This action might not be possible to undo. Are you sure you want to continue?
DISPLACED LOGNORMAL AND DISPLACED HESTON VOLATILITY SKEWS:
ANALYSIS AND APPLICATIONS TO STOCHASTIC VOLATILITY SIMULATIONS
A DISSERTATION SUBMITTED TO
THE FACULTY OF THE DIVISION OF THE PHYSICAL SCIENCE
IN CANDIDACY FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
DEPARTMENT OF STATISTICS
BY
DAN WANG
CHICAGO, ILLINOIS
JUNE 2010
To my parents
TABLE OF CONTENTS
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
PUBLICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 DISPLACED LOGNORMAL PROCESS . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Implied Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Displaced Lognormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Global behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.3 Atthemoney behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.4 Shortexpiry behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Displaced antiLognormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Global behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.3 Atthemoney behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.4 Shortexpiry behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 DISPLACED HESTON PROCESS . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Displaced Heston . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 Implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.2 Atthemoney behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.3 Shortexpiry behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Displaced antiHeston . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 Implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.2 Shortexpiry behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.3 Generalization of shortexpiry behavior . . . . . . . . . . . . . . . . . 20
iii
4 CALIBRATION OF DL AND DH PROCESS . . . . . . . . . . . . . . . . . . . . 22
4.1 Calibration of DL and DH Process . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.1 Calibration of DL process . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 Calibration of DH process . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Calibration DL and DH to CEV/SABR . . . . . . . . . . . . . . . . . . . . . 23
4.2.1 CEV and SABR stochastic volatility models . . . . . . . . . . . . . . 23
4.2.2 Calibration of DL to CEV/SABR . . . . . . . . . . . . . . . . . . . . 25
4.2.3 Calibration of DH to SABR . . . . . . . . . . . . . . . . . . . . . . . 26
5 VARIANCE REDUCTION IN MONTE CARLO SIMULATION . . . . . . . . . . 27
5.1 Variance Reduction Using Control Variate . . . . . . . . . . . . . . . . . . . 27
5.1.1 DL or DH as a control variate . . . . . . . . . . . . . . . . . . . . . . 28
5.1.2 Example I: Discretely sampled barrier option under CEV/SABR dy
namics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1.3 Numerical results I: Discretely sampled barrier option under CEV/SABR
dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.4 Example II: Discretely sampled arithmetic Asian option under SABR
dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.1.5 Numerical results II: Discretely sampled arithmetic Asian option under
SABR dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.2 Variance Reduction Combining Control Variate and Importance Sampling . . 35
5.2.1 Importance sampling on options pricing . . . . . . . . . . . . . . . . 35
5.2.2 Drifted DH/DL process . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2.3 Combine control variate with importance sampling . . . . . . . . . . 37
5.2.4 Example: Discretely sampled barrier option under SABR dynamics . 38
5.2.5 Numerical results: Discretely sampled barrier option . . . . . . . . . 40
6 DISCRETISATION SCHEME . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.1 Partial Strong Convergency of Stochastic Volatility Process . . . . . . . . . . 44
6.2 Strong Convergence of Meanreverting CEV Process . . . . . . . . . . . . . . 46
6.3 Discretisation Schemes Used in the Monte Carlo Simulation . . . . . . . . . 48
7 LARGEEXPIRY IMPLIED VOLATILITY OF DISPLACED LOGNORMAL . . 50
7.1 Largestrike and Largeexpiry Behavior . . . . . . . . . . . . . . . . . . . . . 50
7.1.1 Case one: K = S
0
e
xT
, x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
] . . . . . . . . . . . . . . . 50
7.1.2 Case two: K = S
0
e
xT
, x ∈ (−σ
2
/2, σ
2
/2) . . . . . . . . . . . . . . . 52
7.1.3 Case three: K = S
0
e
xT
α
. . . . . . . . . . . . . . . . . . . . . . . . . 55
7.2 Fixedstrike Largeexpiry Implied Volatility . . . . . . . . . . . . . . . . . . . 56
8 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Appendix
iv
A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
A.1 Appendix: Proof of Theorem 1–Global Behavior of Displaced Lognormal . . 63
A.2 Appendix: Proof of Theorem 2–Atthemoney Behavior of Displaced Lognormal 69
A.3 Appendix: Proof of Theorems 3 and 6–Shortexpiry Behavior of DL . . . . . 70
A.4 Appendix: Proof of Theorem 4–Global Behavior of Displaced anti Lognormal 71
A.5 Appendix: Proof of Theorem 5–Atthemoney Behavior of Displaced anti 
Lognormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
B.1 Appendix: Proof of Theorem 7–Atthemoney Behavior of Displaced Indepen
dent Stochastic Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
B.2 Appendix: Proof of Theorems 8 and 9– Shortexpiry Behavior of DH . . . . 78
B.3 Appendix: Proof of Propositions 3.1.2 and 3.2.1–Level, Slope and Convexity
of DH Shortexpiry Implied Volatility . . . . . . . . . . . . . . . . . . . . . . 80
C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
C.1 Appendix: Proof of Theorem 12–Partial Strong Convergency of Stochastic
Volatility Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
C.2 Appendix: Coeﬃcients of SABR Satisfy the Local Lipschtiz Condition (*) . . 88
C.3 Appendix: Proof of Proposition 6.2.2 and Theorem 13–Strong Convergence
of Meanreverting CEV Process . . . . . . . . . . . . . . . . . . . . . . . . . 89
D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
D.1 Appendix: Proof of Theorem 14 –Largestrike and Largeexpiry Asymptotic
of Displaced Lognormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
D.2 Appendix: Proof of Theorem 15–First Order Approximation of Largestrike
Largeexpiry Implied Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . 98
D.3 Appendix: Proof of Theorem 16–Asymptotic Formula of BlackScholes Call
Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
D.4 Appendix: Proof of Theorems 17 and 18– Second and Third Order Approxi
mation of Largestrike Largeexpiry Implied Volatility . . . . . . . . . . . . 101
D.5 Appendix: Proof of Theorem 19–Largestrike Largeexpiry Atthemoney Im
plied Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
D.6 Appendix: Proof of Theorem 20–Approximation of Implied Volatility when
K = S
0
exp(xT
α
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
D.7 Appendix: Proof of Theorem 21–Fixedstrike Largeexpiry Implied Volatility 103
D.7.1 Asymptotic behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
D.7.2 Monotonicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
D.8 Appendix: Proof of Theorem 22–Approximation of FixedStrike Largeexpiry
Implied Volatility. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
v
LIST OF FIGURES
2.2.1 Theorem 3 formula and Exact σ
imp
for T = 1.0. . . . . . . . . . . . . . . . 8
2.3.1 Theorem 5.1 bounds of σ
atm
for T = 3.0. . . . . . . . . . . . . . . . . . . 11
3.2.1 Theorem 9 formula and exact σ
imp
. . . . . . . . . . . . . . . . . . . . . . . 19
7.1.1 Theorem 14 formula and Exact σ
imp
for T = 10. . . . . . . . . . . . . . . 51
7.1.2 Theorem 17 formula and Exact σ
imp
. . . . . . . . . . . . . . . . . . . . . 53
7.1.3 Theorems 17 and 18 formula and Exact σ
imp
. . . . . . . . . . . . . . . . . 58
7.1.4 Theorems 17 and 18 formula and Exact σ
imp
. . . . . . . . . . . . . . . . . 58
7.1.5 Theorem 20 formula and Exact σ
imp
for T = 10. . . . . . . . . . . . . . . 59
7.1.6 Theorem 20 formula and Exact σ
imp
for T = 7. . . . . . . . . . . . . . . . 59
7.2.1 Theorem 22 formula and Exact σ
imp
for T = 20. . . . . . . . . . . . . . . 60
vi
LIST OF TABLES
5.1.1 Percentage reduction of variance, using DL for downandout call on CEV 32
5.1.2 Percentage reduction of variance, using DL for downandout call on SABR 32
5.1.3 Percentage reduction of variance, using DH for downandout call on SABR 32
5.1.4 Percentage reduction of variance, using DL as control for Asian Option on
SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.1.5 Percentage reduction of variance, using LN as control for Asian Option on
SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2.1 Percentage reduction of variance, using importance sampling for downand
out call on SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2.2 Percentage reduction of variance, using ISDL and ISDH for downandout
call on SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2.3 Percentage reduction of variance, using importance sampling for downand
in call on SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2.4 Percentage reduction of variance, using ISDL and ISDH for downandin call
on SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
vii
ABSTRACT
We analyze the displaced (anti)lognormal (DL) and displaced (anti)Heston (DH) volatility
skew. In particular, for the displaced lognormal, we prove the global monotonicity of the
implied volatility, and an atthemoney bound on the steepness of the downward volatility
skews, which therefore cannot reproduce some features observed in the equity market. A
variant, the displaced antilognormal, overcomes this steepness constraint, but its state space
is bounded above and unbounded below. We prove the global monotonicity of its implied
volatility too. For the displaced Heston dynamics, we show that the atthemoney slope has
the same sign as the displacement. What’s more, we give an explicit formula for the DL and
DH’s shortexpiry limiting volatility skew, which allows direct calibration of their parameters
to volatility skews implied by market data or by other models. In the end, we analyze the
largeexpiry limiting volatility of the displaced lognormal and give an asymptotic formula of
it in the region of largestrike and ﬁxedstrike respectively.
We propose using the DL/DH dynamics as a control variate, to reduce variance in Monte
Carlo simulations of the CEV and SABR local/stochastic volatility models. We give simula
tion results to show that a carefully constructed control variate can signiﬁcantly reduce the
variance in the Monte Carlo simulations. We further propose a combination of the impor
tance sampling and the control variate to reduce the variance. Numerical simulations show
that signiﬁcant variance reduction can be achieved.
Finally we discuss the convergency of the discretisation schemes of the stochastic pro
cesses encountered in the Monte Carlo simulations. Under some regularity conditions, we
give a partial strong convergency result for the stochastic volatility process. Moreover, we
give a strong convergency result for the meanreverting CEV process.
viii
ACKNOWLEDGEMENTS
First and foremost, I am deeply grateful to my advisor, Professor Roger Lee, for sharing
with me his insights and wisdom, for his instructive guidance and invaluable advice, for his
constant support and encouragement over the past few years. Without him, this dissertation
would not be possible.
I am grateful to Professor Steven Lalley and Professor Per Mykland for being on my
committee and for their valuable suggestions and comments. In addition, I thank Professor
Per Mykland for many inspiring discussion on topics of application of Statistics to Finance.
Also, I would like to express my sincere thanks to all the faculty, stuﬀ and students
of the Department of Statistics at the University of Chicago. Because of them, I have a
wonderful education here. I feel fortunate to study and pursue my doctoral degree in such
an intellectually stimulating and friendly environment.
I would also like to thank my friends who made my time so enjoyable that I will al
ways cherish my time here. They are my source of advice and discussion. Particularly, I
would like to thank Xiaohui, Han, Changgee, Wenlong, William, Han, Yibi, Yingying, Oli,
Omar, Baoguan, Zuoheng, Christina, Dale, Winfried, Mike, Minsun and Darongsae. I thank
Nathaniel for proof reading my thesis.
Finally, to my parents, to whom this thesis is dedicated, goes my deepest appreciation.
I thank them for their love, care and support over the years. Words cannot express the
gratitude I owe them. I owe a special thanks to my dearest husband, Zhuo, for his love and
care, for his understanding, support, and for accompanying me through this work.
ix
PUBLICATION
Part of the thesis has been published in Lee and Wang (2009), which include the following
materials: Chapter 2.1, 2.2, 2.3.4, Chapter 4.1.1, 4.2.1, 4.2.2, and the parts related to the
DL process in Chapter 5.
x
CHAPTER 1
INTRODUCTION
Given an empiricallyobserved or modelgenerated price of a call or put, the implied volatility
is by deﬁnition the volatility parameter for which the BlackScholes formula recovers the
given option price. Regardless of what process actually underlies the given option price, the
implied volatility provides a canonical language or scale by which option prices are commonly
quoted and compared. At any expiry, the volatility skew – meaning the implied volatility as
a function of all strikes – captures the full riskneutral underlying distribution at that expiry,
and hence constitutes a natural framework to understand and to compare distributions.
In particular, comparison can occur between an empiricallyobserved volatility skew and
a modelgenerated volatility skew, for the purpose of calibrating the model’s parameters, or
for the purpose of understanding what empirical features can or cannot be reproduced by
the model. Comparison can also occur, between volatility skews generated by two diﬀerent
models, for the purpose of approximating the features of a more complex model, using a
simpler model.
The lognormal (BlackScholes 1973) model generates a ﬂat implied volatility skew, which
does not agree with the sloping skews observed empirically in equity, FX, and interest
rate markets. Displacing the lognormal (Rubinstein 1983) does generate a sloping implied
volatility skew. Marris (1999), Brigo and Mercurio (2002), Joshi and Rebonato (2003),
and SvobodaGreenwood (2009), have investigated the displaced lognormal (and extensions
thereof), as a pricing model or as a analytical approximation to other models, motivated
largely by applications to interest rate derivatives. In contrast, we draw motivation mainly
from problems arising in equity markets, such as how to calibrate to volatility skews that
slope downward more steeply than all displaced lognormal skews; and we intend to use the
calibrated process less for its analytical pricing than for its applicability to Monte Carlo
pricing.
First, we bound the level and slope of the implied volatility skews generated by dis
placed lognormal diﬀusions in various regimes (global, or atthemoney, or shortexpiry).
1
We prove, among other results, the global monotonicity of implied volatility, and an atthe
money upper bound on the absolute slope of downward volatility skews, under displaced
lognormal dynamics, which therefore cannot model some features (nonmonotonicity and a
steep downward slope) observed in equity market volatility skews. A variant, the displaced
anti lognormal, overcomes the steepness constraint, but its state space is bounded above
and unbounded below, unlike stock prices. For the displaced anti lognormal, we prove the
global monotonicity of implied volatility, and an atthemoney bound for the level and slope
of the implied volatility.
The Heston model (Heston 1993) generates the volatility smile and has an analytical
solution for call and put options. We propose a displaced (anti)Heston diﬀusion and analyze
its implied volatility. We show that the slope of its atthemoney implied volatility has the
same sign as the displacement and its shortexpiry implied volatility lacks the skewness
observed in the equity market; while the displaced anti Heston diﬀusion overcomes the
skewness problem, its state space is bounded above and unbounded below.
For both the displaced (anti)lognormal (DL) and the displaced (anti)Heston (DH) dif
fusions, we ﬁnd an explicit formula for their shortexpiry limiting volatility skew. This helps
to calibrate the model to the empirical data or other models. Moreover, for any general
process which is a positive martingale, and its corresponding (anti)displaced process, we
ﬁnd an explicit formula between the shortexpiry implied volatilities of the two processes.
In light of these restrictions on what features the DL and the DH can model, we then
exploit the DL or DH, not as a model, but as a control variate, to reduce variance in Monte
Carlo simulation of other models, such as the CEV and SABR local/stochastic volatility
models. Fisher and Tataru (2010) state that “During the past few years, practitioners have
settled on a consensus of using a mixed stochastic/local volatility (SLV) model as the market
standard for pricing barrier options.” CEV/SABR belong to the SLV family of models so
pricing options under them is of practical importance. We therefore consider pricing barrier
option and Asian option under the CEV/SABR model. We give numerical examples which
illustrate signiﬁcant variance reductions, when the control variate on DL/DH is carefully
constructed. What’s more, we explore the combination of the control variate and other
variance reduction technique, particularly importance sampling. We give numerical examples
to show the signiﬁcance of the variance reduction.
In the Monte Carlo simulation, most of the stochastic processes can not be exactly
sampled. Thus we refer to discretisation schemes to conduct the Monte Carlo simulation.
2
We discuss the conditions which guarantee the convergence of the discretized process to the
underlying continuous process. These convergence results serve as theoretical foundations
for the Monte Carlo simulation.
Finally, we analyze the largeexpiry implied volatility of displaced lognormal dynamics.
We give explicit approximation formula of the implied volatility in the largestrike large
expiry case and the ﬁxedstrike largeexpiry case.
Chapter 2 discusses the features of the implied volatility of the DL diﬀusion. Chapter 3
discusses the features of the implied volatility of the DH diﬀusion. Chapter 4 illustrates how
to calibrate the parameters of the DL and DH processes. Chapter 5 discusses how to use DL
or DH processes to construct control variate to reduce variance in Monte Carlo simulations
of option pricing; it also discusses how to combine the control variate with the importance
sampling to achieve variance reduction. Chapter 6 discusses the discretisation schemes used
in the Monte Carlo simulations. Chapter 7 discusses the largeexpiry behavior of the implied
volatility of displaced lognormal dynamics. Appendix contains most of the proofs.
3
CHAPTER 2
DISPLACED LOGNORMAL PROCESS
2.1 Implied Volatility
We work under martingale measure, and we either assume zero interest rates, or stipulate
that all prices are quoted as forward prices.
Our deﬁnition of the implied volatility skew will refer to the function C
BS
, speciﬁed as
follows.
Deﬁne C
BS
: R
3
∗
× R
+
→ R and P
BS
: R
3
∗
× R
+
→ R, where R
∗
:= R\{0} and
R
+
:= (0, ∞), by
C
BS
(s, k, Σ, T) := sN(d
+
) −kN(d
−
) (2.1)
P
BS
(s, k, Σ, T) := kN(−d
−
) −sN(−d
+
) (2.2)
d
±
:=
log(s/k)
Σ
√
T
±
Σ
√
T
2
(2.3)
where N denotes the standard normal CDF. We may suppress the last argument (T) of C
BS
.
Deﬁnition 1 (Implied volatility). Let X be a process with X
0
> 0. For all positive K, T
such that
(X
0
−K)
+
< E(X
T
−K)
+
< X
0
, (2.4)
deﬁne the implied volatility of X at (K, T) to be the σ
imp
> 0 such that
C
BS
(X
0
, K, σ
imp
, T) = E(X
T
−K)
+
. (2.5)
Write σ
X
imp
(K, T) for this implied volatility, which is welldeﬁned, because C
BS
(X
0
, K, ·, T)
is strictly increasing on R
+
, and has range ((X
0
−K)
+
, X
0
).
We refer to the function σ
X
imp
(·, T) as the implied volatility skew of X at expiry T.
4
2.2 Displaced Lognormal
Deﬁnition 2. A process S follows displaced lognormal dynamics, with displacement θ ∈ R,
if
dS
t
= σ(S
t
−θ)dW
t
, S
0
> θ, σ > 0, (2.6)
where W is Brownian motion.
Thus S − θ is a driftless geometric Brownian motion with volatility σ, and the interval
of points attainable by S is (θ, ∞). If modeling a nonnegative underlying such as a stock
price, this model for θ < 0 will misprice deepoutofthemoney puts, due to the possibility
of S
T
< 0. This model has further limitations, even for atthemoney contracts, as we will
see later.
For K > θ, a Kstrike Texpiry European call option on a displaced lognormal S has
price
E(S
T
−K)
+
= E((S
T
−θ) −(K −θ))
+
= C
BS
(S
0
−θ, K −θ, σ, T). (2.7)
In general, payoﬀs invariant to parallel shifts of the S path and the contract parameters can
be priced using BlackScholes model valuation methods, but applied to displaced arguments.
2.2.1 Implied volatility
Let us apply Deﬁnition 1 to the case that X = S, a displaced lognormal. If θ ≥ 0 then
(2.4) holds for all K, T positive. If θ < 0, then the ﬁrst inequality in (2.4) holds for all K, T
positive, but the second inequality E(S
T
−K)
+
< S
0
may fail for small positive K, due to
the nonzero probability that S
T
< 0. We therefore take care to deﬁne the strike interval on
which implied volatility exists. For displaced lognormal S, let
K
S
(T) := {K > θ
+
: C
BS
(S
0
−θ, K −θ, σ, T) < S
0
} (2.8)
which is a semiinﬁnite interval, by monotonicity of C
BS
in K. For each T > 0 and each
K ∈ K
S
(T), equation (2.5) deﬁnes the implied volatility of S to be the σ
imp
such that
C
BS
(S
0
, K, σ
imp
, T) = C
BS
(S
0
−θ, K −θ, σ, T). (2.9)
5
To abbreviate the σ
S
imp
(K, T) and K
S
(T) notations for displaced lognormal S, we will sup
press the S superscript, and possibly also the T argument.
2.2.2 Global behavior
Theorem 1 establishes the following global properties of σ
imp
: If θ < 0 then σ
imp
is every
where strictly decreasing in K, and bounded below by σ. If θ > 0 then σ
imp
is everywhere
strictly increasing in K, and bounded above by σ. In both cases, the global bounds are also
asymptotes.
Theorem 1 (Global behavior). Implied volatilities in the displaced lognormal model (2.6)
have the following global properties.
1. (Monotonicity in strike). For all T > 0 and K ∈ K(T),
sgn
∂σ
imp
∂K
(K) = sgn θ. (2.10)
2. (Upper/lower bound). For all T > 0 and K ∈ K(T):
If θ > 0 then σ
imp
(K) < σ. If θ < 0 then σ
imp
(K) > σ. (2.11)
3. (Sharpness of bound). For all T > 0, we have σ
imp
(K) → σ as K → ∞.
Hence sup
K∈K(T)
σ
imp
(K) = σ if θ > 0; and inf
K∈K(T)
σ
imp
(K) = σ if θ < 0.
Proof. Appendix A.1.
BrigoMercurio (2001) proves the K = S
0
case of (2.10, 2.11). Theorem 1 extends to all
K ∈ K(T).
Remark 2.2.1. Empirical volatility skews are typically not monotonic over the entire range of
strikes; a volatility skew which slopes downward in the central portion of the strike range will
usually still turn upward at suﬃciently large strikes. Theorem 1 proves that the displaced
lognormal cannot reproduce this empirical feature.
6
2.2.3 Atthemoney behavior
Theorems 2 and 3 focus on two diﬀerent subsets of the (K, T) domain.
Theorem 2 examines the atthemoney strike K = S
0
. Speciﬁcally, if T > 0 and S
0
∈
K(T), then deﬁne the atthemoney implied volatility σ
atm
(T) := σ
imp
(S
0
, T), which may
be abbreviated as σ
atm
. We bound the level σ
atm
and also the slope of σ
imp
atthemoney.
By “slope” we always mean ∂ log σ
imp
/∂ log K, the strikeelasticity of implied volatility.
Theorem 2 (Atthemoney behavior). Atthemoney implied volatilities in the displaced
lognormal model (2.6) have the following properties.
1. (Atthemoney level). If T > 0 and S
0
∈ K(T) then
σ
atm
≥
_
1 +
θ
S
0
_
σ if θ ≤ 0;
σ
atm
≤
_
1 −
θ
S
0
_
σ if θ ≥ 0.
(2.12)
2. (Atthemoney slope). If θ < 0 and T > 0 and S
0
∈ K(T) then
1
2
θ
θ + S
0
≤
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
=
N(σ
atm
√
T/2) −N(σ
√
T/2)
φ(σ
atm
√
T/2)
√
Tσ
atm
<
1
2
e
σ
2
atm
T/8
. (2.13)
Proof. Appendix A.2.
Remark 2.2.2. If T ≤ 1 and σ
atm
≤ 100%, then (1/2)e
σ
2
atm
T/8
< 0.57. Empirically, however,
equity volatility skews typically slope downward more steeply than −0.57. Indeed, in S&P500
daily data from all dates (1996–2008) in the OptionMetrics database, the approximate
1
3
monthexpiry atthemoney slope ∂ log σ
imp
/∂ log K is more negative than −1.29 on 90% of
the days in the sample. Theorem 2 proves that the displaced lognormal cannot reproduce
steepness of this magnitude.
1. Our source is the “Volatility Surface” data set, which contains volatility skews in
terpolated by OptionMetrics using kernel smoothing. We approximate the slope as
log(σ
imp
(K
1
)/σ
imp
(K
0
))/ log(K
1
/K
0
), where K
1
is the strike of a 0.50delta call, and K
0
is the
strike of a 0.55delta call, as computed by OptionMetrics.
7
2.2.4 Shortexpiry behavior
Theorem 3 takes the shortexpiry T ↓ 0 limit of the implied volatility skew, and expresses the
solution explicitly. The K = S
0
case is known (indeed Rebonato (2004) reﬁnes the K = S
0
formula, to address the case of T large). The contribution of Theorem 3 is to ﬁnd and prove
a shortexpiry σ
imp
formula valid for every strike K > θ
+
.
Figure 2.2.1: Theorem 3 formula and Exact σ
imp
for T = 1.0.
60 70 80 90 100 110 120 130 140 150 160
17.5
18
18.5
19
19.5
20
20.5
21
21.5
22
K
P
e
r
c
e
n
t
a
g
e
p
o
i
n
t
s
S
0
=100, σ(1−θ/S
0
)=20%
θ=25
θ=−50
Exact σ
imp
Theorem 3 formula
Theorem 3. For all K > θ
+
, in the displaced lognormal model (2.6),
lim
T→0
σ
imp
(K, T) =
_
¸
_
¸
_
σ log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
σ(1 −θ/S
0
) if K = S
0
.
(2.14)
Proof. Appendix A.3.
Remark 2.2.3. The formula in the righthand side of (2.14) provides, moreover, a remarkably
accurate approximation to σ
imp
(K, T) even for some T not close to 0. Figure 2.2.1 compares
the Theorem 3 formula and the exact σ
imp
(K, T), at expiry T = 1.0.
8
For some expirations of moderate length, therefore, the σ log(S
0
/K)/ log((S
0
−θ)/(K−θ))
formula may still facilitate calibration of the displaced lognormal parameters (σ, θ) to an
empirically observed volatility skew, or to a modelgenerated volatility skew.
2.3 Displaced antiLognormal
The previous section’s results show that with θ < 0, the displaced lognormal produces
downwardsloping implied volatility, but not of steepness commensurate with typical equity
options data – regardless of how large a negative value θ takes.
A related process, however, does generate arbitrarily large downward slopes.
Deﬁnition 3. A process S follows displaced anti lognormal dynamics if
dS
t
= σ(S
t
−θ)dW
t
, 0 < S
0
< θ, σ < 0, (2.15)
where W is a Brownian motion.
Thus θ − S is a driftless geometric Brownian motion with volatility −σ > 0, and the
interval of points attainable by S is (−∞, θ).
Displaced antilognormal pricing calculations have the tractability of the displaced log
normal. For instance, to price a Texpiry call struck at K < θ, on a displaced antilognormal
S,
E(S
T
−K)
+
= E(θ −K −(θ −S
T
))
+
= P
BS
(θ −S
0
, θ −K, −σ, T) = C
BS
(S
0
−θ, K −θ, σ, T).
(2.16)
So we have formally the same C
BS
formula as in the displaced lognormal case. Here its
ﬁrst three arguments are negative, which presents no problem; the C
BS
function is still
welldeﬁned by (2.1). An equivalent way to express the result, without negative arguments,
is C
BS
(θ −K, θ −S
0
, −σ, T).
Recognizing the similarities between the displaced lognormal and antilognormal, the
following terminology groups them together:
Deﬁnition 4 (DL). A process S which satisﬁes either the displaced lognormal (2.6) or the
displaced antilognormal (2.15) speciﬁcation is said to be a DL process.
9
2.3.1 Implied volatility
Implied volatility for displaced anti lognormal S is deﬁned on the strike interval
K
ADL
(S, T) := {K ∈ (0, θ) : C
BS
(S
0
−θ, K −θ, σ, T) < S
0
}. (2.17)
For K ∈ K
ADL
(S, T), we have
S
0
> C
BS
(S
0
−θ, K−θ, σ) = P
BS
(θ −S
0
, θ −K, −σ) > ((θ −K) −(θ −S
0
))
+
= (S
0
−K)
+
,
so (2.4) holds and σ
S
imp
(K, T) is thereby welldeﬁned. To abbreviate the σ
S
imp
(K, T) and
K
ADL
(S, T) notations for displaced antilognormal S, we will suppress the S superscript,
and possibly also the T argument. Note that we have
P
BS
(S
0
, K, σ
imp
) = E(K −S
0
)
+
= C
BS
(θ −S
0
, θ −K, σ). (2.18)
Consequently,
C
BS
(S
0
, K, σ
imp
) = E(S
T
−K) +P
BS
(S
0
, K, σ
imp
) = (S
0
−K) +C
BS
(θ −S
0
, θ −K, σ).
(2.19)
2.3.2 Global behavior
For the implied volatility under the displaced anti lognormal, we have the global monotonic
ity behavior similar to the behavior under the displaced lognormal. This is given by Theorem
4.
Theorem 4 (Global behavior). Implied volatilities in the antilognormal model (2.15) de
creases monotonically in strike. That is, for all T > 0 and K ∈ K
ADL
(T),
sgn
∂σ
imp
∂K
(K) = −sgn θ. (2.20)
Proof. Appendix A.4.
10
2.3.3 Atthemoney behavior
Theorem 5 examines the atthemoney strike K = S
0
. σ
atm
(T) is abbreviated as σ
atm
. We
bound the level σ
atm
and also the slope of σ
imp
atthemoney.
100 150 200 250 300 350
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
θ
σ
a
t
m
σ
atm
(θ/S
0
−1)σ
σ
Figure 2.3.1: Theorem 5.1 bounds of σ
atm
for T = 3.0. Process parameters are S
0
= 100, K =
100, σ = −0.3. The solid blue line is σ
atm
, the black dashed line is (θ/S
0
− 1)σ and the red
horizontal line is at the level of σ. As we can see from the ﬁgure, when θ ≥ 2S
0
, we have
σ
atm
> (θ/S
0
− 1)σ; though when S
0
< θ ≤ 2S
0
, it looks like the two lines coincide with each
other, if we zoom in, we could see that σ
atm
< (θ/S
0
−1)σ.
Theorem 5 (Atthemoney behavior). Atthemoney implied volatilities in the antilognormal
model (2.15) have the following properties.
1. (Atthemoney level). If θ > S
0
, T > 0 and S
0
∈ K
ADL
(T) then
σ
atm
≤
_
θ
S
0
−1
_
σ if S
0
< θ ≤ 2S
0
;
σ
atm
≥
_
θ
S
0
−1
_
σ if θ ≥ 2S
0
.
(2.21)
2. (Atthemoney slope). If θ > S
0
, T > 0 and S
0
∈ K
ADL
(T) then
1
2
θ
θ −S
0
≤
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
≤
θ
2(θ −S
0
)
e
σ
2
atm
T/8
. (2.22)
11
Proof. Appendix A.5.
Remark 2.3.1. θ = 2S
0
is the critical case. When θ < 2S
0
, we have σ
atm
<
_
θ
S
0
− 1
_
σ;
when θ > 2S
0
, we have σ
atm
>
_
θ
S
0
−1
_
σ; the equality happens when θ = 2S
0
. This could
be explained by using the putcall parity as follows. When K = S
0
and θ = 2S
0
, we have
(S
T
−K)
+
= (K −(θ −S
T
))
+
. (2.23)
If S
t
follows the displaced anti lognormal dynamics with parameters (θ = 2S
0
, σ), then using
K = S
0
= θ−K, the righthand side of (2.23) is P
BS
(θ−S
0
, θ−K, σ) = P
BS
(S
0
, K, σ) =
C
BS
(S
0
, K, σ). Therefore the implied volatility of the process S is exactly σ.
Figure 2.3.1 illustrates the Theorem 5.1 bounds of σ
atm
. The ﬁgure also suggests that
we will not have σ as the upper or lower bound for σ
imp
as in the displaced lognormal case.
2.3.4 Shortexpiry behavior
Theorem 3 extends to all DL processes:
Theorem 6 (Shortexpiry behavior). For all K > θ
+
in the displaced lognormal model (2.6),
as well as for all K ∈ (0, θ) in the displaced antilognormal model (2.15), the conclusion
(2.14) holds.
Proof. Appendix A.3.
The Theorem 6 conclusion, and its derivative with respect to log K, yield the shortexpiry
limiting volatility skew’s level and slope
σ
imp
¸
¸
¸
¸
T↓0, K=S
0
=
σ(S
0
−θ)
S
0
,
∂ log σ
imp
∂ log K
¸
¸
¸
¸
T↓0, K=S
0
=
θ
2(S
0
−θ)
. (2.24)
This holds under all DL dynamics. The distinction is that under displaced lognormal dynam
ics, we have θ < S
0
, hence the shortexpiry slope cannot be more negative than −1/2. Under
displaced anti lognormal dynamics, we have S
0
< θ, hence (2.24) can produce arbitrarily
steep negative slopes.
12
CHAPTER 3
DISPLACED HESTON PROCESS
3.1 Displaced Heston
Deﬁnition 5. A process S follows displaced Heston dynamics, with displacement θ ∈ R, if
dS
t
= σ
t
(S
t
−θ)dW
t
, S
0
> θ,
dσ
2
t
= κ(µ −σ
2
t
)dt + σ
t
dB
t
,
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
.
(3.1)
where W
t
, W
∗
t
are i.i.d. Brownian motions and σ
t
is nonnegative for all t > 0.
Thus S−θ follows the Heston model, and the interval of points attainable by S is (θ, ∞).
As with the displaced lognormal model, if modeling a nonnegative underlying process such
as a stock price, for θ < 0 this model will misprice deepoutofthemoney puts, due to the
possibility of S
T
< 0.
We would like to characterize the implied volatility of displaced Heston dynamics. We will
have similar atthemoney behavior and shortexpiry behavior as in the DL case. However,
we will not have the global behavior of implied volatility.
3.1.1 Implied volatility
Implied volatility for the displaced Heston S is deﬁned on the following strike interval:
K
DH
(T) := {K > θ
+
: (S
0
−K)
+
< E(S
T
−K)
+
< S
0
} (3.2)
For each T > 0 and each K ∈ K
DH
(T), (2.5) deﬁnes the implied volatility of the displaced
Heston to be σ
DH
imp
(K, T) > 0 such that
C
BS
(S
0
, K, σ
DH
imp
(K, T), T) = E(S
T
−K)
+
. (3.3)
We will suppress the T argument of σ
imp
(K, T).
13
3.1.2 Atthemoney behavior
Although for the displaced Heston process, we will not have the global behavior that the
slope of implied volatility is determined by the sign of the displacement θ, we are able to
show that when ρ = 0, this feature still holds. We will prove this for a more general DISV
process, deﬁned as follows.
Deﬁnition 6. (DISV) A process S follows displaced independent stochastic volatility dy
namics, with displacement θ ∈ R, if
dS
t
= σ
t
(S
t
−θ)dW
t
, S
0
> θ,
dσ
t
= f(σ
t
)dt + g(σ
t
)dW
∗
t
,
(3.4)
where W
t
, W
∗
t
are i.i.d. Brownian Motion.
Implied volatility for the DISV S is deﬁned on the following strike interval:
K
DISV
(T) := {K > θ
+
: (S
0
−K)
+
< E(S
T
−K)
+
< S
0
}. (3.5)
Using (2.5), for each T > 0 and each K ∈ K
DISV
(T), the implied volatility of the DISV
model σ
DISV
(K, T) is deﬁned such that
C
BS
(S
0
, K, σ
DISV
imp
(K, T), T) = E(S
T
−K)
+
. (3.6)
We can suppress the T in σ
DISV
imp
(K, T).
Theorem 7. If T > 0 and S
0
∈ K
DISV
(T), the slope of the atthemoney implied volatility
in the DISV model (3.4) has the following property:
sgn
∂σ
DISV
imp
∂K
(K)
¸
¸
¸
¸
K=S
0
= sgn θ. (3.7)
Proof. Appendix B.1.
Remark 3.1.1. The slope of atthemoney implied volatility depends on the assumption that
W
t
, W
∗
t
are independent. It may not hold when this condition is violated.
14
Corollary 3.1.1. If T > 0, S
0
∈ K
DH
(T) and ρ = 0, the slope of the atthemoney implied
volatility in the displaced Heston model (3.1) has the following property:
sgn
∂σ
imp
∂K
(K)
¸
¸
¸
¸
K=S
0
= sgn θ. (3.8)
Proof. With ρ = 0, the displaced Heston model belongs to the DISV family of models.
3.1.3 Shortexpiry behavior
Durrleman (2004) gives the shortexpiry implied volatility of the Heston model as a function
of the strike. We will derive the shortexpiry implied volatility of the displaced Heston
model using his results. Particularly, we connect the shortexpiry implied volatility of the
displaced Heston model with that of the Heston model, and we give an approximation of the
shortexpiry implied volatility level, slope and convexity.
Shortexpiry behavior of Heston process
Recall that the Heston process proposed by Heston (1993) as follows.
dS
t
= σ
t
S
t
dW
t
,
dσ
2
t
= κ(µ −σ
2
t
)dt + σ
t
dB
t
,
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
,
(3.9)
where W
t
, W
∗
t
are i.i.d. Brownian motions and σ
t
is nonnegative for all t > 0.
For S
t
under the displaced Heston dynamics (3.1), denote
˜
S
t
= S
t
− θ, then
˜
S
t
follows
the Heston dynamics (3.9).
The implied volatility of the Heston process
˜
S
t
is deﬁned on the following strike interval:
K
H
(T) := {K > θ
+
: (
˜
S
0
−(K −θ))
+
< E(
˜
S
T
−(K −θ))
+
<
˜
S
0
}. (3.10)
For each T > 0 and each K ∈ K
H
(T), (2.5) deﬁnes the implied volatility of the displaced
Heston to be σ
˜
S,H
imp
(
˜
K, T) > 0 such that
C
BS
(
˜
S
0
,
˜
K, σ
˜
S,H
imp
(
˜
K, T), T) = E(
˜
S
T
−
˜
K)
+
, (3.11)
15
where
˜
K = K −θ. We will suppress the
˜
S and T argument of σ
˜
S,H
imp
(
˜
K, T) as σ
H
imp
(
˜
K).
A call option on the displaced Heston process starting at S
0
with strike K can be con
sidered as a call option on a Heston process starting at
˜
S
0
with strike
˜
K:
E(S
T
−K)
+
= E((S
T
−θ) −(K −θ))
+
= E(
˜
S
T
−
˜
K)
+
. (3.12)
We take care here to deﬁne the domain of K to make sure that the implied volatilities
of both the Heston model and the displaced Heston model exist. Combine (3.2) and (3.10)
and deﬁne
K
DH∩H
(T) = K
DH
(T)∩K
H
(T) := {K > θ
+
: (S
0
−K)
+
< E(S
T
−K)
+
< max(S
0
, (S
0
−θ))}
to be the domain of K.
The notation A(T) ∼ B(T) means that A(T)/B(T) → 1 as T ↓ 0. Durrleman (2004)
shows that, for the Heston model, the nearmoney shortexpiry implied volatility of the
vanilla call with
˜
Kstrike, Tmaturity and initial value
˜
S
0
can be approximated as
σ
H
imp
(
˜
K, T) ∼
_
σ
2
0
+ a
0
log(
˜
S
0
/
˜
K) + b
0
T
2
+
c
0
2
log
2
(
˜
S
0
/
˜
K), (3.13)
where
a
0
= −
ρ
2
, b
0
= κ(µ −σ
2
0
) +
ρ
2
σ
2
0
−
2
6
(1 −ρ
2
/4), c
0
=
2
6σ
2
0
(1 −
7ρ
2
4
).
Shortexpiry behavior of displaced Heston process
When near the expiry, the σ
imp
is welldeﬁned for both the Heston process and the displaced
Heston process, so K ∈ K
DH∩H
(T). Theorem 8 says that the shortexpiry implied volatility
of the displaced Heston model can be expressed in terms of the shortexpiry implied volatility
of the Heston model. Proposition 3.1.2 gives the shortexpiry implied volatility’s level, slope
and convexity atthemoney.
Theorem 8. For K > θ
+
, the shortexpiry relationship between the implied volatilities of
16
the displaced Heston model (3.1) and the Heston model (3.9) is given as follows.
lim
T→0
σ
DH
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
σ
H
imp
(K −θ, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
σ
H
imp
(K −θ, T) ×(1 −θ/S
0
) if K = S
0
.
(3.14)
Proof. Appendix B.2.
Proposition 3.1.2. For the displaced Heston model (3.1) with ρ = 0, for K > θ
+
, the
atthemoney implied volatility level, slope and convexity when T ↓ 0 are given as follows.
level: (σ
DH
imp
(K, T))
2
¸
¸
¸
¸
K=S
0
=
_
σ
2
0
+ κ(µ −σ
2
0
)
T
2
−
2
T
12
_
×(1 −θ/S
0
)
2
+ O(T
2
),
slope:
∂ log σ
DH
imp
(K, T)
∂ log K
¸
¸
¸
¸
K=S
0
=
θ
2(S
0
−θ)
+ O(T),
convexity:
∂
2
σ
DH
imp
(K, T)
∂K
2
¸
¸
¸
¸
K=S
0
=
c
0
_
σ
H
imp
(S
0
−θ, T)
_
−1
2(S
0
−θ)S
0
+ σ
H
imp
(S
0
−θ, T)H(S
0
, θ) + O(T),
(3.15)
where c
0
=
2
6σ
2
0
and H(S
0
, θ) =
4θ
2
−5θS
0
6S
3
0
(S
0
−θ)
.
Proof. Appendix B.3.
Remark 3.1.2. The shortexpiry atthemoney slope is
θ
2(S
0
−θ)
, suggesting that when θ > 0,
the atthemoney slope is positive; when θ < 0, the atthemoney slope is negative, and
is bounded by
1
2
. We have mentioned in Remark 2.2.2 that the empirical equity volatility
skews typically slope downward more steeply than −0.57. Theorem 3.1.2 suggests that the
shortexpiry displaced Heston cannot reproduce steepness of this magnitude.
3.2 Displaced antiHeston
Remark 3.1.2 says that with θ < 0, the steepness of the shortexpiry implied volatility from
the displaced Heston has an upper bound of
1
2
atthemoney. This means that the displaced
Heston is not a suitable model for the equity data. A related process, however, which is
analogous to the displaced anti lognormal process, will generate arbitrary large downward
slopes.
17
Deﬁnition 7. A process S follows displaced anti Heston if:
dS
t
= (−σ
t
)(S
t
−θ)dW
t
, 0 < S
0
< θ,
dσ
2
t
= κ(µ −σ
2
t
)dt + σ
t
dB
t
,
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
,
(3.16)
where W
t
, W
∗
t
are i.i.d. Brownian motions and σ
t
is nonnegative for all t > 0.
Thus θ − S follows the Heston dynamics with volatility process σ
t
, and the interval of
points attainable by S is (−∞, θ).
Recognizing the similarities between the displaced Heston and antiHeston, the following
terminology groups them together:
Deﬁnition 8. (DH) A process S which satisﬁes either the displaced Heston (3.1) or the
displaced antiHeston (3.16) speciﬁcation is said to be a DH process.
3.2.1 Implied volatility
Implied volatility for the displaced anti Heston S is deﬁned on the following strike interval:
K
ADH
(T) := {0 < K < θ : (S
0
−K)
+
< E(S
T
−K)
+
< S
0
}. (3.17)
For each T > 0 and each K ∈ K
ADH
(T), (2.5) deﬁnes the implied volatility of the displaced
anti Heston to be σ
ADH
imp
(K, T) > 0 such that
C
BS
(S
0
, K, σ
ADH
imp
(K, T), T) = E(S
T
−K)
+
. (3.18)
Denote
¯
S
0
= θ −S
0
,
¯
K = θ −K, then
¯
S is the Heston process (3.9). Implied volatility for
the Heston
¯
S is deﬁned on the following strike interval
K
AH
(T) := {0 < K < θ : (
¯
S
0
−
¯
K)
+
< E(
¯
S
T
−
¯
K)
+
<
¯
S
0
}. (3.19)
For each T > 0 and each K ∈ K
AH
(T), (2.5) deﬁnes the implied volatility of the Heston to
be σ
H
imp
(K, T) > 0 such that
C
BS
(
¯
S
0
,
¯
K, σ
H
imp
(
¯
K, T), T) = E(
¯
S
T
−
¯
K)
+
. (3.20)
18
Again, to make sure the implied volatilities of both the Heston model and the displaced
anti Heston model exist, we deﬁne a domain of K as
K
ADH∩AH
(T) = K
ADH
(T) ∩K
AH
(T) = {K : 0 < K < θ, σ
ADH
imp
(K, T) ∃, σ
H
imp
(
¯
K, T) ∃.}
We can suppress the T in σ
ADH
imp
(K, T) as σ
ADH
imp
(K), and σ
H
imp
(
¯
K, T) as σ
H
imp
(
¯
K).
3.2.2 Shortexpiry behavior
When near the expiry, the σ
imp
is welldeﬁned for both the Heston process and the displaced
anti Heston process, so K ∈ K
ADH∩AH
(T). Theorem 9 gives a relationship between the
shortexpiry implied volatilities of the displaced anti Heston model and the Heston model.
Proposition 3.2.1 gives the shortexpiry atthemoney implied volatility’s level, slope and
convexity.
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
0.18
0.185
0.19
0.195
0.2
0.205
0.21
0.215
0.22
0.225
0.23
I
m
p
l
i
e
d
V
o
l
a
t
i
l
i
t
y
log(K/S)
T=1/12
S=100, σ
0
=0.04
Th−ImpVol
ImpVol
−0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
0.18
0.185
0.19
0.195
0.2
0.205
0.21
0.215
0.22
0.225
0.23
I
m
p
l
i
e
d
V
o
l
a
t
i
l
i
t
y
log(K/S)
T=4/12
S=100, σ
0
=0.04
Th−ImpVol
ImpVol
Figure 3.2.1: Theorem 9 formula and exact σ
imp
. Parameters of the displaced anti Heston dy
namics are S
0
= 100, σ
0
= 0.04, κ = 0.2, µ = 0.002, = 0.0175, θ = 600, ρ = 0. In the left plot,
T = 1 month and in the right plot, T = 4 month. The DH model parameters are chosen to be close
to the ones calibrated from the SABR model using S = 100, β = 0.2, αS
β
0
= 0.2, ρ = −0.4, ν =
0.2, T = 1 year. The details of the calibration are in section 4.1. The dashed black line is the exact
σ
imp
by (3.18). The solid green line is Theorem 9 formula, where σ
H
imp
(θ − K, T) is approximated
using (3.13).
19
Theorem 9. For K ∈ (0, θ), the shortexpiry relationship between the implied volatilities of
the displaced antiHeston model (3.16) and the Heston model (3.9) is given as follows.
lim
T→0
σ
DH
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
−σ
H
imp
(θ −K, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
−σ
H
imp
(θ −K, T) ×(1 −θ/S
0
) if K = S
0
.
(3.21)
Proof. Appendix B.2
Remark 3.2.1. Figure 3.2.1 compares the Theorem 9 formula and the exact σ
imp
(K, T), at
expiry T = 1/12 and T = 4/12. It suggests that the formula in the righthand side of (3.21)
provides, a remarkably accurate approximation to σ
imp
(K, T) even for T not close to 0.
Proposition 3.2.1. For the displaced antiHeston (3.16) with ρ = 0, for K ∈ (0, θ), the
atthemoney implied volatility level, slope and convexity when T ↓ 0 are given in the follows.
level: (σ
ADH
imp
(K, T))
2
¸
¸
¸
¸
K=S
0
=
_
σ
2
0
+ κ(µ −σ
2
0
)
T
2
−
2
T
12
_
×(1 −θ/S
0
)
2
+ O(T
2
),
slope:
∂ log σ
ADH
imp
(K, T)
∂ log K
¸
¸
¸
¸
K=S
0
=
θ
2(S
0
−θ)
+ O(T),
convexity:
∂
2
σ
ADH
imp
(K, T)
∂K
2
¸
¸
¸
¸
K=S
0
= −
c
0
_
σ
H
imp
(θ −S
0
, T)
_
−1
2(S
0
−θ)S
0
−σ
H
imp
(θ −S
0
, T)H(S
0
, θ) + O(T),
(3.22)
where c
0
=
2
6σ
2
0
and H(S
0
, θ) =
4θ
2
−5θS
0
6S
3
0
(S
0
−θ)
.
Proof. Appendix B.3.
3.2.3 Generalization of shortexpiry behavior
The relationship between the shortexpiry σ
imp
of the Heston process and the DH process
given in Theorems 8 and 9 can be generalized to a relationship between the shortexpiry
σ
imp
of any process which is a positive martingale and its corresponding (anti)displaced
process.
Denote the process X to be a positive martingale.
20
Deﬁnition 9. A process S is said to be a displacedX process, with displacement θ ∈ R, if
S
t
−θ = X
t
, for any t ≥ 0.
Deﬁnition 10. A process S is said to be an displacedantiX process, with displacement
θ ∈ R, if θ −S
t
= X
t
, for any t ≥ 0.
Denote σ
X
imp
(K, T) to be the implied volatility of the process X by (2.5). Denote
σ
DX
imp
(K, T) and σ
ADX
imp
(K, T) to be the implied volatilities of the displacedX process and
the displacedantiX process. Denote K
X
(T) to be the domain of K where σ
X
imp
(K, T) and
σ
DX
imp
(K, T) are welldeﬁned as T ↓ 0. Denote K
ADX
(T) to be the domain of K where
σ
X
imp
(K, T) and σ
ADX
imp
(K, T) are welldeﬁned as T ↓ 0. Theorems 10 and 11 give the
shortexpiry relationship between σ
X
imp
(K, T) and σ
DX
imp
(K, T), as well as σ
X
imp
(K, T) and
σ
ADX
imp
(K, T).
Theorem 10. For a process X which is a positive martingale, for K > θ
+
and K ∈
K
DX
(T), the shortexpiry relationship between the implied volatilities of X and its corre
sponding displacedX process is given as follows.
lim
T→0
σ
DX
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
σ
X
imp
(K −θ, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
σ
X
imp
(K −θ, T) ×(1 −θ/S
0
) if K = S
0
.
(3.23)
Proof. Similar to the proof of Theorem 8.
Theorem 11. For a process X which is a positive martingale, for K ∈ (0, θ) and K ∈
K
ADX
(T), the shortexpiry relationship between the implied volatilities of X and its corre
sponding displacedantiX process is given as follows.
lim
T→0
σ
ADX
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
−σ
X
imp
(θ −K, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
−σ
X
imp
(θ −K, T) ×(1 −θ/S
0
) if K = S
0
.
(3.24)
Proof. Similar to the proof of Theorem 9.
21
CHAPTER 4
CALIBRATION OF DL AND DH PROCESS
4.1 Calibration of DL and DH Process
Whether one chooses to use the DL/DH as a model, or as an approximation of another
model, or (as we will) as a control variate for another model, in any case the DL parameters
(σ, θ) and DH parameters (σ
0
, κ, ν, ρ, , θ) require estimation/calibration. We use Theorem
6 implications (2.24) to ﬁt the DL parameters to a given implied volatility level and slope.
We use Proposition 3.1.2 and Proposition 3.2.1 to ﬁt the DH parameters to a given implied
volatility’s level, slope and curvature.
4.1.1 Calibration of DL process
Given a shortexpiry atthemoney skew level a and slope b (either from some model, or from
direct empirical measurement), and given an underlying level S
0
, there exists a DL process,
with S
0
= S
0
and parameters (θ, σ), such that the DL skew’s shortexpiry level and slope
(2.24) match the given level a and slope b, provided that b = −1/2. Explicitly, we ﬁnd
σ = a(1 + 2b), θ =
2b
1 + 2b
S
0
. (4.1)
In the case of slope b > −1/2, the calibrated DL process is a displaced lognormal. In the case
of slope b < −1/2, the calibrated DL process is a displaced anti lognormal. The singular
case of slope b = −1/2 can be matched by normal or “Bachelier” dynamics dS
t
= aS
0
dW
t
.
The DL and Bachelier models belong to the family dS
t
= (σS
t
+A)dW
t
, where σ = 0 in the
case of DL, and σ = 0 in the singular case of Bachelier.
Remark 4.1.1. Although (2.24) is a shortexpiry limit, Remark 2.2.3 indicates its accuracy
at T of moderate length. Therefore (4.1) may still facilitate calibration of (σ, θ) to volatility
skews (a, b) even at moderately long expiries.
22
4.1.2 Calibration of DH process
Given a shortexpiry atthemoney skew level a, slope b and convexity c (either from some
model, or from direct empirical measurement), and given an underlying level S
0
, there exists
a DH process, with S
0
= S
0
and parameters (σ
0
, κ, ν, ρ, , θ), such that the DH skew’s
shortexpiry level, slope and convexity (3.15) or (3.22) match the given level a, slope b and
convexity c, provided that b = −1/2. Recall that we have restricted ρ = 0 in (3.15) and
(3.22). In application, we will also preﬁx κ. With (ρ, κ) preﬁxed, we ﬁnd explicitly,
θ =
2b
1 + 2b
S
0
, σ
0
= a(2b + 1), µ =
2
6κ
+ σ
2
0
=
¸
6σ
2
0
×2(S
0
−θ)S
0
_
c −a(2b + 1)H(S
0
, θ)
__
a(2b + 1)
_
,
(4.2)
If b > −
1
2
, we calibrate the parameters to the displaced Heston model; if b < −
1
2
, we calibrate
the parameters to the displaced anti Heston model. For the singular case b = −
1
2
, we take
the same approach as in the DL case.
Compared with the calibration of the DL parameters given in (4.1), we see that the θ is
the same in both the DL and the DH case. In the DL case, we have σ = a(2b + 1); while in
the DH case, we have σ
0
= a(2b + 1).
Remark 4.1.2. Although (3.14) and (3.21) are shortexpiry limit, Remark 3.2.1 indicates
their accuracy at T of moderate length. Therefore (4.2) may still facilitate calibration of
(σ
0
, κ, ν, ρ, , θ) to volatility skews (a, b, c) even at moderately long expiries.
4.2 Calibration DL and DH to CEV/SABR
4.2.1 CEV and SABR stochastic volatility models
For many local or stochastic volatility models, there exist explicit short maturity approxi
mations of implied volatility, such as in Lewis (2000) and Berestycki et al. (2004), making it
easy to calculate the implied volatility level a, slope b and convexity c, and to calibrate DL
parameters σ and θ via (4.1), or to calibrate DH parameters (σ
0
, κ, ν, ρ, , θ) via (4.2).
Two such models capable of generating realistically steep atthemoney implied volatility
skews are the Constant Elasticity of Variance (CEV) model and the SABR model. In the
23
CEV model (Cox 1996),
dS
t
= αS
β
t
dW
t
, S
0
> 0, (4.3)
where β ≤ 1, and absorption is imposed at S = 0.
Remark 4.2.1. Boundary Conditions for the CEV:
• β < 1/2: 0 is attainable. The origin is a regular boundary point and is speciﬁed as a
killing boundary by adjoining a killing boundary condition.
• β = 1/2: 0 is attainable and strong reﬂecting.
• β > 1/2: 0 is not attainable. 0 is exist boundary.
The CEV model can generate a steep downward implied volatility skew atthemoney.
Indeed, by Berestycki et al. (2002) and Roper (2009), for all K > 0,
lim
T→0
σ
CEV
imp
(K, T) =
_
¸
¸
_
¸
¸
_
α(1 −β) log(S
0
/K)
S
1−β
0
−K
1−β
if K = S
0
αS
β−1
0
if K = S
0
.
(4.4)
Diﬀerentiating with respect to log K, we have
∂ log σ
CEV
imp
∂ log K
¸
¸
¸
¸
T↓0, K=S
0
=
β −1
2
, (4.5)
which can take arbitrarily large negative values.
The widelyused SABR model (Hagan et al. 2002) generalizes the CEV, by making
the coeﬃcient α stochastic, with volatilityofvolatility ν ≥ 0, and correlation ρ ∈ [−1, 1]
between S and α:
dS
t
= α
t
S
β
t
dW
t
, S
0
> 0
dα
t
= να
t
dB
t
, α
0
> 0
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
(4.6)
where W and W
∗
are independent Brownian motions, and absorption is imposed at S = 0.
Taking ν = 0 in the SABR model reduces to the CEV case.
24
According to a shortexpiry approximation in Hagan et al. (2002, eq. 3.1a),
σ
SABR
imp
(K, T) ≈ α
0
S
β−1
0
_
1 +
_
β −1
2
+
ρν
2αS
β−1
0
_
log(K/S
0
)
_
. (4.7)
The approximation’s slope at K = S
0
is
∂ log σ
SABR
imp
∂ log K
¸
¸
¸
¸
T↓0, K=S
0
≈
β −1
2
+
ρν
2αS
β−1
0
, (4.8)
reﬂecting the contributions to the SABR volatility skew, not just from the functional rela
tionship between price levels and volatility, as expressed by β, but also from the correlation
between price increments and volatility, as expressed by ρ.
4.2.2 Calibration of DL to CEV/SABR
For the SABR process, a = α
0
S
β−1
0
, and b is given by (4.8), so we have
σ = α
0
βS
β−1
0
+ ρν, θ = S
0
−
α
0
S
β
0
α
0
βS
β−1
0
+ ρν
. (4.9)
The ν = 0 special case of (4.9) gives the DL parameters that match the CEV level and slope:
σ = αβS
β−1
0
, θ = S
0
(β −1)/β. (4.10)
In the CEV case, Marris (1999) and SvobodaGreenwood (2009) have already investigated
displaced lognormal approximation, by an approach which chooses parameters such that the
displaced lognormal instantaneous volatility approximates the CEV instantaneous volatility
function S → αS
β
, in contrast to our approach which matches the implied volatility functions.
Their approach arrived at the same result (4.10) as our approach, in the CEV case.
A distinction is that our implied volatility approach is intended to apply moreover to
models, such as SABR, where instantaneous volatility varies not just as a function of S, but
also other stochastic factors. The implied volatility skew reﬂects the dependence of volatility
on the S level together with the other stochastic factors in the model, such as α in the SABR
case.
25
4.2.3 Calibration of DH to SABR
Durrleman (2004) shows that (3.13) can also be used as an approximation for the shortexpiry
implied volatility of SABR model (4.6) with
σ
0
= α
0
S
β−1
0
, a
0
= −(νρ + (β −1)σ
0
)σ
0
,
c
0
=
ν
2
6
(4 −3ρ
2
) + νρ(β −1)σ
0
+
5
6
(β −1)
2
σ
2
0
,
b
0
=
σ
2
0
6
(ν
2
(2 −3ρ
2
) + 6νρβσ
0
+ (β −1)
2
σ
2
0
).
(4.11)
From this, we derive the approximation of the shortexpiry SABR implied volatility level,
slope and convexity as follows.
a(T) =
_
σ
2
0
+ b
0
T
2
,
b(T) = −
(σ
atm
)
−2
a
0
2
,
c(T) =
(σ
atm
)
−2
a
0
2K
∂σ
atm
∂K
+
a
0
+ c
0
2K
2
(σ
atm
)
−1
,
(4.12)
where σ
atm
is approximated by (3.13) with K = S
0
. The parameters of DH are chosen by
plugging (4.12) into (4.2).
Remark 4.2.2. In (4.12), lim
T→0
a(T) = αS
β−1
0
, lim
T→0
b(T) =
β−1
2
+
ρν
2αS
β−1
0
, the same
as the level and slope given in (4.7) and (4.8).
26
CHAPTER 5
VARIANCE REDUCTION IN MONTE CARLO SIMULATION
5.1 Variance Reduction Using Control Variate
Theorems 1 and 2 imply that the displaced lognormal is inconsistent with the steep downward
slopes (Remark 2.2.2) and nonmonotonicity (Remark 2.2.1) typical of stock market volatility
skews. The displaced anti lognormal, by (2.24), overcomes the steepness constraint, but
introduces other drawbacks: its paths, which take values in (−∞, θ), are bounded above and
unbounded below – the opposite of the behavior desirable in a model of stock prices. Such
paradox applies to the DH process too. Proposition 3.1.2 implies that the displaced Heston
is inconsistent with the steep downward slopes (Remark 3.1.2). The displaced anti Heston
overcomes the steepness constraint but its paths are unbounded from below.
For these reasons, we do not generally advocate the DL/DH to model stock price pro
cesses. Rather, we propose the DL/DH to generate control variates to reduce variance in
the Monte Carlo pricing of derivative contracts under commonlyused dynamics which do
match the empirical atthemoney volatility skew.
Indeed, suppose the underlying S dynamics follow some speciﬁcation that a modeler
deems appropriate, such as the CEV or the SABR stochastic volatility model. Suppose the
modeler intends to price a derivative contract for which the desired model lacks analytical
pricing formulas, such as a discretelymonitored barrier option on the CEV/SABR process
S. Let the contract’s payoﬀ Y be given by a speciﬁed function of the S path. In the absence
of analytical solutions, consider the use of Monte Carlo simulation to estimate the price EY.
The basic Monte Carlo estimator is the sample average
ˆ
C :=
1
M
m
Y
m
(5.1)
where the simulations Y
1
, . . . Y
M
are iid as Y. To improve accuracy, in the sense of reducing
variance, let us apply the control variate technique, where the control comes from a DL
process calibrated by (4.1), or from a DH process calibrated by (4.2).
27
There exist, of course, other variance reduction methods, combinable with a DL/DH
control variate. We do not investigate them in this section; rather we maintain focus on
the DL/DH control, with the intent of illustrating how much variance reduction the DL/DH
control brings by itself. In section 5.2 we will discuss combining the DL/DH control with
other techniques.
5.1.1 DL or DH as a control variate
To make explanation clear, we ﬁrst focuses on using DL as control variate. Using DH as
control variate follows the same spirit.
The control variate estimator of EY, using a control Y , where Y has a known expectation
C := EY and a known simulation methodology, is deﬁned by
ˆ
C
cv
:=
1
M
m
_
Y
m
−βY
m
+ βC
_
, (5.2)
where the simulated pairs (Y
1
, Y
1
), . . . , (Y
M
, Y
M
) are iid as (Y, Y ). Good choices of Y have
large correlation ρ
Y,Y
with Y, because increasing ρ
Y,Y
 decreases the estimator’s variance.
Speciﬁcally,
Var
ˆ
C
cv
= (1 −ρ
2
Y,Y
) Var
ˆ
C (5.3)
for the optimal choice of the β coeﬃcient, namely β = Cov(Y, Y )/ Var(Y ), which may also
be estimated by simulation. For further details see, for instance, Boyle et al. (1997).
Because the payoﬀ Y is a speciﬁed function of the S path, we choose Y to be that same
payoﬀ function applied to the S path, where S follows a DL process driven by a Brownian
motion that also drives S. Aiming to produce high correlation ρ
Y,Y
, we choose the S process
parameters by taking S
0
= S
0
and applying (4.1) to ﬁnd (σ, θ) such that the shortexpiry
atthemoney volatility skews implied by S and by S agree in both level and slope.
The suitability of the DL process S to serve in this role stems from a conﬂuence of
ﬂexibility and tractability; the DL is potentially ﬂexible enough to generate signiﬁcant cor
relation between Y and Y (by linking the parameters of S and S, as discussed above), and
yet potentially tractable enough to allow analytic evaluation of EY and unbiased simulation
of Y , as discussed below.
For shiftinvariant contracts (including barriers and lookbacks), exact evaluation of C =
EY under DL dynamics is just as easy as under BlackScholes dynamics; more precisely,
28
if the contract’s payoﬀ is invariant to parallel shifts of the underlying price path and the
contract parameters (such as strike and barrier level), then BlackScholes model valuation
methods, applied to shifted arguments, produce the contract’s DL valuation. In the DL case,
if, moreover, we can simulate the exact distribution of Y – which is often the case, because
S is a transformed Gaussian – then Y can serve as a control variate that reduces variance
without introducing any bias.
When using DH as control variate, we have one more source of ﬂexibility, namely the
stochastic volatility process. If Y depends on the path S where S depends on a stochastic
volatility process Σ, S is driven by a Brownian motion W and Σ is driven by W and W
∗
,
where W and W
∗
are independent, we choose Y to be the same payoﬀ function applied to
the S path, where S follows a DH process driven by W, and the volatility process of S is
driven by W
∗
. We choose the S process parameters by taking S
0
= S
0
and applying (4.2)
to ﬁnd (σ
0
, κ, ν, ρ, , θ) such that the shortexpiry atthemoney volatility skews implied by
S and by S agree in level, slope and convexity.
The DH control variate has an advantage over the DL control variate in that the DH
model is more ﬂexible and the S path tracks the S path more closely. Consequently, Y will
have higher correlation with Y so we can achieve greater variance reduction. On the other
hand, the option pricing formula on the DH model is more complicated than on the DL
model.
5.1.2 Example I: Discretely sampled barrier option under CEV/SABR
dynamics
To take a concrete example, consider a discretely sampled barrier option on S, which follows
CEV (4.3) or SABR (4.6) dynamics. In particular, let the contract be a downandout call
with expiry T, barrier H, strike K, sampling dates t
1
< t
2
< · · · < T
N
= T, and payoﬀ
Y := (S
T
−K)
+
1(min
n
S
t
n
> H). (5.4)
Analytical solutions exist for continuous barriers in the CEV model (DavydovLinetsky
2001), but not for discrete barriers, nor for the SABR model, so we turn to Monte Carlo
simulation.
29
DL as control variate
To generate a control variate on DL, we apply the same payoﬀ function to a DL process,
driven by the same Brownian motion W = W. More precisely,
Y := (S
T
−K)
+
1(min
n
S
t
n
> H)
dS
t
= σ(S
t
−θ)dW
t
,
(5.5)
where S
0
= S
0
, and (σ, θ) are calibrated by (4.9) in the SABR case, or (4.10) in the CEV
case.
In the DL case, This Y is easily simulated without bias, and the value of C = EY can
be computed by shifting any of the fast and exact (up to numerical truncation/quadrature
error) solutions for discrete barrier option prices in the Gaussian framework, such as Broadie
and Yamamoto (2005), or in the L´evy framework, such as Petrella and Kou (2004) or Feng
and Linetsky (2008).
An alternative to (5.5) is to choose instead a continuouslymonitored control
Y
∗
:= (S
T
−K)
+
1( min
t∈[0,T]
S
t
> H). (5.6)
The control expectation EY
∗
has a simple exact formula, and the control Y
∗
can be simulated
without bias, using Brownian bridge techniques of Beaglehole et al. (1997).
DH as control variate
To generate a control variate on DH, we apply the same payoﬀ function to a DH process,
driven by the same Brownian motion W = W, W
∗
= W
∗
. The DH process parameters
(σ
0
, κ, ν, ρ, , θ) are calibrated by (4.2) where a, b, c are determined by (4.12). More precisely,
if b > −
1
2
, the control variate is based on the displaced Heston model,
Y := (S
T
−K)
+
1(min
n
S
t
n
> H)
dS
t
= σ
t
(S
t
−θ)dW
t
, S
0
> θ,
dσ
2
t
= κ(ν −σ
2
t
)dt + σ
t
dW
∗
t
, dW
t
dW
∗
t
= 0
(5.7)
30
If b < −
1
2
, the control variate is based on displaced anti Heston model:
Y := (S
T
−K)
+
1(min
n
S
t
n
> H)
dS
t
= (−σ
t
)(S
t
−θ)dW
t
, 0 < S
0
< θ,
dσ
2
t
= κ(ν −σ
2
t
)dt + σ
t
dW
∗
t
, dW
t
dW
∗
t
= 0
(5.8)
In both (5.7) and (5.8), S
0
= S
0
and σ
t
is nonnegative for all t > 0.
In the DH case, the Y can be simulated from an eﬃcient discretisation scheme which
guarantees strong convergence, such as Lord et, al (2006). We discuss this more in detail in
Chapter 6. The value of C = EY can be computed by shifting any of the fast solutions for
discrete barrier option price under the Heston model, such as Griebsch and Wystup (2008).
5.1.3 Numerical results I: Discretely sampled barrier option under
CEV/SABR dynamics
DL as Control Variate
Our experiments simulate the payoﬀ (5.4), where
K = S
0
= 100, H = 95, T = 4/12, N = 84. (5.9)
In the CEV case we take β ∈ {−0.5, −1.0, −1.5}, with α such that αS
β−1
0
∈ {0.15, 0.20, 0.25},
based on HirsaCourtadonMadan’s (2003) estimates of S&P500 CEV parameters; our β is
what they denote as β + 1. We use the control (5.5), where (σ, θ) are tuned to the CEV by
(4.10).
In the SABR case we take β = 0.2, with α
0
such that α
0
S
β−1
0
= 0.2, with an array of
choices for (ν, ρ). We use the control (5.5), where (θ, σ) are tuned to the SABR by (4.9).
We run 1000 paths. Ten equal spaced points are sampled each day. Tables 5.1.1 and
5.1.2 report the estimated “percentage reduction of variance”
100%−
¯
Var(
ˆ
C
cv
)
¯
Var(
ˆ
C)
, (5.10)
31
Table 5.1.1: Percentage reduction of variance, using DL for downandout call on CEV
T = 4 months
β αS
β−1
0
0.15 0.20 0.25
0.50 99.99% 99.99% 99.99%
1.00 99.99% 99.98% 99.93%
1.50 99.98% 99.89% 98.67%
Payoﬀ: (5.4) with (5.9). Control: (5.5) with (4.10).
Table 5.1.2: Percentage reduction of variance, using DL for downandout call on SABR
T = 4 months
ρ ν
0.20 0.40 0.60
0.4 98.74% 96.07% 92.42%
0.6 98.83% 97.17% 93.70%
0.8 99.64% 98.71% 96.76%
Payoﬀ: (5.4) with (5.9). Control: (5.5) with (4.9).
where each
¯
Var is the scaled sample variance of the summands in (5.1) and (5.2) respec
tively. Equivalently, the percentage reduction of variance equals the “Rsquared” of an OLS
regression of the CEV/SABR barrieroption payoﬀ Y on the DL control payoﬀ Y .
DH as Control Variate
Our next experiments simulate the payoﬀ (5.4) with parameters given by (5.9), using DH
as control variate for the downandout discrete barrier under the SABR process. The
parameters of the SABR process are β = 0.2, with α
0
such that α
0
S
β−1
0
= 0.2, with an
Table 5.1.3: Percentage reduction of variance, using DH for downandout call on SABR
T = 4 months
ρ ν
0.20 0.40 0.60
0.4 99.76% 99.03% 99.41%
0.6 99.73% 98.95% 99.35%
0.8 99.00% 98.60% 98.73%
Payoﬀ: (5.4) with (5.9). Control: (5.7) with (4.2) or (5.8) with (4.2).
In (4.2), a, b and c are calculated from (4.12).
32
array of choices for (ν, ρ). The results are shown in Table 5.1.3.
Comparing the results in Table 5.1.2 and 5.1.3, we can see that using DH as control
variate achieves more variance reduction. As we noted before, the reason is that the DH
process is a twofactor stochastic dynamics which enable the process S to better track the
original process, S.
Remark 5.1.1. In the reported Table 5.1.3 and any other reported results related to DH
process in this paper, we arbitrarily chose κ = 0.2. Numerical simulations suggest that other
choices of κ give relatively same results.
5.1.4 Example II: Discretely sampled arithmetic Asian option under
SABR dynamics
The next example we consider is the discretely sampled arithmetic Asian option on S, which
follows the SABR dynamics (4.6). Let the contract be an Asian call option with expiry T,
strike K, sampling dates t
1
< t
2
< · · · < t
N
= T, and payoﬀ
Y := (
1
N
N
i=1
S
i
−K)
+
(5.11)
There is no analytical solution for the arithmetic Asian option price under the SABR process,
so we turn to Monte Carlo simulation.
DL as control variate
For the control variate, we could apply the same payoﬀ to the DL process. Although there
is no exact analytical solution for the value of the discrete arithmetic Asian option under
the DL process, it could be computed by shifting the numerical solution for the discrete
arithmetic Asian option under Geometric Brownian motion, such as Milevsky and Posner
(1998), Fusai and Meucci (2008).
Here we consider a slightly diﬀerent payoﬀ which we know how to compute exactly under
the DL process. We call the new payoﬀ discrete exponential Asian option. The DL process
is driven by the same Brownian motion which drive the SABR process, W = W. More
33
precisely, the control is,
Y
∗
:=
_
_
_
(θ + exp(
1
N
N
i=1
log(S
i
−θ)) −K)
+
if displaced lognormal is used
(θ −exp(
1
N
N
i=1
log(θ −S
i
)) −K)
+
if displaced anti lognormal is used
dS
t
= σ(S
t
−θ)dW
t
(5.12)
The value C
∗
= EY
∗
can be calculated analytically under the DL process.
Lognormal as control variate
To make a comparison, we also use the control variate under the lognormal(LN) process.
The control variate is constructed by applying the discrete exponential Asian option to the
lognormal process, driven by the same Brownian motion W = W. More precisely,
Y
∗
:= (exp(
1
N
N
i=1
log(S
i
)) −K)
+
dS
t
= σ
LN
S
t
dW
t
(5.13)
The value C
∗
= EY
∗
can be calculated analytically. In (5.13), we choose σ
LN
to be the
shortexpiry atthemoney implied volatility of the SABR process, which is
σ
LN
= α
0
S
β−1
0
. (5.14)
5.1.5 Numerical results II: Discretely sampled arithmetic Asian option
under SABR dynamics
Our experiments simulate the payoﬀ (5.11), where
K = S
0
= 100, T = 4/12, N = 84, β = 0.2, α
0
S
β−1
0
= 0.2 (5.15)
and an array of choices for (ν, ρ). We use control (5.12) where (θ, σ) are tuned to SABR by
(4.9). We also report variance reduction results using control (5.13) with parameters (5.14).
We run 1000 paths. Ten equal spaced points are sampled each day. Table 5.1.4 and 5.1.5
report the estimated “percentage reduction of variance”. The results suggest that the DL
control achieves signiﬁcantly larger variance reduction than the LN control.
34
Table 5.1.4: Percentage reduction of variance, using DL as control for Asian Option on
SABR
T = 4 months
ρ ν
0.20 0.40 0.60
0.4 99.57% 98.43% 96.19%
0.6 99.69% 98.74% 97.36%
0.8 99.83% 99.24% 98.21%
Payoﬀ: (5.11) with (5.15). Exponential Asian option on DL control: (5.12) with (4.9).
Table 5.1.5: Percentage reduction of variance, using LN as control for Asian Option on SABR
T = 4 months
ρ ν
0.20 0.40 0.60
0.4 75.34% 78.61% 76.06%
0.6 76.61% 77.97% 78.95%
0.8 78.96% 80.76% 79.18%
Payoﬀ: (5.11) with (5.15). Exponential Asian option on LN control: (5.13) with (5.14).
5.2 Variance Reduction Combining Control Variate and
Importance Sampling
In this section, we study the issue of combining the importance sampling with the control
variate. Importance sampling itself is a popular variance reduction technique, see Bolye,
Broadie and Glasserman (1997) for more details.
5.2.1 Importance sampling on options pricing
Recall the SABR process in (4.6), if we deﬁne
d
˜
W
t
:= dW
t
−a
t
dt,
then by the Girsanov Theorem, there is a measure
˜
P such that under which
˜
W
t
, W
∗
t
are
independent standard Brownian motions, and W
t
is a drifted Brownian motion. Under
˜
P,
35
the original process S becomes a drifted SABR process:
dS
t
= α
t
S
β
t
(d
˜
W
t
+ a
t
dt), S
0
> 0
dα
t
= να
t
dB
t
, α
0
> 0
dB
t
= ρd
˜
W
t
+
_
1 −ρ
2
dW
∗
t
(5.16)
Again by the Girsanov Theorem, the likelihood ratio or RadonNikodym derivative be
tween the original measure P and the new measure
˜
P is
r(S) :=
dP
d
˜
P
= exp
_
−
_
T
0
a
s
dW
s
+
1
2
_
T
0
a
2
s
ds
_
. (5.17)
For any function G(.) : C →R, where C is the domain of S, the Girsanov Theorem says
EG(S) =
˜
Er(S)G(S), (5.18)
where
˜
E is the expectation which is taken under
˜
P. Therefore, the importance sampling
estimator (ISestimator) is:
ˆ
C
IS
:=
1
M
m
r(S
m
)G(S
m
). (5.19)
r(S
m
)G(S
m
) can be considered as a weighted payoﬀ.
The drift a
t
is chosen in a heuristic way similar to Bolye et al (1997). Denote
a
0
=
_
 log(S/K) + log(K/H)
_
/T +
1
2
α
2
0
S
2β−2
0
α
0
S
β−1
0
. (5.20)
We would like to control the variance of the likelihood ratio, which is exp(
_
T
0
a
2
s
ds), so that
it does not explode in real applications. We will arbitrarily set a constant C
a
< ∞ and
set ¯ a
0
= min(C
a
, a
0
). Next, for the downandout call options, we choose a
t
= ¯ a
0
; for the
downandin call options, we choose a
t
= −¯ a
0
before the barrier is reached and a
t
= ¯ a
0
after
the barrier is reached. The intuition behind this is to make the path less(more) likely to
reach the barrier in the downandout(downandin) case.
36
5.2.2 Drifted DH/DL process
The drifted DH/DL dynamics is derived similarly as the drifted SABR dynamics. Recall that
if we deﬁne d
˜
W
t
:= dW
t
−a
t
dt, then under the new measure
˜
P,
˜
W
t
and W
∗
t
are independent
standard Brownian motions. Consequently, the DL dynamics deﬁned in (2.6) and (2.15),
DH dynamics deﬁned in (3.1) and (3.16) become drifted DL/DH processes under the new
measure
˜
P. These drifted processes are listed below.
Under
˜
P, the displaced lognormal dynamics becomes the drifted displaced lognormal dy
namics:
dS
t
= σ(S
t
−θ)(d
˜
W
t
+ a
t
dt), S
0
> θ, σ > 0. (5.21)
The drifted displaced antilognormal dynamics is
dS
t
= σ(S
t
−θ)(d
˜
W
t
+ a
t
dt), 0 < S
0
< θ, σ < 0. (5.22)
The drifted displaced Heston dynamics is
dS
t
= σ
t
(S
t
−θ)(d
˜
W
1t
+ a
t
dt), S
0
> θ
dσ
2
t
= κ(ν −σ
2
t
)dt + σ
t
dB
t
,
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
.
(5.23)
The drifted displaced antiHeston dynamics is
dS
t
= (−σ
t
)(S
t
−θ)(d
˜
W
1t
+ a
t
dt), 0 < S
0
< θ
dσ
2
t
= κ(ν −σ
2
t
)dt + σ
t
dB
t
,
dB
t
= ρdW
t
+
_
1 −ρ
2
dW
∗
t
.
(5.24)
In both the drifted displaced Heston model and the drifted displaced anti Heston model,
σ
t
is nonnegative for all t > 0.
5.2.3 Combine control variate with importance sampling
Since the ISestimator
ˆ
C
IS
has smaller variance than the naive estimator
ˆ
C, we can construct
a control variate for the ISestimator. We call the new estimator ISCVestimator, and denote
it as
ˆ
C
ISCV
.
ˆ
C
ISCV
will reduce variance further than
ˆ
C
IS
.
37
The detail of the ISCVestimator is as follows. First, considering the weighted payoﬀ in
(5.19) as one payoﬀ function on S:
Y = H(S) := r(S)G(S), (5.25)
where S is sampled from (5.16). As in section 5.1.2, we want to choose a control variate Y
that is highly correlated with Y. Y is constructed by applying the new payoﬀ function H(.)
onto the drifted DL/DH process S:
Y := H(S) = r(S)G(S). (5.26)
It turns out that H(S) = r(S)G(S) can also be considered as an ISestimator for the non
drifted DL/DH process S, because by the Girsanov Theorem,
EG(S) =
˜
Er(S)G(S). (5.27)
Finally, we propose the combined estimator of the importance sampling and the control
variate as follows.
ˆ
C
ISCV
=
1
M
m
r(S
m
)G(S
m
) −β
_
1
M
m
r(S
m
)G(S
m
) −EG(S)
_
. (5.28)
where S
i
is sampled from (5.16), and S
i
is sampled from any of the drifted DL/DH process
(5.21)(5.24).
Remark 5.2.1. This estimator coincides with the estimator proposed by Hesterberg (1996),
but comes from a diﬀerent angle. In Hesterberg’s approach, both the importance sampling
estimator and the control variate estimator are regarded as weighted sums and are com
bined together as a double weighted sum. Our approach is to ﬁnd a control variate for the
importance sampling estimator.
5.2.4 Example: Discretely sampled barrier option under SABR dynamics
We consider a discretely sampled barrier option on S, which follows the SABR (4.6) dynamics.
In particular, let the contract be a downandin call with expiry T, barrier H, strike K,
38
sampling dates t
1
< t
2
< · · · < t
N
= T, and payoﬀ:
G(S) := (S
T
−K)
+
1( min
1≤n≤N
S
t
n
< H). (5.29)
To construct the
ˆ
C
ISCV
, we apply the payoﬀ function G(.) to a drifted DL/DH process S,
where S is driven by the Brownian motion W = W. If using DH, the volatility of S is driven
by W
∗
= W
∗
. in the DL case, (σ, θ) are tuned by (4.9); in the DH case, (σ
0
, κ, ν, ρ, , θ) are
tuned by (4.2) where a, b, c are determined by (4.12).
In practice, there is no way of exactly sampling from the continuous SABR process and
the correlated DH process, so discretisation schemes are used. We discretize the SABR
process into I subintervals and sample discretely. The discretisation of the SABR process is:
log S
i+1
= log S
i
−
α
2
i
2
S
2β−2
i
∆t + α
i
S
β−1
i
W
i
log σ
i+1
= log σ
i
−
ν
2
2
∆t + ν(ρ(W
i
−E(W
i
)) +
_
1 −ρ
2
W
∗
i
)
(5.30)
where W
i
, W
∗
i
∼ i.i.d. N(0, ∆t), i = 1, · · · , I. To sample from the drifted SABR process,
we take W
i
∼ N(a
i
∆t, ∆t) in (5.30). a
i
is chosen by the rule at the end of section 5.2.1.
The likelihoodratio is calculated as
r(S) =
I
i=1
f(W
i
)
f
α
(W
i
)
= exp(−
I
i=1
a
i
W
i
+
1
2
I
i=1
a
2
i
∆t). (5.31)
where f(.) is the density function of N(0, ∆t) and f
α
(.) is the density function of N(a
i
∆t, ∆t).
To construct ISCV with DL control, we sample from the drifted DL process and apply
the payoﬀ G(.) to it:
G(S) := (S
T
−K)
+
1( min
1≤n≤N
S
t
n
< H),
log(S
i+1
−θ) = log(S
i
−θ) −
σ
2
2
∆t + σW
i
, if b > −
1
2
log(θ −S
i+1
) = log(θ −S
i
) −
σ
2
2
∆t + σW
i
, if b < −
1
2
(5.32)
where S
0
= S
0
, W = W ∼ N(a
i
∆t, ∆t). We are not showing the degenerate case where
b =
1
2
and S
i
is sampled from a geometric Brownian motion.
39
Similarly, to construct ISCV with DH control, we sample from the drifted DH process
and apply the same payoﬀ G(.) to it:
G(S) := (S
T
−K)
+
1( min
1≤n≤N
S
t
n
< H),
log(S
i+1
−θ) = log(S
i
−θ) −
σ
2
i
2
∆t +σ
i
W
i
, if b > −
1
2
log(θ −S
i+1
) = log(θ −S
i
) −
σ
2
i
2
∆t −σ
i
W
i
, if b < −
1
2
σ
2
i+1
= σ
2
i
+ κ(µ −σ
2
i
)∆t +
_
σ
2
i
(ρ(W
i
−E(W
i
)) +
_
1 −ρ
2
W
∗
i
),
(5.33)
where S
0
= S
0
, W = W ∼ N(a
i
∆t, ∆t), W
∗
= W
∗
∼ N(0, ∆t), and σ
t
is nonnegative for
all t > 0.
Remark 5.2.2. The volatility process of DH is a meanreverting square process. There are
various discretisation schemes, see Lord et al (2008) for more detail. We choose the ﬁrst
order Euler discretisation scheme with reﬂection principal proposed by Higham and Mao
(2005), where the partial strong convergence of S
i
is ensured. More discussion of this is in
Chapter 6.
The likelihood ratio in both (5.32) and (5.33) is the same as in (5.31):
r(S) = exp(−
I
i=1
a
i
W
i
+
1
2
I
i=1
a
2
i
∆t) (5.34)
Finally, the ISCVestimator is constructed as
ˆ
C
ISCV
=
1
M
m
_
G(S
m
)r(S
m
) −βG(S
m
)r(S
m
) + βEG(S
m
)
_
. (5.35)
5.2.5 Numerical results: Discretely sampled barrier option
In this section, we compare the variance reduction of the importance sampling, with that of
combining importance sampling with the control variate. Denote ISDL to be the estimator
which combines the importance sampling and DL control variate from (5.32) as in (5.35).
Denote ISDH to be the estimator which combines the importance sampling and the DH
control variate from (5.33) as in (5.35).
40
The experiments simulate the payoﬀs (5.4) and (5.29), where
K = S
0
= 100, H = 95, T = 4/12, N = 84, β = 0.2, α
0
S
β−1
0
= 0.2. (5.36)
We simulate 1000 pathes. For the DH parameters in ISDH, preﬁx ρ = 0, κ = 0.2. Table 5.2.1
and 5.2.2 show the variance reduction of the downandout call options using importance
sampling, ISDL and ISDH. Table 5.2.3 and 5.2.4 show the variance reduction of the down
andin call options using the three estimators. For both options, we see that ISDL and ISDH
achieve more variance reduction than importance sampling.
Table 5.2.1: Percentage reduction of variance, using importance sampling for downandout
call on SABR
T = 4 months, Importance Sampling
ρ ν
0.20 0.40 0.60
0.4 69.46% 68.32% 67.00%
0.6 69.07% 67.70% 66.26%
0.8 68.77% 67.31% 65.69%
Payoﬀ: (5.4) with (5.36).
Table 5.2.2: Percentage reduction of variance, using ISDL and ISDH for downandout call
on SABR
T = 4 months, ISDL
ρ ν
0.20 0.40 0.60
0.4 97.02% 89.99% 83.31%
0.6 97.32% 90.91% 85.71%
0.8 97.48% 94.55% 89.30%
T = 4 months, ISDH
ρ ν
0.20 0.40 0.60
0.4 98.16% 98.87% 97.20%
0.6 97.91% 97.34% 96.87%
0.8 97.69% 95.55% 96.45%
Payoﬀ: (5.4) with (5.36).
ISDL: Combined importance sampling and DL control variate: (5.32) and (5.35). Parameters
of DL: (4.9).
ISDH: Combined importance sampling and DH control variate: (5.33) and (5.35). Parame
ters of DH: (4.2).
41
Table 5.2.3: Percentage reduction of variance, using importance sampling for downandin
call on SABR
T = 4 months, Importance Sampling
ρ ν
0.20 0.40 0.60
0.4 33.71% 32.35% 32.13%
0.6 35.85% 33.09% 30.97%
0.8 35.95% 33.25% 32.10%
Payoﬀ: (5.29) with (5.36).
Table 5.2.4: Percentage reduction of variance, using ISDL and ISDH for downandin call on
SABR
T = 4 months, ISDL
ρ ν
0.20 0.40 0.60
0.4 96.06% 87.59% 79.68%
0.6 96.58% 88.80% 82.50%
0.8 96.77% 93.42% 87.49%
T = 4 months, ISDH
ρ ν
0.20 0.40 0.60
0.4 97.89% 98.75% 95.96%
0.6 97.36% 97.00% 94.46%
0.8 97.06% 94.01% 94.63%
Payoﬀ: (5.29) with (5.36).
ISDL: Combined importance sampling and DL control variate: (5.32) and (5.35). Parameters
of DL: (4.9).
ISDH: Combined importance sampling and DH control variate: (5.33) and (5.35). Parame
ters of DH: (4.2).
We summarize the discussion of this Chapter here. As mentioned by Fisher and Tataru
(2010), the stochastic/local volatility (SLV) model has become market standard for barrier
option pricing. The CEV/SABR model belongs to the family of SLV models, so pricing
barrier options under them is of practical interest. This section uses DL and DH as control
variate to reduce variance in the Monte Carlo simulation of the barrier options. More
speciﬁcally, we give numerical examples to demonstrate that the DL and DH controls can
provide signiﬁcant variance reduction for barrier option pricing under the CEV/SABR model,
as shown in Tables 5.1.1, 5.1.2 and 5.1.3. Moreover, we show that DL/DH in concert with
importance sampling is superior to the importance sampling alone, by comparing Table 5.2.1
with 5.2.2, as well as Table 5.2.3 with 5.2.4. Finally, we show that the DL controls also yield
signiﬁcant variance reduction for Asian options.
42
CHAPTER 6
DISCRETISATION SCHEME
For most general stochastic processes, exact simulation methods do not exist and we have
to refer to discretisation schemes to simulate them. There are three basic criteria in the
discretisation schemes: rate of convergence, stability and positivity. These criteria have been
widely and intensively discussed in the literature. For more detail, see Lord et al (2008),
Kahl (2004), Zhang et al (2004) and Higham and Mao (2005). We use some discretisation
schemes in section 5.2.5 and this chapter we will validate their usage.
There are many numerical schemes. EulerMaruyama (Euler for short) scheme is a
straightforward discretisation scheme, wherein one discretises the time interval of interest,
and simulates the process at the discretisation points. Under certain conditions, it can be
shown that the Euler scheme converges to the true process as the time intervals are made
ﬁner and ﬁner. Sometimes, in order to reserve the positivity of the underlying process, the
LogEuler scheme is used, where one simulates from the discretisation of the logarithm of
the original process.
Consider a stochastic process:
dS
t
= A(t, S
t
)dt + B(t, S
t
)dW
t
, (6.1)
where W
t
is Brownian motion. Denote the discrete approximation of the process as S
∆t
t
.
Most existing proofs of the convergence of S
∆t
t
to S
t
rely on the linear growth and global
Lipschitz conditions (Kloeden and Platen (1992)):
(linear growth) there exists a positive constant L
1
such that
B(t, x)
2
≤ L
1
(1 + x
2
); (6.2)
(global Lipschitz ) there exists a positive constant L
2
such that
B(t, x) −B(t, y)
2
≤ L
2
x −y
2
. (6.3)
43
However, for the CEV/SABR processes, the strong convergence rule shown in Kloeden
and Platen (1992) can not be applied since S
β
t
does not satisfy the global Lipschitz condition.
Zhang et al (2004) relax the condition of the convergence to local Lipschitz. For the process
(6.1), they show that its Euler discretisation converges to the underlying continuous process
in L
2
sense before a stopping time, if S
t
satisﬁes a local Lipschitz condition before the
stopping time. Since the convergence they prove is up to a stopping time, we refer to it
as partial strong convergence. The result can be extended to the LogEuler discretisation:
if the logarithm of the underlying process satisﬁes the local Lipschitz condition before the
stopping time, then the LogEuler discretisation converges to the logarithm of the underlying
process. We extend the result further to a stochastic volatility case, where A(.), B(.) could
depend on a stochastic volatility process σ
t
. This result serves as the theoretical foundation
of the Monte Carlo simulation of the CEV/SABR process.
6.1 Partial Strong Convergency of Stochastic Volatility Process
Consider the stochastic process
dS
t
= A(t, S
t
, σ
t
)dt + B(t, S
t
, σ
t
)dW
1t
, (6.4)
dσ
t
= C(σ
t
)dW
2t
, dW
1t
dW
2t
= ρdt (6.5)
Approximate S
t
, σ
t
using Euler scheme on time points t
n
= n∆t, where n = 1, ...N, and
N∆t = T. The discretisation scheme is given as
S
∆t
t+∆t
= S
∆t
t
+ A(t, S
∆t
t
, σ
∆t
t
)∆t + B(t, S
∆t
t
, σ
∆t
t
)∆W
1,t
,
σ
∆t
t+∆t
= f(σ
∆t
t
, ∆t, ∆W
2,t
),
(6.6)
where ∆W
i,t
= W
i,t+∆t
−W
i,t
, i = 1, 2.
Remark 6.1.1. We do not specify the discretisation scheme of σ
t
here. It could be any
reasonable discretisation scheme, for example, the Euler scheme as:
σ
∆t
t+∆t
= σ
∆t
t
+ C(σ
∆t
t
)∆W
2,t
; (6.7)
44
or the LogEuler scheme as:
σ
∆t
t+∆t
= σ
∆t
t
exp
_
C(σ
∆t
t
)
σ
∆t
t
∆W
2,t
−
1
2
_
C(σ
∆t
t
)
σ
∆t
t
_
2
∆t
_
. (6.8)
For the SABR process, the σ
t
can be simulated exactly using the LogEuler scheme.
Deﬁne the local Lipschitz condition (*) as follows: let Ω
1
, Ω
2
be compact sets, there exist
positive constants K
1
(Ω
1
×Ω
2
) and K
2
(Ω
1
×Ω
2
), such that for any x
1
, x
2
∈ Ω
1
, z
1
, z
2
∈ Ω
2
,
A(t, x
1
, z
1
)−A(t, x
2
, z
2
)
2
∨B(t, x
1
, z
1
)−B(t, x
2
, z
2
)
2
≤ K
1
(Ω
1
×Ω
2
)x
1
−x
2

2
+K
2
(Ω
1
×Ω
2
)z
1
−z
2

2
.
(6.9)
If local Lipschitz condition (*) holds, then for the bounded Ω
1
, Ω
2
, there exists a positive
constant K
3
(Ω
1
×Ω
2
), such that for all x ∈ Ω
1
, z ∈ Ω
2
,
A(t, x, z)
2
∨ B(t, x, z)
2
≤ K
3
(Ω
1
×Ω
2
). (6.10)
Theorem 12 says that under the local Lipschitz condition (*) and some other regularity
conditions, the discretized process S
∆t
t
will have L
2
convergency to the true process S before
a stopping time.
Theorem 12. Let Ω
1
, Ω
2
be bounded regions and denote the stopping time
τ = inf{t ≥ 0 : S
∆t
t
/ ∈ Ω
1
or S
t
/ ∈ Ω
1
or σ
t
/ ∈ Ω
2
}.
For discretisation scheme in (6.6), if the following conditions are satisﬁed:
i. A(t, S
t
, σ
t
) and B(t, S
t
, σ
t
) satisfy the local Lipschitz condition (*);
ii. for ∆t small enough and t ∈ [t
n
, t
n+1
), there is a constant D(Ω
2
) such that
E(σ
t
−σ
∆t
t
n
)
2
≤ D(Ω
2
)∆t; (6.11)
then for ∆tT < 1, ∃ constant C(Ω
1
×Ω
2
), such that
E
_
sup
0≤t≤τ∧T
S
∆t
t
−S
t

2
_
≤ C(Ω
1
×Ω
2
)∆t. (6.12)
45
In other words, as long as S
∆t
t
and S
t
remain in the domain Ω
1
and σ
t
remains in the
domain Ω
2
, the Euler scheme S
∆t
t
converges to the true process S
t
as ∆t → 0.
Proof. Appendix C.1.
Remark 6.1.2. If σ
t
follows the geometric Brownian motion such that σ
t
= νσ
t
dW
t
, then
σ
∆t
t
n
can be simulated exactly as σ
t
n
. Since σ
t
= σ
t
n
e
∆sνX−
1
2
∆sν
2
where ∆s = t − t
n
and
X ∼ N(0, 1), substitute σ
∆t
t
with σ
t
n
, and we have
E(σ
t
−σ
∆t
t
n
)
2
= E(σ
t
n
)
2
(e
∆sν
2
−1) = e
t
n
ν
2
(e
∆sν
2
−1).
Since ∆s < ∆t, for any t
n
< T and small enough ∆t, there is a constant K such that
e
t
n
ν
2
(e
∆sν
2
−1) < K∆t. Therefore σ
t
, σ
∆t
t
satisfy the condition (ii) in Theorem 12.
Remark 6.1.3. In the SABR process, we have A(t, S
t
, σ
t
) = 0 and B(t, S
t
, σ
t
) = σ
t
S
β
t
. In the
logarithm of the SABR process, we have A(t, S
t
, σ
t
) =
1
2
σ
2
t
S
2β−2
t
and B(t, S
t
, σ
t
) = σ
t
S
β−1
t
.
In both cases, A(.) and B(.) satisfy the local Lipschitz condition (*). (see appendix C.2
for detail.) So, Theorem 12 indicates that the Euler (LogEuler) discretisation scheme of
the SABR process (with σ
t
simulated exactly) will converge to the true process on the time
interval τ ∨T, where τ = inf{t ≥ 0 : S
∆t
t
/ ∈ [a
1
, a
2
] or S
t
/ ∈ [a
1
, a
2
] or σ
t
/ ∈ [b
1
, b
2
]; 0 < a
1
<
a
2
, 0 < b
1
< b
2
}. This holds for the CEV process too since the CEV process is a special
case of the SABR process. In terms of our Monte Carlo simulation of the CEV/SABR
process, if S
t
starts not close to 0, we could construct an interval [a
1
, a
2
] such that with
large probability, the simulated path will be within the range, and we could consider the
discretized path as a good approximation of the true continuous path.
6.2 Strong Convergence of Meanreverting CEV Process
Kahl and Jackel (2006) consider a stochastic volatility process which they call it mean
reverting CEV process. It is deﬁned as follows:
dV
t
= λ(µ −V
t
)dt + σV
β
t
dW
t
. (6.13)
They point out that the boundary behavior of the process as follows.
1. 0 is an attainable boundary for 0 < β <
1
2
and for β =
1
2
if λµ <
σ
2
2
.
46
2. 0 is unattainable for β >
1
2
.
3. ∞ is unattainable for all β > 0.
They also develop a numerical scheme to simulate the stock process whose volatility process
is the meanreverting CEV process.
There is a lot literature on the particular case of the meanreverting CEV process, when
β =
1
2
, which is called the meanreverting squareroot process. In the Heston model, the
volatility process is the meanreverting squareroot process. For the meanreverting square
root process, Higham and Mao (2005) propose a discretisation scheme called the reﬂection
Euler scheme:
V
∆t
t+∆t
= V
∆t
t
+ λ(µ −V
∆t
t
)∆t + σ
_
V
∆t
t
∆W
t
, ∆W
t
= W
t+∆t
−W
t
. (6.14)
They prove that this discretisation scheme has the strong convergency property. They also
propose a discretisation scheme for the Heston model as follows:
S
∆t
t+∆t
=
_
V
∆t
t
S
∆t
t
∆W
1,t
,
V
∆t
t+∆t
= V
∆t
t
+ λ(µ −V
∆t
t
)∆t + σ
_
V
∆t
t
∆W
2,t
, corr(∆W
1,t
, ∆W
2,t
) = ρ.
(6.15)
They prove that such discretized process S
∆t
will converge to the true process S. This result
serves as the theoretical foundation for our Monte Carlo simulation of the DH process.
We extend the strong convergence to the general meanreverting CEV process under the
following conditions:
Condition A: 0 < λ∆t < 2 and (1 −λ∆t)
2
+ βσ
2
∆t < 1.
Condition B:
1
2
≤ β < 1.
We consider the discretisation scheme with reﬂection principle on the time points t
n
=
n∆t, with n = 1, ..., N and N∆t = T as follows:
V
∆t
t+∆t
= V
∆t
t
+ λ(µ −V
∆t
t
)∆t + σV
∆t
t

β
∆W
t
, ∆W
t
= W
t+∆t
−W
t
. (6.16)
Proposition 6.2.1 gives the ﬁrst moment of the discretized process (6.16). Proposition 6.2.2
says that the second moment of the discretized process (6.16) is bounded. Theorem 13 says
that the discretisation scheme (6.16) converges to the true process (6.13) as ∆t → 0 in L
1
sense.
47
Proposition 6.2.1. For the discretized process (6.16),
EV
∆t
n
= (1 −λ∆t)
n
(EV
∆t
0
−µ) + µ, (6.17)
so that when 0 < λ∆t < 2,
lim
t→∞
EV
∆t
t
→ µ. (6.18)
Proof. For SDE in (6.16), take the expectation of both sides,
EV
∆t
n+1
= EV
∆t
n
(1 −λ∆t) + λ∆tµ (6.19)
⇒ EV
∆t
n+1
−µ = (1 −λ∆t)(EV
∆t
n
−µ) (6.20)
⇒ EV
∆t
n+1
= (1 −λ∆t)
n
(EV
∆t
0
−µ) + µ. (6.21)
So when 0 < λ∆t < 2, lim
t→∞
EV
∆t
t
→ µ; otherwise, lim
t→∞
EV
∆t
t
→ ∞.
Proposition 6.2.2. For the discretized process (6.16), under conditions A and B, E(V
∆t
n
)
2
and E(V
∆t
n
)
2β
is bounded for all n.
Proof. Appendix C.3.
Theorem 13. For the meanreverting CEV process in (6.13), under conditions A and B,
the discretisation process in (6.16) will have the L
1
convergence:
lim
∆t→0
sup
0≤t≤T
EV
t
−V
∆t
t
 = 0. (6.22)
Proof. Appendix C.3.
6.3 Discretisation Schemes Used in the Monte Carlo Simulation
In the Monte Carlo simulation conducted in Chapter 5, we use discretisation schemes to
simulate the processes whose exact simulation is diﬃcult. Based on the discussion in section
6.1 and 6.2, the ﬁrst order LogEuler scheme is applied to simulate the DL process and
the CEV/SABR process. The ﬁrst order Euler scheme with reﬂection principle is applied to
simulate the DH process. The discretisation schemes for each process are summarized below.
48
First order LogEuler scheme:
Denote z ∼ N(0, ∆t), w ∼ N(0, ∆t), w⊥z.
Discretisation of the displaced lognormal process (exact simulation):
∆log(S
∆t
t
−θ) = −
σ
2
2
∆t + σz
Discretisation of the displaced anti lognormal process (exact simulation):
∆log(θ −S
∆t
t
) = −
σ
2
2
∆t + σz.
Discretisation of the CEV/SABR process:
∆log S
∆t
t
= α
t
(S
∆t
t
)
β−1
∆t −
1
2
α
2
t
(S
∆t
t
)
2β−2
z,
∆log α
t
= −
ν
2
2
∆t + ν(ρz +
_
1 −ρ
2
w).
First order LogEuler scheme with reﬂection principle:
Denote z ∼ N(0, ∆t), w ∼ N(0, ∆t), w⊥z.
Discretisation of the displaced Heston process:
∆log(S
∆t
t
−θ) = −
V
∆t
t
2
∆t +
_
V
∆t
t
z;
∆V
∆t
t
= κ(µ −V
∆t
t
)∆t +
_
V
∆t
t
w.
Discretisation of the displaced anti Heston process:
∆log(θ −S
∆t
t
) = −
V
∆t
t
2
∆t −
_
V
∆t
t
z;
∆V
∆t
t
= κ(µ −V
∆t
t
)∆t +
_
V
∆t
t
w.
49
CHAPTER 7
LARGEEXPIRY IMPLIED VOLATILITY OF DISPLACED
LOGNORMAL
7.1 Largestrike and Largeexpiry Behavior
In this section, we analyze the implied volatility of the displaced lognormal dynamics (2.6)
in the largestrike and largeexpiry region. Our goal is to give a largeexpiry asymptotes of
σ
imp
(K, T).
7.1.1 Case one: K = S
0
e
xT
, x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
]
First, consider the strike to be K = S
0
e
xT
, where x is constant. Denote ˜ x(T) =
1
T
log((K−
θ)/(S
0
−θ)). Deﬁne the domain of T as follows:
K
T
:= {T > 0 : S
0
e
xT
> θ
+
and (S
0
−K)
+
< C
BS
(S
0
−θ, S
0
e
xT
−θ, σ, T) < S
0
}. (7.1)
The implied volatility σ
imp
(x, T) is deﬁned on K
T
such that
C
BS
(S
0
, K, σ
imp
(x, T), T) = C
BS
(S
0
−θ, K −θ, σ, T). (7.2)
We can suppress the x in σ
imp
(x, T) to be σ
imp
(T) when no ambiguity arises.
Theorems 14, 15 and 17 give the largeexpiry asymptotic approximation of σ
imp
(x, T),
where Theorem 14 focuses on the region where x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
], Theorems 15 and 17 on
the region where x ∈ (−σ
2
/2, σ
2
/2).
Denote
A(T) =
(−˜ x(T) +
1
2
σ
2
)
2
2σ
2
(7.3)
and deﬁne the domain
K
x
:= {T ∈ R
+
: A(T) + x ≥ 0} (7.4)
50
Theorem 14. For the displaced lognormal mordel (2.6), if x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
] and T ∈
K
T
∩ K
x
, then the following asymptotic limit of σ
2
imp
(x, T) holds.
σ
2
imp
(x, T) = σ
2
∞
(x, T) + ˆ a
1
(x, T)/T + o(1/T) as T → ∞, (7.5)
where
σ
2
∞
(x, T) := 2
_
2A(T) + x −2
_
A
2
(T) + A(T)x
_
, if x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
] (7.6)
A
BS
(x, σ, a
1
) := exp
_
1
8
a
1
_
4x
2
σ
4
−1
__
σ
3
x
2
−σ
4
/4
1
x=±σ
2
/2
(7.7)
ˆ a
1
(x, T) = 2(
x
2
σ
4
∞
(x)
−
1
4
)
−1
log
_
˜
S
0
S
0
A
BS
(˜ x, σ, 0)
A
BS
(x, σ
∞
, 0)
_
. (7.8)
Proof. Appendix D.1.
We suppress the T in ˜ x(T), A(T), σ
2
∞
(x, T), ˆ a
1
(x, T) to be ˜ x, A, σ
2
∞
(x), ˆ a
1
(x), with
the understanding that these variables all depend on T.
Remark 7.1.1. Figure 7.1.1 shows the exact σ
imp
and Theorem 14 formula. We can see that
the approximation is reasonably accurate.
Figure 7.1.1: Theorem 14 formula and Exact σ
imp
for T = 10.
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0.05
0.1
0.15
0.2
0.25
0.3
x
σ
im
p
S
0
=100, σ=0.2, θ=−20, T=10
S
0
=100, σ=0.2, θ=60, T=10
σ
imp
Theorem 14 formula
Remark 7.1.2. We can write A + x =
(−˜ x+σ
2
/2)
2
σ
2
+ (x − ˜ x) and lim
T→∞
˜ x(T) = x, so if
x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
], the condition A + x > 0 always holds for large T.
51
Remark 7.1.3. The derivation of Theorem 14 depends on whether ˜ x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
] or
not. However, since ˜ x will be in a close neighborhood of x when T is large, we express the
condition in terms of x.
Remark 7.1.4. If θ > 0 and x < 0, the strike K will eventually become less than θ as T
increases. Thus S
T
> θ > K, the option price will always be S
0
− K for large T. (7.5)
should be regarded as an approximation of the implied volatility on the domain of T where
S
0
e
xT
> θ
+
.
7.1.2 Case two: K = S
0
e
xT
, x ∈ (−σ
2
/2, σ
2
/2)
Theorem 15 gives the largeexpiry asymptotic approximation of σ
imp
(x, T) in the region
which is complement to that in Theorem 14. Theorems 17 and 18 reﬁne the approximation
to higher orders.
Theorem 15. For the displaced lognormal process (2.6), when x ∈ (−σ
2
/2, σ
2
/2),
1. If θ < 0, then σ
imp
(x, T) is not deﬁned for large T.
2. If θ > 0, for T ∈ K
T
and x ∈ [0, σ
2
/2), then
lim
T→∞
σ
imp
(x, T) = 2x. (7.9)
Proof. Appendix D.2.
Before stating Theorem 17, we give an asymptotic approximation of the BlackScholes
call option in the largestrike, largeexpiry case.
Theorem 16. For x > 0, for ∀B ∈ R, lim
T→∞
a(T) = const, we have the asymptotic
behavior for the BlackScholes call option formula in the largestrike, largeexpiry case as
follows.
1
S
0
C
BS
(S
0
, S
0
e
xT
,
¸
2x + 2B
_
2x
T
+
a(T)
T
, T) = N(B)+
e
−
1
2
B
2
(a(T)/2 −B
2
−1)
√
2πT
√
2x
(1+O(1/
√
T)).
(7.10)
Proof. Appendix D.3.
52
Remark 7.1.5. Forde et al (2009) give the asymptotic behavior of the BlackScholes call
option when the volatility is expressed as
_
σ
2
+ a/T. Theorem 16 is diﬀerent from theirs
in that we have order of 1/
√
T and we restrict the leading term σ
2
to be 2x.
Theorem 17 reﬁnes the asymptotic behavior given in Theorem 15. It elaborates to higher
order of approximation.
Figure 7.1.2: Theorem 17 formula and Exact σ
imp
.
2 3 4 5 6 7 8 9 10
x 10
4
0.0762
0.0763
0.0764
0.0765
0.0766
0.0767
0.0768
0.0769
0.077
Time
σ
i
m
p
S
0
=100, σ=0.2, θ=60, x=0.003
σ
imp
approx σ
imp
Theorem 17. For the displaced lognormal process (2.6), if θ > 0, x ∈ (0, σ
2
/2) and T ∈ K
T
,
the second order approximation of σ
imp
(x, T) when T → ∞ is given as:
σ
2
imp
(x, T) = 2x + 2B
_
2x
T
+
a(T)
T
+ o(
1
T
), (7.11)
where B = N
−1
(
S
0
−θ
S
0
), a(T) = 2(B
2
+1+C(T)e
1
2
B
2
√
2x), C(T) =
˜
S
0
S
0
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
(
σ
3
˜ x
2
−σ
4
/4
).
Proof. See appendix D.4.
Remark 7.1.6. Figure 7.1.2 shows the approximation in Theorem 17 is highly accurate when
T is large. In the plot, we have T ∈ [2 ×10
4
, 10
5
].
53
Theorem 18 reﬁnes Theorem 17’s approximation to the order of
1
T
√
T
.
Theorem 18. For the displaced lognormal process (2.6), if θ > 0, x ∈ (0, σ
2
/2) and T ∈ K
T
,
the third order approximation of σ
imp
(x, T) as T → ∞ is given as:
σ
2
imp
(x, T) = 2x + 2B
_
2x
T
+
a
T
+
d(T)
T
√
T
+ o(
1
T
√
T
) (7.12)
where B = N
−1
(
S
0
−θ
S
0
), a = 2(B
2
+ 1), d(T) =
5B
3
+3B+C(T)e
B
2
/2
4x
√
T
√
2x
,
C(T) =
˜
S
0
S
0
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
(
σ
3
˜ x
2
−σ
4
/4
).
Proof. Appendix D.4.
Figures 7.1.3 and 7.1.4 compare Theorem 17 and Theorem 18’s approximations of σ
imp
.
In both ﬁgures, the left plot is the Theorem 17 formula and the exact σ
imp
, and the right
plot is the Theorem 18 formula and the exact σ
imp
. In Figure 7.1.3, T ∈ [2000, 10000] and
in Figure 7.1.4, T ∈ [200, 1000]. All the other parameters are the same. We can see that for
T ∈ [2000, 10000], Theorem 18 signiﬁcantly improves the results of Theorem 17. However,
neither of them are a good approximation for T ∈ [200, 1000] . To sum up, Theorem 18 gives
a good approximation of σ
imp
when T is large.
Theorem 19 gives an explicit formula for the atthemoney implied volatility.
Theorem 19. For the displaced lognormal process (2.6), the atthemoney implied volatility
on T ∈ K
T
is
σ
imp
(0, T) =
2
√
T
N
−1
_
S
0
−θ
S
0
N(σ
√
T/2) +
θ
2S
0
_
. (7.13)
Thus
lim
T→∞
σ
imp
(0, T)
√
T = 2N
−1
_
1 −
θ
2S
0
_
(7.14)
Proof. Appendix D.5.
Remark 7.1.7. For σ
imp
(0, T) to exist, we need
S
0
−θ
S
0
N(σ
√
T/2) +
θ
2S
0
∈ [0, 1]. So, when
θ < 0, σ
imp
(0, T) becomes undeﬁned for large T.
54
7.1.3 Case three: K = S
0
e
xT
α
Theorem 14 can be extended to the region where K approaches ∞ such that K = S
0
e
xT
α
for x = 0 and α > 1. This is given in Theorem 20.
Denote ˜ x =
1
T
α
log(
S
0
e
xT
α
−θ
S
0
−θ
), M = xT
α−1
,
˜
M = ˜ xT
α−1
, A =
_
−
˜
M+σ
2
/2
σ
_
2
. Deﬁne
the domain of T as
K
T
α
:= {T > 0 : S
0
e
xT
α
> θ
+
and (S
0
−S
0
e
xT
α
)
+
< C
BS
(S
0
−θ, S
0
e
xT
α
−θ, σ, T) < S
0
},
(7.15)
the implied volatility σ
imp
(x, T) is well deﬁned by (2.5) when T ∈ K
T
α
.
Theorem 20. For displaced lognormal dynamics (2.6), for x > 0, T ∈ K
T
α
, we have
σ
2
imp
(x, T) = σ
2
∞
+ ˆ a
1
(T)/T + o(1/T), (7.16)
where
σ
2
∞
= 2(M + A −
_
2MA + A
2
), (7.17)
ˆ a
1
= 2(
M
2
σ
4
∞
−
1
4
)
−1
log
_
˜
S
0
S
0
A
BS
(
˜
M, σ, 0)
A
BS
(M, σ
∞
, 0)
_
, (7.18)
and A
BS
(.) is deﬁned by (7.7).
Proof. Appendix D.6.
Figure 7.1.5 and 7.1.6 compare the Theorem 20 formula and the exact σ
imp
. We can see
the approximation is accurate with moderate T. In Figure 7.1.5, the smallest K is about 700;
in Figure 7.1.6, the smallest K is about 200. Letting K approaches inﬁnity in the fashion of
K = S
0
e
xT
α
, we could have a highly accurate approximation of σ
imp
with reasonable expiry
time T.
Remark 7.1.8. In both Theorems 14 and 20, we have to make the strike K large to have
reasonably accurate approximation. Theorem 14 uses a large x (e.g. x = 0.2, T = 10, α = 1)
to obtain a good approximation; Theorem 20 uses a small x but a large α (e.g. x =
0.003, T = 10, α = 3).
55
7.2 Fixedstrike Largeexpiry Implied Volatility
In this section, we ﬁx the strike, and analyze the implied volatility of the displaced lognormal
dynamics when time to expiry is large. We show the asymptotic behavior and the mono
tonicity of the implied volatility. In this section, denote K = S
0
e
x
, where x is a constant
and deﬁne the domain of T as
¯
K
T
:= {T > 0, (S
0
−K)
+
< C
BS
(S
0
−θ, K −θ, σ, T) < S
0
}. (7.19)
σ
imp
(x, T) is welldeﬁned when T ∈
¯
K
T
by (2.5).
Theorem 21. For the displaced lognormal process (2.6), the ﬁxedstrike largeexpiry implied
volatility has the following properties.
1. (Asymptotic behavior).
(a) For all K > θ
+
, T ∈ K
T
and 2S
0
> θ > 0,
lim
T→∞
σ
imp
(x, T)
√
T = σ
blsimpv
(S
0
, K, S
0
−θ), (7.20)
where σ
blsimpv
(S
0
, K, P) is the implied volatility of the oneyear call option with
price P deﬁned as follows:
C
BS
(S
0
, K, σ
blsimpv
(S
0
, K, P), 1) = P. (7.21)
(b) For θ < 0, the σ
imp
(x, T) is not deﬁned when T → ∞.
2. (Monotonicity in time) For all θ > 0, K > θ
+
and T ∈ K
T
lim
T→∞
sgn
∂σ
imp
∂T
(x, T) = −sgn θ (7.22)
Proof. Appendix D.7.
Theorem 22 reﬁnes the results of Theorem 21.1. The notation A(T) ∼ B(T) means
A(T)/B(T) → 1 as T ↑ ∞.
56
Theorem 22. For the displaced lognormal dynamics with K = S
0
e
x
, the following holds for
the ﬁxedstrike, largetime implied volatility:
σ
imp
(x, T) ∼
1
√
T
_
σ
blsimpv
(S
0
, K, S
0
−θ) + z
_
, (7.23)
where z =
−1+
_
1+4M(−
1
8
+
x
2
8a
2
)
−
1
4
+
x
2
4a
2
, a =
σ
2
blsimpv
(S
0
,K,S
0
−θ)
2
,
M =
√
2π exp(
1
2
(
−x+a
√
2a
)
2
)
_
1
S
0
C
BS
(S
0
−θ, K −θ, σ, T) −
S
0
−θ
S
0
_
.
Proof. Appendix D.8.
Figure 7.2.1 shows Theorem 22 formula and exact σ
imp
. We can see that for θ close to
S
0
the approximation is quite accurate even for moderate large T.
57
Figure 7.1.3: Theorems 17 and 18 formula and Exact σ
imp
.
2000 3000 4000 5000 6000 7000 8000 9000 10000
0.076
0.0765
0.077
0.0775
0.078
0.0785
0.079
0.0795
Time
σ
im
p
S
0
=100, σ=0.2, θ=60, x=0.003
Exact σ
imp
approximated σ
imp
2000 3000 4000 5000 6000 7000 8000 9000 10000
0.0762
0.0764
0.0766
0.0768
0.077
0.0772
0.0774
0.0776
0.0778
0.078
Time
σ
im
p
S
0
=100, σ=0.2, θ=60, x=0.003
Exact σ
imp
approximated σ
imp
Figure 7.1.4: Theorems 17 and 18 formula and Exact σ
imp
.
200 300 400 500 600 700 800 900 1000
0.08
0.085
0.09
0.095
0.1
0.105
Time
σ
im
p
S
0
=100, σ=0.2, θ=60, x=0.003
Exact σ
imp
approximated σ
imp
200 300 400 500 600 700 800 900 1000
0.076
0.077
0.078
0.079
0.08
0.081
0.082
0.083
0.084
0.085
0.086
Time
σ
im
p
S
0
=100, σ=0.2, θ=60, x=0.003
Exact σ
imp
approximated σ
imp
58
Figure 7.1.5: Theorem 20 formula and Exact σ
imp
for T = 10.
3 4 5 6 7 8 9 10
x 10
−3
0.15
0.155
0.16
0.165
0.17
0.175
0.18
0.185
x
σ
i
m
p
S
0
=100, σ=0.2, θ=60 ,T=10, α=3
Exact σ
imp
approximated σ
imp
Figure 7.1.6: Theorem 20 formula and Exact σ
imp
for T = 7.
3 4 5 6 7 8 9 10
x 10
−3
0.105
0.11
0.115
0.12
0.125
0.13
0.135
0.14
0.145
0.15
x
σ
i
m
p
S
0
=100, σ=0.2, θ=60, α=2.8 ,T=7
Exact σ
imp
approximated σ
imp
59
Figure 7.2.1: Theorem 22 formula and Exact σ
imp
for T = 20.
−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
x
σ
i
m
p
S
0
=100, σ=0.4, θ=60, T=20
σ
imp
approx σ
imp
60
CHAPTER 8
CONCLUSION
By establishing properties shared by all displaced lognormal volatility skews, Theorems 1 and
2, in eﬀect, exhibit limitations on what phenomena the displaced lognormal can faithfully
model. In particular, the displaced lognormal’s everywheremonotonic skews (Theorem 1)
and its state space (θ, ∞) = R
+
may be drawbacks, when pricing contracts having sensitivity
to the tail behavior of an underlying process whose true state space is R
+
or whose true
volatility skew is nonmonotonic. The slopeconstrained skew (Theorem 2) of the displaced
lognormal may be a drawback, when modeling markets, such as typical equity markets,
which exhibit downward skews of steepness greater than the Theorem 2 upper bound. The
displaced anti lognormal overcomes the slope constraint, but it imposes bounds from above
on S, while allowing unbounded negative S. Furthermore, the displaced anti lognormal has
the everywheremonotonic skews (Theorem 4).
We therefore do not endorse the DL as a model of equity markets, but we do propose the
DL as a control variate to improve the accuracy of Monte Carlo pricing under alternative
dynamics (such as CEV and SABR) that do model equity prices. The DL is eﬀective as a
control variate, because it is simple enough to admit both unbiased simulation and explicit
pricing formulas for many contracts, yet ﬂexible enough to generate high correlations and
hence signiﬁcant variance reduction with respect to the CEV/SABR dynamics. We give
two examples. The ﬁrst is the downandout call option under the CEV/SABR dynamics.
The DL control generates large variance reduction as shown in Table 5.1.1 and 5.1.2. The
second example is the discrete arithmetic Asian option. We compare the variance reduction
of the DL control and the lognormal control. The DL control generates signiﬁcant more
satisfactory variance reduction than LN control as shown in Table 5.1.4 and 5.1.5.
Another displaced process, the displaced (anti )Heston process, will generate more vari
ance reduction when used as control variate. We are able to have explicit pricing formulas
for the calls/puts and barrier options on the DH process. Though no exact simulations of
the DH process are available, we have a discretisation scheme which guarantees convergence
61
to the true process. We show that DH control improves the variance reduction of DL control
when applied to the SABR dynamics in Table 5.1.3.
Moreover, we analyze the variance reduction by combining the DL/DH control variate
with importance sampling. The combined technique is superior to the pure importance
sampling when applied to options under the SABR dynamics, as shown in Table 5.2.1 and
5.2.2, Table 5.2.3 and 5.2.4.
Toward either purpose – as a model, or as a computational device on behalf of another
model – the DL calibrates easily to a given volatility skew level and slope, via the short
expiry limiting implied volatility formula of Theorems 3 and 6; and the DH calibrates easily
to a given volatility skew level, slope and convexity, via the shortexpiry limiting implied
volatility formula of Theorems 8 and 9.
Furthermore, we give an explicit relationship between the shortexpiry implied volatilities
of any process which is a positive martingale with its corresponding (anti)displaced process
(Theorems 10 and 11). Combined with Durrleman’s (2004) approximation of the shortexpiry
nearthemoney implied volatility, we are able to give an approximation of the shortexpiry
nearthemoney implied volatility of any (anti)displaced process.
In contrast to the shortexpiry limiting implied volatility, the largeexpiry limiting implied
volatility of the displaced lognormal is also analyzed. We ﬁrst give the approximation of the
implied volatility in the largestrike case (Theorems 14, 15, 17 and 18 where K = S
0
e
xT
;
Theorem 20 where K = S
0
e
xT
α
). Secondly, we give the approximation of the implied
volatility in the ﬁxedstrike case (Theorems 21 and 22).
Finally, we discuss some issues in the discretisation scheme of the Monte Carlo simulation.
Under some regularity conditions, we prove the partial strong convergency of the discretized
stochastic volatility process to the continuous process (Theorem 12). We also prove a strong
convergency result of the meanreverting CEV process (Theorem 13).
62
APPENDIX A
A.1 Appendix: Proof of Theorem 1–Global Behavior of
Displaced Lognormal
Deﬁne for all K > θ
+
d
2
(K) :=
log[(S
0
−θ)/(K −θ)]
σ
√
T
−
σ
√
T
2
.
Let K
S
(T) := inf K
S
(T). Our notation may suppress the S or T. Note that K = (K, ∞).
If θ ≥ 0 then K = θ. If θ < 0 then K > 0 because C
BS
(S
0
−θ, 0−θ, σ) > S
0
−θ+θ = S
0
.
Proposition A.1.1. For all T > 0 and K ∈ K(T):
If θ > 0 then σ
imp
(K) < σ. If θ < 0 then σ
imp
(K) > σ.
Proof. For all K > θ and S
0
> θ,
d
dθ
C
BS
(S
0
−θ, K −θ, σ) = −
∂C
BS
∂S
−
∂C
BS
∂K
= −N(d
2
+ σ
√
T) + N(d
2
) < 0 (A.1)
so, according as θ ≷ 0, we have C
BS
(S
0
−θ, K −θ, σ) ≶ C
BS
(S
0
, K, σ). By (2.9),
σ
imp
(K) ≶ σ. (A.2)
for all K ∈ K.
Proposition A.1.2. For all T > 0, we have σ
imp
(K) → σ as K → ∞.
Hence sup
K∈K(T)
σ
imp
(K) = σ if θ > 0; and inf
K∈K(T)
σ
imp
(K) = σ if θ < 0.
Proof. We prove for θ > 0. The proof for θ < 0 is similar.
By Proposition A.1.1, σ
imp
(K) < σ for all K ∈ K. So it suﬃces to show that for all
δ > 0, there exists k such that for all K > k,
σ −δ < σ
imp
(K)
63
or equivalently
C
BS
(S
0
, K, σ −δ) < C
BS
(S
0
, K, σ
imp
(K)) = C
BS
(S
0
−θ, K −θ, σ).
Hence it suﬃces that for K large enough,
0 < M(K) := C
BS
(S
0
−θ, K −θ, σ) −C
BS
(S
0
, K, σ −δ), (A.3)
which we verify as follows. For all K > θ
+
,
∂M
∂K
= −N
_
log[(S
0
−θ)/(K −θ)]
σ
√
T
−
σ
√
T
2
_
+ N
_
log(S
0
/K)
(σ −δ)
√
T
−
(σ −δ)
√
T
2
_
. (A.4)
So
sgn
∂M
∂K
= sgn
_
log(S
0
/K)
(σ −δ)
√
T
−
log[(S
0
−θ)/(K −θ)]
σ
√
T
+
δ
√
T
2
_
(A.5)
= sgn
_
log(S
0
/K)
σ
−log[(S
0
−θ)/(K −θ)]
σ−δ
(σ −δ)σ
√
T
+
δ
√
T
2
_
(A.6)
= sgn
_
log[(S
0
×(K −θ))/(K ×(S
0
−θ))]
σ
−log[(K −θ)/(S
0
−θ)]
δ
(σ −δ)σ
√
T
+
δ
√
T
2
_
.
(A.7)
The part inside the sgn in (A.7) approaches −∞ as K → ∞.
So for K suﬃciently large, ∂M/∂K(K) < 0.
Moreover lim
K→∞
M(K) = 0 −0 = 0.
Therefore M(K) > 0 for K suﬃciently large, which proves (A.3).
Lemma A.1.3. For all K ∈ K,
sgn
∂σ
imp
∂K
(K) = sgn F(σ
imp
(K), K)
where F : R ×(θ
+
, ∞) →R is deﬁned by
F(y, K) := −y
2
T/2 −y
√
Td
2
(K) + log(S
0
/K).
64
Proof. Take the K derivative of (2.9),
∂C
BS
∂K
(S
0
, K, σ
imp
(K)) +
∂C
BS
∂σ
(S
0
, K, σ
imp
(K))
∂σ
imp
∂K
(K) =
∂C
BS
∂K
(S
0
−θ, K −θ, σ).
(A.8)
Therefore
sgn
∂σ
imp
∂K
(K) = sgn
_
∂C
BS
∂K
(S
0
−θ, K −θ, σ) −
∂C
BS
∂K
(S
0
, K, σ
imp
(K))
_
= sgn
__
log(S
0
/K)
σ
imp
√
T
−
σ
imp
√
T
2
_
−d
2
(K)
_
= sgn
_
−σ
2
imp
(K)T/2 −σ
imp
(K)
√
Td
2
(K) + log(S
0
/K)
¸
= sgn F(σ
imp
(K), K),
as claimed.
Lemma A.1.4. There exists K ∈ K such that
sgn
∂σ
imp
∂K
(K) = sgn θ.
Proof. If θ = 0 then this is obvious.
If θ < 0 then as K ↓ K we have C
BS
(S
0
−θ, K −θ, σ) → S
0
hence σ
imp
→ ∞. So there
exists K > K such that ∂σ
imp
/∂K(K) < 0, as claimed.
If θ > 0 then the atthemoney strike K = S
0
satisﬁes the conclusion, because
sgn
∂σ
imp
∂K
(S
0
) = sgn
_
−σ
imp
(S
0
)
√
T/2 + σ
√
T/2
¸
> 0
by Proposition A.1.1.
Proposition A.1.5. For all T > 0 and K ∈ K(T),
sgn
∂σ
imp
∂K
(K) = sgn θ. (A.9)
Proof. If θ = 0 then this is obvious. Otherwise, by Lemma A.1.4 the conclusion holds
for at least one K ∈ K. By continuity of
∂σ
imp
∂K
on K, the conclusion will hold for all
K ∈ K, if we can show that ∂σ
imp
/∂K = 0 on K, or equivalently (by Lemma A.1.3) that
65
F(σ
imp
(K), K) = 0 for all K ∈ K. The remaining lemmas complete the proof, by verifying
that F(σ
imp
(K), K) = 0.
For K > θ
+
deﬁne
∆(K) := d
2
2
(K) + 2 log(S
0
/K) (A.10)
h(K) := d
2
(K) + σ
√
T
K −θ
K
. (A.11)
Let
D := {K > θ
+
: ∆(K) ≥ 0},
and for all K ∈ D, deﬁne
y
±
(K) :=
−d
2
(K) ±
_
∆(K)
√
T
. (A.12)
Lemma A.1.6. For (y, K) ∈ R × K, we have F(y, K) = 0 if and only if K ∈ D and
y = y
±
(K).
Proof. If K / ∈ D then F(·, K) is a quadratic with no real roots. If K ∈ D then F(·, K) has
roots y = y
±
(K).
Lemma A.1.7. For all K ∈ D deﬁne
H
±
(K) := S
0
N(±
_
∆(K)) −(S
0
−θ)N(d
2
(K) + σ
√
T) −θN(d
2
(K))
Then for all K ∈ D ∩ K,
H
±
(K) = C
BS
(y
±
(K)) −C
BS
(σ
imp
(K))
where C
BS
(·) is shorthand for C
BS
(S
0
, K, ·).
Hence y
+
(K) ≷ σ
imp
(K) if H
+
(K) ≷ 0. Likewise y
−
(K) ≷ σ
imp
(K) if H
−
(K) ≷ 0.
66
Proof. Using log(S
0
/K)/(y
±
(K)
√
T) −y
±
(K)
√
T/2 = d
2
(K) and (2.9), we have
C
BS
(y
±
(K)) −C
BS
(σ
imp
) =
_
S
0
N
_
log(S
0
/K)
y
±
(K)
√
T
+
y
±
(K)
√
T
2
_
−KN(d
2
(K))
_
−
_
(S
0
−θ)N(d
2
(K) + σ
√
T) −(K −θ)N(d
2
(K))
_
= S
0
N(d
2
(K) + y
±
(K)
√
T) −(S
0
−θ)N(d
2
(K) + σ
√
T) −θN(d
2
(K))
= H
±
(K)
The remaining conclusion is by monotonicity of C
BS
.
Lemma A.1.8. If θ ≷ 0 then for all K ∈ D we have
_
∆(K) ≷ h(K).
Proof. Consider only the case θ > 0. The proof for θ < 0 is similar.
We need to show that for all K ∈ D,
d
2
2
(K) + 2 log(S
0
/K) > d
2
2
(K) +
_
K −θ
K
_
2
σ
2
T + 2d
2
(K)σ
√
T
K −θ
K
or equivalently that
__
K −θ
K
_
2
−
_
K −θ
K
__
σ
2
T +2
K −θ
K
log[(S
0
−θ)/(K −θ)] −2 log(S
0
/K) < 0. (A.13)
The ﬁrst term is negative, because 0 < (K − θ)/K < 1; so it suﬃces to show that for all
K > θ,
L(K) :=
K −θ
K
log[(S
0
−θ)/(K −θ)] −log(S
0
/K) ≤ 0.
This is veriﬁed by L(S
0
) = 0 and ∂L/∂K = (θ/K
2
) log[(S
0
−θ)/(K−θ)] ≶ 0 for K ≷ S
0
.
Lemma A.1.9. For all K > θ
+
:
• If θ > 0 then K ∈ D and
∂H
−
∂K
> 0,
∂H
+
∂K
> 0.
• If θ < 0, h(K) > 0 then
∂∆
∂K
(K) < 0. If moreover K ∈ D then
∂H
−
∂K
(K) > 0,
∂H
+
∂K
(K) <
0.
• If θ < 0, h(K) < 0 then
∂∆
∂K
(K) > 0. If moreover K ∈ D then
∂H
−
∂K
(K) < 0,
∂H
+
∂K
(K) >
0.
67
Proof. The ∆ conclusions hold because
∂∆
∂K
(K) = −
2h(K)
(K −θ)σ
√
T
.
If θ > 0, the K ∈ D conclusion clearly holds for K ∈ (S
0
, ∞), and also holds for K ∈ (θ, S
0
]
because
log[(S
0
−θ)/(K −θ)]/(σ
√
T) −σ
√
T/2 < log(S
0
/K)/(σ
√
T) −σ
√
T/2 < 0
implies
∆(K) > (log(S
0
/K)/(σ
√
T)−σ
√
T/2)
2
+2 log(S
0
/K) = (log(S
0
/K)/(σ
√
T)+σ
√
T/2)
2
≥ 0.
Lastly, in all cases, the H
±
conclusions hold because of
∂H
±
∂K
(K) =
1
√
2π
Ke
−d
2
2
(K)/2
(K −θ)σ
√
T
_
∆(K)
_
_
∆(K) ∓h(K)
_
and Lemma A.1.8.
Lemma A.1.10. If θ > 0 then for all K ∈ K = (θ, ∞), we have K ∈ D and
y
1
(K) < σ
imp
(K) < y
2
(K).
Proof. By Lemma A.1.9 we have K ∈ D and y
±
(K) are welldeﬁned.
To prove σ
imp
< y
+
, note that as K ↓ θ we have d
2
(K) → ∞ hence H
+
(K) → 0.
Moreover, ∂H
+
/∂K > 0 on K, by Lemma A.1.9. So on K we have H
+
> 0, hence σ
imp
< y
+
by Lemma A.1.7.
To prove y
−
< σ
imp
, note that as K → ∞ we have d
2
(K) → −∞ hence H
−
(K) → 0.
Moreover, ∂H
−
/∂K > 0 on K, by Lemma A.1.9. So on K we have H
−
< 0, hence y
−
< σ
imp
by Lemma A.1.7.
Lemma A.1.11. If θ < 0 then for each K ∈ K∩ D we have
σ
imp
(K) / ∈ [y
−
(K), y
+
(K)].
68
Proof. Because θ < 0, it is clear that h is decreasing on (0, ∞).
Because K ∈ D, we have h(K) = 0 by Lemma A.1.8.
If h(K) < 0, then for all k > K we have h(k) < 0 and ∂∆/∂K(k) > 0 by Lemma
A.1.9. So for all k > K we have k ∈ D and
∂H
−
∂K
(k) < 0 by Lemma A.1.9. Moreover
lim
k→∞
H
−
(k) = 0. Therefore H
−
(K) > 0, hence σ
imp
(K) < y
−
(K) by Lemma A.1.7.
If h(K) > 0, then for all k ∈ (0, K) we have h(k) > 0 and ∂∆/∂K(k) < 0 by Lemma
A.1.9. So for all k ∈ (0, K) we have k ∈ D and
∂H
+
∂K
(k) < 0 by Lemma A.1.9. Moreover, as
k ↓ 0, we have ∆(k) → ∞, hence
H
+
(k) → S
0
−(S
0
−θ)N
_
log[(S
0
−θ)/(−θ)]
σ
√
T
+
σ
√
T
2
_
−θN
_
log[(S
0
−θ)/(−θ)]
σ
√
T
−
σ
√
T
2
_
= S
0
−C
BS
(S
0
−θ, −θ, σ) < S
0
−(S
0
−θ + θ) = 0.
Therefore H
+
(K) < 0, hence y
+
(K) < σ
imp
(K) by Lemma A.1.7.
Lemma A.1.12. For all K ∈ K we have F(σ
imp
(K), K) = 0.
Proof. Combine Lemmas A.1.10, A.1.11, and A.1.6.
A.2 Appendix: Proof of Theorem 2–Atthemoney Behavior of
Displaced Lognormal
Proof of Theorem 2.1. Taking K = S
0
in (2.9), we have
C
BS
(S
0
, S
0
, σ
atm
) = C
BS
(S
0
−θ, S
0
−θ, σ), (A.14)
hence
S
0
[2N(σ
atm
√
T/2) −1] = (S
0
−θ)[2N(σ
√
T/2) −1], (A.15)
and
N(σ
atm
√
T/2) =
S
0
−θ
2S
0
[2N(σ
√
T/2) −1] +
1
2
= (1 −
θ
S
0
)N(σ
√
T/2) +
θ
S
0
N(0) ≤ N((1 −
θ
S
0
)d
1
)
(A.16)
for θ ≥ 0, because N(x) is concave on x ≥ 0. Monotonicity of N implies σ
atm
≤ (1−θ/S
0
)σ.
Similarly, for θ ≤ 0, we have σ
atm
≥ (1 −θ/S
0
)σ.
69
Proof of Theorem 2.2. Using (A.8), we have
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
=
¸
¸
¸
¸
S
0
σ
imp
∂σ
imp
∂K
¸
¸
¸
¸
K=S
0
=
N(σ
atm
√
T/2) −N(σ
√
T/2)
φ(σ
atm
√
T/2)σ
atm
√
T
, (A.17)
because θ < 0 implies σ
atm
> σ. By concavity of N on [0, ∞),
φ(σ
atm
√
T/2) ≤
N(σ
atm
√
T/2) −N(σ
√
T/2)
σ
atm
√
T/2 −σ
√
T/2
≤ φ(σ
√
T/2). (A.18)
Combining (A.17), (A.18), and Theorem 2.1 produces the lower bound
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
≥
σ
atm
−σ
2σ
atm
≥
1
2
_
1 −
1
1 −θ/S
0
_
=
θ
2(S
0
+θ)
. (A.19)
Combining (A.17) and (A.18) produces the upper bound
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
≤
σ
atm
−σ
2σ
atm
×
φ(σ
√
T/2)
φ(σ
atm
√
T/2)
≤
1
2
e
σ
2
atm
T/8
. (A.20)
as claimed.
A.3 Appendix: Proof of Theorems 3 and 6–Shortexpiry
Behavior of DL
The notation A(T) ∼ B(T) means that A(T)/B(T) → 1 as T ↓ 0.
Proof of Theorem 3. For all K > θ
+
, we have C
BS
(S
0
− θ, K − θ, σ, T) ↓ (S
0
− K)
+
as
T ↓ 0, so for all T suﬃciently small, C
BS
(S
0
− θ, K − θ, σ, T) < S
0
hence K ∈ K
S
(T).
Applying RoperRutkowski (2007) Proposition 5.1 to the displaced lognormal model, we
have, as T ↓ 0,
σ
imp
(K, T) ∼
 log(S
0
/K)
_
−2T log(C
BS
(S
0
−θ, K −θ, σ, T) −(S
0
−K)
+
)
. (A.21)
On the other hand, applying it to the BlackScholes model, we have, as T ↓ 0,
σ ∼
 log((S
0
−θ)/(K −θ))
_
−2T log(C
BS
(S
0
−θ, K −θ, σ, T) −(S
0
−K)
+
)
. (A.22)
70
Therefore
σ
imp
(K, T) ∼
σ log(S
0
/K)
 log((S
0
−θ)/(K −θ))
=
σ log(S
0
/K)
log((S
0
−θ)/(K −θ))
(A.23)
as T ↓ 0.
Proof of Theorem 6. For all K ∈ (0, θ), we have C
BS
(θ −K, θ −S
0
, −σ, T) ↓ (S
0
−K)
+
as
T ↓ 0, so for all T suﬃciently small, C
BS
(θ − K, θ − S
0
, −σ, T) < S
0
hence K ∈ K
S
(T).
Applying RoperRutkowski (2007) Proposition 5.1 to the displaced antilognormal model,
we have, as T ↓ 0,
σ
imp
(K, T) ∼
 log(S
0
/K)
_
−2T log(C
BS
(θ −K, θ −S
0
, −σ, T) −(S
0
−K)
+
)
. (A.24)
On the other hand, applying it to the BlackScholes model, we have, as T ↓ 0,
−σ ∼
 log((θ −K)/(θ −S
0
))
_
−2T log(C
BS
(θ −K, θ −S
0
, −σ, T) −(S
0
−K)
+
)
. (A.25)
Therefore (A.23) holds.
A.4 Appendix: Proof of Theorem 4–Global Behavior of
Displaced anti Lognormal
Deﬁne for all K < θ
d
2
(K) :=
log[(θ −S
0
)/(θ −K)]
σ
√
T
−
σ
√
T
2
.
Lemma A.4.1. For all K ∈ K
ADL
,
sgn
∂σ
imp
∂K
(K) = sgn F(σ
imp
(K), K)
where F : R ×(0, θ) →R is deﬁned by
F(y, K) := −y
2
T/2 + y
√
Td
2
(K) + log(S
0
/K).
71
Proof. Taking the K derivative of (2.19),
∂C
BS
∂K
(S
0
, K, σ
imp
(K))+
∂C
BS
∂σ
(S
0
, K, σ
imp
(K))
∂σ
imp
∂K
(K) = −1−
∂C
BS
∂K
(θ−S
0
, θ−K, σ).
(A.26)
Therefore
sgn
∂σ
imp
∂K
(K) = sgn
_
−1 −
∂C
BS
∂K
(θ −S
0
, θ −K, σ) −
∂C
BS
∂K
(S
0
, K, σ
imp
(K))
_
= sgn
_
−1 + N(d
2
(K)) + N
_
log(S
0
/K)
σ
imp
(K)
√
T
−
σ
imp
(K)
√
T
2
__
= sgn
__
log(S
0
/K)
σ
imp
(K)
√
T
−
σ
imp
(K)
√
T
2
_
+ d
2
(K)
_
= sgn
_
−σ
2
imp
(K)T/2 + σ
imp
(K)
√
Td
2
(K) + log(S
0
/K)
¸
= sgn F(σ
imp
(K), K),
as claimed.
Lemma A.4.2. θ > S
0
> 0, there exists K ∈ K
ADL
such that
sgn
∂σ
imp
∂K
(K) = −sgn θ.
Proof. The atthemoney strike K = S
0
satisﬁes the conclusion, because
sgn
∂σ
imp
∂K
(S
0
) = sgn
_
−σ
imp
(S
0
)
√
T/2 −σ
√
T/2
¸
< 0.
Proposition A.4.3. For all T > 0, K ∈ K
ADL
(T) and θ > S
0
,
sgn
∂σ
imp
∂K
(K) = −sgn θ. (A.27)
Proof. By Lemma A.4.2 the conclusion holds for at least one K ∈ K
ADL
. By continuity of
∂σ
imp
∂K
on K
ADL
, the conclusion will hold for all K ∈ K
ADL
, if we can show that ∂σ
imp
/∂K =
0 on K
ADL
, or equivalently (by Lemma A.4.1) that F(σ
imp
(K), K) = 0 for all K ∈ K. The
remaining lemmas complete the proof, by verifying that F(σ
imp
(K), K) = 0.
72
For K < θ deﬁne
∆(K) := d
2
2
(K) + 2 log(S
0
/K) (A.28)
h(K) := d
2
(K) −σ
√
T
θ −K
K
. (A.29)
Let
D := {0 < K < θ : ∆(K) ≥ 0},
and for all K ∈ D, deﬁne
y
±
(K) :=
d
2
(K) ±
_
∆(K)
√
T
. (A.30)
Lemma A.4.4. For (y, K) ∈ R × K
ADL
, we have F(y, K) = 0 if and only if K ∈ D and
y = y
±
(K).
Proof. If K / ∈ D then F(·, K) is a quadratic with no real roots. If K ∈ D then F(·, K) has
roots y = y
±
(K).
Lemma A.4.5. For all K ∈ D deﬁne
H
±
(K) := −S
0
N(∓
_
∆(K)) −(θ −S
0
)N(d
2
(K) +σ
√
T) + θN(d
2
(K)) (A.31)
Then for all K ∈ D ∩ K
ADL
,
H
±
(K) = P
BS
(y
±
(K)) −P
BS
(σ
imp
(K))
where P
BS
(·) is shorthand for P
BS
(S
0
, K, ·).
Hence y
+
(K) ≷ σ
imp
(K) if H
+
(K) ≷ 0. Likewise y
−
(K) ≷ σ
imp
(K) if H
−
(K) ≷ 0.
Proof. Using log(S
0
/K)/(y
±
(K)
√
T) −y
±
(K)
√
T/2 = −d
2
(K) and (2.18), we have
P
BS
(y
±
(K)) −P
BS
(σ
imp
) =
_
KN
_
−
log(S
0
/K)
y
±
(K)
√
T
+
y
±
(K)
√
T
2
_
−S
0
N
_
−
log(S
0
/K)
y
±
(K)
√
T
−
y
±
(K)
√
T
2
__
−
_
(θ −S
0
)N(d
2
(K) +σ
√
T) −(θ −K)N(d
2
(K))
_
= KN(d
2
(K)) −S
0
N(d
2
(K) −y
±
(K)
√
T) −(θ −S
0
)N(d
2
(K) +σ
√
T)
+ (θ −K)N(d
2
(K))
= H
±
(K)
73
The remaining conclusion is by monotonicity of P
BS
.
Lemma A.4.6. θ > S
0
, for all K ∈ D we have
_
∆(K) < h(K).
Proof. We want to show:
h
2
(K)−∆(K) =
__
θ −K
K
_
2
+
_
θ −K
K
__
σ
2
T−2
θ −K
K
log[(θ−S
0
)/(θ−K)]−2 log(S
0
/K) > 0.
(A.32)
The ﬁrst term is positive, because (θ −K)/K > 0; so it suﬃces to show that for all K > θ,
L(K) := −
θ −K
K
log[(θ −S
0
)/(θ −K)] −log(S
0
/K) ≥ 0.
This is veriﬁed by L(S
0
) = 0 and ∂L/∂K = (θ/K
2
) log[(θ−S
0
)/(θ−K)] ≶ 0 for K ≶ S
0
.
Lemma A.4.7. For all 0 < K < θ:
• If h(K) > 0 then
∂∆
∂K
(K) > 0. If moreover K ∈ D then
∂H
−
∂K
(K) < 0,
∂H
+
∂K
(K) > 0.
• If h(K) < 0 then
∂∆
∂K
(K) < 0. If moreover K ∈ D then
∂H
−
∂K
(K) > 0,
∂H
+
∂K
(K) < 0.
Proof. The ∆ conclusion hold because:
∂∆(K)
∂K
=
2h(K)
(θ −K)σ
√
T
. (A.33)
The H
±
conclusions hold because of
∂H
±
∂K
(K) =
1
√
2π
Ke
−d
2
2
(K)/2
(θ −K)σ
√
T
_
∆(K)
_
_
∆(K) ±h(K)
_
.
Lemma A.4.8. If θ > 0 then for each K ∈ K
ADL
∩ D we have
σ
imp
(K) / ∈ [y
−
(K), y
+
(K)].
Proof. Because θ > 0 and
∂h(K)
∂K
=
1
θ−K
1
σ
√
T
+
σ
√
Tθ
K
2
> 0, h(K) is monotonically in
creasing on (0, θ). Furthermore h(0) = −∞, h(θ) = ∞, so there is a point k
0
∈ (0, θ) such
that h(k
0
) = 0. By (A.33), ∆(K) is decreasing on {K : 0 < K < k
0
} and increasing on
{K : k
0
< K < θ}.
74
Because K ∈ D, we have h(K) = 0 by Lemma A.4.6.
For a particular K ∈ D, if h(K) < 0, then for all k < K, we have h(k) < 0 and
∂∆/∂K(k) < 0 by Lemma A.4.7. So for all k < K and k ∈ D, we have
∂H
+
∂K
(k) < 0 by
Lemma A.4.7. Moreover, k ↓ 0, ∆(k) → ∞, hence by (A.31)
lim
k→0
H
+
(k) = 0 −(θ −S
0
)N(d
2
(0) +σ
√
T) + θN(d
2
(0)) = 0 −C
BS
(θ −S
0
, θ, σ, T) < 0.
Therefore H
+
(K) < 0 hence y
+
(K) < σ
imp
(K) when K ∈ K
ADL
by Lemma A.4.5.
If h(K) > 0, then for all k ∈ (K, θ), we have h(k) > 0 and ∂∆/∂K(k) > 0 by Lemma
A.4.7. So for all k > K and k ∈ D, we have
∂H
−
∂K
(k) < 0 by Lemma A.4.7.
k = θ ⇒ d
2
(k) = ∞ ⇒ ∆(k) = ∞,
⇒ H
−
(θ) = −S
0
N(−
_
∆(k))−(θ−S
0
)N(d
2
(k)+σ
√
T)+θN(d
2
(k)) = −S
0
−(θ−S
0
)+θ = 0.
Therefore H
−
(K) > 0 hence y
−
(K) > σ
imp
(K) when K ∈ K
ADL
by Lemma A.4.5.
Lemma A.4.9. For all K ∈ K
ADL
we have F(σ
imp
(K), K) = 0.
Proof. Combine Lemmas A.4.4, and A.4.8.
A.5 Appendix: Proof of Theorem 5–Atthemoney Behavior of
Displaced anti Lognormal
Proof of Theorem 5.1. Consider atthemoney in (2.19), we have,
S
0
[2N(σ
atm
√
T/2) −1] = (θ −S
0
)[2N(σ
√
T/2) −1], (A.34)
and
N(σ
atm
√
T/2) =
θ −S
0
2S
0
[2N(σ
√
T/2) −1] +
1
2
= (
θ
S
0
−1)N(σ
√
T/2) + (2 −
θ
S
0
)N(0).
(A.35)
75
If
θ
S
0
−1 ∈ [0, 1], then
(
θ
S
0
−1)N(σ
√
T/2) + (2 −
θ
S
0
)N(0) ≤ N
_
(
θ
S
0
−1)
σ
√
T
2
_
. (A.36)
Because N(x) is concave on x ≥ 0. Monotonicity of N implies σ
atm
≤ (θ/S
0
−1)σ.
Similarly, if
θ
S
0
−1 ∈ [1, ∞), then
(
θ
S
0
−1)N(σ
√
T/2) + (2 −
θ
S
0
)N(0) ≥ N
_
(
θ
S
0
−1)
σ
√
T
2
_
. (A.37)
Using the monotonicity of N, we have σ
atm
≥ (θ/S
0
−1)σ.
Proof of Theorem 5.2. Reorder the entries in (A.26) and plug in (A.34), we have,
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
= =
¸
¸
¸
¸
−1 + N(−σ
√
T/2) + N(−σ
atm
√
T/2)
φ(σ
atm
√
T/2)σ
atm
√
T
¸
¸
¸
¸
K=S
0
(A.38)
=
¸
¸
¸
¸
N(σ
√
T/2) + N(σ
atm
√
T/2) −1
φ(σ
atm
√
T/2)σ
atm
√
T
¸
¸
¸
¸
K=S
0
(A.39)
=
¸
¸
¸
¸
θ
θ−S
0
_
N(σ
atm
√
T/2) −1/2
_
φ(σ
atm
√
T/2)σ
atm
√
T
¸
¸
¸
¸
K=S
0
. (A.40)
By the concavity of N on [0, ∞):
φ(σ
atm
√
T/2) ≤
N(σ
atm
√
T/2) −N(0)
σ
atm
√
T/2
≤ φ(0). (A.41)
Combine (A.40) and (A.41), we have
θ
2(θ −S
0
)
≤
¸
¸
¸
¸
∂ log σ
imp
∂ log K
¸
¸
¸
¸
K=S
0
≤
θ
2(θ −S
0
)
exp(σ
2
atm
T/8) (A.42)
76
APPENDIX B
B.1 Appendix: Proof of Theorem 7–Atthemoney Behavior of
Displaced Independent Stochastic Volatility
Proof of Theorem 7. Deﬁne the average volatility of the DISV process:
Σ =
¸
_
T
0
σ
2
t
dt
T
. (B.1)
Since conditioning on the average volatility Σ, the process S
T
−θ is lognormal process,
E(S
T
−K)
+
= E{E[(S
T
−θ) −(K −θ)]
+
Σ} = E(C
BS
(S
0
−θ, K −θ, Σ)).
For K ∈ K
DISV
, the implied volatility σ
imp
satisﬁes
C
BS
(S
0
, K, σ
imp
) = E(C
BS
(S
0
−θ, K −θ, Σ)). (B.2)
Diﬀerentiating with respect to K in (B.2) and exchanging the expectation and the diﬀeren
tiation (validate later), we have
∂C
BS
(S
0
, K, σ
imp
)
∂σ
imp
∂σ
imp
∂K
= E
_
∂C
BS
(S
0
−θ, K −θ, Σ)
∂K
−
∂C
BS
(S
0
, K, σ
imp
)
∂K
_
. (B.3)
Since
∂C
BS
(S
0
,K,σ
imp
)
∂σ
imp
> 0, we have
sgn
∂σ
imp
∂K
¸
¸
¸
¸
K=S
0
= sgn E
_
∂C
BS
(S
0
−θ, K −θ, Σ)
∂K
−
∂C
BS
(S
0
, K, σ
atm
)
∂K
_
K=S
0
(B.4)
= sgn E
_
N(−
σ
atm
√
T
2
) −N(−
Σ
√
T
2
)
_
(B.5)
= sgn
_
EN(
Σ
√
T
2
) −N(
σ
atm
√
T
2
)
_
. (B.6)
77
What remains is to show θ ≷ 0 ⇒ N(
σ
atm
√
T
2
) ≶ EN(
Σ
√
T
2
). Assuming the exchange
ability of the expectation and the diﬀerentiation, we have
∂E(C
BS
(S
0
−θ, K −θ, Σ))
∂θ
= E
∂C
BS
(S
0
−θ, K −θ, Σ)
∂θ
= E(−N(d
1
) + N(d
2
)) ≤ 0,
where d
1,2
=
log((S
0
−θ)/(K−θ))±Σ
2
T/2
Σ
√
T
. So E(C
BS
(S
0
− θ, K − θ, Σ)) is monotonically de
creasing as a function of θ. Thus,
θ ≷ 0 ⇒E(C
BS
(S
0
−θ, K −θ, Σ)) ≶ E(C
BS
(S
0
, K, Σ)),
⇒ C
BS
(S
0
, K, σ
imp
) ≶ E(C
BS
(S
0
, K, Σ)).
Particularly when atthemoney, we have
N(
σ
atm
√
T
2
) ≶ EN(
Σ
√
T
2
). (B.7)
To complete the proof, we need to validate the exchangeability of the expectation and
the diﬀerentiation. Since
¸
¸
¸
¸
∂C
BS
(S
0
−θ, K −θ, Σ)
∂θ
¸
¸
¸
¸
=  −N(d
1
) + N(d
2
) ≤ 2,
¸
¸
¸
¸
∂C
BS
(S
0
−θ, K −θ, Σ)
∂K
¸
¸
¸
¸
=  −N(d
2
) ≤ 1,
we can use dominant convergent theorem to exchange the expectation and the diﬀerentiation.
B.2 Appendix: Proof of Theorems 8 and 9– Shortexpiry
Behavior of DH
Proof of Theorem 8. Denote C
DH
(S
0
, K, T) the call option price under the DH process,
with strike K, maturity T and initial value S
0
; and similarly C
H
(S
0
, K, T) the call option
price under the Heston process. We have C
DH
(S
0
, K, T) = C
H
(S
0
−θ, K−θ, T). Applying
78
RoperRutkowski (2007) corollary 5.1 to the displaced Heston model, we have, as T ↓ 0,
σ
DH
imp
(K, T) ∼
_
¸
¸
_
¸
¸
_
 log(S
0
/K)
_
−2T log(C
DH
(S
0
,K,T)−(S
0
−K)
+
)
if K = S
0
_
2πC
DH
(S
0
, K, T)/(S
0
√
T) if K = S
0
(B.8)
On the other hand, applying it to the Heston model, we have, as T ↓ 0,
σ
H
imp
(K −θ, T) ∼
_
¸
¸
_
¸
¸
_
 log((S
0
−θ)/(K−θ))
_
−2T log(C
H
(S
0
−θ,K−θ,T)−(S
0
−K)
+
)
if K = S
0
_
2πC
H
(S
0
−θ, K −θ, T)/((S
0
−θ)
√
T) if K = S
0
(B.9)
By comparing (B.8) and (B.9), and using (3.12), we have
lim
T→0
σ
DH
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
σ
H
imp
(K −θ, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
σ
H
imp
(K −θ, T) ×(1 −θ/S
0
) if K = S
0
.
(B.10)
Proof of Theorem 9. Denote C
ADH
(S
0
, K, T) the call option price under the displaced anti 
process, with strike K, maturity T and initial value S
0
; and similarly C
H
(S
0
, K, T) the call
option price under the Heston process. Let
¯
S = θ − S,
¯
K = θ − K. Using putcall parity,
we have
E(S
T
−K)
+
−(S
0
−K)
+
= E(
¯
K−
¯
S
T
)
+
−(
¯
K−
¯
S
0
)
+
= E(
¯
S
T
−
¯
K)
+
−(
¯
S
0
−
¯
K)
+
. (B.11)
If atthemoney, then,
E(S
T
−K)
+
= E(
¯
S
T
−
¯
K)
+
. (B.12)
Applying RoperRutkowski (2007) corollary 5.1 to the displaced antiHeston model, we
have, as T ↓ 0,
σ
ADH
imp
(K, T) ∼
_
¸
¸
_
¸
¸
_
 log(S
0
/K)
_
−2T log(C
ADH
(S
0
,K,T)−(S
0
−K)
+
)
if K = S
0
_
2πC
ADH
(S
0
, K, T)/(S
0
√
T) if K = S
0
(B.13)
79
Applying it to the Heston model, we have, as T ↓ 0,
σ
H
imp
(θ −K, T) ∼
_
¸
¸
_
¸
¸
_
 log((S
0
−θ)/(K−θ))
_
−2T log(C
H
(θ−S
0
,θ−K,T)−((θ−S
0
)−(θ−K))
+
)
if K = S
0
_
2πC
H
(θ −S
0
, θ −K, T)/((θ −S
0
)
√
T) if K = S
0
(B.14)
By comparing (B.13) and (B.14), and using (B.11) and (B.12), we have
lim
T→0
σ
ADH
imp
(K, T) =
_
¸
_
¸
_
lim
T→0
−σ
H
imp
(θ −K, T) ×
log(S
0
/K)
log((S
0
−θ)/(K −θ))
if K = S
0
lim
T→0
−σ
H
imp
(θ −K, T) ×(1 −θ/S
0
) if K = S
0
.
(B.15)
B.3 Appendix: Proof of Propositions 3.1.2 and 3.2.1–Level,
Slope and Convexity of DH Shortexpiry Implied Volatility
Proof of Proposition 3.1.2. We utilize the shortexpiry approximation formula of the implied
volatility given by Durrleman (2004) and the relationship between the implied volatilities of
the DH model and the Heston model (Theorems 8 and 9).
Denote
˜
S
t
= S
t
− θ,
˜
K = K − θ. Durrleman (2004) shows that, the nearmoney short
expiry implied volatility with strike
˜
K, maturity T and initial value
˜
S
0
is
σ
H
imp
(
˜
K, T) =
_
σ
2
0
+ a
0
log(
˜
S
0
/
˜
K) + b
0
T
2
+
c
0
2
log
2
(
˜
S
0
/
˜
K) + O(log(
˜
S
0
/
˜
K)T + T
2
).
(B.16)
It can be approximated as
σ
H
imp
(
˜
K, T) ∼
_
σ
2
0
+ a
0
log(
˜
S
0
/
˜
K) + b
0
T
2
+
c
0
2
log
2
(
˜
S
0
/
˜
K), (B.17)
where A(T) ∼ B(T) means A(T)/B(T) → 1 as T ↓ 0 and
a
0
= −
ρ
2
, c
0
=
2
6σ
2
0
(1 −
7ρ
2
4
), b
0
= κ(µ −σ
2
0
) +
ρ
2
σ
2
0
−
2
6
(1 −ρ
2
/4). (B.18)
Since we have speciﬁed that ρ = 0 for the Heston process, we have a
0
= 0. In the following,
we give a limit approximation of the level, slope and convexity as T ↓ 0. Though we do not
80
explicitly write the lim
T→0
out.
i. Intercept
Using (3.13)(3.14), the nearexpiry atthemoney implied volatility of the displaced
Heston is:
(σ
DH
imp
(S))
2
= (σ
H
imp
(S
0
−θ)(1−θ/S
0
))
2
= (σ
2
0
+κ(µ−σ
2
0
)
T
2
−
2
T
12
)×(1−θ/S
0
)
2
+O(T
2
).
(B.19)
ii. Slope
∂σ
DH
imp
(K)
∂K

K=S
0
= lim
K→S
0
σ
DH
imp
(K) −σ
DH
imp
(S
0
)
K −S
0
= lim
K→S
0
σ
H
imp
(K −θ)
log(S
0
/K)
log((S
0
−θ)/(K−θ))
−σ
H
imp
(S
0
−θ)(1 −θ/S
0
)
K −S
0
= lim
K→S
0
∂σ
H
imp
(K −θ)
∂K
log(S
0
/K)
log((S
0
−θ)/(K −θ))
+
σ
H
imp
(K −θ)
−
1
K
log((S
0
−θ)/(K −θ)) +
1
K−θ
log(S
0
/K)
(log((S
0
−θ)/(K −θ)))
2
(L’Hopital’s rule)
= lim
K→S
0
∂σ
H
imp
(K −θ)
∂K
(1 −θ/S
0
) + σ
H
imp
(S
0
−θ)
θ
2S
2
0
where in the last step we use
lim
K→S
0
log(S
0
/K)
log((S
0
−θ)/(K −θ))
= 1 −θ/S
0
and
lim
K→S
0
−
1
K
log((S
0
−θ)/(K −θ)) +
1
K−θ
log(S
0
/K)
(log((S
0
−θ)/(K −θ)))
2
=
θ
2S
2
0
.
Applying (B.17) for the Heston model, we have,
∂σ
H
imp
(
˜
K)
∂
˜
K
¸
¸
¸
¸
˜
K=
˜
S
0
=
1
2
(σ
H
imp
(
˜
K))
−1
_
−a
0
1
˜
K
−c
0
log(
˜
S
0
/
˜
K)
1
˜
K
_¸
¸
¸
¸
˜
K=
˜
S
0
= 0 + O(T).
(B.20)
81
So
∂σ
DH
imp
(K)
∂K

K=S
0
= σ
H
imp
(S
0
−θ)
θ
2S
2
0
+ O(T). (B.21)
Consequently the slope is
∂ log σ
DH
imp
(K)
∂ log K
¸
¸
¸
¸
K=S
0
=
θ
2(S
0
−θ)
+ O(T). (B.22)
iii. Convexity
For the second derivative, using (3.14), we have
∂
2
σ
DH
imp
(K)
∂K
2
=
∂
2
σ
H
imp
(K −θ)
∂K
2
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
+ 2
∂σ
H
imp
(K −θ)
∂K
∂
∂K
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
+ σ
H
imp
(K −θ)
∂
2
∂K
2
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
.
Using
∂σ
H
imp
(K−θ)
∂K
¸
¸
¸
¸
K=S
0
= O(T) in (B.20), we only need to calculate the ﬁrst and
third term. The ﬁrst term is straightforward to calculate, and the third term requires
more attention.
(a) The ﬁrst term
Combining
lim
K→S
0
∂
2
σ
H
imp
(
˜
K)
∂
˜
K
2
=
1
2
_
σ
H
imp
(S
0
−θ)
_
−1
c
0
(S
0
−θ)
2
and lim
K→S
0
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
=
S
0
−θ
S
0
,
we have
lim
K→S
0
∂
2
σ
H
imp
(K −θ)
∂K
2
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
=
1
2
_
σ
H
imp
(S
0
−θ)
_
−1
c
0
(S
0
−θ)S
0
. (B.23)
(b) The third term
82
With some calculation, we have
H(S
0
, θ) := lim
K→S
0
∂
2
∂K
2
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
=
4θ
2
−5θS
0
6S
3
0
(S
0
−θ)
.
⇒ lim
K→S
0
σ
H
imp
(K −θ)
∂
2
∂K
2
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
= σ
H
imp
(S
0
−θ)H(S
0
, θ). (B.24)
Combining (B.23) and (B.24) gives the second derivative of the implied volatility
∂
2
σ
DH
imp
(K)
∂K
2
¸
¸
¸
¸
K=S
0
=
1
2
_
σ
H
imp
(S
0
−θ)
_
−1
c
0
(S
0
−θ)S
0
+σ
H
imp
(S
0
−θ)H(S
0
, θ)+O(T).
(B.25)
Proof of Proposition 3.2.1. We will use the relationship between the implied volatilities of
the displaced anti Heston model and the Heston model. Other parts of the proof are similar
to those in Proposition 3.1.2. Denote
¯
K = θ −K and
¯
S = θ −S
0
.
i. Intercept:
The shortexpiry atthemoney implied volatility is easy to get:
(σ
ADH
imp
(S
0
))
2
= (σ
H
imp
(θ−S
0
)(1−θ/S
0
))
2
= (σ
2
0
+κ(µ−σ
2
0
)
T
2
−
2
T
12
)×(1−θ/S
0
)
2
+O(T
2
).
(B.26)
ii. Slope
The ﬁrst derivative
∂σ
DH
imp
(K)
∂K
is calculated as in the case of displaced Heston, which is
∂σ
ADH
imp
(K)
∂K
¸
¸
¸
¸
K=S
0
= −σ
H
imp
(θ −S
0
)
θ
2S
2
0
+ O(T). (B.27)
Consequently, we have
∂ log σ
ADH
imp
(K)
∂ log K
¸
¸
¸
¸
K=S
0
=
θ
2(S
0
−θ)
+ O(T). (B.28)
iii. Convexity
83
Use (3.14), we have
∂
2
σ
ADH
imp
(K)
∂K
2
= −
_
∂
2
σ
H
imp
(θ −K)
∂K
2
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
+ 2
∂σ
H
imp
(θ −K)
∂K
∂
∂K
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
+ σ
H
imp
(θ −K)
∂
2
∂K
2
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
__
.
Again since
∂σ
H
imp
(θ−K)
∂K
¸
¸
¸
¸
K=S
0
= O(T), we only need to calculate the ﬁrst and the
third term.
(a) The ﬁrst term
Combining
lim
K→S
0
∂
2
σ
H
imp
(
¯
K)
∂
¯
K
2
=
1
2
_
σ
H
imp
(θ−S
0
)
_
−1
c
0
(θ −S
0
)
2
and lim
K→S
0
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
=
S
0
−θ
S
0
,
we have
lim
K→S
0
∂
2
σ
H
imp
(θ −K)
∂K
2
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
=
1
2
_
σ
H
imp
(θ −S
0
)
_
−1
c
0
(S
0
−θ)S
0
. (B.29)
(b) The third term
This is exactly the same as in the displaced Heston case:
lim
K→S
0
σ
H
imp
(θ −K)
∂
2
∂K
2
_
log(
S
0
K
)
log(
S
0
−θ
K−θ
)
_
= σ
H
imp
(θ −S
0
)H(S
0
, θ). (B.30)
Combing (B.29) and (B.30) gives the convexity as
∂
2
σ
ADH
imp
(K)
∂K
2
¸
¸
¸
¸
K=S
0
= −
1
2
_
σ
H
imp
(θ−S
0
)
_
−1
c
0
(S
0
−θ)S
0
−σ
H
imp
(θ−S
0
)H(S
0
, θ)+O(T).
(B.31)
84
APPENDIX C
C.1 Appendix: Proof of Theorem 12–Partial Strong
Convergency of Stochastic Volatility Process
Proof. The proof here has the same ﬂavor as the proof of Theorem 1 in Zhang et al (2004),
except for two major diﬀerences:
1. In Zhang et al(2004) Theorem 1, their stopping time is τ = ρ ∧ θ where ρ = inf{t ≥
0 : S
∆t
t
/ ∈ Ω
1
}, θ = inf{t ≥ 0 : S
t
/ ∈ Ω
1
}. We add one more piece to the stopping
time τ, which is τ = ρ ∧ θ ∧ γ, where γ = inf{t ≥ 0 : σ
t
/ ∈ Ω
2
}.
2. We substitute the local Lipschitz condition (*) with their local Lipschitz condition
(15).
To make it complete, we sketch the outline of the proof below.
First, we can write a continuous version of the Euler approximation deﬁned in (6.6) as
follows.
S
∆t
t
= S
0
+
_
t
0
A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)du +
_
t
0
B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)dW
u
, (C.1)
where we introduce the piecewise constant process
ˆ
S
∆t
t
= S
∆t
t
n
, for t ∈ [t
n
, t
n+1
),
ˆ σ
∆t
t
= σ
∆t
t
n
, for t ∈ [t
n
, t
n+1
).
(C.2)
Note that S
∆t
t
and
ˆ
S
∆t
t
coincide at the discrete points t
n
= n∆t.
Let T
1
∈ [0, T] be an arbitrary time. For any t ∈ [0, τ ∧ T
1
], we have
E
_
sup
0≤t≤τ∧T
1
S
∆t
t
−S
t

2
_
≤ 2E sup
0≤t≤τ∧T
1
¸
¸
¸
¸
_
t
0
(A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −A(u, S
u
, σ
u
))du
¸
¸
¸
¸
2
+ 2E sup
0≤t≤τ∧T
1
¸
¸
¸
¸
_
t
0
(B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −B(u, S
u
, σ
u
))dW
u
¸
¸
¸
¸
2
.
(C.3)
85
Applying the Holder inequality to the ﬁrst term of (C.3) gives
2E sup
0≤t≤τ∧T
1
¸
¸
¸
¸
_
t
0
(A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −A(u, S
u
, σ
u
))du
¸
¸
¸
¸
2
≤ 2TE sup
0≤t≤τ∧T
1
_
τ∧T
1
0
¸
¸
¸
¸
(A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −A(u, S
u
, σ
u
))
¸
¸
¸
¸
2
du.
(C.4)
Applying the Doob inequality to the second term of (C.3) gives
2E sup
0≤t≤τ∧T
1
¸
¸
¸
¸
_
t
0
(B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −B(u, S
u
, σ
u
))du
¸
¸
¸
¸
2
≤ 8C
1
E sup
0≤t≤τ∧T
1
_
τ∧T
1
0
¸
¸
¸
¸
(B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −B(u, S
u
, σ
u
))
¸
¸
¸
¸
2
du.
(C.5)
If A(.) and B(.) satisfy the local Lipschitz condition (*), we have
(A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −A(u, S
u
, σ
u
))
2
∨ (B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
) −B(u, S
u
, σ
u
))
2
≤ K
1
(Ω
1
×Ω
2
)
ˆ
S
∆t
u
−S
u

2
+ K
2
(Ω
1
×Ω
2
)ˆ σ
∆t
u
−σ
u

2
.
(C.6)
Substituting (C.4)(C.6) into (C.3) and use condition (6.11), we have
E
_
sup
0≤t≤τ∧T
1
S
∆t
t
−S
t

2
_
≤ 2(T + 4C
1
)K
1
(Ω
1
×Ω
2
)E
_
τ∧T
1
0

ˆ
S
∆t
u
−S
u

2
du + K
4
(Ω
1
×Ω
2
)∆t
= 2(T + 4C
1
)K
1
(Ω
1
×Ω
2
)E
_
τ∧T
1
0

ˆ
S
∆t
u
−S
∆t
u
+ S
∆t
u
−S
u

2
du
+K
4
(Ω
1
×Ω
2
)∆t
≤ 4(T + 4C
1
)K
1
(Ω
1
×Ω
2
)E
_
τ∧T
1
0
(
ˆ
S
∆t
u
−S
∆t
u

2
+S
∆t
u
−S
u

2
)du
+K
4
(Ω
1
×Ω
2
)∆t
≤ 4(T + 4C
1
)K
1
(Ω
1
×Ω
2
)E
_
τ∧T
1
0

ˆ
S
∆t
u
−S
∆t
u

2
du + K
4
(Ω
1
×Ω
2
)∆t
+4(T + 4C
1
)K
1
(Ω
1
×Ω
2
)
_
T
1
0
E
_
sup
0≤u
≤τ∧u
S
∆t
u
−S
u

2
_
du,
(C.7)
where K
4
(Ω
1
×Ω
2
) = 2(T + 4C
1
)K
2
(Ω
1
×Ω
2
)D(Ω
2
)T.
86
If the sum of the ﬁrst term and the second term on the righthand side is bounded, then we
could apply the Gronwall inequality which leads to a bound on E
_
sup
0≤t≤τ∧T
1
S
∆t
t
−S
t

2
_
.
So all we need to prove now is the ﬁrst term on the righthand side is bounded. By the
deﬁnition of
ˆ
S
∆t
t
, we know that
ˆ
S
∆t
t
= S
∆t
[t/∆t]∆t
, where [t/∆t] is the integer part of t/∆t.
Use (C.1), we have
E
ˆ
S
∆t
t
−S
∆t
t

2
= ES
∆t
[t/∆t]∆t
−S
∆t
t

2
= E
_¸
¸
¸
¸
_
[t/∆t]∆t
0
A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)du −
_
t
0
A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)du
+
_
[t/∆t]∆t
0
B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)dW
u
−
_
t
0
B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)dW
u
¸
¸
¸
¸
2
_
≤ 2E
_¸
¸
¸
¸
_
t
[t/∆t]∆t
A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)du
¸
¸
¸
¸
2
+
¸
¸
¸
¸
_
t
[t/∆t]∆t
B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)dW
u
¸
¸
¸
¸
2
_
.
(C.8)
Applying (6.10) to the ﬁrst term on the righthand side and using the Holder inequality leads
to
¸
¸
¸
¸
_
t
[t/∆t]∆t
A(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)du
¸
¸
¸
¸
2
≤ K
3
(Ω
1
×Ω
2
)∆t
2
. (C.9)
Applying (6.10) to the second term on the righthand side and using the Ito isometry
(E(
_
t
0
f(X
u
)dW
u
)
2
=
_
t
0
Ef
2
(X
u
)du) leads to
E
¸
¸
¸
¸
_
t
[t/∆t]∆t
B(u,
ˆ
S
∆t
u
, ˆ σ
∆t
u
)dW
u
¸
¸
¸
¸
2
≤ K
3
(Ω
1
×Ω
2
)∆t. (C.10)
If T∆t < 1, (C.9) and (C.10) leads to
E
ˆ
S
∆t
t
−S
∆t
t

2
≤ 2K
3
(Ω
1
×Ω
2
)∆t
2
+ 2K
3
(Ω
1
×Ω
2
)∆t. (C.11)
Thus
E
_
t∧T
1
0

ˆ
S
∆t
u
−S
∆t
u

2
du ≤ 2T
1
K
3
(Ω
1
×Ω
2
)∆t
2
+ 2T
1
K
3
(Ω
1
×Ω
2
)∆t = C
0
(Ω
1
×Ω
2
)∆t,
(C.12)
87
where C
0
(Ω
1
×Ω
2
) = 2K
3
(Ω
1
×Ω
2
)(1 + T
1
)∆t. Plugging (C.12) into (C.7), we have
E
_
sup
0≤t≤τ∧T
1
S
∆t
t
−S
t

2
_
≤ C
1
(Ω
1
×Ω
2
)∆t+C
2
(Ω
1
×Ω
2
)E
_
T
1
O
E
_
sup
0≤u
≤τ∧u
S
∆t
u
−S
u

2
_
du
(C.13)
where C
1
(Ω
1
×Ω
2
) = C
0
(Ω
1
×Ω
2
)4(T +4C
1
)K
1
(Ω
1
×Ω
2
) +K
4
(Ω
1
×Ω
2
), C
2
(Ω
1
×Ω
2
) =
4(T + 4C
1
)K
1
(Ω
1
×Ω
2
). Now applying the Gronwall inequality to (C.13), we have
E
_
sup
0≤t≤τ∧T
S
∆t
t
−S
t

2
_
≤ C
1
(Ω
1
×Ω
2
)e
C
2
(Ω
1
×Ω
2
)T
∆t = C(Ω
1
×Ω
2
)∆t. (C.14)
Hence the claim is proved.
C.2 Appendix: Coeﬃcients of SABR Satisfy the Local Lipschtiz
Condition (*)
The σ
t
of the SABR process can be simulated exactly, so σ
∆t
t
n
= σ
t
n
. Using this, we have
¸
¸
¸
¸
σ
t
S
β
t
− ˆ σ
∆t
t
(
ˆ
S
∆t
t
)
β
¸
¸
¸
¸
2
=
¸
¸
¸
¸
σ
t
S
β
t
−σ
∆t
t
n
(S
∆t
t
n
)
β
¸
¸
¸
¸
2
=
¸
¸
¸
¸
σ
t
S
β
t
−σ
t
n
(S
∆t
t
n
)
β
¸
¸
¸
¸
2
=
¸
¸
¸
¸
σ
t
S
β
t
−σ
t
(S
∆t
t
n
)
β
+ σ
t
(S
∆t
t
n
)
β
−σ
t
n
(S
∆t
t
n
)
β
¸
¸
¸
¸
2
≤ 2
¸
¸
¸
¸
σ
t
S
β
t
−σ
t
(S
∆t
t
n
)
β
¸
¸
¸
¸
2
+ 2
¸
¸
¸
¸
σ
t
(S
∆t
t
n
)
β
−σ
t
n
(S
∆t
t
n
)
β
¸
¸
¸
¸
2
= 2σ
2
t
_
S
β
t
−(S
∆t
t
n
)
β
_
2
+ 2(σ
t
−σ
t
n
)
2
(S
∆t
t
n
)
2β
.
(C.15)
When Ω
1
is bounded compact set and does not contain 0, for any x, y ∈ Ω
1
there ∃z ∈ (x, y),
such that
(x
β
−y
β
)
2
= (βz
β−1
)
2
(x −y)
2
≤ C(Ω
1
)(x −y)
2
. (C.16)
Thus when Ω
1
, Ω
2
are bounded, B(t, S
t
, σ
t
) = σ
t
S
β
t
satisﬁes the local Lipschitz condition
(*).
We can show that the coeﬃcients of the logarithm of the SABR process satisfy the local
Lipschitz condition (*) in a similar way.
88
C.3 Appendix: Proof of Proposition 6.2.2 and Theorem
13–Strong Convergence of Meanreverting CEV Process
Proof of Proposition 6.2.2. Equation (6.16) gives the recursive relationship of the second
moment of V
∆t
n
:
E(V
∆t
n+1
)
2
= (1−λ∆t)
2
E(V
∆t
n
)
2
+2λ∆tµ(1−λ∆t)EV
∆t
n
+λ
2
∆t
2
µ
2
+σ
2
∆tEV
∆t
n

2β
. (C.17)
For β < 1, the Holder inequality gives
(EV
∆t
n

2β
)
1
2β
≤ (EV
∆t
n

2
)
1
2
⇒ EV
∆t
n

2β
≤ (EV
∆t
n

2
)
β
. (C.18)
Plugging the EV
∆t
n
from (6.17) into (C.17) and use (C.18), we have
E(V
∆t
n+1
)
2
≤ (1−λ∆t)
2
E(V
∆t
n
)
2
+2λ∆tµ(1−λ∆t)
_
(1−λ∆t)
n
(EV
∆t
0
−µ)+µ
_
+λ
2
∆t
2
µ
2
+σ
2
∆t(E(V
∆t
n
)
2
)
β
.
(C.19)
When 0 < β < 1, f(x) = x
β
is a concave function; so ∀x, the tangent line of f(x) is above
itself. Thus we can choose the tangent line of f(x) which pass the point (1, 1) as
g(x) = β(x −1) + 1,
such that x
β
< β(x −1) + 1. Let Z
0
= E(V
∆t
0
)
2
and Z
n
evolves by the following schemes:
Z
n+1
= (1−λ∆t)
2
Z
n
+2λ∆tµ(1−λ∆t)
_
(1−λ∆t)
n
(EV
∆t
0
−µ)+µ
_
+λ
2
∆t
2
µ
2
+σ
2
∆t
_
β(Z
n
−1)+1
_
,
(C.20)
The Z
n
derived in this way will be bigger than E(V
∆t
n
)
2
for all n > 1.
Simplify the notation and write Z
n
as:
Z
n+1
= aZ
n
+ b + cr
n
, (C.21)
where
a = (1 −λ∆t)
2
+ βσ
2
∆t, b = 2λ∆tµ
2
(1 −λ∆t) + λ
2
∆t
2
µ
2
+ (1 −β)σ
2
∆t,
c = 2λ∆tµ(1 −λ∆t)E(V
∆t
0
−µ).
(C.22)
89
It is well established that if a < 1, r < 1, then
lim
n→∞
Z
n
=
b
1 −a
. (C.23)
So when (1 − λ∆t)
2
+ βσ
2
∆t < 1 and 1 − λ∆t < 1, the limit of Z
n
exists, therefore the
Z
n
will be bounded and E(V
∆t
n
)
2
will be bounded.
Proof of Theorem 13. The proof is in similar spirit to the ones in Higham and Mao (2004).
We use a
k
, Φ
k
, Ψ
k
as deﬁned by them:
1. a
0
= 1, a
k
= e
−k(k+1)/2
.
2. For each k ≥ 1, Ψ
k
(µ) is a continuous function with support in (a
k
, a
k−1
) such that
0 ≤ Ψ
k
(µ) ≤
2
kµ
, for a
k
< µ < a
k−1
. Also we have
_
a
k−1
a
k
Ψ
k
(µ)dµ = 1.
3. Φ
k
(x) :=
_
x
0
dy
_
y
0
Ψ
k
(µ)dµ. Three properties of Φ
k
(x) will be used:
(a)
Φ
k
(x) ≤ 1. (C.24)
(b)
Φ
k
(x)
_
¸
_
¸
_
≤
2
kx
, for a
k
< x < a
k−1
= 0, otherwise
(C.25)
(c)
x −a
k−1
≤ Φ
k
(x) ≤ x, for all x ∈ R. (C.26)
First, we denote a continuous version of the Euler approximation (6.16) as
V
∆t
t
:= V
0
+
_
t
0
λ(µ −
ˆ
V
∆t
u
)du +
_
t
0
σ
ˆ
V
∆t
u

β
dW
u
(C.27)
where we introduce the piecewise constant process
ˆ
V
∆t
t
:= V
∆t
t
n
, for t ∈ [t
n
, t
n+1
). (C.28)
Using (C.27), we have
V
t
−V
∆t
t
= −λ
_
t
0
(V
u
−
ˆ
V
∆t
u
)du + σ
_
t
0
(V
u

β
−
ˆ
V
∆t
u

β
)dW
u
. (C.29)
90
Applying Ito’s rule to Φ
k
(V
u
−V
∆t
u
):
dΦ
k
(V
u
−V
∆t
u
) = Φ
k
(V
u
−V
∆t
u
)d(V
u
−V
∆t
u
) +
1
2
Φ
k
(V
u
−V
∆t
u
)(d(V
u
−V
∆t
u
))
2
. (C.30)
Plugging (C.29) into (C.30), we have
Φ
k
(V
t
−V
∆t
t
) =
_
t
0
Φ
k
(V
u
−V
∆t
u
)(−λ)(V
u
−
ˆ
V
∆t
u
)du + σ
_
t
0
Φ
k
(V
u
−V
∆t
u
)(V
u

β
−
ˆ
V
∆t
u

β
)dW
u
+
σ
2
2
_
t
0
Φ
k
(V
u
−V
∆t
u
)(V
u

β
−
ˆ
V
∆t
u

β
)
2
du.
(C.31)
Taking the expectation of both sides gives
EΦ
k
(V
t
−V
∆t
t
) = −λE
_
t
0
Φ
k
(V
u
−V
∆t
u
)(V
u
−
ˆ
V
∆t
u
)du +
σ
2
2
E
_
t
0
Φ
k
(V
u
−V
∆t
u
)(V
u

β
−
ˆ
V
∆t
u

β
)
2
du
≤ λ
_
t
0
E(V
u
−
ˆ
V
∆t
u
)du +
σ
2
2
I(t),
(C.32)
where the ﬁrst inequality comes from using (C.24), and
I(t) = E
_
t
0
Φ
k
(V
u
−V
∆t
u
)(V
u

β
−
ˆ
V
∆t
u

β
)
2
du
≤ E
_
t
0
Φ
k
(V
u
−V
∆t
u
)C(β)V
u
−
ˆ
V
∆t
u

2β
du
= E
_
t
0
Φ
k
(V
u
−V
∆t
u
)C(β)V
u
−V
∆t
u
+ V
∆t
u
−
ˆ
V
∆t
u

2β
du
≤ E
_
t
0
Φ
k
(V
u
−V
∆t
u
)C(β)C(2β)(V
u
−V
∆t
u

2β
+V
∆t
u
−
ˆ
V
∆t
u

2β
)du
= C
_
E
_
t
0
Φ
k
(V
u
−V
∆t
u
)V
u
−V
∆t
u

2β
du +
_
t
0
Φ
k
(V
u
−V
∆t
u
)V
∆t
u
−
ˆ
V
∆t
u

2β
du
_
≤ CE
_
t
0
2
k
1
(a
k
<V
u
−V
∆t
u
<a
k−1
)
V
u
−V
∆t
u

2β−1
du + CE
_
t
0
2
ka
k
V
∆t
u
−
ˆ
V
∆t
u

2β
du
≤ C
2T
k
+ C
_
t
0
2
ka
k
EV
∆t
u
−
ˆ
V
∆t
u

2β
du
≤ C
2T
k
+ C
2T
ka
k
(∆tD)
β
,
91
where the ﬁrst and second inequality use x + y
β
≤ C(x
β
+ y
β
) as proved in Lemma
C.3.1; the second to last and third to last inequality uses (C.25) and the last inequality use
Lemma C.3.2.
Plugging the above results into (C.32), we have
EΦ
k
(V
t
−V
∆t
t
) ≤ λ
_
t
0
EV
u
−V
∆t
u
du + λ
_
t
0
EV
∆t
u
−
ˆ
V
∆t
u
du + C
σ
2
T
k
+ C
σ
2
T
ka
k
(∆tD)
β
≤ λ
_
t
0
EV
u
−V
∆t
u
du + λ(∆tD)
1/2
+ C
σ
2
T
k
+ C
σ
2
T
ka
k
(∆tD)
β
,
where the second inequality uses Lemma C.3.2 again. Combining this with the lefthand
side of (C.26) gives
EV
t
−V
∆t
t
 ≤ a
k−1
+C
σ
2
T
k
+C
σ
2
T
ka
k
(∆tD)
β
+λ(∆tD)
1/2
+λ
_
t
0
E(V
t
−V
∆t
t
)du. (C.33)
Applying the Gronwall’s inequality to (C.33) gives the upper bound for EV
t
−V
∆t
t
,
sup
0≤t≤T
EV
t
−V
∆t
t
 ≤ e
λT
_
a
k
+ C
σ
2
T
k
+ C
σ
2
T
ka
k
(∆tD)
β
+ λ(∆tD)
1/2
_
. (C.34)
The terms in the parenthesis went to 0 as ∆t → 0. This proves (6.22).
Lemma C.3.1. ∀β > 0, there is a constant C(β), such that ∀X, Y ,
X + Y 
β
≤ C(β)(X
β
+Y 
β
). (C.35)
Proof. Note that
X + Y  ≤ X +Y  ⇒ ∀β > 0, X + Y 
β
≤ (X +Y )
β
.
So if we can show
1
C(β)
≤ (
X
X+Y 
)
β
+ (
Y 
X+Y 
)
β
, we are done. Let p =
X
X+Y 
, then
1 −p =
Y 
X+Y 
. The function f(p) = p
β
+(1 −p)
β
is continuous on the interval [0, 1], so it
can achieve the minimum. Moreover, the minimum is bigger than 0. So there is C(β) such
that the inequality holds.
92
Lemma C.3.2. Under conditions A and B, for t ∈ [t
k
, t
k+1
), there is a constant D such
that
EV
∆t
t
−
ˆ
V
∆t
t

2β
≤ (∆tD)
β
, (C.36)
EV
∆t
t
−
ˆ
V
∆t
t
 ≤
√
∆tD. (C.37)
Proof. When β < 1, the Holder inequality gives
EV
∆t
t
−
ˆ
V
∆t
t

2β
≤ (EV
∆t
t
−
ˆ
V
∆t
t

2
)
β
, (C.38)
so it is enough to show that there is a constant D such that E(V
∆t
t
−
ˆ
V
∆t
t
)
2
≤ ∆tD.
From (C.27), we have
(V
∆t
t
−
ˆ
V
∆t
t
)
2
=
_
(t −t
k
)λ(µ −V
∆t
k
) + σV
∆t
k

β
(W
t
−W
t
k
)
_
2
= (t −t
k
)
2
λ
2
(µ −V
∆t
k
)
2
+ σ
2
V
∆t
k

2β
(W
t
−W
t
k
)
2
+2(t −t
k
)λ(µ −V
∆t
k
)σV
∆t
k

β
(W
t
−W
t
k
).
Taking the expectation of both sides gives
E(V
∆t
t
−
ˆ
V
∆t
t
)
2
= (t −t
k
)
2
λ
2
E(µ −V
∆t
k
)
2
+ σ
2
(t −t
k
)EV
∆t
k

2β
≤ ∆t
2
λ
2
E(µ −S
k
)
2
+ σ
2
∆tES
k

2β
= ∆t
_
∆tλ(µ
2
−2µES
k
+ES
2
k
) + σ
2
ES
k

2β
_
≤ ∆t
_
λ(µ
2
−2µES
k
+ES
2
k
) + σ
2
ES
k

2β
_
≤ ∆tD
where the last inequality uses Proposition 6.2.2. Combining with (C.38), we have
EV
∆t
t
−
ˆ
V
∆t
t

2β
≤ (∆tD)
β
.
Finally, EV
∆t
t
−
ˆ
V
∆t
t
 ≤ (E(V
∆t
t
−
ˆ
V
∆t
t
)
2
)
1
2
≤
√
∆tD.
93
APPENDIX D
D.1 Appendix: Proof of Theorem 14 –Largestrike and
Largeexpiry Asymptotic of Displaced Lognormal
Lemma D.1.1. For A deﬁned in (7.3), T ∈ K
x
, deﬁne
σ
2
∞
(x, T) :=
_
_
_
2(2A + x −2
√
A
2
+ Ax) if x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
]
2(2A + x + 2
√
A
2
+ Ax) if x ∈ [−
1
2
σ
2
,
1
2
σ
2
]
(D.1)
we have
1
2
σ
2
∞
(x, T)
_
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
_
≤ x if x >
1
2
σ
2
≥ x if x ∈ [0,
1
2
σ
2
]
≥ (−x) if x ∈ [−
1
2
σ
2
, 0]
≤ (−x) if x < −
1
2
σ
2
(D.2)
For each x, the equalities in the ﬁrst and last equation in (D.2) hold only at most ﬁnite T,
so for T large enough, the strict inequalities hold and we have
1
2
σ
2
∞
(x, T)
_
¸
_
¸
_
< x if x >
1
2
σ
2
< (−x) if x < −
1
2
σ
2
(D.3)
Proof. We suppress the T in σ
2
∞
(x, T) to be σ
2
∞
(x). Note that σ
2
∞
(x) deﬁned in (D.1) is in
fact the solution of the function
A =
(−x + σ
∞
(x)
2
/2)
2
2σ
∞
(x)
2
. (D.4)
Denote
V
∗
BS
(x, σ) =
(x + σ
2
/2)
2
2σ
2
, (D.5)
94
then A = V
∗
BS
(˜ x, σ) − ˜ x and (D.4) can be written as
V
∗
BS
(˜ x, σ) − ˜ x = V
∗
BS
(x, σ
∞
(x)) −x. (D.6)
When x ∈ R/[−
1
2
σ
2
,
1
2
σ
2
], we choose σ
∞
(x) to be the smaller root which satisﬁes (D.4),
so
σ
2
∞
(x) = 2(2A + x −2
_
A
2
+ Ax). (D.7)
If x >
1
2
σ
2
> 0, using A =
(−˜ x+σ
2
/2)
2
2σ
2
≥ 0, we have
1
2
σ
2
∞
(x) −x = 2A −2
_
A
2
+ Ax ≤ 0. (D.8)
The equality holds if and only if ˜ x = σ
2
/2, which happens at most at one T. If x < −
1
2
σ
2
< 0,
using A + x =
(x+σ
2
∞
/2)
2
2σ
2
∞
≥ 0 and 0 ≤ A + x < A, we have
1
2
σ
2
∞
(x) −(−x) = 2(A + x) −2
_
A
2
+ Ax ≤ 0. (D.9)
The equality hold if and only if A+x =
(−˜ x+σ
2
/2)
2
σ
2
+x = 0, which again, happens at most
at two T’s.
When x ∈ [−
1
2
σ
2
,
1
2
σ
2
], we choose σ
∞
(x) to be the larger root which satisﬁes (D.4), so
σ
2
∞
(x) = 2(2A + x + 2
_
A
2
+ Ax). (D.10)
If 0 ≤ x ≤
1
2
σ
2
, using A ≥ 0 gives
1
2
σ
2
∞
(x) −x = 2A + 2
_
A
2
+ Ax ≥ 0. (D.11)
If −
1
2
σ
2
≤ x < 0, using A + x ≥ 0 gives
1
2
σ
2
∞
(x) −(−x) = 2(A + x) + 2
_
A
2
+ Ax ≥ 0. (D.12)
Proof. [Proof of Theorem 14.]
95
We ﬁrst show the case when x > σ
2
/2. For S from displaced lognormal dynamics, denote
˜
S = S −θ, then
˜
S will be a lognormal process and
1
˜
S
0
E(
˜
S
T
−
˜
S
0
e
˜ xT
)
+
= N(z
+
) −e
˜ xT
N(z
−
), (D.13)
where z
±
=
(−˜ x±σ
2
/2)
√
T
σ
.
Recall an approximation of the cumulative density function of standard normal N(z)
from Olver(1974):
N(−z) = 1 −N(z) =
exp(−z
2
/2)
z
√
2π
(1 + O(1/z
2
)), as z → +∞. (D.14)
When x > σ
2
/2, there is a T
0
such that when T > T
0
, ˜ x > σ
2
/2. Then as T → ∞, we
have z
±
→ −∞. Using (D.14) and A
BS
in (7.7), we have
N(z
+
) −e
˜ xT
N(z
−
) =
1
√
2πT
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
_
σ
˜ x −σ
2
/2
−
σ
˜ x + σ
2
/2
_
(1 + O(
1
T
))
=
1
√
2πT
e
−
(˜ x−σ
2
/2)
2
2σ
2
T
A
BS
(˜ x, σ, 0)(1 + O(
1
T
))
=
1
√
2πT
e
−(V
∗
BS
(˜ x,σ)−˜ x)T
A
BS
(˜ x, σ, 0)(1 + O(
1
T
)).
(D.15)
Consequently,
1
S
0
E(S
T
−S
0
e
xT
)
+
=
˜
S
0
S
0
1
˜
S
0
E(
˜
S
T
−
˜
S
0
e
˜ xT
)
+
=
˜
S
0
S
0
1
√
2πT
e
−(V
∗
BS
(˜ x,σ)−˜ x)T
A
BS
(˜ x, σ, 0)(1 + O(
1
T
)).
(D.16)
We choose σ
∞
(x) and ˆ a
1
so that
V
∗
BS
(˜ x, σ) − ˜ x = V
∗
BS
(x, σ
∞
(x)) −x, (D.17)
˜
S
0
S
0
A
BS
(˜ x, σ, 0) = A
BS
(x, σ
∞
(x), ˆ a
1
(x)). (D.18)
96
Plugging (D.17) and (D.18) into (D.16), we have
1
S
0
E(S
T
−S
0
e
xT
)
+
=
1
√
2πT
e
−(V
∗
BS
(x,σ
∞
(x))−x)T
A
BS
(x, σ
∞
(x), ˆ a
1
(x))(1+O(
1
T
)). (D.19)
Now if we could show that −x + (σ
∞
(x)
2
+ ˆ a
1
(x)/T)/2 > 0 for large enough T, then we
could apply (D.14) to show that (D.19) is an asymptotic formula for the BlackScholes call
price with volatility
_
σ
2
∞
(x) + ˆ a
1
(x)/T.
Denote ˆ σ
2
T
= σ
2
∞
(x)+
ˆ a
1
(x)
T
. By Lemma D.1.1 we know that x > σ
2
∞
(x)/2. lim
T→∞
ˆ a
1
(x)
converges to a constant. So for T large enough, we have x > ˆ σ
2
T
/2. Using (D.14), the Black
Scholes call price formula can be approximated as
C
BS
(S
0
, S
0
e
xT
, ˆ σ
T
, T) =
1
√
2πT
exp
_
−
1
2
(
x
2
ˆ σ
2
T
−x +
ˆ σ
2
T
4
)T
__
ˆ σ
T
x − ˆ σ
2
T
/2
−
ˆ σ
T
x + ˆ σ
2
T
/2
_
(1 + O(1/T))
=
1
√
2πT
e
−(V
∗
BS
(x,σ
∞
(x))−x)T
A
BS
(x, σ
∞
(x), ˆ a
1
(x))(1 + O(
1
T
)).
(D.20)
Intuitively, comparing (D.19) and (D.20), we can see that the implied volatility of the
displaced lognormal process is
σ
2
imp
(x, T) = σ
2
∞
(x) + ˆ a
1
(x)/T + o(1/T). (D.21)
The rigorous proof is as follows. From (D.19), ∀ > 0,
1
S
0
E(S
T
−S
0
e
xT
)
+
=
A
BS
(x, σ
∞
, ˆ a
1
)
√
2πT
exp(−(V
∗
BS
(x, σ
∞
) −x)T)(1 + O(1/T)) (D.22)
≤
A
BS
(x, σ
∞
, ˆ a
1
)
√
2πT
exp(−(V
∗
BS
(x, σ
∞
) −x)T)e
. (D.23)
For any δ > 0, Lemma D.1.1 shows that (δ) = (
4x
2
σ
4
∞
(x)
− 1)
δ
16
> 0. Thus, we can choose
> 0 in the previous step so that −(δ) < 0, and
exp
_
1
8
ˆ a
1
(x)(
4x
2
σ
2
∞
−1) +
_
≤ exp
_
1
8
(ˆ a
1
(x) + δ)(
4x
2
σ
2
∞
−1) + −(δ)
_
.
97
Using (D.23) and (D.20), ∀δ > 0, there exists T
1
, such that when T > T
1
,
1
S
0
E(S
T
−S
0
e
xT
)
+
≤
A
BS
(x, σ
∞
, ˆ a
1
+ δ)
√
2πT
exp
_
−(V
∗
BS
(x, σ
∞
) −x)T + −(δ)
_
≤ C
BS
(S
0
, S
0
e
xT
, σ
2
∞
(x) + (ˆ a
1
(x) + δ)/T, T)
Using the monotonicity of the BlackScholes formula, the implied volatility of the displaced
lognormal process σ
imp
(x) is
σ
2
imp
(x, T) ≤ σ
2
∞
(x) + (ˆ a
1
(x) + δ)/T.
Similarly, we can prove the lower bound, so the equality in (7.5) holds.
For the x < −σ
2
/2 case, we consider the put price on the displaced lognormal dynamics,
and apply similar methods.
D.2 Appendix: Proof of Theorem 15–First Order Approximation
of Largestrike Largeexpiry Implied Volatility
Lemma D.2.1. For x ∈ (−σ
2
/2, σ
2
/2), we have
lim
T→∞
E(S
T
−S
0
e
xT
)
+
= S
0
−θ. (D.24)
Proof. Apply the asymptotic approximation of N(.) to get the results. More speciﬁcally,
1
S
0
E(S
T
−S
0
e
xT
)
+
=
˜
S
0
S
0
1
˜
S
0
E(
˜
S
T
−
˜
S
0
e
˜ xT
)
+
=
˜
S
0
S
0
_
1 +
1
√
2πT
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
_
σ
˜ x −σ
2
/2
−
σ
˜ x + σ
2
/2
_
(1 + O(
1
T
))
_
= 1 −
θ
S
0
+
˜
S
0
S
0
1
√
2πT
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
_
σ
˜ x −σ
2
/2
−
σ
˜ x + σ
2
/2
_
(1 + O(
1
T
))
(D.25)
Proof. [Proof of Theorem 15.]
98
When θ < 0, lim
T→∞
E(S
0
−S
0
e
xT
) = S
0
−θ > S
0
, so σ
imp
(T) is not well deﬁned when
T is large.
When θ > 0, denote B := N
−1
(
S
0
−θ
S
0
). Deﬁne ¯ σ(T) as
¯ σ(T) :=
B +
√
B
2
+ 2xT
√
T
.
Abbreviate ¯ σ(T) as ¯ σ, it satisﬁes the equation:
−x + ¯ σ
2
/2
¯ σ
√
T = B, (D.26)
and lim
T→∞
(−x−¯ σ
2
/2)
√
T
¯ σ
= lim
T→∞
−
√
B
2
+ 2xT = −∞ . So
lim
T→∞
C
BS
(S
0
, S
0
e
xT
, ¯ σ, T) = lim
T→∞
S
0
N
_
(−x + ¯ σ
2
/2)
√
T
¯ σ
_
−S
0
e
xT
N
_
(−x − ¯ σ
2
/2)
√
T
¯ σ
_
= S
0
−θ.
(D.27)
Recall that lim
T→∞
C
BS
(S
0
, S
0
e
xT
, σ
imp
(T)) = S
0
−θ, so
lim
T→∞
σ
imp
(x, T) = lim
T→∞
¯ σ(T) = 2x. (D.28)
D.3 Appendix: Proof of Theorem 16–Asymptotic Formula of
BlackScholes Call Option
Proof.
1
S
0
C
BS
(S
0
, S
0
e
xT
,
¸
2x + 2B
_
2x
T
+
a(T)
T
, T)
=
_
N
_
(B
_
2x
T
+
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
−N(B)
_
+ N(B) −e
xT
N
_
(−2x −B
_
2x
T
−
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
(D.29)
99
Applying Taylor expansion to the ﬁrst term, we have
N
_
(B
_
2x
T
+
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
−N(B)
=
1
√
2πT
exp
_
−
1
2
_
(B
_
2x
T
+
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
2
__
(B
_
2x
T
+
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
−B
_
√
T(1 + O(
1
T
))
=
1
√
2πT
exp(−
1
2
B
2
)
a(T)/2 −B
2
√
2x
(1 + O(1/
√
T)).
(D.30)
For the last term, using (D.14) gives
e
xT
N
_
(−2x −B
_
2x
T
−
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
=
1
√
2πT
exp
_
−
1
2
_
(B
_
2x
T
+
a(T)
2T
)
√
T
_
2x + 2B
_
2x
T
+
a(T)
T
_
2
_
_
2x + 2B
_
2x
T
+
a(T)
T
2x + B
_
2x/T +
a(T)
2T
(1 + O(
1
T
)
=
1
√
2πT
exp(−
1
2
B
2
)
1
√
2x
(1 + O(1/
√
T)).
(D.31)
Plugging (D.30) and (D.31) into (D.29), we have
1
S
0
C
BS
(S
0
, S
0
e
xT
,
¸
2x + 2B
_
2x
T
+
a(T)
T
, T)
= N(B) +
1
√
2πT
e
−
1
2
B
2 1
√
2x
(a(T)/2 −B
2
−1)(1 + O(1/
√
T)).
(D.32)
100
D.4 Appendix: Proof of Theorems 17 and 18– Second and Third
Order Approximation of Largestrike Largeexpiry Implied
Volatility
Proof of Theorem 17. Using (D.25) and the deﬁnition of a(T) and B, we have ∀ > 0,
1
S
0
E(S
T
−S
0
e
xT
)
+
= 1 −
θ
S
0
+
˜
S
0
S
0
1
√
2πT
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
_
σ
˜ x −σ
2
/2
−
σ
˜ x + σ
2
/2
_
(1 + O(
1
T
))
≤ 1 −
θ
S
0
+
˜
S
0
S
0
1
√
2πT
e
−
1
2
(
˜ x
2
σ
2
−˜ x+
σ
2
4
)T
_
σ
˜ x −σ
2
/2
−
σ
˜ x + σ
2
/2
_
e
= N(B) +
1
√
2πT
exp(−
1
2
B
2
)
1
√
2x
(a(T)/2 −B
2
−1)e
(D.33)
From the deﬁnition of a(T), we know that a(T) > B
2
+ 1. ∀ > 0, ∀δ > 0, there ∃(δ) > 0,
such that ∃T
0
, ∀T > T
0
(a(T)/2 −B
2
−1)e
< ((a(T) + δ)/2 −B
2
−1)e
−(δ)
. (D.34)
Using (D.33), (D.34) and proposition 16, we have
1
S
0
E(S
T
−S
0
e
xT
)
+
≤ N(B) +
1
√
2πT
exp(−
1
2
B
2
)
1
√
2x
(a(T)/2 −B
2
−1)e
≤ N(B) +
1
√
2πT
exp(−
1
2
B
2
)
1
√
2x
((a(T) + δ)/2 −B
2
−1)e
−(δ)
≤
1
S
0
C
BS
(S
0
, S
0
e
xT
,
¸
2x + 2B
_
2x
T
+
a(T) + δ
T
, T).
(D.35)
So
σ
2
imp
(x, T) ≤ 2x + 2B
_
2x
T
+
a(T) + δ
T
.
Similarly, we can show that
σ
2
imp
(x, T) ≥ 2x + 2B
_
2x
T
+
a(T) −δ
T
.
101
This completes the proof.
Proof of Theorem 18. Similar to the proof of Theorem 17.
D.5 Appendix: Proof of Theorem 19–Largestrike Largeexpiry
Atthemoney Implied Volatility
Proof. Suppress the 0 in σ
imp
(0, T) to be σ
imp
(T). When atthemoney, for the displaced
lognormal, we have
C
BS
(S
0
, S
0
e
xT
, σ
imp
(T), T) = S
0
(2N(σ
imp
√
T/2) −1),
On the other hand, S −θ is lognormal process, so
E(S
T
−K)
+
= E((S
T
−θ) −(K −θ))
+
= (S
0
−θ)(2N(σ
√
T/2) −1).
So
S
0
(2N(σ
imp
√
T/2) −1) = (S
0
−θ)(2N(σ
√
T/2) −1),
⇒ N(σ
imp
(T)
√
T/2) =
S
0
−θ
S
0
N(σ
√
T/2) +
θ
S
0
;
⇒ σ
imp
(T) =
2
√
T
N
−1
_
S
0
−θ
S
0
N(σ
√
T/2) +
θ
2S
0
_
.
Taking the limit of T, we have (7.14).
D.6 Appendix: Proof of Theorem 20–Approximation of Implied
Volatility when K = S
0
exp(xT
α
)
Proof. Here we sketch the outline of the proof. The rigorous proof is similar to the proof of
Theorem 17. We prove the x > 0 case. Denote ˜ x =
1
T
α
log(
S
0
e
xT
α
−θ
S
0
−θ
), M = xT
α−1
,
˜
M =
˜ xT
α−1
and A =
_
−
˜
M+σ
2
/2
σ
_
2
. σ
2
∞
satisﬁes
_
−M+σ
2
∞
/2
σ
∞
_
2
=
_
−
˜
M+σ
2
/2
σ
_
2
. Choose the
smaller root σ
2
∞
= 2(M + A) −
√
2MA + A
2
so that
σ
2
∞
2
− M < 0. Abbreviate σ
imp
(x, T)
102
to be σ
imp
. Using the approximation of N(.), we have
1
S
0
E(S
T
−S
0
e
xT
α
)
+
=
˜
S
0
S
0
1
˜
S
0
E(
˜
S
T
−
˜
S
0
e
˜ xT
α
)
+
=
˜
S
0
S
0
_
N
_
(−˜ xT
α−1
+ σ
2
/2)
√
T
σ
_
−e
xT
α
N
_
(−˜ xT
α−1
−σ
2
/2)
√
T
σ
__
=
˜
S
0
S
0
1
√
2πT
exp
_
−
1
2
_
−˜ xT
α−1
+ σ
2
/2
σ
_
2
T
__
σ
˜ xT
α−1
−σ
2
/2
−
σ
˜ xT
α−1
+ σ
2
/2
_
(1 + o(1/T))
=
˜
S
0
S
0
1
√
2πT
exp
_
−
1
2
_
−M + σ
2
∞
/2
σ
∞
_
2
T
_
A
BS
(
˜
M, σ, 0)(1 + o(1/T))
=
1
√
2πT
exp
_
−
1
2
_
−M + σ
2
∞
/2
σ
∞
_
2
T
_
exp
_
a
8
(
4M
2
σ
2
∞
) −1
_
A
BS
(M, σ
∞
, 0)(1 + o(1/T))
=
1
√
2πT
exp
_
−
1
2
_
−M + σ
2
imp
/2
σ
imp
_
2
T
__
σ
imp
xT
α−1
−σ
2
imp
/2
−
σ
imp
xT
α−1
+ σ
2
imp
/2
_
(1 + o(1/T))
→
1
S
0
C
BS
(S
0
, S
0
e
xT
α
, σ
imp
, T).
(D.36)
The x < 0 case can be similarly proved.
D.7 Appendix: Proof of Theorem 21–Fixedstrike Largeexpiry
Implied Volatility
D.7.1 Asymptotic behavior
Proof. Suppress the x in the σ
imp
(x, T) to be σ
imp
(T). On the one hand, we have
lim
T→∞
E(S
T
−K)
+
= lim
T→∞
S
0
×N
_
log(S
0
/K)
σ
imp
(T)
√
T
+
σ
imp
(T)
√
T
2
_
−K×N
_
log(S
0
/K)
σ
imp
(T)
√
T
−
σ
imp
(T)
√
T
2
_
.
(D.37)
103
On the other hand,
lim
T→∞
E(S
T
−K)
+
= lim
T→∞
E((S
T
−θ) −(K −θ))
+
= lim
T→∞
(S
0
−θ)N(
log((S
0
−θ)/(K −θ)) + σ
2
T/2
σ
√
T
)
−(K −θ)N(
log((S
0
−θ)/(K −θ)) −σ
2
T/2
σ
√
T
)
= S
0
−θ.
(D.38)
Comparing (D.37) and (D.38), we know that the limit of σ
imp
(T)
√
T exists and satisfy
the following condition
lim
T→∞
S
0
×N
_
log(S
0
/K)
σ
imp
(T)
√
T
+
σ
imp
(T)
√
T
2
_
−K×N
_
log(S
0
/K)
σ
imp
(T)
√
T
−
σ
imp
(T)
√
T
2
_
= S
0
−θ.
(D.39)
Recall the deﬁnition of σ
blsimpv
(S
0
, K, P) in (7.21) is
S
0
×N
_
log(S
0
/K)
σ
blsimpv
(S
0
, K, P)
+
σ
blsimpv
(S
0
, K, P)
2
_
−K×N
_
log(S
0
/K)
σ
blsimpv
(S
0
, K, P)
−
σ
blsimpv
(S
0
, K, P)
2
_
= P,
(D.40)
comparing (D.39) and (D.40), we have (7.20).
D.7.2 Monotonicity
Proof. Recall the σ
imp
(T) is deﬁned as follows
C
BS
(S
0
, K, σ
imp
(T), T) = C
BS
(S
0
−θ, K −θ, σ, T). (D.41)
Diﬀerentiating both sides with respect to T gives us
∂C
BS
(S
0
, K, σ
imp
(T), T)
∂T
+
∂C
BS
(S
0
, K, σ
imp
(T), T)
∂σ
imp
(T)
∂σ
imp
(T)
∂T
=
∂C
BS
(S
0
−θ, K −θ, σ)
∂T
.
(D.42)
104
Using
∂C
BS
(S
0
,K,σ
imp
(T),T)
∂σ
imp
(T)
> 0, we have
sgn
∂σ
imp
(T)
∂T
= sgn
_
∂C
BS
(S
0
−θ, K −θ, σ)
∂T
−
∂C
BS
(S
0
, K, σ
imp
(T))
∂T
_
= sgn
_
(S
0
−θ)N
_
log((S
0
−θ)/(K −θ))
σ
√
T
+
σ
√
T
2
_
σ
−S
0
N
_
log(S
0
/K)
σ
imp
(T)
√
T
+
σ
imp
(T)
√
T
2
_
σ
imp
(T)
_
= sgn
_
(S
0
−θ) exp
_
−
1
2
_
log((S
0
−θ)/(K −θ))
σ
√
T
+
σ
√
T
2
_
2
_
σ
−S
0
exp
_
−
1
2
_
log(S
0
/K)
σ
imp
(T)
√
T
+
σ
imp
(T)
√
T
2
_
2
_
σ
imp
(T)
_
.
(D.43)
From (D.43), we have
lim
T→∞
sgn
∂σ
imp
(T)
∂T
= lim
T→∞
sgn
_
(S
0
−θ) exp
_
−
1
2
_
log((S
0
−θ)/(K −θ))
σ
√
T
+
σ
√
T
2
_
2
_
σ
√
T
−S
0
exp
_
−
1
2
_
log(S
0
/K)
σ
imp
(T)
√
T
+
σ
imp
(T)
√
T
2
_
2
_
σ
imp
(T)
√
T
_
= lim
T→∞
sgn
__
K −θ
S
0
−θ
_
1/2
exp
_
−
1
2
σ
2
T
4
_
σ
√
T
−
_
K
S
0
_
1/2
exp
_
−
1
2
σ
2
imp
T
4
_
σ
imp
(T)
√
T
_
.
(D.44)
Note that exp
_
−
1
2
σ
2
T
4
_
σ
√
T as T → ∞ is dominated by the exponential term, so the ﬁrst
term goes to 0. From Theorem 21.1, lim
T→∞
σ
imp
(T)
√
T exists, so the second term goes to
some positive constant. Therefore,
lim
T→∞
sgn
∂σ
imp
(T)
∂T
< 0.
105
D.8 Appendix: Proof of Theorem 22–Approximation of
FixedStrike Largeexpiry Implied Volatility.
Proof. Denote a =
σ
2
blsimpv
(S
0
,K,S
0
−θ)
2
. From Theorem 21, we know that the ﬁrst term of
σ
imp
(x, T) is σ
blsimpv
(S
0
, K, S
0
−θ)/
√
T = 2a/T. Here we want a more reﬁned approxima
tion of the residual term. So denote σ
2
imp
(x, T) =
2a
T
+ 2r. Denote
˜
S
0
= S
0
−θ and use the
Taylor approximation of N(.) to the second order, we have
1
S
0
E(S
0
−K)
+
=
˜
S
0
S
0
+
_
N(
−x + a + rT
√
2a + 2rT
) −N(
−x + a
√
2a
)
_
+ e
x
_
N(
−x −a −rT
√
2a + 2rT
) −N(
−x −a
√
2a
)
_
=
˜
S
0
S
0
+
1
√
2π
exp(−
1
2
(
−x + a
√
2a
)
2
)
__
√
2a + 2rT −
√
2a
_
+
_
√
2a + 2rT −
√
2a
_
2
_
x
2
4a(a + rT)
_
_
a + rT
a
−
1
2
_
−
1
8
__
+ o((rT)
2
)
(D.45)
Let
√
2a + 2rT −
√
2a = z, then
1
S
0
E(S
0
−K)
+
=
˜
S
0
S
0
+
1
√
2π
exp(−
1
2
(
−x + a
√
2a
)
2
)
_
z + z
2
_
x
2
2a(z +
√
2a)
2
_
z
√
2a
+
1
2
_
−
1
8
__
+ o((rT)
2
)
≈
˜
S
0
S
0
+
1
√
2π
exp(−
1
2
(
−x + a
√
2a
)
2
)
_
(−
1
8
+
x
2
8a
2
)z
2
+ z
_
(D.46)
On the other hand, we have
1
S
0
E(S
0
−K)
+
=
1
S
0
C
BS
(S
0
−θ, K −θ, σ, T) (D.47)
Compare (D.46) with (D.47) and solve for z.
106
REFERENCES
[1] Beaglehole, D., Dybvig, P., Zhou, G.: Going to extremes: Correcting simulation bias in
exotic option valuation. Financial Anal. J. 53(1), 6268 (1997)
[2] Berestycki, H., Busca, J., Florent, I.: Asymptotics and calibration of local volatility
models. Quant. Finance 2, 6169 (2002)
[3] Berestycki, H., Busca, J., Florent, I.: Computing the implied volatility in stochastic
volatility models. Commun. Pure Appl. Math. 57, 13521373 (2004)
[4] Black, F., Scholes, M.: The pricing of options and corporate liabilities. J. Political Econ.
81, 637659 (1973)
[5] Boyle, P., Broadie, M., Glasserman, P.: Monte Carlo methods for security pricing. J.
Econ. Dyn. Control 21, 12671321 (1997)
[6] Brigo, D., Mercurio, F.: Displaced and mixture diﬀusions for analyticallytractable smile
models. Mathematical Finance – Bachelier Congress 2000, 151174, Springer Finance,
Springer, Heidelberg (2001)
[7] Broadie, M., Kaya, O.: Exact simulation of option greeks under stochastic volatility and
jump diﬀusion models. Proceedings of the 2004 Winter Simulation Conference, (2004)
[8] Broadie, M., Kaya, O.: Exact simulation of stochastic volatility and other aﬃne jump
diﬀusion processes. Oper. Res. 54(2), 217231 (2006)
[9] Broadie, M., Yamamoto, Y.: A doubleexponential fast Gauss transform algorithm for
pricing discrete pathdependent options. Oper. Res. 53(5), 764779 (2005)
[10] Cox, J.: The constant elasticity of variance option pricing model. J. Portf. Manag. 23,
1517 (1996)
[11] Davydov, D., Linetsky, V.: Pricing and hedging pathdependent options under the cev
process. Manag. Sci. 47(7), 949965 (2001)
107
[12] Durrleman, V.: From implied to spot volatilities. PhD dissertation. Dept of Operational
Resesarch and Financial Engineering, Princeton (2004)
[13] Feng, L., Linetsky, V.: Pricing discretely monitored barrier options and defaultable
bonds in L´evy process models. Math. Finance 18(3), 337384 (2008)
[14] Fisher, T., Tataru, G.: Mixing Models. Bloomberg Markets Apr 2010, 136137
[15] Forde, M., Jacquier, A.: The largematurity smile for the Heston Model.
[16] Forde, M., Jacuier, A., Mijatovic, A.: Asymptotic formulae for implied volatility in the
Heston model.
[17] Fusai, G., Meucci, A.: Pricing discretely monitored Asian options under Levy Processes.
Journal of Banking and Finance. 32, 20762088. (2008).
[18] Griebsch, S., Wystup, U.: On the valuation of fader and discrete barrier options in
Heston’s stochastic volatility model. submitted for publication. (2008)
[19] Hagan, P., Kumar, D., Lesniewski, A., Woodward, D.: Managing smile risk. Wilmott,
1(8), 84108 (2002)
[20] Hesterberg, T. C.: Control variate and importance sampling for eﬃcient boostrap sam
pling. Stat. Comput. 6, 147157 (1996)
[21] Heston, S.: A ClosedForm Solution for Options with Stochastic Volatility with Appli
cations to Bond and Currency Options. The Review of Financial Studies. 6(2), 327343
(1993)
[22] Higham, D. J., Mao, X.: Convergence of the Monte Carlo simulations involving the
meanreverting square root process. J. Comput. Finance 8(3) 3562 (2005)
[23] Hirsa, A., Courtadon, G., Madan, D.: The eﬀect of model risk on the valuation of
barrier options. J. Risk Finance 4, 4755 (2003)
[24] Joshi, M., Rebonato, R.: A displaceddiﬀusion stochastic volatility LIBOR market
model: motivation, deﬁnition and implementation. Quant. Finance 3, 458469 (2003)
[25] Kahl, C.: Positive numerical integration of stochastic diﬀerential equations. Diploma
Thesis, University of Wuppertal (2004)
108
[26] Kahl, C., Jackel, P.: Fast astrong approximation Monte Carlo Schemes for stochastic
volatiity models. working paper (2006)
[27] Kloeden, P. E., Platen, E.: Numerical solution of stochastic diﬀerential equations.
SpringerVerlag (1992)
[28] Lee, R., Wang, D.: Displaced lognormal volatility skews: analysis and applications to
stochastic volatility simulations. Ann. Finance (2009), DOI 10.1007/s1043600901457
[29] Lewis, A.: Option Valuation under Stochastic Volatility. Finance Press, Newport Beach,
California, USA (2000)
[30] Lord, R., Koekkoek, R., Dijk, D. V.: A comparison of biased simulation schemes for
stochastic volatility models. Working Paper (2008)
[31] Marris, D.: Financial option pricing and skewed volatility. MPhil Thesis, Statistical
Laboratory, University of Cambridge (1999)
[32] Milevsky, M. A., Posner, S.: Asian options, the sum of lognormals, and the reciprocal
Gamma distribution. Jour of Financial and Quantitative Analysis. Vol 33(3). (1998).
[33] Petrella, G., Kou, S.: Numerical pricing of discrete barrier and lookback options via
Laplace transforms. J. Comput. Finance 8, 137 (2004)
[34] Rebonato, R. Volatility and Correlation. John Wiley & Sons, 2nd edition (2004)
[35] Roper, M.: Implied volatility: General properties and asymptotics. PhD Thesis, Uni
versity of New South Wales (2009)
[36] Roper, M., Rutkowski, M.: A note on the behaviour of the BlackScholes implied volatil
ity close to expiry. University of New South Wales (2007)
[37] Rubinstein, M.: Displaced diﬀusion option pricing. J. Finance 38(1), 213217 (1983)
[38] SvobodaGreenwood, S.: Displaced diﬀusion as an approximation of the constant elas
ticity of variance. Appl. Math. Finance (2009) Forthcoming.
[39] Zhang, Q., Zhang, W., Nie Z.: Convergence of the Euler scheme for stocahstic functional
partial diﬀerential equations. Appl. Math. Comput 155, 497492 (2004)
109
To my parents
TABLE OF CONTENTS
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PUBLICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 DISPLACED LOGNORMAL PROCESS 2.1 Implied Volatility . . . . . . . . . . 2.2 Displaced Lognormal . . . . . . . . 2.2.1 Implied volatility . . . . . . 2.2.2 Global behavior . . . . . . . 2.2.3 Atthemoney behavior . . . 2.2.4 Shortexpiry behavior . . . . 2.3 Displaced antiLognormal . . . . . 2.3.1 Implied volatility . . . . . . 2.3.2 Global behavior . . . . . . . 2.3.3 Atthemoney behavior . . . 2.3.4 Shortexpiry behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi vii viii ix x 1 4 4 5 5 6 7 8 9 10 10 11 12 13 13 13 14 15 17 18 19 20
3 DISPLACED HESTON PROCESS . . . . . . . . . 3.1 Displaced Heston . . . . . . . . . . . . . . . . 3.1.1 Implied volatility . . . . . . . . . . . . 3.1.2 Atthemoney behavior . . . . . . . . . 3.1.3 Shortexpiry behavior . . . . . . . . . . 3.2 Displaced antiHeston . . . . . . . . . . . . . 3.2.1 Implied volatility . . . . . . . . . . . . 3.2.2 Shortexpiry behavior . . . . . . . . . . 3.2.3 Generalization of shortexpiry behavior
iii
. . . . . . . . . . .1. 5. . . . . . . 4. 8 CONCLUSION . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 LARGEEXPIRY IMPLIED VOLATILITY OF DISPLACED LOGNORMAL 7. 5. . . . . . . . . . 7.2. . . . . . . . . . . . . . . . . . . . 5. 4. .2. . . 5. . . . . . . . . . . σ 2 /2) . .1. . . . . . . . . . . . .1 Importance sampling on options pricing . .1. . . 1 σ 2 ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . α 7. 5. . . .2 Variance Reduction Combining Control Variate and Importance Sampling .2 Fixedstrike Largeexpiry Implied Volatility . . . . . . . . . .2.1. .5 Numerical results II: Discretely sampled arithmetic Asian option under SABR dynamics . . . . .3 Calibration of DH to SABR . . x ∈ (−σ 2 /2. . . . . . . . . .2. . . . . . . . . . . . . .2 Case two: K = S0 exT . . . . . . . . . . . . . . .1. . . . . . 4. . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . 4. . 5. 7. . . .1. . . . . . . .1 Partial Strong Convergency of Stochastic Volatility Process .3 Combine control variate with importance sampling . 5. . . 6. . . . . . . . . . . .2 Calibration of DH process .2.3 Discretisation Schemes Used in the Monte Carlo Simulation . . .1 Calibration of DL process . . . . . . . . . . . . . . . . . . . . . . . . . .2 Calibration DL and DH to CEV/SABR . . .4 Example: Discretely sampled barrier option under SABR dynamics . . . . . .1 Case one: K = S0 exT . .5 Numerical results: Discretely sampled barrier option . .1 CEV and SABR stochastic volatility models 4. . . . . .1. . . . . 6 DISCRETISATION SCHEME . .1 Largestrike and Largeexpiry Behavior . . . . . . . . . . .1 DL or DH as a control variate . . . . . . . . . . . . . . . .2. . . 6.2 Example I: Discretely sampled barrier option under CEV/SABR dynamics . . . . . .2. 22 22 22 23 23 23 25 26 27 27 28 29 31 33 34 35 35 37 37 38 40 43 44 46 48 50 50 50 52 55 56 61 5 VARIANCE REDUCTION IN MONTE CARLO SIMULATION . . 4. . . . . . . . .4 CALIBRATION OF DL AND DH PROCESS . . . . . . . . . . . . . . . . . . . 6. . . . . . . . . .3 Case three: K = S0 exT . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . x ∈ R/[− 1 σ 2 . . . 2 2 7. . . . . . . . . . . . .1. .1 Calibration of DL and DH Process . . . .2. . .2 Strong Convergence of Meanreverting CEV Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Variance Reduction Using Control Variate . . . . . . . . . . . . . . . . . . .1. . . . .2 Calibration of DL to CEV/SABR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . 5. . 4.2 Drifted DH/DL process . . . Appendix iv . . . . . . . . . . .3 Numerical results I: Discretely sampled barrier option under CEV/SABR dynamics . . . . .4 Example II: Discretely sampled arithmetic Asian option under SABR dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . .
. . . . . . . . . . . . . C. . . . . . . . . . .7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Appendix: Coeﬃcients of SABR Satisfy the Local Lipschtiz Condition (*) . . .5 . . . . . . . . . . . . . . . . . .1 Asymptotic behavior . . . .3 A.5 Appendix: Proof of Theorem 19–Largestrike Largeexpiry Atthemoney Implied Volatility . . . .3 Appendix: Proof of Theorem 16–Asymptotic Formula of BlackScholes Call Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Appendix: Proof of Theorem 20–Approximation of Implied Volatility when K = S0 exp(xT α ) . . . . . . . . . . . . . . . . . . . . . . D. . . . . .3 Appendix: Proof of Proposition 6. . . . . . . . . . . . . . .3 Appendix: Proof of Propositions 3. . . D. . . . . . . . . . . . . . Appendix: Proof of Theorem 2–Atthemoney Behavior of Displaced Lognormal Appendix: Proof of Theorems 3 and 6–Shortexpiry Behavior of DL . . D. . .1 Appendix: Proof of Theorem 7–Atthemoney Behavior of Displaced Independent Stochastic Volatility . . . . . .4 A. . . . . . . . . . . . . . . .2. 63 63 69 70 71 75 77 77 78 80 85 85 88 89 94 94 98 99 101 102 102 103 103 104 106 B . . .A . . . .1 Appendix: Proof of Theorem 12–Partial Strong Convergency of Stochastic Volatility Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . A. . . . . . D. . . . . . . . . . . . . . . . . . . . . . . . . . Appendix: Proof of Theorem 1–Global Behavior of Displaced Lognormal . . . . . . . . . . . . . . . . . . . . . D. . . . . . .2 and Theorem 13–Strong Convergence of Meanreverting CEV Process . C. . . . . . . . . . . . C. . . . . . . . .2 Appendix: Proof of Theorem 15–First Order Approximation of Largestrike Largeexpiry Implied Volatility . . . . . . . . . . . . D. . . . . Slope and Convexity of DH Shortexpiry Implied Volatility . . . . . . . . . . . .2 A.7 Appendix: Proof of Theorem 21–Fixedstrike Largeexpiry Implied Volatility D. . . . . . . . B. . .2 and 3. . . . . . . . . . . . . . . . . . . . . . . .1 Appendix: Proof of Theorem 14 –Largestrike and Largeexpiry Asymptotic of Displaced Lognormal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . B. . . . . D. . . . . . . . . . . . . C D v . . . . . . . . . . . . . . . . D. . . . . . . . . . . . . .1 A. . .7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Appendix: Proof of Theorem 4–Global Behavior of Displaced anti Lognormal Appendix: Proof of Theorem 5–Atthemoney Behavior of Displaced anti Lognormal . . . .4 Appendix: Proof of Theorems 17 and 18– Second and Third Order Approximation of Largestrike Largeexpiry Implied Volatility . . . . . . . . . . .8 Appendix: Proof of Theorem 22–Approximation of FixedStrike Largeexpiry Implied Volatility. . . . . .2 Monotonicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1–Level. . . . . . . . . . . . . . D. . . . . . . B.2 Appendix: Proof of Theorems 8 and 9– Shortexpiry Behavior of DH .
and Exact σimp for T = 20. . . . . . .1 7.1. and Exact σimp . . . . and Exact σimp for T = 7. . . . . . . . . . . . . . . . . . . . . . .0. . . . .1. . . . . .2. . . . . . .2 7. . .1 Theorem 9 formula and exact σimp .1. . . . . . . . . . .4 7. . . . . . . .1. . . . . . . . 7. . . . . formula and Exact σimp . . . . . . 3. . . . . . . . . . . . .1 Theorem 14 formula Theorem 17 formula Theorems 17 and 18 Theorems 17 and 18 Theorem 20 formula Theorem 20 formula Theorem 22 formula and Exact σimp for T = 10. . . 8 11 19 51 53 58 58 59 59 60 vi . . . . . formula and Exact σimp . . . . . . . . . . . . . . . . . . .1 Theorem 3 formula and Exact σimp for T = 1. . . . .0.3. . . . . . . . . . .2.1.1. . .5 7.1 Theorem 5. . .3 7.6 7. . . . . . . . .1 bounds of σatm for T = 3. . . 2. . . . . . . and Exact σimp for T = 10. . . . . . . .LIST OF FIGURES 2. . . . .2. . .
. . .1 5. . . . . . . . . . . .4 5. . . . . .2. . . using LN as control for Asian Option on SABR . . .1. . . . . . . .1. . . . . . . . . . . . using DH for downandout call on SABR Percentage reduction of variance. . . using DL as control for Asian Option on SABR . 32 32 32 35 35 41 41 42 42 vii . . . Percentage reduction of variance. . . . . . . Percentage reduction of variance. . . .1 5. . . . . . . . . . . . . . . .1. . using DL for downandout call on CEV Percentage reduction of variance. . . . .3 5. . . . . . . . . using DL for downandout call on SABR Percentage reduction of variance. . . . . . . .2. .3 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . .2 5.5 5. Percentage reduction of variance. . . . . .LIST OF TABLES 5. . . .4 Percentage reduction of variance. . . . . using ISDL and ISDH for downandout call on SABR . . . . . .1. . . . . . . . . . using ISDL and ISDH for downandin call on SABR . Percentage reduction of variance. . . . . . . . . . . . . . . . Percentage reduction of variance. . using importance sampling for downandout call on SABR . . . . . . . . . . . . . . .2. . . . using importance sampling for downandin call on SABR . . . .2 5. . . . . . .1.
but its state space is bounded above and unbounded below. What’s more. We propose using the DL/DH dynamics as a control variate. Moreover.ABSTRACT We analyze the displaced (anti)lognormal (DL) and displaced (anti)Heston (DH) volatility skew. we show that the atthemoney slope has the same sign as the displacement. we give a strong convergency result for the meanreverting CEV process. we prove the global monotonicity of the implied volatility. which therefore cannot reproduce some features observed in the equity market. Under some regularity conditions. For the displaced Heston dynamics. for the displaced lognormal. viii . which allows direct calibration of their parameters to volatility skews implied by market data or by other models. the displaced antilognormal. to reduce variance in Monte Carlo simulations of the CEV and SABR local/stochastic volatility models. and an atthemoney bound on the steepness of the downward volatility skews. In particular. Numerical simulations show that signiﬁcant variance reduction can be achieved. We give simulation results to show that a carefully constructed control variate can signiﬁcantly reduce the variance in the Monte Carlo simulations. Finally we discuss the convergency of the discretisation schemes of the stochastic processes encountered in the Monte Carlo simulations. We prove the global monotonicity of its implied volatility too. we give a partial strong convergency result for the stochastic volatility process. We further propose a combination of the importance sampling and the control variate to reduce the variance. we give an explicit formula for the DL and DH’s shortexpiry limiting volatility skew. In the end. A variant. overcomes this steepness constraint. we analyze the largeexpiry limiting volatility of the displaced lognormal and give an asymptotic formula of it in the region of largestrike and ﬁxedstrike respectively.
Han. They are my source of advice and discussion. Changgee. In addition. Christina. I thank Professor Per Mykland for many inspiring discussion on topics of application of Statistics to Finance. Yibi. Han. Dale. Because of them. Winfried. goes my deepest appreciation. Also. for his constant support and encouragement over the past few years. Particularly. and for accompanying me through this work. Baoguan. I thank Nathaniel for proof reading my thesis. I would like to thank Xiaohui. Omar. Wenlong. support. stuﬀ and students of the Department of Statistics at the University of Chicago. Without him. I have a wonderful education here. Words cannot express the gratitude I owe them. Mike. William. for sharing with me his insights and wisdom. Zhuo. to whom this thesis is dedicated. for his understanding. Finally. this dissertation would not be possible. Oli. to my parents. Zuoheng. I am grateful to Professor Steven Lalley and Professor Per Mykland for being on my committee and for their valuable suggestions and comments. I would like to express my sincere thanks to all the faculty. Professor Roger Lee. I feel fortunate to study and pursue my doctoral degree in such an intellectually stimulating and friendly environment. ix . for his love and care. I thank them for their love. Yingying. for his instructive guidance and invaluable advice. I would also like to thank my friends who made my time so enjoyable that I will always cherish my time here. Minsun and Darongsae. care and support over the years. I am deeply grateful to my advisor.ACKNOWLEDGEMENTS First and foremost. I owe a special thanks to my dearest husband.
2.1.4. 2.2.2. and the parts related to the DL process in Chapter 5.PUBLICATION Part of the thesis has been published in Lee and Wang (2009).1. Chapter 4. 4.1.2. 2.1. 4.3. x . which include the following materials: Chapter 2.
for the purpose of approximating the features of a more complex model. Comparison can also occur. between volatility skews generated by two diﬀerent models. or for the purpose of understanding what empirical features can or cannot be reproduced by the model. and interest rate markets. as a pricing model or as a analytical approximation to other models. which does not agree with the sloping skews observed empirically in equity. In particular. The lognormal (BlackScholes 1973) model generates a ﬂat implied volatility skew. and SvobodaGreenwood (2009). we bound the level and slope of the implied volatility skews generated by displaced lognormal diﬀusions in various regimes (global. Regardless of what process actually underlies the given option price. Joshi and Rebonato (2003). Marris (1999). the volatility skew – meaning the implied volatility as a function of all strikes – captures the full riskneutral underlying distribution at that expiry. for the purpose of calibrating the model’s parameters. have investigated the displaced lognormal (and extensions thereof). and hence constitutes a natural framework to understand and to compare distributions. In contrast. Displacing the lognormal (Rubinstein 1983) does generate a sloping implied volatility skew. we draw motivation mainly from problems arising in equity markets. and we intend to use the calibrated process less for its analytical pricing than for its applicability to Monte Carlo pricing. such as how to calibrate to volatility skews that slope downward more steeply than all displaced lognormal skews. using a simpler model. or atthemoney.CHAPTER 1 INTRODUCTION Given an empiricallyobserved or modelgenerated price of a call or put. motivated largely by applications to interest rate derivatives. the implied volatility provides a canonical language or scale by which option prices are commonly quoted and compared. 1 . Brigo and Mercurio (2002). comparison can occur between an empiricallyobserved volatility skew and a modelgenerated volatility skew. FX. First. the implied volatility is by deﬁnition the volatility parameter for which the BlackScholes formula recovers the given option price. At any expiry. or shortexpiry).
In the Monte Carlo simulation. What’s more. We give numerical examples to show the signiﬁcance of the variance reduction. under displaced lognormal dynamics. A variant. particularly importance sampling. we then exploit the DL or DH. overcomes the steepness constraint. we ﬁnd an explicit formula between the shortexpiry implied volatilities of the two processes. practitioners have settled on a consensus of using a mixed stochastic/local volatility (SLV) model as the market standard for pricing barrier options. among other results.” CEV/SABR belong to the SLV family of models so pricing options under them is of practical importance. for any general process which is a positive martingale. Fisher and Tataru (2010) state that “During the past few years.We prove. but its state space is bounded above and unbounded below. we prove the global monotonicity of implied volatility. 2 . while the displaced anti Heston diﬀusion overcomes the skewness problem. not as a model. unlike stock prices. This helps to calibrate the model to the empirical data or other models. We show that the slope of its atthemoney implied volatility has the same sign as the displacement and its shortexpiry implied volatility lacks the skewness observed in the equity market. The Heston model (Heston 1993) generates the volatility smile and has an analytical solution for call and put options. we explore the combination of the control variate and other variance reduction technique. We therefore consider pricing barrier option and Asian option under the CEV/SABR model. For the displaced anti lognormal. such as the CEV and SABR local/stochastic volatility models. most of the stochastic processes can not be exactly sampled. Moreover. we ﬁnd an explicit formula for their shortexpiry limiting volatility skew. Thus we refer to discretisation schemes to conduct the Monte Carlo simulation. In light of these restrictions on what features the DL and the DH can model. and an atthemoney bound for the level and slope of the implied volatility. when the control variate on DL/DH is carefully constructed. We give numerical examples which illustrate signiﬁcant variance reductions. the displaced anti lognormal. but as a control variate. to reduce variance in Monte Carlo simulation of other models. We propose a displaced (anti)Heston diﬀusion and analyze its implied volatility. which therefore cannot model some features (nonmonotonicity and a steep downward slope) observed in equity market volatility skews. and its corresponding (anti)displaced process. and an atthemoney upper bound on the absolute slope of downward volatility skews. its state space is bounded above and unbounded below. For both the displaced (anti)lognormal (DL) and the displaced (anti)Heston (DH) diffusions. the global monotonicity of implied volatility.
Finally. Chapter 4 illustrates how to calibrate the parameters of the DL and DH processes. We give explicit approximation formula of the implied volatility in the largestrike largeexpiry case and the ﬁxedstrike largeexpiry case. Chapter 7 discusses the largeexpiry behavior of the implied volatility of displaced lognormal dynamics. Chapter 5 discusses how to use DL or DH processes to construct control variate to reduce variance in Monte Carlo simulations of option pricing. Chapter 3 discusses the features of the implied volatility of the DH diﬀusion. Chapter 6 discusses the discretisation schemes used in the Monte Carlo simulations. we analyze the largeexpiry implied volatility of displaced lognormal dynamics.We discuss the conditions which guarantee the convergence of the discretized process to the underlying continuous process. Chapter 2 discusses the features of the implied volatility of the DL diﬀusion. it also discusses how to combine the control variate with the importance sampling to achieve variance reduction. These convergence results serve as theoretical foundations for the Monte Carlo simulation. 3 . Appendix contains most of the proofs.
where R∗ := R\{0} and ∗ ∗ R+ := (0.3) where N denotes the standard normal CDF.1) (2. k. T ) := sN (d+ ) − kN (d− ) P BS (s.4) X Write σimp (K. X0 ). ·. σimp . T ) for this implied volatility. X We refer to the function σimp (·. deﬁne the implied volatility of X at (K. Our deﬁnition of the implied volatility skew will refer to the function C BS . and has range ((X0 − K)+ . because C BS (X0 . T ) := kN (−d− ) − sN (−d+ ) √ log(s/k) Σ T √ d± := ± 2 Σ T (2. K. Deﬁne C BS : R3 × R+ → R and P BS : R3 × R+ → R. (2. T ) to be the σimp > 0 such that C BS (X0 . speciﬁed as follows.2) (2. by C BS (s. Σ. or stipulate that all prices are quoted as forward prices. 4 . ∞).CHAPTER 2 DISPLACED LOGNORMAL PROCESS 2.1 Implied Volatility We work under martingale measure. T ) as the implied volatility skew of X at expiry T . T such that (X0 − K)+ < E(XT − K)+ < X0 .5) (2. Deﬁnition 1 (Implied volatility). and we either assume zero interest rates. For all positive K. which is welldeﬁned. K. k. Let X be a process with X0 > 0. T ) = E(XT − K)+ . We may suppress the last argument (T ) of C BS . Σ. T ) is strictly increasing on R+ .
σ. equation (2.1 Implied volatility Let us apply Deﬁnition 1 to the case that X = S.4) holds for all K. For displaced lognormal S. K − θ. σ.4) holds for all K.6) In general. then the ﬁrst inequality in (2. but the second inequality E(ST − K)+ < S0 may fail for small positive K. σ > 0. as we will see later. (2. K. with displacement θ ∈ R. payoﬀs invariant to parallel shifts of the S path and the contract parameters can be priced using BlackScholes model valuation methods. if dSt = σ(St − θ)dWt . T ). Thus S − θ is a driftless geometric Brownian motion with volatility σ. σ. ∞). K − θ. T positive.2. T ) < S0 } (2. (2. T positive.7) S0 > θ. let KS (T ) := {K > θ+ : C BS (S0 − θ. σimp . If modeling a nonnegative underlying such as a stock price. but applied to displaced arguments. a displaced lognormal. by monotonicity of C BS in K. If θ < 0. If θ ≥ 0 then (2. where W is Brownian motion.5) deﬁnes the implied volatility of S to be the σimp such that C BS (S0 . A process S follows displaced lognormal dynamics.9) 5 . We therefore take care to deﬁne the strike interval on which implied volatility exists. T ) = C BS (S0 − θ.8) which is a semiinﬁnite interval. this model for θ < 0 will misprice deepoutofthemoney puts. (2. even for atthemoney contracts. a Kstrike T expiry European call option on a displaced lognormal S has price E(ST − K)+ = E((ST − θ) − (K − θ))+ = C BS (S0 − θ. 2. This model has further limitations.2. K − θ. For K > θ. due to the possibility of ST < 0.2 Displaced Lognormal Deﬁnition 2. and the interval of points attainable by S is (θ. due to the nonzero probability that ST < 0. T ). For each T > 0 and each K ∈ KS (T ).
If θ < 0 then σimp (K) > σ. a volatility skew which slopes downward in the central portion of the strike range will usually still turn upward at suﬃciently large strikes. and bounded below by σ. Proof. and inf K∈K(T ) σimp (K) = σ if θ < 0. If θ > 0 then σimp is everywhere strictly increasing in K. T ) and KS (T ) notations for displaced lognormal S. Hence supK∈K(T ) σimp (K) = σ if θ > 0.2. Implied volatilities in the displaced lognormal model (2. the global bounds are also asymptotes. Appendix A. (2. Theorem 1 extends to all K ∈ K(T ).S To abbreviate the σimp (K. and possibly also the T argument. (Upper/lower bound). we will suppress the S superscript. 2. BrigoMercurio (2001) proves the K = S0 case of (2.1.10) 2. In both cases.11). 2. For all T > 0 and K ∈ K(T ). For all T > 0. Theorem 1 (Global behavior). 1.2 Global behavior Theorem 1 establishes the following global properties of σimp : If θ < 0 then σimp is everywhere strictly decreasing in K. Theorem 1 proves that the displaced lognormal cannot reproduce this empirical feature.10. 3.11) 6 . Empirical volatility skews are typically not monotonic over the entire range of strikes. and bounded above by σ. Remark 2.2.1. (Sharpness of bound).6) have the following global properties. ∂K (2. we have σimp (K) → σ as K → ∞. sgn ∂σimp (K) = sgn θ. (Monotonicity in strike). For all T > 0 and K ∈ K(T ): If θ > 0 then σimp (K) < σ.
then (1/2)eσatm T /8 < 0. If T > 0 and S0 ∈ K(T ) then θ σ S0 θ σ σatm ≤ 1 − S0 σatm ≥ 1 + if θ ≤ 0. the approximate1 3monthexpiry atthemoney slope ∂ log σimp /∂ log K is more negative than −1. We bound the level σatm and also the slope of σimp atthemoney.2. (Atthemoney level). 1. Empirically. Atthemoney implied volatilities in the displaced lognormal model (2. in S&P500 daily data from all dates (1996–2008) in the OptionMetrics database.2. Remark 2. Speciﬁcally. as computed by OptionMetrics.57.6) have the following properties.3 Atthemoney behavior Theorems 2 and 3 focus on two diﬀerent subsets of the (K.12) if θ ≥ 0. which contains volatility skews interpolated by OptionMetrics using kernel smoothing. 2 7 . We approximate the slope as log(σimp (K1 )/σimp (K0 ))/ log(K1 /K0 ).13) = ≤ 2 θ + S0 ∂ log K K=S 2 φ(σatm T /2) T σatm 0 Proof. equity volatility skews typically slope downward more steeply than −0. (2. if T > 0 and S0 ∈ K(T ).2. Theorem 2 examines the atthemoney strike K = S0 .55delta call. 1. If θ < 0 and T > 0 and S0 ∈ K(T ) then √ √ ∂ log σimp 1 2 1 θ N (σatm T /2) − N (σ T /2) √ √ < eσatm T /8 . and K0 is the strike of a 0. 2. then deﬁne the atthemoney implied volatility σatm (T ) := σimp (S0 . where K1 is the strike of a 0.29 on 90% of the days in the sample. however. which may be abbreviated as σatm . If T ≤ 1 and σatm ≤ 100%.2.2. (Atthemoney slope).57. T ) domain. Our source is the “Volatility Surface” data set. (2. Theorem 2 (Atthemoney behavior). By “slope” we always mean ∂ log σimp /∂ log K. Theorem 2 proves that the displaced lognormal cannot reproduce steepness of this magnitude. Indeed. Appendix A. T ).50delta call. the strikeelasticity of implied volatility.
and expresses the solution explicitly.2. σ(1−θ/S )=20% 18 0 0 17.0.5 θ=−50 19 18.14) σ(1 − θ/S ) 0 Proof. Figure 2.3.6).2.3. Remark 2. to address the case of T large).5 S =100. Appendix A. T ) even for some T not close to 0. at expiry T = 1. The K = S0 case is known (indeed Rebonato (2004) reﬁnes the K = S0 formula. in the displaced lognormal model (2.5 Theorem 3 formula 21 θ=25 20. The contribution of Theorem 3 is to ﬁnd and prove a shortexpiry σimp formula valid for every strike K > θ+ . 22 Exact σimp 21.14) provides.1: Theorem 3 formula and Exact σimp for T = 1.4 Shortexpiry behavior Theorem 3 takes the shortexpiry T ↓ 0 limit of the implied volatility skew. The formula in the righthand side of (2.0. moreover.5 60 70 80 90 100 110 K 120 130 140 150 160 Theorem 3.2.2.1 compares the Theorem 3 formula and the exact σimp (K.2. For all K > θ+ .5 Percentage points 20 19. T ). σ log(S0 /K) log((S0 − θ)/(K − θ)) if K = S0 if K = S0 . 8 . T →0 lim σimp (K. a remarkably accurate approximation to σimp (K. T ) = (2. Figure 2.
is C BS (θ − K. E(ST − K)+ = E(θ − K − (θ − ST ))+ = P BS (θ − S0 .15) speciﬁcation is said to be a DL process. Displaced antilognormal pricing calculations have the tractability of the displaced lognormal. −σ. K − θ. An equivalent way to express the result. on a displaced antilognormal S. Recognizing the similarities between the displaced lognormal and antilognormal. Here its ﬁrst three arguments are negative. does generate arbitrarily large downward slopes. the displaced lognormal produces downwardsloping implied volatility. Thus θ − S is a driftless geometric Brownian motion with volatility −σ > 0. T ).15) (2. without negative arguments.16) So we have formally the same C BS formula as in the displaced lognormal case. however. but not of steepness commensurate with typical equity options data – regardless of how large a negative value θ takes.3 Displaced antiLognormal The previous section’s results show that with θ < 0. 9 . T ) = C BS (S0 − θ. σ. T ). Deﬁnition 3. σ < 0. the following terminology groups them together: Deﬁnition 4 (DL). A related process. θ) to an empirically observed volatility skew. (2.6) or the displaced antilognormal (2. A process S follows displaced anti lognormal dynamics if dSt = σ(St − θ)dWt . A process S which satisﬁes either the displaced lognormal (2. and the interval of points attainable by S is (−∞. −σ. the σ log(S0 /K)/ log((S0 −θ)/(K−θ)) formula may still facilitate calibration of the displaced lognormal parameters (σ. 0 < S0 < θ. where W is a Brownian motion. θ − K.1). or to a modelgenerated volatility skew. For instance.For some expirations of moderate length. therefore. θ). 2. which presents no problem. θ − S0 . to price a T expiry call struck at K < θ. the C BS function is still welldeﬁned by (2.
Note that we have P BS (S0 . T ) notations for displaced antilognormal S.18) 2.3.4) holds and σimp (K. θ − K.4. σ) = P BS (θ − S0 .15) decreases monotonically in strike. That is. T ) := {K ∈ (0. To abbreviate the σimp (K. σimp ) = (S0 − K) + C BS (θ − S0 . T ) is thereby welldeﬁned.20) 10 . T ) and KADL (S. σ. T ). for all T > 0 and K ∈ KADL (T ). (2.19) (2. Theorem 4 (Global behavior). Appendix A. ∂K (2. σimp ) = E(K − S0 )+ = C BS (θ − S0 . Consequently.17) and possibly also the T argument. Implied volatilities in the antilognormal model (2.2 Global behavior For the implied volatility under the displaced anti lognormal. σ). −σ) > ((θ − K) − (θ − S0 ))+ = (S0 − K)+ . K. we have the global monotonicity behavior similar to the behavior under the displaced lognormal. T ) < S0 }. K − θ. This is given by Theorem 4. θ − K. (2. θ) : C BS (S0 − θ.3.1 Implied volatility Implied volatility for displaced anti lognormal S is deﬁned on the strike interval KADL (S. K. σimp ) = E(ST − K) + P BS (S0 . σ). θ − K. S S so (2. sgn Proof.2. we will suppress the S superscript. For K ∈ KADL (S. K. we have S0 > C BS (S0 − θ. ∂σimp (K) = − sgn θ. K − θ. C BS (S0 .
K = 100. when θ ≥ 2S0 .1: Theorem 5. Process parameters are S0 = 100. S0 θ σatm ≥ − 1 σ if θ ≥ 2S0 .2 0.2. 0. σatm (T ) is abbreviated as σatm . we could see that σatm < (θ/S0 − 1)σ. Theorem 5 (Atthemoney behavior). though when S0 < θ ≤ 2S0 .1 0 100 150 200 θ 250 300 350 Figure 2. T > 0 and S0 ∈ KADL (T ) then ∂ log σimp 2 1 θ θ ≤ ≤ eσatm T /8 .3.3 Atthemoney behavior Theorem 5 examines the atthemoney strike K = S0 .7 0.5 σatm 0.6 0. If θ > S0 . S0 (2. The solid blue line is σatm .22) .3 0. 2 θ − S0 ∂ log K K=S 2(θ − S0 ) 0 11 (2.0.4 0. 1.1 bounds of σatm for T = 3. the black dashed line is (θ/S0 − 1)σ and the red horizontal line is at the level of σ. We bound the level σatm and also the slope of σimp atthemoney.3. it looks like the two lines coincide with each other. T > 0 and S0 ∈ KADL (T ) then σatm ≤ θ − 1 σ if S0 < θ ≤ 2S0 . (Atthemoney slope). Atthemoney implied volatilities in the antilognormal model (2. If θ > S0 . σ = −0.3. we have σatm > (θ/S0 − 1)σ. As we can see from the ﬁgure.15) have the following properties.21) 2.8 σatm (θ/S0−1)σ σ 0. if we zoom in. (Atthemoney level).
When θ < 2S0 .3. θ) in the displaced antilognormal model (2. θ when θ > 2S0 . we have S0 < θ. we have σatm > S − 1 σ. This could 0 be explained by using the putcall parity as follows. θ = 2S0 is the critical case. σ) = C BS (S0 . Figure 2. hence the shortexpiry slope cannot be more negative than −1/2. and its derivative with respect to log K. For all K > θ+ in the displaced lognormal model (2.15). Remark 2. yield the shortexpiry limiting volatility skew’s level and slope σimp = T ↓0.23) If St follows the displaced anti lognormal dynamics with parameters (θ = 2S0 . σ).24) can produce arbitrarily steep negative slopes. (2. The ﬁgure also suggests that we will not have σ as the upper or lower bound for σimp as in the displaced lognormal case. σ) = P BS (S0 . Under displaced anti lognormal dynamics. we have σatm < θ S0 − 1 σ. the righthand side of (2. as well as for all K ∈ (0. we have (ST − K)+ = (K − (θ − ST ))+ . then using K = S0 = θ −K.Proof. σ). the conclusion (2.24) This holds under all DL dynamics. K. S0 ∂ log σimp θ = . 2. Appendix A.1 illustrates the Theorem 5. K=S 2(S0 − θ) 0 (2. The distinction is that under displaced lognormal dynamics.3. K=S0 σ(S0 − θ) . Appendix A. When K = S0 and θ = 2S0 . K. The Theorem 6 conclusion.4 Shortexpiry behavior Theorem 3 extends to all DL processes: Theorem 6 (Shortexpiry behavior). ∂ log K T ↓0.6).5. the equality happens when θ = 2S0 . hence (2. we have θ < S0 . Therefore the implied volatility of the process S is exactly σ.3.1. Proof.14) holds. 12 .1 bounds of σatm .3.23) is P BS (θ −S0 . θ −K.
(3. we will not have the global behavior of implied volatility. Wt∗ are i. T ) > 0 such that DH C BS (S0 . with displacement θ ∈ R. 13 . if modeling a nonnegative underlying process such as a stock price. T ) = E(ST − K)+ . (2.1 Displaced Heston Deﬁnition 5.i. 3.CHAPTER 3 DISPLACED HESTON PROCESS 3. due to the possibility of ST < 0.1. As with the displaced lognormal model.5) deﬁnes the implied volatility of the displaced DH Heston to be σimp (K. A process S follows displaced Heston dynamics. T ).1 Implied volatility Implied volatility for the displaced Heston S is deﬁned on the following strike interval: KDH (T ) := {K > θ+ : (S0 − K)+ < E(ST − K)+ < S0 } (3. ∞). dBt = ρdWt + 1 − ρ2 dWt∗ . We would like to characterize the implied volatility of displaced Heston dynamics.3) We will suppress the T argument of σimp (K. K. Thus S − θ follows the Heston model. Brownian motions and σt is nonnegative for all t > 0. (3. where Wt . if dSt = σt (St − θ)dWt . and the interval of points attainable by S is (θ. for θ < 0 this model will misprice deepoutofthemoney puts. We will have similar atthemoney behavior and shortexpiry behavior as in the DL case.2) For each T > 0 and each K ∈ KDH (T ).d. S0 > θ. T ). However.1) 2 2 dσt = κ(µ − σt )dt + σt dBt . σimp (K.
The slope of atthemoney implied volatility depends on the assumption that Wt . we are able to show that when ρ = 0. K.4) (3.2 Atthemoney behavior Although for the displaced Heston process. S0 > θ. (3.1.3.4) has the following property: DISV ∂σimp sgn Proof. where Wt . dσt = f (σt )dt + g(σt )dWt∗ . We will prove this for a more general DISV process. T ) = E(ST − K)+ . ∂K (K) K=S0 = sgn θ.1.1.5) Using (2. It may not hold when this condition is violated.6) Theorem 7. (DISV) A process S follows displaced independent stochastic volatility dynamics. Wt∗ are independent.i.1.5). the slope of the atthemoney implied volatility in the DISV model (3. T ). If T > 0 and S0 ∈ KDISV (T ). this feature still holds. we will not have the global behavior that the slope of implied volatility is determined by the sign of the displacement θ. σimp (K. the implied volatility of the DISV model σ DISV (K. Implied volatility for the DISV S is deﬁned on the following strike interval: KDISV (T ) := {K > θ+ : (S0 − K)+ < E(ST − K)+ < S0 }. (3. (3. Wt∗ are i. with displacement θ ∈ R. Appendix B. deﬁned as follows. Deﬁnition 6. for each T > 0 and each K ∈ KDISV (T ). T ).d. DISV We can suppress the T in σimp (K. Brownian Motion. 14 . if dSt = σt (St − θ)dWt .7) Remark 3. T ) is deﬁned such that DISV C BS (S0 .
σimp (K.1. Brownian motions and σt is nonnegative for all t > 0. the displaced Heston model belongs to the DISV family of models.1.5) deﬁnes the implied volatility of the displaced ˜ S. slope and convexity. (3. (3. Wt∗ are i.d. We will derive the shortexpiry implied volatility of the displaced Heston model using his results. T ) = E(ST − K)+ . 3. T ). (2. Particularly.H (3. If T > 0.i.H ˜ Heston to be σ (K. ˜ S. ∂K K=S0 (3. S0 ∈ KDH (T ) and ρ = 0.9).1. the slope of the atthemoney implied volatility in the displaced Heston model (3. dSt = σt St dWt . ˜ ˜ For St under the displaced Heston dynamics (3.Corollary 3. and we give an approximation of the shortexpiry implied volatility level. Shortexpiry behavior of Heston process Recall that the Heston process proposed by Heston (1993) as follows.10) For each T > 0 and each K ∈ KH (T ). With ρ = 0.3 Shortexpiry behavior Durrleman (2004) gives the shortexpiry implied volatility of the Heston model as a function of the strike.9) dBt = ρdWt + 1 − ρ2 dWt∗ . where Wt . we connect the shortexpiry implied volatility of the displaced Heston model with that of the Heston model. K. T ) > 0 such that imp ˜ ˜ ˜ ˜ ˜ C BS (S0 . ˜ The implied volatility of the Heston process St is deﬁned on the following strike interval: ˜ ˜ ˜ KH (T ) := {K > θ+ : (S0 − (K − θ))+ < E(ST − (K − θ))+ < S0 }.1) has the following property: sgn ∂σimp (K) = sgn θ.1). 2 2 dσt = κ(µ − σt )dt + σt dBt .11) 15 .8) Proof. then St follows the Heston dynamics (3. denote St = St − θ.
13) where a0 = − 2 2 ρ ρ 2 7ρ2 2 . Theorem 8. We will suppress the S and T argument of σimp (K. the shortexpiry relationship between the implied volatilities of 16 . The notation A(T ) ∼ B(T ) means that A(T )/B(T ) → 1 as T ↓ 0. Durrleman (2004) shows that. slope and convexity atthemoney. A call option on the displaced Heston process starting at S0 with strike K can be con˜ ˜ sidered as a call option on a Heston process starting at S0 with strike K: ˜ ˜ E(ST − K)+ = E((ST − θ) − (K − θ))+ = E(ST − K)+ .10) and deﬁne KDH∩H (T ) = KDH (T )∩KH (T ) := {K > θ+ : (S0 −K)+ < E(ST −K)+ < max(S0 . 2 2 (3. T ) ∼ c T 2 ˜ ˜ ˜ ˜ σ0 + a0 log(S0 /K) + b0 + 0 log2 (S0 /K). For K > θ+ . the σimp is welldeﬁned for both the Heston process and the displaced Heston process. b0 = κ(µ − σ0 ) + σ0 − (1 − ρ2 /4).2) and (3. c0 = 2 (1 − ). T ) as σimp (K). Proposition 3. T maturity and initial value S0 can be approximated as H ˜ σimp (K.H ˜ H ˜ ˜ ˜ where K = K − θ. so K ∈ KDH∩H (T ). Theorem 8 says that the shortexpiry implied volatility of the displaced Heston model can be expressed in terms of the shortexpiry implied volatility of the Heston model. (S0 −θ))} to be the domain of K.1.12) We take care here to deﬁne the domain of K to make sure that the implied volatilities of both the Heston model and the displaced Heston model exist.˜ S. 2 2 6 4 6σ0 Shortexpiry behavior of displaced Heston process When near the expiry. for the Heston model.2 gives the shortexpiry implied volatility’s level. the nearmoney shortexpiry implied volatility of the ˜ ˜ vanilla call with Kstrike. Combine (3. (3.
T ) ∂ log σimp 2 2 σ0 + κ(µ − σ0 ) 2T T − 2 12 × (1 − θ/S0 )2 + O(T 2 ). T ))2 = slope: K=S0 DH (K. ∂K 2 = K=S0 2(S0 − θ)S0 (3.2 says that with θ < 0. For the displaced Heston model (3. This means that the displaced 2 Heston is not a suitable model for the equity data. θ) = 4θ2 −5θS0 . lim σ H (K − θ. The shortexpiry atthemoney slope is 2(S θ−θ) . however. T )H(S0 . Appendix B. slope and convexity when T ↓ 0 are given as follows. T ) × imp T →0 DH lim σimp (K. DH level: (σimp (K. and is bounded by 1 .3. 17 . Proposition 3.14) if K = S0 .15) where c0 = 2 2 6σ0 and H(S0 . A related process.2. We have mentioned in Remark 2.1. for K > θ+ . 0 the atthemoney slope is positive. will generate arbitrary large downward slopes. the atthemoney slope is negative.1. θ) + O(T ).2. T ) H + σimp (S0 − θ. Remark 3. 2(S0 − θ) −1 H c0 σimp (S0 convexity: DH ∂ 2 σimp (K. suggesting that when θ > 0.2 suggests that the shortexpiry displaced Heston cannot reproduce steepness of this magnitude. Appendix B. the atthemoney implied volatility level. T ) − θ.the displaced Heston model (3. which is analogous to the displaced anti lognormal process.57. T ) × (1 − θ/S0 ) imp log(S0 /K) log((S0 − θ)/(K − θ)) if K = S0 (3.2. 3 6S0 (S0 −θ) Proof. 3. when θ < 0. Proof.1) and the Heston model (3.2 Displaced antiHeston Remark 3.1.2 that the empirical equity volatility 2 skews typically slope downward more steeply than −0. the steepness of the shortexpiry implied volatility from the displaced Heston has an upper bound of 1 atthemoney.1. Theorem 3.1) with ρ = 0. T ) = T →0 T →0 lim σ H (K − θ. ∂ log K = K=S0 θ + O(T ).9) is given as follows.2.
18) ¯ ¯ ¯ Denote S0 = θ − S0 .9). T ) = E(ST − K)+ .5) deﬁnes the implied volatility of the Heston to H be σimp (K. (3. K.16) speciﬁcation is said to be a DH process. the following terminology groups them together: Deﬁnition 8. 3. K = θ − K. Brownian motions and σt is nonnegative for all t > 0. K. (3. 0 < S0 < θ. and the interval of points attainable by S is (−∞. σimp (K. Thus θ − S follows the Heston dynamics with volatility process σt . Recognizing the similarities between the displaced Heston and antiHeston.i. then S is the Heston process (3.19) For each T > 0 and each K ∈ KAH (T ). T ). where Wt . T ) > 0 such that ¯ ¯ H ¯ ¯ ¯ C BS (S0 . Implied volatility for ¯ the Heston S is deﬁned on the following strike interval ¯ ¯ ¯ ¯ ¯ KAH (T ) := {0 < K < θ : (S0 − K)+ < E(ST − K)+ < S0 }.1) or the displaced antiHeston (3. (3. 18 (3. Wt∗ are i. θ). A process S follows displaced anti Heston if: dSt = (−σt )(St − θ)dWt .16) dBt = ρdWt + 1 − ρ2 dWt∗ .2. (DH) A process S which satisﬁes either the displaced Heston (3. T ) > 0 such that ADH C BS (S0 .17) For each T > 0 and each K ∈ KADH (T ).1 Implied volatility Implied volatility for the displaced anti Heston S is deﬁned on the following strike interval: KADH (T ) := {0 < K < θ : (S0 − K)+ < E(ST − K)+ < S0 }.d.5) deﬁnes the implied volatility of the displaced ADH anti Heston to be σimp (K. σimp (K.20) . T ) = E(ST − K)+ . 2 2 dσt = κ(µ − σt )dt + σt dBt .Deﬁnition 7. T ). (3. (2. (2.
θ = 600.215 0. T ) is approximated using (3.18 −0.1 −0.Again.195 0.04.0175.4.2 −0.05 0.225 0.25 −0. αS0 = 0. Parameters of the displaced anti Heston dynamics are S0 = 100.2 0. σ0 = 0. we deﬁne a domain of K as ADH H ¯ KADH∩AH (T ) = KADH (T ) ∩ KAH (T ) = {K : 0 < K < θ. ρ = 0. In the left plot. to make sure the implied volatilities of both the Heston model and the displaced anti Heston model exist.1.215 S=100.2 −0. Theorem 9 gives a relationship between the shortexpiry implied volatilities of the displaced anti Heston model and the Heston model. = 0. T = 4 month.21 0.225 0.22 T=4/12 0.1: Theorem 9 formula and exact σimp . the σimp is welldeﬁned for both the Heston process and the displaced anti Heston process.1 −0. T = 1 month and in the right plot.2.19 0. where σimp (θ − K.15 0.1 0.2.18). and σimp (K. µ = 0. σ0=0. 19 .2 Figure 3. T ) ∃. 0. 3.2.2.15 −0. κ = 0.21 S=100.13).205 0. Proposition 3. slope and convexity.25 −0.1 gives the shortexpiry atthemoney implied volatility’s level. β = 0. The solid green line is Theorem 9 formula. ν = 0. σimp (K.2 0.185 0.19 0. T ) ∃.205 Implied Volatility 0. so K ∈ KADH∩AH (T ).2 0.2. The dashed black line is the exact H σimp by (3.195 0.05 0.05 0 log(K/S) 0.18 −0. T ) as σimp (K). σ0=0.22 T=1/12 0. ρ = −0.15 −0. The details of the calibration are in section 4.2.23 Th−ImpVol ImpVol 0. T = 1 year.04 0. The DH model parameters are chosen to be close β to the ones calibrated from the SABR model using S = 100.} ADH ADH H H ¯ ¯ We can suppress the T in σimp (K.2 Shortexpiry behavior When near the expiry.05 0 log(K/S) 0.185 0.15 0.23 Th−ImpVol ImpVol 0. σimp (K.1 0.2.002.04 Implied Volatility 0. T ) as σimp (K).
2. T )H(S0 . for K ∈ (0.1 compares the Theorem 9 formula and the exact σimp (K.2 Remark 3. Figure 3. T ) ∂ log σimp 2 2 σ0 + κ(µ − σ0 ) 2T T − 2 12 × (1 − θ/S0 )2 + O(T 2 ).16) and the Heston model (3.21) if K = S0 .2. slope and convexity when T ↓ 0 are given in the follows. T ))2 = slope: K=S0 ADH (K. T ) × (1 − θ/S0 ) imp T →0 if K = S0 (3.22) 0 .2. T ) × imp log((S0 − θ)/(K − θ)) DH lim σimp (K.3. For K ∈ (0. Appendix B. T ). where c0 = 2 and H(S0 . the shortexpiry relationship between the implied volatilities of the displaced antiHeston model (3. Appendix B. Denote the process X to be a positive martingale. T ) even for T not close to 0. Proposition 3. θ) + O(T ).Theorem 9. a remarkably accurate approximation to σimp (K. θ).21) provides. For the displaced antiHeston (3.1. Proof. ∂ log K = K=S0 θ + O(T ).9) is given as follows.1. ADH level: (σimp (K. log(S0 /K) lim −σ H (θ − K. It suggests that the formula in the righthand side of (3. convexity: ADH ∂ 2 σimp (K. T ) ∂K 2 =− K=S0 2(S0 − θ)S0 (3. θ) = 3 6σ0 6S0 (S0 −θ) 2 4θ2 −5θS Proof.2.16) with ρ = 0.3 Generalization of shortexpiry behavior The relationship between the shortexpiry σimp of the Heston process and the DH process given in Theorems 8 and 9 can be generalized to a relationship between the shortexpiry σimp of any process which is a positive martingale and its corresponding (anti)displaced process. θ). 3. the atthemoney implied volatility level. 20 . T ) = T →0 T →0 lim −σ H (θ − K. 2(S0 − θ) −1 H c0 σimp (θ − S0 . at expiry T = 1/12 and T = 4/12. T ) H − σimp (θ − S0 .
log(S0 /K) lim σ X (K − θ. Deﬁnition 10. if θ − St = Xt . θ) and K ∈ KADX (T ). for K ∈ (0. T ).24) if K = S0 .5). A process S is said to be a displacedX process. 21 . For a process X which is a positive martingale. with displacement θ ∈ R. T ) = T →0 T →0 X (K − θ. Denote KADX (T ) to be the domain of K where X ADX σimp (K. T ) are welldeﬁned as T ↓ 0. for K > θ+ and K ∈ KDX (T ). X Denote σimp (K. lim −σ X (θ − K. lim −σ X (θ − K. T ) and σimp (K. as well as σimp (K. T ) and ADX σimp (K. Similar to the proof of Theorem 8. for any t ≥ 0. A process S is said to be an displacedantiX process. T ) and σimp (K. For a process X which is a positive martingale. T ) are welldeﬁned as T ↓ 0. Denote DX ADX σimp (K. T ) and σimp (K. T ) × imp log((S0 − θ)/(K − θ)) DX lim σimp (K. Theorems 10 and 11 give the X DX X shortexpiry relationship between σimp (K. the shortexpiry relationship between the implied volatilities of X and its corresponding displacedX process is given as follows. T ) = log(S0 /K) log((S0 − θ)/(K − θ)) if K = S0 (3.Deﬁnition 9. Proof. Similar to the proof of Theorem 9. Theorem 10. T ) to be the implied volatilities of the displacedX process and X the displacedantiX process. Denote KX (T ) to be the domain of K where σimp (K. T ) × (1 − θ/S ) lim σ 0 imp T →0 if K = S0 (3. T ) to be the implied volatility of the process X by (2. the shortexpiry relationship between the implied volatilities of X and its corresponding displacedantiX process is given as follows. for any t ≥ 0. Theorem 11.23) if K = S0 . T ). if St − θ = Xt . with displacement θ ∈ R. T ) × imp T →0 T →0 T →0 ADX lim σimp (K. T ) and DX σimp (K. T ) × (1 − θ/S0 ) imp Proof.
or as an approximation of another model. and σ = 0 in the singular case of Bachelier. In the case of slope b < −1/2. with S0 = S0 and parameters (θ. Explicitly. .24) is a shortexpiry limit.1. where σ = 0 in the case of DL. or from direct empirical measurement). We use Theorem 6 implications (2.CHAPTER 4 CALIBRATION OF DL AND DH PROCESS 4.1 to ﬁt the DH parameters to a given implied volatility’s level.1 Calibration of DL process Given a shortexpiry atthemoney skew level a and slope b (either from some model. and given an underlying level S0 . σ). Therefore (4. θ) and DH parameters (σ0 . the calibrated DL process is a displaced anti lognormal.2 and Proposition 3. Remark 4.1 Calibration of DL and DH Process Whether one chooses to use the DL/DH as a model. we ﬁnd σ = a(1 + 2b). θ= 2b S . slope and curvature.1) In the case of slope b > −1/2.24) match the given level a and slope b. θ) to volatility skews (a. 22 .1) may still facilitate calibration of (σ. The DL and Bachelier models belong to the family dSt = (σSt + A)dWt .1. in any case the DL parameters (σ. the calibrated DL process is a displaced lognormal.24) to ﬁt the DL parameters to a given implied volatility level and slope. The singular case of slope b = −1/2 can be matched by normal or “Bachelier” dynamics dSt = aS0 dWt . or (as we will) as a control variate for another model. there exists a DL process.1. ν. θ) require estimation/calibration. We use Proposition 3.2. Remark 2. 1 + 2b 0 (4. ρ.3 indicates its accuracy at T of moderate length. b) even at moderately long expiries.2. Although (2. κ. provided that b = −1/2.1. 4. such that the DL skew’s shortexpiry level and slope (2.
ν.1 CEV and SABR stochastic volatility models For many local or stochastic volatility models. Although (3. × 2(S0 − θ)S0 c − a(2b + 1)H(S0 . provided that b = −1/2.2 Calibration DL and DH to CEV/SABR 4. θ) to volatility skews (a. slope and convexity (3. ν. . slope b and convexity c.2. ρ. b. ν. and given an underlying level S0 . ρ. In the DL case. there exist explicit short maturity approximations of implied volatility. In the 23 . κ. In application. we take 2 the same approach as in the DL case. Two such models capable of generating realistically steep atthemoney implied volatility skews are the Constant Elasticity of Variance (CEV) model and the SABR model. µ = + σ0 1 + 2b 6κ 2 6σ0 (4.22).4. .1. θ) 1 If b > − 2 . there exists a DH process. we ﬁnd explicitly. or to calibrate DH parameters (σ0 . making it easy to calculate the implied volatility level a.2).1). κ) preﬁxed. we will also preﬁx κ. we have σ0 = a(2b + 1). For the singular case b = − 1 .14) and (3.1.22) match the given level a.1). we see that the θ is the same in both the DL and the DH case. 4. θ) via (4. we have σ = a(2b + 1). Remark 4.15) or (3.15) and (3. with S0 = S0 and parameters (σ0 .2) a(2b + 1) . κ. or from direct empirical measurement).2 Calibration of DH process Given a shortexpiry atthemoney skew level a. With (ρ. ρ. if b < − 1 . we calibrate 2 the parameters to the displaced anti Heston model. θ= = 2 2b 2 S0 . Therefore (4.2. and to calibrate DL parameters σ and θ via (4. slope b and convexity c (either from some model.2. Recall that we have restricted ρ = 0 in (3. such that the DH skew’s shortexpiry level. while in the DH case. c) even at moderately long expiries.21) are shortexpiry limit. θ).1 indicates their accuracy at T of moderate length. κ. . we calibrate the parameters to the displaced Heston model. (2004). Remark 3. Compared with the calibration of the DL parameters given in (4. σ0 = a(2b + 1). slope b and convexity c.2) may still facilitate calibration of (σ0 . such as in Lewis (2000) and Berestycki et al.
by Berestycki et al.3) if K = S0 (4.5) which can take arbitrarily large negative values.2. we have CEV ∂ log σimp ∂ log K = T ↓0. The CEV model can generate a steep downward implied volatility skew atthemoney. Taking ν = 0 in the SABR model reduces to the CEV case. and correlation ρ ∈ [−1.1. dSt = αSt dWt . Remark 4. Diﬀerentiating with respect to log K. 2002) generalizes the CEV. where β ≤ 1. 0 is exist boundary. for all K > 0. • β = 1/2: 0 is attainable and strong reﬂecting. dBt = ρdWt + β S0 > 0 α0 > 0 ∗ 1 − ρ2 dWt (4. and absorption is imposed at S = 0. T ) = S0 − K 1−β lim σimp β−1 T →0 αS 0 β S0 > 0. 1] between S and α: dSt = αt St dWt . by making the coeﬃcient α stochastic.4) if K = S0 . (2002) and Roper (2009). dαt = ναt dBt . α(1 − β) log(S0 /K) 1−β CEV (K. and absorption is imposed at S = 0. Indeed. The origin is a regular boundary point and is speciﬁed as a killing boundary by adjoining a killing boundary condition. 24 . The widelyused SABR model (Hagan et al. • β > 1/2: 0 is not attainable. 2 (4. with volatilityofvolatility ν ≥ 0. (4. Boundary Conditions for the CEV: • β < 1/2: 0 is attainable. K=S0 β−1 .CEV model (Cox 1996).6) where W and W∗ are independent Brownian motions.
where instantaneous volatility varies not just as a function of S.8) reﬂecting the contributions to the SABR volatility skew.2 Calibration of DL to CEV/SABR For the SABR process.9) The ν = 0 special case of (4. by an approach which chooses parameters such that the displaced lognormal instantaneous volatility approximates the CEV instantaneous volatility function S → αSβ . + β−1 2 2αS0 (4. not just from the functional relationship between price levels and volatility. K=S0 ρν β−1 + . in the CEV case. as expressed by ρ. SABR σimp (K. Their approach arrived at the same result (4.9) gives the DL parameters that match the CEV level and slope: σ = αβS0 β−1 . as expressed by β. 4. 25 .According to a shortexpiry approximation in Hagan et al. (2002. β−1 2 2αS0 (4. Marris (1999) and SvobodaGreenwood (2009) have already investigated displaced lognormal approximation. + ρν (4. θ = S0 (β − 1)/β. eq. in contrast to our approach which matches the implied volatility functions. 3. but also from the correlation between price increments and volatility.1a).8). and b is given by (4. A distinction is that our implied volatility approach is intended to apply moreover to models. a = α0 S0 σ= β−1 α0 βS0 β−1 . (4. so we have α0 S0 α0 βS0 β−1 β + ρν.2. such as α in the SABR case.10) In the CEV case.7) The approximation’s slope at K = S0 is SABR ∂ log σimp ∂ log K ≈ T ↓0. such as SABR. The implied volatility skew reﬂects the dependence of volatility on the S level together with the other stochastic factors in the model. T ) ≈ α0 S0 β−1 1+ β−1 ρν log(K/S0 ) .10) as our approach. θ = S0 − . but also other stochastic factors.
13) can also be used as an approximation for the shortexpiry implied volatility of SABR model (4.2. slope and convexity as follows. limT →0 b(T ) = β−1 + 2 as the level and slope given in (4.8).12) into (4. a(T ) = T 2 σ0 + b0 . 2αS0 the same 26 . we derive the approximation of the shortexpiry SABR implied volatility level. (4. 2 −2 a (σatm ) 0 b(T ) = − . ρν β−1 . 6 6 σ2 2 b0 = 0 (ν 2 (2 − 3ρ2 ) + 6νρβσ0 + (β − 1)2 σ0 ).11) 5 ν2 2 (4 − 3ρ2 ) + νρ(β − 1)σ0 + (β − 1)2 σ0 . 6 From this. limT →0 a(T ) = αS0 β−1 . The parameters of DH are chosen by plugging (4.6) with σ0 = α0 S0 c0 = β−1 . a0 = −(νρ + (β − 1)σ0 )σ0 .2).13) with K = S0 .12).4. 2 (σatm )−2 a0 ∂σatm a0 + c0 c(T ) = + (σatm )−1 . 2K ∂K 2K 2 (4.12) where σatm is approximated by (3.2. In (4.2.7) and (4.3 Calibration of DH to SABR Durrleman (2004) shows that (3. Remark 4.
Indeed. The displaced anti Heston overcomes the steepness constraint but its paths are unbounded from below. .1 Variance Reduction Using Control Variate Theorems 1 and 2 imply that the displaced lognormal is inconsistent with the steep downward slopes (Remark 2. Rather. overcomes the steepness constraint.1) where the simulations Y1 . The basic Monte Carlo estimator is the sample average 1 ˆ Ym C := M m (5. To improve accuracy. In the absence of analytical solutions.1. Proposition 3. The displaced anti lognormal. we propose the DL/DH to generate control variates to reduce variance in the Monte Carlo pricing of derivative contracts under commonlyused dynamics which do match the empirical atthemoney volatility skew.24). 27 . we do not generally advocate the DL/DH to model stock price processes. Let the contract’s payoﬀ Y be given by a speciﬁed function of the S path. consider the use of Monte Carlo simulation to estimate the price EY.1).2. or from a DH process calibrated by (4. are bounded above and unbounded below – the opposite of the behavior desirable in a model of stock prices. For these reasons.CHAPTER 5 VARIANCE REDUCTION IN MONTE CARLO SIMULATION 5. θ). such as the CEV or the SABR stochastic volatility model.2 implies that the displaced Heston is inconsistent with the steep downward slopes (Remark 3. but introduces other drawbacks: its paths. which take values in (−∞. .2). by (2.2) and nonmonotonicity (Remark 2. where the control comes from a DL process calibrated by (4. YM are iid as Y. let us apply the control variate technique. such as a discretelymonitored barrier option on the CEV/SABR process S. Such paradox applies to the DH process too.1.1) typical of stock market volatility skews. Suppose the modeler intends to price a derivative contract for which the desired model lacks analytical pricing formulas. suppose the underlying S dynamics follow some speciﬁcation that a modeler deems appropriate.2.2). . in the sense of reducing variance.
There exist, of course, other variance reduction methods, combinable with a DL/DH control variate. We do not investigate them in this section; rather we maintain focus on the DL/DH control, with the intent of illustrating how much variance reduction the DL/DH control brings by itself. In section 5.2 we will discuss combining the DL/DH control with other techniques.
5.1.1 DL or DH as a control variate
To make explanation clear, we ﬁrst focuses on using DL as control variate. Using DH as control variate follows the same spirit. The control variate estimator of EY, using a control Y , where Y has a known expectation C := EY and a known simulation methodology, is deﬁned by 1 ˆ Ym − βYm + βC , C cv := M m (5.2)
where the simulated pairs (Y1 , Y1 ), . . . , (YM , YM ) are iid as (Y, Y ). Good choices of Y have large correlation ρY,Y with Y, because increasing ρY,Y  decreases the estimator’s variance. Speciﬁcally, ˆ ˆ Var C cv = (1 − ρ2 ) Var C Y,Y (5.3)
for the optimal choice of the β coeﬃcient, namely β = Cov(Y, Y )/ Var(Y ), which may also be estimated by simulation. For further details see, for instance, Boyle et al. (1997). Because the payoﬀ Y is a speciﬁed function of the S path, we choose Y to be that same payoﬀ function applied to the S path, where S follows a DL process driven by a Brownian motion that also drives S. Aiming to produce high correlation ρY,Y , we choose the S process parameters by taking S0 = S0 and applying (4.1) to ﬁnd (σ, θ) such that the shortexpiry atthemoney volatility skews implied by S and by S agree in both level and slope. The suitability of the DL process S to serve in this role stems from a conﬂuence of ﬂexibility and tractability; the DL is potentially ﬂexible enough to generate signiﬁcant correlation between Y and Y (by linking the parameters of S and S, as discussed above), and yet potentially tractable enough to allow analytic evaluation of EY and unbiased simulation of Y , as discussed below. For shiftinvariant contracts (including barriers and lookbacks), exact evaluation of C = EY under DL dynamics is just as easy as under BlackScholes dynamics; more precisely, 28
if the contract’s payoﬀ is invariant to parallel shifts of the underlying price path and the contract parameters (such as strike and barrier level), then BlackScholes model valuation methods, applied to shifted arguments, produce the contract’s DL valuation. In the DL case, if, moreover, we can simulate the exact distribution of Y – which is often the case, because S is a transformed Gaussian – then Y can serve as a control variate that reduces variance without introducing any bias. When using DH as control variate, we have one more source of ﬂexibility, namely the stochastic volatility process. If Y depends on the path S where S depends on a stochastic volatility process Σ, S is driven by a Brownian motion W and Σ is driven by W and W∗ , where W and W∗ are independent, we choose Y to be the same payoﬀ function applied to the S path, where S follows a DH process driven by W, and the volatility process of S is driven by W∗ . We choose the S process parameters by taking S0 = S0 and applying (4.2) to ﬁnd (σ0 , κ, ν, ρ, , θ) such that the shortexpiry atthemoney volatility skews implied by S and by S agree in level, slope and convexity. The DH control variate has an advantage over the DL control variate in that the DH model is more ﬂexible and the S path tracks the S path more closely. Consequently, Y will have higher correlation with Y so we can achieve greater variance reduction. On the other hand, the option pricing formula on the DH model is more complicated than on the DL model.
5.1.2 Example I: Discretely sampled barrier option under CEV/SABR dynamics
To take a concrete example, consider a discretely sampled barrier option on S, which follows CEV (4.3) or SABR (4.6) dynamics. In particular, let the contract be a downandout call with expiry T , barrier H, strike K, sampling dates t1 < t2 < · · · < TN = T , and payoﬀ Y := (ST − K)+ 1(min Stn > H).
n
(5.4)
Analytical solutions exist for continuous barriers in the CEV model (DavydovLinetsky 2001), but not for discrete barriers, nor for the SABR model, so we turn to Monte Carlo simulation.
29
DL as control variate
To generate a control variate on DL, we apply the same payoﬀ function to a DL process, driven by the same Brownian motion W = W. More precisely, Y := (ST − K)+ 1(min Stn > H)
n
(5.5)
dSt = σ(St − θ)dWt , where S0 = S0 , and (σ, θ) are calibrated by (4.9) in the SABR case, or (4.10) in the CEV case. In the DL case, This Y is easily simulated without bias, and the value of C = EY can be computed by shifting any of the fast and exact (up to numerical truncation/quadrature error) solutions for discrete barrier option prices in the Gaussian framework, such as Broadie and Yamamoto (2005), or in the L´vy framework, such as Petrella and Kou (2004) or Feng e and Linetsky (2008). An alternative to (5.5) is to choose instead a continuouslymonitored control Y ∗ := (ST − K)+ 1( min St > H).
t∈[0,T ]
(5.6)
The control expectation EY ∗ has a simple exact formula, and the control Y ∗ can be simulated without bias, using Brownian bridge techniques of Beaglehole et al. (1997).
DH as control variate
To generate a control variate on DH, we apply the same payoﬀ function to a DH process, driven by the same Brownian motion W = W, W ∗ = W∗ . The DH process parameters (σ0 , κ, ν, ρ, , θ) are calibrated by (4.2) where a, b, c are determined by (4.12). More precisely,
1 if b > − 2 , the control variate is based on the displaced Heston model,
Y := (ST − K)+ 1(min Stn > H)
n
dSt = σt (St − θ)dWt ,
S0 > θ,
(5.7)
2 2 dσt = κ(ν − σt )dt + σt dWt∗ , dWt dWt∗ = 0
30
dWt dWt∗ = 0 In both (5. σ) are tuned to the SABR by (4.5). 0.25}. such as Lord et.If b < − 1 . T = 4/12. such as Griebsch and Wystup (2008). with α such that αS0 β−1 (5. We discuss this more in detail in Chapter 6. 5. We use the control (5. −1. In the CEV case we take β ∈ {−0.9). where (σ.2. H = 95.1. S0 = S0 and σt is nonnegative for all t > 0. where K = S0 = 100. our β is what they denote as β + 1. ρ).15.1 and 5.1. θ) are tuned to the CEV by (4.9) ∈ {0. with α0 such that α0 S0 β−1 = 0.4). with an array of choices for (ν. where (θ. In the DH case. The value of C = EY can be computed by shifting any of the fast solutions for discrete barrier option price under the Heston model. −1.3 Numerical results I: Discretely sampled barrier option under CEV/SABR dynamics DL as Control Variate Our experiments simulate the payoﬀ (5. (5.20. based on HirsaCourtadonMadan’s (2003) estimates of S&P500 CEV parameters. 0.8) 2 2 dσt = κ(ν − σt )dt + σt dWt∗ . 0 < S0 < θ.5}.5).5. al (2006). ˆ Var(C) (5. N = 84. the control variate is based on displaced anti Heston model: 2 Y := (ST − K)+ 1(min Stn > H) n dSt = (−σt )(St − θ)dWt .1. In the SABR case we take β = 0. We run 1000 paths.10) 31 . We use the control (5.8).2.10). Ten equal spaced points are sampled each day. the Y can be simulated from an eﬃcient discretisation scheme which guarantees strong convergence.0. Tables 5.2 report the estimated “percentage reduction of variance” ˆ Var(C cv ) 100% − .7) and (5.
where each Var is the scaled sample variance of the summands in (5.8) with (4.71% 96. 32 .1: Percentage reduction of variance.00% 98.10).17% 93.4) with (5.7) with (4.76% 99. Table 5.2).73% Payoﬀ: (5.Table 5.9).4) with (5. using DH as control variate for the downandout discrete barrier under the SABR process.25 0.03% 99.2: Percentage reduction of variance. Equivalently.60% 98.5) with (4. using DH for downandout call on SABR T = 4 months ρ ν 0. Control: (5.70% 0.95% 99. the percentage reduction of variance equals the “Rsquared” of an OLS regression of the CEV/SABR barrieroption payoﬀ Y on the DL control payoﬀ Y .07% 92.3: Percentage reduction of variance.99% 1.15 0. Control: (5.2.12).4) with (5.35% 0. using DL for downandout call on SABR T = 4 months ρ ν 0.8 99.6 99. Control: (5.89% 98.60 0.4 98.2) or (5.67% Payoﬀ: (5.00 99. b and c are calculated from (4.2).1. with α0 such that α0 S0 β−1 = 0.74% 96.20 0.4 99.9).9).2) respectively.60 0.1.9).2.99% 99. with an Table 5.98% 99.20 0.99% 99.99% 99.1.41% 0.50 99.8 99.76% Payoﬀ: (5. In (4.20 0.83% 97.40 0.1) and (5.93% 1.64% 98. The parameters of the SABR process are β = 0.73% 98.9).4) with parameters given by (5.50 99. a.6 98.42% 0.40 0. DH as Control Variate Our next experiments simulate the payoﬀ (5.5) with (4. using DL for downandout call on CEV T = 4 months β−1 β αS0 0.98% 99.
S. Here we consider a slightly diﬀerent payoﬀ which we know how to compute exactly under the DL process. strike K. we can see that using DH as control variate achieves more variance reduction. which follows the SABR dynamics (4. Numerical simulations suggest that other choices of κ give relatively same results.3.6). W = W. DL as control variate For the control variate. Remark 5. Let the contract be an Asian call option with expiry T .array of choices for (ν. ρ). As we noted before.3.1. In the reported Table 5. the reason is that the DH process is a twofactor stochastic dynamics which enable the process S to better track the original process. we could apply the same payoﬀ to the DL process.1. The results are shown in Table 5. Fusai and Meucci (2008). More 33 .11) There is no analytical solution for the arithmetic Asian option price under the SABR process.1.3 and any other reported results related to DH process in this paper. so we turn to Monte Carlo simulation. Comparing the results in Table 5. such as Milevsky and Posner (1998). Although there is no exact analytical solution for the value of the discrete arithmetic Asian option under the DL process. we arbitrarily chose κ = 0.1.4 Example II: Discretely sampled arithmetic Asian option under SABR dynamics The next example we consider is the discretely sampled arithmetic Asian option on S. 5. it could be computed by shifting the numerical solution for the discrete arithmetic Asian option under Geometric Brownian motion.2 and 5.2. sampling dates t1 < t2 < · · · < tN = T . We call the new payoﬀ discrete exponential Asian option.1.1. The DL process is driven by the same Brownian motion which drive the SABR process. and payoﬀ 1 N N Y := ( Si − K)+ i=1 (5.1.
9). α0 S0 β−1 = 0.13) with parameters (5. β = 0.13).2 (5. 34 . Table 5.4 and 5.1. Lognormal as control variate To make a comparison. More precisely.12) N N + i=1 log(θ − Si )) − K) dSt = σ(St − θ)dWt The value C ∗ = EY ∗ can be calculated analytically under the DL process. Ten equal spaced points are sampled each day. the control is.11). where K = S0 = 100. In (5.15) and an array of choices for (ν.14).14) 5.precisely. we also use the control variate under the lognormal(LN) process. The control variate is constructed by applying the discrete exponential Asian option to the lognormal process.1.12) where (θ.2.1. σ) are tuned to SABR by (4. ρ). which is σLN = α0 S0 β−1 . Y∗ 1 := (exp( N N log(Si )) − K)+ i=1 (5.5 Numerical results II: Discretely sampled arithmetic Asian option under SABR dynamics Our experiments simulate the payoﬀ (5.13) dSt = σLN St dWt The value C ∗ = EY ∗ can be calculated analytically. (θ + exp( 1 N ∗ := Y (θ − exp( 1 N i=1 log(Si − θ)) − K)+ if displaced lognormal is used if displaced anti lognormal is used (5. N = 84. T = 4/12. we choose σLN to be the shortexpiry atthemoney implied volatility of the SABR process. We run 1000 paths. driven by the same Brownian motion W = W. We use control (5. The results suggest that the DL control achieves signiﬁcantly larger variance reduction than the LN control. (5. We also report variance reduction results using control (5.5 report the estimated “percentage reduction of variance”.
61% 77. see Bolye.76% 79.20 0.43% 96.18% Payoﬀ: (5.60 0.74% 97. Table 5.9).1.11) with (5.2 Variance Reduction Combining Control Variate and Importance Sampling In this section. we study the issue of combining the importance sampling with the control variate.1.13) with (5. if we deﬁne ˜ dWt := dWt − at dt.24% 98. Exponential Asian option on DL control: (5.1 Importance sampling on options pricing Recall the SABR process in (4.8 78.4 75.12) with (4.6 99.06% 0.11) with (5. 5. Exponential Asian option on LN control: (5.8 99.2. and Wt is a drifted Brownian motion.36% 0. using LN as control for Asian Option on SABR T = 4 months ρ ν 0.15).40 0.69% 98.40 0.19% 0.6 76. 35 . Broadie and Glasserman (1997) for more details.5: Percentage reduction of variance. Wt are ˜ independent standard Brownian motions.14).83% 99.57% 98. Under P.Table 5. ∗ ˜ ˜ then by the Girsanov Theorem.21% Payoﬀ: (5. 5.95% 0.4: Percentage reduction of variance.96% 80. using DL as control for Asian Option on SABR T = 4 months ρ ν 0. there is a measure P such that under which Wt .60 0.6).34% 78.20 0. Importance sampling itself is a popular variance reduction technique.97% 78.4 99.15).61% 76.
the likelihood ratio or RadonNikodym derivative be˜ tween the original measure P and the new measure P is r(S) := dP = exp ˜ dP T − 0 as dWs + 1 T 2 a ds . the importance sampling estimator (ISestimator) is: 1 ˆ r(Sm )G(Sm ). which is exp( 0 a2 ds). C IS := M m r(Sm )G(Sm ) can be considered as a weighted payoﬀ. Next. (5. Therefore. The intuition behind this is to make the path less(more) likely to reach the barrier in the downandout(downandin) case.19) a0 = α0 S0 β−1 . where C is the domain of S. We will arbitrarily set a constant Ca < ∞ and set a0 = min(Ca .16) dαt = ναt dBt .20) T We would like to control the variance of the likelihood ratio. the Girsanov Theorem says ˜ EG(S) = Er(S)G(S). so that s it does not explode in real applications. for the ¯ ¯ downandin call options. The drift at is chosen in a heuristic way similar to Bolye et al (1997). ˜ dBt = ρdWt + α0 > 0 ∗ 1 − ρ2 dWt Again by the Girsanov Theorem. we choose at = −¯0 before the barrier is reached and at = a0 after a ¯ the barrier is reached.the original process S becomes a drifted SABR process: β ˜ dSt = αt St (dWt + at dt). for the downandout call options. (5. we choose at = a0 . S0 > 0 (5. a0 ).18) ˜ ˜ where E is the expectation which is taken under P.) : C → R.17) For any function G(. 2 0 s (5. 36 . Denote 2  log(S/K) +  log(K/H) /T + 1 α0 S0 2 2β−2 (5.
2 Drifted DH/DL process The drifted DH/DL dynamics is derived similarly as the drifted SABR dynamics. 2 2 dσt = κ(ν − σt )dt + σt dBt .22) S0 > θ (5. 0 < S0 < θ (5. 2 2 dσt = κ(ν − σt )dt + σt dBt .1) and (3.3 Combine control variate with importance sampling ˆ ˆ Since the ISestimator C IS has smaller variance than the naive estimator C. The drifted displaced antiHeston dynamics is ˜ dSt = (−σt )(St − θ)(dW1t + at dt). then under the new measure P. and denote ˆ ˆ ˆ it as C ISCV . ˜ Under P. σ < 0. Consequently. the displaced lognormal dynamics becomes the drifted displaced lognormal dynamics: ˜ dSt = σ(St − θ)(dWt + at dt).2. These drifted processes are listed below. (5. We call the new estimator ISCVestimator. The drifted displaced antilognormal dynamics is ˜ dSt = σ(St − θ)(dWt + at dt).5. we can construct a control variate for the ISestimator. The drifted displaced Heston dynamics is ˜ dSt = σt (St − θ)(dW1t + at dt). the DL dynamics deﬁned in (2. 37 . DH dynamics deﬁned in (3.2. 5.21) 0 < S0 < θ. σt is nonnegative for all t > 0. C ISCV will reduce variance further than C IS .16) become drifted DL/DH processes under the new ˜ measure P.15). S0 > θ. σ > 0. In both the drifted displaced Heston model and the drifted displaced anti Heston model.23) dBt = ρdWt + 1 − ρ2 dWt∗ .24) dBt = ρdWt + 1 − ρ2 dWt∗ .6) and (2. Recall that ˜ ˜ ˜ if we deﬁne dWt := dWt − at dt. Wt and W∗ are independent t standard Brownian motions. (5.
and Si is sampled from any of the drifted DL/DH process (5.4 Example: Discretely sampled barrier option under SABR dynamics We consider a discretely sampled barrier option on S. This estimator coincides with the estimator proposed by Hesterberg (1996). Our approach is to ﬁnd a control variate for the importance sampling estimator. (5.The detail of the ISCVestimator is as follows.) onto the drifted DL/DH process S: Y := H(S) = r(S)G(S).21)(5. First.25) where S is sampled from (5. C ISCV = M m M m (5. In particular. let the contract be a downandin call with expiry T .24). which follows the SABR (4. considering the weighted payoﬀ in (5. (5. both the importance sampling estimator and the control variate estimator are regarded as weighted sums and are combined together as a double weighted sum. As in section 5.1. strike K. Y is constructed by applying the new payoﬀ function H(.16). 38 .2.6) dynamics. In Hesterberg’s approach. ˜ EG(S) = Er(S)G(S). we propose the combined estimator of the importance sampling and the control variate as follows. 5. we want to choose a control variate Y that is highly correlated with Y.1. Remark 5. barrier H. but comes from a diﬀerent angle.19) as one payoﬀ function on S: Y = H(S) := r(S)G(S). 1 1 ˆ r(Sm )G(Sm ) − β r(Sm )G(Sm ) − EG(S) .2.2.26) It turns out that H(S) = r(S)G(S) can also be considered as an ISestimator for the nondrifted DL/DH process S.28) where Si is sampled from (5.27) Finally. because by the Girsanov Theorem. (5.16).
we take Wi ∼ N (ai ∆t.1. If using DH.d. and payoﬀ: G(S) := (ST − K)+ 1( min Stn < H). i = 1. The discretisation of the SABR process is: α2 2β−2 β−1 ∆t + αi Si Wi log Si+1 = log Si − i Si 2 ν2 log σi+1 = log σi − ∆t + ν(ρ(Wi − E(Wi )) + 2 (5. κ.) to it: G(S) := (ST − K)+ 1( min Stn < H). ν. (σ0 . We discretize the SABR process into I subintervals and sample discretely. ∆t). · · · . Wi ∼ i.9). To construct ISCV with DL control.29) ˆ To construct the C ISCV . we sample from the drifted DL process and apply the payoﬀ G(. ∆t).sampling dates t1 < t2 < · · · < tN = T . so discretisation schemes are used. We are not showing the degenerate case where b = 1 and Si is sampled from a geometric Brownian motion. if b > − 1≤n≤N σ2 (5. if b < − 2 2 log(Si+1 − θ) = log(Si − θ) − ∆t + σWi . The likelihoodratio is calculated as I r(S) = i=1 f (Wi ) = exp(− fα (Wi ) I ai Wi + i=1 1 2 I a2 ∆t). N (0. 2 39 . θ) are tuned by (4. . i i=1 (5. W = W ∼ N (ai ∆t. c are determined by (4. ∆t) and fα (. in the DL case. where S is driven by the Brownian motion W = W. In practice.) to a drifted DL/DH process S.30) ∗ 1 − ρ2 Wi ) ∗ where Wi .31) where f (. ai is chosen by the rule at the end of section 5. in the DH case. ∆t).) is the density function of N (0. we apply the payoﬀ function G(. the volatility of S is driven by W ∗ = W∗ . b.2.30). there is no way of exactly sampling from the continuous SABR process and the correlated DH process. 1≤n≤N (5. ∆t) in (5.2) where a.12).) is the density function of N (ai ∆t. (σ. 1 2 2 2 σ 1 log(θ − Si+1 ) = log(θ − Si ) − ∆t + σWi .i. ρ. To sample from the drifted SABR process. θ) are tuned by (4.32) where S0 = S0 . I.
where S0 = S0 . we compare the variance reduction of the importance sampling.35) 5. The volatility process of DH is a meanreverting square process.31): I r(S) = exp(− i=1 1 ai Wi + 2 I a2 ∆t) i i=1 (5. ∆t).33) as in (5. the ISCVestimator is constructed as 1 ˆ C ISCV = M m G(Sm )r(Sm ) − βG(Sm )r(Sm ) + βEG(Sm ) . W = W ∼ N (ai ∆t. 40 . 1≤n≤N 2 σi ∆t + σ if b > − (5.34) Finally. and σt is nonnegative for all t > 0. We choose the ﬁrst order Euler discretisation scheme with reﬂection principal proposed by Higham and Mao (2005). to construct ISCV with DH control. More discussion of this is in Chapter 6.2. if b < − 2 2 i Wi . (5.) to it: G(S) := (ST − K)+ 1( min Stn < H).35).2.35). where the partial strong convergence of Si is ensured. ∆t). W ∗ = W∗ ∼ N (0.32) and (5. Remark 5. log(Si+1 − θ) = log(Si − θ) − 1 2 2 2 σi 1 log(θ − Si+1 ) = log(θ − Si ) − ∆t − σi Wi .Similarly.33) 2 2 2 σi+1 = σi + κ(µ − σi )∆t + 2 σi (ρ(Wi − E(Wi )) + 1 − ρ2 Wi∗ ). The likelihood ratio in both (5.33) is the same as in (5. with that of combining importance sampling with the control variate.2. we sample from the drifted DH process and apply the same payoﬀ G(. There are various discretisation schemes.32) as in (5.5 Numerical results: Discretely sampled barrier option In this section. Denote ISDH to be the estimator which combines the importance sampling and the DH control variate from (5. see Lord et al (2008) for more detail. Denote ISDL to be the estimator which combines the importance sampling and DL control variate from (5.
55% 89.34% 96. Table 5.2 show the variance reduction of the downandout call options using importance sampling.4 98.4 show the variance reduction of the downandin call options using the three estimators.8 97.29). Parameters of DH: (4.2.20 0.9).8 68.2.87% 97.2. T = 4/12.33) and (5.20 0. 41 .55% 96. ISDH: Combined importance sampling and DH control variate: (5.4) with (5.77% 67.60 0. Table 5.1: Percentage reduction of variance.40 0.3 and 5.02% 89. using importance sampling for downandout call on SABR T = 4 months.20% 0. Table 5.6 97.36) We simulate 1000 pathes.2.71% 0.45% Payoﬀ: (5.31% 65.2: Percentage reduction of variance.31% 0. ISDL T = 4 months. preﬁx ρ = 0.6 97. κ = 0. Parameters of DL: (4.2.91% 97.91% 85. N = 84. ISDL and ISDH.4) with (5.2. α0 S0 β−1 = 0.48% 94. ISDH ρ ν ρ ν 0.07% 67.60 0.30% 0.26% 0. β = 0. where K = S0 = 100.87% 0.36).4 97.70% 66.32% 90.46% 68.16% 98.4 69.69% Payoﬀ: (5.69% 95.60 0.99% 83. H = 95. For the DH parameters in ISDH.6 69.32% 67. Importance Sampling ρ ν 0.20 0.4) and (5.36).32) and (5. ISDL: Combined importance sampling and DL control variate: (5.2.2).35). Table 5.2.2. we see that ISDL and ISDH achieve more variance reduction than importance sampling.00% 0. using ISDL and ISDH for downandout call on SABR T = 4 months. For both options.40 0.The experiments simulate the payoﬀs (5. (5.35).8 97.40 0.1 and 5.
49% 0. ISDL: Combined importance sampling and DL control variate: (5.4.4: Percentage reduction of variance.2. we show that DL/DH in concert with importance sampling is superior to the importance sampling alone.3 with 5.85% 33.95% 33.97% 0.96% 0.20 0.06% 87. 42 . ISDH: Combined importance sampling and DH control variate: (5.4 97.29) with (5.4 33. using importance sampling for downandin call on SABR T = 4 months. so pricing barrier options under them is of practical interest.1 with 5. As mentioned by Fisher and Tataru (2010). This section uses DL and DH as control variate to reduce variance in the Monte Carlo simulation of the barrier options.77% 93.40 0.32) and (5. The CEV/SABR model belongs to the family of SLV models.89% 98.36% 97.2.1. Parameters of DL: (4.4 96.8 97. using ISDL and ISDH for downandin call on SABR T = 4 months.59% 79.6 35.01% 94.8 96.50% 0. as shown in Tables 5.36). ISDL T = 4 months.80% 82.1. More speciﬁcally. the stochastic/local volatility (SLV) model has become market standard for barrier option pricing.13% 0.2).2. Moreover.3: Percentage reduction of variance. We summarize the discussion of this Chapter here.60 0. Importance Sampling ρ ν 0. we show that the DL controls also yield signiﬁcant variance reduction for Asian options.35).2 and 5.6 97.09% 30.20 0.58% 88. Finally.6 96.2.00% 94.68% 0.63% Payoﬀ: (5. Table 5.Table 5.2.3. by comparing Table 5.60 0. we give numerical examples to demonstrate that the DL and DH controls can provide signiﬁcant variance reduction for barrier option pricing under the CEV/SABR model.40 0.60 0.71% 32.40 0. as well as Table 5. 5.8 35.75% 95.1.25% 32. Parameters of DH: (4.33) and (5.1.46% 0.35).06% 94.20 0.10% Payoﬀ: (5.35% 32.29) with (5. ISDH ρ ν ρ ν 0.2.9).42% 87.36).2.
it can be shown that the Euler scheme converges to the true process as the time intervals are made ﬁner and ﬁner. the LogEuler scheme is used. St )dWt . and simulates the process at the discretisation points. Zhang et al (2004) and Higham and Mao (2005). exact simulation methods do not exist and we have to refer to discretisation schemes to simulate them. ∆t Most existing proofs of the convergence of St to St rely on the linear growth and global Lipschitz conditions (Kloeden and Platen (1992)): (linear growth) there exists a positive constant L1 such that B(t. see Lord et al (2008). (6.2.3) (6. where one simulates from the discretisation of the logarithm of the original process. Kahl (2004). Denote the discrete approximation of the process as St . Sometimes. Consider a stochastic process: dSt = A(t. There are three basic criteria in the discretisation schemes: rate of convergence. St )dt + B(t. (global Lipschitz ) there exists a positive constant L2 such that B(t. x) − B(t. stability and positivity.CHAPTER 6 DISCRETISATION SCHEME For most general stochastic processes. y)2 ≤ L2 x − y2 . For more detail.2) 43 . wherein one discretises the time interval of interest. in order to reserve the positivity of the underlying process. Under certain conditions. EulerMaruyama (Euler for short) scheme is a straightforward discretisation scheme.5 and this chapter we will validate their usage. x)2 ≤ L1 (1 + x2 ). These criteria have been widely and intensively discussed in the literature. (6. There are many numerical schemes.1) ∆t where Wt is Brownian motion. We use some discretisation schemes in section 5.
t ). for example.However. Since the convergence they prove is up to a stopping time.t = Wi.t . the Euler scheme as: ∆t ∆t ∆t σt+∆t = σt + C(σt )∆W2. 2.). St ..5) Approximate St . i = 1. (6. σt )dt + B(t.1). The discretisation scheme is given as ∆t ∆t ∆t ∆t ∆t ∆t St+∆t = St + A(t. β 6. they show that its Euler discretisation converges to the underlying continuous process in L2 sense before a stopping time. and N ∆t = T . σt )∆W1. St . where A(. then the LogEuler discretisation converges to the logarithm of the underlying process. if St satisﬁes a local Lipschitz condition before the stopping time. We do not specify the discretisation scheme of σt here. the strong convergence rule shown in Kloeden and Platen (1992) can not be applied since St does not satisfy the global Lipschitz condition.t . The result can be extended to the LogEuler discretisation: if the logarithm of the underlying process satisﬁes the local Lipschitz condition before the stopping time.t . Remark 6.1. This result serves as the theoretical foundation of the Monte Carlo simulation of the CEV/SABR process. ∆W2.. We extend the result further to a stochastic volatility case. we refer to it as partial strong convergence.6) where ∆Wi. It could be any reasonable discretisation scheme.4) (6.1. σt using Euler scheme on time points tn = n∆t.1 Partial Strong Convergency of Stochastic Volatility Process Consider the stochastic process dSt = A(t. dW1t dW2t = ρdt (6. B(.) could depend on a stochastic volatility process σt . ∆t. σt )dW1t . (6.7) 44 .N. .t+∆t − Wi. σt )∆t + B(t. St . for the CEV/SABR processes. ∆t ∆t σt+∆t = f (σt . Zhang et al (2004) relax the condition of the convergence to local Lipschitz. dσt = C(σt )dW2t . St . For the process (6. where n = 1.
or the LogEuler scheme as: ∆t ∆t σt+∆t = σt exp ∆t C(σt ) ∆t 1 C(σt ) 2 ∆W2. z)2 ∨ B(t.10) Theorem 12 says that under the local Lipschitz condition (*) and some other regularity ∆t conditions. x.6). Ω2 be compact sets. x. Ω2 be bounded regions and denote the stopping time ∆t / τ = inf{t ≥ 0 : St ∈ Ω1 or St ∈ Ω1 or σt ∈ Ω2 }. z2 )2 ∨B(t. σt ) satisfy the local Lipschitz condition (*). St . ∆t ∆t 2 σt σt (6. z1 . ii.12) 45 . such that for any x1 . (6. / / For discretisation scheme in (6. A(t. z2 )2 ≤ K1 (Ω1 ×Ω2 )x1 −x2 2 +K2 (Ω1 ×Ω2 )z1 −z2 2 . (6. x2 . tn+1 ). A(t. A(t. such that for all x ∈ Ω1 . x1 . (6. Theorem 12. σt ) and B(t. the discretized process St will have L2 convergency to the true process S before a stopping time. the σt can be simulated exactly using the LogEuler scheme. there is a constant D(Ω2 ) such that ∆t E(σt − σtn )2 ≤ D(Ω2 )∆t. z ∈ Ω2 .8) For the SABR process. z1 )−A(t. such that E sup 0≤t≤τ ∧T ∆t St − St 2 ≤ C(Ω1 × Ω2 )∆t. then for the bounded Ω1 . (6. x2 ∈ Ω1 .11) then for ∆tT < 1. Deﬁne the local Lipschitz condition (*) as follows: let Ω1 . Ω2 . St . z)2 ≤ K3 (Ω1 × Ω2 ). Let Ω1 . z2 ∈ Ω2 .9) If local Lipschitz condition (*) holds. z1 )−B(t.t − ∆t . there exists a positive constant K3 (Ω1 × Ω2 ). if the following conditions are satisﬁed: i. for ∆t small enough and t ∈ [tn . ∃ constant C(Ω1 × Ω2 ). x1 . x2 . there exist positive constants K1 (Ω1 ×Ω2 ) and K2 (Ω1 ×Ω2 ).
1. substitute σt with σtn . 1). we have A(t. This holds for the CEV process too since the CEV process is a special case of the SABR process. Since σt = σtn e∆sνX− 2 ∆sν where ∆s = t − tn and ∆t X ∼ N (0.13) 46 . Therefore σt . St . as long as St and St remain in the domain Ω1 and σt remains in the ∆t domain Ω2 . a2 ] such that with large probability. we have A(t. 0 is an attainable boundary for 0 < β < 2 and for β = 1 if λµ < σ . 2 2 2 Since ∆s < ∆t. 2 1 1.∆t In other words. we could construct an interval [a1 . logarithm of the SABR process. σt ) = 1 σt St t t t t 2 In both cases. Theorem 12 indicates that the Euler (LogEuler) discretisation scheme of the SABR process (with σt simulated exactly) will converge to the true process on the time ∆t / interval τ ∨ T . St . If σt follows the geometric Brownian motion such that σt = νσt dWt .2.1. S . Remark 6. 2 2 β (6. 6. (see appendix C. In terms of our Monte Carlo simulation of the CEV/SABR process. the simulated path will be within the range. the Euler scheme St converges to the true process St as ∆t → 0.3. It is deﬁned as follows: dVt = λ(µ − Vt )dt + σVt dWt . In the 2 2β−2 and B(t. a2 ] or σt ∈ [b1 . In the SABR process.2 Strong Convergence of Meanreverting CEV Process Kahl and Jackel (2006) consider a stochastic volatility process which they call it meanreverting CEV process.) So. b2 ]. if St starts not close to 0. there is a constant K such that ∆t etn ν (e∆sν − 1) < K∆t. σt satisfy the condition (ii) in Theorem 12. 0 < b1 < b2 }. 2 2 Remark 6. then 1 2 ∆t σtn can be simulated exactly as σtn .1. They point out that the boundary behavior of the process as follows. σt ) = σt St . σ ) = σ S β−1 . σt ) = 0 and B(t. and we have ∆t E(σt − σtn )2 = E(σtn )2 (e∆sν − 1) = etn ν (e∆sν − 1).2 β for detail. for any tn < T and small enough ∆t. 0 < a1 < / / a2 . A(. where τ = inf{t ≥ 0 : St ∈ [a1 . Proof. and we could consider the discretized path as a good approximation of the true continuous path. St . a2 ] or St ∈ [a1 . Appendix C.) and B(.) satisfy the local Lipschitz condition (*).
16) is bounded.t . . with n = 1. There is a lot literature on the particular case of the meanreverting CEV process. Proposition 6. ∆t Vt+∆t = Vt∆t + λ(µ − Vt∆t )∆t + σ They prove that such discretized process S ∆t will converge to the true process S. ∆Wt = Wt+∆t − Wt .. Condition B : 1 ≤ β < 1. the 2 volatility process is the meanreverting squareroot process.2 says that the second moment of the discretized process (6.15) Vt∆t ∆W2. 2 We consider the discretisation scheme with reﬂection principle on the time points tn = n∆t. when β = 1 . N and N ∆t = T as follows: ∆t Vt+∆t = Vt∆t + λ(µ − Vt∆t )∆t + σVt∆t β ∆Wt .2.13) as ∆t → 0 in L1 sense. They also propose a discretisation scheme for the Heston model as follows: ∆t St+∆t = ∆t Vt∆t St ∆W1.. ∆W2. In the Heston model. We extend the strong convergence to the general meanreverting CEV process under the following conditions: Condition A: 0 < λ∆t < 2 and (1 − λ∆t)2 + βσ 2 ∆t < 1.t . ∞ is unattainable for all β > 0.1 gives the ﬁrst moment of the discretized process (6.16). Theorem 13 says that the discretisation scheme (6.16) Proposition 6. ∆Wt = Wt+∆t − Wt .16) converges to the true process (6.t .. corr(∆W1. which is called the meanreverting squareroot process.t ) = ρ.2. (6. This result serves as the theoretical foundation for our Monte Carlo simulation of the DH process. Higham and Mao (2005) propose a discretisation scheme called the reﬂection Euler scheme: ∆t Vt+∆t = Vt∆t + λ(µ − Vt∆t )∆t + σ Vt∆t ∆Wt . For the meanreverting squareroot process. 2 3. 47 . (6. 0 is unattainable for β > 1 .2. They also develop a numerical scheme to simulate the stock process whose volatility process is the meanreverting CEV process.14) They prove that this discretisation scheme has the strong convergency property. (6.
take the expectation of both sides. For SDE in (6. the ﬁrst order LogEuler scheme is applied to simulate the DL process and the CEV/SABR process. The discretisation schemes for each process are summarized below. Based on the discussion in section 6. 6. E(Vn )2 ∆t and E(Vn )2β is bounded for all n. Appendix C. For the discretized process (6. (6. ∆t ∆t EVn+1 = EVn (1 − λ∆t) + λ∆tµ ∆t ∆t ⇒ EVn+1 − µ = (1 − λ∆t)(EVn − µ) ∆t ⇒ EVn+1 = (1 − λ∆t)n (EV0∆t − µ) + µ.21) So when 0 < λ∆t < 2. 48 .13). Proof.17) so that when 0 < λ∆t < 2.18) Proof.16). For the meanreverting CEV process in (6.19) (6.2. (6.1 and 6.16). ∆t Proposition 6. we use discretisation schemes to simulate the processes whose exact simulation is diﬃcult. ∆t EVn = (1 − λ∆t)n (EV0∆t − µ) + µ.2. limt→∞ EVt∆t → ∞. (6. Theorem 13.3.Proposition 6. The ﬁrst order Euler scheme with reﬂection principle is applied to simulate the DH process. otherwise.20) (6.2.16).16) will have the L1 convergence: lim sup EVt − Vt∆t  = 0. Appendix C. (6.3.2. limt→∞ EVt∆t → µ.1. under conditions A and B.3 Discretisation Schemes Used in the Monte Carlo Simulation In the Monte Carlo simulation conducted in Chapter 5.22) ∆t→0 0≤t≤T Proof. the discretisation process in (6. For the discretized process (6. t→∞ lim EVt∆t → µ. under conditions A and B.
2 Discretisation of the CEV/SABR process: 1 2 ∆t ∆t ∆t ∆ log St = αt (St )β−1 ∆t − αt (St )2β−2 z. Discretisation of the displaced lognormal process (exact simulation): ∆t ∆ log(St − θ) = − σ2 ∆t + σz 2 Discretisation of the displaced anti lognormal process (exact simulation): ∆t ∆ log(θ − St ) = − σ2 ∆t + σz. Vt∆t w. Discretisation of the displaced Heston process: V ∆t ∆t ∆ log(St − θ) = − t ∆t + 2 ∆Vt∆t = κ(µ − Vt∆t )∆t + Discretisation of the displaced anti Heston process: V ∆t ∆t ∆ log(θ − St ) = − t ∆t − 2 ∆Vt∆t = κ(µ − Vt∆t )∆t + Vt∆t z. w⊥z. ∆t). Vt∆t w. Vt∆t z. 49 . ∆t). ∆t). ∆t).First order LogEuler scheme: Denote z ∼ N (0. 2 ∆ log αt = − ν2 ∆t + ν(ρz + 2 1 − ρ2 w). First order LogEuler scheme with reﬂection principle: Denote z ∼ N (0. w⊥z. w ∼ N (0. w ∼ N (0.
σ. Our goal is to give a largeexpiry asymptotes of σimp (K. 15 and 17 give the largeexpiry asymptotic approximation of σimp (x. (7. σimp (x. K. T ) to be σimp (T ) when no ambiguity arises. T ).1 Case one: K = S0 exT . we analyze the implied volatility of the displaced lognormal dynamics (2.6) in the largestrike and largeexpiry region. Theorems 15 and 17 on 2 the region where x ∈ (−σ 2 /2. T ). consider the strike to be K = S0 exT . 1 σ 2 ]. Theorems 14. T ). (7.1 Largestrike and Largeexpiry Behavior In this section.1. where x is constant. x ∈ R/[− 2 σ 2 . T ) = C BS (S0 − θ. K − θ. σ 2 /2). Denote A(T ) = and deﬁne the domain Kx := {T ∈ R+ : A(T ) + x ≥ 0} (7. S0 exT − θ. T ) < S0 }. σ. Deﬁne the domain of T as follows: KT := {T > 0 : S0 exT > θ+ and (S0 − K)+ < C BS (S0 − θ.4) 1 (−˜(T ) + 2 σ 2 )2 x 2σ 2 (7. T ). 1 1 7.CHAPTER 7 LARGEEXPIRY IMPLIED VOLATILITY OF DISPLACED LOGNORMAL 7. Denote x(T ) = T log((K − ˜ θ)/(S0 − θ)).2) We can suppress the x in σimp (x.1) The implied volatility σimp (x. T ) is deﬁned on KT such that C BS (S0 . 1 where Theorem 14 focuses on the region where x ∈ R/[− 2 σ 2 .3) 50 . 2 σ 2 ] 1 First.
2. 0.3 0.1 0. Figure 7. σ∞ .1.4 Remark 7. 51 . We can write A + x = x∈ 1 R/[− 2 σ 2 . A. T=10 0. 2 2 σimp (x.5) where 1 1 2 σ∞ (x.1. if x ∈ R/[− 2 σ 2 . σ.1 shows the exact σimp and Theorem 14 formula. θ=−20. S0 ABS (x.35 0. σ∞ (x. a1 (x). so if ˜ ˜ the condition A + x > 0 always holds for large T . σ. T ) = σ∞ (x. A(T ). σ=0.6) (7. if x ∈ R/[− σ 2 .1. θ=60.2 S0=100.2. 1 σ 2 ]. T )/T + o(1/T ) as T → ∞. σ=0.1 1 Theorem 14. 0) (7.8) 1 x2 − )−1 log a1 (x. For the displaced lognormal mordel (2.25 0. Figure 7. We can see that the approximation is reasonably accurate. T ) holds.1 0.25 0. Remark 7.1. ˆ (7. 2 σ 2 ] and T ∈ 2 KT ∩ Kx .15 S0=100. σ∞ (x). T ) + a1 (x. T ) to be x.05 0. then the following asymptotic limit of σimp (x. σ 2 ] 2 2 ABS (x.1: Theorem 14 formula and Exact σimp for T = 10. a1 (x.1. T ). 0) x .2 x 0. T ) = 2( 4 ˆ σ∞ (x) 4 Proof. T ) := 2 2A(T ) + x − 2 A2 (T ) + A(T )x . T=10 σimp 0. 2 (−˜+σ 2 /2)2 x σ2 + (x − x) and limT →∞ x(T ) = x. with ˜ the understanding that these variables all depend on T .3 σimp Theorem 14 formula 0.6). 2 2 ˆ ˆ ˜ We suppress the T in x(T ).05 0. Appendix D.7) (7. a1 ) := exp 1 4x2 −1 a1 8 σ4 σ3 1 2 x2 − σ 4 /4 x=±σ /2 ˜ S0 ABS (˜.2.15 0.1.
If θ > 0 and x < 0. Before stating Theorem 17. 2. for ∀B ∈ R.1. However. S0 √ 2x a(T ) e− 2 B (a(T )/2 − B 2 − 1) √ √ + . 52 .1. we express the ˜ condition in terms of x. Theorems 17 and 18 reﬁne the approximation to higher orders. For x > 0. largeexpiry case. σ 2 /2) Theorem 15 gives the largeexpiry asymptotic approximation of σimp (x. Theorem 15. then lim σimp (x.3. T T 2πT 2x (7. 1 σ 2 ] or ˜ 2 2 not. the strike K will eventually become less than θ as T increases. then σimp (x. when x ∈ (−σ 2 /2. we have the asymptotic behavior for the BlackScholes call option formula in the largestrike.5) should be regarded as an approximation of the implied volatility on the domain of T where S0 exT > θ+ . since x will be in a close neighborhood of x when T is large. The derivation of Theorem 14 depends on whether x ∈ R/[− 1 σ 2 . σ 2 /2). the option price will always be S0 − K for large T . (7. σ 2 /2). If θ > 0. Appendix D. for T ∈ KT and x ∈ [0. T ) in the region which is complement to that in Theorem 14. 1 BS C (S0 . T ) = 2x. T ) = N (B)+ (1+O(1/ T )). T ) is not deﬁned for large T .6). largeexpiry case as follows.2 Case two: K = S0 exT .4. If θ < 0. limT →∞ a(T ) = const. 1. Thus ST > θ > K.9) T →∞ Proof.1. we give an asymptotic approximation of the BlackScholes call option in the largestrike. For the displaced lognormal process (2. Theorem 16. 7. Remark 7. Appendix D.3. x ∈ (−σ 2 /2. S0 exT .Remark 7.10) 1 2 2x + 2B Proof.2. (7.
11) where B = S −θ N −1 ( 0 ).4.0767 σimp 0. Forde et al (2009) give the asymptotic behavior of the BlackScholes call option when the volatility is expressed as σ 2 + a/T . we have T ∈ [2 × 104 .5. x ˜ Proof. It elaborates to higher order of approximation. Theorem 16 is diﬀerent from theirs √ in that we have order of 1/ T and we restrict the leading term σ 2 to be 2x. 0. C(T ) = S e ( 0 σ3 2 −σ 4 /4 ). For the displaced lognormal process (2.1.0762 S0=100. Figure 7. T ) = 2x + 2B 2x a(T ) 1 + + o( ).0769 0.0764 0. the second order approximation of σimp (x.077 σimp 0.1. T ) when T → ∞ is given as: 2 σimp (x.Remark 7. θ=60. Theorem 17 reﬁnes the asymptotic behavior given in Theorem 15.2.0768 0.6.1. See appendix D. In the plot.1.0766 0. σ 2 /2) and T ∈ KT . Figure 7. S0 a(T ) = 1 2 2(B 2 +1+C(T )e 2 B 1 x2 x σ 2 ˜ √ ˜ S0 − 2 ( σ 2 −˜+ 4 )T 2x). x=0.0763 0. x ∈ (0. Remark 7.2: Theorem 17 formula and Exact σimp . if θ > 0.2 shows the approximation in Theorem 17 is highly accurate when T is large. σ=0.0765 0.6). T T T (7. 105 ]. 53 .003 approx σimp 2 3 4 5 6 Time 7 8 9 10 x 10 4 Theorem 17.
To sum up. T ) = 2x + 2B 2x 1 a d(T ) + + √ + o( √ ) T T T T T T √ 2 5B 3 +3B+C(T )eB /2 4x T √ . S0 x −σ /4 ˜ 2 2 S −θ Proof. Theorem 19. In Figure 7. Theorem 18 gives a good approximation of σimp when T is large. We can see that for T ∈ [2000. the left plot is the Theorem 17 formula and the exact σimp . Appendix D. and the right plot is the Theorem 18 formula and the exact σimp . Figures 7.1. Theorem 19 gives an explicit formula for the atthemoney implied volatility.1. Appendix D. For the displaced lognormal process (2.Theorem 18 reﬁnes Theorem 17’s approximation to the order of 1 √ . However.4.6).1. 54 . T ) as T → ∞ is given as: 2 σimp (x.5. All the other parameters are the same. So. Theorem 18 signiﬁcantly improves the results of Theorem 17. 1000]. x ∈ (0. if θ > 0.3. the atthemoney implied volatility on T ∈ KT is √ S −θ 2 θ σimp (0. 10000]. T ) T = 2N −1 1 − 2S0 T →∞ (7.1. S0 2S0 T √ θ lim σimp (0. T ∈ [200. we need 0 N (σ T /2) + 2S ∈ [0. neither of them are a good approximation for T ∈ [200.12) where B = N −1 ( 0 ).3 and 7. σ 2 /2) and T ∈ KT . T ) becomes undeﬁned for large T .7. T ) = √ N −1 0 N (σ T /2) + . √ S −θ θ Remark 7.13) Thus (7. For the displaced lognormal process (2.4 compare Theorem 17 and Theorem 18’s approximations of σimp .1. T T Theorem 18. T ∈ [2000. 2x (7. a = 2(B 2 + 1). T ) to exist. 1000] . d(T ) = S0 C(T ) = 1 x ˜ ˜ 3 x σ S0 − 2 ( σ 2 −˜+ 4 )T e ( 2 σ 4 ). when S0 0 θ < 0.4. For σimp (0. the third order approximation of σimp (x. 1]. In both ﬁgures. σimp (0. 10000] and in Figure 7.6).14) Proof.
) is deﬁned by (7. α α ˜ M = xT α−1 .18) M2 1 a1 = 2( 4 − )−1 log ˆ 4 σ∞ and ABS (.15) the implied volatility σimp (x.6 compare the Theorem 20 formula and the exact σimp . for x > 0. the smallest K is about 700.1.6. we have 2 2 σimp (x. Denote x = T1α log( ˜ the domain of T as S0 exT −θ S0 −θ ).3 Case three: K = S0 exT α Theorem 14 can be extended to the region where K approaches ∞ such that K = S0 exT for x = 0 and α > 1. α = 1) to obtain a good approximation. Theorem 20 uses a small x but a large α (e. T ) < S0 }. the smallest K is about 200. x = 0. S0 exT − θ. we have to make the strike K large to have reasonably accurate approximation. M = xT α−1 .1.003. T ) = σ∞ + a1 (T )/T + o(1/T ).17) (7. Theorem 14 uses a large x (e. σ. Letting K approaches inﬁnity in the fashion of K = S0 exT . We can see the approximation is accurate with moderate T .1.7.5) when T ∈ KTα .1.g.6. Remark 7. A = ˜ ˜ −M +σ 2 /2 σ 2 . In both Theorems 14 and 20.16) where 2 σ∞ = 2(M + A − 2M A + A2 ). x = 0.5. 0) . (7.1. σ. Proof.7).1. S0 ABS (M. Theorem 20. 0) (7. in Figure 7. σ∞ . we could have a highly accurate approximation of σimp with reasonable expiry time T . ˜ ˜ S0 ABS (M . Figure 7. T ∈ KTα . α 55 . In Figure 7.2.g. For displaced lognormal dynamics (2.6). Deﬁne KTα := {T > 0 : S0 exT > θ+ and (S0 − S0 exT )+ < C BS (S0 − θ. T ) is well deﬁned by (2. Appendix D.5 and 7. T = 10. α = 3). This is given in Theorem 20. ˆ α α α (7.8. T = 10.
22) Proof. We show the asymptotic behavior and the monotonicity of the implied volatility. 1) = P. (a) For all K > θ+ . (Monotonicity in time) For all θ > 0. Appendix D.5).7. T ) < S0 }. K. K − θ. and analyze the implied volatility of the displaced lognormal dynamics when time to expiry is large. the σimp (x. the ﬁxedstrike largeexpiry implied volatility has the following properties. (7.21) T →∞ lim sgn (7. T ) = − sgn θ ∂T (7.19) T →∞ where σblsimpv (S0 . σblsimpv (S0 . T ) is welldeﬁned when T ∈ KT by (2. S0 − θ). σ.20) (7. we ﬁx the strike. where x is a constant and deﬁne the domain of T as ¯ KT := {T > 0. K > θ+ and T ∈ KT ∂σimp (x. 1. (S0 − K)+ < C BS (S0 − θ.7. ¯ σimp (x. K. T ) T = σblsimpv (S0 . P ). (Asymptotic behavior). P ) is the implied volatility of the oneyear call option with price P deﬁned as follows: C BS (S0 . Theorem 21. For the displaced lognormal process (2. K. 2. Theorem 22 reﬁnes the results of Theorem 21. In this section.1.2 Fixedstrike Largeexpiry Implied Volatility In this section. The notation A(T ) ∼ B(T ) means A(T )/B(T ) → 1 as T ↑ ∞. T ) is not deﬁned when T → ∞.6). K. √ lim σimp (x. (b) For θ < 0. T ∈ KT and 2S0 > θ > 0. denote K = S0 ex . 56 .
a= 1 √ 2π exp( 1 ( −x+a )2 ) S C BS (S0 − θ.S0 −θ) . σ. 57 . For the displaced lognormal dynamics with K = S0 ex . Appendix D.8.23) where z = M= √ . the following holds for the ﬁxedstrike.K. largetime implied volatility: 1 σimp (x. T 2 −1+ 1+4M (− 1 + x 2 ) 8 8a 2 −1+ x 2 4 4a 2 σblsimpv (S0 .1 shows Theorem 22 formula and exact σimp . We can see that for θ close to S0 the approximation is quite accurate even for moderate large T . S0 − θ) + z .2. K − θ.Theorem 22. T ) ∼ √ σblsimpv (S0 . Figure 7. 2 (7. T ) − 0 2 S0 2a 0 S −θ . K. Proof.
3: Theorems 17 and 18 formula and Exact σimp . x=0.003 Exact σimp approximated σimp 0.003 0.003 0.078 0.0764 0.085 0.083 0.0772 0.105 Exact σimp approximated σimp 0.0774 0.084 0. σ=0. θ=60.076 200 300 400 500 600 Time 700 800 900 1000 S0=100.4: Theorems 17 and 18 formula and Exact σimp .077 0. σ=0. σ=0.2.0778 0.082 0.2.003 58 .09 S0=100. 0.085 0.0765 0.0766 0.077 0.2.078 σimp σimp 0.0795 Exact σimp 0.0776 0. θ=60.0775 0. x=0. σ=0.0768 0. x=0.08 200 300 400 500 600 Time 700 800 900 1000 0.079 0.081 0.0762 2000 S0=100.095 σimp σimp 0. x=0.0785 0.1 0.086 0.1.076 2000 3000 4000 5000 6000 Time 7000 8000 9000 10000 3000 4000 5000 6000 Time 7000 8000 9000 10000 Figure 7. θ=60.079 approximated σimp 0.077 S0=100.Figure 7.2. 0.078 Exact σimp approximated σimp 0. θ=60.1.08 0.
θ=60 .2.155 S0=100.T=10.11 0. σ=0.1. α=3 0.115 0.105 3 Exact σimp approximated σimp σimp S0=100.16 0.13 0.6: Theorem 20 formula and Exact σimp for T = 7.2.12 0.17 σimp 0. θ=60.185 Exact σimp approximated σimp 0.1. α=2.8 .175 0.5: Theorem 20 formula and Exact σimp for T = 10.135 0.165 0.18 0.15 0.T=7 4 5 6 x 7 8 9 x 10 10 −3 59 . σ=0. 0.15 3 4 5 6 x 7 8 9 x 10 10 −3 Figure 7.14 0.Figure 7.125 0.145 0. 0.
Figure 7.2.1: Theorem 22 formula and Exact σimp for T = 20.
0.18 0.17 0.16 0.15 σimp 0.14 0.13 0.12 0.11 0.1 −0.4 σimp approx σimp
S0=100, σ=0.4, θ=60, T=20
−0.3
−0.2
−0.1 x
0
0.1
0.2
0.3
60
CHAPTER 8 CONCLUSION
By establishing properties shared by all displaced lognormal volatility skews, Theorems 1 and 2, in eﬀect, exhibit limitations on what phenomena the displaced lognormal can faithfully model. In particular, the displaced lognormal’s everywheremonotonic skews (Theorem 1) and its state space (θ, ∞) = R+ may be drawbacks, when pricing contracts having sensitivity to the tail behavior of an underlying process whose true state space is R+ or whose true volatility skew is nonmonotonic. The slopeconstrained skew (Theorem 2) of the displaced lognormal may be a drawback, when modeling markets, such as typical equity markets, which exhibit downward skews of steepness greater than the Theorem 2 upper bound. The displaced anti lognormal overcomes the slope constraint, but it imposes bounds from above on S, while allowing unbounded negative S. Furthermore, the displaced anti lognormal has the everywheremonotonic skews (Theorem 4). We therefore do not endorse the DL as a model of equity markets, but we do propose the DL as a control variate to improve the accuracy of Monte Carlo pricing under alternative dynamics (such as CEV and SABR) that do model equity prices. The DL is eﬀective as a control variate, because it is simple enough to admit both unbiased simulation and explicit pricing formulas for many contracts, yet ﬂexible enough to generate high correlations and hence signiﬁcant variance reduction with respect to the CEV/SABR dynamics. We give two examples. The ﬁrst is the downandout call option under the CEV/SABR dynamics. The DL control generates large variance reduction as shown in Table 5.1.1 and 5.1.2. The second example is the discrete arithmetic Asian option. We compare the variance reduction of the DL control and the lognormal control. The DL control generates signiﬁcant more satisfactory variance reduction than LN control as shown in Table 5.1.4 and 5.1.5. Another displaced process, the displaced (anti )Heston process, will generate more variance reduction when used as control variate. We are able to have explicit pricing formulas for the calls/puts and barrier options on the DH process. Though no exact simulations of the DH process are available, we have a discretisation scheme which guarantees convergence
61
to the true process. We show that DH control improves the variance reduction of DL control when applied to the SABR dynamics in Table 5.1.3. Moreover, we analyze the variance reduction by combining the DL/DH control variate with importance sampling. The combined technique is superior to the pure importance sampling when applied to options under the SABR dynamics, as shown in Table 5.2.1 and 5.2.2, Table 5.2.3 and 5.2.4. Toward either purpose – as a model, or as a computational device on behalf of another model – the DL calibrates easily to a given volatility skew level and slope, via the shortexpiry limiting implied volatility formula of Theorems 3 and 6; and the DH calibrates easily to a given volatility skew level, slope and convexity, via the shortexpiry limiting implied volatility formula of Theorems 8 and 9. Furthermore, we give an explicit relationship between the shortexpiry implied volatilities of any process which is a positive martingale with its corresponding (anti)displaced process (Theorems 10 and 11). Combined with Durrleman’s (2004) approximation of the shortexpiry nearthemoney implied volatility, we are able to give an approximation of the shortexpiry nearthemoney implied volatility of any (anti)displaced process. In contrast to the shortexpiry limiting implied volatility, the largeexpiry limiting implied volatility of the displaced lognormal is also analyzed. We ﬁrst give the approximation of the implied volatility in the largestrike case (Theorems 14, 15, 17 and 18 where K = S0 exT ; Theorem 20 where K = S0 exT ). Secondly, we give the approximation of the implied volatility in the ﬁxedstrike case (Theorems 21 and 22). Finally, we discuss some issues in the discretisation scheme of the Monte Carlo simulation. Under some regularity conditions, we prove the partial strong convergency of the discretized stochastic volatility process to the continuous process (Theorem 12). We also prove a strong convergency result of the meanreverting CEV process (Theorem 13).
α
62
For all T > 0 and K ∈ K(T ): If θ > 0 then σimp (K) < σ.2. σ) > S0 −θ +θ = S0 . σ) σimp (K) for all K ∈ K. By Proposition A.1. If θ ≥ 0 then K = θ.1. we have C BS (S0 − θ. K − θ. 0−θ. √ ∂C BS ∂C BS d BS C (S0 − θ. K − θ. Our notation may suppress the S or T . By (2. σimp (K) < σ for all K ∈ K. we have σimp (K) → σ as K → ∞. C BS (S0 . according as θ 0. K. For all T > 0. We prove for θ > 0. and inf K∈K(T ) σimp (K) = σ if θ < 0.9).2) (A.1 Appendix: Proof of Theorem 1–Global Behavior of Displaced Lognormal Deﬁne for all K > θ+ √ log[(S0 − θ)/(K − θ)] σ T √ . If θ < 0 then σimp (K) > σ. (A. there exists k such that for all K > k.APPENDIX A A. σ).1. Note that K = (K. σ − δ < σimp (K) 63 σ. If θ < 0 then K > 0 because C BS (S0 −θ. Proof. For all K > θ and S0 > θ. − d2 (K) := 2 σ T Let K S (T ) := inf KS (T ). Proposition A. Hence supK∈K(T ) σimp (K) = σ if θ > 0. Proof.1) .1. ∞). The proof for θ < 0 is similar. σ) = − − = −N (d2 + σ T ) + N (d2 ) < 0 dθ ∂S ∂K so. So it suﬃces to show that for all δ > 0. Proposition A.1.
64 . K − θ. Lemma A.or equivalently C BS (S0 . ∞) → R is deﬁned by √ F (y. σ − δ) < C BS (S0 .6) √ log(S0 /K) (σ − δ) T √ − +N . 2 (σ − δ) T (A. K) := −y 2 T /2 − y T d2 (K) + log(S0 /K). Therefore M (K) > 0 for K suﬃciently large.3) √ log(S0 /K)σ − log[(S0 − θ)/(K − θ)]σ−δ δ T √ + = sgn 2 (σ − δ)σ T √ log[(S0 × (K − θ))/(K × (S0 − θ))]σ − log[(K − θ)/(S0 − θ)]δ δ T √ = sgn + .3.5) (A. 2 (σ − δ)σ T (A. K) ∂K where F : R × (θ+ . Moreover limK→∞ M (K) = 0 − 0 = 0. For all K > θ+ . σimp (K)) = C BS (S0 − θ. Hence it suﬃces that for K large enough. which we verify as follows. σ) − C BS (S0 . σ − δ). K.7) The part inside the sgn in (A. ∂M/∂K(K) < 0.4) (A. K. √ ∂M log[(S0 − θ)/(K − θ)] σ T √ − = −N ∂K 2 σ T So √ ∂M log(S0 /K) log[(S0 − θ)/(K − θ)] δ T √ − √ sgn = sgn + ∂K 2 (σ − δ) T σ T (A.1. sgn ∂σimp (K) = sgn F (σimp (K).7) approaches −∞ as K → ∞. σ). K. which proves (A. For all K ∈ K. So for K suﬃciently large. K − θ.3). 0 < M (K) := C BS (S0 − θ.
1. sgn ∂σimp (K) = sgn θ. the conclusion will hold for all K ∈ K. Lemma A. Take the K derivative of (2.9) √ √ ∂σimp (S0 ) = sgn − σimp (S0 ) T /2 + σ T /2 > 0 ∂K ∂σimp (K) = sgn θ. σimp (K)) + (S0 . If θ = 0 then this is obvious.1. There exists K ∈ K such that sgn Proof. If θ < 0 then as K ↓ K we have C BS (S0 − θ. K.4. Otherwise. σimp (K)) ∂K ∂K ∂K √ log(S0 /K) σimp T √ = sgn − − d2 (K) 2 σimp T √ 2 = sgn − σimp (K)T /2 − σimp (K) T d2 (K) + log(S0 /K) = sgn F (σimp (K). σ) − (S0 .4 the conclusion holds for at least one K ∈ K.8) Therefore sgn ∂σimp ∂C BS ∂C BS (K) = sgn (S0 − θ. K − θ. K − θ.1. ∂K ∂σ ∂K ∂K (A. or equivalently (by Lemma A. K.Proof. K). K − θ. as claimed. by Lemma A.3) that 65 . If θ > 0 then the atthemoney strike K = S0 satisﬁes the conclusion.5. σimp (K)) (K) = (S0 − θ. as claimed. ∂σimp ∂C BS ∂C BS ∂C BS (S0 . σ) → S0 hence σimp → ∞. If θ = 0 then this is obvious.9).1. ∂K Proof. because sgn by Proposition A. ∂K (A. Proposition A.1. So there exists K > K such that ∂σimp /∂K(K) < 0. K.1. σ). For all T > 0 and K ∈ K(T ). By continuity of ∂σimp ∂K on K. if we can show that ∂σimp /∂K = 0 on K.
K) = 0.6.11) Lemma A.F (σimp (K). K) ∈ R × K.7. K. ·). 66 . deﬁne y± (K) := −d2 (K) ± ∆(K) √ . K) is a quadratic with no real roots. If K ∈ D then F (·. Likewise y− (K) σimp (K) if H− (K) 0.1. K Let D := {K > θ+ : ∆(K) ≥ 0}. H± (K) = C BS (y± (K)) − C BS (σimp (K)) where C BS (·) is shorthand for C BS (S0 . by verifying that F (σimp (K). If K ∈ D then F (·.10) (A. For K > θ+ deﬁne ∆(K) := d2 (K) + 2 log(S0 /K) 2 √ K −θ h(K) := d2 (K) + σ T . K) = 0 for all K ∈ K. K) has / roots y = y± (K). and for all K ∈ D. Lemma A. Hence y+ (K) σimp (K) if H+ (K) 0. K) = 0 if and only if K ∈ D and y = y± (K). For (y. we have F (y. T (A. Proof. The remaining lemmas complete the proof. For all K ∈ D deﬁne √ H± (K) := S0 N (± ∆(K)) − (S0 − θ)N (d2 (K) + σ T ) − θN (d2 (K)) Then for all K ∈ D ∩ K.12) (A.1.
Proof. d2 (K) + 2 log(S0 /K) > d2 (K) + 2 2 or equivalently that K −θ 2 − K K −θ K σ2T + 2 K −θ log[(S0 − θ)/(K − θ)] − 2 log(S0 /K) < 0. Lemma A. ∂K (K) > 0. If moreover K ∈ D then ∂K (K) < 0. If θ 0 then for all K ∈ D we have ∆(K) h(K). (A. K 0 for K S0 . This is veriﬁed by L(S0 ) = 0 and ∂L/∂K = (θ/K 2 ) log[(S0 −θ)/(K −θ)] Lemma A.8. We need to show that for all K ∈ D. because 0 < (K − θ)/K < 1. ∂H ∂H ∂H− ∂H+ ∂∆ • If θ < 0. we have C BS (y BS (σ ± (K)) − C imp ) √ log(S0 /K) y± (K) T √ + = S0 N − KN (d2 (K)) 2 y± (K) T √ − (S0 − θ)N (d2 (K) + σ T ) − (K − θ)N (d2 (K)) √ √ = S0 N (d2 (K) + y± (K) T ) − (S0 − θ)N (d2 (K) + σ T ) − θN (d2 (K)) = H± (K) The remaining conclusion is by monotonicity of C BS . h(K) > 0 then ∂K (K) < 0. The proof for θ < 0 is similar.√ √ Proof.1. so it suﬃces to show that for all K > θ. L(K) := K −θ log[(S0 − θ)/(K − θ)] − log(S0 /K) ≤ 0. Consider only the case θ > 0. For all K > θ+ : − + • If θ > 0 then K ∈ D and ∂K > 0. If moreover K ∈ D then ∂K (K) > 0. h(K) < 0 then ∂K (K) > 0.1. ∂H− ∂H+ ∂∆ • If θ < 0.9).13) K √ K −θ K −θ 2 2 σ T + 2d2 (K)σ T K K The ﬁrst term is negative. 67 . Using log(S0 /K)/(y± (K) T ) − y± (K) T /2 = d2 (K) and (2.9. ∂K > 0. ∂K (K) < 0.
8.11. the K ∈ D conclusion clearly holds for K ∈ (S0 . Lemma A. ∞). ∂H− /∂K > 0 on K. Lemma A. (K) = − ∂K (K − θ)σ T If θ > 0. So on K we have H+ > 0.9.7. note that as K → ∞ we have d2 (K) → −∞ hence H− (K) → 0. ∞).1. By Lemma A. Moreover. Moreover. To prove y− < σimp . S0 ] because √ √ √ √ log[(S0 − θ)/(K − θ)]/(σ T ) − σ T /2 < log(S0 /K)/(σ T ) − σ T /2 < 0 implies √ √ √ √ ∆(K) > (log(S0 /K)/(σ T )−σ T /2)2 +2 log(S0 /K) = (log(S0 /K)/(σ T )+σ T /2)2 ≥ 0.1. we have K ∈ D and y1 (K) < σimp (K) < y2 (K). by Lemma A. hence σimp < y+ by Lemma A. note that as K ↓ θ we have d2 (K) → ∞ hence H+ (K) → 0.1.9. / 2 ∆(K) h(K) 68 .1. If θ < 0 then for each K ∈ K ∩ D we have σimp (K) ∈ [y− (K). The ∆ conclusions hold because 2h(K) ∂∆ √ . To prove σimp < y+ .7. the H± conclusions hold because of ∂H± Ke−d2 (K)/2 1 √ (K) = √ ∂K 2π (K − θ)σ T ∆(K) and Lemma A.1.9 we have K ∈ D and y± (K) are welldeﬁned.1.1. hence y− < σimp by Lemma A. in all cases.1. So on K we have H− < 0.Proof. Lastly. If θ > 0 then for all K ∈ K = (θ. ∂H+ /∂K > 0 on K.10. and also holds for K ∈ (θ. by Lemma A. y+ (K)]. Proof.
12.6.11. then for all k > K we have h(k) < 0 and ∂∆/∂K(k) > 0 by Lemma − A. and A. Because K ∈ D. Combine Lemmas A. ∞). Proof. Lemma A.2 Appendix: Proof of Theorem 2–Atthemoney Behavior of Displaced Lognormal Proof of Theorem 2. we have h(K) = 0 by Lemma A. (A. σ) < S0 − (S0 − θ + θ) = 0. So for all k ∈ (0.1. A. A. we have ∆(k) → ∞.1. 69 .1.1. ∂H If h(K) > 0.1.7.9. then for all k ∈ (0. K) we have h(k) > 0 and ∂∆/∂K(k) < 0 by Lemma + A.10. So for all k > K we have k ∈ D and ∂K (k) < 0 by Lemma A. Moreover limk→∞ H− (k) = 0. Therefore H+ (K) < 0. S0 . Similarly. σ). for θ ≤ 0.14) √ √ S0 [2N (σatm T /2) − 1] = (S0 − θ)[2N (σ T /2) − 1].16) for θ ≥ 0. For all K ∈ K we have F (σimp (K). Moreover.1.9. hence y+ (K) < σimp (K) by Lemma A. because N (x) is concave on x ≥ 0. hence σimp (K) < y− (K) by Lemma A.1.1. Monotonicity of N implies σatm ≤ (1 − θ/S0 )σ. S0 − θ. If h(K) < 0. hence ∂H √ log[(S0 − θ)/(−θ)] σ T √ H+ (k) → S0 − (S0 − θ)N + 2 σ T √ log[(S0 − θ)/(−θ)] σ T √ − θN − 2 σ T = S0 − C BS (S0 − θ. hence (A.1.1. as k ↓ 0.1.Proof. K) = 0.8. K) we have k ∈ D and ∂K (k) < 0 by Lemma A.15) and √ √ S −θ 1 N (σatm T /2) = 0 [2N (σ T /2) − 1] + 2S0 2 √ θ θ θ = (1 − )N (σ T /2) + N (0) ≤ N ((1 − )d1 ) S0 S0 S0 (A.7.1. it is clear that h is decreasing on (0. Taking K = S0 in (2.9). σatm ) = C BS (S0 − θ. −θ. Because θ < 0. we have C BS (S0 .9. Therefore H− (K) > 0. we have σatm ≥ (1 − θ/S0 )σ.9.
19) (A. T ) − (S0 − K)+ ) .18) (A.21) On the other hand.8). σ. ∂ log K K=S 2σatm 2 φ(σatm T /2) 0 as claimed. K − θ. ∞). = = ∂ log K K=S σimp ∂K K=S φ(σatm T /2)σatm T 0 0 because θ < 0 implies σatm > σ. K − θ. T ) < S0 hence K ∈ KS (T ).1 produces the lower bound ∂ log σimp σatm − σ 1 1 ≥ ≥ 1− ∂ log K K=S 2σatm 2 1 − θ/S0 0 Combining (A.18) produces the upper bound √ ∂ log σimp φ(σ T /2) σatm − σ 1 2 √ × ≤ ≤ eσatm T /8 . Proof of Theorem 3. we have.Proof of Theorem 2. as T ↓ 0. (A. σimp (K. Applying RoperRutkowski (2007) Proposition 5.22) . applying it to the BlackScholes model. σ.17). T ) ↓ (S0 − K)+ as T ↓ 0. K − θ. T ) − (S0 − K)+ ) 70 . as T ↓ 0. σ∼  log((S0 − θ)/(K − θ)) −2T log(C BS (S0 − θ. σ. we have. C BS (S0 − θ. we have C BS (S0 − θ. For all K > θ+ . 2(S0 + θ) (A.18). K − θ.2.3 Appendix: Proof of Theorems 3 and 6–Shortexpiry Behavior of DL The notation A(T ) ∼ B(T ) means that A(T )/B(T ) → 1 as T ↓ 0. (A.17) and (A. σ. √ √ √ √ N (σatm T /2) − N (σ T /2) √ √ ≤ φ(σ T /2). By concavity of N on [0. so for all T suﬃciently small. Using (A.1 to the displaced lognormal model. T ) ∼  log(S0 /K) −2T log(C BS (S0 − θ.17) A.20) = θ . (A. we have √ √ ∂ log σimp S0 ∂σimp N (σatm T /2) − N (σ T /2) √ √ . (A. φ(σatm T /2) ≤ σatm T /2 − σ T /2 Combining (A. and Theorem 2.
(A. θ − S0 .Therefore σimp (K. Applying RoperRutkowski (2007) Proposition 5. K) ∂K where F : R × (0. T ) ∼ as T ↓ 0. we have. σimp (K. we have C BS (θ − K. θ − S0 . −σ. θ).4 Appendix: Proof of Theorem 4–Global Behavior of Displaced anti Lognormal Deﬁne for all K < θ √ log[(θ − S0 )/(θ − K)] σ T √ d2 (K) := − . −σ. For all K ∈ (0. θ) → R is deﬁned by √ F (y. (A. we have. −σ.25) A.  log((θ − K)/(θ − S0 )) −2T log(C BS (θ − K.4. θ − S . applying it to the BlackScholes model. C BS (θ − K.1. as T ↓ 0.23) −2T − K)+ ) On the other hand. Proof of Theorem 6. T ) ∼  log(S0 /K) BS (θ − K. as T ↓ 0. T ) − (S0 − K)+ ) . T ) − (S log(C 0 0 .23) holds. θ − S0 . T ) < S0 hence K ∈ KS (T ). 71 .1 to the displaced antilognormal model. −σ ∼ Therefore (A. so for all T suﬃciently small. −σ. sgn ∂σimp (K) = sgn F (σimp (K). For all K ∈ KADL . 2 σ T Lemma A. K) := −y 2 T /2 + y T d2 (K) + log(S0 /K).24) σ log(S0 /K) σ log(S0 /K) =  log((S0 − θ)/(K − θ)) log((S0 − θ)/(K − θ)) (A. T ) ↓ (S0 − K)+ as T ↓ 0.
the conclusion will hold for all K ∈ KADL . K) = 0. θ−K. σ). 72 . By Lemma A. sgn ∂σimp (K) = − sgn θ. as claimed. K. K). σimp (K)) ∂K ∂K ∂K √ σimp (K) T log(S0 /K) √ − = sgn − 1 + N (d2 (K)) + N 2 σimp (K) T √ σimp (K) T log(S0 /K) √ − + d2 (K) = sgn 2 σimp (K) T √ 2 = sgn − σimp (K)T /2 + σimp (K) T d2 (K) + log(S0 /K) = sgn F (σimp (K). by verifying that F (σimp (K).2. The remaining lemmas complete the proof. Taking the K derivative of (2.19). σimp (K))+ (S0 . The atthemoney strike K = S0 satisﬁes the conclusion.26) Therefore sgn ∂σimp ∂C BS ∂C BS (K) = sgn − 1 − (θ − S0 . there exists K ∈ KADL such that sgn ∂σimp (K) = − sgn θ. ∂K Proof. θ − K. K. ∂K Proposition A. By continuity of ∂σimp ∂K on KADL . σimp (K)) (K) = −1− (θ−S0 . Lemma A.4. because sgn √ √ ∂σimp (S0 ) = sgn − σimp (S0 ) T /2 − σ T /2 < 0.Proof. or equivalently (by Lemma A.4. σ) − (S0 . K. K ∈ KADL (T ) and θ > S0 .2 the conclusion holds for at least one K ∈ KADL .1) that F (σimp (K). ∂K ∂σ ∂K ∂K (A.4. ∂K (A.27) Proof. if we can show that ∂σimp /∂K = 0 on KADL . K) = 0 for all K ∈ K. θ > S0 > 0.3. For all T > 0.4. ∂σimp ∂C BS ∂C BS ∂C BS (S0 .
K) = 0 if and only if K ∈ D and y = y± (K). Using log(S0 /K)/(y± (K) T ) − y± (K) T /2 = −d2 (K) and (2. Proof. ·). K) has / roots y = y± (K). K Let D := {0 < K < θ : ∆(K) ≥ 0}. Lemma A.5. For (y. σimp (K) if H+ (K) 0.4. K) ∈ R × KADL .30) (A.18).For K < θ deﬁne ∆(K) := d2 (K) + 2 log(S0 /K) 2 √ θ−K h(K) := d2 (K) − σ T . If K ∈ D then F (·. and for all K ∈ D. For all K ∈ D deﬁne H± (K) := −S0 N ( √ ∆(K)) − (θ − S0 )N (d2 (K) + σ T ) + θN (d2 (K)) (A.29) Lemma A. we have P BS (y BS (σ ± (K)) − P imp ) Hence y+ (K) √ √ log(S0 /K) y± (K) T log(S0 /K) y± (K) T √ + √ − = KN − − S0 N − 2 2 y± (K) T y± (K) T √ − (θ − S0 )N (d2 (K) + σ T ) − (θ − K)N (d2 (K)) √ √ = KN (d2 (K)) − S0 N (d2 (K) − y± (K) T ) − (θ − S0 )N (d2 (K) + σ T ) + (θ − K)N (d2 (K)) = H± (K) 73 . K.4. Likewise y− (K) σimp (K) if H− (K) 0. H± (K) = P BS (y± (K)) − P BS (σimp (K)) where P BS (·) is shorthand for P BS (S0 . If K ∈ D then F (·.28) (A.4. √ √ Proof. (A. we have F (y.31) Then for all K ∈ D ∩ KADL . deﬁne y± (K) := d2 (K) ± √ ∆(K) T . K) is a quadratic with no real roots.
θ) such that h(k0 ) = 0. This is veriﬁed by L(S0 ) = 0 and ∂L/∂K = (θ/K 2 ) log[(θ−S0 )/(θ−K)] Lemma A.4. Lemma A. = ∂K (θ − K)σ T The H± conclusions hold because of ∂H± 1 Ke−d2 (K)/2 √ (K) = √ ∂K 2π (θ − K)σ T ∆(K) 2 (A. h(K) is monotonically in creasing on (0. By (A. K 0 for K S0 .33) ∆(K) ± h(K) . ∆(K) is decreasing on {K : 0 < K < k0 } and increasing on {K : k0 < K < θ}. because (θ − K)/K > 0. L(K) := − θ−K log[(θ − S0 )/(θ − K)] − log(S0 /K) ≥ 0.4. ∂K (K) < 0. / Proof. If θ > 0 then for each K ∈ KADL ∩ D we have σimp (K) ∈ [y− (K). 74 . The ﬁrst term is positive. ∂∆ − + • If h(K) < 0 then ∂K (K) < 0. so there is a point k0 ∈ (0.The remaining conclusion is by monotonicity of P BS . If moreover K ∈ D then ∂K (K) > 0. Lemma A.8.6. Because θ > 0 and ∂h(K) ∂K = 1 1 √ θ−K σ T + √ σ T θ K2 > 0. y+ (K)]. We want to show: h2 (K)−∆(K) = θ−K 2 θ−K + K K σ 2 T −2 θ−K log[(θ−S0 )/(θ−K)]−2 log(S0 /K) > 0. K (A.4. If moreover K ∈ D then ∂K (K) < 0. The ∆ conclusion hold because: 2h(K) ∂∆(K) √ .7. θ > S0 .33). θ). so it suﬃces to show that for all K > θ. for all K ∈ D we have Proof. ∂K (K) > 0. Furthermore h(0) = −∞. ∂H ∂H ∂H ∂H Proof. h(θ) = ∞.32) ∆(K) < h(K). For all 0 < K < θ: ∂∆ − + • If h(K) > 0 then ∂K (K) > 0.
6. θ).4. Combine Lemmas A.5. T ) < 0.34) (A.4. we have ∂K (k) < 0 by Lemma A.4. Consider atthemoney in (2. we have. k ↓ 0.35) 75 . we have h(k) < 0 and + ∂∆/∂K(k) < 0 by Lemma A.Because K ∈ D.7.4. So for all k > K and k ∈ D. K) = 0. Proof.4. we have h(K) = 0 by Lemma A. So for all k < K and k ∈ D. Therefore H− (K) > 0 hence y− (K) > σimp (K) when K ∈ KADL by Lemma A. we have ∂K (k) < 0 by Lemma A. hence by (A. Lemma A. ∂H k = θ ⇒ d2 (k) = ∞ ⇒ ∆(k) = ∞.7.4. ⇒ H− (θ) = −S0 N (− √ ∆(k))−(θ−S0 )N (d2 (k)+σ T )+θN (d2 (k)) = −S0 −(θ−S0 )+θ = 0.19). √ √ S0 [2N (σatm T /2) − 1] = (θ − S0 )[2N (σ T /2) − 1]. we have h(k) > 0 and ∂∆/∂K(k) > 0 by Lemma − A. For all K ∈ KADL we have F (σimp (K). then for all k < K. Moreover. and √ √ θ − S0 1 N (σatm T /2) = [2N (σ T /2) − 1] + 2S0 2 √ θ θ = ( − 1)N (σ T /2) + (2 − )N (0).4.5.4. σ.5 Appendix: Proof of Theorem 5–Atthemoney Behavior of Displaced anti Lognormal Proof of Theorem 5.4. then for all k ∈ (K. and A.1.7. if h(K) < 0. For a particular K ∈ D. S0 S0 (A.31) ∂H k→0 √ lim H+ (k) = 0 − (θ − S0 )N (d2 (0) + σ T ) + θN (d2 (0)) = 0 − C BS (θ − S0 .4. θ.8. A.7.9. ∆(k) → ∞.4. If h(K) > 0. Therefore H+ (K) < 0 hence y+ (K) < σimp (K) when K ∈ KADL by Lemma A.
2.θ If S − 1 ∈ [0. φ(σatm T /2)σatm T K=S0 (A.41) (A.40) and (A.37) Proof of Theorem 5.34). ∞). S0 S0 S0 2 Using the monotonicity of N . Monotonicity of N implies σatm ≤ (θ/S0 − 1)σ.26) and plug in (A. we have ∂ log σimp θ θ 2 ≤ ≤ exp(σatm T /8) 2(θ − S0 ) ∂ log K K=S 2(θ − S0 ) 0 (A. Reorder the entries in (A.42) 76 .39) = By the concavity of N on [0. 1]. if S − 1 ∈ [1. then 0 √ √ θ θ θ σ T ( − 1)N (σ T /2) + (2 − )N (0) ≤ N ( − 1) . φ(σatm T /2) ≤ σatm T /2 Combine (A. θ Similarly.40) √ √ N (σatm T /2) − N (0) √ ≤ φ(0).36) √ √ θ θ θ σ T ( − 1)N (σ T /2) + (2 − )N (0) ≥ N ( − 1) . S0 S0 S0 2 Because N (x) is concave on x ≥ 0. (A. then 0 (A. ∞): (A.41). we have.38) (A. we have σatm ≥ (θ/S0 − 1)σ. ∂ log σimp = = ∂ log K K=S 0 = √ √ −1 + N (−σ T /2) + N (−σatm T /2) √ √ φ(σatm T /2)σatm T K=S0 √ √ N (σ T /2) + N (σatm T /2) − 1 √ √ φ(σatm T /2)σatm T K=S0 √ θ θ−S0 N (σatm T /2) − 1/2 √ √ .
1 Appendix: Proof of Theorem 7–Atthemoney Behavior of Displaced Independent Stochastic Volatility Proof of Theorem 7. T Σ= (B. (B.2) Diﬀerentiating with respect to K in (B. (B. the process ST − θ is lognormal process. we have sgn ∂σimp ∂C BS (S0 − θ. Σ)). σatm ) = sgn E − (B. K − θ. Σ) ∂C BS (S0 . K. the implied volatility σimp satisﬁes C BS (S0 . we have ∂C BS (S0 .1) Since conditioning on the average volatility Σ.4) ∂K K=S ∂K ∂K K=S0 0 √ √ σatm T Σ T = sgn E N (− ) − N (− ) (B.σimp ) ∂σimp > 0. σimp ) = E(C BS (S0 − θ. K − θ. K − θ.APPENDIX B B.5) 2 2 √ √ Σ T σatm T = sgn EN ( ) − N( ) . σimp ) ∂σimp ∂C BS (S0 − θ. K.K. (B.6) 2 2 77 . Σ)). K. σimp ) =E − . K.2) and exchanging the expectation and the diﬀerentiation (validate later).3) ∂σimp ∂K ∂K ∂K Since ∂C BS (S0 . K − θ. Deﬁne the average volatility of the DISV process: T 0 2 σt dt . E(ST − K)+ = E{E[(ST − θ) − (K − θ)]+ Σ} = E(C BS (S0 − θ. Σ) ∂C BS (S0 . For K ∈ KDISV .
K. Σ)) ∂C BS (S0 − θ. Σ)). K. Σ)).What remains is to show θ 0 ⇒ N ( σatm T ) 2 √ EN ( Σ 2 T ).7) To complete the proof. Thus. Denote C DH (S0 . σimp ) E(C BS (S0 . Particularly when atthemoney. maturity T and initial value S0 . Since ∂C BS (S0 − θ. K − θ. K − θ. K − θ.2 = Σ T creasing as a function of θ. T ) the call option price under the DH process. with strike K.2 Appendix: Proof of Theorems 8 and 9– Shortexpiry Behavior of DH Proof of Theorem 8. So E(C BS (S0 − θ. ∂K we can use dominant convergent theorem to exchange the expectation and the diﬀerentiation. we have ∂E(C BS (S0 − θ. we need to validate the exchangeability of the expectation and the diﬀerentiation. E(C BS (S0 . B. ∂θ ∂θ 0 √ where d1. K. T ) = C H (S0 − θ. Σ) =  − N (d1 ) + N (d2 ) ≤ 2. K. we have √ σatm T N( ) 2 √ Σ T EN ( ). Σ)) is monotonically de θ 0 ⇒ E(C BS (S0 − θ. 2 (B. Σ) =  − N (d2 ) ≤ 1. K − θ. ∂θ ∂C BS (S0 − θ. K − θ. K − θ. and similarly C H (S0 . K − θ. log((S −θ)/(K−θ))±Σ2 T /2 . Σ) =E = E(−N (d1 ) + N (d2 )) ≤ 0. We have C DH (S0 . K. Assuming the exchange √ ability of the expectation and the diﬀerentiation. T ) the call option price under the Heston process. T ). K. Applying 78 . Σ)) ⇒ C BS (S0 .
T )/(S0 T ) On the other hand. K. (B. Proof of Theorem 9.K.T )−(S0 −K)+ ) H σimp (K − θ. we have ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ E(ST − K)+ − (S0 − K)+ = E(K − ST )+ − (K − S0 )+ = E(ST − K)+ − (S0 − K)+ . Denote C ADH (S0 .12) Applying RoperRutkowski (2007) corollary 5. K. T ) the call ¯ ¯ option price under the Heston process. we have. K − θ.11) If atthemoney. and similarly C H (S0 .10) if K = S0 . K = θ − K. (B. then. as T ↓ 0. maturity T and initial value S0 .RoperRutkowski (2007) corollary 5. with strike K.8) and (B. Using putcall parity. T ) if K = S0 (B. we have lim σ H (K − θ. K.1 to the displaced antiHeston model. applying it to the Heston model.  log((S0 −θ)/(K−θ)) −2T log(C H (S0 −θ. T ) ∼ if K = S0 (B.T )−(S0 −K)+ ) ADH σimp (K. T )/((S0 − θ) T ) if K = S0 By comparing (B. T ) the call option price under the displaced anti process.8) if K = S0 ∼ √ 2πC DH (S0 . as T ↓ 0. T ) ∼ if K = S0 (B.T )−(S0 −K)+ ) DH σimp (K.9) √ 2πC H (S0 − θ. K.  log(S0 /K) −2T log(C ADH (S0 . as T ↓ 0.13) if K = S0 √ 2πC ADH (S0 . we have. we have. T ) = T →0 T →0 lim σ H (K − θ.1 to the displaced Heston model. T )/(S0 T ) 79 .K−θ.9). Let S = θ − S. T ) × (1 − θ/S0 ) imp log(S0 /K) log((S0 − θ)/(K − θ)) if K = S0 (B. ¯ ¯ E(ST − K)+ = E(ST − K)+ . and using (3.  log(S0 /K) −2T log(C DH (S0 .12). T ) × imp T →0 DH lim σimp (K.K.
17) where A(T ) ∼ B(T ) means A(T )/B(T ) → 1 as T ↓ 0 and 2 2 ρ 7ρ2 2) + ρ σ2 − ).11) and (B. we have a0 = 0. c0 = 2 (1 − 0 2 4 2 6 6σ0 (B. T ) ∼ if K = S0 (B.T )−((θ−S0 )−(θ−K))+ ) H σimp (θ − K. we give a limit approximation of the level.15) if K = S0 . the nearmoney short˜ ˜ expiry implied volatility with strike K.1–Level.Applying it to the Heston model.2 and 3.18) Since we have speciﬁed that ρ = 0 for the Heston process. slope and convexity as T ↓ 0. 2 2 (B.1. as T ↓ 0. we have.16) It can be approximated as H ˜ σimp (K. Slope and Convexity of DH Shortexpiry Implied Volatility Proof of Proposition 3. T ) ∼ T c 2 ˜ ˜ ˜ ˜ σ0 + a0 log(S0 /K) + b0 + 0 log2 (S0 /K). 2 2 (B. T ) = c T 2 ˜ ˜ ˜ ˜ ˜ ˜ σ0 + a0 log(S0 /K) + b0 + 0 log2 (S0 /K) + O(log(S0 /K)T + T 2 ).13) and (B. B. a0 = − .  log((S0 −θ)/(K−θ)) −2T log(C H (θ−S0 .2. K = K − θ.2.14). θ − K. Though we do not 80 . T )/((θ − S0 ) T ) By comparing (B. T ) = T →0 T →0 H (θ − K.θ−K.12).1. b0 = κ(µ − σ0 (1 − ρ2 /4). maturity T and initial value S0 is H ˜ σimp (K. ˜ ˜ Denote St = St − θ. We utilize the shortexpiry approximation formula of the implied volatility given by Durrleman (2004) and the relationship between the implied volatilities of the DH model and the Heston model (Theorems 8 and 9). we have log(S0 /K) lim −σ H (θ − K. and using (B.14) if K = S0 √ 2πC H (θ − S0 . T ) × imp log((S0 − θ)/(K − θ)) ADH lim σimp (K.3 Appendix: Proof of Propositions 3. In the following. T ) × (1 − θ/S ) lim −σ 0 imp T →0 if K = S0 (B. Durrleman (2004) shows that.
17) for the Heston model. the nearexpiry atthemoney implied volatility of the displaced Heston is: 2T 2 T DH H 2 (σimp (S))2 = (σimp (S0 −θ)(1−θ/S0 ))2 = (σ0 +κ(µ−σ0 ) − )×(1−θ/S0 )2 +O(T 2 ).19) ii.13)(3.20) 81 . 2 (log((S0 − θ)/(K − θ)))2 K→S0 2S0 lim Applying (B. we have. ˜ ˜ K=S0 (B.14). H ˜ ∂σimp (K) 1 H ˜ = (σimp (K))−1 ˜ 2 ˜ ˜ ∂K K=S0 − a0 1 ˜ ˜ 1 − c0 log(S0 /K) ˜ ˜ K K = 0 + O(T ). i.explicitly write the limT →0 out. Slope DH ∂σimp (K) ∂K K=S0 = K→S0 lim DH DH σimp (K) − σimp (S0 ) K − S0 log(S /K) H H 0 σimp (K − θ) log((S −θ)/(K−θ)) − σimp (S0 − θ)(1 − θ/S0 ) 0 = lim K − S0 K→S0 = K→S0 H σimp (K lim H ∂σimp (K − θ) ∂K log(S0 /K) + log((S0 − θ)/(K − θ)) 1 1 − K log((S0 − θ)/(K − θ)) + K−θ log(S0 /K) − θ) (L’Hopital’s rule) (log((S0 − θ)/(K − θ)))2 = K→S0 lim H ∂σimp (K − θ) ∂K H (1 − θ/S0 ) + σimp (S0 − θ) θ 2 2S0 where in the last step we use log(S0 /K) = 1 − θ/S0 K→S0 log((S0 − θ)/(K − θ)) lim and 1 1 − K log((S0 − θ)/(K − θ)) + K−θ log(S0 /K) θ = . Intercept Using (3. 2 12 (B.
and the third term requires more attention. Convexity = K=S0 θ + O(T ). The ﬁrst term is straightforward to calculate. 2 S0 −θ ˜ 2 S0 (S0 − θ) K→S0 log( K→S0 ∂K 2 ) K−θ we have H −1 ∂ 2 σimp (K − θ) log( S0 ) 1 H c0 K = lim σimp (S0 − θ) .So DH ∂σimp (K) ∂K Consequently the slope is θ H K=S0 = σimp (S0 − θ) 2 + O(T ).20). 2 S0 −θ 2 (S0 − θ)S0 ∂K K→S0 log( K−θ ) (B.21) DH ∂ log σimp (K) ∂ log K iii. ∂K 2 log( S0 −θ ) K−θ S Using H ∂σimp (K−θ) ∂K = O(T ) in (B.22) For the second derivative.23) (b) The third term 82 . 2S 0 (B. using (3.14). 2(S0 − θ) (B. we have DH ∂ 2 σimp (K) H ∂ 2 σimp (K − θ) log( S0 ) K = 2 S0 −θ ∂K log( K−θ ) ∂K 2 + 2 H ∂σimp (K − θ) ∂ ∂K ∂K 0 log( K ) 0 log( K−θ ) S S −θ H + σimp (K − θ) 0 log( K ) ∂2 . we only need to calculate the ﬁrst and K=S0 third term. (a) The ﬁrst term Combining S0 H ˜ −1 ∂ 2 σimp (K) log( K ) 1 H c0 S −θ and lim lim = σimp (S0 −θ) = 0 .
1. 2 2S0 (B. Intercept: The shortexpiry atthemoney implied volatility is easy to get: 2T ADH H 2 2 T (σimp (S0 ))2 = (σimp (θ−S0 )(1−θ/S0 ))2 = (σ0 +κ(µ−σ0 ) − )×(1−θ/S0 )2 +O(T 2 ). = 2 (S0 − θ)S0 K=S0 (B.26) ii.28) 83 .2. θ). we have 0 log( K ) ∂2 H(S0 . (B. Convexity = K=S0 θ + O(T ). θ) := lim K→S0 ∂K 2 log( S0 −θ ) K−θ 0 log( K ) ∂2 − θ) ∂K 2 log( S0 −θ ) K−θ S = 4θ2 − 5θS0 . We will use the relationship between the implied volatilities of the displaced anti Heston model and the Heston model.2. we have K=S0 θ + O(T ). Slope The ﬁrst derivative DH ∂σimp (K) ∂K is calculated as in the case of displaced Heston.1. i.23) and (B. Other parts of the proof are similar ¯ ¯ to those in Proposition 3.With some calculation. 2 12 (B. θ)+O(T ). 2(S0 − θ) (B.25) ∂K 2 Proof of Proposition 3. which is H = −σimp (θ − S0 ) ADH ∂σimp (K) ∂K Consequently. Denote K = θ − K and S = θ − S0 . 3 6S0 (S0 − θ) S ⇒ lim K→S0 H σimp (K H = σimp (S0 − θ)H(S0 .24) gives the second derivative of the implied volatility DH ∂ 2 σimp (K) −1 c0 1 H H σimp (S0 −θ) +σimp (S0 −θ)H(S0 .27) ADH ∂ log σimp (K) ∂ log K iii.24) Combining (B.
29) (b) The third term This is exactly the same as in the displaced Heston case: ∂2 H lim σimp (θ − K) ∂K 2 K→S 0 0 log( K ) S S0 −θ log( K−θ ) H = σimp (θ − S0 )H(S0 . θ)+O(T ). θ).14).30) Combing (B. (B. we have ADH ∂ 2 σimp (K) H ∂ 2 σimp (θ − K) log( S0 ) K = − S0 −θ ∂K 2 log( K−θ ) H ∂σimp (θ − K) ∂ + 2 ∂K ∂K H ∂σimp (θ−K) ∂K 0 log( K ) 0 log( K−θ ) ∂K 2 S S −θ ∂2 H + σimp (θ − K) ∂K 2 0 log( K ) 0 log( K−θ ) S S −θ . Again since third term.31) ∂K 2 =− K=S0 84 . we only need to calculate the ﬁrst and the K=S0 (a) The ﬁrst term Combining S0 H ¯ −1 ∂ 2 σimp (K) log( K ) 1 H c0 S −θ = and lim = 0 σimp (θ−S0 ) . 2 (S0 − θ)S0 imp (B. = O(T ).Use (3.29) and (B. lim = 2 S0 −θ 2 (S0 − θ)S0 ∂K K→S0 log( K−θ ) (B. ¯ 2 S0 ∂K 2 (θ − S0 )2 K→S0 K→S0 log( S0 −θ ) K−θ lim we have H −1 ∂ 2 σimp (θ − K) log( S0 ) 1 H c0 K σimp (θ − S0 ) .30) gives the convexity as ADH ∂ 2 σimp (K) −1 1 H c0 σimp (θ−S0 ) −σ H (θ−S0 )H(S0 .
σu ) − B(u. (C. ˆ ∆t ∆t ˆ∆t Note that St and St coincide at the discrete points tn = n∆t. Su . where γ = inf{t ≥ 0 : σt ∈ Ω2 }. (C. which is τ = ρ ∧ θ ∧ γ. their stopping time is τ = ρ ∧ θ where ρ = inf{t ≥ ∆t / 0 : St ∈ Ω1 }.3) 2 85 . ∆t St = S0 + 0 t ˆ∆t ˆ ∆t A(u. σu ))dWu . for t ∈ [tn . for t ∈ [tn . tn+1 ). To make it complete. First. we sketch the outline of the proof below. / 2. Su .6) as follows. σu )du + t 0 ˆ∆t ˆ ∆t B(u. ∆t σt = σtn . we can write a continuous version of the Euler approximation deﬁned in (6.2) Let T1 ∈ [0. τ ∧ T1 ].APPENDIX C C. Su . We add one more piece to the stopping / time τ . σu ))du t 2 sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (B(u. The proof here has the same ﬂavor as the proof of Theorem 1 in Zhang et al (2004). We substitute the local Lipschitz condition (*) with their local Lipschitz condition (15). we have sup 0≤t≤τ ∧T1 ∆t St E − St 2 t ≤ 2E + 2E sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (A(u. tn+1 ). Su . For any t ∈ [0. except for two major diﬀerences: 1. Su . σu )dWu . (C. In Zhang et al(2004) Theorem 1.1) where we introduce the piecewise constant process ∆t ˆ∆t St = Stn . σu ) − A(u. Su . θ = inf{t ≥ 0 : St ∈ Ω1 }.1 Appendix: Proof of Theorem 12–Partial Strong Convergency of Stochastic Volatility Process Proof. T ] be an arbitrary time.
Su . 86 . we have sup 0≤t≤τ ∧T1 ∆t St − St 2 (C.4) Applying the Doob inequality to the second term of (C. σu ) − A(u. σu )) du. σu ) − B(u. σu )) du. σ ∆t Substituting (C. σu ))2 ˆ∆t ≤ K1 (Ω1 × Ω2 )Su − Su 2 + K2 (Ω1 × Ω2 )ˆu − σu 2 . σu ) − B(u. Su .3) gives t 2E sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (B(u. Su . (C.3) gives t 2E sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (A(u. σu ) − A(u. Su .) and B(.6) E ≤ 2(T + 4C1 )K1 (Ω1 × Ω2 )E = 2(T + 4C1 )K1 (Ω1 × Ω2 )E +K4 (Ω1 × Ω2 )∆t ≤ 4(T + 4C1 )K1 (Ω1 × Ω2 )E +K4 (Ω1 × Ω2 )∆t ≤ 4(T + 4C1 )K1 (Ω1 × Ω2 )E +4(T + 4C1 )K1 (Ω1 × Ω2 ) τ ∧T1 0 τ ∧T1 0 ˆ∆t Su − Su 2 du + K4 (Ω1 × Ω2 )∆t ∆t ∆t ˆ∆t Su − Su + Su − Su 2 du τ ∧T1 0 ∆t ∆t ˆ∆t (Su − Su 2 + Su − Su 2 )du τ ∧T1 0 T1 0 ∆t ˆ∆t Su − Su 2 du + K4 (Ω1 × Ω2 )∆t E sup 0≤u ≤τ ∧u ∆t Su − Su 2 du. Su .5) If A(. σu ) − B(u. σu ) − A(u.) satisfy the local Lipschitz condition (*).6) into (C. Su . Su .Applying the Holder inequality to the ﬁrst term of (C.7) where K4 (Ω1 × Ω2 ) = 2(T + 4C1 )K2 (Ω1 × Ω2 )D(Ω2 )T . σu ))du τ ∧T1 2 ≤ 8C1 E sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (B(u.11). σu ))du τ ∧T1 2 ≤ 2T E sup 0≤t≤τ ∧T1 0 ˆ∆t ˆ ∆t (A(u.4)(C. 2 (C. Su .3) and use condition (6. Su . Su . σu ))2 ∨ (B(u. 2 (C. we have ˆ∆t ˆ ∆t ˆ∆t ˆ ∆t (A(u. Su . Su .
9) and (C. σu )dWu 2 ≤ K3 (Ω1 × Ω2 )∆t. σu )dWu t 2 ˆ∆t ˆ ∆t B(u. Su . By the ∆t ˆ∆t ˆ∆t deﬁnition of St . Su .1). σu )du ˆ∆t ˆ ∆t B(u. σu )du 2 ≤ K3 (Ω1 × Ω2 )∆t2 . Su . we have ∆t ∆t ∆t ˆ∆t ESt − St 2 = ES[t/∆t]∆t − St 2 [t/∆t]∆t =E + 0 0 [t/∆t]∆t ˆ∆t ˆ ∆t A(u.10) If T ∆t < 1.12) 87 .10) leads to ∆t ˆ∆t ESt − St 2 ≤ 2K3 (Ω1 × Ω2 )∆t2 + 2K3 (Ω1 × Ω2 )∆t. then we ∆t could apply the Gronwall inequality which leads to a bound on E sup0≤t≤τ ∧T1 St −St 2 . σu )du − t 0 t 0 ˆ∆t ˆ ∆t A(u. (C. σu )dWu 2 . Su . (C.10) to the second term on the righthand side and using the Ito isometry t t (E( 0 f (Xu )dWu )2 = 0 Ef 2 (Xu )du) leads to t E [t/∆t]∆t ˆ∆t ˆ ∆t B(u. Su . Use (C. where [t/∆t] is the integer part of t/∆t.If the sum of the ﬁrst term and the second term on the righthand side is bounded. we know that St = S[t/∆t]∆t . σu )dWu − ˆ∆t ˆ ∆t A(u. (C.10) to the ﬁrst term on the righthand side and using the Holder inequality leads to t [t/∆t]∆t ˆ∆t ˆ ∆t A(u. (C. Su . σu )du 2 t ≤ 2E [t/∆t]∆t + [t/∆t]∆t ˆ∆t ˆ ∆t B(u. So all we need to prove now is the ﬁrst term on the righthand side is bounded.8) Applying (6. (C.11) Thus t∧T1 E 0 ∆t ˆ∆t Su − Su 2 du ≤ 2T1 K3 (Ω1 × Ω2 )∆t2 + 2T1 K3 (Ω1 × Ω2 )∆t = C0 (Ω1 × Ω2 )∆t. (C.9) Applying (6. Su . Su .
C. Plugging (C.where C0 (Ω1 × Ω2 ) = 2K3 (Ω1 × Ω2 )(1 + T1 )∆t. C2 (Ω1 × Ω2 ) = 4(T + 4C1 )K1 (Ω1 × Ω2 ). we have β σt St − σt (St )β ˆ ∆t ˆ∆t 2 ∆t ∆t = σt St − σtn (Stn )β β 2 = = β σ t St β σ t St ∆t − σtn (Stn )β ∆t − σt (Stn )β β 2 ∆t + σt (Stn )β 2 ∆t − σtn (Stn )β 2 (C. we have E sup 0≤t≤τ ∧T ∆t St − St 2 ≤ C1 (Ω1 × Ω2 )eC2 (Ω1 ×Ω2 )T ∆t = C(Ω1 × Ω2 )∆t. we have E sup 0≤t≤τ ∧T1 ∆t St −St 2 ≤ C1 (Ω1 ×Ω2 )∆t+C2 (Ω1 ×Ω2 )E T1 O E sup 0≤u ≤τ ∧u ∆t Su −Su 2 du (C. 2 When Ω1 is bounded compact set and does not contain 0. for any x.15) 2 ∆t ≤ 2 σt St − σt (Stn )β 2 ∆t = 2σt St − (Stn )β β ∆t ∆t + 2 σt (Stn )β − σtn (Stn )β ∆t + 2(σt − σtn )2 (Stn )2β .16) Thus when Ω1 . y ∈ Ω1 there ∃z ∈ (x. Ω2 are bounded. such that (xβ − y β )2 = (βz β−1 )2 (x − y)2 ≤ C(Ω1 )(x − y)2 . We can show that the coeﬃcients of the logarithm of the SABR process satisfy the local Lipschitz condition (*) in a similar way. so σtn = σtn .12) into (C. Now applying the Gronwall inequality to (C. Using this. 88 . St . B(t.2 Appendix: Coeﬃcients of SABR Satisfy the Local Lipschtiz Condition (*) ∆t The σt of the SABR process can be simulated exactly.7).13).13) where C1 (Ω1 × Ω2 ) = C0 (Ω1 × Ω2 )4(T + 4C1 )K1 (Ω1 × Ω2 ) + K4 (Ω1 × Ω2 ). y). β (C.14) Hence the claim is proved. σt ) = σt St satisﬁes the local Lipschitz condition (*). (C.
2. ∆t Plugging the EVn from (6.3 Appendix: Proof of Proposition 6.17) into (C.19) When 0 < β < 1. Thus we can choose the tangent line of f (x) which pass the point (1.17) and use (C. where a = (1 − λ∆t)2 + βσ 2 ∆t. f (x) = xβ is a concave function.2 and Theorem 13–Strong Convergence of Meanreverting CEV Process Proof of Proposition 6.22) 89 .20) ∆t The Zn derived in this way will be bigger than E(Vn )2 for all n > 1.18). the tangent line of f (x) is above itself. (C. 1 1 (C. b = 2λ∆tµ2 (1 − λ∆t) + λ2 ∆t2 µ2 + (1 − β)σ 2 ∆t.2.C.21) (C. (C. Let Z0 = E(V0∆t )2 and Zn evolves by the following schemes: Zn+1 = (1−λ∆t)2 Zn +2λ∆tµ(1−λ∆t) (1−λ∆t)n (EV0∆t −µ)+µ +λ2 ∆t2 µ2 +σ 2 ∆t β(Zn −1)+1 . the Holder inequality gives ∆t ∆t ∆t ∆t (EVn 2β ) 2β ≤ (EVn 2 ) 2 ⇒ EVn 2β ≤ (EVn 2 )β .2. so ∀x.16) gives the recursive relationship of the second ∆t moment of Vn : ∆t ∆t ∆t ∆t E(Vn+1 )2 = (1−λ∆t)2 E(Vn )2 +2λ∆tµ(1−λ∆t)EVn +λ2 ∆t2 µ2 +σ 2 ∆tEVn 2β . we have ∆t ∆t ∆t E(Vn+1 )2 ≤ (1−λ∆t)2 E(Vn )2 +2λ∆tµ(1−λ∆t) (1−λ∆t)n (EV0∆t −µ)+µ +λ2 ∆t2 µ2 +σ 2 ∆t(E(Vn )2 )β . c = 2λ∆tµ(1 − λ∆t)E(V0∆t − µ).17) For β < 1. Simplify the notation and write Zn as: Zn+1 = aZn + b + crn . (C. such that xβ < β(x − 1) + 1. 1) as g(x) = β(x − 1) + 1.18) (C. Equation (6.
then lim Zn = b .27). (b) ≤ 2 . for t ∈ [tn .28) ˆ ∆t (Vu − Vu )du + σ 90 t 0 ˆ ∆t (Vu β − Vu β )dWu . 1−a (C. Also we have a k−1 Ψk (µ)dµ = 1. the limit of Zn exists. we denote a continuous version of the Euler approximation (6. for ak < µ < ak−1 . for a < x < a k k−1 kx Φk (x) = 0. The proof is in similar spirit to the ones in Higham and Mao (2004). n Using (C. k 3. First.16) as Vt∆t := V0 + t 0 x y (C. Φk (x) := 0 dy 0 Ψk (µ)dµ. (C. Ψk as deﬁned by them: 1. 2. Φk . Ψk (µ) is a continuous function with support in (ak .25) (C. ak = e−k(k+1)/2 .24) (C.It is well established that if a < 1. for all x ∈ R. therefore the ∆t Zn will be bounded and E(Vn )2 will be bounded. a0 = 1.29) . Proof of Theorem 13. we have Vt − Vt∆t = −λ t 0 (C.27) where we introduce the piecewise constant process ˆ Vt∆t := Vt∆t . tn+1 ). For each k ≥ 1.26) ˆ ∆t λ(µ − Vu )du + t 0 ˆ ∆t σVu β dWu (C. otherwise (c) x − ak−1 ≤ Φk (x) ≤ x. We use ak . Three properties of Φk (x) will be used: (a) Φk (x) ≤ 1. ak−1 ) such that a 2 0 ≤ Ψk (µ) ≤ kµ . r < 1.23) n→∞ So when (1 − λ∆t)2 + βσ 2 ∆t < 1 and 1 − λ∆t < 1.
(C.31) Taking the expectation of both sides gives EΦk (Vt − Vt∆t ) = −λE t σ2 ∆t ∆t ˆ ∆t ˆ ∆t Φk (Vu − Vu )(Vu β − Vu β )2 du Φk (Vu − Vu )(Vu − Vu )du + E 2 0 0 t 2 σ ˆ ∆t E(Vu − Vu )du + I(t). ≤λ 2 0 (C.29) into (C.30) 2 Plugging (C.30). k kak 0 t 2 91 . we have Φk (Vt − Vt∆t ) = t 0 ∆t ˆ ∆t Φk (Vu − Vu )(−λ)(Vu − Vu )du + σ t 0 ∆t ˆ ∆t Φk (Vu − Vu )(Vu β − Vu β )dWu + σ2 2 0 t ∆t ˆ ∆t Φk (Vu − Vu )(Vu β − Vu β )2 du. and t I(t) = E ≤ E = E ≤ E 0 t 0 t 0 t 0 ∆t ˆ ∆t Φk (Vu − Vu )(Vu β − Vu β )2 du ∆t ˆ ∆t Φk (Vu − Vu )C(β)Vu − Vu 2β du ∆t ∆t ∆t ˆ ∆t Φk (Vu − Vu )C(β)Vu − Vu + Vu − Vu 2β du ∆t ∆t ∆t ˆ ∆t Φk (Vu − Vu )C(β)C(2β)(Vu − Vu 2β + Vu − Vu 2β )du t ∆t ∆t Φk (Vu − Vu )Vu − Vu 2β du + t 0 ∆t ∆t ˆ ∆t Φk (Vu − Vu )Vu − Vu 2β du = C E ≤ CE t 2 ∆t ˆ ∆t 1(a <V −V ∆t <a ) Vu − Vu 2β−1 du + CE V ∆t − Vu 2β du k k u u kak u k−1 0 0 t 2 2T ∆t ˆ ∆t ≤ C +C EVu − Vu 2β du k 0 kak 2T 2T ≤ C +C (∆tD)β . (C.∆t Applying Ito’s rule to Φk (Vu − Vu ): 1 ∆t ∆t ∆t ∆t ∆t dΦk (Vu − Vu ) = Φk (Vu − Vu )d(Vu − Vu ) + Φk (Vu − Vu )(d(Vu − Vu ))2 .32) t where the ﬁrst inequality comes from using (C.24).
92 . there is a constant C(β). the second to last and third to last inequality uses (C.33) gives the upper bound for EVt − Vt∆t .3. 1].32). we have EΦk (Vt − Vt∆t ) ≤ λ σ2T σ2T ∆t ˆ ∆t EVu − Vu du + C +C (∆tD)β k kak 0 0 2 2 t ∆t du + λ(∆tD)1/2 + C σ T + C σ T (∆tD)β .25) and the last inequality use Lemma C. EVu − Vu ≤ λ k kak 0 t ∆t EVu − Vu du + λ t where the second inequality uses Lemma C. then Y  1 − p = X+Y  .2. The function f (p) = pβ + (1 − p)β is continuous on the interval [0.22).34) (C. such that ∀X. (C. Y .where the ﬁrst and second inequality use x + yβ ≤ C(xβ + yβ ) as proved in Lemma C. 1 So if we can show C(β) ≤ ( X+Y  )β + ( X+Y  )β . Let p = X+Y  .33) k kak 0 Applying the Gronwall’s inequality to (C. This proves (6. the minimum is bigger than 0. Combining this with the lefthand side of (C. Plugging the above results into (C. σ2T σ2T sup EVt − Vt∆t  ≤ eλT ak + C +C (∆tD)β + λ(∆tD)1/2 .3. we are done. so it X Y  X (C. X + Y β ≤ C(β)(Xβ + Y β ). So there is C(β) such that the inequality holds. Proof.2 again. Lemma C.3. k kak 0≤t≤T The terms in the parenthesis went to 0 as ∆t → 0. ∀β > 0.1.1. Moreover. Note that X + Y  ≤ X + Y  ⇒ ∀β > 0.35) can achieve the minimum.3.26) gives EVt − Vt∆t  ≤ ak−1 + C t σ2T σ2T +C (∆tD)β + λ(∆tD)1/2 + λ E(Vt − Vt∆t )du. X + Y β ≤ (X + Y )β .
we have ˆ (Vt∆t − Vt∆t )2 = ∆t ∆t (t − tk )λ(µ − Vk ) + σVk β (Wt − Wtk ) 2 √ ∆tD. Under conditions A and B. we have ˆ EVt∆t − Vt∆t 2β ≤ (∆tD)β . the Holder inequality gives ˆ ˆ EVt∆t − Vt∆t 2β ≤ (EVt∆t − Vt∆t 2 )β . From (C.2. Taking the expectation of both sides gives ∆t ∆t ˆ E(Vt∆t − Vt∆t )2 = (t − tk )2 λ2 E(µ − Vk )2 + σ 2 (t − tk )EVk 2β ≤ ∆t2 λ2 E(µ − Sk )2 + σ 2 ∆tESk 2β 2 = ∆t ∆tλ(µ2 − 2µESk + ESk ) + σ 2 ESk 2β 2 ≤ ∆t λ(µ2 − 2µESk + ESk ) + σ 2 ESk 2β ≤ ∆tD where the last inequality uses Proposition 6.27).2.Lemma C. there is a constant D such that ˆ EVt∆t − Vt∆t 2β ≤ (∆tD)β . ˆ so it is enough to show that there is a constant D such that E(Vt∆t − Vt∆t )2 ≤ ∆tD. (C. ˆ ˆ Finally. Combining with (C.2. tk+1 ).37) (C. (C.38) ∆t ∆t = (t − tk )2 λ2 (µ − Vk )2 + σ 2 Vk 2β (Wt − Wtk )2 ∆t ∆t +2(t − tk )λ(µ − Vk )σVk β (Wt − Wtk ). When β < 1.38).36) ˆ EVt∆t − Vt∆t  ≤ Proof. EVt∆t − Vt∆t  ≤ (E(Vt∆t − Vt∆t )2 ) 2 ≤ 1 √ ∆tD. 93 . for t ∈ [tk .3.
T ) ≥ (−x) 2 ∞ ≤ (−x) ≤ x ≥ x 1 if x > σ 2 2 1 if x ∈ [0. T ) := 2(2A + x + 2√A2 + Ax) we have 1 if x ∈ R/[− 2 σ 2 .APPENDIX D D.1 Appendix: Proof of Theorem 14 –Largestrike and Largeexpiry Asymptotic of Displaced Lognormal Lemma D. 0] 2 1 2 if x < − σ 2 (D.5) 94 . 2σ 2 (D.3) 2 2 2 Proof. Note that σ∞ (x) deﬁned in (D.1.1) is in fact the solution of the function (−x + σ∞ (x)2 /2)2 .3). σ) (D. T ) < (−x) 2 ∞ < x 1 if x > σ 2 2 1 if x < − σ 2 2 (D.1. deﬁne √ 2(2A + x − 2 A2 + Ax) 2 σ∞ (x.4) (x + σ 2 /2)2 = . σ 2 ] 2 1 2 if x ∈ [− σ .1) 1 2 σ (x. A= 2σ∞ (x)2 Denote ∗ VBS (x. We suppress the T in σ∞ (x. 1 σ 2 ] 2 (D. T ∈ Kx . the equalities in the ﬁrst and last equation in (D. the strict inequalities hold and we have 1 2 σ (x. For A deﬁned in (7.2) For each x. T ) to be σ∞ (x). so for T large enough. 1 σ 2 ] 2 if x ∈ 1 [− 2 σ 2 .2) hold only at most ﬁnite T .
[Proof of Theorem 14. 2 ∞ The equality hold if and only if A + x = at two T ’s. we choose σ∞ (x) to be the smaller root which satisﬁes (D. which happens at most at one T .12) Proof. (−˜+σ 2 /2)2 x σ2 (D.11) 1 2 σ (x) − (−x) = 2(A + x) + 2 2 ∞ A2 + Ax ≥ 0. σ∞ (x)) − x. we choose σ∞ (x) to be the larger root which satisﬁes (D.9) + x = 0. happens at most 1 When x ∈ [− 2 σ 2 . using A = 2 2σ (D.10) 1 2 σ (x) − x = 2A + 2 2 ∞ 1 If − 2 σ 2 ≤ x < 0. (D. which again.] 95 . If x < − 2 σ 2 < 0. we have If x > 2 σ 2 > 0. (−˜+σ 2 /2)2 x 1 ≥ 0. 2 ∞ 2 (x+σ∞ /2)2 2 2σ∞ (D. ˜ (D. σ) − x = VBS (x.4) can be written as ˜ ∗ x ∗ VBS (˜. (D. 2 σ 2 ].4). 1 σ 2 ]. so 2 2 σ∞ (x) = 2(2A + x + 2 1 If 0 ≤ x ≤ 2 σ 2 . σ) − x and (D. using A + x ≥ 0 gives A2 + Ax ≥ 0. (D.7) 1 2 σ (x) − x = 2A − 2 A2 + Ax ≤ 0. we have 1 2 σ (x) − (−x) = 2(A + x) − 2 A2 + Ax ≤ 0.8) 1 The equality holds if and only if x = σ 2 /2. using A ≥ 0 gives A2 + Ax). ˜ using A + x = ≥ 0 and 0 ≤ A + x < A. so 2 σ∞ (x) = 2(2A + x − 2 A2 + Ax).6) 1 1 When x ∈ R/[− 2 σ 2 .∗ x then A = VBS (˜.4).
as z → +∞. For S from displaced lognormal dynamics.18) ˜ S0 A (˜.7). x > σ 2 /2.13) Recall an approximation of the cumulative density function of standard normal N (z) from Olver(1974): N (−z) = 1 − N (z) = exp(−z 2 /2) √ (1 + O(1/z 2 )). we have ˜ N (z+ ) − exT N (z− ) =√ ˜ σ 1 σ 1 x 4 − 1 ( x −˜+ σ )T − e 2 σ2 (1 + O( )) 2 /2 2 /2 T x−σ ˜ x+σ ˜ 2πT 2 2 2 2 (˜−σ /2) x 1 1 T − 2σ 2 ABS (˜. denote ˜ ˜ S = S − θ. 0) = ABS (x. Then as T → ∞. σ∞ (x).14) and ABS in (7. there is a T0 such that when T > T0 . z 2π (D. a1 (x)). 0)(1 + O( )). 0)(1 + O( )). σ.17) (D. σ∞ (x)) − x. x T 2πT (D.14) When x > σ 2 /2. ˜ 1 S 1 ˜ ˜ ˜ E(ST − S0 exT )+ = 0 E(ST − S0 exT )+ ˜0 S0 S0 S ˜ ∗ x 1 1 S x e−(VBS (˜. x = 0√ S0 2πT T We choose σ∞ (x) and a1 so that ˆ ∗ x ∗ VBS (˜.σ)−˜)T ABS (˜. σ. σ.σ)−˜)T ABS (˜. x ˆ S0 BS 96 . σ.15) Consequently. ˜0 S where z± = √ (−˜±σ 2 /2) T x . we ˜ have z± → −∞. σ) − x = VBS (x. σ (D. then S will be a lognormal process and 1 ˜ ˜ ˜ ˜ E(ST − S0 exT )+ = N (z+ ) − exT N (z− ). ˜ (D.We ﬁrst show the case when x > σ 2 /2. Using (D.16) (D. 0)(1 + O( )) x =√ e T 2πT ∗ x 1 1 x =√ e−(VBS (˜.
1 ABS (x.1. limT →∞ a1 (x) ˆ2 converges to a constant.14). From (D.20) Intuitively. S0 exT . σ∞ . σ∞ (x). σT . a1 ) ˆ ∗ √ E(ST − S0 exT )+ = exp(−(VBS (x. ˆ =√ T 2πT (D. σ∞ ) − x)T )(1 + O(1/T )) (D. we have ∗ 1 1 1 E(ST −S0 exT )+ = √ e−(VBS (x. comparing (D.1 we know that x > σ∞ (x)/2. Lemma D.19) and (D. So for T large enough.17) and (D.23) 2πT 4x2 − 1) δ > 0. By Lemma D.14) to show that (D.σ∞ (x))−x)T ABS (x. ˆ (D. we have x > σT /2.20). a 8 σ∞ 97 .19) is an asymptotic formula for the BlackScholes call price with volatility 2 σ∞ (x) + a1 (x)/T . (D. T ) = σ∞ (x) + a1 (x)/T + o(1/T ).Plugging (D. σ∞ (x). (D. ˆ a (x) ˆ 2 2 ˆ Denote σT = σ∞ (x)+ 1T . σ∞ ) − x)T )e .19) ˆ S0 T 2πT Now if we could show that −x + (σ∞ (x)2 + a1 (x)/T )/2 > 0 for large enough T . σ∞ . Thus.19). we can see that the implied volatility of the displaced lognormal process is 2 2 σimp (x. the Blackˆ2 Scholes call price formula can be approximated as C BS (S0 .16). ∀ > 0. a1 ) ˆ ∗ √ ≤ exp(−(VBS (x. a1 (x))(1 + O( )).21) The rigorous proof is as follows.1 shows that (δ) = ( 4 16 σ∞ (x) > 0 in the previous step so that − (δ) < 0. T ) = √ ˆ σ2 ˆ 1 x2 σT ˆ σT ˆ 1 exp − ( 2 − x + T )T − (1 + O(1/T )) 2 /2 2 σT 4 ˆ x − σT ˆ x + σT /2 ˆ2 2πT ∗ 1 1 e−(VBS (x. Using (D. we can choose For any δ > 0.18) into (D.1. a1 (x))(1+O( )).σ∞ (x))−x)T ABS (x.22) S0 2πT ABS (x. and exp 1 4x2 a1 (x)( 2 − 1) + ˆ 8 σ∞ ≤ exp 1 4x2 (ˆ1 (x) + δ)( 2 − 1) + − (δ) . then we ˆ could apply (D.
20). a Similarly. σ∞ ) − x)T + − (δ) E(ST − S0 exT )+ ≤ S0 2πT 2 a ≤ C BS (S0 .1. and apply similar methods. [Proof of Theorem 15. we consider the put price on the displaced lognormal dynamics. so the equality in (7. σ∞ (x) + (ˆ1 (x) + δ)/T.] 98 . For x ∈ (−σ 2 /2. S0 exT . a1 + δ) ˆ ∗ √ exp − (VBS (x. More speciﬁcally. the implied volatility of the displaced lognormal process σimp (x) is 2 2 σimp (x. 1 E(ST − S0 exT )+ S0 ˜ S 1 ˜ ˜ ˜ = 0 E(ST − S0 exT )+ ˜ S0 S0 2 ˜2 x ˜ S0 1 σ 1 σ − 1 ( x2 −˜+ σ )T 2 σ 4 1+ √ − (1 + O( )) = e 2 /2 2 /2 S0 T x−σ ˜ x+σ ˜ 2πT 2 2 ˜ ˜ θ S 1 σ σ 1 − 1 ( x −˜+ σ )T x 4 + 0√ e 2 σ2 − (1 + O( )) =1− S0 S0 2πT T x − σ 2 /2 x + σ 2 /2 ˜ ˜ (D. ∀δ > 0. For the x < −σ 2 /2 case. T ) ≤ σ∞ (x) + (ˆ1 (x) + δ)/T. 1 ABS (x.2. we can prove the lower bound.) to get the results.5) holds.24) T →∞ Proof. T ) Using the monotonicity of the BlackScholes formula. σ 2 /2). σ∞ .25) Proof.23) and (D.2 Appendix: Proof of Theorem 15–First Order Approximation of Largestrike Largeexpiry Implied Volatility Lemma D. Apply the asymptotic approximation of N (. we have lim E(ST − S0 exT )+ = S0 − θ. (D. there exists T1 .Using (D. D. such that when T > T1 .
limT →∞ E(S0 − S0 exT ) = S0 − θ > S0 . Deﬁne σ (T ) as S0 σ (T ) := ¯ B+ √ B 2 + 2xT √ . σimp (T )) = S0 − θ. S0 (B = N 2x T 2x + 2B 2x a(T ) + . it satisﬁes the equation: ¯ ¯ −x + σ 2 /2 √ ¯ T = B. 1 BS C (S0 . so σimp (T ) is not well deﬁned when T is large. so lim σimp (x.28) D.29) 99 .T) T T − N (B) + N (B) − exT N (−2x − B 2x + 2B 2x T a(T ) √ − 2T ) T 2x T a(T ) √ + 2T ) T 2x T 2x + 2B + T a(T ) + T a(T ) (D. ¯ T →∞ T →∞ (D. σ . S0 exT . T S −θ Abbreviate σ (T ) as σ .When θ < 0.27) Recall that limT →∞ C BS (S0 . T ) = lim σ (T ) = 2x. σ ¯ σ ¯ T →∞ T →∞ (D.3 Appendix: Proof of Theorem 16–Asymptotic Formula of BlackScholes Call Option Proof. So √ √ (−x + σ 2 /2) T ¯ (−x − σ 2 /2) T ¯ lim C BS (S0 .26) √ = limT →∞ − B 2 + 2xT = −∞ . T ) = lim S0 N ¯ −S0 exT N = S0 −θ. ¯ When θ > 0. S0 exT . denote B := N −1 ( 0 ). S0 exT . σ ¯ and √ (−x−¯ 2 /2) T σ limT →∞ σ ¯ (D.
31) into (D. we have 1 BS 2x a(T ) C (S0 . using (D.29).30) For the last term.T) S0 T T √ 1 2 1 1 = N (B) + √ e− 2 B √ (a(T )/2 − B 2 − 1)(1 + O(1/ T )).14) gives (−2x − B 2x + 2B 1 =√ exp 2πT 1 − 2 2x T a(T ) √ − 2T ) T 2x T exT N + T a(T ) a(T ) √ + 2T ) T 2x T a(T ) a(T ) (B 2x T 2 2x + 2B 2x + B + T √ 1 1 1 =√ exp(− B 2 ) √ (1 + O(1/ T )).31) Plugging (D. S0 exT . 2x + 2B + . 2x 2πT (D. 2 2x 2πT 2x + 2B 1 (1 + O( ) a(T ) T 2x/T + 2T 2x T + T (D.32) 100 .Applying Taylor expansion to the ﬁrst term.30) and (D. we have (B N 2x T a(T ) √ + 2T ) T 2x T − N (B) a(T ) √ + 2T ) T 2x T a(T ) √ + 2T ) T 2x T 2x + 2B 1 =√ exp 2πT + a(T ) T − 1 2 (B 2x T 2 (B 2x T −B √ 2x + 2B + T a(T ) 2x + 2B + T a(T ) 1 T (1 + O( )) T √ 1 1 a(T )/2 − B 2 √ =√ exp(− B 2 ) (1 + O(1/ T )). 2 2x 2πT (D.
T ) ≥ 2x + 2B 2x a(T ) − δ + . ∀δ > 0.4 Appendix: Proof of Theorems 17 and 18– Second and Third Order Approximation of Largestrike Largeexpiry Implied Volatility Proof of Theorem 17. ∀T > T0 (a(T )/2 − B 2 − 1)e < ((a(T ) + δ)/2 − B 2 − 1)e− (δ) . T ).33) From the deﬁnition of a(T ). 1 E(ST − S0 exT )+ S0 2 ˜2 x ˜ θ S0 1 σ 1 σ − 1 ( x2 −˜+ σ )T 2 σ 4 √ =1− + e (1 + O( )) − 2 /2 2 /2 S0 S0 2πT T x−σ ˜ x+σ ˜ 2 2 ˜ ˜ θ S 1 σ σ − 1 ( x −˜+ σ )T x 4 ≤1− + 0√ e 2 σ2 − e 2 /2 S0 S0 2πT x−σ ˜ x + σ 2 /2 ˜ 1 1 1 = N (B) + √ exp(− B 2 ) √ (a(T )/2 − B 2 − 1)e 2 2x 2πT (D. S0 2x + 2B 2x a(T ) + δ + . ∀ > 0.25) and the deﬁnition of a(T ) and B. T T 101 .35) 1 BS C (S0 .34) (D. we have ∀ > 0. (D.34) and proposition 16. S0 exT . T T Similarly. Using (D.33).D. we have 1 E(ST − S0 exT )+ S0 1 1 1 ≤ N (B) + √ exp(− B 2 ) √ (a(T )/2 − B 2 − 1)e 2 2x 2πT 1 2 1 1 exp(− B ) √ ((a(T ) + δ)/2 − B 2 − 1)e− (δ) ≤ N (B) + √ 2 2x 2πT ≤ So 2 σimp (x. there ∃ (δ) > 0. such that ∃T0 . T ) ≤ 2x + 2B (D. T T 2x a(T ) + δ + . we can show that 2 σimp (x. we know that a(T ) > B 2 + 1. Using (D.
σ∞ = . T ) = S0 (2N (σimp T /2) − 1). S0 exT . D. for the displaced lognormal.14). The rigorous proof is similar to the proof of α 1 log( S0 exT −θ ). S − θ is lognormal process. T ) 2 2 smaller root σ∞ = 2(M + A) 2 imp 102 . S0 2S0 T Taking the limit of T . D. √ √ S −θ θ N (σ T /2) + . So √ √ S0 (2N (σimp T /2) − 1) = (S0 − θ)(2N (σ T /2) − 1). σimp (T ). Proof of Theorem 18. M = ˜ Theorem 17. On the other hand. Choose the σ∞ σ √ σ2 − 2M A + A2 so that ∞ − M < 0. Abbreviate σ (x. we have (7. When atthemoney. Denote x = T α ˜ S −θ xT α−1 and A = ˜ ˜ −M +σ 2 /2 σ 0 2 2 /2 2 ˜ −M +σ 2 /2 2 satisﬁes −M +σ∞ .6 Appendix: Proof of Theorem 20–Approximation of Implied Volatility when K = S0 exp(xT α ) Proof. We prove the x > 0 case. we have √ C BS (S0 . T ) to be σimp (T ). M = xT α−1 . Similar to the proof of Theorem 17.5 Appendix: Proof of Theorem 19–Largestrike Largeexpiry Atthemoney Implied Volatility Proof.This completes the proof. so √ E(ST − K)+ = E((ST − θ) − (K − θ))+ = (S0 − θ)(2N (σ T /2) − 1). ⇒ N (σimp (T ) T /2) = 0 S0 S0 √ 2 θ S −θ ⇒ σimp (T ) = √ N −1 0 N (σ T /2) + . Suppress the 0 in σimp (0. Here we sketch the outline of the proof.
S0 xT α−1 − σimp xT α−1 2 + σimp /2 (1 + o(1/T )) (D.to be σimp .7. D. lim E(ST −K)+ = lim S0 ×N 2 2 T →∞ T →∞ σimp (T ) T σimp (T ) T (D. we have α 1 E(ST − S0 exT )+ S0 ˜ S 1 ˜ ˜ ˜ α = 0 E(ST − S0 exT )+ ˜ S0 S0 √ √ ˜ (−˜T α−1 + σ 2 /2) T x x α−1 − σ 2 /2) T S0 xT α N (−˜T N −e = S0 σ σ 2 ˜ S 1 −˜T α−1 + σ 2 /2 1 x σ σ = 0√ exp − (1 + o(1/T )) T − α−1 − σ 2 /2 α−1 + σ 2 /2 S0 2πT 2 σ xT ˜ xT ˜ 2 ˜ S0 1 1 −M + σ∞ /2 2 ˜ √ = exp − T ABS (M .7 Appendix: Proof of Theorem 21–Fixedstrike Largeexpiry Implied Volatility D. we have √ √ log(S0 /K) σimp (T ) T log(S0 /K) σimp (T ) T √ + √ − −K×N . On the one hand. T ).). S0 exT . T ) to be σimp (T ).36) The x < 0 case can be similarly proved. σimp . σ∞ . Using the approximation of N (. Suppress the x in the σimp (x.1 Asymptotic behavior Proof.37) 103 . 0)(1 + o(1/T )) 8 σ∞ σimp 2 − σimp /2 2 1 −M + σimp /2 2 1 T exp − =√ 2 σimp 2πT α 1 BS → C (S0 . 0)(1 + o(1/T )) S0 2πT 2 σ∞ 1 exp =√ 2πT 2 1 −M + σ∞ /2 2 − T 2 σ∞ exp a 4M 2 ( 2 ) − 1 ABS (M. σ.
On the other hand. K. P ) in (7. T ).38) = S0 − θ. K. √ Comparing (D. Diﬀerentiating both sides with respect to T gives us ∂C BS (S0 . K − θ. K. K.41) 104 .39) and (D.37) and (D.40). ∂T ∂σimp (T ) ∂T ∂T (D. K. T ) = C BS (S0 − θ.39) Recall the deﬁnition of σblsimpv (S0 . P ) 2 σblsimpv (S0 . σ.42) (D. K.2 Monotonicity Proof. Recall the σimp (T ) is deﬁned as follows C BS (S0 . P ) σblsimpv (S0 . σimp (T ). σ) + = .38). P ) 2 (D. σimp (T ). lim E(ST − K)+ = lim E((ST − θ) − (K − θ))+ T →∞ T →∞ = lim (S0 − θ)N ( log((S0 − θ)/(K − θ)) + σ 2 T /2 √ ) T →∞ σ T log((S0 − θ)/(K − θ)) − σ 2 T /2 √ − (K − θ)N ( ) σ T (D. P ) log(S0 /K) log(S0 /K) + − −K×N σblsimpv (S0 . we know that the limit of σimp (T ) T exists and satisfy the following condition √ log(S0 /K) σimp (T ) T √ + lim S0 × N 2 T →∞ σimp (T ) T √ log(S0 /K) σimp (T ) T √ − −K ×N 2 σimp (T ) T = S0 − θ.40) = P. we have (7. σimp (T ). D. K − θ.21) is S0 ×N σblsimpv (S0 . K. (D. K. T ) ∂σimp (T ) ∂C BS (S0 − θ. T ) ∂C BS (S0 . comparing (D.7.20).
lim sgn ∂σimp (T ) < 0. 2 σimp (T ) T 2 (D. S0 2 4 (D. K. we have ∂σimp (T ) sgn = sgn ∂T ∂C BS (S0 − θ. ∂T T →∞ 105 .43) From (D.1. we have ∂σimp (T ) lim sgn = lim sgn (S0 − θ) exp ∂T T →∞ T →∞ − S0 exp √ 1 log((S0 − θ)/(K − θ)) σ T 2 √ √ − + σ T 2 2 σ T √ √ σimp (T ) T 2 1 log(S0 /K) √ + σimp (T ) T − 2 σimp (T ) T 2 √ K − θ 1/2 1 σ2T = lim sgn exp − σ T S0 − θ 2 4 T →∞ 2 T 1/2 √ K 1 σimp exp − − σimp (T ) T .43).Using ∂C BS (S0 .σimp (T ). σ) ∂C BS (S0 . σimp (T )) − ∂T ∂T √ log((S0 − θ)/(K − θ)) σ T √ = sgn (S0 − θ)N + σ 2 σ T √ σimp (T ) T log(S0 /K) √ + − S0 N σimp (T ) 2 σimp (T ) T √ 1 log((S0 − θ)/(K − θ)) σ T 2 √ + σ = sgn (S0 − θ) exp − 2 2 σ T √ σimp (T ) T 2 1 log(S0 /K) √ + − S0 exp − σimp (T ) .44) √ 1 2 − 2 σ 4T σ T as T → ∞ is dominated by the exponential term. From Theorem 21. limT →∞ σimp (T ) T exists.K.T ) ∂σimp (T ) > 0. so the ﬁrst √ term goes to 0. Therefore. K − θ. so the second term goes to Note that exp some positive constant.
we have 1 1 BS E(S0 − K)+ = C (S0 − θ.47) 106 .S0 −θ) . 2 Taylor approximation of N (.) to the second order. we have 1 E(S0 − K)+ S0 ˜ S −x + a + rT −x + a −x − a − rT −x − a = 0 + N( √ ) − N( √ ) + ex N ( √ ) − N( √ ) S0 2a + 2rT 2a + 2rT 2a 2a ˜ √ √ 1 1 −x + a 2 S 2a + 2rT − 2a ) ) = 0 + √ exp(− ( √ S0 2 2π 2a 2 √ √ x2 a + rT 1 1 + 2a + 2rT − 2a − − + o((rT )2 ) 4a(a + rT ) a 2 8 (D.46) On the other hand. σ. (D. then 1 E(S0 − K)+ S0 ˜ 1 −x + a 2 z 1 1 x2 S √ √ + ) ) z + z2 = 0 + √ exp(− ( √ S0 2 2π 2a 2a(z + 2a)2 2a 2 ˜ S 1 −x + a 2 1 1 x2 ≈ 0 + √ exp(− ( √ ) ) (− + 2 )z 2 + z S0 2 8 8a 2π 2a − 1 8 + o((rT )2 ) (D. T ) is σblsimpv (S0 . we know that the ﬁrst term of √ σimp (x. K. K − θ.K.47) and solve for z. T ) = 2a + 2r. From Theorem 21. So denote σ 2 (x. Denote a = 2 σblsimpv (S0 . T ) S0 S0 Compare (D.46) with (D.D. Here we want a more reﬁned approxima˜ tion of the residual term.8 Appendix: Proof of Theorem 22–Approximation of FixedStrike Largeexpiry Implied Volatility.45) Let √ 2a + 2rT − √ 2a = z. Denote S0 = S0 − θ and use the imp T Proof. S0 − θ)/ T = 2a/T .
: Going to extremes: Correcting simulation bias in exotic option valuation. M. 12671321 (1997) [6] Brigo. 949965 (2001) 107 . 57.: Monte Carlo methods for security pricing. F.. Financial Anal. Dyn. (2004) [8] Broadie. I. Manag.. M. 151174. Linetsky. 13521373 (2004) [4] Black. Math. J. D. Commun. Sci. 6268 (1997) [2] Berestycki. V. I. Dybvig. P. Finance 2. O. Florent. D. Quant. P. Res. 764779 (2005) [10] Cox.: A doubleexponential fast Gauss transform algorithm for pricing discrete pathdependent options. Heidelberg (2001) [7] Broadie. J. J. Florent. Yamamoto. Pure Appl. Res. 1517 (1996) [11] Davydov. Econ.. Mathematical Finance – Bachelier Congress 2000.. Y. 54(2).: Pricing and hedging pathdependent options under the cev process. Kaya. 53(1). 23.. H. J. Glasserman.: Computing the implied volatility in stochastic volatility models. H. P.: Displaced and mixture diﬀusions for analyticallytractable smile models.: The constant elasticity of variance option pricing model.. Busca. Oper. Proceedings of the 2004 Winter Simulation Conference. Portf.. Manag. Scholes. Control 21. Zhou.: Exact simulation of option greeks under stochastic volatility and jump diﬀusion models..: Asymptotics and calibration of local volatility models. M. G. Springer Finance. M. Springer. 217231 (2006) [9] Broadie. Political Econ. Oper. M. D. 6169 (2002) [3] Berestycki. J. 47(7).: The pricing of options and corporate liabilities.... F. J.. Mercurio. Kaya. 637659 (1973) [5] Boyle. 81. Busca.. O. 53(5). Broadie.: Exact simulation of stochastic volatility and other aﬃne jump diﬀusion processes.. J.REFERENCES [1] Beaglehole.
1(8).: The eﬀect of model risk on the valuation of barrier options.[12] Durrleman. Jacquier. U. A. Tataru. Bloomberg Markets Apr 2010. J. Linetsky. Wilmott. M. deﬁnition and implementation. Math.: Positive numerical integration of stochastic diﬀerential equations. Madan. J. 20762088. A. Lesniewski. 147157 (1996) [21] Heston. (2008).: Convergence of the Monte Carlo simulations involving the meanreverting square root process. 6(2). Mao. A. M.: Asymptotic formulae for implied volatility in the Heston model. D..: A displaceddiﬀusion stochastic volatility LIBOR market model: motivation. Princeton (2004) [13] Feng. Wystup.: Mixing Models. Comput. The Review of Financial Studies.. 458469 (2003) [25] Kahl.. 84108 (2002) [20] Hesterberg. Woodward.: Pricing discretely monitored Asian options under Levy Processes. Dept of Operational Resesarch and Financial Engineering. G..: Pricing discretely monitored barrier options and defaultable bonds in L´vy process models. submitted for publication. Jacuier.: Managing smile risk. (2008) [19] Hagan.. J. 32... Mijatovic. [18] Griebsch. A.. R. Diploma Thesis. Finance 18(3). T. [17] Fusai.: From implied to spot volatilities. Stat. Finance 3. A.: The largematurity smile for the Heston Model. 4755 (2003) [24] Joshi. L. G. A. Kumar.. Courtadon. C. Meucci. Risk Finance 4. M.. C. 337384 (2008) e [14] Fisher. [16] Forde. 327343 (1993) [22] Higham. PhD dissertation.. D. S.. 6. V. P. Rebonato. S. T. Comput. G. Quant.. University of Wuppertal (2004) 108 . 136137 [15] Forde. D.: A ClosedForm Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options. Finance 8(3) 3562 (2005) [23] Hirsa. V.: On the valuation of fader and discrete barrier options in Heston’s stochastic volatility model. Journal of Banking and Finance. X. D.: Control variate and importance sampling for eﬃcient boostrap sampling..
. R.: Convergence of the Euler scheme for stocahstic functional partial diﬀerential equations. G. University of New South Wales (2007) [37] Rubinstein. R.: Option Valuation under Stochastic Volatility. Appl. Comput. Kou.: Numerical solution of stochastic diﬀerential equations. D. J. Finance Press. John Wiley & Sons. P. Posner. Appl. D. M. 213217 (1983) [38] SvobodaGreenwood. Q.. M. the sum of lognormals. Koekkoek.: A note on the behaviour of the BlackScholes implied volatility close to expiry. W. DOI 10. University of New South Wales (2009) [36] Roper.. Math. E. S. S. Finance 8. M. [39] Zhang.: Financial option pricing and skewed volatility. Dijk. (1998). SpringerVerlag (1992) [28] Lee. A.: Displaced lognormal volatility skews: analysis and applications to stochastic volatility simulations. M. Ann. Jour of Financial and Quantitative Analysis. R... 497492 (2004) 109 . J. working paper (2006) [27] Kloeden.: Numerical pricing of discrete barrier and lookback options via Laplace transforms. Statistical Laboratory. USA (2000) [30] Lord.: A comparison of biased simulation schemes for stochastic volatility models.. [33] Petrella. Platen. Math. Comput 155.[26] Kahl. and the reciprocal Gamma distribution. Jackel. Finance (2009) Forthcoming. Finance 38(1). University of Cambridge (1999) [32] Milevsky. Volatility and Correlation.. Wang. 2nd edition (2004) [35] Roper.: Fast astrong approximation Monte Carlo Schemes for stochastic volatiity models.: Asian options. V. C.. P. 137 (2004) [34] Rebonato. Vol 33(3). M. Rutkowski. R. Zhang.1007/s1043600901457 [29] Lewis. Finance (2009). S. PhD Thesis. California. D.. Nie Z. Working Paper (2008) [31] Marris.: Implied volatility: General properties and asymptotics. Newport Beach. A.: Displaced diﬀusion as an approximation of the constant elasticity of variance..: Displaced diﬀusion option pricing. MPhil Thesis. E.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.