Estimating the Volatility in the Black-Scholes Formula

Rebecca Keenan1* , Rachel Lane2* , Josh Matti3* , Hui Gong4†

Affiliation
1. Department of Mathematics and Computer Science, Eastern Connecticut State University, Willimantic,
CT 06226
2. Department of Mathematics and Computer Science, Concordia College, Moorhead, MN 56562
3. Division of Mathematics and Computer Information Sciences, Indiana Wesleyan University, Marion,
IN 46953
4. Department of Mathematics and Computer Science, Valparaiso University, Valparaiso, IN 46383

Abstract
The Black-Scholes formula is one of the most popular option pricing models; however, one of the in-
puts, volatility, is not deterministic and thus not available for immediate application in the formula. In our
research, we examine four different approaches for better estimating the volatility: smoothing, deriving the
distribution of the volatility, building time series models, and nonparametric approaches. We employ both
single and double exponential smoothing techniques on European call option valuations for the S&P 100
Index. We derive a function for σ 2 and σ and calculate their expected values. Secondly, we derive the proba-
bility distributions of the volatility with a transformation technique. The expectations of the volatility from
the probability distributions are then applied back to the Black-Scholes formula. Additionally, we extend
the cumulative normal distribution functions in the Black-Scholes formula using a Taylor series expansion to
arrive at functions of volatility. With time series volatility models, we apply Autoregressive Conditional Het-
eroscedasticity (ARCH) and Generalized Autoregressive Conditional Heteroscedasticity (GARCH) volatility
for application into the formula. Similarly to these two well-discussed volatility models, we purpose three
new time series volatility models: Moving Average Conditional Heteroscedasticity (MACH), Autoregressive
Moving Average Conditional Heteroscedasticity (ARMACH), and Generalized Autoregressive Moving Av-
erage Conditional Heteroscedasticity (GARMACH). With the nonparametric approaches we do not assume
any probability distributions and instead calculate volatility both from average sample variance as well as
weighted sample variance.

Keywords: Black-Scholes formula, option pricing, volatility models, exponential smoothing

1 Introduction
An option is a type of financial contract where the owner has the right, but not the obligation, to buy
or sell a stock at a certain price (strike price) before a certain date (expiration date). The Black-Scholes
∗ All of these authors made equal contributions to the study and the publication.
† Correspondence Author: hui.gong@valpo.edu

1
formula is one of the most popular option pricing models due to its compact form and computational ease.
It was first introduced by Fischer Black and Myron Scholes in their 1973 paper,“The Pricing of Options
and Corporate Liabilities” [2]. From their stochastic partial differential equation model, the Black-Scholes
formula can be deduced.

C = Φ(d1 )S − Φ(d2 )Ke−rT

S
ln( K
2
)+(r+ σ2 )T √
d1 = √
σ T
and d2 = d1 − σ T
where
C = premium for call option
Φ = cumulative density function of a
standard normal distribution
T = expiration date
S = stock price at time t0
K = strike price
r = risk-free interest rate
σ = volatility

It gives a theoretical estimate of the price of a European-style option. The formula’s arrival led to an
explosion in options trading that shaped the world’s financial markets. However, despite its simplicity and
widespread use, it is not a perfect formula. One of the inputs, volatility, or σ, is not deterministic and thus
not available for immediate application in the formula. A generic approach is to compute the volatility of
the stock price, St , prior to time t0 . However, this approach may not produce the ideal result. A search for
a better estimate for the volatility has generated several different approaches each with varying degrees of
success. One of the most popular methods is using time series models that allow for volatility to change over
time. Allowing for volatility in one time period to be dependent on the volatility during previous time periods
is important in financial modeling due to volatility clustering. Following the groundbreaking introduction of
the Autoregressive Conditional Heteroscedasticity (ARCH) model in Engle (1982) [4], subsequent research
has explored new models that allow for volatility to change over time.
In this paper, we derive a theoretical apparatus for four different approaches: smoothing, deriving the
distribution of the volatility, building time series models, and nonparametric techniques. We test the dif-
ferent approaches numerically using three different numerical comparisons: sum of squared errors of predic-
tion (SSE), root-mean-square error (RMSE), and percent bias. Additionally, we propose three new times
series models: Moving Average Conditional Heteroscedasticity (MACH), Autoregressive Moving Average
Conditional Heteroscedasticity (ARMACH), and Generalized Autoregressive Moving Average Conditional
Heteroscedasticity (GARMACH).
The following paper is organized as follows: In Section 2, we present new single and double exponential
smoothing techniques. Section 3 contains time series models used to estimate the variance from returns data.
Section 4 focuses on deriving the distribution of the volatility. In this section we derive the expectation of σ
and σ 2 using a Jacobian transformation, the expectation of Φ(d1 ) and Φ(d2 ) using a Taylor series expansion,
as well as the expectation for five different volatility models. Section 5 details our nonparametric approach
with average as well as weighted sample variances. Section 6 discloses the findings of our study and Section
7 concludes with a summary of those findings as well as recommendations for future work.

2 Smoothing the Data

In this approach, we modify the standard single and double exponential smoothing formulas. Smoothing
transforms the data to make it less volatile. Additionally, it makes it easier to compare over time. With

2
exponential smoothing we are able to apply more weight to more recent values. In this section and beyond,
instead of working with the stock price, St , we will work with the returns, which are defined as xt = log( SSt−1
t
).

2.1 Single Exponential Smoothing

Consider the standard single exponential formula,
At = αxt−1 + (1 − α)At−1 , with A1 = x1
where xt = observed value at time t, At = smoothed value at time t, and α is a coefficient between 0 and 1.
This formula is recursive, but through repeated iterations we are able to generalize the formula. Addi-
tionally, we set At = xt rather than A1 = x1 . In this way we are setting the most recent smoothed value
equal to the most recent observed value.

At − αxt−1
At−1 =
1−α
At−1 − αxt−2
At−2 =
1−α
At −αxt−1
1−α − αxt−2
=
1−α
At αxt−1 αxt−2
= − −
(1 − α)2 (1 − α)2 1−α
At−2 − αxt−3
At−3 =
1−α
At αxt−1 αxt−2 αxt−3
= − − −
(1 − α)3 (1 − α)3 (1 − α)2 1−α
..
.
At − αxt−1 αxt−2 αxt−3 αxt−i
At−i = i
− i−1
− i−2
− ··· −
(1 − α) (1 − α) (1 − α) (1 − α)

2.2 Double Exponential Smoothing

When the time series data displays a trend, double exponential smoothing is more appropriate. Consider
the standard double exponential formulas ,

At = αxt + (1 − α)(At−1 + Bt−1 ), with A1 = x1

Bt = β(At − At−1 ) + (1 − β)Bt−1 , with B1 = x2 − x1

where xt = observed value at time t, At = smoothed level at time t, Bt = smoothed trend at time t, and α
and β are coefficients between 0 and 1.
The formula is similar to single exponential smoothing but with an added term, Bt to account for trend.
However, despite this added term we can use a similar process of repeated iterations to arrive at a generalized,
rather than recursive, formula.

3
 At − αxt
At−1 = − Bt−1
(1 − α)
At−1 − αxt−1
At−2 = − Bt−2
1−α
[ At −αxt − Bt−1 ] − αxt−1
= 1−α − Bt−2
1−α
At − αxt Bt−1 + αxt−1
= − − Bt−2
(1 − α)2 1−α
..
.
At − αxt Bt−1 + αxt−1 Bt−2 + αxt−2 Bt−i+1 + αxt−i+1
At−i = − − − ··· − − Bt−i
(1 − α)i (1 − α)i−1 (1 − α)i−2 (1 − α)

Bt − β(At − At−1 )
Bt−1 =
1−β
Bt−1 − β(At−1 − At−2 )
Bt−2 =
1−β
[ Bt −β(At −At−1 )
1−β ] − β(At−1 − At−2 )
=
1−β
Bt − β(At − At−1 ) β(At−1 − At−2 )
= −
(1 − β)2 1−β
..
.
Bt − β(At − At−1 ) β(At−1 − At−2 ) β(At−2 − At−3 ) β(At−i−1 − At−i )
Bt−i = − − − ··· −
(1 − β)i (1 − β)i−1 (1 − β)i−2 1−β

Additionally, we can set At = xt and Bt = xt − xt−1 so that the most recent observed value will be
consistent with the most recent smoothed value. Although the new single and double exponential smoothing
formulas have been proposed, the estimation of the smoothing coefficients, α and β, are still under develop-
ment. For purposes of illustration, in the numerical comparison, we use the conventional single and double
exponential smoothing formulas with A1 = x1 and B1 = x2 − x1 .

3 Modeling the Returns

Considering the return, xt , we can apply some time series models to derive the volatility. The Autore-
gressive Moving Average (ARMA) model is a typical time series model. In this section, we apply three
different ARMA models to obtain an estimate of the volatility.

3.1 ARMA(1,1)
We begin with the ARMA(1,1) model. The ARMA(1,1) model accounts for the most recent return and
a random component.
xt = φ1 xt−1 + θ1 et−1 + et
where et ∼ N (0, σe2 ).

4
 To find the variance in this model we start with cov(xt , xt ), put the ARMA(1,1) model in for xt , and use
the following properties to solve:

cov(xt , et ) = σ2
cov(xs , et ) = 0 if s < t
var(xt ) = var(xs )

After several steps, we are able to arrive at a formula for the variance of returns:

var(xt ) = cov(xt , xt )
= cov(φ1 xt−1 + et + θ1 et−1 , φ1 xt−1 + et + θ1 et−1 )
= cov(φ1 xt−1 , φ1 xt−1 ) + cov(φ1 xt−1 , et ) + cov(φ1 xt−1 , θ1 et−1 ) + cov(et , φ1 xt−1 ) + cov(et , et )
+ cov(et , θ1 et−1 ) + cov(θ1 et−1 , φ1 xt−1 ) + cov(θ1 et−1 , et ) + cov(θ1 et−1 , θ1 et−1 )
= φ21 cov(xt−1 , xt−1 ) + cov(et , et ) + θ12 cov(et−1 , et−1 ) + 2cov(φ1 xt−1 , θ1 et−1 )
= φ21 var(xt−1 ) + var(et ) + θ12 var(et−1 ) + cov(et−1 , et−1 ) + cov(et1 , φ1 xt−2 ) + cov(et−1 , θ1 et−2 )
= φ21 var(xt ) + var(et ) + θ12 var(et ) + 2θ1 φ1 σ 2
= φ21 var(xt ) + σ 2 + θ12 σ 2 + 2θ1 φ1 σ 2

So

σ 2 (1 + θ12 + 2θ1 φ1 )
var(xt ) =
1 − φ21

3.2 Higher-Order ARMA Models

In similar fashion to the ARMA(1,1), we also examine the ARMA(2,2):

xt = φ1 xt−1 + φ2 xt−2 + θ1 et−1 + θ2 et−2 + et

and the ARMA(3,3):

xt = φ1 xt−1 + φ2 xt−2 + φ3 xt−3 + θ1 et−1 + θ2 et−2 + θ3 et−3 + et

and arrive at
σe2 [(θ12 + θ22 + 1) + 2(φ1 θ1 + φ2 θ2 + φ1 θ1 θ2 )]
var(xt ) =
1 − φ22 − φ22
for the variance of the ARMA(2,2) model and
3 3
σe2 [(1 + θi2 ) + 2[( φi θi ) + (φ1 θ1 )(φ1 θ2 + φ2 θ1 + φ21 θ3 ) + (φ1 φ2 θ3 + φ1 θ2 θ3 )]]
P P
i=1 i=1
var(xt ) =
1 − φ21 − φ22 − φ23

for the variance of the ARMA(3,3) model. Due to the number of steps required to derive these formulas,
the steps are shown in Appendix 1. Given the complexity beyond the ARMA(3,3) model, in practice, time
series data will not be fit to a higher order ARMA model than ARMA(3,3). The variance of ARMA(3,3)
can be downgraded to any lower order model, by simply setting the corresponding φ and θ equal to 0.

5
4 Deriving the Distribution of the Volatility
In this section we derive the expected value for σ and σ 2 , Φ(d1 ) and Φ(d2 ), as well as five volatility
models. Each expected value is then plugged back into the Black-Scholes formula to yield numerical results.

4.1 Functions of Volatility

One method for estimating σ 2 in the Black-Scholes formula is to start by deriving the probability density
function for σ 2 . Then, we can find the expected value of this function and apply the result back to the
Black-Scholes formula. Besides using this method with σ 2 , we also work with σ.

4.1.1 Expected value of σ 2

Given the setting of the Black-Scholes formula, the returns, xt , follow a normal distribution. Thus, we
2
know (t−1)s
σ2 follows a chi-squared distribution with t − 1 degrees of freedom [5], where s2 is the sample
variance of the returns and2 t is the total number of returns.
Let’s assume y = (t−1)s
σ2 . Then the probability density function of y, fy (y), can be written as
v −y
y ( 2 −1) e( 2 )
fy (y) = v
2( 2 ) Γ( v2 )
where v = t − 1. Employing a Jacobian transformation [7], we can derive the probability density function
for σ 2 :
2 t−1 −(t−1)s2

2 [ (t−1)s
σ2 ]( 2 −1) e( 2σ2 ) (t − 1)s2
f (σ ) = t−1 ·
2( 2 ) Γ( t−1 (σ 2 )2
2 )
. Proof. Let u = σ 2 . By Jacobian transformation we have,
dh−1
fu (u) = fy [h−1 (u)]| |
du
(t−1)s2 dh−1 −(t−1)s2
where h−1 (u) = u and du = u2 ,
After plugging all factors back into the transformation equation, we have the probability density function of
f (σ 2 ) as shown above.
The expectation of σ 2 can be derived by integration techniques [7]:
(t − 1)s2
E[σ 2 ] =
(t − 3)
. Work is shown in Appendix 2.

4.1.2 Expected value of σ

Using the same process as with E[σ 2 ] but allowing u = σ rather than u = σ 2 , we are able to arrive at an
expected value for σ: √
t − 1s( t−4 )!
E[σ] = √ t−3 2
2( 2 )!
Steps for this process are also shown in Appendix 2.

4.2 Taylor Series Approximations

Rather than finding the expected value of σ, we can derive the expected value of Φ(d1 ) and Φ(d2 ) which
are both functions of σ. We start by deriving a general form for the expected value of Φ(d1 ) and Φ(d2 ) by
using a Taylor series expansion. We can obtain numerical results by looking at some specific functions of s2 .

6
4.2.1 General Formulas
The generic Taylor series is given by:
f 0 (a) f 00 (a)
f (x) = f (a) + (x − a) + (x − a)2 + Rn
1! 2!
We modify the formula to solve for Φ(d1 ) and allow a to equal any function of s2 , the sample variance:
Φ0 (h(t)(s2 )) 2 Φ00 (h(t)(s2 )) 2
Φ(d1 ) = Φ(h(t)(s2 )) + (σ − (h(t)(s2 ))) + (σ − (h(t)(s2 ))2 ) + Rn
1! 2!
Thus,
Φ0 (h(t)(s2 )) 2 Φ00 (h(t)(s2 )) 2
E[Φ(d1 )] = E[Φ(h(t)(s2 )) + (σ − (h(t)(s2 ))) + (σ − (h(t)(s2 ))2 )]
1! 2!
Φ0 (h(t)(s2 )) Φ00 (h(t)(s2 ))
= Φ(h(t)(s2 )) + E[(σ 2 − (h(t)(s2 )))] + E[(σ 2 − (h(t)(s2 ))2 )]
1! 2!

We can simplify this formula by solving for E[(σ 2 − (h(t)(s2 )))2 ],

E[(σ 2 − (h(t)(s2 )))2 ] = E[σ 4 − 2σ 2 (h(t)(s2 )) + (h(t)(s2 ))2 ]
= E[σ 4 ] − 2(h(t)(s2 ))E[σ 2 ] + (h(t)(s2 ))2
Thus, we can arrive at a simplified general Taylor series expansion for functions of s2 :
1
E[Φ(d1 )] = Φ(h(t)(s2 ))+Φ0 (h(t)(s2 ))[E[σ 2 ]−(h(t)(s2 ))]+ Φ00 (h(t)(s2 ))[E[σ 4 ]−2(h(t)(s2 ))E[σ 2 ]+(h(t)(s2 ))2 ]
2
where
2
S
ln( K ) + (r + σ2 )T
Z d1
1 −d21 /2 d1 = √
Φ(d1 ) = √ e dd1 [8] σ T
−∞ 2π S
σ 2 T − 2 ln( K ) − 2rT
1 −d1 /2
2 d01 = √
0
Φ (d1 ) = √ e 4σ 2 σ 2 T
2π t−1 2
2 E[σ 2 ] = s
−d −d1 /2 t−3
00
Φ (d1 ) = √ 1e ∗ d01 (t − 1)2 s4
2π E[σ 4 ] =
(t − 3)(t − 5)

For the derivation of E[σ 4 ], see Appendix 3. We can use a similar method to find
1
E[Φ(d2 )] = Φ(h(t)(s2 ))+Φ0 (h(t)(s2 ))[E[σ 2 ]−(h(t)(s2 ))]+ Φ00 (h(t)(s2 ))[E[σ 4 ]−2(h(t)(s2 ))E[σ 2 ]+((h(t)(s2 ))2 ]
2
Work is also shown in Appendix 3.

4.2.2 Two Possible Selections

To yield numerical results, we allow h(t)(s2 ) to equal two different functions of s2 . First, we can set
h(t)(s2 ) = s2 , and this way:
Φ0 (s2 ) 2 Φ00 (s2 ) 2
E[Φ(d1 )] = E[Φ(s2 ) + (σ − s2 ) + (σ − s2 )2 + Rn ]
1! 2!
Φ00 (s2 )
= Φ(s2 ) + Φ0 (s2 )E[σ 2 − s2 ] + E[(σ 2 − s2 )2 ]
2!

7
To simplify further we can solve for E[σ 2 − s2 ] and E[(σ 2 − s2 )2 ]:

t−1 2 2s2
E[σ 2 − s2 ] = E[σ 2 ] − s2 = s − s2 =
t−3 t−3

E[(σ 2 − s2 )2 ] = E[σ 4 ] − 2s2 E[σ 2 ] + s4

t−1 2
= E[σ 4 ] − (2s2 )( s ) + s4
t−3
t−1 4
= E[σ 4 ] − 2s + s4
t−3
t+1 4
= E[σ 4 ] − s
t−3
(t − 1)2 t+1 4
= s4 − s
(t − 3)(t − 5) t−3
2s4 (t + 3)
=
(t − 3)(t − 5)

Thus, we arrive at a simplified form for E[Φ(d1 )]:

2s2 s4 (t + 3)
E[Φ(d1 )] = Φ(s2 ) + Φ0 (s2 )( ) + Φ00 (s2 )
t−3 (t − 3)(t − 5)

A similar process can be used to solve for E[Φ(d2 )]:

2s2 s4 (t + 3)
E[Φ(d2 )] = Φ(s2 ) + Φ0 (s2 )( ) + Φ00 (s2 )
t−3 (t − 3)(t − 5)
(t−1) 2
Work is shown in Appendix 4. Another selection of h(t)(s2 ) is (t−3) s . Since E[σ 2 ] = (t−1) 2
(t−3) s , as proved
previously, the second component of the right side of the Taylor series is zero and the computation is
simplified. Thus, E[Φ(d1 )] is given by:

t−1 2 Φ0 ( t−1 2
t−3 s ) 2 t−1 2 Φ00 ( t−1 2
t−3 s ) 2 t−1 2 2
E[Φ(d1 )] = E[Φ( s )+ [σ − ( s )] + [σ − ( s )] ]
t−3 1! t−3 2! t−3
t−1 2 Φ00 ( t−1 2
t−3 s ) t−1 2 2
= Φ( s )+ E[(σ 2 − ( s )) ]
t−3 2! t−3

8
To simplify further we only need to solve for E[(σ 2 − ( t−1 2 2
t−3 s )) ]:

t−1 2 2 t−1 2 t−1 2

E[(σ 2 − ( s )) ] = E[(σ 2 − ( s ))(σ 2 − ( s ))]
t−3 t−3 t−3
t−1 2 (t − 1)2 4
= E[σ 4 − 2( s )E[σ 2 ] + s ]
t−3 (t − 3)2
t−1 2 t−1 2 (t − 1)2 4
= E[σ 4 − 2( s )( s )+ s ]
t−3 t−3 (t − 3)2
(t − 1)2 s4
= E[σ 4 − ]
(t − 3)2
(t − 1)2 s4 (t − 1)2 s4
= −
(t − 3)(t − 5) (t − 3)2
2s4 (t − 1)2
=
(t − 3)2 (t − 5)

Thus, we arrive at a simplified form for E[Φ(d1 )]:

t−1 2 t−1 2 s4 (t − 1)2
E[Φ(d1 )] = Φ( s ) + φ00 ( s )
t−3 t−3 (t − 3)2 (t − 5)
A similar process can be used to solve for E[Φ(d2 )]:
t−1 2 t−1 2 s4 (t − 1)2
E[Φ(d2 )] = Φ( s ) + φ00 ( s )
t−3 t−3 (t − 3)2 (t − 5)
Work is shown in Appendix 4.

4.3 Modeling Volatility [3]

Due to the highly volatile data, the volatility of time series data usually tends to change as time progresses.
Hence, we would like to set the volatility as a volatility model. Meanwhile, the volatility model can be used
to account for volatility clustering too.

4.3.1 ARCH
A widely used volatility model is the Autoregressive Conditional Heteroscedasticity (ARCH) model. The
general ARCH model is of the form:
p
X
ARCH(p) : σt2 = α0 + 2
αi zt−i
i=1

where zt ∼ N (0, σt2 )

and α is a coefficient between 0 and 1. The volatility at time t depends on some random
effects from the past.
We are able to solve for E[σt2 ] and arrive at a solution dependent on only the coefficients [6].
p
X p
X p
X p
X p
X
E[σt2 ] = E[α0 + 2
αi zt−i ] = α0 + αi 2
E[zt−i ] = α0 + αi 2
E[σt−i ]
i=1 i=1 i=1 i=1 i=1
p p p
X X X α0
= α0 + αi E[σt2 ] = α0 + ( αi )E[σt2 ] = p
P
i=1 i=1 i=1 1− αi
i=1

9
The values of the coefficients, αi , can be computed by the statistical analysis software R.
α0
For numerical testing, we derived E[σt2 ] for the ARCH(1) model, σt2 = α0 + α1 zt−12
: E[σt2 ] = 1−α1

4.3.2 GARCH
The ARCH model can be expanded to include the previous volatility. This model is named the Gener-
alized Autoregressive Conditional Heteroscedasticity (GARCH) model. The general form is given by:
p
X q
X
GARCH(p, q) : σt2 = α0 + 2
αi zt−i + 2
βj σt−j
i=1 j=1

where zt ∼ N (0, σt2 ) and αi and βj are coefficients between 0 and 1.

In similar fashion to the ARCH model, we are able to derive the expected value of the GARCH model:
p
X q
X
E[σt2 ] = E[α0 + 2
αi zt−i + 2
βj σt−j ]
i=1 j=1
p
X p
X q
X q
X
2 2
= α0 + αi E[zt−i ]+ βj E[σt−j ]
i=1 i=1 j=1 j=1
Xp Xp q
X q
X
= α0 + αi E[σt2 ] + βj E[σt2 ]
i=1 i=1 j=1 j=1
Xp q
X
= α0 + ( αi )E[σt2 ] + ( βj )E[σt2 ]
i=1 j=1
α0
= p q
P P
1− αi − βj
i=1 j=1

2
For numerical testing, we derived E[σt2 ] for the GARCH(1,1) model, σt2 = α0 + α1 zt−1 2
+ β1 σt−1 :
α0
E[σt2 ] =
1 − α 1 − β1

4.3.3 MACH
The ARCH and GARCH models have been well discussed in the literature. Correspondingly, in this
paper, we propose three new volatility models. The first proposed model is the Moving Average Conditional
Heteroscedasticity (MACH) model defined as:
r
X
M ACH(r) : σt2 = α0 + e2t + θi e2t−i
i=1

where et ∼ N (0, λ2 ) and is i.i.d and α0 and θi are coefficients between 0 and 1.

10
 Similar to the ARCH and GARCH models, we can derive the expression of E[σt2 ].
r
X r
X r
X
E[σt2 ] = E[α0 + e2t + θi e2t−i ] = α0 + E[e2t ] + θi E[e2t−i ]
i=1 i=1 i=1
Xr r
X Xr
= α0 + E[e2t ] + θi E[e2t ] = α0 + E[e2t ] + ( θi )E[e2t ]
i=1 i=1 i=1
Xr r
X
= α0 + (1 + ( θi ))E[e2t ] = α0 + λ2 (1 + θi )
i=1 i=1

4.3.4 ARMACH
By adding random components, the MACH model can be rewritten as the Autoregressive Moving Average
Conditional Heteroscedasticity (ARMACH) model. The general expression is shown as:
r
X p
X
ARM ACH(p, r) : σt2 = α0 + e2t + θi e2t−i + 2
αj zt−j
i=1 j=1

where zt ∼ N (0, σt2 ), et ∼ N (0, λ2 ) and is i.i.d, and αj and θi are coefficients between 0 and 1.
The expectation of the model, E[σt2 ], is still of interest to us. The derivation process is similar.
r
X p
X
E[σt2 ] = E[α0 + e2t + θi e2t−i + 2
αj zt−j ]
i=1 j=1
r
X r
X p
X p
X
= α0 + E[e2t ] + θi E[e2t−i ] + αj 2
E[zt−j ]
i=1 i=1 j=1 j=1
r
X r
X p
X p
X
= α0 + E[e2t ] + θi E[e2t ] + αj E[σt2 ]
i=1 i=1 j=1 j=1
r
X p
X
= α0 + E[e2t ] + ( θi )E[e2t ] + ( αj )E[σt2 ]
i=1 j=1
r
α0 + λ2 (1 +
P
θi )
i=1
= p
P
1− αj
j=1

4.3.5 GARMACH
Combining the GARCH model and the ARMACH model, our third proposed model is the Generalized
Autoregressive Moving Average Conditional Heteroscedasticity (GARMACH) model. It accounts for pre-
vious random components with fixed variance, previous volatilities, and previous random components with
variance of volatility for the volatility. This model is defined as:
r
X p
X q
X
GARM ACH(p, q, r) : σt2 = α0 + e2t + θi e2t−i + 2
αj zt−j + 2
βk σt−k
i=1 j=1 k=1

where zt ∼ N (0, σt2 ), et ∼ N (0, λ2 ) and is i.i.d, and αj , βk , and θi are coefficients between 0 and 1.

11
 We are able to solve for E[σt2 ] in a similar way as the MACH(r) and ARMACH(p,r) models.
r
X p
X q
X
E[σt2 ] = α0 + e2t + θi e2t−i + 2
αj zt−j + 2
βk σt−k
i=1 j=1 k=1
r
X r
X p
X p
X q
X q
X
= α0 + E[e2t ] + θi E[e2t−i ] + αj 2
E[zt−j ]+ βk 2
E[σt−k ]
i=1 i=1 j=1 j=1 k=1 k=1
r
X r
X p
X p
X q
X q
X
= α0 + E[e2t ] + θi E[e2t ] + αj E[σt2 ] + βk E[σt2 ]
i=1 i=1 j=1 j=1 k=1 k=1
r
X p
X q
X
= α0 + E[e2t ] + ( θi )E[e2t ] + ( αj )E[σt2 ] + ( βk )E[σt2 ]
i=1 j=1 k=1
Xr p
X q
X
2
= α0 + (1 + ( θi ))E[et ] + (( αj ) + ( βk ))E[σt2 ]
i=1 j=1 k=1
r
α0 + λ2 (1 +
P
θi )
i=1
= p q
P P
1− αj − βk
j=1 k=1

5 Nonparametric Approaches
Our last approach is nonparametric. The approaches discussed in the previous sections all assume the
use of parametric statistics. In this section, we choose a nonparametric approach where we do not make any
assumption about the data’s probability distribution.

5.1 Average Sample Variance

Rather than making an assumption about the data’s probability distribution, we calculate both average
and weighted sample variances. We start by calculating the average sample variance given by the equation:
2 2 2
c2 = s1 + s2 + · · · + st−1
σt
t−1
where s2i = sample variance of x1 through xi+1 , t = number of returns, and t − 1 = number of sample
variances.

5.2 Weighted Sum of Sample Variances

Although the previous method is simple to work with, it may not be best since sample variances that
include only the earliest data are weighted the same as sample variances that include the most recent data.
In order to weight more heavily sample variances that include more recent data, we propose to estimate the
volatility as a function of weighted sample variances.

c2 = α s2 + α s2 + · · · + α s2
σ t 1 1 2 2 t−1 t−1

t−1
P
where α1 ≤ α2 ≤ · · · ≤ αt−1 , αi = 1, and st is defined as in the previous subsection.
i=1

12
 There are plenty of selections for the weight of αi . For numerical illustration, we propose two series of
weights, which satisfy both conditions with a large enough number of data points:
1 1 1
α1 = , α2 = t−2 , . . . , αt−1 = 1
2t−1 2 2
9 9 9
α1 = t−1 , α2 = t−2 , . . . , αt−1 = 1
10 10 10

6 Numerical Analysis
In this section, we provide numerical comparisons between our different approaches to determine which
approach estimates the price of the options the best. We start with a dataset of S&P 100 daily index series
from January 2, 1991 through December 29, 2000. However, since our second dataset contains 36 different
S&P 100 call options dating back to June 11, 1997, we only use data from the S&P 100 daily index series from
January 2, 1991 through June 11, 1997. Thus, we are left with 1630 observations. The strike price, stock
price, and expiration date vary between different options; however, we use the same three month treasury
yield rate of 4.98% for each option [1].
The observed call prices are compared to the estimated prices obtained from our different approaches.
Most of our approaches do not perform as well as the sample variance method that used an annualized vari-
ance of .01309; however, the GARCH(1,1), single exponential smoothing, and double exponential smoothing
outperform the sample variance method. The strength of each of our approaches is also dependent on the
expiration date and value of the call option. Our approaches improve as the duration of the option contract
decreases since the RMSE is lower for 24 day expiration dates than it is for either 87 or 115 day expiration
dates. This is an expected finding since a shorter time period leads to less uncertainty. Also, as the value of
the call increases, our approaches perform better.
Table 1 displays the sum of squared errors of prediction (SSE) for each approach as well as the root-
mean-square error (RMSE) for each expiration date where:
36
X
SSE = [(actual call price)i − (predicted call price)i ]2
i=1
r
SSE
RM SE(T ) =
# of data points with expiration date T

Table 1: SSE and RMSE

Approaches M ethods SSE RM SE(24) RM SE(87) RM SE(115)
Baseline Sample V ariance 983.79 2.12 5.59 7.33
Smoothing Single 331.83 1.19 3.21 4.33
Double 326.21 1.02 3.14 4.42
T ime Series M odels ARM A(3) 993.70 2.13 5.62 7.36
ARM A(2) 991.85 2.13 5.61 7.36
ARM A(1) 992.87 2.13 5.62 7.36
Deriving the Distribution E(σ 2 ) 985.40 2.12 5.60 7.33
E(σ) 985.17 2.12 5.59 7.33
Φ(s2 ) 983.79 2.12 5.59 7.33
Φ[E(σ 2 )] 985.40 2.12 5.60 7.33
ARCH(1) 991.62 2.13 5.61 7.36
GARCH(1, 1) 867.67 1.96 5.24 6.92
ARCH(2) 995.55 2.14 5.63 7.37
GARCH(2, 2) 982.64 2.12 5.59 7.33
N onparametric Approaches Average Sample V ariance 1350.36 2.59 6.59 8.48
W eighted (αt−1 = .5) 984.25 2.12 5.59 7.33
W eighted (αt−1 = .9) 984.02 2.12 5.59 7.33

13
 Table 2 displays percent bias for each option for the most relevant methods where:
|observed − predicted|
%bias =
observed
Table 2: Percent Bias for Most Relevant Methods Only

T S K Observed Sample V ar % bias GARCH(1, 1) % bias Single % bias Double % bias

24 425.73 395 30.75 32.66 6.21 32.64 6.15 32.60 6.01 32.60 6.01
24 425.73 400 25.88 27.78 7.35 27.74 7.20 27.62 6.73 27.62 6.74
24 425.73 405 21.00 23.03 9.67 22.95 9.30 22.65 7.84 22.65 7.87
24 425.67 410 16.50 18.44 11.76 18.31 10.98 17.61 6.73 17.66 7.02
24 425.68 415 11.88 14.26 20.08 14.07 18.51 12.64 6.47 12.87 8.41
24 425.65 420 7.69 10.52 36.86 10.28 33.73 7.64 0.65 8.45 9.91
24 425.65 425 4.44 7.40 66.69 7.13 60.53 2.92 34.19 4.85 9.13
24 425.68 430 2.10 4.94 135.71 4.66 122.53 0.34 83.90 2.35 12.06
24 425.65 435 0.78 3.09 295.79 2.84 263.87 0.00 99.39 0.92 17.82
24 425.16 440 0.25 1.71 585.43 .52 507.17 0.00 100.00 0.25 1.22
24 424.78 445 0.10 0.89 836.78 0.75 693.17 0.00 100.00 0.05 43.40
24 425.19 450 0.10 0.48 406.57 0.39 309.94 0.00 100.00 0.01 87.82
87 425.73 380 46.75 52.47 12.24 52.40 12.09 52.21 11.67 52.21 11.68
87 425.73 385 42.00 47.73 13.64 47.63 13.40 47.29 12.60 47.30 12.61
87 425.73 390 37.50 43.07 14.86 42.93 14.48 42.38 13.01 42.39 13.04
87 425.73 395 33.00 38.53 16.76 38.34 16.19 37.46 13.52 37.50 13.63
87 425.73 400 28.50 34.14 19.81 33.90 18.95 32.55 14.20 32.64 14.54
87 425.73 405 24.13 29.95 24.10 29.64 22.84 27.63 14.52 27.87 15.48
87 425.26 410 20.38 25.60 25.63 25.23 23.82 22.25 9.20 22.79 11.84
87 425.86 415 16.13 22.36 38.66 21.94 36.06 17.93 11.22 18.92 17.35
87 425.68 420 12.82 18.81 46.76 18.34 43.11 12.85 0.30 14.69 14.65
87 425.42 425 9.32 15.55 66.90 15.04 61.48 7.84 15.79 10.89 16.90
87 425.62 430 6.51 12.91 98.44 12.38 90.36 3.93 39.57 7.95 22.29
87 425.82 435 4.51 10.58 134.86 10.05 123.08 1.39 69.26 5.57 23.55
87 425.68 440 2.75 8.42 206.01 7.90 187.13 0.27 90.15 3.60 30.97
87 425.75 445 1.60 6.66 317.71 6.17 286.73 0.03 98.00 2.25 41.23
87 425.78 450 0.85 5.19 513.80 4.73 459.73 0.00 99.76 1.33 57.28
87 425.39 455 0.44 3.88 781.02 3.47 688.68 0.00 99.99 0.70 59.26
115 425.73 380 47.25 54.75 15.87 54.64 15.63 54.27 14.85 54.27 14.86
115 425.73 390 38.13 45.57 19.50 45.37 18.98 44.49 16.69 44.52 16.77
115 425.73 400 29.38 36.88 25.52 36.57 24.47 34.72 18.17 34.88 18.73
115 425.73 410 21.19 28.91 36.42 28.48 34.39 24.94 17.71 25.61 20.84
115 425.41 420 13.88 21.65 56.00 21.11 52.14 14.86 7.11 16.94 22.12
115 425.63 430 8.13 15.86 95.14 15.26 87.78 5.96 26.69 10.25 26.21
115 425.28 440 3.88 10.91 181.50 10.30 165.86 0.78 79.91 5.22 34.72
115 425.13 450 1.50 7.23 382.09 6.67 344.80 0.02 98.66 2.30 53.16

7 Conclusion
In this paper, we explored different approaches to obtain the volatility embedded in the Black-Scholes formula:
smoothing, deriving the distribution of the volatility, building time series models, and using nonparametric techniques.
Some numerical results were computed to compare the closeness of the estimated call prices with the observed call
prices. We discovered that the GARCH(1,1), single exponential smoothing, and double exponential smoothing
methods are better methods for estimating the volatility than immediately plugging the sample variance back into
the Black-Scholes formula. This paper adds to the literature by proposing new formulas for single and double
exponential smoothing as well as three new time-series models: the MACH, ARMACH, and GARMACH. However,
due to complexity and page limitation at the current stage, the estimation procedures for the parameters in these
models are still under development and were not discussed in this paper. We would like to present the updates of
our proposed models in another paper.

14
References
[1] “Daily Treasury Yield Curve Rates”. U.S. Department of the Treasury. http://www.treasury.gov/resource-
center/data-chart-center/interest- rates/Pages/TextView.aspx?data=yieldYearyear=1997.
[2] Fischer Black and Myron Scholes. “The Pricing of Options and Corporate Liabilities”. The Journal of Political
Economy, pages 637–654, 1973.
[3] J.D. Cryer and K.S. Chan. Time Series Analysis: With Applications in R. Springer Texts in Statistics Series.
Springer-Verlag New York, 2008.
[4] Robert F. Engle. “Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom
inflation”. Econometrica, 50(4):987–1007, 1982.
[5] James Jones. “Stats: Chi-square Distribution”. June 2013. http://people.richland.edu/james/lecture/m170/ch12-
int.html.
[6] Rob Reider. “Volatility forecasting I: GARCH models”. pages 1–16, 2009.
[7] D.D. Wackerly, W. Mendenhall, and R.L. Scheaffer. Mathematical Statistics With Applications. Thomson
Brooks/Cole, 2008.
[8] Eric W. Weisstein. “Normal Distribution”. MathWorld-A Wolfram Web Resource.
http://mathworld.wolfram.com/NormalDistribution.html.

Acknowledgements
This research is supported by National Science Foundation Grant DMS-1262852

15
Appendix 1
Computing variance for ARMA(2,2)

xt = φ1 xt−1 + φ2 xt−2 + θ1 et−1 + θ2 et−2 + et

var(xt ) = cov(φ1 xt−1 , φ1 xt−1 ) + cov(φ1 xt−1 , φ2 xt−2 ) + cov(φ1 xt−1 , θ1 et−1 )
+ cov(φ1 xt−1 , θ2 et−2 ) + cov(φ1 xt−1 , et ) + cov(φ2 xt−2 , φ1 xt−1 )
+ cov(φ2 xt−2 , φ2 xt−2 ) + cov(φ2 xt−2 , θ1 et−1 ) + cov(φ2 xt−2 , θ2 et−2 ) + cov(φ2 xt−2 , et )
+ cov(θ1 et−1 , φ1 xt−1 ) + cov(θ1 et−1 , φ2 xt−2 )+
+ cov(θ1 et−1 , θ1 et−1 ) + cov(θ1 et−1 , θ2 et−2 ) + cov(θ1 et−1 , et )
+ cov(θ2 et−2 , φ1 xt−1 ) + cov(θ2 et−2 , φ2 xt−2 ) + cov(θ2 et−2 , θ1 et−1 )
+ cov(θ2 et−2 , θ2 et−2 ) + cov(θ2 et−2 , et ) + cov(et , φ1 xt−1 ) + cov(et , φ2 xt−2 )
+ cov(et , θ1 et−1 ) + cov(et , θ2 et−2 ) + cov(et , et )
= cov(φ1 xt−1 , φ1 xt−1 ) + cov(φ1 xt−1 , θ1 et−1 ) + cov(φ1 xt−1 , θ2 et−2 )
+ cov(φ2 xt−2 , φ2 xt−2 ) + cov(φ2 xt−2 , θ2 et−2 ) + cov(θ1 et−1 , φ1 xt−1 )
+ cov(θ1 et−1 , θ1 et−1 ) + cov(θ2 et−2 , φ1 xt−1 ) + cov(θ2 et−2 , φ2 xt−2 )
+ cov(θ2 et−2 , θ2 et−2 ) + cov(et , et )
= φ21 var(xt ) + φ22 var(xt ) + θ12 σ 2 + θ22 σ 2 + σ 2 + 2φ1 θ1 σ 2
+ 2φ2 θ2 σ 2 + 2φ21 θ2 σ 2 + 2φ1 θ1 θ2 σ 2

σ 2 [(θ12 + θ22 + 1) + 2(φ1 θ1 + φ2 θ2 + φ1 θ1 θ2 )]

var(xt ) =
1 − φ21 − φ22

Computing variance for ARMA(3,3)

xt = φ1 xt−1 + φ2 xt−2 + φ3 xt−3 + θ1 et−1 + θ2 et−2 + θ3 et−3 + et

var(xt ) = cov(φ1 xt−1 , φ1 xt−1 ) + cov(φ1 xt−1 , φ2 xt−2 ) + cov(φ1 xt−1 , φ3 xt−3 ) + cov(φ1 xt−1 , θ1 et−1 )
+ cov(φ1 xt−1 , θ2 et−2 ) + cov(φ1 xt−1 , θ3 et−3 ) + cov(φ1 xt−1 , et ) + cov(φ2 xt−2 , φ1 xt−1 )
+ cov(φ2 xt−2 , φ2 xt−2 ) + cov(φ2 xt−2 , φ3 xt−3 ) + cov(φ2 xt−2 , θ1 et−1 ) + cov(φ2 xt−2 , θ2 et−2 )
+ cov(φ2 xt−2 , θ3 et−3 ) + cov(φ2 xt−2 , et ) + cov(φ3 xt−3 , φ1 xt−1 ) + cov(φ3 xt−3 , φ2 xt−2 )
+ cov(φ3 xt−3 , φ3 xt−3 ) + cov(φ3 xt−3 , θ1 et−1 ) + cov(φ3 xt−3 , θ2 et−2 ) + cov(φ3 xt−3 , θ3 et−3 )
+ cov(φ3 xt−3 , et ) + cov(θ1 et−1 , φ1 xt−1 ) + cov(θ1 et−1 , φ2 xt−2 ) + cov(θ1 et−1 , φ3 xt−3 )
+ cov(θ1 et−1 , θ1 et−1 ) + cov(θ1 et−1 , θ2 et−2 ) + cov(θ1 et−1 , θ3 et−3 ) + cov(θ1 et−1 , et )
+ cov(θ2 et−2 , φ1 xt−1 ) + cov(θ2 et−2 , φ2 xt−2 ) + cov(θ2 et−2 , φ3 xt−3 ) + cov(θ2 et−2 , θ1 et−1 )
+ cov(θ2 et−2 , θ2 et−2 ) + cov(θ2 et−2 , θ3 et−3 ) + cov(θ2 et−2 , et ) + cov(θ3 et−3 , φ1 xt−1 )
+ cov(θ3 et−3 , φ2 xt−2 ) + cov(θ3 et−3 , φ3 xt−3 ) + cov(θ3 et−3 , θ1 et−1 ) + cov(θ3 et−3 , θ2 et−2 )
+ cov(θ3 et−3 , θ3 et−3 ) + cov(θ3 et−3 , et ) + cov(et , φ1 xt−1 ) + cov(et , φ2 xt−2 ) + cov(et , φ3 xt−3 )
+ cov(et , θ1 et−1 ) + cov(et , θ2 et−2 ) + cov(et , θ3 et−3 ) + cov(et , et )
= cov(φ1 xt−1 , φ1 xt−1 ) + cov(φ2 xt−2 , φ2 xt−2 ) + cov(φ3 xt−3 , φ3 xt−3 ) + cov(θ1 et−1 , θ1 et−1 )
+ cov(θ2 et−2 , θ2 et−2 ) + cov(θ3 et−3 , θ3 et−3 ) + cov(et , et ) + 2cov(φ1 xt−1 , φ2 xt−2 )
+ 2cov(φ1 xt−1 , φ3 xt−3 ) + 2cov(φ1 xt−1 , θ1 et−1 ) + 2cov(φ1 xt−1 , θ2 et−2 ) + 2cov(φ1 xt−1 , θ3 et−3 )
+ 2cov(φ2 xt−2 , φ3 xt−3 ) + 2cov(φ2 xt−2 , θ2 et−2 ) + 2cov(φ2 xt−2 , θ3 et−3 ) + 2cov(φ3 xt−3 , θ3 et−3 )

16
 = φ21 var(xt−1 ) + φ22 var(xt−2 ) + φ23 var(xt−3 ) + θ12 var(et−1 ) + θ22 var(et t − 2) + θ32 var(et−3 )
+ var(et ) + 2φ1 θ1 σ 2 + 2φ2 θ2 σ 2 + φ3 θ3 σ 2 + 2φ1 θ2 σ 2 (φ1 + θ1 )
+ 2φ2 θ3 σ 2 (φ1 + θ1 ) + 2φ1 θ3 σ 2 (φ1 (φ1θ 1 ) + φ2 + θ2 )

σ 2 [(θ12 + θ22 + θ32 + 1) + 2[(φ1 θ1 + φ2 θ2 + φ3 θ3 ) + [(φ1 θ1 )(φ1 θ2 + φ2 θ1 + φ21 θ3 )] + (φ1 φ2 θ3 + φ1 θ2 θ3 )]]

var(xt ) =
1 − φ21 − φ22 − φ23

Appendix 2
Finding E[σ 2 ]

Z ∞
E[u] = fu (u) · u du
0
2 t−1 −(t−1)s2
∞
[ (t−1)s ]( )
e( )
Z 2 2u
u
= t−1 · u du
0 2( 2
)
Γ( t−1
2
)u
2 t−1 −(t−1)s2
∞
[ (t−1)s ]( )
e( )
Z 2 2u
u
= du
( t−1
0 2 2
)
Γ( t−1
2
)

(t − 1)s2
u=
y
−(t − 1)s2
du = dy
y2

t−1 −y
∞
−(t − 1)s2
Z
y 2 e 2
E[u] = t−1 · dy
0 2 2 Γ( t−1 ) y2
2
v −y
Z ∞
y2e 2
= −(t − 1)s2 v dy
0 2 2 Γ( v2 )
2 Z ∞
−(t − 1)s v −y
= v y2e 2 dy
2 2 Γ( v2 ) 0

−(t − 1)s2 v v
= v v
· [−2 2 −1 Γ( − 1)]
2 Γ( 2 )
2 2
(t − 1)s2
=
2( v2 − 1)
(t − 1)s2
=
(t − 3)

Finding E[σ]

(t − 1)s2
y=
σ2
u=σ
v =t−1
dh−1
fu (u) = fy [h−1 (u)]| |
du

17
 (t − 1)s2
h−1 (u) =
u2
−1
dh −2(t − 1)s2
=
du u3
2 t−1 −(t−1)s2
( )
−1 [ (t−1)s
u2
]( 2
−1)
e 2u2
fy [h (u)] =
( t−1
2 2
)
Γ( t−1
2
)

2 t−1 −(t−1)s2
( )
[ (t−1)s
u2
]( 2
−1)
e 2u2 −2(t − 1)s2
fu (u) = ·
( t−1 u3
2 2
)
Γ( t−1
2
)
2 t−1 −(t−1)s2
( )
2[ (t−1)s
u2
]( 2
)
e 2u2
=
( t−1
u2 2
)
Γ( t−1
2
)

Z ∞
E[u] = fu (u) · u du
0
2 t−1 −(t−1)s2
( )
∞ 2[ (t−1)s ]( )
e 2u2
Z 2
u2
= t−1 · u du
0 u2( 2
)
Γ( t−1
2
)
2 t−1 −(t−1)s2
( )
∞ 2[ (t−1)s ]( )
e 2u2
Z 2
u2
= t−1
0 2( 2
)
Γ( t−1
2
)

p
(t − 1)s
u= √
y
√
− t − 1s
du = √ dy
2y y

Z ∞ t−1 −y √
2y 2 e 2 − t − 1s
E[u] = t−1 · √ dy
0 Γ( t−12 ) 2 2y y
2
Z ∞ v − 3 −y
√ y2 2e 2
= − t − 1s v
0 2 2 Γ( v2 )
√
− t − 1s ∞ v2 − 32 −y
Z
= v y e 2 dy
2 2 Γ( v2 ) 0
√
− t − 1s v 1 v 1
= v v
· [−(2 2 − 2 Γ( − )]
2 Γ( 2 )
2 2 2
√
t − 1sΓ( v2 − 12 )
= √
2Γ( v2 )
√
t − 1s( t−4 )!
= √ t−3 2
2( 2 )!

√
t − 1s( t−4 )!
E[σ] = √ t−3 2
2( 2 )!

18
Appendix 3
Derivation of E[σ 4 ]

E[σ 4 ] = E[(σ 2 )2 ] = E[u2 ]

Z ∞
E[u2 ] = fu (u) · u2 du
0
2 t−1 −(t−1)s2
∞
[ (t−1)s ]( )
e( )
Z 2 2u
= u
· u2 du
( t−1
0 2 2
)
Γ( t−1
2
)u
2 t−1 −(t−1)s 2
∞
u[ (t−1)s ]( 2 ) e( 2u )
Z
u
= t−1 du
0 2( 2 ) Γ( t−1
2
)

(t − 1)s2
u=
y
−(t − 1)s2
du = dy
y2

t−1 −y
∞
(t − 1)s2 −(t − 1)s2
Z
2 y 2 e 2
E[u ] = t−1 · · dy
0 Γ( t−1
2 2) y y2
2
Z ∞ v −3 −y
y2 e 2
= −((t − 1)s2 )2 v
0 2 2 Γ( v2 )
−((t − 1)s2 )2 ∞ v2 −3 −y
Z
= v y e 2 dy
2 2 Γ( v2 ) 0

−((t − 1)s2 )2 v v
= v · [−2 2 −2 Γ( − 2)]
2 2 Γ( v2 ) 2

(t − 1)2 s4
E[σ 4 ] =
(t − 3)(t − 5)

General Formula for E[Φ(d2 )]

Φ0 (h(t)(s2 )) 2 Φ00 (h(t)(s2 )) 2

Φ(d2 ) = Φ(h(t)(s2 )) + (σ − (h(t)(s2 ))) + (σ − (h(t)(s2 ))2 ) + Rn
1! 2!

Φ0 (h(t)(s2 )) 2 Φ00 (h(t)(s2 )) 2

E[Φ(d2 )] = E[Φ(h(t)(s2 )) + (σ − (h(t)(s2 ))) + (σ − (h(t)(s2 ))2 )]
1! 2!
Φ0 (h(t)(s2 )) Φ00 (h(t)(s2 ))
= Φ(h(t)(s2 )) + E[(σ 2 − (h(t)(s2 )))] + E[(σ 2 − (h(t)(s2 ))2 )]
1! 2!

We can simplify this formula by solving for E[(σ 2 − (h(t)(s2 )))2 ],

E[(σ 2 − (h(t)(s2 )))2 ] = E[σ 4 − 2σ 2 (h(t)(s2 )) + (h(t)(s2 ))2 ]

= E[σ 4 ] − 2(h(t)(s2 ))E[σ 2 ] + (h(t)(s2 ))2

Thus,
1 00
E[Φ(d2 )] = Φ(h(t)(s2 )) + Φ0 (h(t)(s2 ))[E[σ 2 ] − (h(t)(s2 ))] + Φ (h(t)(s2 ))[E[σ 4 ] − 2(h(t)(s2 ))E[σ 2 ] + (h(t)(s2 ))2 ]
2

19
where
d2 2
1 −d2 /2
Z
Φ(d2 ) = √ e dd2
−∞ 2π
2
1 −d2 /2
Φ0 (d2 ) = √ e
2π
2
−d2 −d2 /2
Φ00 (d2 ) = √ e ∗ d02
2π
2
S
ln( K ) + (r − σ2 )T
d2 = √
σ T
S
2 ln( K ) + 2rT + σ 2 T
d02 = − √
4σ 2 σ 2 T
t−1 2
E[σ 2 ] = s
t−3
(t − 1)2 s4
E[σ 4 ] =
(t − 3)(t − 5)

Appendix 4
E[Φ(d2 )] with h(t)(s2 ) = s2

Φ0 (s2 ) 2 Φ00 (s2 ) 2

E[Φ(d2 )] = E[Φ(s2 ) + (σ − s2 ) + (σ − s2 )2 + Rn ]
1! 2!
Φ00 (s2 )
= Φ(s2 ) + Φ0 (s2 )E[σ 2 − s2 ] + E[(σ 2 − s2 )2 ]
2!
To simplify further we can solve for E[σ 2 − s2 ] and E[(σ 2 − s2 )2 ]:

E[σ 2 − s2 ] = E[σ 2 ] − s2
t−1 2
= s − s2
t−3
2s2
=
t−3

E[(σ 2 − s2 )2 ] = E[σ 4 ] − 2s2 E[σ 2 ] + s4

t−1 2
= E[σ 4 ] − (2s2 )( s ) + s4
t−3
t−1 4
= E[σ 4 ] − 2s + s4
t−3
t+1 4
= E[σ 4 ] − s
t−3
(t − 1)2 t+1 4
= s4 − s
(t − 3)(t − 5) t−3
2s4 (t + 3)
=
(t − 3)(t − 5)

20
Thus, we arrive at a simplified form for E[Φ(d2 )]:

2s2 s4 (t + 3)
E[Φ(d2 )] = Φ(s2 ) + Φ0 (s2 )( ) + Φ00 (s2 )
t−3 (t − 3)(t − 5)

E[Φ(d2 )] with h(t)(s2 ) = E[σ 2 ]

t−1 2 Φ0 ( t−3
t−1 2
s ) 2 t−1 2 Φ00 ( t−1
t−3
s2 ) 2 t−1 2 2
E[Φ(d2 )] = E[Φ( s )+ [σ − ( s )] + [σ − ( s )] ]
t−3 1! t−3 2! t−3
t−1 2 t−1 2 t−1 2 Φ00 ( t−1 s2 ) t−1 2 2
= Φ( s ) + Φ0 ( s )E[σ 2 − ( s )] + t−3
E[(σ 2 − ( s )) ]
t−3 t−3 t−3 2! t−3

To simplify further we can solve for E[σ 2 − ( t−1

t−3
s2 )] and E[(σ 2 − ( t−1
t−3
s2 ))2 ]:

t−1 2 t−1 2
E[σ 2 − ( s )] = E[σ 2 ] − ( s )
t−3 t−3
t−1 2 t−1 2
=( s )−( s )
t−3 t−3
=0

t−1 2 2 t−1 2 t−1 2

E[(σ 2 − ( s )) ] = E[(σ 2 − ( s ))(σ 2 − ( s ))]
t−3 t−3 t−3
t−1 2 (t − 1)2 4
= E[σ 4 − 2( s )E[σ 2 ] + s ]
t−3 (t − 3)2
t−1 2 t−1 2 (t − 1)2 4
= E[σ 4 − 2( s )( s )+ s ]
t−3 t−3 (t − 3)2
(t − 1)2 s4
= E[σ 4 − ]
(t − 3)2
(t − 1)2 s4 (t − 1)2 s4
= −
(t − 3)(t − 5) (t − 3)2
2s4 (t − 1)2
=
(t − 3)2 (t − 5)

Thus, we arrive at a simplified form for E[Φ(d2 )]:

t−1 2 t−1 2 s4 (t − 1)2

E[Φ(d2 )] = Φ( s ) + φ00 ( s )
t−3 t−3 (t − 3)2 (t − 5)

21