 displays the price behaviour as a function of stock price and interest rates at a
reference time of zero. The pricing grid is set up such that an upper interest rate level of 25%
applies. Note how the convertible bond price is driven by the interest rate at lower values of
stock price. In this region the instrument behaves more like a corporate bond. At high enough
stock price levels, the price impact of interest rates is less discernible.

We present price results for the COCB in Figure 3. Recall that callability kicks in after 3
years, and we see a more pronounced fall in COCB price as a function of stock price after this
date. While the convertible is protected against call (0-3 years) the COCB price declines more
smoothly as a function of stock price as conversion (but not call) becomes likely. The COCB
price falls in a step fashion across coupon dates as coupons are paid out.






31
5
4
.
5
0
5
4
.
1
3
.
7
3
.
3
2
.
9
2
.
5
2
.
0
0
5
1
.
6
1
.
2
0
.
8
0
.
4
0
2
4
4
8
7
2
9
6
1
2
0
1
4
4
1
6
8
1
9
2
0
20
40
60
80
100
120
COCB price
Calendar years
Stock price


Figure 3. Cash only part of the convertible (COCB) evolution from maturity (5 years) to today (which
corresponds to point 0 on the y-axis above. Note the sharp (but smooth) fall in COCB price in high
regions of S as conversion and early call features become the main price drivers.

We turn now to the pricing of a fully fledged cross-currency convertible with local volatility
sensitivity to both foreign exchange and equity implied volatility skew and term structures. In
this mode, the algorithm progresses as follows. First, an implied volatility surface is
calculated for the foreign currency in domestic units using the algorithm sequences in
equations (7)(9). Then local volatility is found using second order accurate discrete
approximations to (10) (although the choice can be made to use either Wolfes (1959)
algorithm or the NAG routines at the expense of some speed). Then the system of linear
algebraic equations (21) and (23) are applied recursively for the convertible and cash only
part.

We begin with the input implied volatility surfaces for FX and Equity in foreign units and the
double integration routines in equations (7)-(9). These surfaces are constructed via simulation
and display the properties of the equity and FX options markets, although a conscious choice


32
is made to make the input display sharp second derivatives to stress the local volatility
routines.

For the FX local volatility, the following USD/GBP implied volatility surface is taken. This is
displayed in Figure 4. The at-the-money FX rate is USD 1.5 per GBP.

0
0
.
3
7
5
0
.
7
5
1
.
1
2
5
1
.
4
5
1
.
5
1
.
5
5
1
.
8
7
5
2
.
2
5
3
4
.
5
0
0
.
1
0
.
2
5
0
.
5
1
1
.
5
2
3
5
10.00%
12.00%
14.00%
16.00%
18.00%
20.00%
22.00%
Implied vol
Strike (ATM = 1.5)
Time to maturity (yrs)


Figure 4. Approximately symmetric implied volatility surface for USD/GBP exchange rate used in
double integration routine (7)-(9)

The equity implied volatility surface is presented in Figure 5. It exhibits monotonic
decreasing behaviour with moneyness which flattens out with maturity.




33
0.0
11.1
22.2
33.3
40.0
44.4
48.9
55.6
66.7
88.9
133.3
0
0.1
0.25
0.5
1
1.5
2
3
5
0.21
0.23
0.25
0.27
0.29
0.31
0.33
0.35
0.37
Impl.Volatility
Stock price (GBP)
ATM=GBP 44.4
To maturity (years)

Figure 5. Implied volatility surface for foreign equity used for double integration routine in (7)-(9).
Notice the flattening in implied volatility with maturity. The first and second derivatives with strike
and maturity above display a degree of non smoothness. This is done deliberately to stress the
algorithm.

We can then obtain the implied volatility surface for the foreign equity in domestic currency
using the numerical integration routines in equations (7)-(9). This is displayed in Figure 6.
Notice the tent-like shape of the implied surface. It now displays a humped centre due to the
combined effect of the FX surface and the stock surface. We use this surface to obtain the
local volatility for the international convertible bond.


34
0.0
16.7
33.3
50.0
60.0
66.7
73.3
83.3
100.0
133.3
0
0
.
1
0
.
2
5
0
.
5
1
1
.
5
2
3
5
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Volatility
Scaled strike (in USD)
Time (to maturity)


Figure 6. Implied volatility surface the scaled variable S
~
obtained from double integration over the
bivariate lognormal densities of S and FX. Displays humped centre due to influence of symmetric FX
smile which has feedback to the joint density in equations (8) (9)

The details of the cross-currency convertible bond are now given. The maturity is still set at
five years, at the money S
~
is US$ 66.66, the initial domestic short rate is 6%, 3 . 0
r , S
~ = ,
the dividend yield 1.54%, the coupon 2% (semi-annual), and the conversion ratio is 1.2. The
convertible is callable after three years at a price of US$ 175.0. There is no soft call
protection. The bond is also putable in a Bermudan fashion on certain discrete times during
its lifetime. The put dates are two and four years from issue date, at prices of US$ 86.0708
and US$ 95.12294 respectively. We use the same parameters for the CIR process for the short
rate as in the previous example. The implied volatility surface for foreign equity in domestic
units is given in Figure 6. The model price is US$ 98.798 versus a parity of US$ 79.999. For
simplicity and for comparison with the example that follows, we impose that the credit spread
is independent of the stock price in the mesh, so the convertible is floored at its bond price.
This example provides us with the first indication of the impact of local volatility on pricing.
To assess this consider a trader using constant volatility who, as a crude indication, might


35
choose the implied volatility corresponding to the maturity and strike of the instrument, which
as can be seen in Figure 6, is in the region of 19% . If he applies this rule, he would value the
convertible bond with constant volatility at US$ 95.076, which is different from the model
price by 3.77%. Alternatively, he might use the implied volatility from Figure 6
corresponding to his expectations of the expected lifetime of the convertible bond. While this
might produce an answer closer to the model price, it is dependent on the traders guesswork
and not consistent in any case with the information available in the vanilla options market
from which the local volatility surface would have been obtained.


Table II
International Convertible Bond
with Local Volatility
Call and Bermudan Put features
Convertible Bond Term Sheet

Maturity 5.0 Years ConvRatio 1.2 Put Bermudan
Coupon 2% Volatility method Local Put (4yrs) US$ 95.12
Frequency Semi-ann Call Protection >3yrs @175 Put (2yrs) US$ 86.07
Spread 3.5% Model Price US$ 98.7982 S-r correl -0.3
ConvPremium 23.5% Price @ flat vol US$ 95.076 S(0) US$ 66.66
Share Currency GBP Conversion Price US$ 80.0 r(0) 0.06
Denominated USD Flat Vol 19% S_Max US$ 1200


The pricing profiles obtained from this example are presented in Figures 7 and 8. Note that
due to the non-collapsing bond floor at low stock price levels, the influence of the put feature
is not very prominent in the figures that follow. This situation is reversed when we allow a
deterministic relationship between credit spreads and the stock price in the next example.
Also notice the swelling in the convertible bond (Figure 7) and COCB price profile (Figure 8)
at times close to zero. This is the influence of the local volatility, which at low maturities and
strikes is most pronounced. This is corroborated by looking at the implied volatility surface
for the scaled stock price in Figure 6.






36
5
4
.
5
4
3
.
5
3
2
.
5
2
1
.
5
1
0
.
5
0
.
0
0
1
6
.
6
7
3
3
.
3
3
5
0
.
0
0
6
6
.
6
7
8
3
.
3
3
1
0
0
.
0
0
1
1
6
.
6
7
1
3
3
.
3
3
1
5
0
.
0
0
1
6
6
.
6
7
0
50
100
150
200
250
CB price
Time to maturity
Stock price


5
4.005
3.1
2.2
1.3
0.4
0
.
0
0
1
6
.
6
7
3
3
.
3
3
5
0
.
0
0
6
6
.
6
7
8
3
.
3
3
1
0
0
.
0
0
1
1
6
.
6
7
1
3
3
.
3
3
1
5
0
.
0
0
1
6
6
.
6
7
1
8
3
.
3
3
2
0
0
.
0
0
0
20
40
60
80
100
120
COCB price
Time to maturity
Stock price

Figures 7 & 8- Price versus short rate profile for time-0 for the cross-currency convertible bond.
Notice the price profile displays exactly what we would expect.


37
For the final example, we impose the following functional form between spreads and stock
prices in all regions of the dual mesh
23
, described by the equations (19) and (20), and
assuming the parameters have been estimated from data:

65 . 0
0 0 c

0 0 c
) S S )( 00545 . 0 ) S ( r ( 00545 . 0 ) S S )( h h ( h ) S ( r


+ = + =

For the example, we take

h (the spread prevailing at sufficiently high stock price levels) at

55 basis points above the risk free rate. In addition, we take different levels of initial spread
) S ( r
0 c
(observable in the market) of 2%, 3.5%, 5%, and 8% into consideration. The pricing
is done at two reference levels,
r , S
~ = -0.30 and -0.05. The contract features, with the
exception of the above, remain the same as in Table II. At a reference spread of 3.5% and
r , S
~
= -0.3, we find the model price as US$ 97.3653 versus a price of US$ 98.7982 when the
functional form (24) is dropped. It is seen that, even for a moderate estimate for like the one
used in (24), the price impact on the convertible can be significant. This is seen most
strikingly when the price profiles are compared with the earlier profiles (Figures 7 and 8).

Table III
International Convertible Bond
with Local Volatility and Spread/Stock Price Relationship
65 . 0
0 0 c

0 0 c
) S S )( 00545 . 0 ) S ( r ( 00545 . 0 ) S S )( h h ( h ) S ( r


+ = + =
Call and Bermudan Put Features
Convertible Bond Term Sheet

Maturity 5.0 Years ConvRatio 1.2 Put Bermudan
Coupon 2% Volatility method Local Put (4yrs) US$ 95.12
Frequency Semi-ann Call Protection >3yrs @175 Put (2yrs) US$ 86.07
Spread 3.5% Model Price US$ 97.3653 S-r correl -0.3
ConvPremium 17.8% Price @ flat vol US$ 94.431 S(0) US$ 66.66
Share Currency GBP Conversion Price US$ 80.0 r(0) 0.06
Denominated USD Flat Vol 19% S_Max US$ 1200


23
We set the parameter k=0.65 ourselves to demonstrate how the model will capture the collapse of the bond floor,
and it has not been estimated from data. For a discussion of how to obtain it empirically, see Olsen (2002)
(24)


38
5
4.2
3.5
2.7
2
1.2
0.5
0
.
0
0
1
6
.
6
7
3
3
.
3
3
5
0
.
0
0
6
6
.
6
7
8
3
.
3
3
1
0
0
.
0
0
1
1
6
.
6
7
1
3
3
.
3
3
1
5
0
.
0
0
1
6
6
.
6
7
1
8
3
.
3
3
0
50
100
150
200
250
Convertible price
Time to maturity (years)
Stock price


Figure 9. Evidence of how the bond floor (as a minimum for the convertible bond value) is collapsing
as stock price drops sharply. This should be part of the reality check of a candidate model, as the
phenomenon depicted above will often be seen in the markets. Note too that the Put feature kicks in at
2 and 4 years, thereby cushioning the sharp fall in price as stock prices tend to zero. We can also
observe the influence of high local volatility at short maturities.

Even more telling is the behaviour of the COCB in low stock price regions. In the original
flavour, the COCB would enjoy some of its highest values in low stock price regions as it is
least likely that the bond will be converted and the future fixed cash payments are most safe.
This is readily seen in figure 2. Now we contrast this with the conspicuous change that can be
observed in the COCB price profile under the parameterization given by (26) and local
volatility:







39
5
4
.
6
4
.
3
4
3
.
6
3
.
3
3
2
.
6
2
.
3
2
1
.
6
1
.
3
1
0
.
6
0
.
3
0
.
0
0
3
3
.
3
3
6
6
.
6
7
1
0
0
.
0
0
1
3
3
.
3
3
1
6
6
.
6
7
2
0
0
.
0
0
0
20
40
60
80
100
120
COCB price
Time to maturity
Stock price


Figure 10. The tent shape of the cash only part of the convertible bond conforms to empirical
expectation- the COCB has highest value when the corresponding convertible bond is least likely to be
converted by the investor, but as stock price tends to zero the COCB value breaks down and the
instrument is most likely to default.

This figure is best contrasted with that in Figure 7. It is seen in addition to the tent shape due
to the spread/stock price relationship, that the Bermudan put feature increases the COCB
value on the put dates. As another check on the model, we probe its sensitivity to various
market spreads and find the correct behaviour is retained as spreads change. Notice too the
persistence of the discrepancies arising from using constant volatility.









40
Table IV
Convertible Bond Prices for a Selection of Credit Spreads
and
S,r

VOLATILITY Local VOLATILITY Constant 19%
-0.3 -0.3
Spread 2% 3.50% 5% 8% Spread 2% 3.50% 5% 8%
Price 102.899 97.3653 92.8311 87.512 Price 99.709 94.431 89.748 82.985
VOLATILITY Local VOLATILITY Constant 19%
-0.05 -0.05
Spread 2% 3.50% 5% 8% Spread 2% 3.50% 5% 8%
Price 103.368 97.761 93.183 87.772 Price 100.08 94.765 90.0451 83.26



V. Conclusion
This paper makes several contributions. First, we present a comprehensive theoretical
framework for pricing convertible bonds in the presence of interest rate, equity, credit,
currency, and volatility risk and employ a quasi five factor model incorporating most of the
contractual features. Given that the market in convertible bonds is in excess of USD 400
billion it is surprising that unlike in the bond or interest rates markets there is little consensus
on the best way to price these instruments. There is little agreement too as to which risk
factors should not be excluded, as using extra risk factors leads to complications arising
invariably from the increase in PDE dimension. We develop and implement a model that
includes all of the most important sources of risk that influence convertible bond value.
Second, we propose and implement a separate, computationally intensive, and
unconditionally stable multifactor model for the credit risk component of convertible bonds,
which moreover is capable of capturing the credit spread - stock price correlation, and the
collapsing bond floor dynamics as stock prices fall. The numerical results indicate that both
the multifactor extension as well as the correlation impacts prices significantly. Our third
contribution is our use of local volatility, which has hitherto not been investigated in
convertible bond pricing. Convertible bonds prices will display the same sensitivity to short
and medium term volatility as other exotics. By marginalizing the issue of volatility
sensitivity for convertible bonds, earlier studies have implicitly endorsed the view that the
(relatively) longer maturity of convertible bonds obviates the need to account for price
sensitivity to the local volatility especially in the short to medium maturity brackets. We
document that using constant volatility creates considerable inconsistencies, not least because
the trader must enter a point estimate of this for the maturity of the convertible bond.
Additionally, the choice of this constant parameter is invariably subjective and in any case


41
inconsistent with the input options data. Our study shows that pricing discrepancies of several
percentage points occur when at the money forward implied volatility (with maturity
coinciding with the bond maturity) is used instead of local volatility.

Given that a considerable number of liquid and large convertible issues display currency
sensitivity, we have included currency risk in our framework by employing a standard
technique from the literature to show that currency risk for international convertible bonds
does not introduce an extra factor in the PDE. Our fourth contribution consists of merging the
local volatility approach with the change of numeraire method which we utilize to keep the
problem dimension lower. The difficulty with this is that options prices on the foreign equity
in domestic currency are non-traded, and it is necessary to back these out to obtain the local
volatilities for convertibles bonds with currency risk. We propose an accurate procedure to
back out the implied volatility of options on the foreign equity in domestic currency, as
implied by the market prices of options on foreign exchange and foreign equity in foreign
currency units. Two methods from the literature are proposed to get the local volatility from
the obtained implied volatility surface. The first of these (Wolfes) can be used to place
smoothness constraints in problematic neighbourhoods of the solution domain. Finally, we
have provided a detailed computational framework for implementing convertible bond pricing
models. A variety of instruments and models can be accommodated in this framework,
including international, domestic, exchangeable, and resetting convertible bonds, not to
mention a larger family of non convertible exotic instruments (barriers, Asians, digitals) that
display sensitivity to two risk factors, with or without credit local volatility, credit, or
currency risk.




42
References

Amman, M., Kind, A.H., and C. Wilde, March 2001, The Pricing of Convertible Bonds: An
Analysis of the French Market. Department of Economics Working Paper no. 2001-02,
University of St. Gallen
Andersen, L.B.G and R. Brotherton-Ratcliffe, Winter 1998, The equity option volatility
smile: an implicit finite-difference approach, Journal of Computational Finance, 1,2, 5-37
Asquith, P., Convertible Bonds are not called late, 1995, Journal of Finance, 50, 1273-1289.
Barone-Adesi, G., Bermudez, A., and J. Hatgioannides, 2002, Two-Factor Convertible Bond
Valuation Using the Method of Characteristics/Finite Elements, preprint, City University
Business School, London
Brennan, M. and E. Schwarz, 1980, Analyzing convertible bonds, Journal of Financial and
Quantitative Analysis, 15, 907-929
Brennan, M. and E. Schwarz, 1977, Convertible Bonds: Valuation and Optimal Strategies for
Call and Conversion Journal of Finance, 5, 1699-1715.
Calamos, J.P, 1998, Convertible Securities (McGraw-Hill).
Dupire, B, July 1996, Pricing with a smile, Risk 45-48
Dunn, K., and K. Eades, 1984, Voluntary conversion of convertible securities and the optimal
call stratategy, Journal of Financial Economics, 23, 273-301
Davis, M. and F.R. Lischka, 1999, Convertible bonds with market risk and credit risk, Tokyo-
Mitsubishi International Plc
Duffie, D., and K. Singleton, 1999, Modelling the Term Structure of Defaultable Bonds.
Review of Financial Studies, 12, 687-720
Geman, H., El Karoui, N., and J.C. Rochet, 1995, Changes of Numeraire, Changes of
Probability Measures and Pricing of Options. Journal of Applied Probability, 32, 443-458
Goldman Sachs. Valuing convertible bonds as derivatives, November 1994, Quantitative
Strategies Research Notes, Goldman Sachs, New York
Grimwood, R. and S. Hodges, December 2001, The Valuation of Convertible Bonds: A Study
of Alternative Pricing Models, preprint, Financial Options Research Centre, Warwick
Business School
Harris, M., and A. Raviv, 1985, A sequential model of convertible debt call policy. Journal of
Finance, 40, 1263-1282
Ho, T., and D.M. Pfeffer, 1996, Convertible bonds: model, value, attribution and analytics,
Financial Analysts Journal, 52 Sept.-Oct, 35-44
Hull, J., and A. White, 1990, Pricing interest-rate derivative securities, Review of Financial
Studies, 3, 573-592
Cox, J.C., Ingersoll, J.E., and S.A Ross, 1985, A Theory of the Term Structure of Interest
Rates, Econometrica 53, 385-407
Hoogland, J., Neumann, D., and D. Bloch, July 2001, Converting the reset, Working Paper
Jackwerth, J., and M. Rubinstein, 1994, Recovering probability distributions from options
prices, Journal of Finance, 51, 1611-1631
Kunzi, H., Tzschach, H.G., and C.A. Zehnder, 1968, Numerical Methods of Mathematical
Optimization (Academic Press)
Lvov, D., and A.B. Yigitbasioglu, 2002, Numerical Issues in an Object-Oriented
Implementation of Local Volatility using PDE methods, ISMA Centre Discussion Papers in
Finance 2002 (Forthcoming)
Mikkelson,W.H, 1985, Capital Structure change and decrease in stock holder wealth: a cross-
sectional study of convertible security calls, in Benjamin M. Friedman , ed.: Corporate
Capital Structure in the United States, University of Chicago Press, pp. 265-296
Musiela, M., and M. Rutkowski, 1998, Martingale Methods in Financial Modelling
(Springer)


43
Olsen, L, 2002, Convertible Bonds: A Technical Introduction, Barclays Capital Research
Paper
ksendal, B, 2000, Stochastic Differential Equations (Springer)
Press, W.H., Teukolsky, S.A., Vetterling, W.T., and B.F. Flannery, 2002, Numerical Recipes
in C++ (Cambridge University Press)
Rebonato, R, 1999, Volatility and Correlation (Wiley)
Rubinstein, M, 1994, Implied Binomial Trees, Journal of Finance, 49, 771-818
Rubinstein, M, November 1991, Somewhere Over the Rainbow, Risk 4, 63-66
Shimko, D, April 1993, Bounds of Probability, Risk 33-37
Sirbu, M., Pikovsky, I., and S. Shreve, 2002, Perpetual Convertible Bonds, Working Paper,
Department of Mathematical Sciences, Carnegie Mellon University
Stroustoup, B, 1997, The C++ Programming Language (Addison Wesley)
Takahashi, A., Kobayashi, T., and N. Nakagawa, 2001, Pricing Convertible Bonds with
Default Risk: A Duffie-Singleton Approach. Preprint, University of Tokyo Discussion Paper
CIRJE-F-140
Tsiveriotis, K., and C. Fernandes, September 1998, Valuing Convertible Bonds with Credit
Risk, The Journal of Fixed Income, 95-102
Tavella, D., and C. Randall, 2000, Pricing Financial Instruments: The Finite Difference
Method (John Wiley & Sons)
Wolfe, P, 1959, The Simplex Method for Quadratic Programming, Econometrica, Vol. 27,
pp. 382- 398
Wilmott, P., J. Dewynne, and S. Howison, 1993, Option Pricing: Mathematical Models and
Computation (Oxford University Press)
Wu, L., and Y-K Kwok, 1997, A Front-fixing Finite Difference Method for the Valuation of
American Options, Journal of Financial Engineering, 6,2,83-97
Zhu, Y-l., and Y. Sun, Fall 1999, The singularity-separating method for two-factor
convertible bonds, Journal of Computational Finance, 3,1, 91-110


44
A. The Discretization Schemes

(i) For m = 0, l = 0, which corresponds to S = 0, r = 0, the partial differential equation
becomes

v r
r
u

t
u
c
+

The spatial discretization is then given by

1 n
0,0 c
r
n
0,2
n
0,1
n
0,0
2
1
r
1 n
0,2
1 n
0,1
1 n
0,0
2
1
n
n
0,0
1 n
0,0 0,0
v r
2
u 4u 3u

2
u 4u 3u

u u
dt
du
+
+ + + +
+
+

+
=

=

This scheme possesses second order accuracy in both
S
and
r
(see TR[00])

(ii) For m = 0, l = 1, , L-1, which corresponds to S = 0 and r 0, the partial differential
equation reduces to:

v r ru
r
u
) (ar
r
u
r w
t
u
c
2
2
2
2
1
+ +

The spatial discretization is then given by

1 n
l , 0 c
n
l , 0
1 n
l , 0
l
n
1 l , 0
n
1 l , 0
l
1 n
1 l , 0
1 n
1 l , 0
l
2
n
1 l , 0
n
l , 0
n
1 l , 0
l
2
2
1
2
1 n
1 l , 0
1 n
l , 0
1 n
1 l , 0
l
2
2
1
n
n
l 0,
1 n
l 0,
n
l , 0
v r
2
u u
r
r 2
u u
2
ar
r 2
u u
2
ar

r 2
u u 2 u
r w
r 2
u u 2 u
r w

u u
dt
du
+
+
+
+

+
+
+
+

+ +
+
+
+
|
|
.
|

\
| +
+
|
|
.
|

\
|
|
.
|

\
|
+
|
|
.
|

\
|
|
.
|

\
|
+
|
|
.
|

\
| +

|
|
.
|

\
| +
=

=

This is second order accurate in both S and r.

(iii) For m = 0, l = L (S = 0, r = r
max
0), the partial differential equation simplifies to:

v r ru
r
u
) (ar
r
u
r w
t
u
c 2
2
2
2
1
+ +

45
The spatial discretization is then given by

L 0, c
1 n
L 0,
1 n
L 0, 2
1
l
r
n
2 L 0,
n
1 L 0,
n
L 0,
2
1
l
r
1 n
2 L 0,
1 n
1 L 0,
1 n
L 0,
2
1
l
2
r
n
2 L 0,
n
1 L 0,
n
L 0,
2
1
l
2
2
1
2
r
1 n
2 L 0,
1 n
1 L 0,
1 n
L 0,
2
1
l
2
2
1
n
n
L 0,
1 n
L 0, L 0,
v r ) u (u r
2
u 4u 3u
) (ar
2
u 4u 3u
) (ar

u 2u u
r w

u 2u u
r w

u u
dt
du
+ + +
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +

|
|
.
|

\
| +
=

=
+ +

+

+

+

+ +


which is second order accurate in S and r.

(iv) For m = 1, , M-1, l = 0 (S 0, r = 0), the partial differential equation can be written as

v r
r
u

S
~
u
S
~

S
~
u
S
~

~

v r
r
u

S
~
u
S
~

S
~
u
S
~
) 2 (
t
u
c
2
2
2 2
S
~
2
1
c
2
2
2 2
S S X X S,
2
X 2
1

+ + =

where
2
S
~
~
refers to the local volatility for foreign equity in domestic currency, obtained
numerically from the integration (see section II.D ). The spatial discretization is then given
by:

0 , m c
r
n
2 , m
n
1 , m
n
0 , m
2
1
r
1 n
2 , m
1 n
1 , m
1 n
0 , m
2
1
S
n
0 , 1 m
n
0 , 1 m
2
1
m
S
1 n
0 , 1 m
1 n
0 , 1 m
2
1
m
2
S
n
0 , 1 m
n
0 , m
n
0 , 1 m
2
1
2
m
2
S
~
2
1
2
S
1 n
0 , 1 m
1 n
0 , m
1 n
0 , 1 m
2
1
2
m
2
S
~
2
1
n
n
l 0,
1 n
l 0, 0 , m
v r
2
u u 4 u 3

2
u u 4 u 3

2
u u
S
~

2
u u
S
~

u u 2 u
S
~

u u 2 u
S
~

u u
dt
du

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
| +

|
|
.
|

\
| +
=

=
+ + +
+
+

+
+
+
+

+ +
+
+

which is second order accurate in S and r.

(v) For m = 1, , M-1, l = 1, , L-1 (S 0, r 0) the partial differential equation can be
written as:



46
v r ru
r
u
) (ar
S
~
u
S
~
r) (
r
u
r w
S
~
r
u
r w
~
S
~
S
~
u
S
~

~
t
u
c
2
2
2
2
1
2
r , S
~
S
~
2
2
2 2
S
~
2
1
+ +

The spatial discretization is then given by:

l , m c
n
l , m
1 n
l , m 2
1
l
r
n
1 l , m
n
1 l , m
2
1
l
r
1 n
1 l , m
1 n
1 l , m
2
1
l
S
n
l , 1 m
n
l , 1 m
2
1
m l
S
1 n
l , 1 m
1 n
l , 1 m
2
1
m l
2
r
n
1 l , m
n
l , m
n
1 l , m
2
1
l
2
2
1
2
r
1 n
1 l , m
1 n
l , m
1 n
1 l , m
2
1
l
2
2
1
S r
n
1 l , 1 m
n
1 l , 1 m
n
1 l , 1 m
n
1 l , 1 m
2
1
r , S
~
l
S
~
m
S r
1 n
1 l , 1 m
1 n
1 l , 1 m
1 n
1 l , 1 m
1 n
1 l , 1 m
2
1
r , S
~
l
S
~
m
2
S
n
l , 1 m
n
l , m
n
l , 1 m
2
1
2
m
2
S
~
2
1
2
S
1 n
l , 1 m
1 n
l , m
1 n
l , 1 m
2
1
2
m
2
S
~
2
1
n
n
l m,
1 n
l m, l , m
v r ) u u ( r
2
u u
) ar (
2
u u
) ar (
2
u u
S
~
) r (
2
u u
S
~
) r (

u u 2 u
r w

u u 2 u
r w
4
u u u u
r w
~
S
~
4
u u u u
r w
~
S
~

u u 2 u
S
~

u u 2 u
S
~

u u
dt
du
+ + +
|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +
=

=
+ +
+

+
+
+
+

+
+
+
+

+ +
+
+ + + +
+

+
+
+
+
+
+ +
+
+

+ +
+
+

which is second order accurate in S and r.

(vi) For m = 1, , M-1, l = L (S 0, r = rmax 0) the partial differential equation can be
written as:

v r ru
r
u
) (ar
S
~
u
S
~
r) (
r
u
r w
S
~
r
u
r w
~
S
~
S
~
u
S
~

~
t
u
c
2
2
2
2
1
2
r , S
~
S
~
m
2
2
2
m
2
S
~
2
1
+ +

The spatial discretization is then given by:



47
L , m c
n
L , m
1 n
L , m L 2
1
r
n
2 L , m
n
1 L , m
n
L , m
2
1
L
r
1 n
2 L , m
1 n
1 L , m
1 n
L , m
2
1
L
S
n
L , 1 m
n
L , 1 m
2
1
m L
S
1 n
L , 1 m
1 n
L , 1 m
2
1
m L
2
r
n
2 L , m
n
1 L , m
n
L , m
2
1
L
2
2
1
2
r
1 n
2 L , m
1 n
1 L , m
1 n
L , m
2
1
L
2
2
1
S r
n
1 L , 1 m
n
L , 1 m
n
1 L , 1 m
n
L , 1 m
2
1
r , S
~
L
S
~
m
S r
1 n
1 L , 1 m
1 n
L , 1 m
1 n
1 L , 1 m
1 n
L , 1 m
2
1
r , S
~
L
S
~
m
2
S
n
L , 1 m
n
L , m
n
L , 1 m
2
1
2
m
2
S
~
2
1
2
S
1 n
L , 1 m
1 n
L , m
1 n
L , 1 m
2
1
2
m
2
S
~
2
1
n
n
L m,
1 n
L m, L , m
v r ) u u ( r
2
u u 4 u 3
) ar (
2
u u 4 u 3
) ar (
2
u u
S
~
) r (
2
u u
S
~
) r (

u u 2 u
r w

u u 2 u
r w
2
u u u u
r w
~
S
~
2
u u u u
r w
~
S
~

u u 2 u
S
~

u u 2 u
S
~

u u
dt
du
+ + +
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +
=

=
+

+

+
+
+

+
+

+

+
+ +
+

+

+
+
+
+
+
+

+ +
+
+

This is second order accurate in S and first order accurate in r.

(vii) For m = M, l = 0, corresponding to S = S
max
, r = 0 the partial differential equation can be
written as:

0
S
~
u
, S
~
S
~
at that assumed we as v r
r
u

S
~
u
S
~

t
u
2
M
2
max c
=

The spatial discretization is then given by:

M,0 c
r
1 n
M,2
1 n
M,1
1 n
M,0
2
1
r
1 n
M,2
1 n
M,1
1 n
M,0
2
1
S
n
2,0 - M
n
1,0 - M
n
M,0
2
1
M
S
1 n
2,0 - M
1 n
1,0 - M
1 n
M,0
2
1
M
n
n
M,0
1 n
M,0 M,0
v r
2
u 4u 3u

2
u 4u 3u

2
u 4u 3u
S
~

2
u 4u 3u
S
~

u u
dt
du

|
|
.
|

\
| +

|
|
.
|

\
| +

|
|
.
|

\
| +
+
|
|
.
|

\
| +
=

=
+ + + + + +
+ + + +

which is second order accurate in S and r



48
(viii) For m = M, l = 1, , L-1 (S = Smax 0, r 0) the partial differential equation
simplifies to

boundary S
~
S
~
S
~
at the
, 0
r S
~
u
as v r ru
r
u
) ar (
S
~
u
S
~
) r (
r
u
r w
t
u
M max
M
2
c
M
M
2
2
2
2
1
= =
=

+ +

The spatial discretization is then given by:

l M, c
n
l M,
1 n
l M, 2
1
l
r
n
1 - l M,
n
1 l M,
2
1
l
r
1 n
1 - l M,
1 n
1 l M,
2
1
l
S
n
l 2, - M
n
l 1, - M
n
l M,
2
1
M l
S
1 n
l 2, - M
1 n
l 1, - M
1 n
l M,
2
1
M l
2
r
n
1 l M,
n
l M,
n
1 l M,
2
1
l
2
2
1
2
r
1 n
1 l M,
1 n
l M,
1 n
1 l M,
2
1
l
2
2
1
l M,
v r ) u (u r
2
u u
) (ar
2
u u
) (ar
2
u 4u 3u
S
~
) r (
2
u 4u 3u
S
~
) r (

u 2u u
r w

u 2u u
r w
dt
du
+ +
|
|
.
|

\
|
+
|
|
.
|

\
|
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +

|
|
.
|

\
| +
=
+
+
+ +
+
+ + +
+
+

+ +
+


This is second order accurate in S and r

(ix) For m = M, l = L (S = Smax 0, r = rmax 0) the partial differential equation becomes:

v r ru
r
u
) ar (
S
u
S
~
) r (
r
u
r w
t
u
c f 2
2
2
2
1
+ +

and the first order discretization is:

L M, c
n
L M,
1 n
L M, 2
1
L
r
n
2 - L M,
n
1 - L M,
n
L M,
2
1
L
r
1 n
2 - L M,
1 n
1 - L M,
1 n
L M,
2
1
L
S
n
L 2, - M
n
L 1, - M
n
L M,
2
1
M L
S
1 n
L 2, - M
1 n
L 1, - M
1 n
L M,
2
1
M L
2
r
n
2 - L M,
n
1 - L M,
n
L M,
2
1
L
2
2
1
2
r
1 n
2 - L M,
1 n
1 - L M,
1 n
L M,
2
1
L
2
2
1
L M,
v r ) u (u r
2
u 4u 3u
) (ar
2
u 4u 3u
) (ar
2
u 4u 3u
S
~
) r (
2
u 4u 3u
S
~
) r (

u 2u u
r w

u 2u u
r w
dt
du
+ + +
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +
+
|
|
.
|

\
| +

|
|
.
|

\
| +
=
+
+ + +
+ + +
+ + +

which is second order accurate in S and first order accurate in r


49
B. Obtaining Local Volatility
Another method for obtaining the local volatility for all ) r , S
~
( points on the pricing grid for
the convertible is via minimization of a quadratic form involving the product of a propagator
matrix of Arrow Debreu securities for FX and equity, and a function of the local volatilities.
This method is described in Andreasen et al. (1998) and Lvov and Yigitbasioglu (2002). It has
the advantage of allowing one to constrain local volatility in the grid to satisfy certain
smoothness criteria, which can improve the numerical solution when the input implied surface
is overly spiked. This is done by imposing a range constraint on local volatility at one or
several time/strike points as a natural part of the formulation of the quadratic program. It can
be invaluable when Arrow Debreu prices are highly jumpy at those points, possibly due to
bad input data. Such spikes are capable of contaminating the pricing algorithms (21) and (23)
in the neighbourhood of the spikes, which can emanate backwards towards todays price
through the time stepping PSOR routine. The numerical framework is flexible to
accommodate such smoothness constraints as part of a minimization algorithm. This might
make the routines described in (21) and (23) more reliable in a production environment. In
this vein, we could do a search on the input implied volatility smile surface before executing
the pricing routine, locating possible local volatility trouble spots via the spikes in Arrow-
Debreu profiles, and regularizing them by imposing smoothness constraints on the local
volatility as part of the quadratic program formulation, and then running the pricing routine.

The minimization problem to be solved is:

j
T
j j j
T
j
2
1
y ) c ( y Q ) y ( Minimize +

subject to
max j
y y 0 where
min j j
v v y = and
min max max
v v y = and v
max
and v
min
are
the maximum and minimum (local) variances, which can be time and space dependent (i.e.
user-defined).

The objective of the quadratic program is to minimize the deviation of the left hand side from
the right hand side in the equation below (where
j
v is the local volatility vector at time step j)

( ) ( ) ( ) ( ) ( )
(maturity) M ..., 0, j for
A A A 1 P P 1 A 1 P P 1 v
j
ini
2
1
1 j
ini
2
1
j
j
ini 1 j j
2
1
1 j
ini 1 j j
2
1
j j
=
+ + =
+
+
+
+



50
where P
j
is the price of a zero coupon bond maturing at time j, and
j , i
ini
A are the initial Arrow
Debreu prices that pay $1 if state i is reached at time j. These are computed from todays
implied volatility surface.

Here, Q is by construction a positive definite matrix, and c is a vector defined as follows:

( ) ( )
( ) ( ) ( )
j j
T
j
T
min j j
T
j
j j
T
j j j
v T c
Q
+ =
=


where, Psi (matrix) and T (vector) above are constructed with the Arrow Debreu securities
corresponding to the relevant time index j. The matrices Q, Psi, and Phi and vectors c, T
j
are
of dimension NxN and Nx1 respectively, where N is the number of asset steps in the PDE
grid. Q, Psi, Phi, c, and T
j
are defined as follows:

( ) ( ) ( ) ( ) ( ) ( )
(maturity) M ..., 0, j for
A A A 1 P P 1 A 1 P P 1 - T
j
ini
2
1
1 j
ini
2
1
j
j
ini 1 j j
2
1
1 j
ini 1 j j
2
1
j
=
+ + =
+
+
+
+



NxN
j , N
ini
j , 1 N
ini
j , N
ini
j , 1 N
ini
j , 2 N
ini
j , 3
ini
j , 2
ini
j , 1
ini
j , 2
ini
j , 1
ini
j
A - A L 0 0 0 0
A U A - A L 0 0 0
0 0 0 A U A - A L
0 0 0 0 A U A -
(
(
(
(
(
(

L
L
M M M M M M M
L
L


Here is the ratio t/x
2
,
1 j , i
ini
2
1
j , i
ini
2
1
j , i
ini
A A A
+
+ = , and

) tanh 1 ( L
) tanh 1 ( U
x 2
1
x 2
1
=
+ =


Furthermore,
(13)


51
NxN
j
j j
j j
j
j
0 0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0 0
(
(
(
(
(
(

L
L
M M M M M M M
L
L


Also,

(
(
(
(
(
(

|
|
.
|

\
|
+
|
|
.
|

\
|
+

=
+
+

+
2
1
2
1
2
1
2
1
2
1
2
1
x
j
1 j
j
1 j
j
1 j
j
1 j
P
j
P
sinh 2
1


There are a variety of ways to solve the quadratic program embedded in the convertible bond
pricing problem. To achieve this in a fast computational environment, we use C++. One way
is to use a commercial C++ library such as NAG and link it to the pricing algorithm. Another
method is to use Wolfes three-phase algorithm, which relies on pivot operations on a Tableau
of approximately four times the size of the Q matrix, formed via the Q matrix and the c vector
above. This method introduces slack variables into the constraint system
max j
y y 0 to
make the constraints binding (if they are not already so)
24
, then adds artificial variables w to
the linear constraints portion of the quadratic program, and artificial variables z
1
or z
2
to the
set of equations that are formed from the Kuhn-Tucker optimality conditions involving the
Nx1 vector c. The first phase of the algorithm involves minimizing the sum of the ws to zero,
and in the second phase the sum of zs are minimized to zero. Once this condition is attained,
an initial feasible solution has been found for the local volatilities and we proceed to
maximize from the set of feasible solutions by a parametric search method (see Wolfe (1959))
and Kunzi et al. (1968) for further details). The algorithm involves numerous pivot operations
to exchange elements in and out of the basis in the search towards an optimal solution, with
restrictions on the set of possible variables eligible for basis entry becoming effective in
various stages of the algorithm to maintain the orthogonality conditions that must be effective
due to the Kuhn-Tucker conditions. We provide implementation details in the Appendix C.




52
C. Wolfes three-phase algorithm
This algorithm implements the long form of Philip Wolfe's (1959) Quadratic programming
method for

Min p'x + 0.5x'Cx
subject to Ax <= b, x >=0

Note that in the initial problem there are n original (x
1
, ..., x
n
) variables and m
1
inequality
constraints. The matrix C is square symmetric positive definite and of rank n.

'p' is a vector of size n.
A is the constraints matrix of size m
1
xn.
The inequalities are transformed into equalities by the addition of m
1
additional slack
variables, extending the x system to (x
1
,...x
n
, x
n+1
, ..., x
n+m
), 'p' to an n+m
1
vector, the
constraints matrix A into m
1
*(n+ m
1
) matrix which looks like [AId] where Id is the m
1
*m
1

identity matrix.

C is transformed into a symmetric (n+ m
1
)x(n+ m
1
) positive semi-definite matrix.
The transformed system is:
Min p'x + 0.5x'Cx
subject to Ax = b, x >= 0

Wolfe's algorithm proceeds in three phases to solve

Ax + w = b m
1
equations
Cx - v - A'u + z
1
- z
2
= * p n+ m
1
equations

Here u and w are m
1
x1 vectors and v is an (n+m
1
) vector. z
1
and z
2
are (n+m
1
)x1 artificial
variable vectors. The Phase 1 objective is to minimize the sum of the w's, while not allowing
u and v and into the basis. The Phase 2 objective is to minimize the sum of the z's, while not
allowing , and x
i
and v
i
to be in the basis simulatneously for each i in 1, ..., n+m
1
. The Phase
3 objective is to solve parametrically for in the vicinity of 1, and the optimum is then a
linear interpolation of two optima, one for < 1, the other for > 1, where the weights are

24



53
chosen on the basis of the proximity of the optima to 1. Wolfe(1959) has shown that this is
the optimal solution to the quadratic program.
The initial basis are the w
i
's, and those z
1i
's for which p
i
is positive, and z
2i
's for which p
i
is
negative.

In Phases 1 and 2 we delete the columns pertaining to any artificial variables that has been
pivoted out of the basis, to prevent its re-entry into the basis at any future stage

In Phase 3, we clean out any artificial variables (z
1
, z
2
) that are still left in the basis (which
will only occur if these z
1i
/ z
2i
's are still present in the Phase 3 basis with a value of zero.

Details of the algorithm can be found in P. Wolfe's (1959) articles in Econometrica, Vol. 27,
Issue 3, pp.382-398























54
D. Convertibles Market
We provide data on the international and US convertible bonds market. Table V provides
market figures for cross currency convertible issues by maturity, issue size, coupon, spread,
liquidity (approximately measured by dividing issue size by bid-ask spread), and other
features such as put and call. We gratefully acknowledge Luke Olsen (Barclays Capital,
London) for providing this data.

Convertible bonds by amount outstanding in the US:

0
10
20
30
40
50
60
70
80
90
100
How Many
Under 25M 25-50M 50-75M 75-100M 100-150M 150-250M 250-500M 500M+
Convertible Bonds by amount outstanding in the US


Figure 10. Convertible Bonds by Amount Outstanding in the US. The figure displays market data
as of end of 2001. Source: ConvertBond.com, a division of Morgan Stanley &Co. Incorporated.



55
0
10
20
30
40
50
60
70
80
How Many
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013+
Maturity

Figure 11. Convertible Bonds by Maturity in the US. The figure displays maturities of outstanding
convertible bonds in the US market as of end of 2001. Source: ConvertBond.com, a division of Morgan
Stanley &Co. Incorporated.







56
0
10
20
30
40
50
60
How Many
Under
5%
6-7% 8-9% 10-
11%
12-
13%
14-
15%
16-
17%
18-
19%
20-
21%
22-
23%
24-
25%
26-
27%
28-
29%
30-
31%
32-
33%
34-
35%
36-
37%
38-
39%
Over
40%
Initial Premium

Figure 12. Convertible Bonds by Initial Premium in the US. The figure displays initial premium
data of outstanding convertible bonds in the US market as of end of 2001. Source: ConvertBond.com, a
division of Morgan Stanley &Co. Incorporated.




















1
Table V
Large and Liquid Cross Currency Convertibles Issues
*

August 2002
Issuer Currency Code Coupon (%) Maturity
Conversion
Premium (%)
Amount
Outstanding (EUR
Millions) Spread
Liquidity =
Size/Spread Callable Puttable
Anglo American USD 3.375 17/04/07 64.5 1122.2 0.5 2244.4 Y
ASM Lithography USD 5.75 15/10/06 64.34 586.6 1 586.6 Y
Corus EUR 3 11/01/07 38.86 307.0 0.5 614.0 Y
Deutsche Bank / Novartis EUR 0 06/12/10 17.58 1400.0 0.25 5600.0 Y Y
Deutsche Bank / Novartis EUR 0 06/12/11 18.31 1400.0 0.25 5600.0 Y Y
Holcim USD 0 10/06/17 54.61 306.1 0.25 1224.2 Y Y
Meridian / Adecco EUR 1.5 25/11/04 52.06 360.0 0.5 720.1 Y
Nestl USD 0 11/06/08 33.08 714.1 0.5 1428.3 Y
Roche USD 0 25/07/21 38.5 2092.8 0.5 4185.6 Y Y
Roche JPY 0.25 25/03/05 58.86 891.8 0.75 1189.1
Roche USD 0 20/04/10 78.52 2193.4 0.125 17547.4 Y Y
South African Breweries USD 4.25 10/08/06 31.73 612.1 0.5 1224.2 Y Y
ST Microelectronics USD 0 22/09/09 64.7 937.1 0.75 1249.4 Y Y
Standard Chartered EUR 4.5 30/03/10 48.76 575.0 0.25 2300.0 Y
Swiss Life EUR 0.5 05/07/06 45.98 300.0 1 300.0
Swiss Life / HSBC EUR 0.5 05/07/06 45.98 300.0 1 300.0
Swiss Re USD 3.25 21/11/21 36.35 1173.2 0.5 2346.5 Y
Verizon / C&W (144A) USD 4.25 15/09/05 68.65 3244.2 0.25 12976.9 Y


*
The author gratefully acknowledges Luke Olsen, Barclays Capital, for the data provided in this table.