Quadrinomial Trees to

Value Options in Stochastic

Volatility Models
Julián A. Pareja-Vasseur and Freddy H. Marín-Sánchez

Julián A. ABSTRACT: This article describes in detail the of other types of options in different con-
Pareja-Vasseur multiplicative quadrinomial tree numerical method texts, through the use of different estimation
is a PhD at Centrum
with nonconstant volatility, based on a system of ­techniques, such as closed analytical solutions,
Catolica in Peru and an
associate professor in stochastic differential equations of the GARCH- the finite difference method, the Adomian
the School of Economy diffusion type. The methodology allowed for the decomposition method, and the numerical
and Finance at EAFIT derivation of the first two moments of the proposed method through binomial trees. However,
University in Medellín, equations to estimate the respective recombination the problem of considering the volatility con-
Colombia. between discrete and continuous processes and, as a stant persists. Recently, some models have
[email protected] and jparejav@
eafit.edu.co
result, a numerical methodological proposal is for- appeared that include stochastic volatility,
mally presented to value, with relative ease, both which were created to avoid considering this
Freddy H. real and financial options, when the volatility is fixed variable in terms of the evaluation time
M arín-Sánchez stochastic. The main findings showed that in the horizon. From an empirical point of view,
is a titular professor in proposed method, when volatility approaches zero, this last assumption is contrasted with stylized
the School of Science
the multiplicative binomial traditional method is facts offered by the market, such as leptokur-
at EAFIT University in
Medellín, Colombia. a particular case, and the results are comparable tosis, heavy tail distribution, and conditional
[email protected] between these methodologies, as well as to the variance that changes randomly over time.
exact solution offered by the Black–Scholes model. Motivated by this empirical evidence, sev-
Finally, the originality of the methodological pro- eral authors, such as Hull and White (1987),
posal is that it allows for the emulation in a simple Scott (1987), Wiggins (1987), Chesney and
way of the presence of a nonconstant volatility in Scott (1989), Stein and Stein (1991), and
the price of the underlying asset, and it can be used Heston (1993), proposed models that involve
to value all kinds of options both in the real world stochastic volatility as a parsimonious exten-
and in risk-neutral situations. sion of the well-known Black–Scholes model
(1973).
TOPICS: Options, derivatives*
Understanding the nature of the vola-
tility phenomenon and its appropriate esti-

S
mation is considered important for economic
ince Black and Scholes’s (1973) sem-
agents and for its use in financial and eco-
inal work, which is considered to be
nomic applications, because models with a
the most commonly used method for
well-estimated volatility play an important
*All articles are now estimating the value of European call
role in decisions that involve financial risk
categorized by topics and put options, multiple models and exten-
and subtopics. View at and that are supported on the observa-
sions have appeared that allow the valuation
PM-Research.com. tion of the price of the underlying asset in

Fall 2019 The Journal of Derivatives   49

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 d­ iscrete time, whose logarithmic returns are assumed of conceiving a normal distribution does not apply
to be stationary and in turn behave like a random walk (Fernández Castaño 2007). Motivated by this empirical
(Posedel 2005). evidence, several authors have proposed models that
One of the great difficulties in confronting the involve stochastic volatility as a parsimonious extension
literature is that the developments in the area of quanti- of the Black–Scholes model, in which the linearity of
fication of volatility, specifically stochastic volatility, and the drift and the diffusion components of the asset price
its respective application in numerical methods, are not are maintained, but incorporate a second equation that
significant, particularly with regard to modifications or allows for the modeling of the behavior of the vari-
extensions of the classic models of asset valuation. Little ance of St (Barone-Adesi et al. 2005; Chesney and Scott
financial engineering has been applied to derivatives and 1989; Chourdakis and Dotsis 2011; Christoffersen et al.
to numerical methods in a timely manner, for the valu- 2010; Drost and Werker 1996; Duan 1996, 1997; Figà-
ation of both financial and real options, which would Talamanca 2009; Heston 1993; Hull and White 1987;
foster genuinely important research developments. Ritchken and Trevor 1999; Scott 1987; Stein and Stein
This article is organized as follows: the next sec- 1991; Wiggins 1987; Wu et al. 2012, 2014).
tion describes the stochastic volatility models and, A model with stochastic volatility describes its
specifically, the GARCH-diffusion type used for the change over time and generalizes the Black–Scholes
development of the document. We then show the two model, defined in a given filtered probability space
first moments derived from the proposed system of sto- (W, F, Ft , P), which generally satisfies a system of sto-
chastic differential equations and display the details of chastic differential equations (SDEs). For example,
the binomial and quadrinomial recombination for the Wilmott (1998) considered a system described as follows:
discrete processes. The section after that offers a brief
description of the options valuation, specifically its use dSt = µStdt + σ tStdWt(1)
according to the method proposed, and is followed by dσ t = f (St , σ t , t )dt + g(St , σ t , t )dWt( 2) (1)
the presentation of a series of numerical experiments and
some examples. The final section includes discussion where m is constant and volatility st is considered as a
and conclusions. dynamic variable in the price St. Functions f and g cor-
respond, respectively, to the tendency and diffusion of
DIFFERENTIAL EQUATIONS WITH volatility. The model also incorporates two sources of
STOCHASTIC VOLATILITY MODELS randomness—Wt(1) and Wt ( 2)—that correspond to Wiener
standard processes with a correlation coefficient r.
One of the strongest conditions considered in the In this way, the prices process {St , 0 ≤ t ≤ T} is not com-
Black–Scholes model (1973) is that the volatility (s) is pletely described by Equation 1, and the value St will
constant; several empirical studies have demonstrated be conditioned to the information of S 0, s0 and to the
that the logarithm of returns on asset prices is uncor- trajectory followed by volatility {ss, 0 ≤ s ≤ t}.
related and that they have leptokurtosis and conditional Three models are considered pioneering in terms
variance that changes in a random way as a function of of including and describing stochastic volatility, the first
time (Grajales Correa and Pérez Ramírez 2008). In addi- of which corresponds to that of Hull and White (1987),
tion, the implied volatilities, that is, those that equal the which examined how stochastic volatility is independent
estimated theoretical price in the Black–Scholes model of the stock price; in other words, it was presented as a
with the market price, are considered nonconstant and solution to the series in which the price is instantaneously
differ between exercise prices and time to maturity, and uncorrelated with volatility. In this model, the price St
their values form what is known as the volatility smile.
of a financial asset has an instantaneous variance Vt = σ t2 ,
To overcome the drawbacks presented in this model,
several extensions have been proposed in the literature whose behavior is expressed in the following SDE:
in which volatility is considered a function of time and
of the price of the underlying asset. The returns on dSt = µStdt + σ tStdWt(1)
the asset exhibit two important empirical conditions: dVt = δVt dt + ξVtdWt( 2)
leptokurtosis and heavy tails; therefore, the ­assumption

50   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 The novelty here is that the parameters d and x are The third model corresponds to that of Wiggins
declared and are considered constant and correspond, (1987), who proposed the following system:
respectively, to the tendency and volatility for Vt , while
Wt (1) and Wt ( 2) are Wiener processes with correlation dSt = µStdt + σ tStdWt(1)
coefficient r. The parameter m depends on St , st and t, dσ t = f (σ t )dt + θσ tdWt( 2)
while d and x can depend on st and t, but it is assumed
that they do not depend on St. In this case, the change was based on the fact that
In addition, an analytical solution was also pre- the author decided to use f(st) as a function that describes
sented to assess a European call or put option, in which the volatility tendency st , while Wt (1) and Wt ( 2) corre-
the following conditions are met: first, it is assumed spond to Wiener processes with correlation coefficient r.
that the derivatives are valued in a risk-neutral world, The findings showed that there are not many differences
and second, it is conjectured that there is no correlation between the valuation of Wiggins’s proposal and the
between volatility and the price of the asset and, conse- one obtained through the Black-Scholes model, but it is
quently, r = 0. This allows one to define the system of emphasized that the latter only turns out to be practical
equations in a risk-neutral situation, as follows: in certain cases, but there are others in which its applica-
tion should be deepened, as for example in warrants and
dSt = rStdt + σ tStdWt(1) junk bonds, and the author researches the stability of the
dσ t2 = δσ t2dt + ξσ t2dWt( 2) firm-level volatility to find a more f lexible functional
form for the drift term for volatility (Wiggins 1987).
where the variable r is assumed to be constant and Later Chesney and Scott (1989) developed a model
corresponds to the risk-free rate, and Wt (1) and Wt ( 2) based on Black–Scholes, which they indicated is rejected
are independent Wiener processes with null corre- if it follows a lognormal behavior, because of empirical
lation coeff icient. The authors concluded that the evidence introduced by Bollerslev (1986) and Wasser-
Black–Scholes model generally overestimates the option fallen and Zimmerman (1986). The explanation for this
value and that the degree of this overvaluation increases rejection is that the variance of ln(St /St −1 ) changes in
with time at expiration. a random way, so that its model was structured based
At the same time, another of the models proposed on Hull and White (1987), Scott (1987), and Wiggins
was that of Scott (1987), which assumed the following (1987), which has the following functional structure:
structure:
dSt = µStdt + σ tStdWt(1)
(1)
dSt = µStdt + σ tStdWt 1
dσ t = σ t  γ 2 + β( a − ln σ t ) dt + γσ tdWt( 2) (2)
∗
dσ t = β(σ − σ t )dt + γdWt (2)  2 

As a novelty, st which is the instantaneous vola- with E(dWt(1)dWt( 2) ) = δdt . The equation dst was con-
tility of the price of the asset, follows a mean-reversion structed such that lnst is a mean-reversion process:
process of the Ornstein–Uhlenbeck type. The param- d ln σ t = β( a − ln σ t )dt + γdWt( 2). To obtain dst in equa-
eters a, b, s*, and g are constant; s* is the mean long- tion (2), let x = lnst and dx = β( a − x )dt + γdWt( 2) ; if Itô’s
term volatility of st; b is the speed reversion from st lemma is applied, then st = exp(x). It is assumed that both
to s*; and g is the volatility of st . At the same time, the foreign and domestic interest rates do not demon-
Wt (1) and Wt ( 2) correspond to Wiener processes with a strate stochastic behavior.
correlation coefficient r. It was concluded that it was Using the fundamental partial differential equation
not possible to find an analytical solution of the pro- used in the contingent claims method that includes the
posed SDE, but that its results were satisfactory when aggregate wealth and state variables of the economy,
the estimated prices of options were contrasted by the but based on Cox et al. (1985, 2005), it was possible to
Monte Carlo simulation method (Scott 1987), and it demonstrate that the wealth variable can be eliminated
was also indicated that using this model is marginally if the utility function includes constant relative risk aver-
better at explaining these prices. sion, and the claim payments do not depend directly on

Fall 2019 The Journal of Derivatives   51

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 wealth. By taking the conditional expectation according 1997) and Drost and Werker (1996), who proposed the
to the information that it has at the moment t, that is, first approach between a GARCH process and a sto-
St and st , and the adjusted risk is incorporated, the fol- chastic volatility model, and indicated that it is possible
lowing system of equations is obtained: to exchange between such processes as required; later,
Ritchken and Trevor (1999) developed the same idea
dSt = (rd − r f )Stdt + σ tStdWt(1) but expressed it in an algorithm to evaluate options,
both European and American, based on the construc-
 1 
dσ t =  σ t  γ 2 + β( a − ln σ t ) − φσ  dt + γσ tdWt( 2) tion of trinomial trees. Subsequently, Barone-Adesi et
 2   
al. (2005) analytically derived the first four moments
of the model and obtained a closed solution to easily
Later, the system of equations presented by Stein
value an option; in addition, they analyzed the implicit
and Stein (1991) to model the price St of a financial asset
volatility surfaces with this solution. An important
with variance Vt = σ t2 , which resembles that of Scott
condition was discovered by Christoffersen et al.
(1987), was as follows:
(2010), which empirically demonstrated, through the
dSt = µStdt + σ tStdWt(1) use of the realized volatilities, returns of the S&P 500
and options data panel; they claimed that the Heston
dσ t = −δ(σ t − θ)dt + κdWt( 2) (1993) model was poorly specified because, in the dif-
fusion model presented by the author, volatility is in
The novelty in this case is that the authors used the square root, instead of being considered linear;
analytical techniques to derive a closed explicit solution the same conclusions were reached by Chourdakis
for when volatility follows an Ornstein-Uhlenbeck pro- and Dotsis (2011), although they suggested that the
cess (or AR1 process), with a reversion tendency to the model should be considered a nonlinear drift against
long-term mean q (Stein and Stein 1991). the linear one.
One of the best-known and referenced models in The most recent research on this model is con-
the literature is that of Heston, presented in 1993, which sidered more empirical and is used successfully in
was described to model the price St of a financial asset applications in different contexts; for example, Figà-
with variance Vt = σ t2 , in the following SDE (Heston Talamanca (2009) used multiple stock indexes to com-
1993): pare the theoretical and empirical autocovariances
according to the aforementioned model, with the
dSt = µStdt + vt StdWt(1) conclusion that this model allows the autocovariance
dvt = κ(θ − vt )dt + σ vt dWt( 2) observed in the data to be captured. At the same time,
Plienpanich et al. (2009) integrated into the diffusion
where m, q, k, and s are constants; m is the mean of St; model a disturbance through fractional noise, and the
q is the mean long-term volatility; k is the speed rever- results showed that the estimation of the stock price of
sion of nt to its mean q; and s is the volatility associ- a commercial bank is better using this model than the
ated with nt , known as the volatility of volatility; while traditional one. Similar conclusions were obtained by
Wt(1) and Wt( 2) are Wiener processes with correlation Wu et al. (2012), who analyzed the Hang Seng index
coefficient r and dWt(1)dWt( 2) = ρdt , or, equivalently, and concluded that the GARCH-diffusion model is
better for predicting the price of the warrants than
dWt( 2) = ρdWt(1) + 1 − ρ2 dZ t , where Zt is an indepen- the classic Black–Scholes volatility model; years later,
dent Wiener process of Wt(1) . Wu et al. (2014) studied the Hong Kong stock market
An SDE model like that of Heston is the GARCH- through American options and found the same advan-
diffusion type, which was first introduced by Wong in tage in using this model.
1964, but its popularity only grew from the works of In general terms, this type of model is usually char-
Nelson (1990a, 1990b). Recent studies have demon- acterized by not having a closed solution and is in the
strated that this model allows a better description of class of non-affine models; in addition, these models’
the behavior of financial time series than other types solutions must be achieved with simulation or numerical
of models, such as that of Heston (1993). Duan (1996, methods (Wu et al. 2012). The system of traditional

52   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 equations presented by this model has the following moments of the SDE, in order to estimate a multiplica-
functional structure (Barone-Adesi et al. 2005): tive tree for each equation, with the purpose of obtaining
the dynamic transition probabilities presented below.
dSt = µStdt + Vt StdWt(1) Proposition 1. To consider the SDE given by
dVt = (c 1 − c 2Vt )dt + c 3VtdWt( 2) (3)
dSt = µStdt + Vt StdBt(1) (5)
where c1, c2 and c3 are positive constants, m is the tendency
parameter, c 2 is the mean-reversion rate, (c1)/(c 2) is the dVt = α(θ − Vt )dt + σVtdBt( 2) (6)
mean long-term variance, and c 3 models the random
behavior of volatility and corresponds usually to the vol- over the time interval [ti, tk], where m ∈ R, a ≥ 0, q > 0, and s >
atility of volatility. For its part, Wt(1) and Wt( 2) correspond 0 are constant, while {Bt(1)}t ≥0 and {Bt( 2)}t ≥0 are independent
to Wiener processes with a correlation coefficient r. one-dimensional standard Brownian motions. Suppose further
If c 2 has a low value, then the reversion to the mean is St = Si and Vt = Vi.
weak, so the process Vt tends to be above or below the
long-term mean for a long period, whereas if c 3 equals 1. For Equation 6, the calculations of the first and second
zero, this implies a deterministic process of volatility, moments are given by
and if c 3 > 0, then the distribution of the returns will
a) E[Vt|Vti ] = θ + (Vti − θ)exp( −α(t − ti ))
be leptokurtic—that is, the kurtosis will be greater
than 3; finally, if c1 = c 2 = 0, the process is reduced to a  2 2αθ2 
V ti + 2
log-normal one without tendency, as shown by Hull  ( σ − 2α ) 
b) E[Vt 2|Vti ] =   exp((σ − 2α )(t − ti ))
2
and White (1987). 2 αθ (V − θ )
 ti 
An alternative way to present the structural form  − 
of Equation 3 is as follows (Wu et al. 2012): α − σ2
2αθ(Vti − θ) 2αθ2
+ exp( −α ( t − t )) −
dSt = µStdt + Vt StdWt(1) α − σ2
i
( σ 2 − 2α )
dVt = α(θ − Vt )dt + σVtdWt( 2) (4) 2. For Equation 5, the calculations of the first and second
moments are given by
where the parameters a, q, and s are constant and are
equivalent to the mean-reversion speed, the mean long- a) E[St|Sti ] = Sti exp(µ(t − ti ))
term volatility or tendency, and the volatility of vola-
tility, respectively.  (2µ + θ)(t − ti ) 
This article focuses in particular on using b) E[ S 2
|
tS ti ] = S 2
exp
ti
 (V t − θ ) 
− i (exp−α( t −ti ) − 1)
Equation 4, that is, the GARCH-diffusion model, to  α 
develop in detail the numerical method for a multiplica-
tive quadrinomial tree that includes stochastic volatility; 3. The first combined moment of Equations 5 and 6 is
this requires an analytical development to derive the given by
first two moments of the SDE, in order to estimate the
a) E[St ,Vt|Sti ,Vti ]
proper recombination between the continuous and dis-
crete processes; then it calculates the respective transition = SV
t t + ( µ − α )SV
t t (t − ti ) + ( αθ)St (t − ti )

factors, which will allow the valuation of a derivative

instrument, such as options, of a financial type as real. Proof.
For all subsequent calculations the higher-order
FIRST MOMENTS FOR THE terms for ∆t were ignored.
GARCH-DIFFUSION MODEL 1a). Equation 6 can be expressed in integral
form as
To construct a multiplicative tree that allows one to
describe the discrete behavior of the price of the under- Vt − V0 = α ∫ tti (θ − Vu )du + σ ∫ tti VudBu
lying asset, it is necessary to determine the first two

Fall 2019 The Journal of Derivatives   53

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 The term E[], corresponds to a conditional Likewise,
expected value, for example E[Vt ]=E[Vt|Vti ], and now
it takes the expected value E[], Q
M (t ) tk = Vti2 + [Vti2 (σ 2 − 2α ) + 2αθVti ]∆t + O( ∆t ). (11)

E[Vt ] = E[Vti ] + α ∫ tti (θ − E[Vu ])du. Now taking m(t) = E[St] and P (t ) = E[St2 ] = E[Yt ],
similarly to 1a) and 1b), the ordinary differential equa-
Let h(t) = E[Vt]; in this way, the ordinary differen- tions are obtained,
tial equation is obtained,
m (t ) − µm(t ) = 0
h(t ) + αh(t ) = αθ (7) and
whose solution is given as P (t ) = (2µ + h(t ))P (t ) (12)
h(t ) = θ + (Vti − θ)exp( −α(t − ti )). (8) whose explicit solutions are 2a) and 2b) respectively:
Likewise, m(t ) = Sti exp(µ(t − ti )) (13)

Q
h(t ) tk = Vti − (Vti − θ)α( ∆t ) + O( ∆t ). (9) and

1b). Let Yt = Vt 2 ; by applying Itô’s formula on Yt  (2µ + θ)(t − ti ) 

it obtains 
P (t ) = S exp (Vti − θ)
2 .
ti
− (exp( −α(t − ti )) − 1)
 α 
dYt = (2αθVt − 2αVt 2 + σ 2Vt 2 )dt + 2σVt 2dBt ,

in an integral form, Likewise,

Yt − Yti = ∫ tti (2αθV s + (σ 2 − 2α )V s2 )ds + ∫ tti 2σV s2dB s . Q

m(t ) tk = Sti exp(µ∆t )

By taking the expected value on both sides of the and

equation, where M (t ) = E[Yt ] = E[Vt 2 ] , it obtains
Q
P (t ) tk = St2i exp(2µ∆t + V j( i ) ∆t + O( ∆t )). (14)
M (t ) = M (ti ) + ∫ (2αθh( s ) + (σ − 2α )M ( s ))ds.
t
ti
2

3a). Let Zt = StVt; if Itô’s formula is applied on Zt

In this way, the ordinary differential equation is it reaches that
obtained, 1
dZ t = VtdSt + StdVt + (dStdVt + dVtdSt ),
M (t ) − (σ 2 − 2α )M (t ) = 2αθh(t ). (10) 2

in the integral form,

Substituting Equation 8 for 10, it follows that
Z t − Z ti = ∫ tti (µSuVu + αθSu − αSuVu )du
 2 2αθ2 
Vti + 2
 ( σ − 2α ) 
3
+ ∫ tti Vu 2SudBu(1) + ∫ tti σSuVudBu( 2) .
 exp((σ − 2α )(t − ti ))
2
M (t )= 
 2αθ(Vti − θ) 
 −  If the expected value is taken and it def ines
α − σ2
N(t) = E [StVt], then
2αθ(Vti − θ) 2αθ2
+ exp( −α(t − ti )) − 2
α−σ 2
( σ − 2α ) N (t ) = N (ti ) + ∫ tti [(µ − α )N (u ) + αθm(u )]du.

54   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 In this way, the ordinary differential equation is Exhibit 1
obtained, Traditional Multiplicative Binomial Tree

N (t ) − (µ − α )N (t ) = αθSti exp(µ(t − ti ))
Xoun
whose solution is given as
Xou3
N (t ) = ( N (ti ) − θSti )exp((µ − α )(t − ti )) + θSti exp(µ(t − ti )). Xoun–1d
Xou 2

As N (ti ) = StiVti , then Xou Xou2d

Q
N (t ) tk = StiVti + (µ − α )StiVti ∆t + αθSti ∆t + O∆t. (15)
Xo Xoud
Xoud 2
Xod

RECOMBINATION FOR DISCRETE Xod 2

PROCESSES Xoud n–1
Xod 3

Binomial Recombination
Xod n
To consider a stochastic process Xt discretely mod-
0 ∆t 2∆t 3∆t ... n∆t = T
eled over some time interval, in each period i the possible Notes: Traditional multiplicative binomial tree, with X0 as the initial
states of the process are given by a fixed number of vec- value of the underlying asset; u is the upward parameter and d is the
decrease parameter. Graph made using Matlab software. (Pareja-Vasseur
tors X ( i ) = ( X (ji ) ), where j is part of a finite indexed set J(i). and Cadavid Pérez 2016).
As the time f lows from period i to i + 1, the discrete
node X (ji ) can go to any X (ji +1), j ∈ J(i+1) (Lari-Lavassani
et al. 2001). AV (Vti , ∆t ) − d (ji )
The discrete process determines a matrix of tran- p =
(i )
j ;
u(ji ) − d (ji )
sition probabilities P ( i ) = (Pj(,ij) ), so that the element Pj(,ij)
represents the probability that X (ji ) moves to X (ji +1). The 1
u(ji ) = exp(cosh −1(CV (Vti , ∆t ))); d (ji ) = .
number of rows in P (i) is the cardinal of J (i), and the u(ji )
number of columns is the cardinal of J(i+1), so that for all
j ∈ J(i), ∑ j ∈ J ( i +1) Pj(,ij) = 1 (Marín Sánchez 2010). Proof.
If the definition of Ed[• | •] is used, the matching of
Calculation of Transition Probability the processes in discrete time and in continuous time
and Growth Factors for Vt turns out to be

Proposition 2. To consider the stochastic differential Ed [V ( i )|V j( i ) ] Q ti +∆t := p(ji )V j(+i +11) + (1 − p(ji ) )V j( i +1)
Equation 6 and the multiplicative recombination in binomial
tree V ( i ) = (V j( i ) ), assume that t = ti and Vti = V j( i ). Over the = E[Vt|Vti ] Q ti +∆t

time interval [ti, ti + Dt], the conditional expectation of the

Q
continuous process is denoted by E[Vt|Vti ] ti +∆t , and for the
(i ) 2
Ed [(V ) |V ] (i )
j Q ti +∆t
(i )
:= p j (V ( i +1) 2
j +1 ) + (1 − p(ji ) )(V j( i +1) )2

Q
discrete process, by Ed [V ( i )|V j( i ) ] ti +∆t . = E[Vt 2|Vti ] Q ti +∆t .
The matching of the first and second moments for these
two processes allows one to obtain, as a result, a recombination
specified by Since V j(+i +11) = V j( i )u(ji ) and V j( i +1) = V j( i )d (ji ), therefore
the previous equations are reduced to

Fall 2019 The Journal of Derivatives   55

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 p(ji )V j( i )u(ji ) + (1 − p(ji ) )V j( i )d (ji ) = E[Vt|Vti ] Q ti +∆t
Proposition 3. To consider the stochastic differential
Equation 6, over the time interval [t0, T]. Let V ( i ) = (V j( i ) ),
p(ji ) (V j( i )u(ji ) )2 + (1 − p(ji ) )(V j( i )d (ji ) )2 = E[Vt 2|Vti ] Q ti +∆t . a recombination in multiplicative binomial trees with a time
step Dt. Assume that there is a lower bound Vmin > 0 such
The above equations can be rewritten as that V j( i ) > Vmin for all i, j. If the first and second moments are
matched for the discrete and continuous processes, this results in a

p(ji )u(ji ) + (1 − p(ji ) )d (ji ) = AV (Vti , ∆t ) :=

E[Vt|Vti ] Q ti +∆t
(16)
recombination whose transition probability is perfectly specified by
(i )
V 1
j
u(ji ) = exp(σ ∆t ), d (ji ) = ,
u(ji )
p(ji ) (u(ji ) )2 + (1 − p(ji ) )(d (ji ) )2   θ  

= BV (Vti , ∆t ) :=
2
E[Vt |Vti ] Q ti +∆t
. (17)
A(V j( i ) ) = AV = exp  α  ( i ) − 1 ∆t  and
 Vj  
(i ) 2
(V ) j
 αθ 1 2
 V ( i ) − α − 2 σ  ∆t
From Equation 16 it follows that 1  j 
p(ji ) = + .
2 2σ
AV (Vti , ∆t ) − d (ji ) Proof.
p(ji ) = . (18) The proof of this proposition is based on the expan-
u(ji ) − d (ji )
sion in Taylor’s series in Dt with respect to V (i). It is also
observed that in this case both u(ji ) = u and d (ji ) = d are
If Equation 18 is replaced into 17, it obtains constant (Marín Sánchez 2010; Lari-Lavassani et al. 2001).
From Proposition 2 and Equation 11, it is found
BV (Vti , ∆t ) = AV (Vti , ∆t )(u(ji ) + d (ji ) ) − u(ji )d (ji ) . (19) that

Without loss of generality and to remove an extra Vti2 + [Vti2 (σ 2 − 2α ) + 2αθVti ]∆t + O( ∆t )
degree of freedom, suppose that u(ji )d (ji ) = 1, that is, BV (Vti , ∆t ) =
(V j( i ) )2
d (ji ) = 1/u(ji ) and, therefore, from Equation 19 it obtains
 2αθ 
u(ji ) + d (ji ) 1 + BV (Vti , ∆t ) BV (Vti , ∆t ) = 1 + (σ 2 − 2α ) + ( i )  ∆t + O( ∆t ).
= CV (Vti , ∆t ) := .  V j 
2 2 AV (Vti , ∆t )
From Equation 9 to AV (Vti , ∆t ) = E[Vt ]/V j( i ) , it
It follows that it is possible to derive from CV (Vti , ∆t ) follows that
an expression to determine u(ji ) in the following way:
Vti − (Vti − θ)α( ∆t ) + O( ∆t )
1 AV (Vti , ∆t ) =
u + (i )
(i )
j V j( i )
uj
= CV (Vti , ∆t ).   θ
2  
AV (Vti , ∆t ) = exp  α  ( i ) − 1 ∆t  . (21)
By taking the positive part of the root  Vj  

(
ln(u(ji ) ) = ln CV (Vti , ∆t ) + CV2 (Vti , ∆t ) − 1 ) (20)
Finally, CV (Vti , ∆t ) can be written as

  2αθ  
so that 1 +  1 + (σ 2 − 2α ) + ( i )  ∆t + O( ∆t )
  Vj  
CV (Vti , ∆t ) =
u(ji ) = exp(cosh −1(CV (Vti , ∆t ))).   θ  
2  1 −  1 − ( i )  α( ∆t ) + O( ∆t )
  Vj  

56   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 where
  σ2   σ2
CV (Vti , ∆t ) = exp    ∆t  ≈ 1 + ( ∆t ) + O( ∆t ). (22)
 2   2
AS (Sti , ∆t ) :=
E[St|Sti ] Q ti +∆t
;
(i )
S
So, j

C (Vti , ∆t ) = exp(σ ∆t ) ≈ 1 + σ ( ∆t ) + O( ∆t ).
2 2 2
(23) BS (Sti , ∆t ) :=
E[St2|Sti ] Q ti +∆t
;
(i ) 2
V
(S ) j
Substituting Equations 22 and 23 into 20, it 1 + BSt (Sti , ∆t )
obtains C S (Sti , ∆t ) := .
2 ASt (Sti , ∆t )
u(ji ) = exp(σ ∆t ). (24)
Proof.
(i )
Finally, it calculates p from Equations 21 and 24 Using the definition of Ed[•|•], the matching of
j
replaced in Equation 18, as follows: the processes in discrete time and continuous time turns
out to be
 1 − σ ∆t 
 θ 
1 −  1 − ( i )  α( ∆t ) + O( ∆t ) −  σ 2 ∆t  Q
Ed [S ( i )|S (ji ) ] ti +∆t := q(ji )S (ji++11) + (1 − q(ji ) )S (ji +1)

p(ji ) =
 Vj  +
 2
+ O( ∆t )
 = E[St|Sti ] Q ti +∆t

 σ 2 ∆t  
 1 − σ ∆t 

2
Ed [S ( i ) |S (ji ) ] Q ti +∆t := q(ji ) (S (ji++11) )2 + (1 − q(ji ) )(S (ji +1) )2
 1 + σ ∆t + 2 + O( ∆t ) −  σ 2 ∆t

+
2
+ O( ∆t )

= E[St2|Sti ] Q ti +∆t .

 αθ 1 2 Since S (ji++11) = S (ji )h (ji ) and S (ji +1) = S (ji )l (ji ) , these equa-
 V ( i ) − α − 2 σ  ∆t tions are reduced to
1  j 
p(ji ) = + + O( ∆t ).
2 2σ q(ji )S (ji )h (ji ) + (1 − q(ji ) )S (ji )l (ji ) = E[St|Sti ] Q ti +∆t

Calculation of Transition Probability

q(ji ) (S (ji )h (ji ) )2 + (1 − q(ji ) )(S (ji )l (ji ) )2 = E[St2|Sti ] Q ti +∆t .
and Growth Factors for St
These equations can be rewritten as
Proposition 4. To consider the stochastic differen-
tial Equation 5 and the multiplicative recombination in bino-
q(ji )h (ji ) + (1 − q(ji ) )l (ji ) = AS (Sti , ∆t ) :=
E[St|Sti ] Q ti +∆t
(25)
mial tree S ( i ) = (S (ji ) ), assume that t = ti and Sti = S (ji ) . Over S (i )
j
the time interval [ti , ti , + Dt], the conditional expectation
Q
of the continuous process is denoted by E[St |Sti ] ti +∆t , and for E[St2|Sti ] Q
the discrete process, by Ed [S ( i ) |S (ji ) ] ti +∆t . Q q(ji ) (h (ji ) )2 + (1 − q(ji ) )(l (ji ) )2 = BS (Sti , ∆t ) :=
(S ) (i ) 2
ti +∆t
.
(26)
j
The matching of the first and second moments for these
two processes allows one to obtain, as a result, a recombination
specified by From Equation 25 it follows that
AS (Sti , ∆t ) − l (ji ) AS (Sti , ∆t ) − l (ji )
q(ji ) = ; q =
(i )
. (27)
h −l
(i )
j
(i )
j
j
h (ji ) − l (ji )
1
h (ji ) = exp(cosh −1(C S (Sti , ∆t ))); l (ji ) =
h (ji ) By substituting Equation 27 for Equation 26,
it obtains

AS (Sti , ∆t )(h (ji ) + l (ji ) ) − h (ji )l (ji ) = BS (Sti , ∆t ). (28)

Fall 2019 The Journal of Derivatives   57

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Without loss of generality and to remove an extra so,
degree of freedom, suppose that h (ji )l (ji ) = 1, that is,
l (ji ) = 1/h (ji ) and, therefore, from Equation 28 it obtains 4 + 4V j( i ) ∆t + O( ∆t ) (32)
C S2 (Vti , ∆t ) = .
4
h (ji ) + l (ji ) 1 + BS (Sti , ∆t )
= C S (Sti , ∆t ) := .
2 2 AS (Sti , ∆t ) Substituting Equations 31 and 32 for Equation 29,
it obtains
It follows that it is possible to derive from C S (Sti , ∆t )
an expression to determine hj(i) in the following way: (
h (ji ) = exp V j( i ) ∆t . ) (33)

1 Finally, q(ji ) is estimated from Equations 30 and 33

h (ji ) +
h (ji ) replaced in Equation 27, and it obtains that
= C S (Sti , ∆t )
2
(
ln(h (ji ) ) = ln C S (Sti , ∆t ) + C S2 (Sti , ∆t ) − 1 ) (29)
q =
(i )
exp(µ∆t ) − exp − ( ( V ∆t )) (i )
j

( ∆t ) − exp ( − ( V ∆t ))
j
exp V j( i ) (i )
j
so that
 V j( i ) 
 µ − 2  ∆t
−1
h = exp(cosh (C S (Sti , ∆t ))).
(i )
j 1
q(ji ) = + .
2 2 V j( i )
Proposition 5. To consider the stochastic Equation 5,
over the time interval [t 0, T ], let S ( i ) = (S (ji ) ), a recombination
in multiplicative binomial trees with a time step Dt. Suppose Quadrinomial Recombination
there is a lower bound Smin > 0 such that S (ji ) > Smin for all
i, j. If the first and second moments are matched for the discrete To consider the recombination of two binomial
and continuous processes, this results in a recombination whose trees S ( i ) = (S (ji ) ) , j ∈ J(i) and V ( i ) = (Vk( i ) ), k ∈ K(i), which
transition probability is perfectly specified by possess the same number of nodes along their time
axes, that is, J (i) = K (i) for all i, denote the transition
(
h (ji ) = exp V j( i ) ∆t , l (ji ) = ) 1
h (ji )
, probabilities of S (i) by q(ji ) with increases up and down
defined by h (ji ) and l (ji ), and for V (i) by pk( i ), uk( i ), and dk( i ),
 V j( i )  respectively. The direct product S (i) × V (i) is defined
1  µ − 2  ∆t by a tree T (i) with a node described by T j(,ik) = (S (ji ) ,Vk( i ) )
AS = exp(µ∆t ) and q(ji ) = + . at time i. In the next step, T j(,ik) generates four nodes
2 2 V j( i ) that are T j(+i +1,1)k +1 = (S (ji )h (ji ) ,Vk( i )uk( i ) ), T j(,ik++1)1 = (S (ji )l (ji ) ,Vk( i )uk( i ) ),
T j(+i +1,1)k = (S (ji )h (ji ) ,Vk( i )dk( i ) ), and T j(,ik+1) = (S (ji )l (ji ) ,Vk( i )dk( i ) ), whose
Proof.
respective probabilities are q(ji ) pk( i ), (1 − q(ji ) ) pk( i ), q j (1 − pk ),
(i ) (i )
From Equation 14, it follows that
and (1 − q(ji ) )(1 − pk( i ) ), respectively.
BS (Sti , ∆t ) = exp((2µ + V j( i ) )∆t ). Other recombination types using trees have been
studied in the literature, in which the dynamic price
If Equation 13 to AS (Sti , ∆t ) = E[St ]/S (ji ) it obtains behavior follows a binomial, trinomial, and multinomial
rule. For example, Nelson and Ramaswamy (1990), based
AS (Sti , ∆t ) = exp(µ∆t ). (30) on Cox et al. (1979), proposed additive binomial recombi-
nations for different diffusion models and also showed how
In consequence, to employ a transformation to produce computationally a
binomial process in a simple way. On the other hand, Flo-
2 + V j( i ) ∆t + O( ∆t ) rescu and Viens (2008) used trinomial and quadrinomial
C S (Sti , ∆t ) = , (31)
2

58   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 2 Exhibit 3
First Step of Proposed Multiplicative Quadrinomial Recombination in the Second Step of the
Tree Model to Stochastic Asset Price and Volatility Multiplicative Quadrinomial Tree Model
L i +2
7MN Tj,k+2
i+2
T j+1,k+2

L
7 MN 
i+1
Tj,k+1
i +2
Tj,k+1
L
7M N i
Tj,k

L
7MN
Tj,ki+2
i+1
Tj+1,k

i+2
Tj+2,k+1 i+2
Tj+1,k

L
7MN

i+2
Notes: This exhibit shows each position for the first step of the tree and Tj+2,k
the respective algebraic condition, with T2,2 = T1,1 uh1 , T1,2 = T1,1 ul1 ,
( 2) (1) (1) ( 2) (1) (1)

T2,1 = T1,1 dh1 , and T1,1 = T1,1 dl1 . Graph made using Photoshop
( 2) (1) (1) ( 2) (1) (1) Notes: This exhibit shows each position for the second step of the tree and
software. the respective algebraic condition. Graph made using Photoshop software.

recombinations in additive form for other types of diffu- 2. To consider S ( i ) = (S j ( i ) ) and V ( i ) = (Vk ( i ) ), defined
sion models. Note that the authors in this research propose in a quadrinomial recombination with a time step Dt,
a methodology that is based on a multiplicative quadri- assume that there are lower and upper limits Smin ≠ 0
nomial recombination from two multiplicative binomial and Vmax < ∞, such that S (ji ) > Smin and Vk( i ) < Vmax for
recombinations, whose values of the discrete process can all i, j, k. If the first and second moments of S(i) and V (i)
be represented in a three-dimensional space (Exhibit 2), are matched with those of St and Vt, that is, the discrete
while for the previously reported models they were raised and continuous processes, this results in a recombination
only under a two-dimensional space (Exhibit 1). whose transition probability is specified for V (i) and S (i),
After all the corresponding parameters have been in their order, as follows:
defined for Vt and for St , it is possible to construct a
tree that emulates the behavior of the price St through  αθ 1 2
the construction of a quadrinomial tree, which in each  V ( i ) − α − 2 σ  ∆t
discrete step has a value of branches equal to 4n , where 1  j 
pk( i ) = + ;
n corresponds to the number of steps, but when the 2 2σ
respective recombination is performed, the number of
branches decreases to n 2, as can be seen in Exhibit 3. ( )1
uk( i ) = exp σ ∆t ; dk( i ) = ( i )
uk
Below is the value of each position for the first step of
the tree, as shown in Exhibit 2.
 V j( i ) 
Proposition 6. To consider the SDE 5 and 6 over µ − ∆t
1  2 
the interval [t 0, T ],
q(ji ) = + ;
2 2 V j( i )
1. Over any time interval [ti, ti + Dt] ⊂ [t 0, T ], the pro-
cesses St and Vt will remain uncorrelated until the order
O( ∆t, Sti ,Vti ).
(
h (ji ) = exp V j( i ) ∆t ; l (ji ) = ) 1
h (ji )

Fall 2019 The Journal of Derivatives   59

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Proof. acquiring the option is known as the premium. Among
the methods used in its valuation, the multiplicative
1. For each (S (ji ) ,Vk( i ) ) the two processes St and Vt binomial method is the most often used and is appro-
remain uncorrelated on [ti, ti + Dt]. Therefore, at priate because it is easy to understand and construct; it
a certain point and on [ti , ti + Dt], St and Vt can allows one to both emulate the price of the underlying
be approximated with independent binomial trees asset in discrete time graphically and offer the respective
and, in turn, can be approximated by the direct valuation of the derivative.
product of these two trees over the same time This section brief ly describes the option valua-
interval. tion method through multiplicative recombination in
2. Assume that t = ti and that the value of St is S (ji ) and quadrinomial trees according to the subjective prob-
that for Vt is Vk( i ). From the first-order approxima- ability P, which is presented in the real world, which
tion of the results obtained in Proposition 1, it has arbitrage opportunities and absence of risk neutrality.
obtains that Finally, the necessary transformation is offered, which
allows one to make the same valuation according to the
E[St|S (ji ) ] Q ti +∆t = Sti exp(µ( ∆t )) risk-neutrality condition.
]Q
Most of the volatility estimation methods are based
E[Vt|Vk( i ) ti +∆t = Vti − (Vti − θ)α( ∆t ) + O( ∆t ) on a single-level stochastic differential equation, which
E[St , St|S (ji ) ] Qti +∆t = St2i exp(2µ∆t + V j( i ) ∆t + O( ∆t )) only allows modeling the asset’s random price, but fails
to include the effect of volatility on this variable. A dif-
Vti2 (σ 2 − 2α )
]Q
ferential equation that is commonly used employs the
ti +∆t = Vti +   ∆t + O( ∆t )
(i ) 2
E[Vt ,Vt|V k
 + 2αθVti  following structure:

E[St ,Vt|S (ji ) ,Vk( i ) ] Q ti +∆t = StiVti + (µ − α )StiVti ∆t dSt = µStdt + σStdBt (35)
+ αθSti ∆t + O( ∆t ).
with t ∈ [0, T ] and with the initial condition S 0 = S,
where m is constant and denotes the asset’s average rate
So
of return; s > 0 corresponds to volatility, and {Bt}t≥0 is a
OV[St ,Vt ] Q ti +∆t = E[St ,Vt |S (ji ) ,Vk( i ) ] Q ti +∆t
one-dimensional standard Brownian motion.
Assume that the life of an option (real or finan-
− E[St |S ] (i )
j Q ti +∆t E[Vt |V ] k
(i )
Q ti +∆t = 0. (34) cial) on an asset that does not pay dividends, with an
initial price S 0 and strike price K, is divided by N sub-
If Equations 15, 13, and 9 are replaced by intervals, each with duration Dt. Define f j(,ik) as the value
Equation 34, it obtains of the option in the node (i, j, k). Based on Marín Sán-
chez (2010), the price of the asset has a quadrinomial
OV[St ,Vt ] = [StiVti + (µ − α )StiVti ∆t + αθSti ∆t + O( ∆t )] recombination in the node (i, j, k), which can be repre-
− [StiVti + (µ − α )StiVti ∆t + αθSti ∆t + O( ∆t )] = 0. sented by the following expression:
k −1 i −k j −1 i− j
Consequently (S (ji ) ,Vk( i ) ), and for higher terms of T j(,ik) = S0 ∏ hww ∏ lk i − w ∗V0 ∏ umm ∏ d j i − m (36)
O( ∆t, Sti ,Vti ), the processes St and Vt will remain uncor- w =1 w =1 m =1 m =1

related on [ti, ti + Dt].

with T0 = T1,1(1) , i = 2, …, N, j = 1, 2, …, i, and k = 1,
2, …, i. It is important to keep in mind that in this case
OPTIONS VALUATION
both u and d are constant, so Equation 36 can be sum-
It is common knowledge that an option is con- marized as follows:
sidered a financial instrument that grants a right to the k −1 i −k

buyer and an obligation to the seller to buy or sell a T j(,ik) = S0 ∏ hww ∏ lk i − w ∗V0u j −1d i − j .
w =1 w =1
certain asset or underlying asset at a price established
on a fixed date; the exchange price that is obtained for Below is an algebraic approach based on the proposed
methodology to assess the basic financial options.

60   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Financial Options Valuation a single probability measure P* must be constructed from
the subjective P, such that the discounted price at interest
• In the case of a European call option, its value at rate r is a martingale and that, as mentioned earlier,
maturity date is given by max(Tt – K, 0); so the discounted expected value at this rate according
to the new probability P * does not present arbitrage
 k −1 N − k 
f j(,Nk ) = max  S0 ∏hww ∏lkN −w ∗V0u j −1d N − j − K ,0 opportunities.
 w =1 w =1  Following Heston (1993), Marín Sánchez (2010),
Wu et al. (2012), and Wu et al. (2014), and applying the
for j = 1, 2, …, i and k = 1, 2, …, i. In addition, its theorems of Cameron-Martin-Girsanov and Girsanov
discounted value is defined as (Mao 1997), the SDE proposed may be written as

 q(ji ) pk( i ) f j(+i +1,1)k +1 + (1 − q(ji ) ) pk( i ) f j(,ik++1)1  dSt = rStdt + Vt StdBt(1)∗ (37)
( i ) −1 −1
f =  (i )
(i )
j ,k ( i ) ( i +1) ( i ) ( i +1)
 A(V j ) AS .
+
 jq (1 − p k ) f j +1,k + (1 − q (i )
j )(1 − p k ) f j ,k 

dVt = α ∗ (θ∗ − Vt )dt + σVtdBt( 2)∗ (38)
• For the European put option, the value at maturity
is evaluated as max(K − Tt , 0), and the discounted where α ∗ = (α + λ ), θ∗ = αθ/(α + λ ), {Bt(1)∗}t ≥0 , and
value is valued in the same way as that used for the {Bt( 2)∗}t ≥0 are independent one-dimensional standard
European call option. Brownian motions over the probability space (W, F, P*).
• For the American financial call option, the value The term l(S, V, t) represents the price of volatility
at its maturity date is given by max(Tt – K), so risk and is independent of the particular asset, and there
exists evidence that this term is nonzero for options
 k −1 N − k  (Heston 1993), while in contrast, Hull and White (1987)
f j(,Nk ) = max  S0 ∏hww ∏lkN −w ∗V0u j −1d N − j − K ,0 proposed to set this term at zero, based on the assump-
 w =1 w =1 
tion that this term is independent from the aggregate
consumption. This term is the variance risk premium
f or j = 1, 2, …, i and k = 1, 2, …, i, while its dis-
as linear function of variance, namely, l(Vt , t) = lVt
counted value is defined as
(Wu et al. 2012; Wu et al. 2014). In theory, following
Heston (1993), this parameter should be determined for
  q(ji ) pk( i ) f j(+i +1,1)k +1
f (i )
= max T j ,k − K , 
(i ) each particular asset.
j ,k ( i ) ( i +1)
 + (1 − q j ) pk f j ,k +1
(i )

NUMERICAL EXPERIMENTS
+ q(ji ) (1 − pk( i ) ) f j(+i +1,1)k
 
( i +1) 
A(V j( i ) )−1 AS−1  . Comparison between the Quadrinomial
+ (1 − q j )(1 − pk ) f j ,k 
(i ) (i )

Method with Stochastic Volatility, the
Binomial Method with Constant Volatility,
• For the American put option, its value at maturity
and the Black–Scholes Equation
is evaluated as max(K − Tt , 0), and the discounted
value is valued in the same way as that used for the Before presenting examples of applying the pro-
American call option. posed methodology for financial options, it is neces-
sary to illustrate graphically the asymptotic behavior of
Risk-Neutral Valuation the quadrinomial and binomial methods by comparing
the results with those obtained with the Black–Scholes
To evaluate options in a risk-neutral world, two model.
conditions must be assumed: the estimated future cash Suppose that the dynamic behavior of daily prices
f lows can be determined by discounting their expected of a given financial asset can be modeled by a geometric
values at the risk-free rate, and the expected return on Brownian motion, using Equation 35 with an initial
derivatives is the risk-free interest rate. To eliminate the condition S 0 = 10 and an annualized volatility sBS =
potential for arbitrage on the discounted expected value, 0.3162. Now suppose that one wants to value a European

Fall 2019 The Journal of Derivatives   61

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 4 Exhibit 5
Comparison between the Quadrinomial Proposed Comparison of the Value of a European Call Option
Methodology, the Traditional Binomial Method, and in the Real World, between the Quadrinomial Trees
the Black–Scholes Method Method and the Monte Carlo Simulation
1.8 0HWKRG?7LPH       
1.75 Binomial Model 4XDGULQRPLDO7UHH       
Quadrinomial Model
1.7 0RQWH&DUORVLPXODWLRQ       
Black–Scholes Model
European Call Option Value

1.65 Notes: This exhibit shows the quadrinomial proposed methodology and
1.6 Monte Carlo simulation with T0 = 10, K = 10, r = 0.05, N = 200,
V0 = 0.11, α = 0.01, q = 0.1, and s = 0.08. Table made using Matlab
1.55 and Excel software.
1.5
1.45 K = 10 and a risk-free rate of r = 0.05, using the quadri-
1.4 nomial method with stochastic volatility, as well as for
1.35
the Monte Carlo simulation method, which includes
stochastic volatility. Assume that the dynamic behavior
1.3
0 20 40 60 80 100 120 140 160 180 200 of the prices can be modeled using a GARCH-diffusion
Steps model as presented above, with an initial condition T0 =
10, with V0 = 0.11, a = 0.01, q = 0.1, s = 0.08, and N =
Notes: This exhibit shows the quadrinomial proposed methodology with 200. To prove the experimental numerical results, the
T0 = 10, α = 0.01, q = (sBS)2 = V0 = 0.1, and s = 0.000000001 and
the traditional binomial method and Black–Scholes method with sBS = value of the option was determined using the Monte
0.3162, S0 = 10, K = 10, and r = 0.05, for N = 200. Graph made Carlo simulation method with a total of 10,000 trajecto-
using Matlab software. ries and 1,000 repetitions, to estimate its average value.
Exhibit 5 presents the results of the value of the
call financial option derived from the underlying asset, option for different periods. As seen, the values are sim-
whose exercise price is K = 10, with risk-free rate r = ilar, enabling the proposed methodology to be verified
0.05 and time to maturity of T = 1; likewise consider with the use of alternative numerical methods such as
a traditional multiplicative recombination in binomial Monte Carlo simulation, applying stochastic differential
trees with N = 200. It can be verified that the binomial Equations 5 and 6.
method is a particular or special case of the multiplicative Now consider the multiplicative quadrinomial
quadrinomial when V0 = (sBS)2 = q and s tends toward tree of the previous example, but with 0.000000001 ≤
zero; suppose further, in the latter method, that α is an s ≤ 0.4 for a fixed expiration time at T = 3, with an
arbitrary constant. exercise price of K = 11 and V0 = 0.10; suppose further
Exhibit 4 displays the results, showing that the that σ BS = θ = 0.3162 . With these conditions, one can
quadrinomial and binomial methods have very close graphically show the increasing behavior of the option
solutions at around N = 100, with values that are also price in the proposed model for different values of the
close to those of the closed solution estimated by the volatility of volatility. In Exhibits 6 and 7, the parameters
Black–Scholes model. contained in the proposed method were sensitized
according to a subjective probability.
Exhibit 6 shows how the value of the European
Example: Valuation of a European Call
call option changes based on different values assigned
Financial Option and Analysis of the
to the parameter σ in the proposed method. If the last
Sensitivity of The Model Variables in the
variable tends toward zero, the value of sBS in the tra-
Real World and in a Risk-Neutral World
ditional method becomes θ , so that the two models
Suppose one wants to valuate a European call (the proposed and the traditional) have the same results;
option at different maturities, with a strike price of in other words, the quadrinomial method with sto-
chastic volatility includes the traditional methodology

62   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Exhibit 6 Exhibit 7
Comparison of the Value of a European Call Option Sensitivity of the European Call Option
in the Real World, between the Quadrinomial Trees for the Proposed Method in the Real World
Method and the Monte Carlo Simulation

 



(XURSHDQ&DOO2SWLRQ9DOXH
 

(XURSHDQ&DOO2SWLRQ9DOXH

 

 
 


 


%ODFN6FKROHV0RGHO
 
%LQRPLDO0RGHO   
 4XDGULQRPLDO0RGHO    

 $OSKD 6LJPD
        
6LJPD Notes: This exhibit shows a sensitivity analysis of the value of a
European call option in the real world with T0 = 10, K = 11, r = 0.05,
Notes: This exhibit shows the value’s behavior in a European call option T = 3, V0 = 0.1, and q = 0.1, with variations of s and α for N = 100.
between the proposed methodology with 0.000000001 ≤ s ≤ 0.4, the Graph made using Matlab software.
traditional binomial one, and the Black–Scholes method with sBS = θ =
0.3162, for N = 200. Graph made using Matlab software.
Exhibit 8
Sensitivity Analysis of the Value of a European Call
through multiplicative binomial trees with constant
Option in a Risk-Neutral World
volatility, with a value of the option in both methods
of 2.37, which turns out to be equal to the exact or 
closed solution estimated by the Black-Scholes model. 
However, after the value of the volatility of volatility

(XURSHDQ&DOO2SWLRQ9DOXH

in the quadrinomial model begins to change, the value

of the call option tends to increase; for example, when 
the parameter s = 0.4, the option has a value of 4.18,

which is 76% higher.
In addition, it is possible to illustrate a sensitivity 
analysis of the parameters α and s, which are part of the

proposed SDE, to observe how the value of the option
changes due to variations in those parameters. Exhibit 7 
shows the surface that generates these changes, in which, 
for example, for a value of α = 0.001 and from a σ that
tends toward zero, the value of the option is 2.37, a 
          
value that increases to 4.10 after σ = 0.4; moreover, when /DPEGD
modifying α, for example, with a value of 0.2, the value
for the option decreases when σ takes values between Notes: This exhibit shows a sensitivity analysis of the value of a Euro-
pean call option in a risk-neutral world with T0 = 10, K = 11, r = 0.05,
0 and 0.054, and then the option increases after the last T = 3, N = 200, V0 = 0.1, α = 0.01, q = 0.1, and s = 0.1131.
variable has positive higher values. Graph made using Matlab software.
In addition, a sensitivity analysis of the same
European call option is presented, but this time in a

Fall 2019 The Journal of Derivatives   63

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 risk-neutralized world, using the initial conditions of an appropriate valuation of financial options is possible
the example, except for the strike price set at K = 11. regardless of whether volatility is constant.
Exhibit 8 shows the sensitivity of the price of the Future research should try to analyze what hap-
option with respect to the parameter l in the SDE pre- pens to the proposed model when there is a correlation
sented in Equations 37 and 38, in which if the value of between Brownian motions—that is, when it is different
parameter l is zero, the values of α* and q* are the same from zero—as well as using different commodities in
as they were initially, but as the value of the market applications of the valuation of options with the appro-
premium increases, these neutralized parameters change, priate use of the proposed methodology.
just as the value of the option does. For example, while
with l = 0 the option equals 2.56, this value will change REFERENCES
to 0.80 when l = 0.2, α* = 0.21, and q* = 0.0048, which
means that the value of the option decreases about 69%, Barone-Adesi, G., H. Rasmussen, and C. Ravanelli. 2005.
if everything else remains constant. “An Option Pricing Formula for the GARCH Diffusion
Model.” Computational Statistics & Data Analysis 49 (2):
287–310.
DISCUSSION AND CONCLUSIONS
Black, F., and M. Scholes. 1973. “The Pricing of Options
This article extends the literature about numerical
and Corporate Liabilities.” Journal of Political Economy 81 (3):
methods for valuing derivatives, specifically for options; 637–654.
the main advantage of the methodology proposed here
is that it includes nonconstant volatility, from a formal Bollerslev, T. 1986. “Generalized Autoregressive Conditional
derivation of the first two moments of a GARCH- Heteroskedasticity.” Journal of Econometrics 31 (3): 307–327.
diffusion system, to determine the probabilities of transi-
tion, growth factors, and discount rates. Throughout the Chesney, M., and L. Scott. 1989. “Pricing European Cur-
article, a rigorous mathematical technique was proposed, rency Options: A Comparison of the Modified Black-Scholes
which offers a relatively easy method for the public to Model and a Random Variance Model.” Journal of Financial
implement and use, so it can be used to assess all types and Quantitative Analysis 24 (3): 267–284.
of options, in both a risk-neutral world and in the real
Chourdakis, K., and G. Dotsis. 2011. “Maximum Likelihood
world, with nonconstant volatility.
Estimation of Non-Affine Volatility Processes.” Journal of
The findings are summarized as follows. First, the
Empirical Finance 18 (3): 533–545.
methodology proposed here is comparable to the tradi-
tional methodology using multiplicative binomial trees Christoffersen, P., K. Jacobs, and K. Mimouni. 2010. “Vola-
as well as the closed solution proposed by Black and tility Dynamics for the S&P500: Evidence from Realized
Scholes. The above fact occurs when the volatility in the Volatility, Daily Returns, and Option Prices.” The Review of
first mentioned method tends toward zero; after the last Financial Studies 23 (8): 3141–3189.
variable has positive values, the value of the option will
vary, and this change will be in function of the values Cox, J. C., J. E. Ingersoll Jr., and S. A. Ross. 1985. “An
assigned to the parameters α, θ, and σ that are part of the Intertemporal General Equilibrium Model of Asset Prices.”
methodology proposed. Second, it is possible to use an Econometrica: Journal of the Econometric Society 53 (2): 363–384.
algebraic expression to depict the evolution of the price
of an asset in the presence of both dynamic or constant — — . “A Theory of the Term Structure of Interest
volatility, in such a way that its effect is captured in an Rates.” In Theory of Valuation, 2nd Edition Sudipto Bhat-
tacharya (London School of Economics, UK) and George
appropriate manner, which allows better modeling of
M Constantinides (University of Chicago, USA), 2005,
the evolution or behavior of the underlying asset in the pp. 129–164. http://doi.org/10.1142/9789812701022_0005.
market. Third, the proposed method can be used in the
presence of risk neutrality through the appropriate use Cox, J. C., S. A. Ross, and M. Rubinstein. 1979. “Option
of a market risk premium, which has a negative effect Pricing: A Simplified Approach.” Journal of Financial Economics
on the value of a call option and a positive effect on 7 (3): 229–263.
a put option. In addition, the examples showed that

64   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Drost, F. C., and B. J. Werker. 1996. “Closing the GARCH Nelson, D. B. 1990a. “ARCH Models as Diffusion Approxi-
Gap: Continuous Time GARCH Modeling.” Journal of Econo- mations.” Journal of Econometrics 45 (1–2): 7–38.
metrics 74 (1): 31–57.
——. 1990b. “Stationarity and Persistence in the GARCH
Duan, J. C. 1996. “A Unified Theory of Option Pricing under (1, 1) Model.” Econometric Theory 6 (3): 318–334.
Stochastic Volatility: From GARCH to Diffusion.” Hong
Kong, Hong Kong University of Science and Technology. Nelson, D. B., and K. Ramaswamy. 1990. “Simple Bino-
Retrieved from https://rmi.nus.edu.sg/duanjc/index_files/ mial Processes as Diffusion Approximations in Financial
files/opm_sv.pdf. Models.” The Review of Financial Studies 3 (3): 393–430.

——. 1997. “Augmented GARCH (p, q) Process and Its Dif- Pareja-Vasseur, J., and C. Cadavid Pérez. 2016. “Pharma-
fusion Limit.” Journal of Econometrics 79 (1): 97–127. ceutical Patents Valuation through Real Options: Certainty
Equivalents and Utility Function.” Contaduría y Administración
Fernández Castaño, H. 2007. “EGARCH: A Model to Esti- 61 (4): 794–814.
mate the Asymmetric Volatility of Financial Series.” Revista
Ingenierías Universidad de Medellín 9 (16): 49–60. Plienpanich, T., P. Sattayatham, and T. H. Thao. 2009. “Frac-
tional Integrated GARCH Diffusion Limit Models.” Journal
Figà-Talamanca, G. 2009. “Testing Volatility Autocorrela- of the Korean Statistical Society 38 (3): 231–238.
tion in the Constant Elasticity of Variance Stochastic Vola-
tility Model.” Computational Statistics & Data Analysis 53 (6): Posedel, P. 2005. “Properties and Estimation of GARCH (1,
2201–2218. 1) Model.” Metodoloski Zvezki 2 (2): 243.

Florescu, I., and F. G. Viens. 2008. “Stochastic Vola- Ritchken, P., and R. Trevor. 1999. “Pricing Options under
tility: Option Pricing Using a Multinomial Recombining Generalized GARCH and Stochastic Volatility Processes.”
Tree.” Applied Mathematical Finance 15 (2): 151–181. The Journal of Finance 54 (1): 377–402.

Grajales Correa, C. A., and F. O. Pérez Ramírez. 2008. “A Scott, L. O. 1987. “Option Pricing when the Variance
Continuous Model and Discrete Model for Estimating the Changes Randomly: Theory, Estimation, and an Appli-
Stochastic Volatility Probability Density of Financial Series cation.” Journal of Financial and Quantitative Analysis 22 (4):
Yields.” Cuadernos de Administración 21 (36): 113–132. 419–438.

Heston, S. L. 1993. “A Closed-Form Solution for Options Stein, E. M., and J. C. Stein. 1991. “Stock Price Distribu-
with Stochastic Volatility with Applications to Bond and tions with Stochastic Volatility: An Analytic Approach.” The
Currency Options.” The Review of Financial Studies 6 (2): Review of Financial Studies 4 (4): 727–752.
327–343.
Wasserfallen, W., and H. Zimmermann. 1986. “The Wiener
Hull, J., and A. White. 1987. “The Pricing of Options on Process, Variance Measurement, and Option Pricing: Evi-
Assets with Stochastic Volatilities.” The Journal of Finance 42 dence from Intra-Daily Data on Foreign Exchange.” Working
(2): 281–300. paper, University of Bern and Hochschule St. Gallen.

Lari-Lavassani, A., M. Simchi, and A. Ware. 2001. “A Discrete Wiggins, J. B. 1987. “Option Values under Stochastic Vola-
Valuation of Swing Options.” Canadian Applied Mathematics tility: Theory and Empirical Estimates.” Journal of Financial
Quarterly 9 (1): 35–74. Economics 19 (2): 351–372.

Mao, X. Stochastic Differential Equations and Applications. Wilmott, P. Derivatives: The Theory and Practice of Financial
Chichester: Horwood Publishing, 1997. Engineering. Chichester: John Wiley & Sons. 1998.

Marín Sánchez, F. 2010. “Binomial Trees for Option Valua- Wu, X. Y., C. Q. Ma, and S. Y. Wang. 2012. “Warrant
tion Process Derived from Stochastic Autonomous Differential Pricing under GARCH Diffusion Model.” Economic Modelling
Equation.” Ingeniería y Ciencia 6 (12): 145–170. 29 (6): 2237–2244.

Fall 2019 The Journal of Derivatives   65

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.
 Wu, X., W. Yang, C. Ma, and X. Zhao. 2014. “American Option used, including direct integration with an approach such as Gaussian
Pricing under GARCH Diffusion Model: An ­E mpirical quadrature, the fast Fourier transform (FFT), or fractional FFT. In
Study.” Journal of Systems Science and Complexity 27 (1): this article, Kilin reviews these methods and proposes a computational
193–207. enhancement to the basic Gaussian quadrature approach that speeds it
up substantially by storing intermediate results and reusing the saved
values rather than recomputing them each time. In the direct comparison
To order reprints of this article, please contact David Rowe at reported here, the modified quadrature approach was around 30 times
[email protected] or 646-891-2157. faster than FFT and 15 times faster than fractional FFT.

A Stochastic-Volatility Model for Pricing Power

Variants of Exchange Options
ADDITIONAL READING Weixuan X ia
The Journal of Derivatives
https://jod.pm-research.com/content/26/4/113
Accelerating the Calibration of Stochastic
­Volatility Models ABSTRACT: In this article, the author presents a model with
Fiodar K ilin jumps and stochastic volatility, based on two correlated variance-
The Journal of Derivatives gamma processes combined with an Ornstein–Uhlenbeck process with
https://jod.pm-research.com/content/18/3/7 gamma innovations. The objective is to analyze pricing methods for
a European-style equity option to exchange one stock for another,
ABSTRACT: Option valuation models keep advancing toward
as well as two important classes of its variants, which raise the stock
greater versatility and realism, by introducing more-flexible stochastic prices and the standard option payoff, respectively, to certain powers.
processes and more-complex payoff structures. Allowing stochastically These option variants are particularly useful in adjusting the risk
time-varying volatility and jumps in the returns process, the volatility level of exchange options and can also be viewed as generalizations
process, or both greatly broadens the range of asset price behavior that of traditional power-type options. The pricing formulas are obtained
can be incorporated in the pricing model. One of the key mathematical under risk neutrality in terms of characteristic functions and are thus
tools for solving more-complicated models is the use of characteristic func- independent from the model distribution. Numerical results are given
tions and related transform methods. But these tools require computing for illustrating the efficiency of the presented formulas along with
integrals in the complex domain, which is complicated and has some various advantages of the proposed stochastic-volatility model.
potential pitfalls. Several different numerical solution techniques are

66   Quadrinomial Trees to Value Options in Stochastic Volatility Models Fall 2019

Downloaded from https://pm-research.com/content/iijderiv/27/1, by guest on June 9, 2024. Copyright 2019 With Intelligence LLC.
It is illegal to make unauthorized copies, forward to an unauthorized user, post electronically, or store on shared cloud or hard drive without Publisher permission.