OpenGamma Quantitative Research

Tree Option Pricing Models

Yukinori Iwashita
[email protected]

OpenGamma Quantitative Research n. 23 | First version: April 2 2014; This version: September 19 2014
 Abstract

Discrete approximation of a continuous asset price process based on tree lattice models allows for
intuitive and flexible valuation of financial derivatives. This note reviews binomial and trinomial
recombining tree option pricing models. Our discussion includes various lattice specifications,
estimation strategies for hedge sensitivities, and two-dimensional extensions. We also introduce
examples of recently proposed techniques for computational efficiency and market consistency.
Contents
1 Introduction 1

2 Binomial Tree Models 2

2.1 Lattice Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Dividend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Time-varying Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Two-dimensional Binomial Tree Model . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Hedge Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Trinomial Tree Models 14

3.1 Lattice Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Dividend . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Time-varying Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Two-dimensional Trinomial Tree Model . . . . . . . . . . . . . . . . . . . . . . . 18
3.5 Hedge Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 Advanced Techniques in Tree Models 20

4.1 Flexible Binomial Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Adaptive Mesh Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 Acceleration and Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Implied Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1 Introduction
There are fundamentally three numerical approaches to pricing of financial derivatives: partial
differential equation (PDE), Monte Carlo, and tree model. Despite its numerical inefficiency the
tree method is widely accepted since it is straightforward to implement and so flexible that it
can be intuitively adapted to nonstandard options.
The tree model is a discrete approximation of a continuous process for the underlying instru-
ment of a financial derivative. At a discrete time step the underlying asset price jumps by one
of the predetermined amounts with respective transition probabilities. The first concrete real-
isation of the tree lattice approach to option valuation is a binomial tree model introduced by
Cox, Ross and Rubinstein [CRR79] presenting an alternative route to the Black-Scholes formula
for European vanilla options and a simple strategy to options with early exercise feature. In
the binomial model the underlying asset price moves up or down at each time step. Improved
valuation of options is expected with more possible states for the asset price evolution. Inserting
a middle transition to the binomial model a trinomial tree was developed in [Boy86] that admits
more flexibility in sake of computational simplicity. Over the years a wide variety of tree methods
have been proposed in order to overcome difficulties in the primary models and refine numerical
procedures for nonstandard option pricing.1
Using discrete time and asset price necessarily introduces approximation errors to resulting
option value. In [FG99] the errors are classified into distribution error and nonlinearity error.
Discrepancy between the distribution in the continuous model and approximated distribution
in the lattice model for underlying asset price produces the distribution error while the two
distributions have the same mean and variance. The nonlinearity error originates from nonlinear
relation of the underlying asset price and option value, which is involved in payoff value function
at expiry in general and on barriers for knock-in and knock-out options.
Although it is mathematically proved that tree lattice methods converge to the theoretical
option value of a continuous-time model, the convergence is not smooth but oscillating due to
the nonlinearity error [DD04]. Moreover, numerical simulation of the lattice model does not
converge to the exact value because all of the numbers are truncated decimals in the numerical
computation and each lattice model has a specific lower bound for the approximation error.
These are the strong motivations for studying smooth and higher order convergence techniques
[LR96, Jos10, Tia99, CP07].
In this note we review the binomial and trinomial tree option pricing models together with
examples of lattice specifications, and enhancement of the computation techniques. Various types
of tree lattice models have been studied from both of the theoretical and practical perspectives,
and they are still in the revolutionary stage developing into a vast research area. Thus we
must excuse that our references are not complete and readers should refer to the following
literature and references therein. Gentle introduction to binomial and trinomial models and
their implementation are given in [CS98] including application to exotic option pricing. Standard
lattice specifications for the binomial tree are almost comprehensively explained in [Cha08]. For
analytic and numerical approaches to vanilla and exotic options the complete guide [Hau06]
summarises formulas and pesudo codes.
This note is organised as follows.
We start a discussion with binomial tree models in section 2 building a two-state model and
1 Binomial and trinomial tree models are closely related to the PDE approach: time evolution of option price

in the trees is equivalent to a finite difference method for solving the PDE (see section 3.1 for the trinomial model
and [Jam03] for other cases).

1
extending it to multi-step cases, followed by definition of lattice specifications in 2.1, modelling
of dividend payments in 2.2, approximation for nonconstant market parameters in 2.3, a two-
dimensional generalisation in 2.4 and estimation of Greek values in 2.5. Then numerical results
for the binomial tree models are summarised in 2.6.
Section 3 describes trinomial tree models. Our construction of the trinomial tree is based
on risk-neutral probability measure. Subsequent discussions are on lattice specifications in 3.1,
dividends of the underlying asset in 3.2, time-varying market parameters in 3.3, a two-dimensional
generalisation in 3.4, estimation strategies for Greek values in 3.5, and numerical results in 3.6.
Section 4 is devoted to introducing alternative lattices and new techniques proposed to resolve
issues identified in the earlier sections: flexible binomial tree in 4.1, adaptive mesh methods in
4.2, acceleration and truncation techniques in 4.3, and market-implied trinomial tree in 4.4.
All the figures are attached at the end of the paper.

2 Binomial Tree Models

In a binomial model underlying asset price follows a binomial process: at every moment the asset
price changes to one of two possible prices during one step period.
We shall start with the simplest example by assuming an asset process St with t = 0, △t such
that the asset price is S0 at t = 0 and it goes up to uS0 with probability pu or down to dS0
with probability pd at t = △t. The conservation of probability requires pu + pd = 1. Consider
an European type option Ct (St ) which matures at t = △t and introduce the notation for payoff
at maturity, Cu = C△t (uS0 ), Cd = C△t (dS0 ).
Let us construct a replicating portfolio of the option in terms of the underlying asset St and
riskless bond Bt with the constant interest rate r, i.e., Bt = ert , and then, derive the current
value of the option C0 (S0 ) from the absence of arbitrage opportunities. With δ notional of the
underlying and n notional of the bond we have the conditions for replication at t = △t,
δuS0 + ner△t = Cu , δdS0 + ner△t = Cd , (2.1)
which are solved by
Cu − Cd uCd − dCu
δ= , n= . (2.2)
(u − d)S0 (u − d)er△t
Absence of arbitrage requires that this portfolio and the option value should be also equal at
t = 0. Thus we obtain
 r△t
u − er△t

e −d
C0 = δS0 + n = e −r△t
Cu + Cd . (2.3)
u−d u−d
If we introduce a new variable,
er△t − d
p= , (2.4)
u−d
the current value of the option is rewritten as
C0 = e−r△t (p Cu + (1 − p)Cd ) . (2.5)
The coefficients p, 1 − p can be, respectively, interpreted as the up probability pu and the down
probability pd under a specific probability measure as far as 0 < p < 1 is satisfied. From (2.4)
this inequality is true if we impose
d < er△t < u . (2.6)

2
This condition is indeed the requirement for no arbitrage. Hence our interpretation for p, 1 − p
turns out to be appropriate.
In addition to that p in (2.4) has the properties of probability, it is actually derived from
option valuation based on the risk-neutral probability: pu = p and pd = 1 − p solve the equation,
E(S△t |F0 ) = S0 er△t , i.e.,
uS0 pd + dS0 pu = S0 er△t , (2.7)
where F0 is the filtration at t = 0. Therefore the current value of the claim C0 is given by the
expected value of the discounted payoff computed with the risk-neutral probability.
Hereafter our discussion is based on the risk-neutral probability measure meaning that the
parameters of the binomial model should satisfy (2.4) in every time step at least asymptotically
as △t → 0. The detailed discussion on the replicating portfolio strategy and risk-neutral pricing
in the discrete model is given in [BR96].
So far we have discussed the two layer binomial model which has one time step and two final
states. We shall consider its extension to a multi-step system. Let N to be the number of time
intervals. For equally spaced intervals between current time 0 and expiry T we have the time
period,
T
△t = . (2.8)
N
Thus the lattice has N + 1 layers separated by the time step △t.
In a general binomial tree the number of nodes in the j-th layer becomes 2j leading to
unacceptable amount of computation in total. So we shall consider a recombining tree which
can be constructed with the same set of u, d for all of the steps.2 A state of the underlying asset
is denoted as
Si,j = S0 uj dj−i , 0 ≤ i ≤ N , 0 ≤ j ≤ i, (2.9)
With the risk-neutral probability p the option price at the node (i, j) is given by the discounted
expectation value,
Ci,j = e−r△t (p Ci+1,j+1 + (1 − p) Ci+1,j ) . (2.10)
Thus the option values in the i-th layer can be obtained once those in the (i + 1)-th layer are
known. The boundary values are given in the N -th layer: CN,j are the payoffs at expiry for the
asset price S0 uj dN −j . The formula (2.10) is modified depending on option types. For example,
if the option has the early exercise feature, the option value Ci,j is written as

Ci,j = max e−r△t (p Ci+1,j+1 + (1 − p) Ci+1,j ) , (intrinsic value) , (2.11)



where the intrinsic value is the amount of cash obtained by the immediate exercise at the node.
Another example of modification is a barrier option which we will discuss later.
In summary our option pricing strategy in the binomial model is as follows. We begin with
the terminal values CN,j and compute the option values in the previous layer CN −1,j with the
equation (2.10) (or (2.11)). Using the recursive formula at every node at ever time step we
eventually obtain C0,0 that is the current value of the option.
2 If the value of the up and down factors is step-specific, that is, u 6= u , d 6= d for i 6= j, the binomial tree
i j i j
is not recombining. Under certain situations, e.g., an underlying paying discrete dividends (sections 2.2 and 3.2)
or time-varying volatility (sections 2.3 and 3.3), it is natural to consider the non-recombining property in order
to capture the asset dynamics precisely. However, the non-recombinant tree requires the significantly increased
amount of computation. For N time intervals, the non-recombining tree contains nodes of order O(2N +1 ) in
contrast with O(N 2 ) for the recombining binomial tree. Therefore, our discussion in this note concentrates on
the recombining tree and approximation methods are used for the cases mentioned above.

3
2.1 Lattice Specifications
In the previous section we have constructed the binomial model and explained the option pricing
strategy with the risk-neutral probability distribution. The model parameters in the binomial
tree are u, d and p which should satisfy the conditions (2.4) and (2.6) asymptotically. This
section is devoted to discussing how to determine these parameters such that the binomial model
approximates a continuous model. While the continuous model we consider in this note is
limited to the Black-Scholes model where the underlying asset price is distributed log-normally,
the binomial model can be incorporated with other distributions (see, e.g., [Rub98]).
In the risk-neutral measure the time evolution of risky asset is governed by the stochastic
differential equation,
dSt
= r dt + σ dWt , (2.12)
St
where Wt is the Brownian motion under the risk-neutral measure and r, σ are constants.
Since the property of the Black-Scholes model is captured by the mean and variance, we fix
the binomial model parameters such that the same mean and variance are reproduced. Given a
time interval △t we have the first and second moments in the Black-Scholes model,
2r+σ 2 △t
(2.13)


2
E (S△t ) = er△t S0 , S02 .


E S△t =e

Matching of these two quantities between the Black-Scholes model and the binomial model
requires the conditions,
2r+σ 2 △t
(2.14)


puS0 + (1 − p)dS0 = er△t S0 , pu2 S02 + (1 − p)d2 S02 = e S02 .

As expected the first equation is exactly the same as the risk-neutrality condition (2.4). Thus
the system is underdetermined and one more condition should be imposed to uniquely determine
the binomial model parameters. While there are various choices for the extra condition, we shall
introduce several useful examples in this note. The analysis across the lattice specifications is
detailed in the paper [Cha08].
For notational simplicity we use a new variable,

1
ν = r − σ2 . (2.15)
2
Below we only express p which is equal to the upper probability pu in the risk-neutral measure.
Since the probability conservation is exactly satisfied for all of the lattice specifications, the down
probability is given by 1 − p.

Cox-Ross-Rubinstein
Imposing an additional constraint ud = 1 and solving the moment matching conditions (2.14),
then the unique solution is given by

1  (r+σ2 )△t 
(2.16)
p p
u=M+ M2 − 1 , d=M− M2 − 1 , M= e + e−r△t .
2
Recalling that our objective is to build a lattice model which converges to the Black-Scholes
model in the N → ∞ limit, it is not needed that subleading terms suppressed in the limit do

4
not exactly match between the two models for finite N . Modifying higher terms in △t leads to
a simple form called Cox-Ross-Rubinstein (CRR) model [CRR79],
 p   p  1 ν p
u = exp σ △t , d = exp −σ △t , p = + △t . (2.17)
2 2σ
At this point it becomes clear what asymptotically means. The solution does not satisfy the
conditions (2.14) exactly, but the equations hold only for its leading terms in △t. In fact one
can confirm that the distribution of the terminal underlying price in the CRR binomial process
surely converges to the log-normal distribution (2.12). Furthermore it was demonstrated in
[CRR79] that the European vanilla option price computed by the CRR model converges to the
Black-Scholes option formula.
It should be noted that the probability p becomes negative for specific choices of r and σ. As
discussed in [Hau04] the negativity of the intermediate probabilities does not necessarily imply
the breakdown of the model. However, it can be a cause of instability issues in option valuation
[Tri91].

Jarrow-Rudd
Jarrow and Rudd (JR) [JR83] (a similar model was proposed soon after CRR from a different
point view [RJRB79]) proposed another lattice specification by imposing p = 1/2,
 p   p  1
u = exp ν△t + σ △t , d = exp ν△t − σ △t , p = . (2.18)
2
Again this solves the equations (2.14) only approximately. The negative probability issue is not
found in this lattice, but u > d is not always true, especially for small number of time steps.
The asset price in the JR tree is growing on average with the constant rate. We call this a
growing tree in contrast with the drift-free nature of the CRR model.

Tian
Tian considered the exact solution to the moment matching conditions [Tia93]. In addition to
the first and second moment conditions (2.14), the condition for the third order moments is
introduced. The third order moment in the Black-Scholes model is given by S0 e3(r+σ )△t which
2

is equated to p u S0 + (1 − p)d S0 . Then the exact solution is

3 3

r̂σ̂   r̂σ̂   r̂ − d
, (2.19)
p p
u= σ̂ + 1 + σ̂ 2 + 2σ̂ − 3 , d = σ̂ + 1 − σ̂ 2 + 2σ̂ − 3 , p =
2 2 u−d
where
r̂ = exp (r△t) , σ̂ = exp σ 2 △t . (2.20)

Although CRR, JR and Tian models adapt the different probability measure for finite time step,
all of them lead to the same result in the N → ∞ limit.

Trigeorgis
It is convenient to impose the moment matching condition (2.14) in the log space by introducing
x = log St , where x is governed by the equation dx = νdt + σdWt . Expressing the up and down
factors, u, d as exp(△xu ), exp(△xd ), then we have
p△xu + (1 − p)△xd = ν△t , p△x2u + (1 − p)△x2d = σ 2 △t + ν 2 △t2 . (2.21)

5
One choice for the third condition in the log space is △xu = △xd . Then the solution is
p   p 
u = exp σ 2 △t + ν 2 △t2 , d = exp − σ 2 △t + ν 2 △t2 ,
(2.22)
1 ν△t
p= + p .
2 2 σ 2 △t + ν 2 △t2

This lattice specification was developed by Trigeorgis [Tri91] in order to achieve a better accuracy
and attain more stability than the CRR and JR models.

Jabbour-Kramin-Young
Another lattice specification based on the log space proposed by Jabbour, Kramin and Young
(JKY) [JKY01] assumes that the tree is growing with the risk free rate. Recalling the the drift
in the log space is ν instead of r we set ud = eν△t . Then the system (2.21) is exactly solved by
√ ! √ !
(1 − p)σ △t pσ △t
u = exp ν△t + p , d = exp ν△t − p ,
p(1 − p) p(1 − p)
√ (2.23)
1 σ △t
p= + p .
2 2 4 + σ 2 △t

Other common choices for the third condition in the system (2.21) include setting equal proba-
bilities p = 1/2. However, there is no advantage in this choice in terms of the accuracy and rate
of convergence [CS98].

Leisen-Reimer
When using any discrete model, there are two kinds of convergences to be taken into account: the
convergence of binomial distribution to the geometric Brownian motion and that of option price
computed in a discrete model to the solution of the continuous model. The relation between
these two convergences is not obvious. It is not necessarily true that one lattice specification
with higher order convergence to the risk-neutral distribution produces a better approximation
to the option price.
In [LR96] authors proved that the order of convergence to the analytic price of the European
option is equal to one for CRR, JR also and Tian. Thus, as far as valuation of the option is
concerned, the efficiency of the computation is the same for all the three methods although the
order of the convergence to the geometric Brownian motion is different.
Leisen and Reimer (LR) proposed a lattice specification with which the European option price
exhibits the second order convergence [LR96]. The order of the convergence of the Leisen-Reimer
model was conjectured in the original paper and proved later in [Jos10].3
The second order convergence is realised only for odd N by the following lattice specification,

r̂p̂ r̂ − p u
u= , d= , p = f (d2 ) , (2.24)
p 1−p
3 In [Jos10] higher order convergence is achieved by extending the Leisen-Reimer lattice specification.

6
where p̂ = f (d1 ) and the function f (z) is defined by
  !2   12
1 sgn (z) 

z 1 
f (z) = + 1 − exp − 1 0.1 N+ ,
2 2 N+3+ 6
N +1 (2.25)
log (S0 /K) + νT √
d1 = √ , d 2 = d1 − σ T .
σ T

The definition of r̂ is given in (2.20). Although the higher order convergence is demonstrated only
for the European options, it is well known that this lattice specification tends to show a better
convergence for other cases. One example for which the Leisen-Reimer lattice works very poorly
is powered options whose payoff is (max{ST − K, 0}) for a call option and (max{K − ST , 0})
n n

for a put option with n > 0.

Remarks on Convergence
We have seen that such a wide range of the lattice specifications is allowed in the binomial
model. This flexibility originates from the fact that the constraints (2.14) are imposed only
asymptotically, and one may consider that higher order terms in △t can be manipulated in
arbitrary ways. From the mathematical point of view a natural question is what is the minimum
requirement for the model convergence to the Black-Scholes model in the large N limit? In
[Hsi83] the author presented the answer to this question and proved that any binomial model
with the risk-neutral probability p such that N p → ∞ as N → ∞ converges to the Black-Scholes
model. This condition is indeed weak since any probability with 0 < p < 1 meets the requirement.
It is the rate of convergence to option prices rather than asset prices that we are more
concerned with. In [DD04] the large N expansion of European vanilla option price in the CRR
model is investigated and the explicit expansion formula is derived for the European call option
price,  
F (N ) 1
CCRR (N ) = BS + +o , (2.26)
N N
where BS is the Black-Scholes price and F (N ) is a bounded function that has no limit when N
tends to infinity. This means that the 1/N 1/2 term is absent, and F (N ) is responsible for the
peculiar oscillatory behaviour (see section 2.6).
For American options, as their simple analytic formula is not available, the rate of the con-
vergence is still uncovered. It is shown in [Lam98] that the valuation error for the American
vanilla put option is bounded between 1/2 and 2/3, while numerical studies suggest that the
convergence is order one [BD96, Lei96].

2.2 Dividend
In this section we consider options on an underlying asset which pays dividend yields. Before
looking into the lattice model case we shall recall prescriptions frequently employed in the con-
tinuous model.
If the asset pays a continuous dividend at a fixed rate q, the Black-Scholes model (2.12) is
modified as
dSt
= (r − q) dt + σ dWt . (2.27)
St

7
Thus the drift is reduced by the dividend rate and any derivative pricing strategies based on the
Black-Scholes model are completed by r → r − q.
What if the dividend payments are discrete? There are two cases we should consider: the
dividends are proportional to the asset price at the dividend payment time, and each dividend is
paid by cash. The former case can be handled within the framework of the Black-Scholes model.
Let δ to be the known proportional dividend paid at the dividend date τ , i.e., the underlying
asset price decreases by δSτ . Rescaling the underlying asset,
Ŝt = St (1 − δ) , (2.28)
then we can assume that Ŝ follows the stochastic process (2.12). The option pricing procedure
with the rescaled underlying price is exactly the same as the no dividend case.
If the discrete dividend is paid by cash at t = τ and the dividend amount is known as D, the
underlying asset price jumps on the dividend date, Sτ → Sτ − D. A simple prescription is to
decompose the asset price,
St = S̃t + D e−r(τ −t) , t<τ <T, (2.29)
and find that S̃t follows the geometric Brownian motion (2.12). It should be noted that resulting
option price on the cash dividend asset is scheme-dependent. Several approximation methods
are compared in [Fri02] and a comprehensive review including American type options is, e.g.,
[NP13].
Below applications of the above techniques to the binomial tree are presented following the
argument in [CS98].

Continuous Dividend
To achieve the convergence to the modified Black-Scholes model (2.27) we should shift the interest
rate parameter in the binomial model,
r → r−q. (2.30)
Thus the lattice specifications in section 2.1 can be used with this shift and the the modification
of equation (2.15), ν = r − q − σ 2 /2.

Discrete Proportional Dividend

The analogous approach to (2.28) is to assume that the dividend date is at i△t and express the
value of the asset at node (i, j) as
Ŝi,j = S0 (1 − δ) uj di−j . (2.31)
By using the rescaled asset price the binomial tree continues to be recombining even after the
dividend payment. This argument can be extended for multi-dividend payment cases by rescaling
the underlying asset price at every node when a dividend is paid.

Cash Dividend
If the underlying pays a cash dividend, the asset price is governed by a nonrecombining tree
because the recombining condition, (Si,j u)d = (Si,j d)u, is indeed smeared,
(Si,j u − D)d 6= (Si,j d − D)u , (2.32)

8
where the dividend time is between (i+1)△t and (i+2)△t. However, using the same prescription
as (2.29) allows one to use a recombining tree. Let us define

S̃t = St − De−r(τ −t) for t ≤ τ , (2.33)

and assume that the volatility of S̃t is constant. Then the following asset price process produces
a combining binomial tree,

for t ≤ τ
(
S̃0 uj di−j + De−r(τ −t)
Si,j = (2.34)
j i−j
S̃0 u d for τ < t

where t = i△t. Although it is possible to extend this to multiple dividend payment cases, the
recombining tree becomes an inaccurate approximation to the nonrecombining tree. Recently
this issue is addressed by using interpolation technique [VN06].

2.3 Time-varying Volatility

Our discussion on the binomial tree approximation of the Black-Scholes model has assumed that
the model parameters r, σ are constant. This is a too strong assumption because it is widely
observed that volatilities in the market are not constant in time.4 In this section we consider
time-dependent interest rate and volatility, r = r(t), σ = σ(t), as a more general setup,

dSt
= r(t)dt + σ(t) dWt . (2.35)
St
Its lattice counterpart is actually a nonrecombining model because time-varying parameters in
the continuous model naively reflect on the layer specific parameters in the binomial model,
ui , di , pi and the recombining condition is broken, ui di+1 Si,j 6= di ui+1 Si,j . With the notation
νi = r(i△t) − 21 [σ(i△t)] and σi = σ(i△t) the moment matching conditions in the log space
2

(2.21) take the form,

pi △xi − (1 − pi ) △xi = νi △ti , pi △x2i + (1 − pi ) △x2i = σi2 △ti + νi2 △t2i , (2.36)

which are rewritten as

 
1
q
2 4 2 2
△ti = 2 −σi + σi + 4νi △xi ,
2νi
(2.37)
1 νi △ti
pi = + .
2 2△xi

However, if we use the following △x for all i,

q N N
2 1 X 1 X
△x = σ 2 △t + ν 2 △t , with σ = σi , ν = νi , (2.38)
N i=1 N i=1

then △t is approximately the average time step T /N [CS98].

4 In addition the market implied volatility exhibit strike dependence σ(K, t). Modelling this property is
discussed in section 4.4.

9
 We should note that the lattice specification above is not a good approximation in general.
One can see this by computing the approximation error,

σ2
 
△t + O △t2 , (2.39)


△ti − △t = 2 − 1
σi

which shows that the error is the order of △t and the leading term is small only if σi ≃ σ. Thus
the recombining lattice approximation is not appropriate if σ(t) changes rapidly in time.
Although a naive extension of the lattice specification to the trinomial model improves the
computation, the same problem still exists and the continuous process is poorly approximated.
Instead the time dependent feature of the model parameters can be nicely captured in the
trinomial model by the implied tree (see section 4.4).

2.4 Two-dimensional Binomial Tree Model

For options whose payoff depends on multi-assets, one should consider multidimensional exten-
sion of the Black-Scholes model. Especially for the two-dimensional case, underlying asset prices
are expressed in terms of correlated geometric Brownian motions,

dStI = r − q I StI dt + σ I StI dWtI , (2.40)

I = 1, 2 ,

where WtI are standard Brownian motion processes with a constant correlation ρ, i.e., dWt1 dWt2 =
ρdt. When constructing an extended binomial tree model converging to the two-dimensional
Black-Scholes model, it is again convenient to introduce the log space for the asset prices, xI =
logStI . Then the equations (2.40) are rewritten as

dxI = ν I dt + σ I dWtI , (2.41)

with ν I = r − q I − 21 (σ I )2 . A two-dimensional binomial model is specified by four probabilities,

puu , pud , pdu , pdd and jump sizes for the two assets, △xIu , △xId , where for example pud represents
the probability that the asset 1 goes up by △x1u and the asset 2 goes down by △x2d in one time
step. Here we shall impose that up and down jump sizes are the same, i.e., △xI ≡ △xIu = △xId
for each I. The first and second moments match if the following conditions are satisfied

(puu + pud ) △x1 − (pdu + pdd ) △x1 = ν 1 △t ,

2
2
2
2
(puu + pud ) △x1 + (pdu + pdd ) △x1 = ν 1 △t2 + σ 1 △t ,
(puu + pdu ) △x2 − (pud + pdd ) △x2 = ν 2 △t , (2.42)

2
2
2
2
(puu + pdu ) △x2 + (pud + pdd ) △x2 = ν 2 △t2 + σ 2 △t ,
(puu − pud − pdu − pdd ) △x1 △x2 = ρσ 1 σ 2 △t .

10
Together with the constraint on the probabilities, puu + pud + pdu + pdd = 1, the system is solved
by
p p
△x1 = σ 1 △t , △x2 = σ 2 △t ,
△x1 △x2 + ν 1 △x2 + ν 2 △x1 + ρσ 1 σ 2 △t


puu = ,
4△x1 △x2
△x1 △x2 + ν 1 △x2 − ν 2 △x1 − ρσ 1 σ 2 △t


pud = , (2.43)
4△x1 △x2
1 2 1 2 2 1 1 2


△x △x − ν △x − ν △x + ρσ σ △t
pdu = ,
4△x1 △x2
△x1 △x2 − ν 1 △x2 + ν 2 △x1 − ρσ 1 σ 2 △t


pdd = .
4△x1 △x2
Again we modified higher order terms in △t to achieve the simple form of the solution.
Since each node in the two-dimensional tree is specified by the prices of the two underlying
assets, we label the option values as Ci,j,k where i represents the time layer with 0 ≤ i ≤ N and
the other two labels j and k respectively correspond to the asset 1 and asset 2 with 0 ≤ j, k ≤ i.
Then the recursive formula which relate the option values in the (i + 1)-th layer to those in the
i-th layer is written as

Ci,j,k = e−r△t (puu Ci+1,j+1,k+1 + pud Ci+1,j+1,k + pdu Ci+1,j,k+1 + pdd Ci+1,j,k ) . (2.44)

Therefore, once the option s payoff is computed at expiry, the current value of the option C0,0,0
is eventually obtained by using this recursive formula at every time layer.

2.5 Hedge Sensitivities

In this section we consider Greek values in the one-dimensional and two-dimensional binomial
tree models. The values of delta, gamma and theta can be estimated by using the option values
at node points Ci,j . Since these quantities have been already computed when pricing the option,
we obtain option price and the three Greek values once a single tree is built.
We shall first discuss the one-dimensional case. One simple approximation of the current
value of delta is using the two nodes in the t = △t layer and taking finite difference ratio of the
option values and asset prices,

∂C C1,1 − C1,0
∆= ≈ . (2.45)
∂S S1,1 − S1,0

To estimate the second order Greek one should step forward to the layer containing three nodes.
Taking finite difference ratio twice gives
C2,2 −C2,1 C2,1 −C2,0
∂2C S2,2 −S2,1 − S2,1 −S2,0
Γ= ≈ 1 . (2.46)
∂S 2 2 (S2,2 − S0,0 )

The above computation of the delta and gamma can be regarded as estimates in the future times.
The estimation at the current time is possible extending the tree up to C−2,0 such that two extra
nodes C0,1 , C0,−1 appear at t = 0.

11
 In the binomial model only up node and down node are in the next layer t = △t. Thus the
option theta is estimated with the middle node in the layer t = 2△t,
∂C C2,1 − C0,0
Θ= ≈ . (2.47)
∂T 2△t
However this not a good approximation if the tree is growing in time, i.e., S2,1 6= S0,0 . In this
case an alternative formula is suggested in [Rub94],
1
Θ ≈ rC0,0 − rS0,0 ∆ − σ 2 S0,0
2
Γ. (2.48)
2
For example the theta value in the CRR model is computed by the former formula (2.47) whereas
the JR model relies on the latter (2.48).
Other Greeks can be computed by repricing the option for small change in the corresponding
parameter. For example, using the central finite difference approximation to vega yields

∂C C(σ + ∆σ) − C(σ − ∆σ)

≈ . (2.49)
∂σ 2∆σ
Note that computing one Greek value requires building two trees and the trees can not be reused
for computation of other Greeks or the option price.
Greeks in the two-dimensional model can be estimated in a similar manner, but more ambi-
guity arises in the estimation. Also one needs to consider sensitivities to both of the two asset
prices, so there are more varieties in the spot sensitivities.
Two kinds of option deltas are defined in the two-dimensional model: the sensitivity to the
asset 1 and the sensitivity to asset 2. In the first time step both of the asset 1 price and asset
2 price change. Then it is in the next layer, t = 2△t, that one is allowed to take the difference
ration of the option values and asset prices with the other asset fixed. Another way to estimate
the deltas is taking the average of the two difference ratios in the t = △t layer,
!
1 ∂C 1 C1,1,1 − C1,0,1 C1,1,0 − C1,0,0
∆ = ≈ 1 − S1 + 1 − S1 ,
∂S 1 2 S1,1 1,0 S1,1 1,0
! (2.50)
2 ∂C 1 C1,1,1 − C1,1,0 C1,0,1 − C1,0,0
∆ = ≈ 2 − S2 + 2 − S2 .
∂S 2 2 S1,1 1,0 S1,1 1,0

For the second order Greeks we first take the average of the three difference ratios corresponding
to the three states of the asset 2,
!
1 1 C2,2,2 − C2,1,2 C2,2,1 − C2,1,1 C2,2,0 − C2,1,0
∆2,1 = 1 − S1 + 1 − S1 + 1 − S1 ,
3 S2,2 2,1 S2,2 2,1 S2,2 2,1
! (2.51)
1 1 C2,1,2 − C2,0,2 C2,1,1 − C2,0,1 C2,1,0 − C2,0,0
∆2,0 = 1 − S1 + 1 − S1 + 1 − S1 ,
3 S2,1 2,0 S2,1 2,0 S2,1 2,0

then gamma for the asset 1 is estimated by

∆12,1 − ∆12,0
Γ1 ≈ 1 1 − S1

. (2.52)
2 S2,2 2,0

12
Using the same prescription we can estimate gamma for the asset 2, Γ2 , and another second
order Greek Γ12 , where the option price is differentiated by asset 1 and then by asset 2. Again
this is not the unique way of estimating gamma. For example one can avoid the approximation
based on taking average, c.f., (2.50), by computing the difference ration in the layer of t = 3△t.
Since uI dI = 1 holds for I = 1, 2 in (2.43), the option theta is estimated in an analogous way
to (2.47),
C2,1,1 − C0,0,0
Θ= . (2.53)
2△t
As in the one-dimensional case, other Greek values are computed by the finite difference approx-
imation.
We should note that the convergence of the Greek values is in general slower than that of
the option price. Also the estimation using the nodes in adjacent layers results in a very poor
approximation to the analytic formula in certain situations, for example, the case when the
relevant nodes are separated by a boundary which causes the large nonlinearity error. While
the central finite difference approximation (2.49) works better in these cases, the computation
time is considerably increased since two distinct trees need to be built for evaluating each Greek
value.

2.6 Numerical Results

In this section we present numerical results for the binomial tree models. All of the analytic
formulas to be compared with our results are available in [Hau06].
The convergence behaviour of option prices and delta values is illustrated for all of the lattice
specifications in Figures 1-10 where we show the normalised difference from the true value
defined by
VTree − VTrue
, (2.54)
VTrue
as a function of the number of time steps N for the range (50, 249). VTree is a resulting value in
a binomial model and VTrue is a value in the continuous counterpart. If an analytic expression
is not available, VTrue is estimated by a tree model with sufficiently large number of time steps.
The convergence of an European vanilla call option price with S0 = 31, K = 30, T = 1,
σ = 0.25, r = 0.1 is depicted for the CRR lattice, the JR lattice and the Tian lattice in Figure
1, for the Trigeorgis lattice and the JKY lattice in Figure 2, and for the LR lattice in Figure 3.
The option delta for the same European option is shown in Figures 4, 5 and 6 for the same set
of the lattice specifications. We have similar lessons from the option price and delta. In addition
to the significant odd-even behaviour, it is observed that the error oscillates with a longer period
causing that the greater approximation error can be produced even if we increase the number of
time steps by 2. Since it produces much smaller errors and the number of time steps should be
odd in the model, the LR lattice is separated in Figure 3 for the option price and in Figure 6 for
the option delta. The figures suggest that the option price and delta in the LR lattice converge
to those of the Black-Scholes model fast and smoothly.
It is not necessarily true that higher order convergence of European vanilla option price
suggests that of nonstandard options. Especially, as pricing of path-dependent options in the
binomial models introduces another source of the nonlinearity error, the LR lattice is not better
than others any more. For example a down-and-out option with barrier H is priced by modifying

13
(2.10) as 
 e−r△t (pCi+1,j+1 + (1 − p) Ci+1,j ) if Si,j > H
Ci,j = (2.55)
 0 if Sij ≤ H

American knock-out options can be valued taking into account early exercise possibility in ad-
dition as in (2.11). If the option is European, knock-in type is priced through the in-out parity,
whereas one should rely on alternative strategies for American knock-in options due to the ab-
sence of the parity (see e.g., [CS98]). We illustrate the case of an European-type down-and-out
call option with S0 = 31, K = 30, T = 1, σ = 0.25, r = 0.1 and the barrier H = 25 for the
CRR lattice, the JR lattice and the Tian lattice in Figure 7, for the Trigeorgis lattice, the JKY
lattice and the LR lattice in Figure 8. The severe oscillations are found for CRR and Trigeorgis,
whereas the convergence is still oscillatory but moderately bounded for the JR, Tian and JKY
cases resulting in a better approximation than the LR lattice. One may conclude that using
a binomial model is not an optimal strategy for constant barrier options because shape of the
barrier realised in the tree lattice is not flat but sawtooth, and the the effective barrier is located
in a wrong place.
One of the most important applications of the binomial trees is to price American options.
Since early exercise is never optimal for an American call option without dividend payments,
we consider an American put option on an asset without dividend payments as the simplest
example. The convergence behaviour is shown in Figure 9 for the CRR lattice, the JR lattice
and the Tian lattice, and in Figure 10 for the Trigeorgis lattice, the JKY lattice and the LR
lattice. Our parameter choice is S0 = 29, K = 30, T = 1, σ = 0.25, r = 0.1, and the true value
in (2.54) is computed by the LR model with N = 10001. The figures are quite similar to the case
of the European vanilla option: the odd-even property and long period oscillation are found. It
is worth mentioning that the advantage in the LR lattice is limited for the present parameter
choice in comparison with the outstanding rate of convergence for the European vanilla option.
An exchange-one-asset-for-another option is the right to exchange asset S 2 for asset S 1 at
expiry. Then the payoff of the option is

max Q1 S 1 − Q2 S 2 , (2.56)

Q1 , Q2 are quantities of the assets 1 and 2, respectively. We price the exchange-one-asset-for-

another option with S01 = 29, S02 = 31, T = 1, σ01 = 0.3, σ02 = 0.2, r = 0.1, ρ = 0.6, Q1 = Q2 = 1
by using the two-dimensional lattice model in section 2.4. The convergence to the analytic for-
mula is depicted in Figure 11. Besides the oscillatory behaviour as in the one-dimensional cases,
the convergence is slow and the result is hardly improved even with a larger number of time
steps. It is considered that this is because more approximation and numerical errors tend to
accumulate in the two-dimensional model. Slightly better convergence is obtained if we use the
trinomial model discussed in section 3.4.

3 Trinomial Tree Models

While the original argument on the binomial lattice method of CRR [CRR79] is based on the
principle of option replication, the replication principle does not necessarily hold in other models.
As shown for the binomial method, however, the option s fair value evaluated by the replication
is the same as the price computed in the risk-neutral world, in which the fair value is expressed
as the expected payoff under the risk-neutral probability discounted back to the valuation date

14
by the risk-free interest rate. This is not specific to the binomial model: any approximation
procedure with a probability distribution that converges to the risk-neutral distribution in the
continuous limit can be used to price options. Thus we shall start a discussion on trinomial tree
models with the risk-neutral pricing.
In a trinomial tree an asset price changes to one of the three possible prices during one
time step period meaning that given a set of three probabilities {pu , pm , pd } and three factors
{u, m, d}, the three possible prices for the asset in the next time step are uS0 with probability
pu , mS0 with probability pm or dS0 with probability pd . If we fix the time step △t, the conditions
for probability conservation and risk-neutral probability distribution are, respectively, written as

p u + pm + p d = 1 , pu uS0 + pm mS0 + pd dS0 = S0 er△t , (3.1)

where r is the risk-free interest rate. Then the option price at t = 0 is expressed in terms of its
terminal values at t = △t, Cu = C△t (uS0 ), Cm = C△t (mS0 ) and Cd = C△t (dS0 ),

C0 = e−r△t (pu Cu + pm Cm + pd Cd ) . (3.2)

Towards the extension to a multi-step lattice we should note that the recombination is not
automatic even if we use common {pu , pm , pd } for all the time steps. In fact one should impose

ud = m2 , (3.3)
which ensures that the trinomial lattice is recombinant. Let Si,j = S0 di−j mj be the asset price
at the node (i, j) with 0 ≤ i ≤ N, 0 ≤ j ≤ 2i and Ci,j be its corresponding option value. Starting
with payoffs at expiry CN,j the values of the option in the next layer are obtained by

Ci,j = e−r△t (pu Ci+1,j+1 + pm Ci+1,j + pd Ci+1,j−1 ) . (3.4)

Using this recursive formula we eventually reach C0,0 which is the current value of the option.
The early exercise and other exotic features of options are captured in the same way as the
binomial model.

3.1 Lattice Specifications

We consider a set of trinomial approximations to the geometric Brownian motion (2.12). Impose
the matching of the mean and variance between the two models then we have the constraints,
2r+σ 2 △t
(3.5)


pu uS0 + pm mS0 + pd dS0 = er△t S0 , pu u2 S02 + pm m2 S02 + pd d2 S02 = e S02 .

Together with the conservation of the probability in (3.1) and the recombining condition (3.3),
the system is underdetermined and has two degrees of freedom which we can make use of when
building a trinomial tree. Also the moment matching conditions (3.5) are imposed only asymp-
totically as done for the binomial models, that leads to a wide variety of the lattice specifications.
Here we shall introduce four examples of fixing the trinomial parameters. Some of other lattice
specifications which are useful in specific contexts are described later in section 4.

Kamrad-Ritchken
Soon after the pioneer work on the trinomial model by Boyle [Boy86] Kamrad and Ritchken
(KR) proposed a simpler and more efficient lattice specification [KR91]. The lattice parameters

15
are expressed in terms of a positive parameter λ,
√ √
u = eλσ △t , m = 1, d = e−λσ △t ,
(3.6)
1 ν p 1 1 ν p
pu = 2 + △t , pm = 1 − 2 , pd = 2 − △t ,
2λ 2λσ λ 2λ 2λσ

If we set λ = 1, the possibility to take the middle brunch vanishes, then this trinomial lattice
specification formally reduces to the CRR binomial model . By this reason the KR model
is regarded as a trinomial extension of the CRR model. The p authors also investigated the λ
dependence in the option valuation and concluded that λ = 3/2 is the optimal choice.

Growing Trinomial Lattice

A growing trinomial (GT) tree is obtained by setting m = eν△t . If we define the up and down
factors by √ √
u = eν△t+λσ △t , m = eν△t , d = eν△t−λσ △t , (3.7)
then the probabilities are written as

σ̂ 2 − (D + 1) σ̂ 1/2 + D σ̂ 2 − (U + 1) σ̂ 1/2 + U
pu = , p m = 1 − pu − p d , pd = , (3.8)
(U − D) (U − 1) (U − D) (1 − D)
√ √
where U , D are rescaled up and down factors, U = eλσ △t , D = e−λσ △t . Noticing that the up
and down factors for λ = 1 coincide with those of the JR lattice specification (2.18), this model
corresponds to a trinomial-tree extension of the JR binomial model.

Tian
In the paper by Tian [Tia93] a trinomial version of the higher moment matching tree is also
proposed. For the trinomial model the third moments match if
r+σ 2 △t
(3.9)


pu u3 + pm m2 + pd d3 = e3 .

Setting the equal probabilities, pu = pm = pd = 1/3, then we obtain

p r̂ (3 − σ̂) p
u=a+ a 2 − m2 , m= , d = a − a 2 − m2 ,
2
1 (3.10)
pu = p m = pd = ,
3
where
r̂ (σ̂ + 3)
a= , (3.11)
4
and r̂, σ̂ are defined in the binomial case (2.20).

Log-transformed Trinomial lattice

We introduce another lattice specification based on convergence and stability arguments which
were originally made for finite-difference methods of solving the Black-Scholes partial differential
equation.

16
 As in the case of the binomial model, an improvement is expected with the log-transformed
asset price x = log St . The moment matching conditions in the log space are written as

pu △xu + pm △xm + pd △xd = ν△t , pu △x2u + pm △x2m + pd △x2d = σ 2 △t + ν 2 △t2 . (3.12)

Once the equal jumps △x ≡ △xu = −△xd are assumed, we have △xm = 0 from the recombining
condition (3.4). Then the probabilities are expressed in terms of △x,

σ 2 △t + ν 2 △t2 σ 2 △t + ν 2 △t2
 
1 ν△t
pu = + , p m = 1 − ,
2 △x2 △x △x2
(3.13)
1 σ 2 △t + ν 2 △t2
 
ν△t
pd = − ,
2 △x2 △x

Then the current value of the option price is obtained by solving the equation (3.3) recursively,
" 
σ 2 △t + ν 2 △t2 σ 2 △t + ν 2 △t2
  
−r△t 1 ν△t
Ci,j = e + Ci+1,j+1 + 1 − Ci+1,j
2 △x2 △x △x2
# (3.14)
1 σ 2 △t + ν 2 △t2
 
ν△t
+ − C i+1,j−1 .
2 △x2 △x

This is indeed analogous to an explicit finite difference scheme of the Black-Scholes equation
discretsing the space and time into the intervals △t, △x.5 It should be noted that the space
step △x can not be chosen independently of the time step △t. The convergence and stability
analysis in the explicit finite-difference scheme shows that the spatial and temporal steps should
satisfy the following condition,
△x2
α≡ 2 ≥ 3. (3.15)
σ △t
Furthermore Ritchken demonstrated in [Rit95] that the optimal choice of α is its minimum value.
Hence we shall employ
(3.16)
p
△x = σ 3△t .
We denote this lattice specification as LT.

3.2 Dividend
Previously we have handled dividend payments by the shift of the risk-free interest rate for con-
tinuous dividend and modification of spot asset price for discrete dividends. As these approaches
are independent of the option type and the shape of lattice, one can apply the same techniques
to the trinomial tree model. See section 2.2 for detail.

3.3 Time-varying Volatility

For time-dependent volatility and drift we follow the argument on the binomial model. The
moment matching conditions corresponding to (2.36) now take the form,

piu △xi − pid △xi = νi △ti , piu △x2i + pid △x2i = σi2 △ti + νi2 △t2i . (3.17)
5 For finite-difference approaches to the Black-Scholes equation see [Whi13] and references therein, and the
relation to the binomial and trinomial models is detailed in [Jam03].

17
We should note that piu , pid are not constrained in the present case but free parameters to be fixed
as far as the condition 0 < pi < 1 is satisfied for all of the up, middle and down probabilities. If
we rewrite the equations (3.17) as

1 σi2 △ti + νi2 △t2i 1 σi2 △ti + νi2 △t2i

   
νi △ti νi △ti
piu = + , p i
d = − , (3.18)
2 △x2i △xi 2 △x2i △xi

and use the approximation as before △ti → △t, νi → ν, σi → σ for all i, then the error is
the same as (2.39) independent of how to approximate △xi . Thus we shall fix △xi → △x by
approximating the third oder moment matching condition,

△x3i piu − △x3i pid = νi3 △t3i + 3σi2 νi △t2i . (3.19)

The replacement of △xi , △ti , νi and σi leads to

q
2
△x = 3 σ 2 △t + ν 2 △t , (3.20)

where ν and σ are defined in (2.38).

Numerical result for this lattice should be compared with the time-varying lattice in the
binomial model. As mentioned earlier in section 2.3 significant improvement is not achieved by
using the trinomial model. An alternative strategy to the nonconstant parameters is an implied
tree discussed in section 4.4.

3.4 Two-dimensional Trinomial Tree Model

For the binomial lattice, the model parameters in the two-dimensional extension are determined
uniquely up to the leading order in 1/N . However, the trinomial lattice possesses more degrees
of freedom and free parameters remain after all of the constraints are imposed. Our construction
of a two-dimensional model here is such that the one-dimensional KR model is recovered when
integrating out one of the two assets,
I√ I√
△t
uI = eλσ , mI = 1 , dI = e−λσ △t
,
1 1 1 1 (3.21)
pµν = p1µ + p2ν − + cµν ρ ,
3 3 9 4
where µ, ν take u, m, d and p1µ , p2ν are one-dimensional trinomial lattice specification (3.6). cµν
are 
 1 µ = ν 6= m
cµν = −1 µ 6= ν and µ 6= m and ν 6= m (3.22)
0 µ = m or ν = m


Although it is true that the trinomial model provides the finer mesh and a better approximation
than the binomial model, usage of the two-dimensional trinomial model is limited due to the
increased computational effort.

3.5 Hedge Sensitivities

Estimation of Greek values is similar to the binomial model in section 2.5. Since three distinct
nodes lie in the layer of t = △t, it is possible to evaluate all of the three Greeks without using

18
the information in the next layer t = 2△t,
 
∂C 1 C1,2 − C1,1 C1,1 − C1,0
∆= ≈ + . (3.23)
∂S 2 S1,2 − S1,1 S1,1 − S1,0

C1,2 −C1,1 C1,1 −C1,0

∂2C S1,2 −S1,1 − S1,1 −S1,0
Γ= ≈ 1 . (3.24)
∂S 2 2 (S1,2 − S0,0 )

∂C C1,1 − C0,0
Θ= ≈ . (3.25)
∂T △t
For a growing tree an modification of the formula (2.48) admits an improved estimation for Θ.
In the lattice model various ways of estimation can be considered reflecting the fact that the
model involves more adjacent nodes than the binomial tree. For instance an alternative for delta
is
∂C C1,2 − C1,0
∆= ≈ , (3.26)
∂S S1,2 − S1,0
which in general produces a different value from (3.23). It is not obvious which choice produces
a better estimation, rather this depends on what lattice and what sample space we are using.
To avoid this ambiguity one can rely on the central finite difference approximation.

3.6 Numerical Results

We present numerical results for the trinomial tree models and discuss the convergence behaviour
based on the normalised error defined in (2.54). Our parameter choice is exactly the same as the
binomial tree cases for the respective options, but the figures are shown with a wider range of
the number of time steps, (50, 649), for the one-dimensional cases.
The European vanilla call option with S0 = 31, K = 30, T = 1, σ = 0.25, r = 0.1 is
shown in Figure 12 for the option price and in Figure 13 for the option delta. Different from
the binomial models, the significant odd-even phenomenon are not found but we obtain mild
oscillations specific to the lattice specifications. It should be noted that the GR and Tian lattice
show exactly the same convergence behaviour, and in fact, this is true for nonstandard options.
Although one concludes that the difference in the two lattices, (3.7), (3.8) and (3.10), is a higher
order terms in △t and suppressed in option pricing, the Tian lattice is widely used as a starting
point of applying advanced techniques (see section 4.3).
The European-type down-and-out call option with S0 = 31, K = 30, T = 1, σ = 0.25, r = 0.1
and the barrier H = 25 is depicted in Figure 14. The jagged curves are obtained for the KR
and LT lattices whereas a slow convergence with mild oscillation is found for the other two.
The former two lattice specifications have the property of m = 1 and this explains the peculiar
convergence behaviour. If Si,j is equal to or slightly less than the barrier H, the effective barrier
in the lattice is located at Si+k,j+k at every time step and is close enough to H. Conversely,
if Si,j is slightly larger than H, the effective barrier in the lattice is at Si+k,j+k−1 away from
the true barrier H, then the nonlinearity error associated with the knock-out feature becomes
significantly large. Proposed solution to this problem is to adjust λ such that the effective barrier
is located at the true barrier H, which is discussed in detail in section 4.2.
For the American vanilla put option with S0 = 29, K = 30, T = 1, σ = 0.25, r = 0.1 we
illustrate the normalised error in Figure 15. Besides the absence of the odd-even property, the
fast convergence is obtained except the LT lattice for this parameter set. One lessen from the

19
comparison between the European option in Figure 12 and the American option in Figure 15 is
that the respective lattice specifications exhibit very different convergence behaviour for these
two options. Thus the fast convergence for European options is not a strong indication that
using the same lattice is optimal for American options.
We apply the two-dimensional trinomial model to the exchange-one-asset-for-another option
defined in (2.56) with exactly the same set of the parameters as before, S01 = 29, S02 = 31, T = 1,
σ01 = 0.3, σ02 = 0.2, r = 0.1, ρ = 0.6, Q1 = Q2 = 1. Reflecting the increased number of nodes,
the slight improvement is achieved in the trinomial model in Figure 16.

4 Advanced Techniques in Tree Models

As shown in sections 2.6 and 3.6 the tree models suffer from the oscillatory convergence that pre-
vents one to estimate precision of computation. This is due to the nonlinearity error concerning
relative positions between the nodes and the strike, the barrier and so on. Several techniques
are proposed in order to remove the nonlinearity error and achieve the smooth convergence.
In section 4.1 we discuss flexible binomial tree in which the degree of freedom λ is chosen
such that one node at expiry exactly matches the strike price.
The nonlinearity error arising from the barrier can be handled in the trinomial model by
using adaptive mesh method. In section 4.2 we consider simple approaches proposed in [Rit95]
and [FG99]. Similar to the flexible binomial tree the extra degree of freedom is used such that
the nodes are on the barrier.
Another approach to smooth convergence is introduced in [BD96] as a part of acceleration
technique for American option valuation. Section 4.3 discusses the acceleration together with
other useful techniques for an American put option including Richardson extrapolation and
truncation of the binomial tree.
Nonconstant feature of the model parameters discussed in sections 2.3 and 3.3 is more nicely
captured within the framework of the recombinant tree by building a trinomial tree consistent
with market prices of European options. This is detailed in section 4.4

4.1 Flexible Binomial Tree

The flexible binomial tree was proposed in [Tia99] and the proof of its smooth
√
convergence was
given for European vanilla option and digital option in [CP07]. Set u0 = eσ △t , d0 = u−1
0 and
find j0 at expiry such that

S0 u0j0 −1 dN
0
−j0 +1
< K < S0 uj00 dN
0
−j0
, (4.1)

then the up factor and down factor are defined as

 p   p 
u = exp σ △t + λσ 2 △t , d = exp σ △t − λσ 2 △t , (4.2)

where λ solves
K = S0 uj00 dN −j0 . (4.3)
While rigorous argument of the smooth convergence is given only for limited options, it is indeed
useful for other types including American options since the adjustment of the lattice resolves at
least one source of the nonlinearity errors.

20
 The convergence behaviour of the flexible lattice is illustrated for the European vanilla call
option in Figure 17 and for the American vanilla put option in Figure 18. Here we use the same
parameter sets as before, i.e., S0 = 31, K = 30, T = 1, σ = 0.25, r = 0.1 for the European
option and S0 = 29, K = 30, T = 1, σ = 0.25, r = 0.1 for the American option. These results
should be compared with the binomial models in Figures 1-3 and Figures 9, 10.

4.2 Adaptive Mesh Methods

For vanilla options the flexible binomial tree model and acceleration technique are clues for
achieving smooth convergence (sections 4.1 and 4.3). For barrier options, however, the barrier is
another source of the nonlinearity error and the convergence still exhibits the sawtooth pattern
even after the nonlinearity error at expiry is handled.
In this section we review the adaptive mesh method to the barrier options proposed in [Rit95]
and improved in [FG99]. The former work demonstrated that computation of the single barrier
option price in the trinomial model is greatly improved by adjusting the lattice such that a layer
of nodes coincides with the barrier. Under the assumption that an option has a√ single barrier H
which is constant in time, we start with the specification (3.13) with △x = λσ △t and find λ,
j such that ||3 − λ|| takes the minimum value and
√
S0 ejλσ △t
=H, (4.4)
for −N < j < N . If j is not within this range, the barrier lies outside of the trinomial tree.
With this adjustment of λ the barrier coincides with one node in each layer after t = j△t.
In this approach we encounter the difficulty in the case the current asset price is close to
the barrier. If the spot and barrier are so close that they are apart only by one space step, for
instance, S0 e−△x ≤ H < S0 for the down-and-out case, the probability to hit the barrier is too
high, and consequently, the option price tends to be undervalued. To insert extra nodes between
H and S0 by increasing N sometimes requires the huge number of the time steps. In [FG99]
this issue is addressed introducing fine resolution of the lattice only in the vicinity of the barrier,
that allows one to improve the computation result without changing N . The model parameters
for the fine mesh are obtained from (3.13) by adjusting △x and △t. Then the local mesh is
consistently connected to the original lattice nodes.
The adaptive mesh methods above are not applicable to double barrier options since no extra
degree of freedom remains for synthesising location of node points to the other barrier. A new
approach to the valuation of these securities is suggested in [DL10] by using a combination of
binomial and trinomial trees.
The normalised error defined in (2.54) for the down-and-out European vanilla call option
with S0 = 31, K = 30, T = 1, σ = 0.25, r = 0.1, H = 25 is shown as a function of number
of time steps in Figure 19. Because the spot is sufficiently distinct from the barrier, we only
apply the version of [Rit95] to the trinomial lattice. Compared with the standard approaches in
Figures 7, 8 and 14 it is clear that the convergence behaviour is remarkably improved with the
adaptive mesh as a consequence of the fact that the nonlinearity error of the barrier is resolved.
Remaining approximation errors in the payoff function at expiry cause small jumps in the graph,
but we see that the errors at expiry are much smaller than those associated with the barrier.

4.3 Acceleration and Truncation

Since early exercise is never optimal for an American call option without dividend payment, we
consider an American put option on an asset without dividend payment as the simplest example

21
of the early exercise feature.
According to [Sta05] tree models for the American put option pricing are more efficient than
finite difference methods in the PDE approach, and a recent paper [CJTY09] demonstrates that
the binomial model has advantages to the trinomial model. In [CJ12] various speed-up techniques
are compared for the binomial model. Our objective in this section is to discuss acceleration and
truncation techniques based on the Tian binomial lattice.
The acceleration consists of two procedures, smoothing and Richardson extrapolation. In
[BD96] the acceleration technique was first applied to the binomial model. There are several
ways to achieve the smooth convergence and one simple method which works effectively for the
American option is to replace the option values in the second last layer by the Black-Scholes
prices. By taking N to be large enough one can avoid the option has exercise opportunity
after the second last step, then this replacement is a good approximation. For a smoothly
converging sequence Richardson extrapolation can be applied. The expansion formula (2.26) is
now expressed with a constant F ,
 
F 1
P (N ) = Exact + +o , (4.5)
N N
The independence of N in F allows one to cancel out the leading-order error term by computing
P (N /2) for even N ,  
1
2P (N ) − P (N /2) = Exact + o . (4.6)
N
By a similar manipulation the leading error terms cancel out for odd N .
In the i-th layer the probability for the asset price to take a value Si,j is given by
 
i
i,j
p = pju pi−j
d . (4.7)
j

Recalling that pu , pd are smaller than 1 and the value of the put option at (i, j), Pi,j , is bounded
from above by the strike price, the contribution to the total option price, Pi,j pi,j , tends to
become very small for large i with large j or small j. In fact only nodes with Smin ≤ 0 ≤ Smax
are relevant to the option valuation and various strategies to determine the bounds Smin , Smax
are compared in [AWDN04]. One of the truncation methods which offer stable option valuation
for both of the positive and negative drifts is driven by
√ √
Smax = Ke−r(T −i△t)+ξσ T −i△t
, Smin = Ke−r(T −i△t)−ξσ T −i△t
, (4.8)

and option values at the truncated notes are replaced by intrinsic values at the respective nodes.
Other ways of the truncation are studied in [CJ12]. In (4.8) we introduced the parameter ξ which
controls the truncation positions. Fixing the truncation introduces another lower bounded for
the approximation error, and then, improvement of the computational results becomes limited
even if a large N is used. Thus the choice of ξ should be synthesised with objective numerical
accuracy.
Figure 20 shows the normalised error (2.54) as a function of computation time (nano second)
for the American put option with the same parameter set S0 = 29, K = 30, T = 1, σ = 0.25,
r = 0.1. Here we compare the result using the acceleration and truncation with that of the Tian
trinomial model (3.10), and find that 10-100 times accuracy is realised for a given computational
time. Note that the truncation method is not greatly advantageous for small N because of the
small number of truncated nodes.

22
4.4 Implied Trees
In sections 2.3 and 3.3, aiming at the consistency with the observed structure of market implied
volatilities, nonconstant volatility was discussed for the binomial and trinomial models where we
smooth out the time-varying paremeters to keep the recomninat property of the lattices, and
it turned out that both the models suffer from the errors which do not decay fast enough in the
large N limit. In addition it is a shortcoming that the strike value dependence of the market
volatilities is not captured by those models.
Implied tree theories were proposed such that a tree model is consistent with shape of implied
volatility smile [DK94, DEK95], and advantages in using the trinomial model were proposed in
[Dup94]. Their strategy is to construct a inhomogeneous trinomial tree, i.e., tree associated with
u , pm , pd , such that European vanilla option value at
node-specific transition probabilities, pi,j i,j i,j

each node C(Si,j−1 , i△t) is consistent with the market value. We shall briefly review how to
determine the probabilities pi,j
u , pm , pd from the option market data.
i,j i,j

Let Qi,j to be Arrow-Debreu (AD) security at node (i, j), i.e., price today of an instrument
whose payoff at t = i△t is unity if the node (i, j) is reached. Then the call option price with
strike K and expiry i△t is written as

2i
max (Si,j − K) Qi,j . (4.9)
X
C(K, i△t) =
j=0

Especially if we choose the strike price as a set of possible terminal asset prices Si,j , the contri-
butions blow the j-th level are zero. Then we have

2i
(4.10)
X
C(Si,j−1 , i△t) = (Si,k − Si,j−1 ) Qi,k .
k=j

Conversely, if the option prices C(Si,j−1 , i△t) are known from the market data, the price of the
AD securities with the same expiry, Qi,j , is derived by solving (4.10) iteratively,

2i
X
C(Si,j−1 , i△t) − (Si,k − Si,j−1 ) Qi,k
(4.11)
k=j+1
Qi,j = .
Si,j − Si,j−1

Starting with the highest node (i, 2i) we compute Qi,j at every node up to j = i, and determine
the remaining AD security values from the lowest node (i, 0) by using the put option prices in
order to avoid accumulated numerical errors.
The next step is to relate the AD securities to the transition probabilities. To accomplish
this we shall impose the probability conservation, risk-neutrality condition for the underlying
and risk-neutrality condition for the AD security,

i,j
pi,j i,j
u + p m + pd = 1 ,

pi,j i,j i,j

u Si+1,j+2 + pm Si+1,j + pd Si,j = Si,j e
r△t
, (4.12)
pi,j−1
u Qi,j−1 + pi,j
m Qi,j + pi,j+1
d Qi,j+1 = Qi+1,j+1 e r△t
.

23
The system has the unique solution, and the up, middle and down probabilities are written as

er△t Qi+1,j+2 − pi,j+2 Qi,j+2 − pi,j+1

m Qi,j+1
pi,j
u =
d
,
Qi,j
er△t Si,j − Si+1,j − pi,j
u (Si+1,j+2 − Si+1,j )
pi,j
m = , (4.13)
Si+1,j+1 − Si+1,j

pi,j i,j i,j

d = 1 − p u − pm .

The up transition probabilities for the highest node and the second highest node are respectively
obtained by

Qi+1,2i+2 er△t Qi+1,2i+1 er△t − pi,2i

m Qi,2i
pui,2i = , pui,2i−1 = . (4.14)
Qi,2i Qi,2i−1

Hence, once the upper probabilities are determined, one can use the formulas in (4.13) for their
middle and down probabilities.
Generally the derived transition probabilities are not positive-definite. One origin of the
negative probability is a large value of the local volatility at a tree node. In fact, with a large
call option price C(Si,j−1 , i△t) we have a large Qi,j , and accordingly, a large up probability pi,j
u .
This tends to produce negative middle and down probabilities.6 If the possibilities are negative,
they are modified as suggested in [DKC96],

Si,j er△t − Si+1,j+1 Si,j er△t − Si+1,j

 
1
pi,j
u = + ,
2 Si+1,j+2 − Si+1,j+1 Si+1,j+2 − Si+1,j
(4.15)
1 Si+1,j+2 − Si,j er△t
pi,j
d = ,
2 Si+1,j+2 − Si+1,j

for Si+1,j+1 < Si,j er△t < Si+1,j+2 or

1 Si,j er△t − Si+1,j

pi,j
u = ,
2 Si+1,j+2 − Si+1,j
(4.16)
1 Si+1,j+2 − Si,j er△t Si+1,j+1 − Si,j er△t
 
pi,j
d = + ,
2 Si+1,j+2 − Si+1,j Si+1,j+1 − Si+1,j

for Si+1,j < Si,j er△t < Si+1,j+1 . The middle probability is determined from the probability
conservation.
Another market-implied tree is constructed within the framework of the binomial model
[Rub94], where the tree is synthesised to market values of European call and put options whose
expiry is the same as that of the target security. Thus the implied binomial tree is consistent
with the market only at expiry in contrast with that the market consistency holds in the implied
trinomial tree in every layer of the time steps.
6 In addition the negativity of the probability is encountered when the forward price of the underlying er△t S
i,j
is outside the sample space spanned by the asset prices in the next layer, Si+1,j , Si+1,j+1 and Si+1,j+2 . One
can avoid this situation with an appropriate choice for the sample space [DKC96].

24
References
[AWDN04] Ari D. Andricopoulos, Martin Widdicks, Peter W. Duck, and David P. Newton. Cur-
tailing the range for lattice and grid methods. Journal of Derivatives, 11(4):55–61,
2004. 22

[BD96] Mark Broadie and Jerome Detemple. American option valuation: new bounds, ap-
proximations, and a comparison of existing methods. Review of Financial Studies,
9(4):1211–1250, 1996. 7, 20, 22

[Boy86] Phelim P. Boyle. Option valuation using a three jump process. International Options
Journal, 3:7–12, 1986. 1, 15

[BR96] Martin Baxter and Andrew Rennie. Financial Calculus: An Introduction to Deriva-
tive Pricing. Cambridge University Press, 1996. 3

[Cha08] Don M. Chance. A synthesis of binomial option pricing models for lognormally
distributed assets. Journal of Applied Finance, 18(1):38, 2008. 1, 4

[CJ12] Ting Chen and Mark Joshi. Truncation and acceleration of the tian tree for the
pricing of american put options. Quantitative Finance, 12(11):1695–1708, 2012. 22

[CJTY09] Jiun Hong Chan, Mark Joshi, Robert Tang, and Chao Yang. Trinomial or bino-
mial: Accelerating american put option price on trees. Journal of Futures Markets,
29(9):8269, 2009. 22

[CP07] Lo-Bin Chang and Ken Palmer. Smooth convergence in the binomial model. Finance
and Stochastics, 11(1):91–105, 2007. 1, 20

[CRR79] John Cox, Stephen Ross, and Mark Rubinstein. Option pricing, a simplifed approach.
Journal of Financial Economics, 7:229–263, 1979. 1, 5, 14

[CS98] Les Clewlow and Chris Strickland. Implementing Derivative Models. John Wiley &
Sons, 1998. 1, 6, 8, 9, 14

[DD04] Francine Diener and Marc Diener. Asymptotics of the price oscillations of a european
call option in a tree model. Mathematical Finance, 14(2):2713, 2004. 1, 7

[DEK95] Emanuel Derman, Deniz Ergener, and Iraj Kani. Static options replication. Journal
of Derivatives, 2:78–95, 1995. 23

[DK94] Emanuel Derman and Iraj Kani. The volatility smile and its implied tree. Quanti-
tative Strategies Research Notes, Goldman Sachs, 1994. 23

[DKC96] Emanuel Derman, Iraj Kani, and Neil Chriss. Implied trinomial trees of the volatility
smile. Journal of Derivatives, 3(4):7–22, 1996. 24

[DL10] Tian-Shyr Dai and Yuh-Dauh Lyuu. The bino-trinomial tree: A simple model for
efficient and accurate option pricing. Journal of Derivatives, 17(4):7–24, 2010. 21

[Dup94] Bruno Dupire. Pricing with a smile. Risk, 7:18–20, 1994. 23

25
 [FG99] Stephen Figlewski and Bin Gao. The adaptive mesh model: A new approach to
efficient option pricing. Journal of Financial Economics, 53(3):313–351, 1999. 1, 20,
21

[Fri02] Volf Frishling. A discrete question. Risk magazine, January 2002. 8

[Hau04] Espen Gaarder Haug. Why so negative to negative probabilities? Wilmott Magazine,
pages 34–38, 2004. 5

[Hau06] Espen Gaarder Haug. The Complete Guide to Option Pricing Formulas. McGraw-
Hill, 2nd edition, 2006. 1, 13

[Hsi83] Chi-Cheng Hsia. On binomial option pricing. Journal of Financial Research, 6:41–46,
1983. 7

[Jam03] Peter James. Option Theory. Wiley, 2003. 1, 17

[JKY01] George M. Jabbour, Marat V. Kramin, and Stephen D. Young. Two-state option
pricing: Binomial models revisited. Journal of Futures Markets, 21(11):98701, 2001.
6

[Jos10] Mark S. Joshi. Achieving higher order convergence for the prices of european options
in binomial trees. Mathematical Finance, 20(1):89–103, January 2010. 1, 6

[JR83] Robert A. Jarrow and Andrew Rudd. Option Pricing. Irwin Professional, 1983. 5

[KR91] Bardia Kamrad and Peter Ritchken. Multinomial approximating models for options
with k state variables. Management Science, 37(12):1640–1652, 1991. 15

[Lam98] Damien Lamberton. Error estimates for the binomial approximation of american put
options. Annals of Applied Probability, 8(1):206–233, 1998. 7

[Lei96] Dietmar P. J. Leisen. Pricing the american put option: A detailed convergence
analysis for binomial models. Discussion Paper Serie, University of Bonn, B-366,
1996. 7

[LR96] Dietmar P. J. Leisen and Matthias Reimer. Binomial models for option valuation -
examining and improving convergence. Applied Mathematical Finance, 3(4):319–346,
1996. 1, 6

[NP13] Martina Nardon and Paolo Pianca. Extracting information on implied volatilities
and discrete dividends from american option prices. Journal of Modern Accounting
and Auditing, 9(1):112–129, 2013. 8

[Rit95] Peter Ritchken. On pricing barrier options. Journal of Derivatives, 3(2):19–28, 1995.
17, 20, 21

[RJRB79] Jr. Richard J. Rendleman and Brit J. Bartter. Two-state option pricing. Journal of
Finance, 34(5):1093–1110, 1979. 5

[Rub94] Mark Rubinstein. Implied binomial trees. Journal of Finance, 49(3):7718, 1994. 12,
24

26
[Rub98] Mark Rubinstein. Edgeworth binomial trees. Journal of Derivatives, 5(3):20–27,
1998. 4
[Sta05] Mike Staunton. Efficient estimates for valuing american options. Best of Wilmott,
2:108–114, 2005. 22

[Tia93] Yisong Tian. A modified lattice approach to option pricing. Journal of Futures
Markets, 13(5):563–577, 1993. 5, 16
[Tia99] Yisong Tian. flexible binomial option pricing model. Journal of Futures Markets,
19:817–813, 1999. 1, 20
[Tri91] Lenos Trigeorgis. A log-transformed binomial numerical analysis method for valuing
complex multi-option investments. Journal of Financial and Quantitative Analysis,
26(3):309–326, 1991. 5, 6
[VN06] M. H. Vellekoop and J. W. Nieuwenhuis. Efficient pricing of derivatives on assets
with discrete dividends. Applied Mathematical Finance, 13(3):265–284, 2006. 9

[Whi13] Richard White. Numerical solutions to pdes with financial ap-

plications. Technical report, OpenGamma, 2013. http://develop-
ers.opengamma.com/quantitative-research/Numerical-Solutions-to-PDEs-with-
Financial-Applications-OpenGamma.pdf. 17

27
 3.00E-03
CRR
JR
Tian
2.00E-03

1.00E-03

0.00E+00
Figure 1: Convergence to the analytic
price formula of the European vanilla -1.00E-03

call option with S = 31, K = 30, T = 1,

σ = 0.25, r = 0.1 for the CRR binomial -2.00E-03

lattice (black line), the JR binomial lat-

tice (blue line) and the Tian binomial
-3.00E-03
40 60 80 100 120 140 160 180 200

lattice (red line).

3.00E-03
TRIG
JKY
2.00E-03

1.00E-03

0.00E+00

Figure 2: Convergence to the analytic -1.00E-03

price formula of the European vanilla

call option with S = 31, K = 30, T = 1, -2.00E-03

σ = 0.25, r = 0.1 for the Trigeorgis bi-

nomial lattice (black line) and the JKY
-3.00E-03
40 60 80 100 120 140 160 180 200

binomial lattice (blue line).

0.00E+00

-1.00E-06

-2.00E-06

-3.00E-06

-4.00E-06

Figure 3: Convergence to the analytic -5.00E-06

price formula of the European vanilla -6.00E-06

call option with S = 31, K = 30, T = 1,

σ = 0.25, r = 0.1 for the LR binomial
-7.00E-06
40 60 80 100 120 140 160 180 200

lattice.

28
3.00E-03
CRR
JR
Tian
2.00E-03

1.00E-03

0.00E+00
Figure 4: Convergence to the Black-
-1.00E-03 Scholes delta of the European vanilla
call option with S = 31, K = 30, T = 1,
-2.00E-03 σ = 0.25, r = 0.1 for the CRR binomial
lattice (black line), the JR binomial lat-
tice (blue line) and the Tian binomial
-3.00E-03
40 60 80 100 120 140 160 180 200

lattice (red line).

3.00E-03
TRIG
JKY
2.00E-03

1.00E-03

0.00E+00

-1.00E-03 Figure 5: Convergence to the Black-

Scholes delta of the European vanilla
-2.00E-03 call option with S = 31, K = 30, T = 1,
σ = 0.25, r = 0.1 for the Trigeorgis bi-
nomial lattice (black line) and the JKY
-3.00E-03
40 60 80 100 120 140 160 180 200

binomial lattice (blue line).

0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

Figure 6: Convergence to the Black-

Scholes delta of the European vanilla
-2.00E-03

call option with S = 31, K = 30, T = 1,

σ = 0.25, r = 0.1 for the LR binomial
-2.50E-03
40 60 80 100 120 140 160 180 200

lattice.

29
 2.50E-02
CRR
JR
Tian
2.00E-02

1.50E-02

Figure 7: Convergence to the ana- 1.00E-02

lytic price formula of the European-type
down-and-out call option with S0 = 31, 5.00E-03

K = 30, T = 1, σ = 0.25, r = 0.1

and H = 25 for the CRR binomial lat- 0.00E+00

tice (black line), the JR binomial lattice

(blue line) and the Tian binomial lattice
-5.00E-03
40 60 80 100 120 140 160 180 200

(red line).

2.50E-02
TRIG
JKY
LR
2.00E-02

1.50E-02

Figure 8: Convergence to the ana- 1.00E-02

lytic price formula of the European-type
down-and-out call option with S0 = 31, 5.00E-03

K = 30, T = 1, σ = 0.25, r = 0.1 and

H = 25 for the Trigeorgis binomial lat- 0.00E+00

tice (black line), the JKY binomial lat-

tice (blue line) and the LR binomial lat-
-5.00E-03
40 60 80 100 120 140 160 180 200

tice (red line).

3.00E-03
CRR
JR
Tian
2.00E-03

1.00E-03

0.00E+00

-1.00E-03

Figure 9: Convergence of the American -2.00E-03

vanilla put option price with S0 = 29, -3.00E-03

K = 30, T = 1, σ = 0.25, r = 0.1 for -4.00E-03

the CRR binomial lattice (black line),
the JR binomial lattice (blue line) and
-5.00E-03
40 60 80 100 120 140 160 180 200

the Tian binomial lattice (red line).

30
3.00E-03
TRIG
JKY
2.00E-03 LR

1.00E-03

0.00E+00

-1.00E-03
Figure 10: Convergence of the Ameri-
-2.00E-03
can vanilla put option price with S0 =
-3.00E-03 29, K = 30, T = 1, σ = 0.25, r = 0.1
-4.00E-03
for the Trigeorgis binomial lattice (black
line), the JKY binomial lattice (blue
line) and the LR binomial lattice (red
-5.00E-03
40 60 80 100 120 140 160 180 200

line).

0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

-2.00E-03

-2.50E-03

-3.00E-03 Figure 11: Convergence to the analytic

-3.50E-03 price formula of the exchange-one-asset-
-4.00E-03
for-another option with S01 = 29, S02 =
31, T = 1, σ01 = 0.3, σ02 = 0.2, r = 0.1,
ρ = 0.6, Q1 = Q2 = 1 for the two-
-4.50E-03
40 60 80 100 120 140 160 180 200

dimensional binomial model.

1.50E-03
KR
GR
Tian
LT
1.00E-03

5.00E-04
Figure 12: Convergence to the analytic
price formula of the European vanilla
call option with S = 31, K = 30,
0.00E+00

T = 1, σ = 0.25, r = 0.1 for the KR

lattice (black line), the growing trino-
-5.00E-04

mial lattice (blue line), Tian trinomial

lattice (black dashed line) and the log-
-1.00E-03
40 140 240 340 440 540 640

transformed trinomial lattice (red line).

31
 0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

-2.00E-03
Figure 13: Convergence to the Black- -2.50E-03
Scholes delta of the European vanilla
call option with S = 31, K = 30,
-3.00E-03

T = 1, σ = 0.25, r = 0.1 for the KR

-3.50E-03

lattice (black line), the growing trino-

-4.00E-03
KR

mial lattice (blue line), Tian trinomial

-4.50E-03 GR
Tian
LT

lattice (black dashed line) and the log-

-5.00E-03
40 140 240 340 440 540 640

transformed trinomial lattice (red line).

3.50E-02
KR
GT
3.00E-02
Tian
LT
2.50E-02

Figure 14: Convergence to the ana- 2.00E-02

lytic price formula of the European-type 1.50E-02

down-and-out call option with S0 =
1.00E-02
31, K = 30, T = 1, σ = 0.25,
r = 0.1 and H = 25 for the KR lat- 5.00E-03

tice (black line), the growing trinomial 0.00E+00

lattice (blue line), Tian trinomial lat-
tice (black dashed line) and the log-
-5.00E-03
40 140 240 340 440 540 640

transformed trinomial lattice (red line).

1.50E-03

1.00E-03

5.00E-04

0.00E+00

-5.00E-04

Figure 15: Convergence of the Ameri- -1.00E-03

can vanilla put option price with S0 = -1.50E-03

29, K = 30, T = 1, σ = 0.25, r = 0.1 -2.00E-03

for KR lattice (black line), the growing -2.50E-03

trinomial lattice (blue line), Tian trino- -3.00E-03

KR

mial lattice (black dashed line) and the -3.50E-03 GT

Tian
LT
log-transformed trinomial lattice (red
-4.00E-03
40 140 240 340 440 540 640

line).

32
0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

-2.00E-03

-2.50E-03

-3.00E-03 Figure 16: Convergence to the analytic

-3.50E-03 price formula of the exchange-one-asset-
-4.00E-03
for-another option with S01 = 29, S02 =
31, T = 1, σ01 = 0.3, σ02 = 0.2, r = 0.1,
ρ = 0.6, Q1 = Q2 = 1 for the two-
-4.50E-03
40 60 80 100 120 140 160 180 200

dimensional trinomial model.

0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

-2.00E-03

Figure 17: Convergence to the analytic

-2.50E-03
price formula of the European vanilla
call option with S = 31, K = 30, T = 1,
σ = 0.25, r = 0.1 for the flexible bino-
-3.00E-03
40 60 80 100 120 140 160 180 200

mial lattice.

0.00E+00

-5.00E-04

-1.00E-03

-1.50E-03

Figure 18: Convergence of the Ameri-

-2.00E-03

can vanilla put option price with S0 =

-2.50E-03
40 60 80 100 120 140 160 180 200 29, K = 30, T = 1, σ = 0.25, r = 0.1
for the flexible binomial lattice.

33
 1.20E-03

1.00E-03

8.00E-04

6.00E-04

4.00E-04

2.00E-04

0.00E+00

Figure 19: Convergence to the ana- -2.00E-04

lytic price formula of the European-type
down-and-out call option with S0 = 31,
-4.00E-04

K = 30, T = 1, σ = 0.25, r = 0.1 and

-6.00E-04
40 140 240 340 440 540 640

H = 25 for the adaptive mesh.

1.00E-02

Tian
Acceleration and Truncation

1.00E-03

1.00E-04
Figure 20: The normalised error as a
function of computation time (nano sec-
ond) for the American put option with
1.00E-05

S0 = 29, K = 30, T = 1, σ = 0.25,

r = 0.1. The Tian trinomial lattice is 1.00E-06

used for the red line whereas the trun-

cation and acceleration techniques are 1.00E-07
1.00E+00 1.00E+01 1.00E+02 1.00E+03 1.00E+04
applied to the lattice for the blue line.

34
 OpenGamma Quantitative Research
1. Marc Henrard. Adjoint Algorithmic Differentiation: Calibration and implicit function theorem.
November 2011.

2. Richard White. Local Volatility. January 2012.

3. Marc Henrard. My future is not convex. May 2012.
4. Richard White. Equity Variance Swap with Dividends. May 2012.
5. Marc Henrard. Deliverable Interest Rate Swap Futures: Pricing in Gaussian HJM Model. Septem-
ber 2012.
6. Marc Henrard. Multi-Curves: Variations on a Theme. October 2012.
7. Richard White. Option pricing with Fourier Methods. April 2012.
8. Richard White. Equity Variance Swap Greeks. August 2012.

9. Richard White. Mixed Log-Normal Volatility Model. August 2012.

10. Richard White. Numerical Solutions to PDEs with Financial Applications. February 2013.
11. Marc Henrard. Multi-curves Framework with Stochastic Spread: A Coherent Approach to STIR
Futures and Their Options. March 2013.

12. Marc Henrard. Algorithmic Differentiation in Finance: Root Finding and Least Square Calibration.
January 2013.
13. Marc Henrard. Multi-curve Framework with Collateral. May 2013.
14. Yukinori Iwashita. Mixed Bivariate Log-Normal Model for Forex Cross. January 2013.

15. Yukinori Iwashita. Piecewise Polynomial Interpolations. May 2013.

16. Richard White. The Pricing and Risk Management of Credit Default Swaps, with a Focus on the
ISDA Model. September 2013.
17. Yukinori Iwashita. Conventions for Single-Name Credit Default Swaps. December 2013.

18. Arroub Zine-eddine. Brazilian Swaps. December 2013.

19. Arroub Zine-eddine. Inflation: Instruments and curve construction. January 2014.
20. Arroub Zine-eddine. Inflation caps and floors. January 2014.

21. Arroub Zine-eddine. Convexity adjustment for inflation derivatives. January 2014.
22. Richard White. Forward CDS, Indices and Options. March 2014.
23. Yukinori Iwashita. Tree Option Pricing Models. April 2014.
24. Richard White and Yukinori Iwashita. Eight ways to strip your caplets: An introduction to caplet
stripping. September 2014.
25. Yukinori Iwashita. Smile Interpolation and Extrapolation. September 2014.
 About OpenGamma
OpenGamma helps financial services firms unify their calculation of analytics across the traditional
trading and risk management boundaries.
The company's flagship product, the OpenGamma Platform, is a transparent system for front-office
and risk calculations for financial services firms. It combines data management, a declarative calcu-
lation engine, and analytics in one comprehensive solution. OpenGamma also develops a modern,
independently-written quantitative finance library that can be used either as part of the Platform, or
separately in its own right.
Released under the open source Apache License 2.0, the OpenGamma Platform covers a range of asset
classes and provides a comprehensive set of analytic measures and numerical techniques.

ABOUT US DOWNLOAD

www.opengamma.com developers.opengamma.com/downloads

Europe North America

OpenGamma OpenGamma
185 Park Street 125 Park Avenue
London SE1 9BL 25th Floor, Suite 2525
United Kingdom New York, NY 10017
United States of America