Equity Variance Swap with Dividends

Richard White∗

First version: 28 May 2012; this version†June 25, 2012

Abstract
We present a discussion paper on how to price equity variance swaps in the presence of
known (cash and proportional) dividends.

1 Variance swaps on non-dividend paying stock

In this section we present the standard theory of pricing a variance swap. Good references for
this are Demeterfi(1999), Neuberger(1992) and Carr(1998).
The payoff of a standard variance swap1 is given by:
" n   2 #
AX Si 2
V S(T ) = Nvar log −K (1)
n i=1 Si−1

where Si is the closing price of a stock or index on the ith observation date, the annualisation
factor A is set by the contract (and usually 252), Nvar is the variance notional and K 2 (the
variance strike) is usually chosen to make the initial value zero. The notional is often quoted in
units of volatility (i.e. $1M per vol), Nvol with Nvar = N2Kvol
.
Since all the other terms are fixed by the contract, we will focus on the realised variance part
n   2
X Si
RV (T ) = log (2)
i=1
Si−1

1.1 The Black-Scholes World

If the stock follows the process
dSt
= rdt + σdW (3)
St
then the returns are given by
 
Si
Ri ≡ log = (r − σ 2 /2)∆t + σ∆W (4)
Si−1
∗ [email protected]
† Version 2.0
1 The name swap is a bit of a misnomer, as it is a forward contract on the realised variance

1
That is, they are normally distributed with mean (r − σ 2 /2)∆t and variance σ 2 ∆t Ignoring the
drift term (which will be negligible next to the diffusion term for a period of one day), we can
write the realised variance as
Xn
2
RV (T ) = σ ∆t Zi2 (5)
i=1

where Zi are standard normal random variables. Hence the realised variance has a chi-squared
distribution with n degrees of freedom scaled by σ 2 ∆t - which is a Gamma distribution Γ(n/2, 2σ 2 ∆t).
The expected variance is simply
E[RV (T )] = σ 2 T (6)
but even in the Black-Scholes world, the realised variance can vary greatly from this.2

1.2 Other diffusions

    2
By considering the Taylor series of log SSi−1
i
and log SSi−1i
about Si−1 , it is easy to show
that 2
(Si − Si−1 )3
     
Si Si Si − Si−1
log ≈ −2 log +2 +O 3 (7)
Si−1 Si−1 Si−1 Si−1
Summing the terms gives Neuberger’s formula, Neuberger(1992).
  n
ST X 1
RV (T ) ≈ −2 log +2 (Si − Si−1 ) (8)
S0 i=1
Si−1

What this says is that you can replicate the realised variance up to the expiry, T , by holding
2 log(S0 ) zero coupon bonds with expiry at T , shorting two log-contracts3 and daily delta hedging
with a portfolio who’s delta is 2/Si−1 .
The expected value of the realised variance (or just expected variance) is

EV (T ) = 2 log(FT ) − 2E[log(ST )] (9)

and therefore the fair value of the variance strike is

r
2A
K= (log(FT ) − E[log(ST )]) (10)
n
where FT is the forward price. This is model independent provided the stock process is a diffusion,
Gatheral(2006).

1.3 Static Replication

Any twice differentiable function, H(x) can be written exactly as (Carr(1998))
Z x0 Z ∞
0 00
H(x) = H(x0 ) + (x − x0 )H (x0 ) + +
H (z)(z − x) dz + H 00 (z)(x − z)+ dz (11)
0 x0
p
2 The fractional error is approximately 2/n
3A log-contact has the payoff log(ST ) at expiry - these are not liquid but can be statically replicated with
strips of Europeans puts and calls

2
for x ≥ 0. The (non-discounted) value of a derivative at time t with a payoff H(ST ) at T is
E[H(ST )|Ft ]. Expanding H(ST ) using equation 11 we find

E[H(ST )|Ft ] = H(s∗ ) + (E[ST ] − s∗ )H 0 (s∗ )

Z s∗ Z ∞ (12)
+ H 00 (k)E[(k − ST )+ ]dk + H 00 (k)E[(ST − k)+ ]dk
0 s∗

Choosing s∗ = FT and recognising that E[(ST − k)+ ] and E[(k − ST )+ ] are (non-discounted) call
and put prices at strike k, we can write
Z FT Z ∞
E[H(ST )|Ft ] = H(FT ) + H 00 (k)P (k)dk + H 00 (k)C(k)dk (13)
0 FT

So in principle, one can replicate any payoff that is a function of the underlying at expiry, by an
infinite strip of puts and calls on the underlying, with weights H 00 (k). For our log contract we
have H 00 (k) = −1/k 2 , so the expected variance can be written as
Z FT Z ∞ !
P (k) C(k)
EV (T ) = 2 dk + dk (14)
0 k2 FT k2

In practise only a finite number of liquid call and put prices are available in the market
and the expiries may not match that of the variance swap. What is needed is an arbitrage free
interpolation of option prices. For a call with expiry T and strike k we have
• A call with a zero strike equals the forward C(T, 0) = FT
∂C(T,k)
• Positive calendar spreads, ∂T ≥0
∂C(T,k)
• Call price is monotonically deceasing in strike, ∂k ≤0
∂ 2 C(T,k)
• Positive butterflies, ∂k2 ≥0
• High strike limit, c(T, ∞) = 0
• If the underlying cannot reach zero (e.g. Exponential Brownian Motion (EBM)) then,
∂C(T,0)
∂k = −1, otherwise (e.g. SABR) ∂C(T,0)
∂k > −1
The corresponding conditions for puts follow from put-call parity. We carry out the interpo-
lation in implied volatility space, using arbitrage free smile models for interpolation in the strike
direction - this is detailed in White(2012). To extrapolate to strikes outside the range of market
quotes, we use a shifted log-normal model. This is just the Black option pricing formula where
the forward and the volatility are treated as free parameters which are calibrated so that either
the last two option prices are matched, or the price and dual delta4 of the last option is matched.
This is done separately for low strikes (puts) and high strikes (calls).
The last practical detail is to find an upper cut-off for the integral. Using the monotonically
decresing property, we know that for a cut-off of A we have
Z ∞ Z ∞
C(k) C(A) C(A)
2
dk ≤ 2
dk = (15)
A k A k A
So having integrated up to A we have an upper limit on the error.
4 dual delta is the rate of change of a option price with respect to the strike

3
 Using L’Hôpital’s rule, we find
P (k) 1 ∂ 2 P (T, 0)
lim 2
= (16)
k→0 k 2 ∂k 2
Which is half the probability density at zero, and is zero for our shifted log-normal extrapolation.
This limit on the integrand should be used in any numerical integration routine. Alternatively
we have for very small A
Z A
P (k) P (A)
2
dk ≈ +θ (17)
0 k 2A
where θ is the probability mass at zero (zero for exponential Brownian motion but non-zero
for SABR etc). This can again be used to set a limit on the error.

2 Dividends
In this section we follow the approach of Buehler (2010) and Bermúdez (2006)5 .
Companies make regular dividend payments (e.g. every six months) of a fixed amount per
share, with the amount announced well in advance6 , so dividends over a 1-2 year horizon can be
treated as fixed cash payments (per share) regardless of the share price. Further in the future, the
dividend payment is likely to depend on company performance, hence we can model payments
as being a fixed proportion of the share price on the ex-dividend date. In general dividends are
paid7 at times τ1 , τ2 , ... from today, with a cash amount αi and a proportional about βi . If the
stock price immediately before the dividend payment is Sτi − , then the price just after must be
Sτi = Sτi − (1 − βi ) − αi (18)
If the stock price follows a diffusion process, then this implies that it can go negative. One
way to avoid this is by making the dividend payment a (non-affine) function of the stock price,
Di (Sτi − ) such that Di (Sτi − ) ≤ Sτi − (see Haug(2003) and Nieuwenhuis(2006)). Bermúdez (2006)
and Buehler (2010) show that their affine dividends assumption leads to a modified stock price
process8 .
Consider a forward contract to deliver one share at maturity T in exchange for a payment
of k. If we start with δt shares then we will receive a dividend of δt (Sτ1 − β1 + α1 ) on the first
dividend date after t. Rewriting this in terms of the post-dividend price Sτ1 we have
 
β1 α1
Dividend Payment1 = δt Sτ1 + (19)
1 − β1 1 − β1
We use the part that is proportional to Sτ1 to buy new stock immediately after the payment,
then the stock holding will be
δt
δ τ1 = (20)
1 − β1
Our residual cash may be written as δτ1 α1 . Q
If we repeat the procedure for each dividend payment, then at T we will own δT = δt / j:t<τj ≤T (1 − βj )
shares. To have exactly one share at maturity, set
Y
δt = (1 − βj ) (21)
j:t<τj ≤T

5 Itshould be noted that both the book and the paper contain considerable typos in the mathematics
6 Although there may be no legal obligation to actually pay this amount if the company’s performance is poor.
7 We assume the ex-dividend date and the payment date coincide.
8 They also consider credit risk and repo rates, which we ignore are present.

4
Then at some time s with t < s ≤ T we own
Y
δs = (1 − βj ) (22)
j:s<τj ≤T

shares. At each dividend date we also have residual cash equal to δτi αi . As these are fixed, we
can realise their value now by selling δτi αi zero coupon bonds expiring at τi . The total value of
the sale is X
δτi αj P (t, τj ) (23)
j:t<τj ≤T

If we define the growth factor as

1 Y
R(t, T ) = (1 − βj ) (24)
P (t, T )
j:t<τj ≤T

then the total initial investment that guarantees one share at time T is
 
X α j
P (t, T )R(t, T ) St −  (25)
R(t, τj )
j:t<τj ≤T

We can fund this by selling k zero coupon bonds with maturity T . The value of k that makes
the initial investment zero is the forward price, therefore
 
X α j
F (t, T ) = R(t, T ) St −  (26)
R(t, τj )
j:t<τj ≤T

SincePthe share price cannot become negative, neither can the forward. This implies that
α
St ≥ j:t<τj ≤T R(t,τj j ) , and since this must be independent of the arbitrary maturity, T , of
a forward, we must have X αj
St ≥ Dt ≡ (27)
j:τ >t
R(t, τj )
j

That is, the dividend assumption imposes a floor on the stock price equal to the growth factor
discounted value of all future cash dividends.9
As a result of the floor on the stock price, the Black model, St = Ft Xt ,10 where Xt is a EBM
with E[Xt ] = 1, is no longer valid since P(St < Dt ) > 0.11
Bermúdez (2006) and Buehler (2010) propose

St = (Ft − Dt )Xt + Dt (28)

with E[Xt ] = 1 and Xt ≥ 0.12 The stock price experiences randomness just on the portion above
the floor Dt . Xt is known as the pure stock process.
9 We have overloaded the letter D - D is the growth factor discounted value of all future cash dividends, while
t
D(τ ) is the dividend paid at time τ
10 F ≡ F (0, t)
t
11 Any other process for X is also invalid for the same reason.
t
12 Their model also includes credit risk, so they impose X > 0 and have the stock price and all future dividends
t
fall to zero on a default. The setup with no credit risk and Xt ≥ 0 implied that the stock becomes a bond with
fixed coupons if xt becomes zero.

5
2.1 European Options
Consider the price at time t of two european call options, one with expiry τ − (i.e. just before
the dividend payment) and one with expiry τ (i.e. just after the payment). We write the prices
as Ct (τ − , k − ) & Ct (τ, k). The first payoff is
 +
− + Sτ + α − 1
+
(Sτ − − k ) = −k = Sτ − [(1 − β)k − − α] (29)
1−β 1−β
If we let k = (1 − β)k − − α then the payoff of the first option is just 1/(1 − β) times that of the
second. It then follows that
(1 − β)Ct (τ − , k − ) = Ct (τ, (1 − β)k − − α) (30)
If the options are priced using Black’s formula with the same volatility, then the condition of
equation 30 does not hold. Conversely, if the condition does hold, and (Black) implied volatilities
are found, they will jump up across the dividend date. Once again, the Black model, or indeed
any model that has a continuous implied volatility surface, is not consistent with the dividend
assumption.
Another consideration is what happens to the price of an option with expiry, T , as the current
time, t, moves over a dividend date. We now write the call price as C(t, St , T, k). Clearly
C(τ − , Sτ − , T, k) = C(τ, Sτ , T, k) (31)
This is true for any European option (e.g. a log-payoff) with T > τ . This condition must be
applied when solving a backwards PDE for option prices, as discussed later.

2.2 Options on the Pure Stock

Define a call option on the pure stock as
C̃(T, x) = E[(XT − x)+ ] (32)
This is related to a real call by
1
C̃(T, x) = C (T, (FT − DT )x + DT )
P (0, T )(FT − DT )
  (33)
k − DT
C(T, k) = P (0, T )(FT − DT )C̃ T,
FT − DT
Hence if the dividend structure is known, market prices for calls (and puts) can be converted
into pure call prices - which is turn can be converted to a pure implied volatility by inverting
the Black formula in the usual way.
Figure 1 shows the implied volatility surface if the underlying prices are generated from a flat
pure implied volatility surface at 40%, and the pure implied volatility surface if the prices are gen-
erated from a flat implied volatility surface at 40%. These shapes are observed in Buehler(2010),
but without the jumps at the dividend dates. The condition leading to equation 30 means
we expect a jump in implied volatility if the pure implied volatility is flat and vice versa. As
Buehler does not state the dividends schedule, we cannot make a direct comparison. However if
we increase the dividend frequency ten-fold and decrease each amount ten-fold (which may be
more the case for indices)13 , then the implied volatility surface (for a flat pure implied volatility
surface) becomes as shown in figure 2.
13 We have cash only dividends in year one, proportional only after 3 years and a linear switch over in year 2

6
 0.39 0.46

0.38 0.45

0.37 0.44

Pure  Implied  Volatility  

 Implied  Volatility  

0.36 0.43

0.35 0.42

0.34 1.9801 0.41 1.9801

1.7612 1.7612
1.5423 1.5423
0.33 1.3234 0.4 1.3234
1.1045 1.1045
0.8856 0.8856
0.32 0.39
0.6667 0.6667
50
54.5

0.5
59

0.545
63.5

0.59
68

0.635
72.5

0.68
77

0.725
0.4478
81.5

0.77
86

0.4478

0.815
90.5

0.86
95

0.905
99.5

0.95
104

0.995
108.5

1.04
113

1.085
117.5

1.13
122

1.175
126.5

1.22
0.2289
131

1.265
0.2289
135.5

1.31
Time  to  Expiry   Time  to  Expiry  
140

1.355
144.5

1.4
149

1.445
153.5

1.49
158

1.535
162.5

1.58
1.625
167
171.5

1.67
0.01
176

1.715
0.01

1.76
180.5
185

1.805
1.85
189.5
194

1.895
1.94
198.5

1.985
Strike   Strike  

Figure 1: Implied Volatility (left) and Pure Implied volatility (right) for an initial stock price is
100, dividends every six months and the next is in one month, with a proportional part equal
to 1% of the stock price (β = 0.01) and a cash part of 1.0, and the risk free rate is 5%. In each
case the prices were generated from a flat pure implied volatility surface (left), or a flat implied
volatility surface (right) - both at 40%

0.41

0.4

0.39
 Implied  Volatility  

0.38

0.37
1.9801
1.7612
1.5423
0.36
1.3234
1.1045
0.8856
0.35
0.6667
30
35.1
40.2
45.3
50.4
55.5
60.6

0.4478
65.7
70.8
75.9
81
86.1
91.2
96.3
101.4
106.5

Time  to  Expiry  

111.6
116.7

0.2289
121.8
126.9
132
137.1
142.2
147.3
152.4
157.5
162.6
167.7

0.01
172.8
177.9
183
188.1
193.2

Strike  
198.3

Figure 2: Implied Volatility surface for a pure implied volatility surface of 40%, and 20 dividend
payments per year

7
2.3 Dupire Local Volatility for Pure Stock Process
Following exactly the same argument as Dupire(1994), if a unique solution to the SDE
dXt
= σ X (t, Xt )dWt (34)
Xt
exists, then the pure stock local volatility is given by:
∂ C̃
σ X (t, Xt )2 = 2 ∂T
2
(35)
x2 ∂∂xC̃2
Once we have the pure option prices (or equivalently the pure implied volatilities) from market
prices, we can proceed as in White(2012) to construct a smooth Pure Local Volatility surface,
and numerically solve PDEs without having to impose jump conditions at the dividend dates.
Conversely, if we solve a PDE expressed in the stock price, then at dividend dates the nodes
must be shifted such that for the ith stock price node siτ = siτ − − D(Sτ − ), where D(Sτ − ) is the
dividend amount, and the option price maps according to equation 31. Between dividend dates
the local volatility is  
St − Dt X St − Dt
σ(t, St ) = 1St >Dt σ t, (36)
St Ft − Dt

3 Expected Variance in the Presence of Dividends

3.1 Proportional Dividends Only
The stock price process can be written as
dSt X
= rt dt + σt (St− )dWt − βj δ(τj ) (37)
St− j

where the volatility is a function of the stock price immediately prior to any jump (and possibly
another stochastic factor), and δ(τj ) indicates that jumps only occur at one of the dividend dates
τj .
If we define yt = log(St ) then
X
dyt = (rt − σt2 /2)dt + σt dWt + log(1 − βj )δ(τj ) (38)
j

and Z S Z S X
ys = yt + (rt0 − σt20 /2)dt + σt0 dWt0 + log(1 − βj ) (39)
t t j:t<τj ≤s

Which is a formal solution since in general σt is a function of yt . The expected squared return is
   Z ti
2 Si X
E ln ≈ σt2 dt + log2 (1 − βj )δτj =ti (40)
Si−1 ti−1 j

which is the quadratic variance of the continuous part of the log-process, plus the contribution
(if any) from the dividend. The total expected variance is then
Z T X
EV (T ) ≈ σt2 dt + log2 (1 − βj ) (41)
0 j:τj ≤T

8
Of course, if dividends are corrected for (as is usual for single stock), then the second term is
absent.
Again from equation 45 the expected value of a payoff of log(ST ) is
Z T
1 T 2
X Z
E[log(ST )] ≡ E[yT ] = log(S0 ) + rt dt + log(1 − βj ) − σ dt
0 2 0 t
j:τj ≤T
(42)
1 T 2
Z
= log(FT ) − σ dt
2 0 t
where the replacement of the first three terms by the forward comes from equation 26. It then
follows immediately that
 
ST X
EV (T ) = −2E log + log2 (1 − βj ) (43)
FT
j:τj ≤T

So for proportional dividends only, where an adjustment is made for the dividend, the expected
variance is the same as the no dividend case,14 and otherwise there is a trivial adjustment.

3.2 With Cash Dividends

Following the argument in Buehler(2006), the realised variance can be expressed as
n
Sτj (1 − βj )
   
X
2 Si− X
2
RV (T ) = log + log (44)
Si−1 Sτj + αj
i=1 j:t<τj ≤T

Si− is the stock price just before the dividend payment (if one occurs on that observation
date), hence the first term is the realised variance of the continuous part, RV (T )cont . If dividends
are corrected for, then the second term disappears, and the realised variance is the same as the
realised variance of the continuous part, as above.
The SDE for the log of the stock price is given by
Sτj (1 − βj )
 
1 1 X
d(log St ) = dSt − dRVcont + log δτj (dt) (45)
St− 2 j
Sτj + αj
 S (1−β ) 
τ j
The last term simply means that at dividend dates the log of the stock jumps by log Sjτ +αj ,
j
but is otherwise continuous. Integrating equation 45 and combining with equation 44 gives
Z T
ST 1
RV (T ) = − 2 ln +2 dSt
S0 0 S t−
Sτj (1 − βj ) Sτj (1 − βj )
X   X  
2
+2 log + log
Sτj + αj Sτj + αj
j:τj ≤T j:t<τj ≤T
Z T Z T (46)
ST
= − 2 ln +2 rt dt + 2 σt dWt
S0 0 0
Sτj (1 − βj ) Sτj (1 − βj )
X   X  
+2 log + log2
Sτj + αj Sτj + αj
j:τj ≤T j:t<τj ≤T

14 The big caveat here is that option prices will not be the same with and without dividends, so the value will

be different.

9
Taking expectations we arrive at the result

 
ST X  
EV (T ) = −2E log − 2 log(P (0, T )) + 2 E Gj (Sτj )
S0
j:τj ≤T
    (47)
s(1 − βi ) 1 2 s(1 − βi )
where Gj (s) ≡ log + log
s + αi 2 s + αi

Unlike the pure proportion case, we cannot rewrite this in terms of the forward as

Sτj (1 − βi )
X   
log(FT ) 6= log(S0 ) − log(P (0, T )) + E log (48)
Sτj + αi
j:τj ≤T

The correction term is a (twice differentiable) function of the stock price just after a dividend
payment, so can be priced with a strip of call and puts with expiry τj as in equation 13. When
dividends are corrected for (the usual case for single stock), then last term is removed. An
alternative derivation of this result is given in appendix A.
We now have a mechanical procedure to value variance swaps when dividends are paid on
the underlying stock and the stock price is a pure diffusion between dividend dates.
• Estimate the future dividend schedule. This can come for declared dividends (for short
time periods), equity dividend futures and calendar spreads.

• Collect prices of all European options on the underlying stock or index15 , and convert to
prices of pure put and call options
• Construct the pure local volatility surface
• solve the forward PDE up to the expiry of the variance swap

• Convert the pure option prices on the PDE grid, to pure implied volatility to form a
continuous (interpolated) surface
• Convert back to option prices at arbitrary expiry and strike
• Compute the value of the log-contact and the correction at each dividend date using equa-
tion ??.

3.3 The Buehler Paper

In Buehler (2010), the realised variance which is not adjusted for dividends is given as
n  
ST X Sti X
RV (T ) = −2 log +2 −1 +2 Ej (Sτj )
S0 Sti−1
i=1 j:τj ≤T
 2 (49)
s(1 − βj ) sβj + αj sβj + αj
where Ej (s) = log + −
s + αj s + αj s + αj
15 Most equity options are American style, but we will ignore this important detail for now.

10
Taking expectations, we find the expected variance is
  n
ST X h i
EV (T ) = −2E log − 2 log(P (0, T )) + 2 E Êi (Sτi )
S0 i=1
 2 (50)
s(1 − βj ) sβj + αj
where Êj (s) = log −
s + αj s + αj
This is not the same as our equation 47, since the last term (which only applies if dividends are
 2  
sβj +αj 1 2 s(1−βi )
not corrected for) is − s+α j
rather than 2 log s+αi . We believe this is a mistake in
the paper, Buehler (2010), and have confirmed this using Monte Carlo methods described in the
next section.

4 Numerical Testing
We present four test cases with exaggerated parameters to highlight any errors. All cases have a
spot of 100, a pure local volatility of 30%, a fixed interest rate of 10%, a single dividend payment
at 0.85 years and expiry of the variance swap at 1.5 years. The first two cases have α = 0 and
β = 0.4, the former with and the latter without dividends corrected for in the realised variance.
The second two cases have α = 30 and β = 0.2, again with and without dividends correction.
Since the pure local volatility is flat, we can find call and put prices from the modified Black
formula of equation 33, which allows us to find the expected variance using equations 47 with
the static replication of equation 13 (once at expiry and once at the dividend date). In the case
of pure proportion dividends (first two cases), we have a simple analytical answer via equation
4316 . All the numerical methods do not depend on a flat pure local volatility surface.
Another method is to solve the backwards PDE with the log-payoff as the initial condition.
This can be done using the pure stock or the actual stock as the spacial variable - in the latter
case the jump condition must be applied at the dividend. Since this method does not calculate
the terms G(S), it is only applicable to the first test case (α = 0 with dividends corrected for in
the realised variance).
Solving the forward PDE for the pure call prices up to expiry, gives actual call and put prices
on a expiry-strike grid, which can be used as in the static replication case to price all four cases.
This in our principle numerical method.
Finally, it is simple to set up Monte Carlo to calculate expected variance in all four cases.17
2
The results arep shown in table 1 below, expressed in terms of K where K is the variance
strike (i.e. K = EV (T )/T ). The agreement to 4 dp of the forward PDE to the Monte Carlo,
gives use confidence in the method.

5 Concussion
We have reviewed the pricing of variance swaps, and the affine dividend model of Buehler et al.
We believe that there is a mistake in the paper Buehler(2010), and that the expression presented
here is correct.
A more general dividend model would require more numerical work to calibrate and price
variance swaps.
h i
ST
16 For a flat pure local volatility, σ x we have E log F = − 12 (σ x )2
T
17 Since the purpose of the Monte Carlo is to test the other methods, we are free to use a very large number of

paths.

11
 Case1 Case2 Case3 Case4
Analytic 0.3 0.5138 N/A N/A
Static Replication 0.3000 0.5138 0.2463 0.6035
Backwards PDE (pure stock) 0.3000 N/A N/A N/A
Backwards PDE (actual stock) 0.3000 N/A N/A N/A
Forward PDE 0.3000 0.5138 0.2463 0.6035
Monte Carlo 0.3000 0.5136 0.2461 0.6035
p
Table 1: The expected variance expressed as K = EV (T )/T for the 4 test cases and 5 numerical
methods discribled above.

A Alternative Derivation
Assuming that Si is the closing price on date ti and that dividends are paid just before this, so
that
Si = Si− − Di (51)
where Di is the amount of the dividend (which is zero except for dividend dates). The realised
variance can be written as
n   X n  
X Si Si− − Di Si−
RV (T ) = log2 = log2
i=1
Si−1 i=1
Si− Si−1
n    
X
2 Si− X
2 Sτj
= log + log (52)
Si−1 Sτj + Dj
i=1 j:t<τj ≤T
!
X 
Sτj
 Sτ −
j
+2 log log
Sτj + Dj Sτ − −∆t
j:t<τj ≤T j

The last term (which is the return of continuous part multiplied by the return due to the dividend
payment) is negligible and can be ignored. This is the same as the result show in equation 44.
The first term (which is the squared returns of the continuous part), can be approximated as in
equation 7.    
Si− Si− S − − Si−1
log2 ≈ −2 log +2 i (53)
Si−1 Si−1 Si−1
We rewrite the first term as
     
Si− Si Si + Di
log = log + log (54)
Si−1 Si−1 Si

Hence
n
X Si− − Si−1
RV (T ) ≈ −2 log(ST /S0 ) + 2
i=1
Si−1
    (55)
X Sτj X
2 Sτj
+2 log + log
Sτj + Dj Sτj + Dj
j:t<τj ≤T j:t<τj ≤T

βj Sτj +αj
If we take the expectation and set Dj = 1−βj we arrive at the result in equation 47.

12
References
Bermúdez, Ana (2006). Equity hybrid derivatives. Wiley Finance.
Buehler, Hans (2006). Options on Variance: Pricing and Hedging. IQPC Volatility Trading
Conference. http://www.quantitative-research.de/dl/IQPC2006-2.pdf
Buehler, Hans (2010). Volatility and Dividends. Working paper available at http://papers.
ssrn.com/sol3/papers.cfm?abstract_id=1141877]
Carr, Madan (1998). Towards a Theory of Volatility Trading. In ”Volatility: New Estimation
Techniques for Pricing Derivatives,” R. Jarrow (ed.) RISK Publications, London.
Demeterfi, Derman, Kamal, Zou (1999). More Than You Ever Wanted To Know About Volatil-
ity Swaps. Goldman Sachs Quantitative Strategies Research Notes.
Dupire, B. (January 1994). Pricing with a Smile. Risk Magazine, Incisive Media.
Gatheral, Jim. (2006). The Volatilty Surface. Wiley Finance.
Haug, Espen; Haug, Jorgen; Lewis, Alan (2003). Back to Basics: a new approach to the dis-
crete dividend problem.
Neuberger, A (1992). Volatility Trading. London Business School working paper.
Nieuwenhuis, J.W.; Vellekoop, M. H. (2006). Efficient Pricing of Derivatives on Assets with
Discreate Dividends. Applied Mathematical Finance.
White, Richard R (2012). Local Volatility. OpenGamma working paper.http://docs.
opengamma.com/display/DOC/Quantitative+Papers+and+Working+Notes

13