Beruflich Dokumente
Kultur Dokumente
June 2016
TABLE OF CONTENTS
The note is prepared for the purpose of summarizing local volatility models frequently
encountered in derivative pricing. It at first derives the Kolmogorov forward and backward equations,
which fundamentally govern the transition probability density of the diffusion process in derivative price
dynamics. Subsequently, it introduces the local volatility model in the context of Dupire formula and
then presents a PDE based local volatility model, in which the local volatility function is parametrized to
The time evolution of the transition probability density function is governed by Kolmogorov
forward and backward equations, which will be introduced as follows, without loss of generality, in
multi-dimension.
dimensional Brownian motion 𝑊𝑡 whose correlation matrix 𝜌 is given by 𝜌𝑑𝑡 = 𝑑𝑊𝑡 𝑑𝑊𝑡′
1 1
𝑑ℎ(𝑆𝑡 ) = 𝐽ℎ 𝑑𝑆𝑡 + 𝑑𝑆𝑡′ 𝐻ℎ 𝑑𝑆𝑡 = 𝐽ℎ 𝐴𝑑𝑡 + 𝐽ℎ 𝐵𝑑𝑊𝑡 + 𝑑𝑊𝑡′ 𝐵′ 𝐻ℎ 𝐵𝑑𝑊𝑡 (2)
1×1 1×𝑚𝑚×1 2 1×𝑚𝑚×𝑚𝑚×1 2
where 𝐽ℎ is the 1 × 𝑚 Jacobian (i.e. the same as gradient if ℎ is a scalar-valued function) and 𝐻ℎ the
𝜕ℎ 𝜕2ℎ
[𝐽ℎ ]𝑖 = and [𝐻ℎ ]𝑖𝑗 = (3)
𝜕𝑆𝑖 𝜕𝑆𝑖 𝜕𝑆𝑗
2
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
𝑚 𝑚 𝑛 𝑚 𝑛
𝜕ℎ 𝜕ℎ 1 𝜕2ℎ
𝑑ℎ = ∑ 𝐴𝑖 𝑑𝑡 + ∑ ∑ 𝐵𝑖𝑘 𝑑𝑊𝑘 + ∑ ∑ 𝐵𝑖𝑘 𝜌𝑖𝑗 𝐵𝑗𝑘 𝑑𝑡
𝜕𝑆𝑖 𝜕𝑆𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗
𝑖=1 𝑖=1 𝑘=1 𝑖,𝑗=1 𝑘=1
(4)
𝑚 𝑚 𝑚 𝑛
2
𝜕ℎ 1 𝜕 ℎ 𝜕ℎ
= (∑ 𝐴𝑖 + ∑ 𝛴𝑖𝑗 ) 𝑑𝑡 + ∑ ∑ 𝐵𝑖𝑘 𝑑𝑊𝑘
𝜕𝑆𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗 𝜕𝑆𝑖
𝑖=1 𝑖,𝑗=1 𝑖=1 𝑘=1
𝑚 𝑚 𝑚 𝑛
𝑇 𝑇
𝜕ℎ 1 𝜕2ℎ 𝜕ℎ
ℎ(𝑆𝑇 ) − ℎ(𝑆𝑡 ) = ∫ (∑ 𝐴𝑖 + ∑ 𝛴𝑖𝑗 ) 𝑑𝑢 + ∫ ∑ ∑ 𝐵𝑖𝑘 𝑑𝑊𝑘 (5)
𝑡 𝜕𝑆𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗 𝑡 𝜕𝑆𝑖
𝑖=1 𝑖,𝑗=1 𝑖=1 𝑘=1
Taking expectation on both sides of (5), we get (using notation 𝔼𝑡 [∙] = 𝔼[∙|ℱ𝑡 ])
𝑇 𝑚 𝑚 𝑇 𝑚 𝑛
𝜕ℎ 1 𝜕2ℎ 𝜕ℎ
RHS = 𝔼𝑡 [∫ (∑ 𝐴𝑖 + ∑ 𝛴𝑖𝑗 ) 𝑑𝑢] + 𝔼𝑡 [∫ ∑ ∑ 𝐵𝑖𝑘 𝑑𝑊𝑘 ]
𝑡 𝜕𝑆𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗 𝑡 𝜕𝑆𝑖 (6)
𝑖=1 𝑖,𝑗=1 ⏟ 𝑖=1 𝑘=1
=0
𝑇 𝑚 𝑚
𝜕ℎ 1 𝑇 𝜕2ℎ
= ∫ ∑ 𝔼𝑡 [ 𝐴𝑖 ] 𝑑𝑢 + ∫ ∑ 𝔼𝑡 [ 𝛴 ] 𝑑𝑢
𝑡 𝜕𝑆𝑖 2 𝑡 𝜕𝑆𝑖 𝜕𝑆𝑗 𝑖𝑗
𝑖=1 𝑖,𝑗=1
where 𝑝𝑇,𝑦|𝑡,𝑥 is the transition probability density having 𝑆𝑇 = 𝑦 at 𝑇 given 𝑆𝑡 = 𝑥 at 𝑡 (i.e. if we solve
the equation (1) with the initial condition 𝑆𝑡 = 𝑥 ∈ ℝ𝑚 , then the random variable 𝑆𝑇 = 𝑦 ∈ 𝛺 has a
density 𝑝𝑇,𝑦|𝑡,𝑥 in the 𝑦 variable at time 𝑇). Differentiating (6) with respect to 𝑇 on both sides, we have
𝑚 𝑚
𝜕𝑝𝑇,𝑦|𝑡,𝑥 𝜕ℎ 1 𝜕2ℎ
∫ ℎ𝑦 𝑑𝑦 = ∑ 𝔼𝑡 [ 𝐴 ] + ∑ 𝔼𝑡 [ 𝛴 ]
𝛺 𝜕𝑇 𝜕𝑆𝑖 𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗 𝑖𝑗
𝑖=1 𝑖,𝑗=1
(7)
𝑚 𝑚 2
𝜕ℎ𝑦 1 𝜕 ℎ𝑦
= ∑∫ 𝐴𝑖 𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦 + ∑ ∫ 𝛴𝑖𝑗 𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦
𝑖=1 𝛺 𝜕𝑦𝑖 2 𝛺 𝜕𝑦𝑖 𝜕𝑦𝑗
𝑖,𝑗=1
3
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
If we assume 𝛺 ≡ ℝ𝑚 and also assume the probability density 𝑝 and its first derivatives 𝜕𝑝/𝜕𝑦𝑖 vanish
at a higher order of rate than ℎ and 𝜕ℎ/𝜕𝑦𝑖 as 𝑦𝑖 → ±∞ ∀ 𝑖 = 1, ⋯ , 𝑚, then we can integrate by parts
for the right hand side of (7), once for the first integral and twice for the second
𝜕ℎ𝑦 +∞ 𝜕(𝐴𝑖 𝑝)
∫ 𝐴𝑖 𝑝𝑑𝑦 = ∫ ℎ𝑦 𝐴𝑖 𝑝|𝑦 =−∞ 𝑑𝑦−𝑖 − ∫ ℎ𝑦 𝑑𝑦 and
𝛺 𝜕𝑦𝑖 𝛺−𝑖
⏟ 𝑖
𝛺 𝜕𝑦𝑖
=0
+∞
𝜕 2 ℎ𝑦 𝜕ℎ𝑦 𝜕ℎ𝑦 𝜕(𝛴𝑖𝑗 𝑝)
∫ 𝛴𝑖𝑗 𝑝𝑑𝑦 = ∫ 𝛴𝑖𝑗 𝑝| 𝑑𝑦̅𝑖 − ∫ 𝑑𝑦
𝛺 𝜕𝑦𝑖 𝜕𝑦𝑗 ̅ 𝑖 𝜕𝑦𝑗
𝛺 ⏟ 𝑦𝑖 =−∞ 𝛺 𝜕𝑦𝑗 𝜕𝑦 𝑖
=0
(8)
+∞
𝜕(𝛴𝑖𝑗 𝑝) 𝜕 2 (𝛴𝑖𝑗 𝑝)
= − ∫ ℎ𝑦 | 𝑑𝑦̅ + ∫ ℎ𝑦 𝑑𝑦
̅𝑗
𝛺 𝜕𝑦𝑖 𝑦 =−∞ 𝑗 𝛺 𝜕𝑦𝑖 𝜕𝑦𝑗
⏟ 𝑗
=0
This is the Multi-dimensional Fokker-Planck Equation (a.k.a. Kolmogorov Forward Equation) [1]. In
this equation, the 𝑡 and 𝑥 are held constant, while the 𝑇 and 𝑦 are variables (called “forward variables”).
𝜕𝑝 𝜕(𝐴𝑝) 1 𝜕 2 (𝐵2 𝑝)
+ − =0 (11)
𝜕𝑇 𝜕𝑦 2 𝜕𝑦 2
4
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Let’s express conditional expectation of ℎ(𝑆𝑡 ) by 𝑔(𝑡, 𝑆𝑡 ) = 𝔼𝑡 [ℎ(𝑆𝑇 )]. Because for 𝑜 ≤ 𝑡 ≤ 𝑇
we have
the 𝑔(𝑡, 𝑆𝑡 ) is a martingale by the tower rule (i.e. If ℋ holds less information than 𝒢 , then
𝜕𝑔 1
𝑑𝑔 = 𝑑𝑡 + 𝐽𝑔 𝑑𝑆𝑡 + 𝑑𝑆𝑡′ 𝐻𝑔 𝑑𝑆𝑡
𝜕𝑡 1×𝑚𝑚×1 2 1×𝑚𝑚×𝑚𝑚×1
(13)
𝜕𝑔 1
= 𝑑𝑡 + 𝐽𝑔 𝐴𝑑𝑡 + 𝐽𝑔 𝐵𝑑𝑊𝑡 + 𝑑𝑊𝑡′ 𝐵′ 𝐻𝑔 𝐵𝑑𝑊𝑡
𝜕𝑡 2
where 𝐽𝑔 is the Jacobian (i.e. the same as gradient if 𝑔 is a scalar-valued function) and 𝐻𝑔 the Hessian of
𝜕2𝑔 𝜕2𝑔
⋯
𝜕𝑔 𝜕𝑔 𝜕𝑆12 𝜕𝑆1 𝜕𝑆𝑚
𝐽𝑔 = ( ⋯ ), 𝐻𝑔 = ⋮ ⋱ ⋮ (14)
𝜕𝑆1 𝜕𝑆𝑚 2 2
𝜕 𝑔 𝜕 𝑔
⋯ 2 )
(𝜕𝑆𝑚 𝜕𝑆1 𝜕𝑆𝑚
𝑚 𝑚 𝑚 𝑛
𝜕𝑔 𝜕𝑔 1 𝜕2𝑔 𝜕𝑔
𝑑𝑔 = ( + ∑ 𝐴𝑖 + ∑ 𝛴𝑖𝑗 ) 𝑑𝑡 + ∑ ∑ 𝐵𝑖𝑘 𝑑𝑊𝑘 (15)
𝜕𝑡 𝜕𝑆𝑖 2 𝜕𝑆𝑖 𝜕𝑆𝑗 𝜕𝑆𝑖
𝑖=1 𝑖,𝑗=1 𝑖=1 𝑘=1
Using the transition probability density 𝑝𝑇,𝑦|𝑡,𝑥 , we can write the expectation as
1
https://en.wikipedia.org/wiki/Feynman-Kac_formula
5
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
This is the multi-dimensional Kolmogorov Backward Equation. In this equation, the 𝑇 and 𝑦 are held
constant, while the 𝑡 and 𝑥 are variables (called “backward variables”). In the 1-D case, it reduces to
𝜕𝑝 𝜕𝑝 1 2 𝜕 2 𝑝
+𝐴 + 𝐵 =0 (20)
𝜕𝑡 𝜕𝑥 2 𝜕𝑥 2
2. LOCAL VOLATILITY
In local volatility models, the volatility process is assumed to be a function of both the spot level
and the time. It is one step generalization of the well-known Black-Scholes model. Under risk neutral
measure, the spot process (e.g. an equity or an FX rate) is assumed to follow a geometric Brownian
motion
𝑑𝑆𝑡
̃𝑡 ,
= 𝜇𝑡 𝑑𝑡 + 𝜎(𝑡, 𝑆𝑡 )𝑑𝑊 𝜇𝑡 = 𝑟𝑡 − 𝑞𝑡 (21)
𝑆𝑡
with cash rate 𝑟𝑡 and dividend rate 𝑞𝑡 (or foreign cash rate for FX).
Under the assumption of deterministic 𝑟𝑡 , the European (vanilla) call option price can be
6
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
∞
𝒞𝑇,𝐾|𝑡,𝑥 ̃ 𝑡 [𝐷𝑡,𝑇 (𝑆𝑇 − 𝐾)+ ] = 𝐷𝑡,𝑇 ∫ (𝑦 − 𝐾)𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦
=𝔼 (22)
𝐾
𝑇
where 𝐷𝑡,𝑇 = exp (− ∫𝑡 𝑟𝑢 𝑑𝑢) is the deterministic discount factor and 𝑝𝑇,𝑦|𝑡,𝑥 is the transition
probability density having spot 𝑆𝑇 = 𝑦 at 𝑇 given initial condition 𝑆𝑡 = 𝑥 at 𝑡. Differentiating (22) with
respect to 𝐾, we have the first order and second order partial derivative
∞
𝜕𝒞𝑇,𝐾|𝑡,𝑥 𝜕 2 𝒞𝑇,𝐾|𝑡,𝑥
= −𝐷𝑡,𝑇 ∫ 𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦 , = 𝐷𝑡,𝑇 𝑝𝑇,𝐾|𝑡,𝑥 (23)
𝜕𝐾 𝐾 𝜕𝐾 2
1 𝜕 2 𝒞𝑇,𝐾|𝑡,𝑥
𝑝𝑇,𝐾|𝑡,𝑥 = (24)
𝐷𝑡,𝑇 𝜕𝐾 2
2
𝜕𝑝𝑇,𝑦|𝑡,𝑥 1 𝜕 2 (𝜎𝑇,𝑦 𝑦 2 𝑝𝑇,𝑦|𝑡,𝑥 ) 𝜕(𝜇 𝑇 𝑦𝑝𝑇,𝑦|𝑡,𝑥 )
= − (26)
𝜕𝑇 2 𝜕𝑦 2 𝜕𝑦
Applying integration by parts to the integrals on the right hand side of (25) yields
∞ 𝜕(𝜇 𝑇 𝑦𝑝𝑇,𝑦|𝑡,𝑥 ) ∞
∞
∫ (𝑦 − 𝐾) 𝑑𝑦 = (𝑦 − 𝐾)𝜇 𝑇 𝑦𝑝𝑇,𝑦|𝑡,𝑥 |𝑦=𝐾 − ∫ 𝜇 𝑇 𝑦𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦
𝐾 𝜕𝑦 ⏟ 𝐾
=0
∞ ∞
𝜇 𝑇 𝐾 𝜕𝒞𝑇,𝐾|𝑡,𝑥 𝜇 𝑇 𝒞𝑇,𝐾|𝑡,𝑥 (27)
= −𝜇 𝑇 (∫ (𝑦 − 𝐾)𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦 + 𝐾 ∫ 𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦) = −
𝐾 𝐾 𝐷𝑡,𝑇 𝜕𝐾 𝐷𝑡,𝑇
and
7
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
2
∞ 𝜕 2 (𝜎𝑇,𝑦 𝑦 2 𝑝𝑇,𝑦|𝑡,𝑥 )
∫ (𝑦 − 𝐾) 𝑑𝑦
𝐾 𝜕𝑦 2
2 ∞ ∞ 𝜕(𝜎 2 𝑦 2 𝑝
𝜕(𝜎𝑇,𝑦 𝑦 2 𝑝𝑇,𝑦|𝑡,𝑥 ) 𝑇,𝑦 𝑇,𝑦|𝑡,𝑥 )
= (𝑦 − 𝐾) | −∫ 𝑑𝑦
𝜕𝑦 𝜕𝑦
⏟ 𝑦=𝐾 𝐾
=0
2
2 ∞ 2
𝜎𝑇,𝐾 𝐾 2 𝜕 2 𝒞𝑇,𝐾|𝑡,𝑥
= 𝜎𝑇,𝑦 𝑦 2 𝑝𝑇,𝑦|𝑡,𝑥 | = 𝜎𝑇,𝐾 𝐾 2 𝑝𝑇,𝐾|𝑡,𝑥 =
𝑦=𝐾 𝐷𝑡,𝑇 𝜕𝐾 2
where we have 𝑝𝑇,∞|𝑡,𝑥 = 0 and 𝜕𝑝𝑇,∞|𝑡,𝑥 /𝜕𝑦 = 0 assuming the probability density 𝑝𝑇,𝑦|𝑡,𝑥 and its first
derivative vanish at a higher order of rate as 𝑦 → ∞. Plugging (27) into (25), we find
2
𝜕𝒞𝑇,𝐾|𝑡,𝑥 𝜎𝑇,𝐾 𝐾 2 𝜕 2 𝒞𝑇,𝐾|𝑡,𝑥 𝜕𝒞𝑇,𝐾|𝑡,𝑥
= −𝑟𝑇 𝒞𝑇,𝐾|𝑡,𝑥 + 2
− 𝜇𝑇 𝐾 + 𝜇 𝑇 𝒞𝑇,𝐾|𝑡,𝑥
𝜕𝑇 2 𝜕𝐾 𝜕𝐾
(28)
2
𝜎𝑇,𝐾 𝐾 2 𝜕 2 𝒞𝑇,𝐾|𝑡,𝑥 𝜕𝒞𝑇,𝐾|𝑡,𝑥
= − 𝜇𝑇 𝐾 − 𝑞𝑇 𝒞𝑇,𝐾|𝑡,𝑥
2 𝜕𝐾 2 𝜕𝐾
and eventually the Dupire formula for the local volatility 𝜎𝑇,𝐾 expressed in terms of vanilla call price
𝜕𝒞𝑇,𝐾|𝑡,𝑥 𝜕𝒞𝑇,𝐾|𝑡,𝑥
2
𝜎𝑇,𝐾 + 𝜇𝑇 𝐾 + 𝑞𝑇 𝒞𝑇,𝐾|𝑡,𝑥
= 𝜕𝑇 𝜕𝐾
2 (29)
2 𝜕 𝒞𝑇,𝐾|𝑡,𝑥
𝐾2
𝜕𝐾 2
Sometimes, it is more convenient to express the Dupire formula in terms of a forward (i.e.
with 𝒞𝑇,𝐾|𝑡,𝑥 given in (22) (note that (30) is true only if the interest rate is deterministic). Following a
8
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Therefore the Dupire formula for 𝜎𝑇,𝐾 expressed in terms of forward call price reads
𝜕𝐶𝑇,𝐾|𝑡,𝑥 𝜕𝐶𝑇,𝐾|𝑡,𝑥
2
𝜎𝑇,𝐾 + 𝜇𝑇 𝐾 − 𝜇 𝑇 𝐶𝑇,𝐾|𝑡,𝑥
= 𝜕𝑇 𝜕𝐾 (32)
2 𝜕 2 𝐶𝑇,𝐾|𝑡,𝑥
𝐾2
𝜕𝐾 2
̃ 𝑡 [(𝑆𝑇 − 𝐾)+ ] = 𝔼
𝐶𝑇,𝐾|𝑡,𝑥 = 𝔼 ̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)(𝑆𝑇 − 𝐾)] = 𝔼
̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)𝑆𝑇 ] − 𝐾𝔼
̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)] (33)
where 𝒽 is the Heaviside step function1. Differentiating once with respect to 𝐾, we get
∞
𝜕𝐶𝑇,𝐾|𝑡,𝑥 𝜕𝐶
̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)𝑆𝑇 ] = 𝐶𝑇,𝐾|𝑡,𝑥 − 𝐾 𝑇,𝐾|𝑡,𝑥 (34)
̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)] ⟹ 𝔼
= − ∫ 𝑝𝑇,𝑦|𝑡,𝑥 𝑑𝑦 = −𝔼
𝜕𝐾 𝐾 𝜕𝐾
𝜕 2 𝐶𝑇,𝐾|𝑡,𝑥
̃ 𝑡 [𝛿(𝑆𝑇 − 𝐾)]
= 𝑝𝑇,𝐾|𝑡,𝑥 = 𝔼 (35)
𝜕𝐾 2
where 𝛿 is the Dirac delta function2. Applying Ito’s Lemma to the terminal payoff of the option gives
the identity
1
𝑑(𝑆𝑇 − 𝐾)+ = 𝒽(𝑆𝑇 − 𝐾)𝑑𝑆𝑇 + 𝛿(𝑆𝑇 − 𝐾)𝜎𝑇2 𝑆𝑇2 𝑑𝑇
2
(36)
1
̃𝑇
= 𝒽(𝑆𝑇 − 𝐾)𝜇 𝑇 𝑆𝑇 𝑑𝑇 + 𝛿(𝑆𝑇 − 𝐾)𝜎𝑇2 𝑆𝑇2 𝑑𝑇 + 𝒽(𝑆𝑇 − 𝐾)𝜎𝑇 𝑆𝑇 𝑑𝑊
2
1
̃ 𝑡 [(𝑆𝑇 − 𝐾)+ ] = 𝜇 𝑇 𝔼
𝑑𝐶𝑇,𝐾|𝑡,𝑥 = 𝑑𝔼 ̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)𝑆𝑇 ]𝑑𝑇 + 𝔼̃ 𝑡 [𝛿(𝑆𝑇 − 𝐾)𝜎𝑇2 𝑆𝑇2 ]𝑑𝑇
2
(37)
𝜕𝐶𝑇,𝐾|𝑡,𝑥 1
⟹ ̃ 𝑡 [𝒽(𝑆𝑇 − 𝐾)𝑆𝑇 ] + 𝔼
= 𝜇𝑇 𝔼 ̃ [𝛿(𝑆𝑇 − 𝐾)𝜎𝑇2 𝑆𝑇2 ]
𝜕𝑇 2 𝑡
Notice that
1 0, 𝑥 < 0
Heaviside step function: 𝒽(𝑥) = {
1, 𝑥 ≥ 0
2 𝑑𝒽(𝑥) ∞, 𝑥 = 0
Dirac delta function can be viewed as the derivative of the Heaviside step function: 𝛿(𝑥) = ={ ,
𝑑𝑥 0, 𝑥 ≠ 0
which is also constrained to satisfy the identity: ∫ℝ 𝛿(𝑥)𝑑𝑥 = 1.
9
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
̃ 𝑡 [𝛿(𝑆𝑇 − 𝐾)𝜎𝑇2 ] = 𝔼
𝔼 ̃ 𝑡 [𝜎𝑇2 |𝑆𝑇 = 𝐾]𝔼
̃ 𝑡 [𝛿(𝑆𝑇 − 𝐾)] (39)
This is identical to (32). It means that the conditional expectation of the stochastic variance must equal
the Dupire local variance [2]. That is, local variance is the risk-neutral expectation of the instantaneous
variance conditional on the final stock price 𝑆𝑇 being equal to the strike price 𝐾 [3].
In real applications, numerical methods are often in favor of log-moneyness to be the spatial
𝐾 𝜕𝑘 1 𝜕𝑘
𝑘 = ln where = , = −𝜇 𝑇 (41)
𝐹𝑡,𝑇 𝜕𝐾 𝐾 𝜕𝑇
We want to express the Dupire formula in the (𝑇, 𝑘)-plane using call option price 𝐶𝑇,𝑘 (short for 𝐶𝑇,𝑘|𝑡,𝑧
𝑥
for 𝑧 = ln ) equivalent to the forward call 𝐶𝑇,𝐾 (short for 𝐶𝑇,𝐾|𝑡,𝑥 ). Note that although the 𝐶𝑇,𝑘 and
𝐹𝑡,𝑇
𝐶𝑇,𝐾 are equivalent, they are two different functions. The conversion from (𝑇, 𝐾)-plane to (𝑇, 𝑘)-plane
10
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Plugging these partial derivatives into (32), we have the Dupire formula expressed in 𝑘
𝜕𝐶𝑇,𝑘
2
𝜎𝑇,𝑘 − 𝜇 𝑇 𝐶𝑇,𝑘
= 𝜕𝑇
2 (43)
2 𝜕 𝐶𝑇,𝑘 𝜕𝐶𝑇,𝑘
−
𝜕𝑘 2 𝜕𝑘
more straightforward to express the local volatility in terms of the implied volatilities rather than option
𝐹𝑡,𝑇
ln
𝑋𝑇,𝐾,𝜉 = 𝐹𝑡,𝑇 Φ(𝑑+ ) − 𝐾Φ(𝑑− ) with 𝑑± = 𝐾 ± 𝜉𝑇,𝐾 √𝜏 , 𝜏 =𝑇−𝑡 (45)
𝜉𝑇,𝐾 √𝜏 2
where 𝜉𝑇,𝐾 is the Black-Scholes implied volatility derived from market quotes of vanilla options and Φ
the standard normal cumulative density function. Its partial derivatives can be derived as
11
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
the local volatility given by implied volatility can be derived from (32) as
𝜕𝐶𝑇,𝐾 𝜕𝐶 𝜕𝑋 𝜕𝑋 𝜕𝜉𝑇,𝐾 𝜕𝜉 𝜕𝑋
+ 𝜇 𝑇 𝐾 𝑇,𝐾 − 𝜇 𝑇 𝐶𝑇,𝐾 + ( + 𝜇 𝑇 𝐾 𝑇,𝐾 ) + 𝜇 𝑇 𝐾 − 𝜇𝑇 𝑋
2 𝜕𝑇 𝜕𝐾 𝜕𝑇 𝜕𝜉 𝜕𝑇 𝜕𝐾 𝜕𝐾
𝜎𝑇,𝐾 = =
𝐾 2 𝜕 2 𝐶𝑇,𝐾 𝐾2 𝜕2𝑋 𝜕 2 𝑋 𝜕𝜉𝑇,𝐾 𝜕𝑋 𝜕 2 𝜉𝑇,𝐾 𝜕 2 𝑋 𝜕𝜉 2
2 𝜕𝐾 2 ( + 2 + + 2( ) )
2 𝜕𝐾 2 𝜕𝐾𝜕𝜉 𝜕𝐾 𝜕𝜉 𝜕𝐾 2 𝜕𝜉 𝜕𝐾
𝐾𝜙(𝑑− )𝜉 𝜕𝑋 𝜕𝜉 𝜕𝜉
𝜇 𝑇 𝐹𝑡,𝑇 Φ(𝑑+ ) + + ( + 𝜇 𝑇 𝐾 ) − 𝜇 𝑇 𝐾Φ(𝑑− ) − 𝜇 𝑇 𝑋
2√𝜏 𝜕𝜉 𝜕𝑇 𝜕𝐾
=
2
𝐾 𝜙(𝑑− ) 𝜙(𝑑− )𝑑+ 𝜕𝜉 𝐾 𝜕𝑋 𝜕 𝜉 𝐾 2 𝑑+ 𝑑− 𝐾𝜙(𝑑− )√𝜏 𝜕𝜉 2
2 2
+ 𝐾2 + + ( )
2 𝐾𝜉 √𝜏 𝜉 𝜕𝐾 2 𝜕𝜉 𝜕𝐾 2 2 𝜉 𝜕𝐾
(50)
𝜕𝑋 𝜉 𝜕𝜉 𝜕𝜉
( + + 𝜇𝑇 𝐾 )
𝜕𝜉 2𝜏 𝜕𝑇 𝜕𝐾
=
1 𝜕𝑋 𝐾𝑑+ 𝜕𝑋 𝜕𝜉 𝐾 2 𝜕𝑋 𝜕 2 𝜉 𝐾 2 𝑑+ 𝑑− 𝜕𝑋 𝜕𝜉 2
+ + + ( )
2𝜉𝜏 𝜕𝜉 𝜉 √𝜏 𝜕𝜉 𝜕𝐾 2 𝜕𝜉 𝜕𝐾 2 2𝜉 𝜕𝜉 𝜕𝐾
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (+ 𝜇𝑇 𝐾 )
= 𝜕𝑇 𝜕𝐾
𝜕𝜉 𝜕𝜉 2 𝜕2𝜉
1 + 2√𝜏𝐾𝑑+ + 𝑑+ 𝑑− 𝜏𝐾 2 ( ) + 𝜉𝜏𝐾 2
𝜕𝐾 𝜕𝐾 𝜕𝐾 2
Numerical methods, e.g. PDE or Monte Carlo simulation, often demand a local volatility
function constructed on a 2D grid to perform pricing. In these methods, it is often more numerically
The local volatility formula in log strike 𝑥 = ln 𝐾 can be derived from (50)
12
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (
+ 𝜇𝑇 )
2
𝜎𝑇,𝑥 = 𝜕𝑇 𝜕𝑥
𝜕𝜉 𝜕𝜉 2 𝜕 2 𝜉 𝜕𝜉
1 + 2√𝜏𝑑+ + 𝑑+ 𝑑− 𝜏 ( ) + 𝜉𝜏 ( 2 − )
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (
+ 𝜇𝑇 )
= 𝜕𝑇 𝜕𝑥
(51)
𝑘 𝜕𝜉 𝑘 2 𝜉 2 𝜏 2 𝜕𝜉 2 𝜕 2 𝜉 𝜕𝜉
1 + (𝜉𝜏 − 2 ) +( 2 − ) ( ) + 𝜉𝜏 ( 2 − )
𝜉 𝜕𝑥 𝜉 4 𝜕𝑥 𝜕𝑥 𝜕𝑥
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (+ 𝜇𝑇 )
= 𝜕𝑇 𝜕𝑥
𝑘 𝜕𝜉 2 𝜉𝜏 𝜕𝜉 2 𝜕2𝜉
(1 − ) −( ) + 𝜉𝜏 2
𝜉 𝜕𝑥 2 𝜕𝑥 𝜕𝑥
𝜕𝜉 𝜕𝜉 𝜕𝑥 1 𝜕𝜉 𝜕2𝜉 𝜕 1 𝜕𝜉 𝜕𝑥 1 𝜕 2 𝜉 𝜕𝜉 𝜕𝑥 1
= = , = ( ) = ( − ), =
𝜕𝐾 𝜕𝑥 𝜕𝐾 𝐾 𝜕𝑥 𝜕𝐾 2 𝜕𝑥 𝐾 𝜕𝑥 𝜕𝐾 𝐾 2 𝜕𝑥 2 𝜕𝑥 𝜕𝐾 𝐾
(52)
−𝑘 𝜉 √𝜏 𝐾
𝑑± = ± , 𝑘 = ln
𝜉 √𝜏 2 𝐹𝑡,𝑇
We may also want to change the spatial variable to log-moneyness 𝑘. Defining a new quantity,
2
implied total variance 𝑣𝑇,𝑘 , which is equivalent to 𝜉𝑇,𝐾 𝜏, the Black-Scholes call price that is equivalent
−𝑘 √𝑣𝑇,𝑘
𝑋𝑇,𝑘,𝑣 = 𝐹𝑡,𝑇 (Φ(𝑑+ ) − exp(𝑘) Φ(𝑑− )) with 𝑑± = ± (53)
√𝑣𝑇,𝑘 2
𝜕𝑋𝑇,𝑘,𝑣 𝑘 2 1 1
= ( 2− − )
𝜕𝑣 2𝑣 2𝑣 8
13
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
𝜕𝑋𝑇,𝑘,𝑣 𝜕𝐹𝑡,𝑇
= (Φ(𝑑+ ) − exp(𝑘) Φ(𝑑− )) = 𝜇 𝑇 𝑋𝑇,𝑘,𝑣
𝜕𝑇 𝜕𝑇
We may connect the local volatility 𝜎𝑇,𝑘 to the implied total variance 𝑣𝑇,𝑘 via two steps. Firstly
𝜕𝑋𝑇,𝑘,𝑣 𝜕𝑋𝑇,𝑘,𝑣 𝜕𝑣
2( + − 𝜇 𝑇 𝑋𝑇,𝑘,𝑣 )
2 𝜕𝑇 𝜕𝑣 𝜕𝑇
𝜎𝑇,𝑘 = 2 (56)
𝜕 𝑋𝑇,𝑘,𝑣 𝜕 2 𝑋𝑇,𝑘,𝑣 𝜕𝑣 𝜕𝑋𝑇,𝑘,𝑣 𝜕 2 𝑣 𝜕 2 𝑋𝑇,𝑘,𝑣 𝜕𝑣 2 𝜕𝑋𝑇,𝑘,𝑣 𝜕𝑋𝑇,𝑘,𝑣 𝜕𝑣
+ 2 + + ( ) − −
𝜕𝑘 2 𝜕𝑘𝜕𝑣 𝜕𝑘 𝜕𝑣 𝜕𝑘 2 𝜕𝑣 2 𝜕𝑘 𝜕𝑘 𝜕𝑣 𝜕𝑘
Secondly we substitute the partial derivatives in (54) into (56) and reach the final equation
𝜕𝑋 𝜕𝑣
2 (𝜇 𝑇 𝑋 +
− 𝜇 𝑇 𝑋)
2
𝜎𝑇,𝑘 = 𝜕𝑣 𝜕𝑇
𝜕𝑋 𝜕𝑋 𝜕𝑋 1 𝑘 𝜕𝑣 𝜕𝑋 𝜕 2 𝑣 𝜕𝑋 𝑘 2 1 1 𝜕𝑣 2 𝜕𝑋 𝜕𝑋 𝜕𝑣
+2 +2 ( − ) + + ( − − ( ) − −
𝜕𝑘 𝜕𝑣 𝜕𝑣 2 𝑣 𝜕𝑘 𝜕𝑣 𝜕𝑘 2 𝜕𝑣 2𝑣 2 2𝑣 8) 𝜕𝑘 𝜕𝑘 𝜕𝑣 𝜕𝑘
(57)
𝜕𝑣
2
⟹ 𝜎𝑇,𝑘 = 𝜕𝑇
𝑘 𝜕𝑣 1 𝑘 2 1 1 𝜕𝑣 2 1 𝜕 2 𝑣
1− + − − ( ) +
𝑣 𝜕𝑘 4 (𝑣 2 𝑣 4) 𝜕𝑘 2 𝜕𝑘 2
14
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
2 2
The 𝜎𝑇,𝐾 in (50) is in fact equivalent to the 𝜎𝑇,𝑘 in (57). This can be shown as follows
𝜕𝑣
2
𝜎𝑇,𝑘 = 𝜕𝑇
𝑘 𝜕𝑣 𝑘2 1 1 𝜕𝑣 2 1 𝜕 2 𝑣
1− +( 2− − )( ) +
𝑣 𝜕𝑘 4𝑣 4𝑣 16 𝜕𝑘 2 𝜕𝑘 2
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (
+ 𝜇𝐾 )
= 𝜕𝑇 𝜕𝐾
𝑘 𝜕𝜉 𝑘2 1 1 𝜕𝜉 2 𝜕𝜉 2 𝜉 𝜕𝜉 𝜕2𝜉
1 − 2𝜉𝜏𝐾 + ( 2 − − ) (𝜉𝜏𝐾 ) + 𝜏𝐾 2 (( ) + +𝜉 )
𝑣 𝜕𝐾 𝑣 𝑣 4 𝜕𝐾 𝜕𝐾 𝐾 𝜕𝐾 𝜕𝐾 2
(58)
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 (
+ 𝜇𝐾 )
= 𝜕𝑇 𝜕𝐾
𝑘 𝜕𝜉 𝑘2 𝑣 𝜕𝜉 2 𝜕2𝜉
1 + (1 − 2 ) 𝜉𝜏𝐾 + ( − ) 𝜏𝐾 2 ( ) + 𝜉𝜏𝐾 2
𝑣 𝜕𝐾 𝑣 4 𝜕𝐾 𝜕𝐾 2
𝜕𝜉 𝜕𝜉
𝜉 2 + 2𝜉𝜏 ( + 𝜇𝐾 )
= 𝜕𝑇 𝜕𝐾 2
= 𝜎𝑇,𝐾
2 2𝜉
𝜕𝜉 𝜕𝜉 𝜕
1 + 2√𝜏𝐾𝑑+ + 𝑑+ 𝑑− 𝜏𝐾 2 ( ) + 𝜉𝜏𝐾 2
𝜕𝐾 𝜕𝐾 𝜕𝐾 2
𝐹𝑡,𝑇 𝜉 2 𝜏
𝐾 ln ± 𝑘2 𝑣
𝑘 = ln , 𝑣 = 𝜉 𝜏, 2
𝑑± = 𝐾 2 = −𝑘 ± √𝑣 , 𝑑+ 𝑑− = − (59)
𝐹𝑡,𝑇 𝜉 √𝜏 √𝑣 2 𝑣 4
𝜕𝑣 𝜕(𝜉 2 𝜏) 𝜕𝜉 𝜕𝜉 𝜕𝐾 𝜕𝜉 𝜕𝜉
= = 𝜉 2 + 2𝜉𝜏 ( + ) = 𝜉 2 + 2𝜉𝜏 ( + 𝜇𝐾 )
𝜕𝑇 𝜕𝑇 𝜕𝑇 𝜕𝐾 𝜕𝑇 𝜕𝑇 𝜕𝐾
𝜕𝜉 𝜕𝜉 (60)
𝜕 2 𝑣 𝜕 (2𝜉𝜏𝐾 𝜕𝐾 ) 𝜕𝑇 𝜕 (2𝜉𝜏𝐾 𝜕𝐾 ) 𝜕𝐾 𝜕𝜉 𝜕𝜉 𝜕𝜉 𝜕2𝜉
= + = 2𝜏𝐾 (𝜉 +𝐾 + 𝜉𝐾 )
𝜕𝑘 2 𝜕𝑇 𝜕𝑘 𝜕𝐾 𝜕𝑘 𝜕𝐾 𝜕𝐾 𝜕𝐾 𝜕𝐾 2
𝜕𝜉 2 𝜉 𝜕𝜉
2
𝜕2𝜉
= 2𝜏𝐾 (( ) + +𝜉 )
𝜕𝐾 𝐾 𝜕𝐾 𝜕𝐾 2
considering the fact that in (𝑇, 𝑘)-plane the 𝑇 and 𝐾 are no longer independent
15
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
In this chapter, we will present a PDE based local volatility model, in which the local volatility
surface is constructed as a 2-D function that is piecewise constant in maturity and piecewise linear in
log-moneyness (for equity) or delta (for FX). Due to great similarity between FX and equity processes,
our interest lies primarily in the context of equity derivatives, the conclusions and formulas drawn from
our discussion here are in general applicable to FX products with minor changes. In contrast to the
traditional way to construct the local volatility by estimating highly sensitive and numerically unstable
partial derivatives in Dupire formulas, this method relies heavily on solving forward PDE’s to calibrate a
parametrized local volatility surface to vanilla option prices in a bootstrapping manner. Once the local
volatility surface is calibrated, the backward PDE can then be used to price exotic options (e.g. barrier
options) that are in consistent with the market observed implied volatility surface.
Before proceeding to the PDE’s, it is important to have an overview of the date conventions for
equity and equity options. The date conventions for FX products are defined in a similar manner.
The diagram illustrates the date definitions for an equity and its associated option. The quantities
16
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
As most of the quantities are self-explanatory, our discussion focuses more on the treatment of
equity dividends.
In our example, we can assume both the short rate and the dividend rate are deterministic and
continuous, e.g. time-dependent 𝑟𝑢 and 𝑞𝑢 as in (21). the equity forward in this case can be calculated by
𝑃𝑞 (𝑡𝑒,𝑠 , 𝑡𝑒,𝑝 )
𝐹(𝑡0 , 𝑡𝑒,𝑚 ) = 𝑆(𝑡0 ) where
𝑃𝑟 (𝑡𝑒,𝑠 , 𝑡𝑒,𝑝 )
(62)
𝑇 𝑇
𝑃𝑞 (𝑡, 𝑇) = exp (− ∫ 𝑞𝑢 𝑑𝑢) , 𝑃𝑟 (𝑡, 𝑇) = exp (− ∫ 𝑟𝑢 𝑑𝑢)
𝑡 𝑡
In a more realistic implementation, we may assume the underlying equity issues a series of
discrete dividends with fixed amounts in a foreseeable future. It is obvious that the equity spot still
follows the SDE (21) with 𝑞𝑢 = 0 in between two adjacent ex-dividend dates (There is discontinuity in
1
Equity settlement delay
17
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
spot process on ex-dividend dates that demands special treatment. This will be discussed in detail in due
where 𝜃𝑖 is the fixed amount of the 𝑖-th dividend issued on ex-dividend date 𝑡𝑖,𝑒 .
Discrete dividend can also be modeled as proportional dividend. It assumes that at each ex-
dividend date, the dividend payment will result in a price drop in equity spot proportional to the spot
level. For example, the equity spot before and after the dividend fall has the relationship
where Δ denotes an infinitesimal amount of time and 𝜂𝑖 the proportional dividend rate at ex-dividend
∏𝑖(1 − 𝜂𝑖 )
𝐹(𝑡0 , 𝑡𝑒,𝑚 ) = 𝑆(𝑡0 ) for 𝑡0 < 𝑡𝑖,𝑒 ≤ 𝑡𝑒,𝑚 (65)
𝑃𝑟 (𝑡𝑒,𝑠 , 𝑡𝑒,𝑝 )
dividends. The conversion can be achieved by equating the equity forward in (63) and in (65), such that
1
∏(1 − 𝜂𝑖 ) = 1 − ∑ 𝜃𝑖 𝑃𝑟 (𝑡𝑒,𝑠 , 𝑡𝑖,𝑝 ) for 𝑡0 < 𝑡𝑖,𝑒 ≤ 𝑡𝑒,𝑚 (66)
𝑆(𝑡0 )
𝑖 𝑖
The proportional dividend 𝜂𝑖 can then be bootstrapped from a series of fixed dividends 𝜃𝑖 starting from
In the following, our derivation is based on the spot process 𝑆𝑡 defined in (21) and its variants.
For example, depending on the application we may write the SDE (21) in terms of log-spot 𝓏𝑢 = ln 𝑆𝑢
𝑆𝑢
or in terms of centered log-spot 𝑧𝑢 = ln
𝐹𝑡,𝑢
1 1
̃𝑢 and 𝑑𝑧𝑢 = − 𝜎(𝑢, 𝑧)2 𝑑𝑢 + 𝜎(𝑢, 𝑧)𝑑𝑊
𝑑𝓏𝑢 = (𝜇𝑢 − 𝜎(𝑢, 𝓏)2 ) 𝑑𝑢 + 𝜎(𝑢, 𝓏)𝑑𝑊 ̃𝑢 (67)
2 2
18
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
where 𝜎(𝑢, 𝓏) and 𝜎(𝑢, 𝑧) are the local volatility function of 𝓏 and 𝑧, respectively.
Let’s denote the forward temporal variable by 𝑢 for 𝑡 < 𝑢 < 𝑇, the spatial variable by log-
𝐾 𝑆𝑢
moneyness 𝑘 = ln (as in (41)) and the spot by 𝑧 = ln . Given that 𝑧𝑡 = 0 , the value of a
𝐹𝑡,𝑢 𝐹𝑡,𝑢
Let 𝜎𝑢,𝑘 be the local volatility function of variable 𝑘 equivalent to 𝜎𝑇,𝐾 . We can derive the forward PDE
𝜕𝑉𝑢,𝑘|𝑡,𝑧 𝜕𝑉𝑢,𝑘|𝑡,𝑧
2
𝜎𝑢,𝑘 𝐹𝑡,𝑢 + 𝜇𝑢 𝐹𝑡,𝑢 𝑉𝑢,𝑘|𝑡,𝑧 − 𝜇𝑢 𝐶𝑢,𝑘|𝑡,𝑧
= 𝜕𝑢 = 2 𝜕𝑢
2 𝜕 2 𝑉𝑢,𝑘|𝑡,𝑧 𝜕𝑉𝑢,𝑘|𝑡,𝑧 𝜕 𝑉𝑢,𝑘|𝑡,𝑧 𝜕𝑉𝑢,𝑘|𝑡,𝑧
𝐹𝑡,𝑢 − 𝐹𝑡,𝑢 −
𝜕𝑘 2 𝜕𝑘 𝜕𝑘 2 𝜕𝑘 (69)
2
𝜕𝑉𝑢,𝑘|𝑡,𝑧 𝜎𝑢,𝑘 𝜕 2 𝑉𝑢,𝑘|𝑡,𝑧 𝜕𝑉𝑢,𝑘|𝑡,𝑧
⟹ = ( − )
𝜕𝑢 2 𝜕𝑘 2 𝜕𝑘
The PDE (69) appears drift-less and provides more robust calibration stability at low volatility and/or
A (discrete) dividend pay-out will typically result in a drop in equity price on the ex-dividend
date. Suppose that time 𝑢 is the ex-dividend date, the no-arbitrage condition states that at 𝑢+ the time
right after the ex-dividend date (e.g. the difference between 𝑢 and 𝑢+ can be infinitesimal), we must
have
19
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
𝑆𝑢+ = 𝑆𝑢 − 𝜃𝑢 (72)
where 𝜃𝑢 is the value of dividend issued at 𝑢 (note that in a rigorous setup the value must take into
account the discounting effect due to dividend payment delay). Since a forward is expectation of spot
̃ 𝑡 [𝑆𝑢 ] = 𝔼
𝐹𝑡,𝑢+ = 𝔼 ̃ 𝑡 [𝑆𝑢 − 𝜃𝑢 ] = 𝐹𝑡,𝑢 − 𝔼
̃ 𝑡 [𝜃𝑢 ] (73)
+
In our finite difference method, the spatial grid for log-moneyness 𝑘 is assumed uniform such that 𝑘𝑖 −
𝑘𝑖−1 is constant for all 𝑖. Dividend payment causes discontinuity in the underlying spot. Evolving the
forward PDE (69) from initial time 𝑡 produces a state vector 𝑉𝑢,𝑘|𝑡,𝑧 at time 𝑢. Immediately after the
issuance of dividend at time 𝑢+ , the spot and forward drop the same 𝜃𝑢 amount and hence the state
vector 𝑉𝑢+,𝑘|𝑡,𝑧 must be realigned to reflect the dividend fall. This can be done using the option no-
+ + +
̃ 𝑡 [(𝑆𝑢 − 𝐾) ] = 𝔼
𝐶𝑢+,𝑘|𝑡,𝑧 = 𝔼 ̃ 𝑡 [(𝑆𝑢 − 𝐹𝑡,𝑢 𝑒 𝑘̂ ) ] = 𝐶𝑢,𝑘̂|𝑡,𝑧
̃ 𝑡 [(𝑆𝑢 − 𝜃𝑢 − 𝐹𝑡,𝑢 𝑒 𝑘 ) ] = 𝔼
+ +
(75)
𝐹𝑡,𝑢+ 𝑒 𝑘 + 𝜃𝑢
where ̂
𝑘 = ln
𝐹𝑡,𝑢
Subsequently we can use 𝑘̂ to interpolate from the 𝑉𝑢,𝑘|𝑡,𝑧 state vector and transform the interpolated
If the dividend is proportional, we must have spot price 𝑆𝑢+ = 𝑆𝑢 (1 − 𝜂𝑢 ) for a rate 𝜂𝑢 and
hence forward price 𝐹𝑡,𝑢+ = 𝐹𝑡,𝑢 (1 − 𝜂𝑢 ) before and after the dividend fall. Because we can show that
1
Strictly speaking, a forward on time 𝑇 spot is an expectation of the spot under 𝑇-forward measure, i.e. 𝐹𝑡,𝑇 =
̂
𝔼𝑇𝑡 [𝑆𝑇 ]. However since the interest rate is assumed deterministic, the 𝑇-forward measure coincides with the risk
neutral measure.
20
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
̃ 𝑡 [(𝑆𝑢 − 𝐾)+ ]
𝔼 ̃ 𝑡 [(𝑆𝑢 − 𝐹𝑡,𝑢 𝑒 𝑘 )+ ]
(1 − 𝜂𝑢 )𝔼 ̃ 𝑡 [(𝑆𝑢 − 𝐹𝑡,𝑢 𝑒 𝑘 )+ ]
𝔼
+
𝑉𝑢+,𝑘|𝑡,𝑧 = = = = 𝑉𝑢,𝑘|𝑡,𝑧 (77)
𝐹𝑡,𝑢+ 𝐹𝑡,𝑢 (1 − 𝜂𝑢 ) 𝐹𝑡,𝑢
the state vector remains unchanged before and after the issuance of dividend.
With continuous dividend 𝑞𝑢 , the realignment of state vector is unnecessary because there is no
Again we assume the spot follows the SDE (21). Without loss of generality, let’s denote
𝐺(𝑆𝑇 |𝐾) an arbitrary payoff function with parameter 𝐾, whose value is contingent on 𝑆𝑇 at time 𝑇. One
example of such function would be the payoff function of a call option: 𝐺(𝑆𝑇 |𝐾) = (𝑆𝑇 − 𝐾)+ . Let
𝑈𝑢,𝑥|𝑇,𝐾 be the expectation of the function 𝐺(𝑆𝑇 |𝐾) at time 𝑢 with spatial variable 𝑥 = 𝑆𝑢 , which can be
written as
where the transition probability 𝑝𝑇,𝑦|𝑢,𝑥 follows the Kolmogorov backward equation (20)
𝜕𝑝𝑇,𝑦|𝑢,𝑥 2
𝜕𝑝𝑇,𝑦|𝑢,𝑥 𝜎𝑢,𝑥 𝑥 2 𝜕 2 𝑝𝑇,𝑦|𝑢,𝑥
= 𝜇𝑢 𝑥 + (79)
𝜕𝑢 𝜕𝑥 2 𝜕𝑥 2
In turn, we can derive the backward PDE for the 𝑈𝑢,𝑥|𝑇,𝐾 such that
𝜕𝑈𝑢,𝑥|𝑇,𝐾 𝜕𝑝𝑇,𝑦|𝑢,𝑥 2
𝜕𝑝𝑇,𝑦|𝑢,𝑥 𝜎𝑢,𝑥 𝑥 2 𝜕 2 𝑝𝑇,𝑦|𝑢,𝑥
= ∫ 𝐺(𝑦|𝐾) 𝑑𝑦 = − ∫ 𝐺(𝑦|𝐾) (𝜇𝑢 𝑥 + ) 𝑑𝑦
𝜕𝑢 ℝ 𝜕𝑢 ℝ 𝜕𝑥 2 𝜕𝑥 2
(80)
𝜕𝑈𝑢,𝑥|𝑇,𝐾 2
𝜎𝑢,𝑥 𝑥 2 𝜕 2 𝑈𝑢,𝑥|𝑇,𝐾
= −𝜇𝑢 𝑥 −
𝜕𝑥 2 𝜕𝑥 2
equivalent to 𝑈𝑢,𝑥|𝑇,𝐾 . The backward PDE (80) can then be transformed into
21
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
𝜕𝑈𝑢,𝑧|𝑇,𝑘 2
𝜎𝑢,𝑧 𝜕 2 𝑈𝑢,𝑧|𝑇,𝑘 𝜕𝑈𝑢,𝑧|𝑇,𝑘
=− ( − ) (82)
𝜕𝑢 2 𝜕𝑧 2 𝜕𝑧
by using the following partial derivatives derived from the chain rule
𝐹𝑡,𝑢 𝑒 𝑧 − 𝜃𝑢 (86)
̃ [𝐺(𝑆𝑇 |𝐹𝑡,𝑇 𝑒 𝑘 )|𝑢+ , 𝐹𝑡,𝑢 𝑒 𝑧̂ ] = 𝑈𝑢 ,𝑧̂ |𝑇,𝑘
=𝔼 where 𝑧̂ = ln
+ +
𝐹𝑡,𝑢+
It is likely that if 𝑧 is sufficiently small (e.g. at lower boundary of spatial grid) we may end up with
𝐹𝑡,𝑢 𝑒 𝑧 − 𝜃𝑢 < 0, which makes the 𝑧̂ not well defined. A solution is to floor it to a small positive number,
e.g. taking max(10−10 , 𝐹𝑡,𝑢 𝑒 𝑧 − 𝜃𝑢 ) . This is valid because equity spot must be positive and the
𝑈𝑢+,𝑧̂ |𝑇,𝑘 flattens as 𝑧̂ goes to negative infinity. After the special treatment, we can use the 𝑧̂ to
interpolate from the 𝑈𝑢,𝑧|𝑇,𝑘 state vector and convert the interpolated value into vector 𝑈𝑢+,𝑧|𝑇,𝑘 .
With proportional dividend, the conclusion drawn for forward PDE still applies here and the
state vector remains unchanged before and after the dividend fall. With continuous dividend, the
22
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Due to the duality between the forward and backward PDE, it is evident that vanilla calls (or
puts) must admit the identity: 𝑈𝑡,𝑧|𝑇,𝑘 = 𝑉𝑇,𝑘|𝑡,𝑧 𝐹𝑡,𝑇 , where 𝑈𝑡,𝑧|𝑇,𝑘 is the forward call solved from
backward PDE (82) and 𝑉𝑇,𝑘|𝑡,𝑧 the normalized forward call solved from forward PDE (69). This
relationship can be used to check the correctness of implementation of the numerical engines of forward
For pricing some exotic options, e.g. barrier options, it is more convenient to use log-spot 𝓏 =
ln 𝑥 as the spatial variable. Similarly we can define 𝓀 = ln 𝐾. Let’s denote the (discounted) price of a
derivative product by
By taking into account the discount factor, it must follow the following backward PDE
𝜕𝑋𝑢,𝓏|𝑇,𝓀 𝜕𝑈𝑢,𝓏|𝑇,𝓀
= 𝑟𝑢 𝑋𝑢,𝓏|𝑇,𝓀 + 𝐷𝑢,𝑇
𝜕𝑢 𝜕𝑢
2
1 𝜕𝑈𝑢,𝓏|𝑇,𝓀 𝜎𝑢,𝓏 1 𝜕 2 𝑈𝑢,𝓏|𝑇,𝓀 𝜕𝑈𝑢,𝓏|𝑇,𝓀
= 𝑟𝑢 𝑋𝑢,𝓏|𝑇,𝓀 + 𝐷𝑢,𝑇 (−𝜇𝑢 𝑥 − ( − )) (88)
𝑥 𝜕𝓏 2 𝑥2 𝜕𝓏 2 𝜕𝓏
2
𝜎𝑢,𝓏 𝜕 2 𝑋𝑢,𝓏|𝑇,𝓀 2
𝜎𝑢,𝓏 𝜕𝑋𝑢,𝓏|𝑇,𝓀
=− 2
+ ( − 𝜇𝑢 ) + 𝑟𝑢 𝑋𝑢,𝓏|𝑇,𝓀
2 𝜕𝓏 2 𝜕𝓏
1 𝜕 2 𝑈𝑢,𝓏|𝑇,𝓀 𝜕𝑈𝑢,𝓏|𝑇,𝓀
= 2( − )
𝑥 𝜕𝓏 2 𝜕𝓏
23
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Again, extremely small 𝓏 may result in 𝓏̂ that is not well defined, we may floor the difference 𝑒 𝓏 − 𝜃𝑢
to a small positive number, e.g. taking max(10−10 , 𝑒 𝓏 − 𝜃𝑢 ) . The vector 𝑋𝑢,𝓏|𝑇,𝓀 can then be
The vector 𝑋𝑢,𝓏|𝑇,𝓀 can be interpolated from the 𝑋𝑢+,𝓏|𝑇,𝓀 using the 𝓏̂ .
With continuous dividend, the realignment of state vector is unnecessary because there is no
This section is devoted to discussing the construction of local volatility surface 𝜎(𝑢, 𝑘). There
are various ways to define the local volatility surface. The one that we would like to discuss is a 2-D
𝐾
function that is piecewise constant in maturity 𝑢 and piecewise linear in log-moneyness 𝑘 = ln (or in
𝐹𝑡,𝑢
delta for FX). The volatility surface comprises a series of volatility smiles 𝜎𝑗 (𝑘) for maturity 𝑡 < 𝑢1 <
⋯ < 𝑢𝑗 < ⋯ < 𝑢𝑚 = 𝑇 . At each maturity 𝑢𝑗 , volatility smile 𝜎𝑗 (𝑘) is constructed by linear
𝐾𝑖
interpolation between log-moneyness pillars 𝑘𝑖 = ln for strikes 𝐾1 < ⋯ < 𝐾𝑖 < ⋯ < 𝐾𝑛 and flat
𝐹𝑡,𝑢
extrapolation where the volatility values at 𝑘1 and 𝑘𝑛 are used for all 𝑘 < 𝑘1 and 𝑘 > 𝑘𝑛 , respectively.
The smile 𝜎𝑗 (𝑘) constructed at 𝑢𝑗 is assumed to remain constant over time for any 𝑢 between the two
24
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
Calibration of the local volatility surface is conducted in a bootstrapping manner starting from
the shortest maturity 𝑢1 . It is done by solving the forward PDE such that the local volatility surface is
able to reproduce the vanilla call prices at the prescribed log-moneyness pillars 𝑘𝑖 for each of the
maturities 𝑢𝑗 . The PDE can be solved using finite difference method1 on a uniform grid defined on log-
moneyness 𝑘 that extends to ±5 standard deviations of the underlying spot. The choice of boundary
condition has little impact to the solutions of vanilla option prices because at ±5 standard deviations the
transition probability becomes negligibly small. Our application uses linearity boundary condition for its
simplicity. To allow a higher tolerance to market data input and smoother calibration process, the
objective function may include a penalty term to suppress unfavorable concavity of a local volatility
smile. Again, there can be many ways to define the objective function as well as the penalty function. In
this essay, we will only focus on the simplest objective (e.g. at maturity 𝑢𝑗 ): the least square
where 𝑈𝑡,𝑧|𝑇,𝑘 is the normalized forward call price defined in (68), the superscript “BS” denotes the
theoretical price by Black-Scholes model and the “PDE” denotes the numerical value by forward PDE.
Note that without a penalty term, the minimization can lead to an exact solution given a proper2 implied
volatility surface.
In contrast to the calibration, the pricing of a barrier option relies on the backward PDE (88) in
line with proper terminal condition (i.e. payoff function) and boundary conditions defined by the
characteristics of the barrier option. Barrier options often demand a spatial grid defined on log-spot 𝓏 =
ln 𝑆𝑢 , which allows an easier fit of time-invariant barrier (e.g. with European or American type of
1
A brief introduction to finite difference method can be found in my notes “Introduction to Interest Rate Models”,
which can be downloaded from http://www.cs.utah.edu/~cxiong/.
2
A proper implied volatility surface should well behave and admit no arbitrage.
25
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
observation window) into the domain. For example, an up-and-out barrier option would be priced on a
domain with upper bound at the barrier level 𝑏 where Dirichlet boundary condition is applied (the lower
bound and its boundary condition remain the same as for vanilla options).
26
Changwei Xiong, October 2017 http://www.cs.utah.edu/~cxiong/
REFERENCES
1. Clark, I., Foreign Exchange Option Pricing - A Practitioner’s Guide, Wiley-Finance, 2011, pp.82
2. Online resource: http://itf.fys.kuleuven.be/~nikos/papers/lect4_localvol.pdf
3. Gatheral, J., The Volatility Surface: A Practitioner’s Guide, Wiley-Finance, 2006, pp. 13-14
27