Advanced Derivative Securities

Chapter 4
Estimating and
Modeling Volatility

Professor Jian Cai

 Outline

1. Statistics Review
2. Estimating a Constant Volatility and Mean
3. GARCH Models
4. Stochastic Volatility Models
5. Smiles and Smirks Again

2 FIN 452 & 537 / Jian Cai / Spring 2025

 Motivation

 Thus far, we have assumed that the volatility of the

underlying asset is constant or varying in a non-random
way during the lifetime of the derivative.

 We will look at models that relax this assumption and

allow the volatility to change randomly. This is very
important, because there is plenty of evidence that
volatilities do change over time in a random way.

 We will first consider the problem of estimating the

volatility, which then leads to a discussion of modeling
a changing volatility.

3 FIN 452 & 537 / Jian Cai / Spring 2025

 Statistics Review (I)

 We begin with a brief review of basic statistics. Given a

random sample 𝑥𝑥1 , … , 𝑥𝑥𝑁𝑁 of size 𝑁𝑁 from a population with
mean 𝜇𝜇 and variance 𝜎𝜎 2 , the best estimate of 𝜇𝜇 is of course
the sample mean
𝑁𝑁
1
𝑥𝑥̅ = � 𝑥𝑥𝑖𝑖 .
𝑁𝑁
𝑖𝑖=1

 The unbiased estimator of the variance 𝜎𝜎 2 is

𝑁𝑁 𝑁𝑁
1 1
𝑠𝑠 2 = � 𝑥𝑥𝑖𝑖 − 𝑥𝑥̅ 2 = � 𝑥𝑥𝑖𝑖2 − 𝑁𝑁𝑥𝑥̅ 2 .
𝑁𝑁 − 1 𝑁𝑁 − 1
𝑖𝑖=1 𝑖𝑖=1

4 FIN 452 & 537 / Jian Cai / Spring 2025

 Statistics Review (II)
 It is important to know how much variation there would be in 𝑥𝑥̅
if one had access to multiple random samples. More variation
means that an 𝑥𝑥̅ computed from a single sample will be a less
reliable estimate of 𝜇𝜇.

 The standard deviation of 𝑥𝑥̅ in repeated samples, called the

“standard error” of 𝑥𝑥,̅ is 𝜎𝜎⁄ 𝑁𝑁. It is estimated by
𝑁𝑁
1
𝑠𝑠⁄ 𝑁𝑁 = � 𝑥𝑥𝑖𝑖2 − 𝑁𝑁𝑥𝑥̅ 2 .
𝑁𝑁 𝑁𝑁 − 1
𝑖𝑖=1

 By the Central Limit Theorem, 𝑥𝑥̅ will be approximately normally

distributed if the sample size 𝑁𝑁 is large enough regardless the
actual distribution of 𝑥𝑥.

5 FIN 452 & 537 / Jian Cai / Spring 2025

 Geometric Brownian Motion Asset Price
 Consider an asset price following
d𝑆𝑆
= 𝜇𝜇 d𝑡𝑡 + 𝜎𝜎 d𝐵𝐵 ,
𝑆𝑆
where 𝜇𝜇 and 𝜎𝜎 are unknown constants, and 𝐵𝐵 is a Brownian
motion.
 Suppose we have observed the asset price 𝑆𝑆 at dates 0 = 𝑡𝑡0
< 𝑡𝑡1 < ⋯ < 𝑡𝑡𝑁𝑁 = 𝑇𝑇, where 𝑡𝑡𝑖𝑖 − 𝑡𝑡𝑖𝑖−1 = ∆𝑡𝑡. The annualized
continuously compounded rate of return 𝑟𝑟𝑖𝑖 is defined by
𝑆𝑆 𝑡𝑡𝑖𝑖
= 𝑒𝑒 𝑟𝑟𝑖𝑖 ∆𝑡𝑡 .
𝑆𝑆 𝑡𝑡𝑖𝑖−1
 This implies that
1 2 𝐵𝐵 𝑡𝑡𝑖𝑖 − 𝐵𝐵 𝑡𝑡𝑖𝑖−1
𝑦𝑦𝑖𝑖 ≡ 𝑟𝑟𝑖𝑖 ∆𝑡𝑡 = 𝜇𝜇 − 𝜎𝜎 ∆𝑡𝑡 + σ . (1)
2 ∆𝑡𝑡

6 FIN 452 & 537 / Jian Cai / Spring 2025

 Estimation
 The sample 𝑦𝑦1 , … , 𝑦𝑦𝑁𝑁 is a sample of independent random
variables each of which is normally distributed with mean
𝜇𝜇 − 𝜎𝜎 2 ⁄2 ∆𝑡𝑡 and variance 𝜎𝜎 2 .
 The best estimate of the mean of 𝑦𝑦 is the sample mean
𝑁𝑁
1
𝑦𝑦� = � 𝑦𝑦𝑖𝑖 ,
𝑁𝑁
𝑖𝑖=1
the best estimate of 𝜎𝜎 2 is
𝑁𝑁
1
𝜎𝜎� 2 = � 𝑦𝑦𝑖𝑖 − 𝑦𝑦� 2
,
𝑁𝑁 − 1
𝑖𝑖=1
and the best estimate of 𝜇𝜇 is
𝑦𝑦� 1 2 1 2
𝜇𝜇� = + 𝜎𝜎� = 𝑟𝑟̅ + 𝜎𝜎� .
∆𝑡𝑡 2 2
7 FIN 452 & 537 / Jian Cai / Spring 2025
 Unreliability of the Estimate of Mean
 Notice that
𝑁𝑁
∑𝑖𝑖=1 log 𝑆𝑆 𝑡𝑡𝑖𝑖 − log 𝑆𝑆 𝑡𝑡𝑖𝑖−1 log 𝑆𝑆 𝑇𝑇 − log 𝑆𝑆 0
𝑟𝑟̅ = = . (2)
𝑁𝑁∆𝑡𝑡 𝑇𝑇

 Therefore, the reliability of this component cannot depend on

how frequently we observe 𝑆𝑆 within the time period 0, 𝑇𝑇 .

 The standard deviation of 𝑟𝑟̅ in repeated samples is 𝜎𝜎⁄ 𝑇𝑇,

which is likely to be quite large.

 For example, with 𝜎𝜎 = 0.3 and 𝑇𝑇 = 10, the standard deviation

of 𝑟𝑟̅ is 9.5%, which means that a 95% confidence interval will
be a band of roughly 38%.

8 FIN 452 & 537 / Jian Cai / Spring 2025

 Reliability of the Estimate of Variance
 Fortunately, it is easier to estimate 𝜎𝜎. We observed that 𝜎𝜎� 2
defined above can be calculated as
∑ 𝑁𝑁 2 2
𝑖𝑖=1 𝑦𝑦𝑖𝑖 𝑁𝑁 𝑦𝑦
� (3)
𝜎𝜎� 2 = − .
𝑁𝑁 − 1 𝑁𝑁 − 1

 As ∆𝑡𝑡 decreases to 0, 𝑦𝑦� = 𝑟𝑟̅ ∆𝑡𝑡 goes to 0 and the second term
in (3) will be negligible. So if ∆𝑡𝑡 is small, 𝜎𝜎� 2 is approximately
∑𝑁𝑁 2
𝑖𝑖=1 𝑦𝑦𝑖𝑖 𝑁𝑁 1 𝑁𝑁
= × ×� log 𝑆𝑆 𝑡𝑡𝑖𝑖 − log 𝑆𝑆 𝑡𝑡𝑖𝑖−1 2 , (4)
𝑁𝑁 − 1 𝑁𝑁 − 1 𝑇𝑇 𝑖𝑖=1
which converges to 𝜎𝜎 2 as ∆𝑡𝑡 → 0 and 𝑁𝑁 → ∞.

 Thus, in theory, we can estimate 𝜎𝜎 2 with any desired degree of

precision by simply observing 𝑆𝑆 sufficiently frequently.

9 FIN 452 & 537 / Jian Cai / Spring 2025

 A GARCH Model
 Consider
1 2
log 𝑆𝑆 𝑡𝑡𝑖𝑖 − log 𝑆𝑆 𝑡𝑡𝑖𝑖−1 = 𝑟𝑟 − 𝑞𝑞 − 𝜎𝜎𝑖𝑖 ∆𝑡𝑡 + 𝜎𝜎𝑖𝑖 ∆𝐵𝐵 , (5)
2
and redefine
1
log 𝑆𝑆 𝑡𝑡𝑖𝑖 − log 𝑆𝑆 𝑡𝑡𝑖𝑖−1 − 𝑟𝑟 − 𝑞𝑞 − 𝜎𝜎𝑖𝑖2 ∆𝑡𝑡
2
𝑦𝑦𝑖𝑖 ≡ .
∆𝑡𝑡
Thus 𝑦𝑦𝑖𝑖 is normally distributed with mean 0 and variance 𝜎𝜎𝑖𝑖2 .

 We assume that the volatility 𝜎𝜎𝑖𝑖 between dates 𝑡𝑡𝑖𝑖−1 and 𝑡𝑡𝑖𝑖 is
given by
𝜎𝜎𝑖𝑖2 = κ𝜃𝜃 + 1 − κ 2
1 − λ 𝑦𝑦𝑖𝑖−1 2
+ λ𝜎𝜎𝑖𝑖−1 , (6)
where κ ∈ 0, 1 , 𝜃𝜃 > 0, and λ ∈ 0, 1 are all constants.
10 FIN 452 & 537 / Jian Cai / Spring 2025
 A GARCH Model (Continued)
 This is called a GARCH(1,1) variance model (GARCH for
“Generalized Autoregressive Conditional Heteroskedasticity”).
2
𝜃𝜃 is the unconditional variance (𝐸𝐸𝑡𝑡𝑖𝑖 𝜎𝜎𝑖𝑖+𝑛𝑛 → 𝜃𝜃, as 𝑛𝑛 → ∞) and
𝜎𝜎𝑖𝑖2 is the conditional variance of 𝑦𝑦𝑖𝑖 (𝐸𝐸𝑡𝑡𝑖𝑖−1 𝑦𝑦𝑖𝑖2 = 𝜎𝜎𝑖𝑖2 ).

 The most interesting feature of the GARCH(1,1) model is that

large returns (in absolute value) lead to an increase in the
variance and hence are likely to be followed by more large
returns (whether positive or negative).

 This captures the phenomenon of “volatility clustering.” This

feature also implies that the distribution of returns will be “fat
tailed” (the probability of extreme returns is higher than under
a normal distribution with the same standard deviation).

11 FIN 452 & 537 / Jian Cai / Spring 2025

 A GARCH Model: Simulation
 We can simulate a path of an asset price that follows a
GARCH process and the path of its volatility as follows.
Sub Simulating_GARCH()
Dim S, sigma, r, q, dt, theta, kappa, lambda, LogS, Sqrdt
Dim a, b, c, y, i, N
S = InputBox("Enter initial stock price")
sigma = InputBox("Enter initial volatility")
r = InputBox("Enter risk-free rate")
q = InputBox("Enter dividend yield")
dt = InputBox("Enter length of each time period (Delta t)")
N = InputBox("Enter number of time periods (N)")
theta = InputBox("Enter theta")
kappa = InputBox("Enter kappa")
lambda = InputBox("Enter lambda")
LogS = Log(S)
Sqrdt = Sqr(dt)
a = kappa * theta
b = (1 - kappa) * (1 - lambda)
c = (1 - kappa) * lambda

12 FIN 452 & 537 / Jian Cai / Spring 2025

 A GARCH Model: Simulation (Continued)
ActiveCell.Value = "Time"
ActiveCell.Offset(0, 1) = "Stock Price"
ActiveCell.Offset(0, 2) = "Volatility"
ActiveCell.Offset(1, 0) = 0 ' initial time
ActiveCell.Offset(1, 1) = S ' initial stock price
ActiveCell.Offset(1, 2) = sigma ' initial volatility
For i = 1 To N
ActiveCell.Offset(i + 1, 0) = i * dt ' next time
y = sigma * RandN()
LogS = LogS + (r - q - 0.5 * sigma * sigma) * dt + Sqrdt * y
S = Exp(LogS)
ActiveCell.Offset(i + 1, 1) = S ' next stock price
sigma = Sqr(a + b * y ^ 2 + c * sigma ^ 2)
ActiveCell.Offset(i + 1, 2) = sigma ' next volatility
Next i
End Sub

 To price European options, we need to compute the usual

probabilities prob𝑉𝑉 𝑆𝑆 𝑇𝑇 ≥ 𝐾𝐾 and prob𝑅𝑅 𝑆𝑆 𝑇𝑇 ≥ 𝐾𝐾 . We will
show in Chapter 5 how to apply Monte-Carlo methods to
compute these probabilities.
13 FIN 452 & 537 / Jian Cai / Spring 2025
 Stochastic Volatility Models
 The volatility is stochastic (random) in a GARCH model, but it
is determined by the changes in the stock price.

 We will now consider models in which the volatility depends on

a second Brownian motion. The most popular model of this
type is Heston [34] (RFS, 1993):
1 2
d log 𝑆𝑆 = 𝑟𝑟 − 𝑞𝑞 − 𝜎𝜎 𝑡𝑡 d𝑡𝑡 + 𝜎𝜎 𝑡𝑡 d𝐵𝐵𝑠𝑠 , (7a)
2
where 𝐵𝐵𝑠𝑠 is a Brownian motion under the risk-neutral measure
and 𝜎𝜎 𝑡𝑡 = 𝑣𝑣 𝑡𝑡 , where
d𝑣𝑣 𝑡𝑡 = κ 𝜃𝜃 − 𝑣𝑣 𝑡𝑡 d𝑡𝑡 + 𝛾𝛾 𝑣𝑣 𝑡𝑡 d𝐵𝐵𝑣𝑣 , (7b)
where Brownian motion 𝐵𝐵𝑣𝑣 has a constant correlation 𝜌𝜌 with
𝐵𝐵𝑠𝑠 , and κ, 𝜃𝜃, and 𝛾𝛾 are positive constants.

14 FIN 452 & 537 / Jian Cai / Spring 2025

 Stochastic Volatility Models (Continued)
 The correlation 𝜌𝜌 is typically negative, which implies that a
large shock in 𝐵𝐵𝑠𝑠 (hence a large stock return) would likely drive
the volatility down and thus the return for the next period
would also tend to be high (“fat tail”).

 The volatility is “mean revert” with unconditional mean of 𝜃𝜃,

and κ determines the mean reverting speed (as in GARCH).

 As 𝑣𝑣 𝑡𝑡 goes to 0, the volatility of 𝑣𝑣 𝑡𝑡 goes to 0 and the drift

goes to κ𝜃𝜃 > 0. So 𝑣𝑣 𝑡𝑡 is always positive.

 We could discretize (7) as:

1
log 𝑆𝑆 𝑡𝑡𝑖𝑖+1 = log 𝑆𝑆 𝑡𝑡𝑖𝑖 + 𝑟𝑟 − 𝑞𝑞 − 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝑡𝑡 + 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝐵𝐵𝑠𝑠 , (8a)
2
𝑣𝑣 𝑡𝑡𝑖𝑖+1 = 𝑣𝑣 𝑡𝑡𝑖𝑖 + κ 𝜃𝜃 − 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝑡𝑡 + 𝛾𝛾 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝐵𝐵𝑣𝑣 . (8b)

15 FIN 452 & 537 / Jian Cai / Spring 2025

 Stochastic Volatility Models (Continued)
 To guarantee that 𝑣𝑣 𝑡𝑡𝑖𝑖+1 will be nonnegative, we simulate the
Heston model as (8a) and
𝑣𝑣 𝑡𝑡𝑖𝑖+1 = max 0, 𝑣𝑣 𝑡𝑡𝑖𝑖 + κ 𝜃𝜃 − 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝑡𝑡 + 𝛾𝛾 𝑣𝑣 𝑡𝑡𝑖𝑖 ∆𝐵𝐵𝑣𝑣 .

 A simple way to simulate the changes ∆𝐵𝐵𝑠𝑠 and ∆𝐵𝐵𝑣𝑣 in the two
correlated Brownian motions is to generate two independent
standard normals 𝑧𝑧1 and 𝑧𝑧2 and take
∆𝐵𝐵𝑠𝑠 = ∆𝑡𝑡𝑧𝑧 and ∆𝐵𝐵𝑣𝑣 = ∆𝑡𝑡𝑧𝑧 ∗ ,
where we define
𝑧𝑧 = 𝑧𝑧1 𝑎𝑎𝑎𝑎𝑎𝑎 𝑧𝑧 ∗ = 𝜌𝜌𝑧𝑧1 + 1 − 𝜌𝜌2 𝑧𝑧2 .

 The random variable 𝑧𝑧 ∗ is also a standard normal, and the

correlation between 𝑧𝑧 and 𝑧𝑧 ∗ is 𝜌𝜌.

16 FIN 452 & 537 / Jian Cai / Spring 2025

 Stochastic Volatility Models: Simulation
Sub Simulating_Stochastic_Volatility()
Dim S, sigma, r, q, dt, theta, kappa, Gamma, rho, LogS, var
Dim Sqrdt, Sqrrho, z1, Z2, Zstar, i, N
S = InputBox("Enter initial stock price")
sigma = InputBox("Enter initial volatility")
r = InputBox("Enter risk-free rate")
q = InputBox("Enter dividend yield")
dt = InputBox("Enter length of each time period (Delta t)")
N = InputBox("Enter number of time periods (N)")
theta = InputBox("Enter theta")
kappa = InputBox("Enter kappa")
Gamma = InputBox("Enter gamma")
rho = InputBox("Enter rho")
LogS = Log(S)
var = sigma * sigma
Sqrdt = Sqr(dt)
Sqrrho = Sqr(1 - rho * rho)
ActiveCell.Value = "Time"
ActiveCell.Offset(0, 1) = "Stock Price"
ActiveCell.Offset(0, 2) = "Volatility"
ActiveCell.Offset(1, 0) = 0 ' initial time
ActiveCell.Offset(1, 1) = S ' initial stock price
ActiveCell.Offset(1, 2) = sigma ' initial volatility

17 FIN 452 & 537 / Jian Cai / Spring 2025

 Stochastic Volatility Models: Simulation
(Continued)
For i = 1 To N
ActiveCell.Offset(i + 1, 0) = i * dt ' next time
z1 = RandN()
LogS = LogS + (r - q - 0.5 * sigma * sigma) * dt + sigma * Sqrdt * z1
S = Exp(LogS)
ActiveCell.Offset(i + 1, 1) = S ' next stock price
Z2 = RandN()
Zstar = rho * z1 + Sqrrho * Z2
var = Application.Max(0, var + kappa * (theta - var) * dt _
+ Gamma * sigma * Sqrdt * Zstar)
sigma = Sqr(var)
ActiveCell.Offset(i + 1, 2) = sigma ' next volatility
Next i
End Sub

 To price European options, we again need to compute the usual

probabilities prob𝑉𝑉 𝑆𝑆 𝑇𝑇 ≥ 𝐾𝐾 and prob𝑅𝑅 𝑆𝑆 𝑇𝑇 ≥ 𝐾𝐾 .
 One can use the same approach as implied volatility to imply all the
parameters 𝜎𝜎 0 , κ, 𝜃𝜃, λ in the GARCH model or 𝜎𝜎 0 , κ, 𝜃𝜃, 𝛾𝛾, 𝜌𝜌 in
the stochastic volatility model from observed option prices.

18 FIN 452 & 537 / Jian Cai / Spring 2025

 Smiles and Smirks Again
 Let 𝜎𝜎am denote the implied volatility from an at-the-money call
option (𝐾𝐾 = 𝑆𝑆 0 ). Recall that “Smiles and Smirks” refer to the
fact that implied volatilities from deep-out-of-money and deep-
in-the-money options are higher than 𝜎𝜎am .

 The high implied volatility means that the market is pricing the
right to buy and sell at 𝐾𝐾 ≫ 𝑆𝑆 0 and 𝐾𝐾 ≪ 𝑆𝑆 0 above the
Black-Scholes price computed from the volatility 𝜎𝜎am . Thus,
the market must attach a higher probability to stock prices
𝑆𝑆 𝑇𝑇 ≫ 𝑆𝑆 0 and 𝑆𝑆 𝑇𝑇 ≪ 𝑆𝑆 0 than the volatility 𝜎𝜎am would
suggest.

 Therefore, the market uses a “fat tailed” distribution for option

pricing, which is why GARCH and stochastic volatility models
fit better the data.

19 FIN 452 & 537 / Jian Cai / Spring 2025