Download as pdf or txt
Download as pdf or txt
You are on page 1of 29

Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Contents lists available at ScienceDirect

Commun Nonlinear Sci Numer Simulat


journal homepage: www.elsevier.com/locate/cnsns

Research paper

Analytically pricing volatility swaps and volatility options


with discrete sampling: Nonlinear payoff volatility derivatives
Sanae Rujivan a,∗, Udomsak Rakwongwan b
a
Center of Excellence in Data Science for Health Study, Division of Mathematics and Statistics, School of Science, Walailak University,
Nakhon Si Thammarat 80161, Thailand
b
Department of Mathematics, Faculty of Science, Kasetsart University, Bangkok 10900, Thailand

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents the first analytical pricing formulas for volatility swaps and volatility
Received 21 October 2020 options with discrete sampling under the Black-Scholes model with time varying risk-free
Revised 5 April 2021
interest rate. Despite numerous analytical works on the pricing of variance swaps with
Accepted 6 April 2021
discrete sampling under different models of asset prices, an analytical pricing formula for
Available online 16 April 2021
volatility swaps as well as volatility options had not been well addressed until now. The
MSC: main challenge in pricing volatility swaps and volatility options is that payoff functions
91G20 contain a square root operator, making their expectations nonlinear. By utilizing proper-
ties of noncentral chi random variables, we can compute expectations of payoff functions
Keywords: analytically and obtain formulas for pricing volatility swaps and volatility options, includ-
Volatility swaps ing variance swaps and variance options. Furthermore, we investigate the accuracy of the
Volatility options
well-known convexity correction formula. Most interestingly, we extend our results to the
Discrete sampling
Analytical pricing formula
Black-Scholes model with time varying parameters and to the Heston stochastic volatility
Approximate formula model in which the variance process is assumed to follow the extended Cox-Ingersoll-
Ross process by constructing simple closed-form approximate formulas for pricing volatil-
ity swaps and demonstrate the accuracy and efficiency of this approach by comparing the
approximated prices against those obtained with Monte Carlo simulations.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction

The pricing of volatility derivatives, based on the definitions of discretely-sampled log-return realized variance, can be
separated into two categories: pricing volatility swaps and variance swaps. However, despite many common characteristics
between these two types of volatility derivatives, the former is clearly viewed to be more difficult to price analytically
than the latter because the payoff function involves a square root operator. More specifically, analytically calculating the
expectation of the payoff function containing a square root of sum of squared log-returns of the underlying asset, which are
random variables, is a very difficult task due to the lack of the explicit form of the joint distribution. As a result, quite a
few analytical formulas have been discovered for the latter under various models of underlying asset prices whereas it is
very rare to see a paper discussing an analytical formula for the former including those papers working on the standard
Black-Scholes (B-S) model [2] including time varying parameters, particularly when the sampling is discretely conducted.


Corresponding author.
E-mail addresses: rsanae@wu.ac.th (S. Rujivan), udomsak.ra@ku.th (U. Rakwongwan).

https://doi.org/10.1016/j.cnsns.2021.105849
1007-5704/© 2021 Elsevier B.V. All rights reserved.
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

In addition to volatility derivatives, options on the realized volatility allure investors in recent years such that they can
use as investment tools to view the future market volatility by allowing them to deal with particular risk exposures without
taking direct positions in the underlying asset. Unlike the payoff of volatility swaps, the payoff of these options is defined
in terms of the positive part of the spread between a square root of the realized variance and volatility strike (quoted as
a volatility). Since the distribution of realized volatility is very difficult to find in explicit form, analytically calculating the
expectation of the payoff function is almost impossible. Consequently, none has proposed an analytical pricing formula for
the option pricing on the realized volatility, in particular with discrete samplings, based on any models of asset prices. From
a mathematical point of view, these pricing problems are actually nonlinear. Therefore, an analytical approach is needed in
order to obtain the exact values of the fair prices and delta-hedging instead of using those approximate values produced
from numerical approaches such as Monte Carlo (MC) simulations.
This paper aims to develop an analytical approach to solve the pricing problems by concentrating on the B-S model with
time-varying risk-free interest rate described as follows.
Under a risk-neutral martingale measure Q, we assume that the underlying asset price St evolves according to the fol-
lowing stochastic differential equation (SDE)

dSt = r (t )St dt + σ St dWt (1.1)


for 0 < t ≤ T and S0 > 0 where r (t ) is a time varying risk-free interest rate, σ > 0 is a price volatility, and (Wt )0≤t≤T is a
Brownian motion under the probability space (, F, Q ), and filtration (Ft )0≤t≤T generated by the price process (St )0≤t≤T .
In addition, we assume that r (t ) is continuous on [0, T ].
In our study, we first consider volatility swaps which are essentially forward contracts on the future realized variance
of the returns of the specified underlying asset. The payoff at expiry for the long position of a volatility swap is equal to
square root of the annualized realized variance over a pre-specified minus a pre-set delivery price of the contract multiplied
by a notional amount of the swap in dollars per annualized volatility, whereas
√ the short position is just the opposite. More
specifically, the value of a volatility swap at expiry can be written as ( RV − Kvol ) × L, where the RV is the annualized
realized variance over the contract life [0, T ] and Kvol is the annualized delivery price for the volatility swap, which is set
to make the value of a volatility swap equal to zero for both long and short positions at the time the contract is initially
entered. The life time of the contract is denoted by T and L is the notional amount of the swap in dollars per annualized
variance point.
This paper focuses on the measure of realized variance known as the log-return realized variance defined by
   
AF  2 1 2
N N
Sti Sti
RV ≡ RVd (t1 , N, T ) := ln 2
× 100 = ln × 1002 (1.2)
N−1 Sti−1 T Sti−1
i=2 i=2

for 0 < t1 < T where Sti is the closing price of the underlying asset at the i-th observation time ti , and there are alto-
gether N ≥ 2 observations. AF is the annualized factor converting this expression to an annualized variance. If the sampling
frequency is every trading day, then AF = 252, assuming that there are 252 trading days in one year, if every week then
AF = 52, if every month then AF = 12 and so on. Typically, we set T = N−1
AF
and assume equally-spaced discrete observations
so that the annualized factor is of a simple expression AF = N−1 1
T = t .
In the risk-neutral world, the value of a volatility swap at time t1 is the expected present value of the future payoff,

− T r (s )ds 
Vt1 = e t1 EtQ [( RVd − Kvol )] × L. This should be zero at the beginning of the contract since there is no cost to enter
1
into a swap. Therefore, the fair strike price of a volatility swap can be defined as

Kvol := EtQ1 [ RVd ], (1.3)

after setting the value of Vt1 = 0 initially where EtQ [X] denotes the conditional expectation of a random variable X with
1
respect to Ft1 under the risk-neutral martingale measure Q. Therefore, the volatility swap valuation problem is reduced to
calculating the expectation value of square root of the future realized variance in the risk-neutral world.
Due to the increasing popularity of volatility and variance swaps in the financial market, there have been numerous
works on valuation of such derivatives. The settings of the valuation of volatility and variance swaps in the literature can
be categorized into two types: continuous sampling and discrete sampling, depending on how the realized volatility and
variance are sampled. Since several comprehensive reviews of papers in the first category have been given in [3,30,31,38,40],
we shall mainly focus on a brief review of the literature in the second category which is generally much more difficult than
that of continuous sampling assumption. In addition, the continuous sampling assumption implies that the results obtained
from a continuous model can only be viewed as an approximation for the real cases in financial practice, in which all
contracts are written with the realized volatility as well as realized variance being evaluated on a set of discrete sampling
points. Another major drawback is that this strategy also requires options with a continuum of exercise prices, which is not
actually available in the marketplace.
In the literature on analytically pricing variance swaps with discrete sampling, the most influential pioneer work is of
Broadie and Jain [3]. They represented a solution for pricing discretely-sampled variance swaps in terms of a sum of the so-
lution obtained from the corresponding continuous model and a correction function obtained by using a direct integration
method. They also investigated the effects of jumps and stochastic volatility on the price of variance swaps by comparing

2
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

calculated prices under various models such as the standard B-S model [2], the Heston stochastic volatility model [13], the
Merton jump diffusion model [21], the Bates and Scott stochastic volatility and jump model [1,28]. Itkin and Carr [15] pro-
posed an asymptotic method to price variance swaps under general time-change LȨvy models.
Unlike Broadie and Jain’s approach [3], Zhu and Lian [38] presented a PDE-based approach to obtain a closed-form pricing
formula for variance swaps under the Heston stochastic volatility model [13] with the realized variance being defined as
the sum of the squared percentage increments of the underlying price. Zhu and Lian [39] also adopted the PDE-based
approach to derive a closed-form formula for variance swaps with the log-return realized variance. However, Zhu and Lian’s
approach [38,39] is still too complicated. Rujivan and Zhu [23], Rujivan and Zhu [24], and Rujivan [25] pointed out that
there is actually a simplified approach, and they applied this approach to the percentage return as well as generalized
log-return cases, respectively, demonstrating that the exact same results can be produced without using the generalized
Fourier transform. Chunhawiksit and Rujivan [5] and Weraprasertsakun and Rujivan [34] extended Rujivan and Zhu’s works
[23,24] to price variance swaps in commodity markets. Elliott and Lian [9] derived the fair strike prices for variance and
volatility swaps under the Heston model with regime switching by using the forward characteristic functions. Zheng and
Kwok [37] proposed another analytical approach, which relies on the availability of the analytical expression of the joint
moment generating function of the underlying processes, for pricing various types of discretely-sampled generalized variance
swaps, including the same realized variance as used by Zhu and Lian [39]. In addition, they extended Zhu and Lian’s work
[39] to price variance swaps under the stochastic volatility models with simultaneous jumps in the asset price and variance
processes. Yuen et al. [36] applied Zheng and Kwok’s work [37] to price various types of exotic discrete variance swaps
under the 3/2-stochastic volatility model with jumps in asset price. Lian et al. [18] derived the characteristic function of the
discretely sampled realized variance and utilized either semi-analytical pricing formulas or an efficient Fourier-cosine series
method to price discrete volatility derivatives.
The valuation of derivatives on the discrete realized variance under exponential Lévy dynamics based on Fourier-Laplace
methods was carried out by Keller and Muhle [17]. Drimus et al. [7] proposed approximate method for pricing derivatives on
the discretely monitored realized variance in stochastic volatility models by one-dimensional MC simulations of the instan-
taneous variance. He and Zhu [11] presented a series-form solution for pricing variance and volatility swaps with stochastic
volatility and stochastic interest rate. Liu and Zhu [19] adopted Rujivan and Zhu’s approach [23,24] to obtain an analytic
solution of variance swaps under the Hawkes jump-diffusion model and showed that the model is effective in capturing
market in crisis. He and Zhu [12] proposed analytical pricing formulas for variance and volatility swaps under a two-factor
stochastic volatility model with regime switching. Yang et al. [35] derived an analytical pricing formula for volatility swaps
when the underlying asset follows a stochastic volatility model with jumps and stochastic intensity. Xu et al. [32] presented
analytical pricing formulas for variance and volatility swaps in the framework of a liquidity-adjusted underlying assets model
with the stochastic liquidity risk. Very recently, Rujivan [27] presented an analytical formula for pricing discretely-sampled
variance swaps in commodity derivative markets under stochastic convenience yields. Nevertheless, the realize volatilities
defined in [11,12,19,32,35] are in the class of linear payoff volatility derivatives. 
Let us now focus on analytically pricing volatility swaps with the realized volatility being defined as RVd , which is
in the class of nonlinear payoff volatility derivatives. Broadie and Jain [3] also presented a solution approach for pricing
volatility swaps under the standard B-S model. However, their solution approach is based on computing the alternative
form of the square root function and the Laplace transform of realized volatility which prevents it from knowing the exact
distribution of the realized volatility, resulting an analytical pricing formula for volatility options as well as variance options
is almost impossible to obtain. Considering the numerous analytical approaches on pricing variance swaps with discretely
sampling as mentioned previously, to the best of our knowledge, no analytical formulas obtained from those analytical
approaches for pricing volatility swaps as well as volatility options based on the B-S model (1.1) have been presented in the
literature to date. 
Our solution approach presented in this paper reveals the exact distribution of the realized volatility RVd including
the exact distribution of the realized variance RVd under the B-S model (1.1). We apply some properties of a noncentral
chi random variable defined from the square root of sum of the squared log returns to obtain the first analytical formulas
for pricing volatility and variance swaps on discretely monitored volatility. We further investigate the accuracy of the well-
known convexity correction formula and prove that the price approximated by the convexity correction formula converges
to the actual price as a sample size increases. The major contribution of this paper is the exact distribution of the realized
volatility is analytically derived, yielding the exact pricing formulas for volatility and variance options. In addition, we extend
our results to approximate the fair discretely monitored volatility strikes for the B-S model with time varying parameters
[21] and for the Heston stochastic volatility model [13] in which the variance process is assumed to follow the extended
Cox-Ingersoll-Ross process. Most interestingly, our closed-form approximate formulas are in simple forms and ready for
practical applications. The accuracy and efficiency of the approximation approach are also demonstrated by comparing the
approximated prices with those obtained with MC simulations.
This paper is organized into six sections. In Section 2, we first derive analytical formulas for volatility and variance swaps
in discrete context, followed by the investigation of the accuracy of the convexity correction formula. The analytical pricing
formulas of options on the realized volatility and variance are given in Section 3. In Section 4, we describe the approaches,
extended from Section 2, to approximate fair volatility strikes under the B-S model with time varying parameters and under
the Heston model. The numerical results are illustrated in Section 5. This includes the investigation of the efficiency and ac-

3
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

curacy of the approximation approaches, detailed in Section 4, and the convexity correction formula, explained in Section 2.
Section 6 is the conclusions.

2. Pricing volatility and variance swaps

2.1. An analytical formula for pricing volatility swaps

Firstly, we set
 ti+1
1 2
μi = r (s )ds − σ t (2.1)
ti 2
for i = 1, . . . , N − 1, where t = ti+1 − ti > 0.

i=1 μi > 0
2.1.1. Volatility swap prices case: N−1 2

In the following theorem, we derive a pricing formula for volatility swaps based on the square root of the discretely-
sampled log-return realized variance (1.2) when the sum of μ2i , i = 1, . . . , N − 1, is strictly positive.
N−1
Theorem 2.1. Suppose that i=1 μ2i > 0. The fair strike price of a volatility swap in (1.3) can be expressed as
 2
√ π ( N−1 λ
2 −1 )
Kvol ≡ Kvol1 (t, N, T ; r (t ), σ ) = σ t L1 − × 100 (2.2)
2T 2 2
N−1
μ2i  (−1 )k (α +ν )!xk
where λ = and Lν(α ) (x ) = ∞
k=0 (ν −k )!(α +k )!k! is a Laguerre function with parameter α and fractional order ν . Moreover,
i=1

σ t
Kvol1 computed by using (2.2) is always positive for all t > 0.

Proof. See Appendix A. 

According to the Proof of Theorem 2.1 in Appendix A, (A.2) and (A.4) imply that the log-return realized volatility is a
 √ √
RVd = β Y where Y ∼ N C χN−1 (λ ) and β = σ Tt × 100.
linear transformation of a noncentral chi random variable, i.e.,
As a result, the first analytical pricing formulas for pricing discretely-sampled volatility and variance options as well as their
properties have been found and intensively studied in the following sections. This is the  major contribution of the paper

that should be pointed out here. The following corollary provides the exact distribution of RVd case: N−1 i=1 μi > 0.
2

Corollary 2.2. According to Theorem 2.1 and setting β = σ t × 100, we have the following results.
T

1. The variance of RVd can be found in explicit form as

VARtQ1 [ RVd ] = β 2 (N − 1 + λ2 ) − Kv2ol1 . (2.3)
2. For 0 ≤ a ≤ b ≤ ∞,
  b
β 1 y
Q (a ≤ RVd ≤ b) = f1,√Y ; N − 1, λ dy (2.4)
a
β
β β
where
y2 +λ2
e− 2 y(N−1) λ
f1,√Y (y; N − 1, λ ) = I N−1 −1 (λy ) (2.5)
(λy )
N−1
2
2

k
x2
∞
is a modified Bessel function of the first kind with parameter ν .
4
and Iν (x ) = x
2 k=0 k! (ν +k+1 )

Proof. See Appendix A 



2.1.2. Volatility swap prices case: N−1
i=1 μ2i = 0
N−1 2 √
In the case that i=1√μi = 0, Y defined in (A.2) is distributed according to the chi distribution with N − 1 degrees
of freedom, denoted by Y ∼ χN−1 (see Johnson et al. [16]). As a result, a pricing formula for volatility swaps based on
the square root of the discretely-sampled log-return realized variance (1.2) has a simple form as proposed in the following
theorem.

Theorem 2.3. Suppose that N−1 i=1 μi = 0. The fair strike price of a volatility swap in (1.3) can be expressed as
2


√ 2
N
Kvol ≡ Kvol2 (t, N, T ; r (t ), σ ) = σ t 2
 × 100 (2.6)
T N−1
2

4
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

∞
where (x ) = 0 t x−1 e−t dt is the gamma function.

Proof. See Appendix A. 


 N−1
The following corollary provides the exact distribution of RVd case: i=1 μ2i = 0.

Corollary 2.4. According to Theorem 2.3 and setting β = σ t × 100, we have the following results.
T

1. The variance of RVd can be found in explicit form as

VARtQ1 [ RVd ] = β 2 (N − 1 ) − Kv2ol2 . (2.7)
2. For 0 ≤ a ≤ b ≤ ∞,
  b
β 1 y
Q (a ≤ RVd ≤ b) = f2,√Y ; N − 1 dy (2.8)
a
β
β β
where
y2
yN−2 e− 2
f2,√Y ( y; N − 1 ) = N−3 N−1
. (2.9)
2 2
2

Proof. The proof is rather trivial, and can be written by analogy with the proof of Corollary 2.2, omitted here. 
N−1
It should be discussed here about the difference between the value of Kvol2 and the limit of Kvol1 when i=1 μ2i ap-
proaches zero. The following corollary clarifies this point.

Corollary 2.5. According to Theorems 2.1 and 2.3, we have



N−1
Kvol1 → Kvol2 as μ2i → 0+ . (2.10)
i=1

Proof. See Appendix A. 

The important points to be noted here are as follows. Our analytical formulas (2.2) and (2.6) for pricing volatility swaps
under the B-S model (1.1) with time-varying risk-free interest rate and constant volatility are in simple forms, which can be
exploited to explore some interesting properties discussed in the following sections. In terms of implementing our formulas,
one can compute values of Laguerre as well as gamma functions for obtaining strike prices of volatility swaps by employing
some numerical calculation packages provided in MATHEMATICA, MATLAB, or MAPLE.
Additionally, for the case that the volatility is either a time-dependent function or stochastic process, we demonstrate
how to apply our analytical formulas (2.2) and (2.6) to construct simple closed-form approximate formulas for pricing
volatility swaps in Sections 4.1 - 4.2. Furthermore, the accuracy and efficiency of the approximations are intensively in-
vestigated in Section 5.1.

2.1.3. Vega of volatility swaps


As introduced by Hull [14], we define the vega of a volatility swap as the first-order variation in the fair strike price with
∂K
respect to the volatility parameter σ , i.e., VKvol := ∂σvol . In risk management, vega is typically expressed as the amount of
money per underlying share that the value of a volatility swap will gain or lose as volatility rises or falls by 1 percentage
point. When vega is highly positive or highly negative, there is a high sensitivity to changes in volatility. If the vega of a
volatility derivative position is close to zero, volatility changes have very little effect on the value of the position.
Utilizing our analytical formulas derived in Theorems 2.1 - 2.3, we present an exact formula for VKvol in the following
corollary.

Corollary 2.6. According to Theorems 2.1 - 2.3, we have the following results.

i=1 μi > 0,
1. Case: N−1 2
   
π 1  2 ( N−1 ) − λ2
N−1
1 N−1
VKvol = Kvol1 + (t σ )2 − 2 Ri L− 12 × 100 (2.11)
1 σ 2T t 4 σ 2
i=1
2
N−1
t μ2i
where Ri = t i+1 r (s )ds for i = 1, . . . , N − 1 and λ =
i=1
√ .
 i σ t
i=1 μi = 0,
2. Case: N−1 2

N

2t 1
VKvol =
2
 × 100 = Kvol2 > 0. (2.12)
2 T N−1
2
σ
Proof. See Appendix A. 

5
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

2.2. An analytical formula for pricing variance swaps

Similar to pricing volatility swaps, the value of a variance swap at time t1 is the expected present value of the future

− T r (s )ds
payoff, Vt1 = e t1 EtQ [(RVd − Kvar )] × L. This should be zero at the beginning of the contract since there is no cost to
1
enter into a swap. Applying (A.2) and (A.4) in Appendix A, the fair strike price of a variance swap under the B-S model
(1.1) can be defined as

Kvar := EtQ1 [RVd ] = EtQ1 [β 2Y ]. (2.13)

2.2.1. Variance swap prices


In the following theorem, we propose an analytical formula for pricing variance swaps based on the EBS model (1.1).

Theorem 2.7. The fair strike prices of a variance swap Kvar can be expressed as

Kvar ≡ Kvar j (t, N, T ; r (t ), σ ) = β 2 (N − 1 + (2 − j )λ2 ) (2.14)


N−1
μ2i
for j = 1, 2, where λ =
i=1
√ and β = σ t × 100. In addition,
σ t T


N−1
Kvar1 → Kvar2 as μ2i → 0+ . (2.15)
i=1

Proof. See Appendix A. 


N−1
In the following corollary, we provide the exact distribution of RVd case: i=1 μ2i > 0.
Corollary 2.8. According to Corollary 2.2, we have the following results.

1. Y is distributed according to the noncentral chi-square distribution with N − 1 degrees of freedom and noncentrality parameter
λ2 , denoted by Y ∼ N C χ 2 (λ2 ) (see Johnson et al. [16]). Moreover, for 0 ≤ a ≤ b ≤ ∞,
N−1

 b
 
β2 1 y
Q(a ≤ RVd ≤ b) = f1,Y ; N − 1, λ2 ) dy (2.16)
β2
a β2 β 2

where
N−1 1
1 − y+ν y 4 −2 √
f1,Y (y; N − 1, ν ) = e 2 I N−1 −1 ( ν y ). (2.17)
2 ν 2

2. The variance of RVd can be found in explicit form as

VARtQ1 [RVd ] = 2β 4 (N − 1 + 2λ2 ). (2.18)

Proof. See Appendix A. 


N−1
We further derive the exact distribution of RVd case: i=1 μ2i = 0 as follows.
Corollary 2.9. According to Corollary 2.4, we have the following results.

1. Y is distributed according to the chi-square distribution with N − 1 degrees of freedom, denoted by Y ∼ χN−1
2 (see Johnson
et al. [16]). Moreover, for 0 ≤ a ≤ b ≤ ∞,
 b
 
β2 1 y
Q(a ≤ RVd ≤ b) = f2,Y ; N − 1 dy (2.19)
β2
a β2 β 2

where
N−3 y
y 2 e− 2
f2,Y (y; N − 1 ) = N−1 N−1
. (2.20)
2 2
2

2. The variance of RVd can be found in explicit form as

VARtQ1 [RVd ] = 2β 4 (N − 1 ). (2.21)



Proof. Applying the property of chi distribution to Y , the proof of this corollary can be written by analogy with the proof
of Corollary 2.16, omitted here. 

6
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

2.2.2. Vega of variance swaps


Like vega of volatility swaps defined in Section 2.1.3, the vega of a variance swap is the first-order variation in the fair
strike price with respect to the volatility parameter σ , i.e., VKvar := ∂∂σ
Kvar
.
Utilizing our analytical formula derived in Theorem 2.7, an exact formula for VKvar is obtained in the following corollary.

Corollary 2.10. According to Theorem 2.7, we have


 
(N − 1 )t σ 1  λ2
N−1
2
VKvar j = Kvar j + (4 − 2 j )β 2 − Ri − (2.22)
σ 2 σ i=1 σ
 ti+1
for j = 1, 2, and Ri = ti r (s )ds for i = 1, . . . , N − 1.

Proof. The proof is rather trivial, and can be written by analogy with the proof of Corollary 2.6, omitted here. 

2.3. Comparison with continuous sampling

As described in Broadie and Jain [3] in the case of the B-S model, the continuous realized volatility and continuous
 T T 2
RVc (t1 , T ) := T −t t σ dt × 100 = σ × 100 and RVc (t1 , T ) := T −t t σ dt × 100 = σ ×
1 2 2 1 2 2
realized variance are given by
1 1 1 1

1002 , respectively. We show in the following theorem that the fair strike prices of volatility and variance swaps in discrete
sampling converge to their corresponding fair strike prices in continuous sampling when the number of sampling frequency
approaches infinity.

Theorem 2.11. Set t = T


N−1 for a positive integer N ≥ 2. Utilizing the pricing formulas in Theorems 2.1 and 2.3, we have the
following results.

1.
T 
lim Kvol j , N, T ; r (t ), σ = σ × 100 = RVc (t1 , T ) (2.23)
N→∞ N−1
for j = 1, 2.
2.
T
lim Kvar j , N, T ; r (t ), σ = σ 2 × 1002 = RVc (t1 , T ) (2.24)
N→∞ N−1
for j = 1, 2.

Proof. See Appendix A. 

2.4. The convergence of the convexity correction

In this section, we investigate the famous convexity correction formula introduced by Brockhaus and Long [4] under the
B-S model (1.1). The formula allows one to approximate a fair volatility strike from a realized variance, which is easier to
compute as the volatility strike is an expectation of a square root function which is nonlinear. The fair volatility strike is
approximated as a risk-neutral expectation of a second order Taylor’s expansion of the square root of the realized variance.
To show the readers the derivation of the convexity correction formula, we consider a second order Taylor’s expansion

of f (x ) = x around x0 which can be expressed as

√ √ (x − x0 ) (x − x0 )2 ( x − x0 )3
x = x0 + 1
− 3
+ f (ε ) (2.25)
2x0 2 8x0 2 3!
where ε is in (x0 , x ).
Substituting x with RVd and x0 with EtQ [RVd ] in (2.25) yields that
1

 (RVd − EtQ1 [RVd ]) (RVd − EtQ1 [RVd ])


2

RVd ≈ EtQ1 [RVd ] + 1


− 3
. (2.26)
2EtQ1 [RVd ] 2 8EtQ1 [RVd ] 2

By taking expectation under the risk-neutral martingale measure Q to (2.26), the approximate fair volatility strike, Kˆvol j
for the cases j = 1, 2 mentioned earlier, can then be written as
 VARtQ1 [RVd ]  VARtQ1 [RVd ]
Kˆvol j := Kvar j − = Kvar j − . (2.27)
8EtQ1 [RVd ]
3 3
2
8Kv2ar j

7
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

In continuous setting, the convexity correct formula provides a decent estimate for a fair volatility strike as long as the
RVd − EtQ [RVd ] is bounded by EtQ [RVd ] (see Broadie and Jain [3]). In other words, the formula gives a good estimate if
1 1

|RVd − EtQ1 [RVd ]| ≤ EtQ1 [RVd ].


However, in discrete context, a sampling size greatly effects an accuracy of such formula. Broadie and Jain [3] showed
that, for small sampling sizes, the formula does not provide decent estimates for various models including the standard B-S
model [2].
In the following theorem, we prove the approximate volatility strike converges to the actual fair strike as the number of
the sampling size increases as well as present a closed-form of the absolute error of the fair volatility strike approximated
by the convex correction formula.

Theorem 2.12. By setting t = T


N−1 , we have
 
lim Kˆvol j − Kvol j  = 0 (2.28)
N→∞

for j = 1, 2.
Furthermore, the absolute errors obtained from approximating Kvol j by Kˆvol j for j = 1, 2, can be computed as follows:
  2 
  1 − 1
+ (24(N−5 )λ2 λ4
+ (N−1
Kˆvol1 − Kvol1  = σ  4(N−1 ) N−1 )2 )2 π ( N−1
−1 ) λ 
− L1 2
2 (N − 1 ) 2
−  × 100 (2.29)
 λ2 ) 23
(1 + N−1 2 

and
  
   N
Kˆvol2 − Kvol2  = σ 1 − 1

2 2
  × 100 (2.30)
 4 ( N − 1) N−1 N−1
2

for N ≥ 2.

Proof. See Appendix A. 

3. Pricing volatility and variance options

In this section, we present analytical pricing formulas for volatility and variance options, respectively. Our options treat-
ment concentrates on calls; results for puts can be obtained from parity relations1
For 0 ≤ t1 < T , a volatility call on [t1 , T ] with strike Kvcol pays the holder at time T some notional amount times

( RVd (t1 , N, T ) − Kvcol )+ (3.1)
and a variance call on [t1 , T ] with strike Kvcar pays the holder at time T some notional amount times
(RVd (t1 , N, T ) − Kvcar + ) (3.2)
where RVd is the log-return realized variance defined in (1.2).
The problem of pricing volatility and variance options is actually computing the conditional expectations of the payoffs
(3.1) and (3.2), respectively, with respect to Ft1 under the risk-neutral martingale measure Q. In other words, we need to
compute to following two quantities:

T
r (s )ds

Cvol := e t1
EtQ1 [( RVd (t1 , N, T ) − Kvcol )+ ] (3.3)
for a Kvcol -strike volatility call and
T
− r (s )ds
Cvar := e t1
EtQ1 [(RVd (t1 , N, T ) − Kvcar )+ ] (3.4)
for a Kvcar -strike variance call.

3.1. An analytical formula for pricing volatility options

3.1.1. Volatility call option prices


The main contributions of Corollaries 2.2 - 2.4 for pricing volatility options are demonstrated in the following theorem.

Theorem 3.1. The time-t1 value of Kvcol -strike volatility call can be written in terms of volatility swap price Kvol as


T
r (s )ds

Cvol ≡ Cvol j (t1 , t, N, T ; r (t ), σ , Kvcol ) = e t1
Kvol j − Kvcol + Ivol j (3.5)

1
The parity relations between put and call prices for European options are derived in [14] on pages 238–241.

8
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

for j = 1, 2, where
 Kvcol
1 y
Ivol1 = (Kvcol − y ) f1,√Y ; N − 1, λ dy (3.6)
0 β β
 Kvcol
1 y
Ivol2 = (Kvcol − y ) f2,√Y ; N − 1 dy (3.7)
0 β β
N−1
μ2i
where λ =
i=1
√ , β=σ t × 100, and the pdfs f √ , j = 1, 2, are given in (2.5) and (2.9), respectively.
σ t T j, Y

Proof. See Appendix B. 

3.1.2. Vega of volatility call options and some properties of Cvol


The vega of a volatility call option is the first-order variation in the fair strike price with respect to the volatility pa-
∂C
rameter σ , i.e., VCvol := ∂σvol . Utilizing our analytical formula derived in Theorem 3.1, a relation between VKvol and VCvol is
immediately obtained in the following corollary.

Corollary 3.2. According to Theorem 3.1, we have


 

T
r (s )ds ∂ Ivol j
VCvol = e t1
VKvol + (3.8)
j j ∂σ
for j = 1, 2.
∂ Ivol
Proof. See Appendix B for the proof of the corollary including explicit formulas for ∂σ j , j = 1, 2. 

The following corollary presents some interesting properties of Cvol with respect to the strike price Kvcol .

Corollary 3.3. According to Theorem 3.1, we have the following properties.


1.
T
− r (s )ds
lim+ Cvol j (t1 , t, N, T ; r (t ), σ , K ) = e t1
Kvol j > 0 (3.9)
K→0

and
lim Cvol j (t1 , t, N, T ; r (t ), σ , K ) = 0. (3.10)
K→∞

for j = 1, 2.
2.

C (t , t, N, T ; r (t ), σ , K ) < 0 (3.11)
∂ K vol j 1
and
∂2
C (t , t, N, T ; r (t ), σ , K ) > 0 (3.12)
∂ K 2 vol j 1
for all K > 0 and j = 1, 2.
3.
T
− r (s )ds
0 < Cvol j (t1 , t, N, T ; r (t ), σ , K ) < e t1
Kvol j (3.13)
for all K > 0 and j = 1, 2.
Proof. See Appendix B. 

3.2. An analytical formula for pricing variance options

3.2.1. Variance call option prices


The next theorem demonstrates the main contributions of Corollaries 2.8 - 2.9 for pricing variance options.

Theorem 3.4. The time-t1 value of Kvcar -strike volatility call can be written in terms of variance swap price Kvar as

T
r (s )ds

Cvar ≡ Cvar j (t1 , t, N, T ; r (t ), σ , Kvcar ) = e t1
Kvar j − Kvcar + Ivar j (3.14)
for j = 1, 2, where
 Kvcar
 
1 y
Ivar1 = ( Kvcar − y) f1,Y ; N − 1, λ 2
dy (3.15)
0 β2 β2

9
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

 Kvcar
 
1 y
Ivar2 = ( Kvcar − y) f2,Y ; N − 1 dy (3.16)
0 β2 β2
N−1
μ2i
where λ =
i=1
√ , β=σ t × 100, and the pdfs f , j = 1, 2, are given in (2.17) and (2.20), respectively.
σ t T j,Y

Proof. See Appendix B. 

3.2.2. Vega of variance call options and some properties of Cvar


The vega of a variance call option is the first-order variation in the fair strike price with respect to the volatility pa-
rameter σ , i.e., VCvar := ∂∂σ
Cvar
. Utilizing our analytical formula derived in Theorem 3.4, a relation between VKvar and VCvar is
obtained in the following corollary.

Corollary 3.5. According to Theorem 3.1, we have


 

T
r (s )ds ∂ Ivar j
VCvar j = e t1
VKvar j + (3.17)
∂σ
for j = 1, 2.
∂ Ivar
Proof. See Appendix B for the proof of the corollary including explicit formulas for ∂σ j , j = 1, 2. 

By mimicking the proof of Corollary 3.3, the properties of Cvar with respect to the strike price Kvcar can be immediately
obtained as follows.

Corollary 3.6. According to Theorem 3.4, we have the following properties.

1.
T
− r (s )ds
lim+ Cvar j (t1 , t, N, T ; r (t ), σ , K ) = e t1
Kvar j > 0 (3.18)
K→0

and

lim Cvar j (t1 , t, N, T ; r (t ), σ , K ) = 0. (3.19)


K→∞

for j = 1, 2.
2.

C (t , t, N, T ; r (t ), σ , K ) < 0 (3.20)
∂ K var j 1
and
∂2
C (t , t, N, T ; r (t ), σ , K ) > 0 (3.21)
∂ K 2 var j 1
for all K > 0 and j = 1, 2.
3.
T
− r (s )ds
0 < Cvar j (t1 , t, N, T ; r (t ), σ , K ) < e t1
Kvar j (3.22)
for all K > 0 and j = 1, 2.

4. Applications of our analytical pricing formulas

In this section, we shall adopt the analytical formulas derived in Section 2 to construct an approximate formula of the
true values of volatility swaps with discrete sampling when the dynamics of the underlying asset price follows the model
described as follows.

4.1. Approximation for the Black-Scholes model with time varying parameters

We assume that the underlying asset price St evolves according to the SDE:

dSt = r (t )St dt + σ (t )St dWt (4.1)


for 0 < t ≤ T and S0 > 0 where r (t ) is a risk-free interest rate at time t , σ (t ) > 0 is price volatility at time t, and (Wt )0≤t≤T
is the standard Brownian motion under a probability space (, F, Q ) with a filtration (Ft )0≤t≤T . In addition, we assume
that r (t ) and σ (t ) are continuous on [0, T ].

10
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

According to (1.3), the fair strike prices of volatility and variance swaps under the model (4.1) can be defined as

Kvσol(t ) := EtQ1 [ RVd ] (4.2)
and
Kvσar(t ) := EtQ1 [RVd ] (4.3)
for 0 < t1 < T , respectively.
In similar way as written in (2.1) but σ (t ) depends on time, we alternatively define
 ti+1
1 2
μ̄i := r (s ) − σ (s ) ds (4.4)
ti 2
for i = 1, 2, . . . , N − 1.
 ti
μ̄i−1 + σ (s )dWs
From (4.1) and (4.4), we have Sti = Sti−1 e ti−1
. Next, we set Z̄i = ln Sti − ln Sti−1 and we obtain
 ti
  ti

Z̄i = μ̄i−1 + σ (s )dWs ∼ N μ̄i−1 , σ 2 (s )ds (4.5)
ti−1 ti−1

for i = 2, . . . , N.
In the following corollary, we derive an analytical formula for Kvσol(t ) defined in (4.2) in some particular cases.

Corollary 4.1. Suppose that the following condition holds:


 ti+1
σ̄ 2 t = σ 2 (s )ds (4.6)
ti

for some σ̄ > 0 for all = i, . . . , N − 1.


 σ (t )
i=1 μ̄i > 0. Then, Kvol
1. If N−1 can be obtained by using (2.2) in Theorem 2.1 with replacing μi by μ̄i and σ by σ̄ .
2
N−1 2
2. If i=1 μ̄i = 0. Then, Kvσol(t ) can be obtained by using (2.6) in Theorem 2.3 with replacing σ by σ̄ .

Proof. See Appendix C. 

4.1.1. Construction of an approximate of Kvσol(t )


Next, we consider for the case that (4.6) is not true. Firstly, we define
 ti+1
 12
σ̄i := σ (s )ds
2
(4.7)
ti

and
N−1  12
j=1 μ̄2j
λ̄i := (4.8)
σ̄i
for i = 1, 2, . . . , N − 1.
Our conceptual idea of constructing an approximate of Kvσol(t ) on the time interval [t1 , T ] is to compute volatility swap
prices by using our analytical formula (2.2) or (2.6) corresponding to μ̄i , σ̄i , and λ̄i . These volatility swap prices shall be
weighted with appropriate values in order to obtain a good approximation for Kvσol(t ) .
In our approximation for the B-S model with time varying parameters (4.1), we choose weights as follows:
  ti+1  12
ti σ 2 (s )ds σ̄i
w̄i := T = (4.9)
(N − 1 ) σ (s )ds
2 N−1
t1 (N − 1 ) k=1 σ̄k2
for i = 1, 2, . . . , N − 1.
A closed-form approximation of Kvσol(t ) can be constructed by using (4.4)-(4.9) as


N−1
K̄vσol(t ) (t, N, T ; r (t ), σ (t )) := w̄i Kvolk (t, N, T ; r (t ), σ̄i ) (4.10)
k
i=1

where Kvolk (t, N, T ; r (t ), σ̄i ) for k = 1, 2, are computed by using (2.2) and (2.6), respectively, with replacing μi by μ̄i , σ t
by σ̄i , and λ by λ̄i .
It should be noticed by the construction of the approximate (4.10) that K̄vσol(t ) and Kvσol(t ) coincide when the condition
k
(4.6) holds.

11
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

4.1.2. An error estimate for K̄vσol(t )


1
To obtain an error estimate for K̄vσol(t ) , we consider a random variable defined by
1
 2

N
Z̄i
Ȳ := (4.11)
i=2
σ̄i−1

where Z̄i and σ̄ are given in (4.5) and (4.7), respectively. Then, Ȳ ∼ N C χN−1 (λ̄ ) where the noncentrality parameter is given
by
  2

N−1
μ̄i
λ̄ = (4.12)
i=1
σ̄i
and μ̄i is given in (4.4).
In the following theorem, we construct lower and upper bounds of Kvσol(t ) using some properties of Ȳ as well as derive an
1
error estimate obtained from approximating Kvσol(t ) by K̄vσol(t ) .
1 1

Theorem 4.2. Let σ̄min = min {σ̄i ; i = 1, . . . , N − 1} and σ̄max = max {σ̄i ; i = 1, . . . , N − 1}. Suppose that λ̄ > 0. We have the fol-
lowing results:

1.
σ̄min K̄vol ≤ Kvσol(1t ) ≤ σ̄max K̄vol (4.13)
where
 2
π ( N−1 −1 ) λ̄
K̄vol = L1 2 − × 100 (4.14)
2T 2 2
and L is the Laguerre function given in Theorem 2.1.
2. If

σ̄min K̄vol ≤ K̄vσol(1t ) ≤ σ̄max K̄vol (4.15)

where K̄vσol(t ) is given in (4.10). Then, we obtain an error estimate


1
 σ (t ) 
K̄ − K σ (t )  ≤ (σ̄max − σ̄min )K̄vol . (4.16)
vol1 vol1

Proof. See Appendix C 

4.1.3. An analytical formula for Kvσar(t1)


It should be noted from (4.5) and (4.7) that
 
2 μ̄2i−1
Ȳi := Z̄i /σ̄i2−1 ∼ N C χ 2 (4.17)
1 σ̄i2−1
1002 N
for all i = 2, . . . , N. As a result, we have RVd = T i=2 σ̄i2−1Ȳi where Y j and Yk are independent for all j, k = 2, . . . , N and
j = k. Hence, the analytical formula for Kvσar(t ) defined in (4.3) and VARtQ [RVd ] can be easily obtained as written in the
1
following theorem.

Theorem 4.3.
 
1002  2
N
μ̄2i−1
Kvσar(t1) = σ̄i−1 1 + , (4.18)
T
i=2
σ̄i2−1
and
 
1004  4
N
μ̄2i−1
VARtQ1 [RVd ] = EtQ1 [ (RVd − Kvσar(t1) )2 ] = 2 × 2 σ̄i−1 1+2 2 , (4.19)
T
i=2
σ̄i−1
where μ̄i−1 and σ̄i−1 are given in (4.4) and (4.7), respectively. In addition, the convexity correction formula for Kvσol(t ) can be
1
derived by using (2.27) as
VARtQ1 [RVd ]
Kˆvσol(t ) := Kvσar(t1) − . (4.20)
8(Kvσar(t1) ) 2
1 3

12
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Proof. The proof is rather trivial by applying the properties of the noncentral chi-square random variables with degree of
freedom 1 and noncentrality parameter μ̄2i−1 /σ̄i2−1 (see Johnson et al. [16]), thus omitted here. 

It should be remarked that we shall use Kˆvσol(t ) as a benchmark to investigate the accuracy of K̄vσol(t ) compared with MC
1 1
simulations in our numerical study in Section 5.1.

4.2. Approximation for the Heston stochastic volatility model

In Heston stochastic volatility model [13], the dynamics of the underlying asset St is assumed to follow a diffusion process
defined on an original probability space with a stochastic instantaneous variance vt . According to the well-known results of
Harrison and Pliska [10], the existence of an equivalent martingale measure guarantees the absence of arbitrage opportuni-
ties. Therefore, we are able to change the original probability measure to a risk-neutral martingale measure Q and describe
the processes as
√ 
dSt = r (t )St dt + vt St dWtS
√ (4.21)
dvt = κ (t )(θ (t ) − vt )dt + σv (t ) vt dWtv

for S0 , v0 > 0 where r (t ) is a time varying risk-free interest rate. The time-dependent functions κ (t ) and θ (t ) are the risk-
neutral parameters. The volatility of volatility σv (t ) is assumed to depend on time. The two-dimensional Brownian motion
(WtS , Wtv )0≤t≤T under a probability space (, F, Q ) with a filtration (Ft )0≤t≤T describes the random noises in asset and
variance. They are assumed to be correlated with a constant correlation coefficient ρ , i.e., dWtS dWtv = ρ dt.
It should be noted that our model for the underlying asset and variance written in (4.21) is more general than the Heston
stochastic volatility model in which the variance process vt is assumed to follow the so-called extended CoxIngersollRoss
(ECIR) process, i.e., the parameter functions are allowed to depend on time, and hereafter we refer to the Heston-ECIR model.
By setting these parameter functions to be constants such that κ (t ) = κ , θ (t ) = θ and σv (t ) = σv , the variance process vt
follows the CoxIngersollRoss (CIR) process and hereafter we refer to the Heston-CIR model.

4.2.1. Approximation for the Heston-ECIR model


According to the theoretical study of the ECIR process by Maghsoodi [20], the following two assumptions are required,
in order to ensure that vt avoids zero a.s. Q for all t ∈ [0, T ] providing that v0 > 0.

Assumption 1. The parameter functions θ (t ), κ (t ) and σv (t ) are strictly positive and continuous on [0, T ] such that the
dimension of the ECIR process, defined by δ (t ) := 4θ (t2)κ (t ) , is bounded.
σv (t )

Assumption 2. The inequality δ (t ) ≥ 2 holds for all t ∈ [0, T ].

According to (1.3), the fair strike price of a volatility swap based on our model can be defined as

KvHol := EtQ1 [ RVd ]. (4.22)

By choosing a two-dimensional Brownian motion (Wt


ˆ v )0≤t≤T such that W
ˆ S, W
t
ˆ v are uncorrelated with respect to
ˆ S and W
t t
(, F, Q ), the SDEs in (4.21) can be rewritten as
√ 
dSt = r (t )St dt + vt St dW ˆ tS

√ (4.23)
dvt = κ (t )(θ (t ) − vt )dt + σv (t ) vt (ρ dW
ˆ tS + 1 − ρ 2 dW
ˆ tv ).

Let Xt = ln St . Applying Itô lemma to (4.23), we have


 √

dXt = r (t ) + (ρσv (t )(1 + 1 − ρ 2 ) − 12 )vt dt + vt dWˆ tS
√  (4.24)
dvt = κ (t )(θ (t ) − vt )dt + σv (t ) vt (ρ dW ˆ tS + 1 − ρ 2 dW
ˆ tv ).
A conceptual idea which is similar to the one presented in Section 3.1 shall be applied to construct an approximate of
KvHol as follows.
Due to the drift term of dXt in (4.24), we approximate vt on the time interval [ti , ti+1 ] by EtQ [vt ] and define
1
 ti+1  1
μˆ i := r (s ) + (ρσv (s )(1 + 1 − ρ 2 ) − )EtQ1 [vs ] ds (4.25)
ti 2
for i = 1, 2, . . . , N − 1. Note that μ
ˆ i is a deterministic function depending on vt1 , providing some values of t1 , ti , ti+1 , and
model parameters.
√ √
Similarly, we approximate vt on the time interval [ti , ti+1 ] by EtQ [ vt ]. Hence from the diffusion term of dXt in (4.24),
1
we define
  12
ti+1 √ 2
σˆ i := EtQ1 [ s ] ds
v (4.26)
ti

13
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

N−1  12
j=1 μˆ 2j
λˆ i := (4.27)
σˆ i
for i = 1, 2, . . . , N − 1.
In our approximation for the Heston-ECIR model (4.21), we set weights as follows:
σˆ
ˆ i := N−1i
w (4.28)
k=1 σˆ k
for i = 1, 2, . . . , N − 1.
To compute the conditional expectations appearing in (4.25)-(4.26), we thank for the work of Rujivan [26] providing the
closed-form formulas for these conditional expectations when the variance vt follows the ECIR process.

Next, we set vt1 = v. Applying Theorems 2.1–2.2 proposed in [26] to compute EtQ [ vs ] and EtQ [vs ], respectively, the
1 1
conditional expectations can be written in terms of v as


√ √
EtQ1 [
1
vs ] = A 21 (s − t1 ) v + A 1 −k (s − t1 )v 2 −k
2
(4.29)
k=1

EtQ1 [vs ] = A1 (s − t1 )v + A0 (s − t1 ) (4.30)


where the coefficient functions depending on time are given by

A 1 ( τ ) = e− 2
1
0 κ (s−u )du (4.31)
2


 τ η
A 1 −k (τ ) = e−( 2 −k )
1
0 κ (s−u )du e( 2 −k )
1
0 κ ( s −ζ ) d ζ P (s − η )A 21 −k+1 (η )dη (4.32)
1
2 2 −k+1
0
  
for k = 1, 2, . . . , where P 1 −k+1 (τ ) := 3
2 −k 1
4 − 2k σv2 (τ ) + κ (τ ) θ (τ ) and
2

A1 ( τ ) = e− 0 κ (s−u )du (4.33)
 τ
A0 ( τ ) = κ ( s − η )θ ( s − η )A1 ( η )d η . (4.34)
0

From (4.29), we have


 
√ 2  ∞
 
k
EtQ1 [ s ] v = A 1 ( s − t1 )v +
2
A 1 −m (s − t1 )A 1 −(k−m ) (s − t1 ) v1−k . (4.35)
2 2 2
k=1 m=0

Applying (4.30) and(4.35) to (4.25)-(4.28), one can show that μ


ˆ i , σˆ i , λ
ˆ i , and w
ˆ i can be written in terms of v. In particular,

μˆ i (v ) = a1,i v + a0,i (4.36)

  12


σˆ i (v ) = b1,i v + b1−k,i v 1−k
(4.37)
k=1

for i = 1, 2, . . . , N − 1 where
 t j=i+1  1
ak,i = (1 − k )r (s ) + (ρσv (s )(1 + 1 − ρ2 ) − )A (s − t1 ) ds (4.38)
ti 2 k
for k = 0, 1 and
 
 ti+1 
k
b1−k,i = A 1 −m (s − t1 )A 1 −(k−m ) (s − t1 ) ds (4.39)
2 2
ti m=0

for k = 0, 1, . . . .
As the results in (4.36)-(4.39), a closed-form approximation of KvHol can be deduced as


N−1 
KˆvHolk (t, N, T ; r (t ), v ) := ˆ i (v )Kvolk t, N, T ; r (t ), σˆ i (v )
w (4.40)
i=1

where vt1 = v and Kvolk t, N, T ; r (t ), σˆ i (v ) for k = 1, 2, are computed by using (2.2) and (2.6), respectively, with replacing

μi by μˆ i (v ), σ t by σˆ i (v ), and λ by λˆ i (v ).

14
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

4.2.2. Approximation for the Heston-CIR model


It should be noticed when the variance vt follows the CIR process that the coefficient functions as written in (4.33)-(4.32)
can be rewritten without integrals by using Theorems 2.3 - 2.4 proposed in [26] as follows:

A1 (τ ) = e−κτ (4.41)

A0 (τ ) = θ e−κτ (eκτ − 1 ) (4.42)

A 1 (τ ) = e− 2 κτ
1
(4.43)
2

 

k
e− 2 κτ
1
eκτ − 1 k
A 1 −k (τ ) = P̄ 1 −m+1 (4.44)
2
m=1
2 k! κ

for k = 1, 2, . . . , where P̄ 1 −m+1 := ( 32 − m )(( 14 − 2 )σv + κθ ).


m 2
2
Consequently, ak,i and b1−k,i in (4.38)-(4.39) can be simplified by using (4.41)-(4.44) to obtain
    
ti+1
θ e(t1 −ti+1 )κ − e(t1 −ti )κ
a0,i = r (s )ds + ρσv (1 + 1 − ρ )t − 2 t + (4.45)
ti 2 κ
  
1 e(t1 −ti )κ − e(t1 −ti+1 )κ
a1,i = ρσv (1 + 1 − ρ 2 )t − (4.46)
2 κ

e(t1 −ti )κ − e(t1 −ti+1 )κ


b1,i = (4.47)
κ
  

k 
m −m
k
1
b1−k,i = P̄ 1 −m1 +1 P̄ 1 −m2 +1 Bk,i (4.48)
m=0
m ! ( k − m )! κ k m1 =1
2
m2 =1
2

for k = 1, 2, . . . , where
 ti+1
Bk,i := e−(s−t1 )κ (e(s−t1 )κ − 1 )k ds. (4.49)
ti

Applying (4.45)-(4.49) to (4.25)-(4.28) for computing μ


ˆ i (v ), σˆ i (v ), λ
ˆ i , and w
ˆ i , respectively, a closed-form approximation
H
of Kvol for the Heston-CIR model can be deduced by using (4.40).

5. Numerical results and discussions

5.1. Comparison with MC simulations

The frameworks presented in Sections 2 - 3 are based on the analytical approach and should therefore reflect the true
theoretical values of a volatility swap and volatility option, respectively. However as pointed out in Section 1, there has
no analytical solution available for the B-S model with time varying parameters. Therefore, some numerical approaches are
necessary for pricing volatility swaps and volatility options with discrete sampling.
In this section, the accuracy as well as the efficiency of our analytical formulas and approximate formulas derived in
Sections 2 - 4, shall be investigated through four numerical examples, respectively. Our volatility swap prices calculated
through the analytical formulas and approximate formulas are benchmarked against by MC prices obtained through MC
simulations.
Our MC simulations are based on the RungeKutta (R-K) method, implemented by the numerical package called “ItoPro-
cess” in MATHEMATICA, to obtain numerical solutions of the SDEs (4.1) and (4.23) as well as approximate values of the
conditional expectations (4.2) and (4.22). In all our calculations, MATHEMAICA 11 and a PC with the following details were
used: Intel(R) Core (TM) i5-6500, CPU @3.20GHz, 16GB RAM, Windows 10, 64 bit operating system2

2
One can perform MC simulations based on the Euler-Maruyama (E-M) as well as Milstein methods to produce numerical solutions of the SDEs (4.1) and
(4.23) by using the numerical package “ItoProcess”. Even though, in regards to our experiments, the average time consumed by R-K method is higher than
the ones consumed by E-M and Milstein methods approximately 2–3 times. We preferred to employ the R-K method in our numerical tests in order to
obtain more accurate results with a higher rate of convergence in MC simulations.

15
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Table 1
Verification for the condition in Theorem 2.1
in which r1 (t ) = 0.05, r2 (t ) = 6 + 0.5t, and
r3 (t ) = 10 + 0.5t + 0.5t 2 for t ∈ [0, T ], T = 1, N =
252, t = N−1 T
, t1 = 0, ti = t1 + (i − 1 )t, i = 2, . . . N,
and all sample points σ0 = 0.10, 0.15, . . . , 1.00.

σ0 M ( r 1 , σ0 ) M ( r 2 , σ0 ) M ( r 3 , σ0 )

0.10 8.067729×10−6 0.155462 0.432221


0.15 5.982321×10−6 0.155151 0.431703
0.20 3.585657×10−6 0.154716 0.430978
0.25 1.400647×10−6 0.154158 0.430046
0.30 9.960159×10−8 0.153478 0.428909
0.35 5.042331×10−7 0.152675 0.427567
0.40 3.585657×10−6 0.151752 0.426022
0.45 0.000010 0.150709 0.424273
0.50 0.000022 0.149548 0.422323
0.55 0.000041 0.148269 0.420173
0.60 0.000067 0.146875 0.417825
0.65 0.000104 0.145368 0.415280
0.70 0.000151 0.143748 0.412540
0.75 0.000213 0.142019 0.409607
0.80 0.000290 0.140182 0.406484
0.85 0.000386 0.138240 0.403172
0.90 0.000502 0.136195 0.399674
0.95 0.000641 0.134049 0.395993
1.00 0.000807 0.131806 0.392131

5.1.1. The Black-Scholes model with time varying parameters


In order to demonstrate the accuracy and efficiency of our approach adopted for the B-S model with time varying pa-
rameters (4.1), we choose3

r (t ) = r̄0 + r̄1 t + r̄2 t 2 ; σ (t ) = σ0 + σ1 cos2 (2π σ0 t ) (5.1)

for t ∈ [0, T ], where r̄i ≥ 0, i = 0, 1, 2, σ0 > 0, and σ1 ≥ 0.



It should be mentioned that our numerical study presented in the section focuses on the case: N−1 i=1 μi > 0. As a result
2
N−1 2 N−1
of Corollary 2.5, for the case: i=1 μi = 0, our numerical results can be used when the value of i=1 μi is very close to 2

zero. Therefore, we omit presenting numerical results for the rare case for brevity reasons.
Example 1. (Accuracy and efficiency of our analytical pricing formula for volatility swaps)
We set T = 1, N = 252, t = N−1 T
, t1 = 0, and ti = t1 + (i − 1 )t for i = 2, . . . N. This example considers three cases of
parameter setting for r (t ), i.e., r1 (t ) = 0.05, r2 (t ) = 6 + 0.5t, and r3 (t ) = 10 + 0.5t + 0.5t 2 for t ∈ [0, T ], while σ0 varies on
(0,1) and σ1 = 0.
With this parameter setting, Table 1 shows that the condition in Theorem 2.1 is fulfilled, i.e.,
 2

N−1 
N−1 ti+1
1
M (r j , σ0 ) := μ =
2
i, j r j (s )ds − σ02 t >0
ti 2
i=1 i=1

for all j = 1, 2, 3, and all sample points σ0 = 0.10, 0.15, . . . , 1.00. Hence, our analytical formula (2.2) can be adopted to obtain
the true values of volatility swaps in this example. In addition, as shown in Table 1, our numerical results obtained from

setting σ0 = 0.10, 0.15, . . . , 0.4 and r (t ) = r1 (t ), can be applied for the case: N−1
i=1 μi = 0 as previously mentioned.
2

The volatility swap prices (Kvol1 ) obtained from (2.2) are plotted against those obtained by MC simulations using 10,0 0 0
sample paths (KvMC ol
) as shown in Fig. 1. One can clearly observe that the results from our analytical formula perfectly match
1
the results from the MC simulations.
To ensure that readers have some quantitative concept of how large the difference between the results from our ana-
lytical solutions and those from the MC simulations is, we have selected r3 (t ) to be our case study and tabulated MC ; the
average of the percentage relative differences of Kvol1 and KvMC
ol
for the 19 sample points of σ0 as a function of number of
1
paths N p , using our analytical solution (2.2) as the reference in the calculation, in Table 2.
Clearly, when the number of paths reaches 1,0 0 0 in MC simulations, the relative difference of the two is less than 0.10%
already. Such an average relative difference is further reduced when the number of paths is increased; showing that the
convergence of the MC simulations towards our analytical solution.

3
The quadratic and periodic functions written in (5.1) can be interpreted as the instantaneous forward rate of maturity t implied from the yield curve
of a risk-free asset and a skewed structure of implied volatilities known as “the smile effect” in the B-S model, respectively. For more details about the
smile effect, readers can study in [6,8].

16
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 1. A comparison between the volatility swap prices obtained by using our analytical formula (2.2) with r j (t ), j = 1, 2, 3, for t ∈ [0, 1] and the MC
simulations.

Table 2
Computational times of the MC simulations (TMC ) compared with the com-
putational time (TS ) obtained by implementing our analytical formula (2.2)
with r3 (t ) = 10 + 0.5t + 0.5t 2 for t ∈ [0, 1], including approximate folds of re-
duction in computational time for one sample point.

Np MC (%) TMC (seconds) TS (seconds) Reduction (folds)

1,000 0.0911 13.0128 0.0446 291


5,000 0.0532 64.7569 0.0446 1451
10,000 0.0488 128.0783 0.0446 2,871
50,000 0.0471 658.2607 0.0446 14,759
100,000 0.0459 1264.9109 0.0446 28,361

On the other hand, in terms of computational time, the MC simulations consume much longer time than our analytical
solution does. To demonstrate it clearly, we compare the computational times of implementing formula (2.2) and the MC
simulations with sampling frequency N = 252. Table 2 illustrates the computational times for different path numbers in
the MC simulations. In contrast to a formidable computational time of 1264.9109 seconds using the MC simulations with
10 0,0 0 0 paths, implementing formula (2.2) just consumes 0.0446 seconds: a roughly 28,361 thousands folds of reduction in
computational time for one data point.
Example 2. (Accuracy of our analytical pricing formulas for volatility and variance options)
In this example, we set the values of T , N, t, and σ1 to be the same values as used in Example 1. We select r2 (t ) given
in Example 1 to compute volatility call option prices (Cvol1 ) and variance call option prices (Cvar1 ) based on the formulas
(3.5) and (3.14), respectively. MC simulations are employed to obtain MC volatility call option prices (CvMC
ol
) and MC variance
1
option prices (CvMC
ar1 ) using 10,0 0 0 sample paths in MC simulations.
Using different values of σ0 = 0.03, 0.30, 0.5, we plot Cvol1 (Kvcol ) against CvMC
ol
(Kvcol ) for strike volatility call Kvcol ∈ (0, 80 ) in
1
Fig. 2 whereas Cvar1 (Kvcar ) is plotted against CvMC
ar1 (Kvar ) for strike variance call Kvar ∈ (10 0 0, 20 0 0 ) in Fig. 3. It is clearly seen
c c

that the results from our analytical pricing formulas (3.5) and (3.14) perfectly match the results from the MC simulations.
Next, we consider the properties of the volatility call option price Cvol1 with respect to the strike volatility call Kvcol as
previously proposed in Corollary 3.3. As shown in Fig. 2, our numerical results confirm (3.9)-(3.13). Similarly, the properties
of the variance call option price Cvar1 with respect to the strike variance call Kvcar written in (3.18)-(3.22) are illustrated in
Fig. 3. Furthermore, Fig. 4 demonstrates that the volatility call option prices never exceed the present values of its corre-

17
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 2. A comparison between the volatility call option prices obtained by using our analytical formula (3.5) with r (t ) = r2 (t ) for t ∈ [0, 1] and the MC
simulations.

Fig. 3. A comparison between the variance call option prices obtained by using our analytical formula (3.14) with r (t ) = r2 (t ) for t ∈ [0, 1] and the MC
simulations.


− T r (s )ds
sponding volatility swap defined by e t1 2 Kvol1 and this can be viewed as an important implication of the inequality
(3.13).
Example 3. (Accuracy and efficiency of K̄vσol(t ) )
From the time varying parameters written in (5.1), we set r¯2 = 0, σ1 = 1, and r¯0 = r¯1 = r0 . This implies r (t ) = r0 (1 + t )
and σ (t ) = σ0 + cos2 (2π σ0 t ). Next, we define a domain for (r0 , σ0 ) by Dr0 ,σ0 := (0, 1 ) × (0, 1 ) and choose 81 sample points
in the domain determined by (r0,m , σ0,n ) = (0.1m, 0.1n ) for m, n = 1, 2, . . . , 9. Let T = 1, N = 252, t = N−1T
, t1 = 0, and ti =
t1 + (i − 1 )t for i = 2, . . . N. Firstly, we apply Theorem 4.2 to compute the lower and upper bounds of Kvσol(t ) derived in
(4.13). These bounds are plotted against our volatility swap prices K̄vσol(t ) (r0 , σ0 ) obtained from our approximate formula

18
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 4. A comparison between the volatility call option prices and the present values of its corresponding volatility swap obtained from our analytical
formulas (3.5) and (2.2), respectively, with r (t ) = r2 (t ) for t ∈ [0, 1].

Fig. 5. A comparison between the approximates of the volatility swap prices obtained by using our approximate formula (4.10) and the lower and upper
bounds of Kvσol(t ) derived in (4.13) with r (t ) = r0 (1 + t ) and σ (t ) = σ0 + cos2 (2πσ0 t ) over Dr0 ,σ0 .

(4.10) over Dr0 ,σ0 in Fig. 5. We clearly see K̄vσol(t ) satisfies the inequality (4.15) on Dr0 ,σ0 , making sure that our approximate
prices are in the feasible region such the errors can be estimated by using (4.16).
Secondly, the MC volatility swap prices corresponding to the 81 sample points computed by using N p = 10 0, 0 0 0 in MC
simulations, are plotted against K̄vσol(t ) (r0 , σ0 ) in Fig. 6. As shown in Fig. 6, the point-wise closeness between our volatility
price and MC volatility swap price is a clear sign of the accuracy of our approximate formula (4.10).
In addition, to illustrate the convergence of the MC volatility swap price to the true value of the conditional expectation
on the RHS of (4.2), we choose three sample points from the 81 sample points in Dr0 ,σ0 and compute MC volatility swap
prices in which N p varies from 10 0 0, 20 0 0, . . . , 10 0 0 0 0. The convergences of the three sequences of the MC volatility swap
prices corresponding to the three sample points are shown in Fig. 7 (a)-(c). In particular, we clearly see that the limits
are closer to the approximate values obtained from our approximate formula (4.10) (the blue lines) than the ones obtained
from the convexity correction formula (4.20) (the green lines). A similar result is also obtained for the remaining sample

19
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 6. A comparison between the approximates of the volatility swap prices obtained by using our approximate formula (4.10) and the MC simulations
with r (t ) = r0 (1 + t ) and σ (t ) = σ0 + cos2 (2πσ0 t ) over Dr0 ,σ0 .

Fig. 7. The convergence of the MC volatility prices to the true value of the conditional expectation on the RHS of (4.2) compared with K̄vσol(t ) obtained from
our approximate formula (4.10) (the blue lines) and Kˆvσol(t ) obtained from the convexity correction formula (4.20) (the green lines) for three different sample
points in Dr0 ,σ0 . (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

20
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 8. The percentage relative differences between the results from our approximate formula (4.40) and those from the MC simulations interpolated by
using contours over Dv0 ,θ0 (The percentage relative differences obtained when θ0 = 0, i.e., the Heston-ECIR model (4.21) reduces to the Heston-CIR model
(5.2), can be determined by the values on the dashed green line.). (For interpretation of the references to colour in this figure legend, the reader is referred
to the web version of this article.)

points but we omit presenting our numerical results for brevity reasons. Therefore, the numerical results shown in Figs. 6 -
7 ensure our formula (4.10) produces volatility swap prices which are very close to the true values of Kvσol(t ) .
To demonstrate the efficiency of the approximate formula (4.10), we compare the computational times of implementing
formula (4.10) and the MC simulations with sampling frequency N = 252 and N p = 10 0, 0 0 0 on Dr0 ,σ0 . Unlike a computa-
tional time of 1264.2978 seconds using the MC simulations, implementing formula (4.10) just consumes 0.0846 seconds: a
roughly 15,180 thousands folds of reduction in computational time for one data point.

5.1.2. The Heston-ECIR model


Example 4. (Accuracy and efficiency of KˆvHol )
To investigate the accuracy of the approximate formula (4.40) for pricing volatility swaps under the Heston-ECIR model
as previously described in Section 4.2, we choose the parameter functions in (4.21) as follows:
2 2θ0 sin (2π t )
r (t ) = 0.05(1 + t ), θ (t ) = 3σ0 e 4κ0 ,
κ (t ) = κ0 , σv (t ) = σ0 eθ0 sin (2π t ) ,
for t ∈ [0, T ] where we set κ0 ∈ (0, 1 ), θ0 ∈ (−1, 1 ), σ0 ∈ (0, 0.1 ), and the correlation coefficient ρ ∈ (−1, 1 ). With these
parameter functions, one can verify that Assumptions 1 - 2 are fulfilled such that δ (t ) = 3 ≥ 2 for all t ∈ [0, T ].
We consider volatility swaps initiated at t1 = 0, maturity date T = 1, sampling frequency N = 252, and the initial variance
v0 ∈ (0, 1 ). From the approximate formula (4.40), KˆvHol can be viewed as a function of v0 when the values of κ0 , θ0 , σ0 , and
ρ are given. To make sure that readers have some quantitative concept of the difference between the results from our
approximate formula (4.40) and those from the MC simulations, we compute the percentage relative difference of the two
volatility swap prices (hereafter denoted as PRD), using our volatility swap price as the reference in the calculation, for
all sample points and interpolate by using contours over a considering domain. In our numerical tests, we consider the

21
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Fig. 9. The convergence of the MC volatility prices to the true value of the conditional expectation on the RHS of (4.22) compared with KˆvHol obtained from
our approximate formula (4.40) for three different sample points in Dv0 ,θ0 (The convergence shown in (b) can be considered as a particular case when the
Heston-ECIR model (4.21) reduces to the Heston-CIR model (5.2).).

PRD of the two volatility swap prices over the domain determined by the key parameter4 θ0 as follows: (v0 , θ0 ) ∈ Dv0 ,θ0 :=
(0, 1 ) × (−1, 1 ) where we fix κ0 = 0.05, σ0 = 0.07, and ρ = 0.5.
According to the numerical results presented in [26,29,33], we set the number of terms in the infinite series on the

RHS of (4.29) used to approximate EtQ [ vs ] to be 2. To obtain the value of KˆvHol , the Gaussian quadrature rule is adopted to
1
approximate the values of the integrals in (4.38) - (4.39). For the domain Dv0 ,θ0 , 81 sample points are uniformly chosen to
obtain two sets of volatility swap prices; one is from our approximate formula (4.40) and the other one is from N p = 10, 0 0 0
in MC simulations. We compute the PRDs, using our volatility swap price as the reference in the calculation, for all sample
points and interpolate by using contours over Dv0 ,θ0 .
As shown in Fig. 8, the maximum of the PRDs is less than 0.45% in Dv0 ,θ0 . It should be noticed from the parameter
functions used in this example that the Heston-ECIR model (4.21) reduces to the Heston-CIR model when θ0 = 0 as follows:

√ 
dSt = r (t )St dt + vt St dWtS
√ (5.2)
dvt = κ0 (θ1 − vt )dt + σ0 vt dWtv

3σ 2
for S0 , v0 > 0 where θ1 = 4κ0 . The PRDs obtained under this particular case are less than 0.30% as shown in Fig. 8 by the
0
values on the dashed green line.
To demonstrate the convergence of the MC volatility swap price to the true value of the conditional expectation on the
RHS of (4.22), we follow the procedure presented in Example 3. The convergences of the three sequences of the MC volatility
swap prices corresponding to the three sample points are shown in Fig. 9 (a)-(c) for Dv0 ,θ0 . We clearly see that the limits
are very close to the approximate values obtained from our approximate formula (4.40). Therefore, the numerical results

4
In our numerical study, we mainly focused on the accuracy and efficiency of KˆvHol with respect to θ0 because this allows us to study a special case such
that the Heston-ECIR model (4.21) reduces to the Heston-CIR model (5.2) when θ0 = 0. We also investigated the accuracy and efficiency of KˆvHol with respect
to κ0 , σ0 and ρ , but we omit to present our results in this paper for brevity reasons.

22
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

 
Fig. 10. The absolute errors e j,N = Kˆvol1 , j,N − Kvol1 , j,N  obtained from (2.29) by setting r (t ) = r j (t ), j = 1, 2, 3, respectively, with three different sampling
frequencies.

obtained ensure the approximate formula (4.40) produces volatility swap prices which are very close to the true values of
KvHol .
Let us discuss the efficiency of Kˆ H over the MC simulations. Although our approximate formula (4.40) requires the
vol
utilisation of a numerical integration technique to compute the integrals (4.38) - (4.39), the MC simulations take a much
longer time than our approximate formula does. For example, to obtain a volatility swap price from one sample point in the
domains, our numerical experiments show that the MC simulations with N p = 10 0, 0 0 0 consume 1901.6532 seconds, while
implementing formula (4.40) just takes 0.3321 seconds: a roughly 5,726 thousands folds of reduction in computational time
for one data point.

5.2. Accuracy of the convexity correction

Example 5. (Accuracy of Kˆvol1 )


In this example, we present some numerical results concerning the accuracy  of the convexity correction as previously
discussed in Section 2.4 as follows. Firstly, we define the absolute error e j,N := Kˆvol1 , j,N − Kvol1 , j,N  obtained from (2.29) by
setting r (t ) = r j (t ), j = 1, 2, 3, given in Example 1. We set T = 1 and assume that the price volatility σ0 varies on (0,30), and
sampling frequency N = 12, 52, 252, representing every month, every week, and every trading day, respectively.
By fixing σ0 , we clearly see from Fig. 10 (a)-(c) that e j,N for j = 1, 2, 3, tend to zero when N increases; demonstrating
that the convexity correction formula (2.27) provides a good approximation for Kvol1, j,N when the sampling frequency is
sufficiently large. In particular, the absolute errors are very close to zero when the sampling frequency is every trading day.
Next, we consider the case that the sampling frequency is every trading day. Notice that r1 (t ) < r2 (t ) < r3 (t ) for all
t ∈ [0, T ]. One may claim that e1,252 < e2,252 < e3,252 for all σ0 > 0. However, our numerical results shown in Fig. 11 clarify
that it is not always true. Therefore, market practitioners should be aware of this point when they adopt the convexity
correction.

23
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

 
Fig. 11. A comparison between the absolute errors e j,252 = Kˆvol1 , j,252 − Kvol1 , j,252  obtained from (2.29) by setting r (t ) = r j (t ), j = 1, 2, 3, respectively, where
the sampling frequency is every trading day.

6. Conclusions

In this paper, we have applied the properties of noncentral chi random variables to derive an analytical pricing formula
for discretely-sampled volatility swaps based on the B-S model with time varying risk-free interest rate. We have found
that the log-return realized volatility is a linear transformation of a noncentral chi random variable. As a result, analytical
pricing formulas for pricing discretely-sampled volatility and variance options as well as their properties have been found
and intensively studied. Furthermore, we have clarified the accuracy of the convexity correction for approximating volatil-
ity swap prices in the case of discrete sampling. To demonstrate some applications of our newly found analytical pricing
formulas, we have constructed simple closed-form approximate formulas for the B-S model with time varying parameters
and the Heston-ECIR model. Through numerical experiments, we have shown that our approximate formulas can improve
the accuracy, including efficiency in pricing volatility swaps such that the computational efficiency of our approach is enor-
mously enhanced in terms of assisting practitioners to price volatility swaps compared with the MC simulations. Clearly,
our newly found analytical pricing formulas as well as approximate formulas proposed in the paper can be very useful tools
for practitioners and researchers when there is obviously increasing demand of trading volatility derivatives in the futures
markets.

Declaration of Competing Interest

No conflict of interest exists.

CRediT authorship contribution statement

Sanae Rujivan: Conceptualization, Methodology, Formal analysis, Investigation, Writing - original draft, Funding acquisi-
tion. Udomsak Rakwongwan: Software, Visualization, Validation, Writing - review & editing.

Acknowledgements

The first author gratefully acknowledges the financial support from Walailak University under grant WU 63250. The
constructive suggestions from the anonymous referees have substantially improved the readability and presentation of the
paper. All errors are the author’s own responsibility.

24
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Appendix A

Proofs of Theorems and Corollaries in Section 2:


Proof of Theorem 2.1
μi−1 +σ (Wti −Wti−1 )
Proof. From (1.1) and (2.1), we have Sti = Sti−1 e . Next, we set Zi = ln Sti − ln Sti−1 and we obtain

Zi = μi−1 + σ (Wti − Wti−1 ) ∼ N (μi−1 , σ t ) 2


(A.1)
for i = 2, . . . , N. In other words, Zi is a normal random variable with mean μi−1 and variance σ 2 t
for all i = 2, . . . , N.
Applying the property of Brownian motion such that the increments (Wt j − Wt j−1 ) ∼ N (0, t ) and (Wtk − Wtk−1 ) ∼ N (0, t )
for all j, k = 2, . . . , N and j = k are independent, we then obtain that the sequence of Zi , i = 2, . . . , N, is independent and
identically distributed (i.i.d.). As a result, a random variable defined by
  2
√ 
N
Zi
Y := √ (A.2)
i=2
σ t
√ according to the noncentral chi distribution with N − 1 degrees of freedom and noncentrality parameter λ,
is distributed
denoted by Y ∼ N C χN−1 (λ ), in which its mean can be found in explicit form as
 2
√ π ( N−1 −1 ) λ
EtQ1 [ Y ] = L1 2 − >0 (A.3)
2 2 2
N−1
μ2i
where λ =
i=1
√ (see Johnson et al. [16]). Note that the conditional expectation above is always positive due to the
σ t
well-known property called the explicitly of the Laguerre functions (see Mirevski and Boyadjiev [22]). From (1.2), one can
show that
 √
RVd = β Y (A.4)

where β = σ t × 100, and this immediately implies (2.2) by using (A.3). In addition, when the risk-free interest rate is
T
assumed to be constant, our formula for Kvol1 can be simplified to the one written in equation (40) of Lian et al. [18]. 

Proof of Corollary 2.2:


 √ √
Proof. We note from (A.4) that VARtQ [ RVd ] = β 2 VARtQ [ Y ]. Since Y ∼ N C χN−1 (λ ). Then, one can verify from the prop-
1 1 √
erty of a noncentral chi random variable such that VARtQ [ Y ] = (N − 1 ) + λ2 − β12 Kv2ol . As a result, we immediately obtain
1
√ 1
(2.3). Moreover, the probability density function (pdf) of Y is f1,√Y (y; N − 1, λ ). It should be also noted from (A.4) that
 √ 
RVd is a linear transformation of Y . Applying the Jacobian transformation, the pdf of RVd is in fact the integrand of the
integral on the RHS of (2.4), and hence, the value of the probability on the LHS of (2.4) can be obtain by computing the
integral. 

Proof of Theorem 2.3:



Proof. Since Y ∼ χN−1 and its mean can be found in explicit form as

√ √ N
EtQ1 [ Y] = 2
2
N−1
. (A.5)
2

Utilizing (A.4) and (A.5), we immediately obtain (2.6). 

Proof of Corollary 2.5:

Proof. By utilizing the continuity and property of the Laguerre functions at the origin such that
(1 + ν + α )
Lν(α ) (0 ) =
( 1 + ν ) (1 + α )
with ν = 12 and α = N−1
2 − 1 (see Mirevski and Boyadjiev [22]), the limit (2.10) is immediately obtained using the formulas
(2.2) and (2.6). 

Proof of Corollary 2.6:



i=1 μi = 0, from Theorem 2.3, we immediately obtain
Proof. In the case that N−1 2


∂ Kvol2 2t
N
=
2
 × 100 > 0. (A.6)
∂σ T N−1
2

25
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

N−1
On the other hand, if i=1 μ2i > 0, from Theorem 2.1, we consider the partial derivative:
  N−1 2    N−1 2 
∂ Kvol1 π t ( N−1
2 −1 )
μ ∂ ( N−1 −1 ) − i=1 μi
= L1 − i=12 i + σ L1 2 × 100. (A.7)
∂σ 2T 2 2σ t ∂σ 2 2σ 2 t

Using the property of Laguerre functions (see Mirevski and Boyadjiev [22]), we arrive
  N−1 2 
∂ ( N−1 −1 ) − i=1 μi d ( N−1 −1 ) du ( N−1 ) du
L1 2 = L1 2 (u ) = −L− 12 (u ) (A.8)
∂σ 2 2σ t
2 du 2 dσ 2 dσ
N−1
μ2i
where u = − i=1
2σ 2 t
. Using (2.1), the derivative on the RHS of (A.8) can be simplified to obtain
 
1 
N−1
du 1 N−1
= R2i − (t )2 σ . (A.9)
dσ t σ3 4
i=1

Inserting (A.9) into (A.8) and (A.7) and simplifying the result obtained yield (2.11). 

Proof of Theorem 2.7:


 
Proof. Using the identity VARtQ [ RVd ] = EtQ [RVd ] − (EtQ [ RVd ] )2 and the results as obtained in (2.3) and (2.7), we arrive
1 1 1

EtQ1 [RVd ] − (EtQ1 [ RVd ] )2 = Kvar j − Kv2ol j = β 2 (N − 1 + (2 − j )λ2 ) − Kv2ol j (A.10)

for j = 1, 2, and this immediately implies (2.14). 

Proof of Corollary 2.8:



Proof. Due to Y ∼ N C χN−1 (λ ), it is clear by using the property of the noncentral chi distribution that Y ∼ N C χ 2 (λ2 ) in
N−1
which the pdf of Y is f1,Y (y; N − 1, ν ). By analogy with the proof of Corollary 2.2, we apply the Jacobian transformation to
RVd and obtain that its pdf is in fact the integrand of the integral on the RHS of (2.16). Hence, the value of the probability
on the LHS of (2.16) can be obtain by computing the integral on the RHS of (2.16). 

Proof of Theorem 2.11:

Proof. 1. Since r (s ) is continuous on [t1 , T ], we apply the Mean Value Theorem for the integral in (2.1) to obtain that
μi = (r (si ) − 12 σ 2 )t for some si ∈ [ti , ti−1 ] for all i = 1, . . . , N − 1. For j = 1, we compute the following limit:
 N−1 
i=1μ2i 1 
N−1
1 T
lim = lim ( r ( si ) − σ 2 )2
N→∞ σ t
2 2 2σ 2 N→∞
i=1
2 N−1
 T
1 1
= (r (s ) − σ 2 )2 ds := p > 0. (A.11)
2σ 2 t1 2
From (2.2), (A.11), and using the continuity of Laguerre function, we arrive

T π pk ( N2 − 1 )!
lim Kvol1 , N, T ; r (t ), σ =σ lim √ × 100. (A.12)
N→∞ N−1 2
k=0
k!( − k )! N→∞ N − 1( N2 − 32 + k )!
1
2

( N −1 )! 
2
Consider the summation on the RHS of (A.12) for k = 0, one can show that limN→∞ √2 = π , otherwise, the
(1/2 )! N−1( N2 − 32 )!
limits vanish for all k ≥ 1. Hence, we immediately obtain (2.23) for j = 1. For j = 2, we use (2.6) to obtain that

T √ N
lim Kvol2 , N, T ; r (t ), σ = σ 2 lim √
2
 × 1002 . (A.13)
N→∞ N−1 N→∞ N−1 N−1
2

N

Consider the limit on the RHS of (A.13), one can use the property of Gamma function to show that limN→∞ √ 2  =
N−1 N−1
√ 2
1/ 2 and then we now obtain (2.23) for j = 2.
2. For j = 2, by the formula of Kvar2 as expressed in (2.14), it is easy to obtain the limit in (2.24). Similarly, for j = 1, by
 
N−1
μ2i
the formula of Kvar1 as expressed in (2.14) and using the fact that limN→∞ i=1
T = 0, one can easily deduce the limit

in (2.24). 

Proof of Theorem 2.12:

26
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Proof. By applying (2.14), (2.18), and (2.21) to (2.27) and simplifying the results obtained for j = 1, 2, we have
 2 
λ2
1 + N−1 − 1 1 λ
+ 2 (N−1
2

4 N−1 )2
Kˆvol1 = σ  32 × 100 (A.14)
λ2
1 + N−1
and
1
Kˆvol2 = σ 1 − × 100 (A.15)
4 (N − 1 )
N−1
μ2i
where λ = and μi for i = 1, . . . , N − 1, are defined in (2.1).
i=1

σ t
From (A.11), we have limN→∞ λ2 = 2 p > 0 and this implies that
 2 
λ2
1 + N−1 − 1 1 λ
+ 2 (N−1
2

4 N−1 )2
lim  32 = 1. (A.16)
N→∞ λ2
1 + N−1
Applying (A.16) to (A.14) for the case j = 1 and computing the limit as N approaches infinity on the RHS of (A.15) for the
case j = 2, we now obtain the limit as N approaches infinity of the convexity correction (2.27) as
lim Kˆvol j = σ × 100 (A.17)
N→∞

for j = 1, 2. Using (2.23) and (A.17), we immediately obtain the limit in (2.28). In addition, (2.29) and (2.30) can be easily
deduced using the analytical formulas of Kvol j , j = 1, 2, as written in (2.2) and (2.6), respectively. 

Appendix B

Proofs of Theorems and Corollaries in Section 3: Proof of Theorem 3.1:



Proof. For j = 1, 2, the pdf of RVd (t1 , N, T ) is the integrand of the integral on the RHS of (2.4) and (2.8), respectively.
Hence,
  ∞
1 y
EtQ1 [( RVd (t1 , N, T ) − Kvcol )+ ] = (y − Kvcol ) f1,√Y ; N − 1, λ dy (B.1)
Kvcol β β
for j = 1, and
  ∞
1 y
EtQ1 [( RVd (t1 , N, T ) − Kvcol )+ ] = (y − Kvcol ) f2,√Y ; N − 1 dy (B.2)
Kvcol β β
for j = 2. Rearranging the domains of the integrals on the RHS of (B.1) and (B.2), the pricing formula (3.4) can be easily
deduced. 

Proof of Corollary 3.2:

Proof. It is easy to obtain (3.8) by differentiating the RHS of (3.5) with respect to σ and applying Corollary 2.6. To determine
N−1
∂ Ivol μ2i (σ )
t × 100 and λ(σ ) :=
∂σ , j = 1, 2, we write β (σ ) := σ as functions of σ where μi (σ ) is defined in (2.1).
j i=1

T σ t
Differentiating (3.6)-(3.7) with respect to σ yields

∂ Ivol1 Kvcol
1 y
= (Kvcol − y ) f1,√Y ; N − 1, λ(σ ) F1,√Y (y )dy (B.3)
∂σ 0 β (σ ) β (σ )
and

∂ Ivol2 Kvcol
1 y
= (Kvcol − y ) f2,√Y ; N − 1 F2,√Y (y )dy (B.4)
∂σ 0 β (σ ) β (σ )
where

(y2 − (N − 1 )β 2 (σ ))β (σ ) yI N−1 yβλ((σσ))
F1,√Y (y ) = − λ ( σ )λ ( σ ) + 2
 (β (σ )λ (σ ) − λ(σ )β (σ )) (B.5)
β 3 (σ ) β 2 (σ )I N−3 yλ(σ )
β (σ )
2

and
(y2 − (N − 1 )β 2 (σ ))β (σ )
F2,√Y (y ) = (B.6)
β 3 (σ )
providing that β (σ ) and λ (σ ) are the first derivatives of β (σ ) and λ(σ ) with respect to σ , respectively. 

27
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

Proof of Corollary 3.3:


∞
Proof. From (3.4), (3.9) is trivial. Using the property of the pdf f such that 0 f (y )dy = 1, (3.10) can be easily derived.
For j = 1, the first and second derivatives of Cvol1 with respect to the strike price K can be analytically derived by using
(3.5)-(3.7) as

∂ K
1 y
C (t , t, N, T ; r (t ), σ , K ) = −1 + f1,√Y ; N − 1, λ dy (B.7)
∂ K vol1 1 0 β β
and
∂2 1 K
C (t , t, N, T ; r (t ), σ , K ) = f1,√Y ; N − 1, λ > 0 (B.8)
∂ K 2 vol1 1 β β
K
for all K > 0 and (B.8) implies (3.12). Since F (K ) = 0 β1 f1,√Y βy ; N − 1, λ dy is the cumulative distribution function (cdf)

of Y . This implies that 0 < F (K ) < 1 for all K > 0. As a result, we immediately obtain (3.11) and similar results can be
obtained for j=2. The inequalities contained in (3.13) can be easily deduced by using (3.9), (3.10), and (3.11). 

Proof of Theorem 3.4:

Proof. For j = 1, 2, the pdf of RVd (t1 , N, T ) is the integrand of the integral on the RHS of (2.16) and (2.19), respectively.
By analogy with the proof of Theorem 3.1, the pricing formula (3.14) can be obtained by rearranging the domains of the
integrals. 

Proof of Corollary 3.5:

Proof. By differentiating the RHS of (3.14) with respect to σ and applying (2.22), we immediately obtain (3.17). To derive
∂ Ivar j
∂σ , j = 1, 2, we follow the method proposed in the Proof of Corollary 3.2 by differentiating (3.15)-(3.16) with respect to σ
to obtain
  
∂ Ivar1 Kvcar
1 y √ 
= ( Kvcar − y) f1,Y ; N − 1, λ (σ ) F1,√Y
2
y dy (B.9)
∂σ 0 β 2 (σ ) β 2 (σ )
and
  
∂ Ivar2 Kvcar
1 y √ 
= ( Kvcar − y) f2,Y ; N − 1 F2,√Y y dy (B.10)
∂σ 0 β 2 (σ ) β 2 (σ )
where F1,√Y (y ) and F2,√Y (y ) are given in (B.5) and (B.6), respectively. 

Appendix C

Proofs of Theorem and Corollary in Section 4:


Proof of Corollary 4.1:

N−1 √ N Z̄
2
Proof. Suppose that i=1 μ̄2i > 0. Since (4.6) holds. Then, we have Y = i=2
√i
σ̄ t
∼ N C χN−1 (λ ) where λ =
N−1
μ̄2i
. By analogy to the proof of Theorem 2.1, Kvσol(t ) can be obtained by using (2.2) with replacing μi by μ̄i and σ
i=1

σ̄ t  √
by σ̄ . On the other hand, when N−1
i=1 μ̄i = 0,
2 Y ∼ χN−1 . By analogy to the proof of Theorem 2.3, one can easily show that
Kvσol(t ) can be computed by using (2.6) with replacing σ by σ̄ . 

Proof of Theorem 4.2:

Proof. From (4.11), the following inequality holds


 
1 
N  1 
N
0< Z̄i2 ≤ Ȳ ≤ Z̄i2 (C.1)
σ̄max i=2
σ̄min i=2

almost surely with respect to the risk-neutral martingale measure Q. This implies
 
 
N 
0 < σ̄ Q
min Et1 [ Ȳ ] ≤ EtQ1 Z̄i2 ] ≤ σ̄max EtQ1 [ Ȳ ]. (C.2)
i=2
  
 N
According to (4.2), we have Kvσol(t ) = EtQ [ RVd ] = √1 EtQ i=2 Z̄i2 ] × 100. Since Ȳ ∼ N C χN−1 (λ̄ ). This implies
1 1 T 1
 π N−1 −1

EtQ [ Ȳ ] = − λ̄2
2 2
2 L1 . Inserting these results into (C.2), the inequality (4.13) can be obtained after a simplification.
1
2
Moreover, the error estimate (4.16) is immediately obtained by applying the inequalities (4.13) and (4.15). 

28
S. Rujivan and U. Rakwongwan Commun Nonlinear Sci Numer Simulat 100 (2021) 105849

References

[1] Bates DS. Jumps and stochastic volatility: exchange rate processes implicit in deutsche mark options. Rev Financ Stud 1996;9(1):69–107.
[2] Black F, Scholes M. The pricing of options and corporate liabilities. J Polit Econ 1973;81:637–54.
[3] Broadie M, Jain A. The effect of jumps and discrete sampling on volatility and variance swaps. Int J Theor Appl Finance 2008;11(8):761–97.
[4] Brockhaus O, Long D. Volatility swaps made simple. Risk 20 0 0;2(1):92–5.
[5] Chunhawiksit C, Rujivan S. Pricing discretely-sampled variance swaps on commodities. Thai J Math 2016;14(3):711–24.
[6] Derman E, Kani I. The volatility smile and its implied tree. Quantitative strategies research notes 1994. http://www.ederman.com/new/docs/
gs-volatility_smile.pdf
[7] Drimus G, Farkas W, Gourier E. Valuation of options on discretely sampled variance: a general analytic approximation. J Comput Finance
2016;20(2):39–66.
[8] Dupire B. Pricing with a smile. Risk 2007:126–9.
[9] Elliott RJ, G-H L. Pricing variance and volatility swaps in a stochastic volatility model with regime switching: discrete observation case. Quantit Finance
2013;13(5):687–98.
[10] Harrison M, Pliska R. Martingales and stochastic integrals in the theory of continuous trading. J Econ Theory 1981;11:215–60.
[11] X-J H, S-P Z. A series-form solution for pricing variance and volatility swaps with stochastic volatility and stochastic interest rate. Comput Math with
Appl 2018;76:2223–34.
[12] X-J H, S-P Z. Variance and volatility swaps under a two-factor stochastic volatility model with regime swithcing. Int J Theor Appl Finance
2019;22(4):1–19.
[13] Heston SL. Rev financ stud. A closed-form solution for options with stochastic volatility with applications to bond and currency options
1993;6(2):327–43.
[14] Hull J.C.. Options, futures, and other derivatives (tenth edition). 2018. New York: Pearson Education.
[15] Itkin A, Carr P. Pricing swaps and options on quadratic variation under stochastic time change models - discrete observations case. Rev Deriv Res
2010;13:141–76.
[16] Johnson N, Kotz S, Balakrishnan N. Continuous univariate distributions. New York: A Wiley-Interscience Publication; 1995.
[17] Keller-Ressel M, Muhle-Karbe J. Asymptotic and exact pricing of options on variance. Finance and Stoch 2013;17:107–33.
[18] Lian G, Chiarella C, Kalev PS. Volatility swaps and volatility options on discretely sampled realized variance. J Econ Dyn Control 2014;47:239–62.
[19] Liu W, S-P Z. Pricing variance swaps under the Hawkes jump-diffusion process. J Futures Mark 2019;39:635–55.
[20] Maghoodi Y. Solution of the extended CIR term structure and bond option valuation. Math Finance 1996;6:89–109.
[21] Merton RC. Theory of rational option pricing. Bell J Econ 1973;4(1):141–83.
[22] Mirevski SP, Boyadjiev L. On some fractional generalizations of the Laguerre polynomials and the Kummer function. Comput Math with Appl
2010;59:1271–7.
[23] Rujivan S, S-P Z. A simplified analytical approach for pricing discretely-sampled variance swaps with stochastic volatility. Appl Math Lett
2012;22:1644–50.
[24] Rujivan S, S-P Z. A simple closed-form formula for pricing discretely-sampled variance swaps under the heston model. ANZIAM J 2014;56:1–27.
[25] Rujivan S. A novel analytical approach for prcing discretely sampled gamma swaps in the heston model. ANZIAM J 2016;57:244–68.
[26] Rujivan S. A closed-form formula for the conditional moments of the extended CIR processes. J Comput Appl Math 2016;297:75–84.
[27] Rujivan S. Analytically pricing variance swaps in commodity derivative markets under stochastic convenience yields. Commun Math Sci
2021;19(1):111–46.
[28] Scott LO. Pricing stock options in a jump diffusion model with stochastic volatility and interest rates: applications of fourier inversion methods. Math
Finance 1997;7(4):413–26.
[29] Sutthimat P, Mekchayz K, Rujivan S. Explicit formula for conditional expectations of product of polynomial and exponential function of affine transform
of extended cox-ingersoll-ross process. J Phys Conf Ser 2019. 1132 Number: Article ID: 012083
[30] Swishchuk A. Modeling of variance and volatility swaps for financial markets with stochastic volatilities. Technical Article 2004. http://people.ucalgary.
ca/∼aswish/StochVolatSwap.pdf
[31] Swishchuk A, Li X. Pricing variance swaps for stochastic volatilities with delay and jumps. Int J Stoch Anal 2011. Article ID 435145
[32] D-X X, B-Z Y, J-H K, N-J H. Variance and volatility swaps valuations with stochastic liquidity risk. Physica A 2021;566:125679.
[33] Thamrongrat N, Rujivan S. A closed-form formula for the conditional expectation of the extended CIR process. Songklanakarin J Sci Technol
2020;42(2):424–9.
[34] Weraprasertsakun A, Rujivan S. A closed-form formula for pricing variance swaps on commodities. Vietnam J Math 2017;45:255–64.
[35] B-Z Y, Yue J, M-H W, N-J H. Volatility swaps valuation under stochastic volatility with jumps and stochastic intensity. Appl Math Comput
2019;355:73–84.
[36] Yuen CH, Zheng W, Kwok YK. Pricing exotic discrete variance swaps under the 3/2-stochastic volatility models. Appl Math Finance 2015;22(5):421–49.
[37] Zheng W, Kwok YK. Closed form pricing formulas for discretely sampled generalized variance swaps. Math Finance 2014;24:855–81.
[38] Zhu S-P, G-H L. A closed-form exact solution for pricing variance swaps with stochastic volatility”. Math Finance 2011;21:233–56.
[39] Zhu S-P, G-H L. On the valuation of variance swaps with stochastic volatility. Appl Math Comput 2012;219:1654–69.
[40] Zhu S-P, G-H L. On the convexity correction approximation in pricing volatility swaps and VIX futures. New Math Nat Comput 2018;14(3):383–401.

29

You might also like