International Journal of Sciences:

Basic and Applied Research

(IJSBAR)
ISSN 2307-4531
(Print & Online)

[Link]
---------------------------------------------------------------------------------------------------------------------------

Volatility Extraction in Information Based Asset Pricing

Framework Via Non-Linear Filtering

Cynthia Ikamaria∗, Philip Ngareb, Patrick Wekec

a
Multimedia University of Kenya, P.O Box 15653-00503, Nairobi, Kenya
b,c
University of Nairobi, P.O Box 30197-00100, Nairobi, Kenya
a
Email: cikamari@[Link]
b
Email: pngare@[Link]
c
Email: pweke@[Link]

Abstract

This study looks at the derivation of a state space model that is applied in non-linear filtering. The model is
based on the Brody, Hughson and Macrina information based asset pricing model, also known as the BHM
approach or BHM model. The objective of this study is to extend the application of a filtering approach used in
estimation of volatilities for the Heston model to the BHM model. The measurement and transition equations
obtained in the state space model are used in the extended kalman filter to extract volatility. The option price is
obtained from the BS-BHM Updated Model by incorporating information in the Black-Scholes Model. This
option price is used to obtain the measurement equation while the variance process is used as the transition
equation.

Keywords: Kalman filter; Extended Kalman Filter; Measurement Equation; Transition Equation; State Space
Model.

1. Introduction

The author in [7] proposed a stochastic volatility model (SVM) referred to as the Heston model where the asset
returns’ volatility are driven by a stochastic variance process.

------------------------------------------------------------------------
* Corresponding author.

147
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

Filtering has been applied in estimating the volatility from an underlying asset price. In this study, in order to
apply filtering, the state space models which consist of a measurement and state transition equation are used. A
similar approach to that proposed by the authors in [16] where filtering is applied to the Heston Model is
followed. The approach is extended to the case of information based asset pricing in [4]. Filtering is an
engineering terminology for extracting information about a signal from partial and noisy observations. Filtering
can be used to estimate a dynamic system’s internal states given that the system has a series of current and past
noisy observations. The observation variables are observable unlike the system’s states which are unobservable.
The system’s states conditional probability distribution can then be estimated using the filtering approach. The
rationale underlying filtering is to obtain the optimal estimate of a hidden state given all the available
information up to that point. The estimation is performed in two steps: The first step involves determining the
prior distribution of the hidden state, denoted by the vector 𝑥𝑥𝑘𝑘 using all the given past information. That is, the
prediction of 𝑥𝑥𝑘𝑘 at time k is given by 𝑥𝑥�𝑘𝑘|𝑘𝑘−1 . An assumption is made that the previous estimates 𝑥𝑥�𝑘𝑘−1 are
known, they are used to predict the state vector 𝑥𝑥�𝑘𝑘|𝑘𝑘−1 . The second step involves using Bayes rule to obtain the
posterior distribution by combining the prior distribution obtained in the first step with the conditional
likelihood of the newest observation. This implies that the predicted states 𝑥𝑥�𝑘𝑘|𝑘𝑘−1 are combined with the current
observations 𝑦𝑦𝑘𝑘 to estimate the current states 𝑥𝑥�𝑘𝑘|𝑘𝑘 . Given that the observations are noisy, the interest is to get the
best estimate 𝑥𝑥�𝑘𝑘|𝑘𝑘 of 𝑥𝑥𝑘𝑘 that minimizes the error, 𝑥𝑥𝑘𝑘 − 𝑥𝑥�𝑘𝑘|𝑘𝑘 . This is done by recursion at each time step k. A
Kalman filter is an example of an optimal filtering method which is applicable in the field of science,
engineering and finance. In finance it can be used in hedging under partial observation, volatility estimation,
optimal asset allocation, etc. It is considered easy to understand with little computational burdens. The Kalman
filter is also ideal when a large volume of information must be taken into account because it is very fast. There
are two basic building blocks of a Kalman Filter, the measurement equation and the transition equation. The
measurement equation relates an unobserved variable to an observable variable. The transition equation is based
on a model that allows the unobserved variable to change through time. The method requires first of all that the
model is expressed on a state space form. A state space model is characterized by a measurement equation and a
transition equation. The Kalman filter is however only limited to linear models with Gaussian noises. Some
non-linear filtering methods that are applicable to non-linear systems include the extended Kalman filter and the
unscented Kalman filter. Particle filters can be applied to non-linear models with non-Gaussian noises. The
main objective of this study is to use filtering to extract volatility from the information based asset pricing
model as proposed in [4], referred to as the BHM model. Since the system of equations in the model are non-
linear and gaussian, the extended kalman filter will be used. This study extends the work by [16] who applied
filtering to the Heston Model and [5] who applied filtering to the Double Heston Model. A discrete dynamical
system is considered with unobservable state vector 𝑥𝑥𝑘𝑘 , for 𝑘𝑘 = 1, 2, … ., where 𝑘𝑘 denotes time

𝑥𝑥𝑘𝑘 = 𝑓𝑓𝑘𝑘 (𝑥𝑥𝑘𝑘−1 , 𝑤𝑤𝑘𝑘 ) (1.1)

and 𝑓𝑓𝑘𝑘 is a possibly non-linear and time-dependent function that represents the evolution of the state

process 𝑥𝑥𝑘𝑘 . The state process is driven by noise denoted by 𝑤𝑤𝑘𝑘 .

Suppose that an observable vector 𝑦𝑦𝑘𝑘 at time k is also given such that:

148
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝑦𝑦𝑘𝑘 = ℎ𝑘𝑘 (𝑥𝑥𝑘𝑘 , 𝑣𝑣𝑘𝑘 ) (1.2)

where ℎ𝑘𝑘 is a possibly non-linear and time-dependent function that defines the measurement 𝑦𝑦𝑘𝑘 . The
observations noise is denoted by 𝑣𝑣𝑘𝑘 . The state process in equation 1.1 is called the state transition equation and
the observation process in equation 1.2 is called measurement equation. The study begins by looking at filtering
techniques starting with the kalman filter and extending it to the extended kalman filter. The Heston and BHM
SVMs are then looked at and the state space representations derived. These state space models are used to
obtain the measurement and transition equation used in the extended kalman filter to extract volatility from the
BHM model.

2. Filtering Techniques

Kalman Filter

Kalman filter is only optimal for linear systems.

Given that the state function 𝑓𝑓𝑘𝑘 from Equation 1.1 and the measurement function ℎ𝑘𝑘 from the Equation 1.2 are
linear and their corresponding noise 𝑤𝑤𝑘𝑘 and 𝑣𝑣𝑘𝑘 respectively, are normally distributed and additive. Equation 1.1
can be expressed as

𝑥𝑥𝑘𝑘 = 𝑀𝑀𝑘𝑘 𝑥𝑥𝑘𝑘−1 + 𝑤𝑤𝑘𝑘 (2.1)

and Equation 1.2 becomes

𝑦𝑦𝑘𝑘 = 𝐻𝐻𝑘𝑘 𝑥𝑥𝑘𝑘 + 𝑣𝑣𝑘𝑘 (2.2)

The matrix 𝑀𝑀𝑘𝑘 is assumed to be known, it defines the state transition evolution, and the matrix 𝐻𝐻𝑘𝑘 defines the
measurement process which is also assumed to be known. An assumption is made that the state noise
𝑤𝑤𝑘𝑘 ~𝑁𝑁(0, 𝑄𝑄𝑘𝑘 ) and the measurement noise 𝑣𝑣𝑘𝑘 ~𝑁𝑁(0, 𝑅𝑅𝑘𝑘 ) are uncorrelated Gaussian random variables. In
addition, 𝑤𝑤𝑘𝑘 , 𝑣𝑣𝑘𝑘 are independent of 𝑥𝑥𝑘𝑘 , 𝑦𝑦𝑘𝑘 respectively. By substituting 𝑥𝑥𝑘𝑘 and 𝑦𝑦𝑘𝑘 from Equations 2.1 and 2.2 in
Equations 1.3 and 1.5, the computations result in the Kalman filtering algorithm with distributions given as:

𝑝𝑝(𝑥𝑥𝑘𝑘 |𝑥𝑥𝑘𝑘−1 )~𝑁𝑁(𝑀𝑀𝑘𝑘 𝑥𝑥𝑘𝑘−1 , 𝑄𝑄𝑘𝑘 ),

𝑝𝑝(𝑦𝑦𝑘𝑘 |𝑥𝑥𝑘𝑘 )~𝑁𝑁(𝐻𝐻𝑘𝑘 𝑥𝑥𝑘𝑘 , 𝑅𝑅𝑘𝑘 ).

The objective is to find an estimate of the state vector 𝑥𝑥𝑘𝑘 given the observations 𝑦𝑦𝑘𝑘 . An estimate of the state
vector 𝑥𝑥𝑘𝑘 is obtained from the past estimated states, 𝑥𝑥�𝑘𝑘−1 in the prediction step. We have that

𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 = 𝑀𝑀𝑘𝑘 𝑥𝑥�𝑘𝑘−1

𝑘𝑘| 𝑘𝑘 − 1 represents the estimated state of the state 𝑥𝑥𝑘𝑘 using previous (𝑘𝑘 − 1) estimated states.

149
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝑘𝑘| 𝑘𝑘 represents the estimates of 𝑥𝑥𝑘𝑘 using estimated states at 𝑘𝑘| 𝑘𝑘 − 1.

The computation of the past states is conducted by making use of the expectation of 𝑥𝑥𝑘𝑘 given in Equation 2.1.

The error in estimation error is obtained from

𝑒𝑒𝑘𝑘− = 𝑥𝑥𝑘𝑘 − 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 (2.4)

and the estimate error covariance

𝑃𝑃𝑘𝑘− = [𝑒𝑒𝑘𝑘− 𝑒𝑒𝑘𝑘−𝑇𝑇 ] (2.5)

The prediction of the observations is computed from

𝑦𝑦�𝑘𝑘 = 𝐻𝐻𝑘𝑘 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 (2.7)

The estimate of 𝑥𝑥�𝑘𝑘| 𝑘𝑘 is obtained from 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 and a measurement residual weighted by Kalman gain 𝐾𝐾𝑘𝑘 in the
update step as

𝑥𝑥�𝑘𝑘| 𝑘𝑘 = 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 + 𝐾𝐾𝑘𝑘 (𝑦𝑦𝑘𝑘 − 𝑦𝑦�𝑘𝑘 )

The measurement residual is computed from 𝑏𝑏𝑘𝑘 = 𝑦𝑦𝑘𝑘 − 𝑦𝑦�𝑘𝑘 . The estimate error is

𝑒𝑒𝑘𝑘 = 𝑥𝑥𝑘𝑘 − 𝑥𝑥�𝑘𝑘| 𝑘𝑘

and the estimate error covariance

𝑃𝑃𝑘𝑘 = [𝑒𝑒𝑘𝑘 𝑒𝑒𝑘𝑘𝑇𝑇 ]

Extended Kalman Filter

This is an extension of the Kalman filter and is used where the dynamical systems are non-linear. Consider a
case where the state transition function 𝑓𝑓𝑘𝑘 and the observation function ℎ𝑘𝑘 given in Equations Equations 1.1 and
1.2 respectively are both non-linear and their corresponding noises are uncorrelated Gaussian random variables
with 𝑤𝑤𝑘𝑘 ~𝑁𝑁(0, 𝑄𝑄𝑘𝑘 ) and 𝑣𝑣𝑘𝑘 ~𝑁𝑁(0, 𝑅𝑅𝑘𝑘 ). Given that the densities in 1.3 and 1.5 are normally distributed, then the
extended Kalman filter can be applied to obtain an estimate of the state v vector 𝑥𝑥𝑘𝑘 given the observations 𝑦𝑦𝑘𝑘 at
time step k. In the extended Kalman filter algorithm, the states in the first step are predicted as:

𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 = 𝑓𝑓𝑘𝑘 (𝑥𝑥�𝑘𝑘−1 , 0) (2.12)

The non-linear functions in the state transition and measurement equations are linearized to obtain the
covariance using Jacobian matrices:

150
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝜕𝜕𝑓𝑓𝑖𝑖 (𝑥𝑥�𝑘𝑘−1 , 0) 𝜕𝜕𝑓𝑓𝑖𝑖 (𝑥𝑥�𝑘𝑘−1 , 0)

𝐴𝐴𝑖𝑖𝑖𝑖 = , 𝑊𝑊𝑖𝑖𝑖𝑖 =
𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗

𝜕𝜕ℎ𝑖𝑖 �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� 𝜕𝜕𝑓𝑓𝑖𝑖 �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0�

𝐻𝐻𝑖𝑖𝑖𝑖 = , 𝑈𝑈𝑖𝑖𝑖𝑖 =
𝜕𝜕𝑥𝑥𝑗𝑗 𝜕𝜕𝑣𝑣𝑗𝑗

The predicted state covariance is thus

𝑃𝑃𝑘𝑘− = 𝐴𝐴𝑘𝑘 𝑃𝑃𝑘𝑘−1 𝐴𝐴𝑇𝑇𝑘𝑘 + 𝑊𝑊𝑘𝑘 𝑄𝑄𝑘𝑘−1 𝑊𝑊𝑘𝑘𝑇𝑇

Prediction of the measurement is given by

𝑦𝑦�𝑘𝑘 = ℎ𝑘𝑘 �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0�

with covariance

𝐹𝐹𝑘𝑘 = 𝐻𝐻𝑘𝑘 𝑃𝑃𝑘𝑘− 𝐻𝐻𝑘𝑘𝑇𝑇 + 𝑈𝑈𝑘𝑘 𝑅𝑅𝑘𝑘 𝑈𝑈𝑘𝑘𝑇𝑇

The state vector 𝑥𝑥𝑘𝑘 is estimated using the predicted states 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 in the update step. The measurement residual
is weighted by the Kalman gain 𝐾𝐾𝑘𝑘 ,

𝑥𝑥�𝑘𝑘| 𝑘𝑘 = 𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 + 𝐾𝐾𝑘𝑘 𝑏𝑏𝑘𝑘

where the measurement residual is 𝑏𝑏𝑘𝑘 = 𝑦𝑦𝑘𝑘 − 𝑦𝑦�𝑘𝑘 .

The optimal gain is

𝐾𝐾𝑘𝑘 = 𝑃𝑃𝑘𝑘− 𝐻𝐻𝑘𝑘𝑇𝑇 (𝐹𝐹𝑘𝑘 )−1

and the updated covariance

𝑃𝑃𝑘𝑘 = 𝑃𝑃𝑘𝑘− −𝐾𝐾𝑘𝑘 𝐻𝐻𝑘𝑘 𝑃𝑃𝑘𝑘−

3. Stochastic Volatility Models

After the October 1987 stock market crash, significant variations from normality have shown up in the term
structure of volatility. Various academicians and traders have taken a keen interest on this observation and as a
result, a lot of work has been done on this area. The danger of models used for pricing based on an incorrect
assumption of log-normality is the risk of obtaining biased prices.

The Black-Scholes model which has been used extensively in the past is considered to be successful in asset
pricing both in terms of approach and applicability. With the assumption of geometric brownian motion, the risk
neutral density for the underlying assets is taken to be lognormal. Asset prices are often observed to have

151
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

random volatility. These observations cannot be accurately be assumed to have a lognormal density since the
density functions are fat-tailed and skewed. Stochastic Volatility Models (SVMs) can be used to address this
weakness of the Black-Scholes Model. SVMs are models in which the variance of a stochastic process is itself
randomly distributed. They are widely used in the finance industry for derivative pricing and hedging. The
Heston model is an example of a SVM. It makes the assumption of stochastic volatility in the pricing of
European call option and obtains a closed-form solution. The model further assumes that the volatility and the
underlying asset price are correlated. In so doing, the Heston model is able to capture various properties of the
financial information which the Black-Scholes model doesn’t. However, the reliability of the model is
questionable because the assumptions on the volatility and the underlying asset price dynamics display an ad-
hoc nature. Another approach suggested later by Brody, Hughson and Macrina(BHM) obtains the asset price
dynamics using a more realistic approach towards the structure of the market unlike the Heston model where
the dynamics of the volatility and price are pre-specified. The approach is based on the assumption that the
market information is incomplete. It specifies a model for the structure of the information available in the
market since asset prices are determined by expectations on the future cash flows given the market information
available. The BHM approach doesn’t assume any dynamic model for the asset prices, it’s observed that the
asset price dynamics derived with the assumed information structure naturally has stochastic volatility giving a
different view of the volatility nature. Accordingly, the model illustrates that the volatility of volatility is
stochastic. The study shows that the Black Scholes asset pricing model can be obtained as a special case of the
BHM model following the approach by [1]. A Black Scholes asset pricing model from an information-based
perspective has been developed by [1], this is known as the BS-BHM model. A different model is obtained by
[2] by applying Gaussian integral on the BS-BHM model, this is referred to as the BS-BHM-Updated model. It
is based on the BlackScholes model from information-based perspective by Brody Hughston Macrina that it is
updated in the results of Gaussian Integrals, more specifically on the analysis of algebra trick of completing
square. Here, the Heston model and the BHM model are looked at in detail. Their state space representations are
also presented which are then used in the filtering to estimate the volatilities.

The Heston Model

In this section, we first present the dynamic system for the Heston model under a risk-neutral measure ℚ. The
model in [7] assumes that an underlying stock price, 𝑆𝑆𝑡𝑡 has a stochastic variance, 𝑉𝑉𝑡𝑡 , that follows a CIR process.
This process is represented by the following dynamical system:

𝑑𝑑𝑆𝑆𝑡𝑡 = (𝑟𝑟 − 𝑞𝑞)𝑆𝑆𝑡𝑡 𝑑𝑑𝑑𝑑 + �𝑉𝑉𝑡𝑡 𝑆𝑆𝑡𝑡 𝑑𝑑𝑊𝑊𝑡𝑡 (3.1)

𝑑𝑑𝑉𝑉𝑡𝑡 = 𝜅𝜅(𝜃𝜃 − 𝑉𝑉𝑡𝑡 )𝑑𝑑𝑑𝑑 + 𝜎𝜎�𝑉𝑉𝑡𝑡 𝑑𝑑𝑍𝑍𝑡𝑡 (3.2)

where r is a constant risk-free interest rate and q is a constant dividend. All the parameters κ, θ and σ are
positive constant. The terms 𝑊𝑊𝑡𝑡 and 𝑍𝑍𝑡𝑡 are Wiener processes that must be correlated with each other, that is;

(𝑑𝑑𝑊𝑊𝑡𝑡 𝑑𝑑𝑍𝑍𝑡𝑡 ) = 𝜌𝜌𝜌𝜌𝜌𝜌 (3.3)

where ρ is the correlation coefficient between the return of the underlying asset and the changes in the variance.

152
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

The correlation, which is often negative, will ensure that the volatility for example will rise if the underlying
asset value falls dramatically. In addition the variance is also mean-reverting, which is also evident in the
market. The mean-reverting process is the term 𝜅𝜅(𝜃𝜃 − 𝑣𝑣). For option valuation, the approach by [9] is followed,
such that the characteristic function of log returns 𝑥𝑥𝑘𝑘 = 𝑙𝑙𝑙𝑙(𝑆𝑆𝑘𝑘 ⁄𝑆𝑆𝑘𝑘−1 ) (𝑓𝑓𝑓𝑓𝑓𝑓 𝑘𝑘 ≤ 𝑡𝑡) of the Heston model is
derived using the so called the little Heston trap. This characteristic function is only slightly different from the
original formulation of [7], but it provides a better computation of the numerical integration. The European call
option price under the Heston model in the one dimensional framework is given by;

𝐶𝐶(𝑆𝑆, 𝑉𝑉, 𝐾𝐾, 𝜏𝜏) = 𝑆𝑆𝑘𝑘 𝑒𝑒 −𝑞𝑞𝑞𝑞 𝑃𝑃1 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 𝑃𝑃2 (3.4)

Where 𝑃𝑃𝑗𝑗 (𝑗𝑗 = 1,2) are the risk-adjusted probabilities of the log of the underlying price 𝑥𝑥𝑡𝑡 = ln(𝑆𝑆𝑡𝑡 /(𝑆𝑆𝑡𝑡−1 ). K
denotes the strike price.

1 1 ∞ 𝑒𝑒 −𝑖𝑖∅ ln 𝐾𝐾 𝑓𝑓𝑗𝑗 (∅; 𝑥𝑥𝑘𝑘 , 𝑉𝑉𝑘𝑘 )

𝑃𝑃𝑗𝑗 = + � 𝑅𝑅𝑅𝑅 � � 𝑑𝑑∅ (3.5)
2 𝜋𝜋 0 𝑖𝑖∅

for 𝑗𝑗 = 1, 2.

The characteristic functions 𝑓𝑓𝑗𝑗 (∅; 𝑥𝑥𝑘𝑘 , 𝑉𝑉𝑘𝑘 ) in the probabilities are given by

𝑓𝑓𝑗𝑗 (∅; 𝑥𝑥𝑘𝑘 , 𝑉𝑉𝑘𝑘 ) = 𝑒𝑒 𝑖𝑖∅ ln 𝐾𝐾+𝐴𝐴𝑗𝑗 (∅,𝜏𝜏)+𝐵𝐵𝑗𝑗 (∅,𝜏𝜏)𝑉𝑉𝑘𝑘 (3.6)

where

𝑏𝑏𝑗𝑗 − 𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 + 𝑑𝑑𝑗𝑗 1 − 𝑒𝑒 𝑑𝑑𝑗𝑗 𝜏𝜏

𝐵𝐵𝑗𝑗 (∅, 𝜏𝜏) = � �,
𝜎𝜎 2 1 − 𝑔𝑔𝑗𝑗 𝑒𝑒 𝑑𝑑𝑗𝑗 𝜏𝜏

𝑎𝑎 1 − 𝑒𝑒 𝑑𝑑𝑗𝑗 𝜏𝜏
𝐴𝐴𝑗𝑗 (∅, 𝜏𝜏) = 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 + ��𝑏𝑏 − 𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 + 𝑑𝑑 �𝜏𝜏 − 2 𝑙𝑙𝑙𝑙 � ��,
𝜎𝜎 2 𝑗𝑗 𝑗𝑗
1 − 𝑔𝑔𝑗𝑗 𝑒𝑒 𝑑𝑑𝑗𝑗 𝜏𝜏

𝑏𝑏𝑗𝑗 − 𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 + 𝑑𝑑𝑗𝑗

𝑔𝑔𝑗𝑗 = ,
𝑏𝑏𝑗𝑗 − 𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 + 𝑑𝑑𝑗𝑗

2
𝑑𝑑𝑗𝑗 = ��𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 − 𝑏𝑏𝑗𝑗 � − 𝜎𝜎 2 �2𝑢𝑢𝑗𝑗 𝜙𝜙𝜙𝜙 − 𝜙𝜙 2 �

1 1
And 𝑖𝑖 = √−1 , 𝜏𝜏 = 𝑇𝑇 − 𝑘𝑘, 𝑢𝑢1 = , 𝑢𝑢2 = − , 𝑎𝑎 = 𝜅𝜅𝜅𝜅, 𝑏𝑏1 = 𝜅𝜅 − 𝜌𝜌𝜌𝜌, 𝑏𝑏2 = 𝜅𝜅 and 𝜙𝜙 is called the integration
2 2

variable or node.

4. The BHM Model

The BHM Model as presented in [4] views asset price movements as an emergent phenomenon. The model’s

153
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

basis is pricing of assets by modelling the flow of market information. The market information in this case
relates to the given assets expected future cash flows. It is different from the other models used in pricing assets
mainly because the stochastic process governing the underlying asset price dynamics is deduced rather than
being imposed in an arbitrary way. The asset price dynamics from the BHM model are used to determine the
asset price dynamics in the BS-BHM Updated model. The BS-BHM-Updated model as presented by [2] uses
Gaussian integrals to determine the equation of the asset pricing model. The result obtained is different from
that of BS-BHM model due to an imprecision made in the Gaussian Integrals by BHM. Consider a single cash
flow occurring at time T, represented by a random variable, 𝑋𝑋𝑇𝑇 . The value of this variable will be revealed at
time T. The flow of market information available to market participants is assumed to be contained in a process
{𝜉𝜉𝑡𝑡 }0≤𝑡𝑡≤𝑇𝑇 given by:

𝜉𝜉𝑡𝑡 = 𝜎𝜎𝜎𝜎𝜎𝜎𝑇𝑇 + 𝛽𝛽𝑡𝑡𝑡𝑡 (3.11)

𝜉𝜉𝑡𝑡 denotes the market information process, its also known as the information process. 𝜎𝜎𝜎𝜎𝜎𝜎𝑇𝑇 contains the ‘true
information’ about the value of the cash flow 𝑋𝑋𝑇𝑇 . 𝜎𝜎 denotes the rate at which the true value of 𝑋𝑋𝑇𝑇 is revealed to
the market participants. {𝛽𝛽𝑡𝑡𝑡𝑡 }0≤𝑡𝑡≤𝑇𝑇 denotes a standard Brownian bridge over the interval [0, 𝑇𝑇] with 𝛽𝛽0𝑇𝑇 = 0
and 𝛽𝛽𝑇𝑇𝑇𝑇 = 0.

𝑡𝑡(𝑇𝑇 − 𝑡𝑡)
𝛽𝛽𝑡𝑡𝑡𝑡 ~ 𝑁𝑁 �0, �
𝑇𝑇

𝑋𝑋𝑇𝑇 and 𝛽𝛽𝑡𝑡𝑡𝑡 are assumed to be independent in the information-based framework. From the market’s point of
view, it is the process 𝑊𝑊𝑡𝑡 that drives the asset price dynamics.

According to [4], the dynamics of the price process are given as:

𝑑𝑑𝑑𝑑𝑡𝑡 = 𝑟𝑟𝑡𝑡 𝑆𝑆𝑡𝑡 𝑑𝑑𝑑𝑑 + Γ𝑡𝑡𝑡𝑡 𝑑𝑑𝑑𝑑𝑡𝑡

In this study, an assumption will be made that 𝑟𝑟𝑡𝑡 is a constant which implies that 𝑟𝑟𝑡𝑡 = 𝑟𝑟. Thus;

𝑑𝑑𝑑𝑑𝑡𝑡 = 𝑟𝑟𝑆𝑆𝑡𝑡 𝑑𝑑𝑑𝑑 + Γ𝑡𝑡𝑡𝑡 𝑑𝑑𝑑𝑑𝑡𝑡 (3.12)

Where Γ𝑡𝑡𝑡𝑡 denotes the absolute volatility process:

𝜎𝜎𝜎𝜎
Γ𝑡𝑡𝑡𝑡 = 𝑃𝑃𝑡𝑡𝑡𝑡 𝑉𝑉 (3.13)
𝑇𝑇 − 𝑡𝑡 𝑡𝑡

𝑃𝑃𝑡𝑡𝑡𝑡 denotes the discount factor and 𝑟𝑟𝑡𝑡 denotes the short rate. By using Gaussian integrals, the equation of asset
pricing model is presented as in [2] in the BS-BHMUpdated model as follows:

1 𝜎𝜎 2 𝜏𝜏 𝜎𝜎𝜎𝜎𝜎𝜎√𝑇𝑇
𝑆𝑆𝑡𝑡 = 𝑆𝑆0 exp �𝑟𝑟𝑟𝑟 − 𝜈𝜈 2 𝑇𝑇 + 𝜉𝜉 � (3.14)
2 𝜎𝜎 2 𝜏𝜏+1 𝑡𝑡(𝜎𝜎 2 𝜏𝜏+1) 𝑡𝑡

154
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝑡𝑡𝑡𝑡
where 𝜏𝜏 = .
𝑇𝑇−𝑡𝑡

𝜈𝜈 is the asset price volatility parameter. 𝜎𝜎 and 𝑣𝑣 cannot be observed directly and are assumed to be constants.
The equation above gives a different final result as compared to that of BHM’s Black-Scholes Model from an
Information-Based Perspective;

1 1 𝜈𝜈√𝑇𝑇 𝜎𝜎𝜎𝜎𝜎𝜎√𝑇𝑇
𝑆𝑆𝑡𝑡 = 𝑃𝑃𝑡𝑡𝑡𝑡 𝑆𝑆0 exp �𝑟𝑟𝑟𝑟 − 𝜈𝜈 2 𝑇𝑇 + + 𝜉𝜉 � (3.15)
2 2 𝜎𝜎 2 𝜏𝜏+1 𝑡𝑡(𝜎𝜎 2 𝜏𝜏+1) 𝑡𝑡

The difference arises due to an imprecision done by BHM using Gaussian Integrals. The authors in [3] show
that the BS-BHM Updated model in equation 3.15 follows the lognormal distribution given as:

2
𝑆𝑆 1 𝜎𝜎 2 𝜏𝜏 𝜎𝜎𝜎𝜎𝜎𝜎√𝑇𝑇 𝑡𝑡(𝑇𝑇−𝑡𝑡)
𝑙𝑙𝑙𝑙𝑙𝑙 � 𝑡𝑡 � ∼ 𝑁𝑁 �𝑟𝑟𝑟𝑟 − 𝜈𝜈 2 𝑇𝑇, � � �𝜎𝜎 2 𝑡𝑡 2 + �� (3.16)
𝑆𝑆0 2 𝜎𝜎 2 𝜏𝜏+1 𝑡𝑡(𝜎𝜎 2 𝜏𝜏+1) 𝑇𝑇

2
1 𝜎𝜎 2 𝜏𝜏 𝜎𝜎𝜎𝜎𝜈𝜈√𝑇𝑇 𝑡𝑡(𝑇𝑇−𝑡𝑡)
Let 𝐴𝐴 = 𝑟𝑟𝑟𝑟 − 𝜈𝜈 2 𝑇𝑇 and 𝐵𝐵2 = � � �𝜎𝜎 2 𝑡𝑡 2 + �. This implies that
2 𝜎𝜎 2 𝜏𝜏+1 𝑡𝑡(𝜎𝜎 2 𝜏𝜏+1) 𝑇𝑇

𝑆𝑆𝑡𝑡
𝑙𝑙𝑙𝑙𝑙𝑙 � � ∼ 𝑁𝑁[𝐴𝐴, 𝐵𝐵2 ]
𝑆𝑆0

2
𝑆𝑆𝑡𝑡 ∼ 𝑁𝑁�𝑆𝑆0 𝑒𝑒 𝐴𝐴 , 𝑆𝑆0 2 𝑒𝑒 2𝐴𝐴 �𝑒𝑒 𝐵𝐵 − 1��

Thus:

𝑆𝑆𝑡𝑡 = 𝑆𝑆0 𝑒𝑒 𝐴𝐴+𝐵𝐵𝐵𝐵 (3.17)

where Z denotes a standard normal random variable. The European call option price under the BS-BHM-
Updated model at time t with expiration date T and strike price K takes the form

𝐶𝐶 = 𝐸𝐸ℚ [𝑚𝑚𝑚𝑚𝑚𝑚{𝑆𝑆𝑡𝑡 − 𝐾𝐾, 0}] (3.18)

where C denotes the European call price. Following a similar approach to [1], an assumption is made that there
exists an established pricing kernel and the absence of arbitrage. These two assumptions ensure the existence of
a unique risk neutral probability measure ℚ. The BS-BHM-Updated model European call price formula takes a
similar form to the Black-Scholes Model given as:

𝐵𝐵2
−�𝐴𝐴+ �
𝐶𝐶 = 𝑆𝑆0 Φ(𝑑𝑑1 ) − 𝐾𝐾𝑒𝑒 2 Φ(𝑑𝑑2 ) (3.19)

𝑆𝑆
log� 𝐾𝐾0 �+𝐴𝐴
where 𝑑𝑑1 = + 𝐵𝐵, 𝑑𝑑2 = 𝑑𝑑1 − 𝐵𝐵 and Φ(x) = P[Z ≤ x], Z being a standard normal random variable.
𝐵𝐵

The author in [1] used to obtain the dynamics for the volatility in the BHM model. Let:

155
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

∞
𝑋𝑋𝑡𝑡𝑡𝑡 = 𝔼𝔼[𝑋𝑋𝑇𝑇 /𝜉𝜉𝑡𝑡 ] = ∫0 𝑥𝑥𝜋𝜋𝑡𝑡 (𝑥𝑥)𝑑𝑑𝑑𝑑 (3.20)

where 𝜋𝜋𝑡𝑡 (𝑥𝑥) denotes the conditional probability density for the random variable XT ;

𝑑𝑑
𝜋𝜋𝑡𝑡 (𝑥𝑥) = ℚ[𝑋𝑋𝑇𝑇 ≤ 𝑥𝑥/𝜉𝜉𝑡𝑡 ] (3.21)
𝑑𝑑𝑑𝑑

𝜎𝜎𝜎𝜎
𝑑𝑑𝑑𝑑𝑡𝑡 (𝑥𝑥) = (𝑥𝑥 − 𝑋𝑋𝑡𝑡𝑡𝑡 )𝜋𝜋𝑡𝑡 (𝑥𝑥)𝑑𝑑𝑑𝑑𝑡𝑡 (3.22)
𝑇𝑇−𝑡𝑡

The SDE for 𝐷𝐷𝑡𝑡𝑡𝑡 is given as

𝜎𝜎𝜎𝜎 1
𝑑𝑑𝑑𝑑𝑡𝑡𝑡𝑡 = 𝑉𝑉 � (𝜉𝜉𝑡𝑡 − 𝜎𝜎𝜎𝜎𝜎𝜎𝑡𝑡𝑡𝑡 )𝑑𝑑𝑑𝑑 + 𝑑𝑑𝜉𝜉𝑡𝑡 �. (3.23)
𝑇𝑇−𝑡𝑡 𝑡𝑡 𝑇𝑇−𝑡𝑡

This leads to:

𝜎𝜎𝜎𝜎
𝑑𝑑𝑑𝑑𝑡𝑡𝑡𝑡 = 𝑉𝑉 𝑑𝑑𝑑𝑑
𝑇𝑇 − 𝑡𝑡 𝑡𝑡 𝑡𝑡

𝑉𝑉𝑡𝑡 denotes the conditional variance of 𝑋𝑋𝑇𝑇 .

𝑉𝑉𝑡𝑡 = 𝔼𝔼[(𝑋𝑋𝑇𝑇 − 𝔼𝔼[𝑋𝑋𝑇𝑇 ])2 ] (3.24)

∞
2
𝑉𝑉𝑡𝑡 = � 𝑥𝑥 2 𝜋𝜋𝑡𝑡 (𝑥𝑥)𝑑𝑑𝑑𝑑 − 𝑋𝑋𝑡𝑡𝑡𝑡 (3.25)
0

Using Ito’s lemma and letting 𝜅𝜅𝑡𝑡 = 𝔼𝔼[(𝑋𝑋𝑇𝑇 − 𝔼𝔼[𝑋𝑋𝑇𝑇 ])3 ], it follows that:

𝜎𝜎𝜎𝜎 𝜎𝜎𝜎𝜎 2 2
𝑑𝑑𝑉𝑉𝑡𝑡 = � 𝜅𝜅𝑡𝑡 𝑑𝑑𝑑𝑑𝑡𝑡 � − � � 𝑉𝑉𝑡𝑡 𝑑𝑑𝑑𝑑
𝑇𝑇 − 𝑡𝑡 𝑇𝑇 − 𝑡𝑡

where deterministic nonnegative process {vt}0≤t≤T is defined by

𝑡𝑡
1
𝑣𝑣𝑡𝑡 = 𝜎𝜎𝑡𝑡 + � 𝜎𝜎𝑠𝑠 𝑑𝑑𝑑𝑑 , (3.26)
𝑇𝑇 − 𝑡𝑡 0

If σ is a constant, then:

𝜎𝜎𝜎𝜎
𝑣𝑣𝑡𝑡 = (3.27)
𝑇𝑇 − 𝑡𝑡

The dynamics for Vt are therefore given as:

𝑑𝑑𝑑𝑑𝑡𝑡 = −𝑣𝑣𝑡𝑡2 𝑉𝑉𝑡𝑡2 𝑑𝑑𝑑𝑑 + 𝑣𝑣𝑡𝑡 𝜅𝜅𝑡𝑡 𝑑𝑑𝑍𝑍𝑡𝑡 (3.28)

Using the approach in [1],the Black-Scholes asset-price model can be recovered from the BHM [Link]

156
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

authors in [2] presented the equation of the asset price model in the BS-BHM-Updated model as given in
equation 3.14, considering the case where σ2T = 1, the asset pricing model reduces to:

1
𝑆𝑆𝑡𝑡 = 𝑆𝑆0 𝑒𝑒𝑒𝑒𝑒𝑒 �𝑟𝑟𝑟𝑟 − 𝑣𝑣 2 𝑡𝑡 + 𝑣𝑣𝜉𝜉𝑡𝑡 � (3.29)
2

This takes a similar form to the Black-Scholes asset pricing model obtained by making use of the assumption
that the underlying share price follows a geometric Brownian motion in [6]. By applying Ito’s lemma to obtain
the SDE for log St and integration both sides from 0 to t it is found that:

1
𝑆𝑆𝑡𝑡 = 𝑆𝑆0 𝑒𝑒𝑒𝑒𝑒𝑒 �𝑟𝑟𝑟𝑟 − 𝜎𝜎 2 𝑡𝑡 + 𝜎𝜎𝐵𝐵𝑡𝑡 � (3.30)
2

Thus, the Black Scholes model can be obtained as a special case of the BHM model.

State Space Representation

The authors in [5,16] use a filtering approach to obtain an estimate for the volatilities in the case of the Heston
model and the Double Heston model respectively. In this study, this approach is extended to the BHM model.
𝑉𝑉𝑘𝑘 is the state variable for the BHM model which is unobserved. The option prices are taken to be the model
observations and the variance processes are taken to be the transition equations. Therefore, in order to estimate
the unobservable factors and the model’s parameters, the relationship between the option prices and the
underlying state variables is used. This is the relationship between the evolution of the measurement equations
and the state transition equations. A system of the measurement and transition equations is called the state space
representation of the model which is referred to as the state space model. The measurement noise and the state
noise are correlated in the heston model. Cholesky decomposition is used to decorrelate the sources of
randomness so as to ensure that for the filters, the process noise and measurement noise are uncorrelated. In
order to formulate the models in the state space representation, the state transition equations and the
measurement equations need to be specified. The state space model for the Heston model is presented first, then
it’s extended to the BHM model. The authors in [16] shows that if the spot prices Sk and option pricesC(Sk,K)
are taken as the observations and the variance Vk as the state, then the measurement equations are represented by

1
𝑦𝑦𝑘𝑘 = ln 𝑆𝑆𝑘𝑘 = ln 𝑆𝑆𝑘𝑘−1 + �𝑟𝑟 − 𝑞𝑞 − 𝑉𝑉𝑘𝑘−1 � Δ𝑘𝑘 + �𝑉𝑉𝑘𝑘−1 √Δ𝑘𝑘𝑊𝑊𝑘𝑘−1 . (3.33)
2

𝑦𝑦𝑘𝑘0 = 𝑔𝑔(𝑆𝑆𝑘𝑘 , 𝑉𝑉𝑘𝑘 , Θ) + 𝜖𝜖𝑡𝑡0 (3.34)

where 𝑦𝑦𝑘𝑘0 is the observable option prices, with identical independent distributed measurement errors 𝜖𝜖𝑡𝑡0 →
𝑁𝑁(0, 𝜎𝜎02 ), independent of 𝑊𝑊𝑘𝑘 and 𝑍𝑍𝑘𝑘 , and 𝑔𝑔(. ) is the theoretical option price computed from the Heston model.
The state transition equations are given by the variance processes

157
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝑉𝑉 κθΔ𝑘𝑘 1 − κΔ𝑘𝑘 0 𝑉𝑉𝑘𝑘−Δ𝑘𝑘

� 𝑘𝑘 � = � �+� �� � + �𝜎𝜎�Δ𝑘𝑘𝑉𝑉𝑘𝑘−Δ𝑘𝑘 � 𝑍𝑍𝑘𝑘−1
𝑉𝑉𝑘𝑘−Δ𝑘𝑘 0 1 0 𝑉𝑉𝑘𝑘−2Δ𝑘𝑘 0

Extending this approach to the BHM Model where the system of stochastic equations are given by equation
3.12 and equation 3.34. An assumption is made that the brownian motions, 𝑊𝑊𝑡𝑡 and 𝑍𝑍𝑡𝑡 are uncorrelated. The
asset price volatility is given as Γ𝑡𝑡𝑡𝑡 = 𝑣𝑣𝑡𝑡 𝑃𝑃𝑡𝑡𝑡𝑡 𝑉𝑉𝑡𝑡 By discretizing equation 3.12, the following measurement
equation is obtained:

𝑆𝑆𝑘𝑘 = 𝑆𝑆𝑘𝑘−1 + 𝑟𝑟𝑆𝑆𝑘𝑘−1 Δ𝑘𝑘 + 𝑣𝑣𝑘𝑘−1 𝑃𝑃(𝑘𝑘−1)𝑇𝑇 𝑉𝑉𝑘𝑘−1 √Δ𝑘𝑘𝑊𝑊𝑘𝑘−1 (3.35)

The variance process is given by:

2 2
𝑉𝑉𝑘𝑘 = 𝑉𝑉𝑘𝑘−1 − 𝑣𝑣𝑘𝑘−1 𝑉𝑉𝑘𝑘−1 Δ𝑘𝑘 + 𝑣𝑣𝑘𝑘−1 𝜅𝜅𝑘𝑘−1 √Δ𝑘𝑘𝑍𝑍𝑘𝑘−1 (3.36)

Thus the state transition equations are given by the variance processes:

𝑉𝑉𝑘𝑘 2
0� �𝑉𝑉𝑘𝑘−1 � + �𝑣𝑣𝑘𝑘−1 𝜅𝜅𝑘𝑘−1 Δ𝑘𝑘 � 𝑍𝑍
� � = �1 − 𝑣𝑣𝑘𝑘−1 𝑉𝑉𝑘𝑘−1 Δ𝑘𝑘 𝑘𝑘−1
𝑉𝑉𝑘𝑘−1 1 0 𝑉𝑉𝑘𝑘−2 0

5. Filtering in Stochastic Volatility Models

In the previous sections, the extended Kalman filter as well as the stochastic volatility models have been
discussed. Here, a similar approach to [16] where the non-linear filtering methods are applied to the Heston
Model is used. The author in [5] also used a similar approach in applying non-linear filtering to the Double
Heston Model. In this study, this approach is extended to the BHM Model. In particular, the extended kalman
filter non-linear filtering method is used to extract volatility in the BHM model.

Filtering in the Heston Model

In the case of the Heston model, the initial mean is given by 𝑥𝑥�0 = 𝑉𝑉0 and the initial covariance 𝑃𝑃� =
𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑(𝜎𝜎 2 Δ𝑘𝑘𝑉𝑉0 , 0). 𝑉𝑉0 denotes the initial variance level. The parameters to be estimated are κ,𝜃𝜃,σ,V0,ρ. These
parameters are first estimated and the kalman filter is then used to estimate volatility. The Jacobian matrices in
the extended Kalman filter are obtained as follows:

𝐴𝐴𝑘𝑘 = 1 − 𝜅𝜅Δ𝑘𝑘. (4.1)

𝑊𝑊𝑘𝑘 = 𝜎𝜎�Δ𝑘𝑘𝑉𝑉𝑘𝑘−1 (4.2)

Thus, the Jacobian matrices under the Heston model are given by equation 4.1 and equation 4.2. The matrix 𝐻𝐻𝑘𝑘
can be viewed as the derivative of the call price with respect to the implied volatility. In the Heston model, the
shape of the surface of the implied volatility is determined by the parameters which drive the variance process.
In order to compute 𝐻𝐻𝑘𝑘 , the derivative is computed based on two parameters, 𝑣𝑣1 = �𝑉𝑉0 and 𝑣𝑣2 = √𝜃𝜃 which
gives the following Jacobian matrices:

158
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� 𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0�

𝐻𝐻1 = = 2�𝑉𝑉0 (4.3)
𝜕𝜕𝑣𝑣1 𝜕𝜕𝑉𝑉0

𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� 𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0�

𝐻𝐻2 = = 2√𝜃𝜃 (4.4)
𝜕𝜕𝑣𝑣2 𝜕𝜕𝜃𝜃

Substituting ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� for the value of a European call option under the Heston model, that is:

ℎ�𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� = 𝑆𝑆𝑒𝑒 −𝑞𝑞𝑞𝑞 𝑃𝑃1 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 𝑃𝑃2 (4.5)

This results in:

𝜕𝜕(𝑆𝑆𝑒𝑒 −𝑞𝑞𝑞𝑞 𝑃𝑃1 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 𝑃𝑃2 )

𝐻𝐻1 = 2�𝑉𝑉0
𝜕𝜕𝑉𝑉0

𝜕𝜕𝑃𝑃1 𝜕𝜕𝑃𝑃2
𝐻𝐻1 = 𝑆𝑆𝑒𝑒 −𝑞𝑞𝑞𝑞 2�𝑉𝑉0 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 2�𝑉𝑉0
𝜕𝜕𝑉𝑉0 𝜕𝜕𝑉𝑉0

where

𝜕𝜕𝑃𝑃𝑗𝑗 1 ∞ 𝑒𝑒 −𝑖𝑖∅ ln 𝐾𝐾 𝑓𝑓𝑗𝑗 (∅; 𝑆𝑆𝑘𝑘 , 𝑉𝑉𝑘𝑘 )𝐵𝐵𝑗𝑗 (𝜏𝜏, ∅)

= � 𝑅𝑅𝑅𝑅 � � 𝑑𝑑∅
𝜕𝜕𝑉𝑉0 𝜋𝜋 0 𝑖𝑖∅

For 𝑗𝑗 = 1,2

𝜕𝜕(𝑆𝑆𝑒𝑒 −𝑞𝑞𝑞𝑞 𝑃𝑃1 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 𝑃𝑃2 )

𝐻𝐻2 = 2√𝜃𝜃
𝜕𝜕𝜕𝜕

𝜕𝜕𝑃𝑃1 𝜕𝜕𝑃𝑃2
𝐻𝐻2 = 𝑆𝑆𝑒𝑒 −𝑞𝑞𝑞𝑞 2√𝜃𝜃 − 𝐾𝐾𝑒𝑒 −𝑟𝑟𝑟𝑟 2√𝜃𝜃
𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕

where

𝜕𝜕𝑃𝑃𝑗𝑗 1 ∞ 𝑒𝑒 −𝑖𝑖∅ ln 𝐾𝐾 𝑓𝑓𝑗𝑗 (∅; 𝑆𝑆𝑘𝑘 , 𝑉𝑉𝑘𝑘 )𝜕𝜕𝐴𝐴𝑗𝑗 (𝜏𝜏, ∅)/𝜕𝜕𝜕𝜕

= � 𝑅𝑅𝑅𝑅 � � 𝑑𝑑∅
𝜕𝜕𝜕𝜕 𝜋𝜋 0 𝑖𝑖∅

and

𝜕𝜕𝐴𝐴𝑗𝑗 (𝜏𝜏, ∅) 𝜅𝜅 1 − 𝑔𝑔𝑗𝑗 𝑒𝑒 𝑑𝑑𝑗𝑗 𝜏𝜏

= 2 ��𝑏𝑏𝑗𝑗 − 𝜌𝜌𝜌𝜌𝜌𝜌𝜌𝜌 + 𝑑𝑑𝑗𝑗 �𝜏𝜏 − 2𝑙𝑙𝑙𝑙 � ��
𝜕𝜕𝜕𝜕 𝜎𝜎 1 − 𝑔𝑔𝑗𝑗

Filtering in the BHM Model

In this case, the mean is initialized as 𝑥𝑥�0 = 𝑉𝑉0 and the initial covariance is 𝑃𝑃0− = 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑(𝑣𝑣02 𝜅𝜅0 ∆𝑘𝑘, 0). For the
BHM model, the parameters to be estimated are σ,v and 𝑉𝑉0 .. The parameters are first initialized and the kalman

159
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

filter is used to estimate volatility.

Using a similar approach to the Heston model, the Jacobian matrices in the extended Kalman filter for the BHM
model can be obtained as follows:

2
𝐴𝐴𝑘𝑘 = 1 − 2𝑣𝑣𝑘𝑘−1 𝑉𝑉𝑘𝑘−1 ∆𝑘𝑘 (4.13)
𝑊𝑊𝑘𝑘 = 𝑣𝑣𝑘𝑘−1 𝜅𝜅𝑘𝑘−1 √∆𝑘𝑘 (4.14)

Thus, the Jacobian matrices under the BHM model are given by equation 4.13 and 4.14.

Under the BHM model, the matrix Hk is computed based on a parameter V0 which gives:

𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� 𝜕𝜕ℎ �𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0�

𝐻𝐻1 = = (4.15
𝜕𝜕𝑥𝑥1 𝜕𝜕𝑉𝑉0

Substituting ℎ�𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� for the value of a European call option under the BS-BHM-Updated model, that is:

𝐵𝐵2
−�𝐴𝐴+ �
ℎ�𝑥𝑥�𝑘𝑘| 𝑘𝑘−1 , 0� = 𝑆𝑆0 Φ(𝑑𝑑1 ) − 𝐾𝐾𝑒𝑒 2 Φ(𝑑𝑑2 )

This results in:

𝐵𝐵2
−�𝐴𝐴+ �
𝜕𝜕 �𝑆𝑆0 Φ(𝑑𝑑1 ) − 𝐾𝐾𝑒𝑒 2 Φ(𝑑𝑑2 )�

𝐻𝐻1 = 2�𝑉𝑉0
𝜕𝜕𝑉𝑉0

𝐵𝐵2
−�𝐴𝐴+ �
𝜕𝜕Φ(𝑑𝑑1 ) 𝜕𝜕𝑒𝑒 2 Φ(𝑑𝑑 )
2
𝐻𝐻1 = 𝑆𝑆0 2�𝑉𝑉0 − 𝐾𝐾 2�𝑉𝑉0
𝜕𝜕𝑉𝑉0 𝜕𝜕𝑉𝑉0

𝐵𝐵2
−�𝐴𝐴+ �
𝜕𝜕 �𝑆𝑆0 Φ(𝑑𝑑1 ) − 𝐾𝐾𝑒𝑒 2 Φ(𝑑𝑑2 )�

𝐻𝐻2 = 2√𝜃𝜃
𝜕𝜕𝜕𝜕

6. Conclusion

In this study, the extended kalman filtering technique has been applied to extract volatility in the information
based asset pricing framework. Using the approach by the authors in [1], the dynamics for the volatility process
are obtained. The asset price dynamics in [3] are then used with the volatility process dynamics to obtain the
state space model. The price obtained from the BS-BHM-Updated Model is used as the measurement equation
and the variance process is used as the state transition [Link] state space model is then used to perform
filtering using the extended Kalman filtering technique since the system of equations in the model are non-
linear. The Brownian motion process is the same for the state transition equation and measurement equation

160
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

under the BHM model. The European Call Option price obtained using the BS-BHM-Updated Model will be
used to determine the option price. This option price is used to obtain the measurement equation.

7. Recommendation

Further studies can be done by using other non-linear filtering methods such as the unscented kalman filter to
extract volatility in the BHM Model. Methods such as maximum likelihood can also be used to estimate the
parameters used in the model. In addition, the case where the assumption that the interest rate is a constant can
be relaxed.

References

[1]. A., Macrina (2006). “An Information-Based Framework for Asset Pricing: X-Factor Theory and
itsApplications”, PhD Thesis, King’s College London.
[2]. Mutijah, Guritno and Gunardi (2012), “A Black Scholes Model from an Information-Based
Perspectiveby Brody Hughston Macrina”, International Conference on Statistics in Science, Business
and Engineering.
[3]. Mutijah, Guritno and Gunardi (2013), “Estimation of Parameters on the BS-BHM Updated
Model.”Journal Applied Mathematical Sciences.
[4]. D., Brody, L., Hughston and A. Macrina (2008). “Information-based asset pricing”. In: International
Journal of Theoretical and Applied finance 11, 107-142.
[5]. Elia Namundjebo (2016). “The Double Heston Model via Filtering Methods”. Msc. Thesis
FinancialMathematics in the Faculty of Science at Stellenbosch University.
[6]. Black, Fischer and Scholes, Myron (1973). ”The Pricing of Options and Corporate Liabilities”. In:
Journal of Political Economy 81 (3) 637–654.
[7]. Heston Steven (1993). “A Closed-Form Solution for Options with Stochastic Volatility with
Applications to Bond and Currency Options”. In The Review of Financial Studies 6(2) 327-343.
[8]. C., Cox,E., Ingersoll and A. Ross (1985). “A Theory of the Term Structure of Interest Rates”. In:
Journal of the Econometric Society 53 (2), 385-407.
[9]. H., Albrecher, P., Mayer, W., Schoutens and J., Tistaert (2006). “The little heston trap.”
[10]. Bakshi, Charles and Zhiwu (1997), “Empirical Performance of Alternative Option Pricing Models”. In:
The Journal of Finance, 5(2), 46-78.
[11]. C., Ball C and [Link] (2000). “Stochastic correlation across international stock markets”. In:
Journal of Empirical Finance 7, 373-388.
[12]. A. Doucet, N., De Freitas, K., Murphy, and S., Russell (2000). “Rao-blackwellised particle filtering
fordynamic bayesian networks”. In: Proceedings of the Sixteenth conference on Uncertainty in
artificial intelligence,” 176-183.
[13]. Gatheral, Jim (2006). “The Volatility Surface. A Practitioner’s Guide”. Wiley Finance
[14]. J. Hull. (2008), “Options, Futures and other Derivatives”. Prentice Hall (7), 91-102.
[15]. A., Javaheri. (2005). ”Inside Volatility Arbitrage: the secret of skewness”, Wiley and Sons
[16]. A., Javaheri, D., Lautier and A., Galli. (2003). ”Filtering in finance.” Wilmott, 3.

161
 International Journal of Sciences: Basic and Applied Research (IJSBAR) (2019) Volume 47, No 1, pp 147-162

[17]. S., Julier and J., Uhlmann (1996). “A general method for approximating nonlinear transformations
ofprobability distributions.” Tech. Rep., Technical Report, RRG.
[18]. Lewis, Alan (2005), “Option Valuation under Stochastic Volatility: With Mathematica Code”
FinancePress (2), Newport Beach, 22-57.
[19]. J., Li (2013). “An unscented kalman smoother for volatility extraction: Evidence from stock prices
andoptions.” In: Computational Statistics and Data Analysis, vol. 58, 15-26.
[20]. L., Lu, M., Yang, et al. (2015). “Study on importance function for particle filter.” International Journal
of Multimedia and Ubiquitous Engineering, 10(2), 249-258.
[21]. Rouah, F.D. (2013). “The Heston model and its extensions in MATLAB and C#.” John Wiley and
Sons.
[22]. R., Van Der Merwe, A., Doucet, N., De Freitas and E., Wan (2000). “The unscented particle filter.” In:
NIPS, 2000, 584-590.

162